1 | /* |
2 | * Copyright (c) 2005, 2019, Oracle and/or its affiliates. All rights reserved. |
3 | * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. |
4 | * |
5 | * This code is free software; you can redistribute it and/or modify it |
6 | * under the terms of the GNU General Public License version 2 only, as |
7 | * published by the Free Software Foundation. |
8 | * |
9 | * This code is distributed in the hope that it will be useful, but WITHOUT |
10 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or |
11 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License |
12 | * version 2 for more details (a copy is included in the LICENSE file that |
13 | * accompanied this code). |
14 | * |
15 | * You should have received a copy of the GNU General Public License version |
16 | * 2 along with this work; if not, write to the Free Software Foundation, |
17 | * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. |
18 | * |
19 | * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA |
20 | * or visit www.oracle.com if you need additional information or have any |
21 | * questions. |
22 | * |
23 | */ |
24 | |
25 | #include "precompiled.hpp" |
26 | #include "c1/c1_CFGPrinter.hpp" |
27 | #include "c1/c1_CodeStubs.hpp" |
28 | #include "c1/c1_Compilation.hpp" |
29 | #include "c1/c1_FrameMap.hpp" |
30 | #include "c1/c1_IR.hpp" |
31 | #include "c1/c1_LIRGenerator.hpp" |
32 | #include "c1/c1_LinearScan.hpp" |
33 | #include "c1/c1_ValueStack.hpp" |
34 | #include "code/vmreg.inline.hpp" |
35 | #include "runtime/timerTrace.hpp" |
36 | #include "utilities/bitMap.inline.hpp" |
37 | |
38 | #ifndef PRODUCT |
39 | |
40 | static LinearScanStatistic _stat_before_alloc; |
41 | static LinearScanStatistic _stat_after_asign; |
42 | static LinearScanStatistic _stat_final; |
43 | |
44 | static LinearScanTimers _total_timer; |
45 | |
46 | // helper macro for short definition of timer |
47 | #define TIME_LINEAR_SCAN(timer_name) TraceTime _block_timer("", _total_timer.timer(LinearScanTimers::timer_name), TimeLinearScan || TimeEachLinearScan, Verbose); |
48 | |
49 | // helper macro for short definition of trace-output inside code |
50 | #define TRACE_LINEAR_SCAN(level, code) \ |
51 | if (TraceLinearScanLevel >= level) { \ |
52 | code; \ |
53 | } |
54 | |
55 | #else |
56 | |
57 | #define TIME_LINEAR_SCAN(timer_name) |
58 | #define TRACE_LINEAR_SCAN(level, code) |
59 | |
60 | #endif |
61 | |
62 | // Map BasicType to spill size in 32-bit words, matching VMReg's notion of words |
63 | #ifdef _LP64 |
64 | static int type2spill_size[T_CONFLICT+1]={ -1, 0, 0, 0, 1, 1, 1, 2, 1, 1, 1, 2, 2, 2, 0, 2, 1, 2, 1, -1}; |
65 | #else |
66 | static int type2spill_size[T_CONFLICT+1]={ -1, 0, 0, 0, 1, 1, 1, 2, 1, 1, 1, 2, 1, 1, 0, 1, -1, 1, 1, -1}; |
67 | #endif |
68 | |
69 | |
70 | // Implementation of LinearScan |
71 | |
72 | LinearScan::LinearScan(IR* ir, LIRGenerator* gen, FrameMap* frame_map) |
73 | : _compilation(ir->compilation()) |
74 | , _ir(ir) |
75 | , _gen(gen) |
76 | , _frame_map(frame_map) |
77 | , _cached_blocks(*ir->linear_scan_order()) |
78 | , _num_virtual_regs(gen->max_virtual_register_number()) |
79 | , _has_fpu_registers(false) |
80 | , _num_calls(-1) |
81 | , _max_spills(0) |
82 | , _unused_spill_slot(-1) |
83 | , _intervals(0) // initialized later with correct length |
84 | , _new_intervals_from_allocation(NULL) |
85 | , _sorted_intervals(NULL) |
86 | , _needs_full_resort(false) |
87 | , _lir_ops(0) // initialized later with correct length |
88 | , _block_of_op(0) // initialized later with correct length |
89 | , _has_info(0) |
90 | , _has_call(0) |
91 | , _interval_in_loop(0) // initialized later with correct length |
92 | , _scope_value_cache(0) // initialized later with correct length |
93 | #ifdef X86 |
94 | , _fpu_stack_allocator(NULL) |
95 | #endif |
96 | { |
97 | assert(this->ir() != NULL, "check if valid" ); |
98 | assert(this->compilation() != NULL, "check if valid" ); |
99 | assert(this->gen() != NULL, "check if valid" ); |
100 | assert(this->frame_map() != NULL, "check if valid" ); |
101 | } |
102 | |
103 | |
104 | // ********** functions for converting LIR-Operands to register numbers |
105 | // |
106 | // Emulate a flat register file comprising physical integer registers, |
107 | // physical floating-point registers and virtual registers, in that order. |
108 | // Virtual registers already have appropriate numbers, since V0 is |
109 | // the number of physical registers. |
110 | // Returns -1 for hi word if opr is a single word operand. |
111 | // |
112 | // Note: the inverse operation (calculating an operand for register numbers) |
113 | // is done in calc_operand_for_interval() |
114 | |
115 | int LinearScan::reg_num(LIR_Opr opr) { |
116 | assert(opr->is_register(), "should not call this otherwise" ); |
117 | |
118 | if (opr->is_virtual_register()) { |
119 | assert(opr->vreg_number() >= nof_regs, "found a virtual register with a fixed-register number" ); |
120 | return opr->vreg_number(); |
121 | } else if (opr->is_single_cpu()) { |
122 | return opr->cpu_regnr(); |
123 | } else if (opr->is_double_cpu()) { |
124 | return opr->cpu_regnrLo(); |
125 | #ifdef X86 |
126 | } else if (opr->is_single_xmm()) { |
127 | return opr->fpu_regnr() + pd_first_xmm_reg; |
128 | } else if (opr->is_double_xmm()) { |
129 | return opr->fpu_regnrLo() + pd_first_xmm_reg; |
130 | #endif |
131 | } else if (opr->is_single_fpu()) { |
132 | return opr->fpu_regnr() + pd_first_fpu_reg; |
133 | } else if (opr->is_double_fpu()) { |
134 | return opr->fpu_regnrLo() + pd_first_fpu_reg; |
135 | } else { |
136 | ShouldNotReachHere(); |
137 | return -1; |
138 | } |
139 | } |
140 | |
141 | int LinearScan::reg_numHi(LIR_Opr opr) { |
142 | assert(opr->is_register(), "should not call this otherwise" ); |
143 | |
144 | if (opr->is_virtual_register()) { |
145 | return -1; |
146 | } else if (opr->is_single_cpu()) { |
147 | return -1; |
148 | } else if (opr->is_double_cpu()) { |
149 | return opr->cpu_regnrHi(); |
150 | #ifdef X86 |
151 | } else if (opr->is_single_xmm()) { |
152 | return -1; |
153 | } else if (opr->is_double_xmm()) { |
154 | return -1; |
155 | #endif |
156 | } else if (opr->is_single_fpu()) { |
157 | return -1; |
158 | } else if (opr->is_double_fpu()) { |
159 | return opr->fpu_regnrHi() + pd_first_fpu_reg; |
160 | } else { |
161 | ShouldNotReachHere(); |
162 | return -1; |
163 | } |
164 | } |
165 | |
166 | |
167 | // ********** functions for classification of intervals |
168 | |
169 | bool LinearScan::is_precolored_interval(const Interval* i) { |
170 | return i->reg_num() < LinearScan::nof_regs; |
171 | } |
172 | |
173 | bool LinearScan::is_virtual_interval(const Interval* i) { |
174 | return i->reg_num() >= LIR_OprDesc::vreg_base; |
175 | } |
176 | |
177 | bool LinearScan::is_precolored_cpu_interval(const Interval* i) { |
178 | return i->reg_num() < LinearScan::nof_cpu_regs; |
179 | } |
180 | |
181 | bool LinearScan::is_virtual_cpu_interval(const Interval* i) { |
182 | #if defined(__SOFTFP__) || defined(E500V2) |
183 | return i->reg_num() >= LIR_OprDesc::vreg_base; |
184 | #else |
185 | return i->reg_num() >= LIR_OprDesc::vreg_base && (i->type() != T_FLOAT && i->type() != T_DOUBLE); |
186 | #endif // __SOFTFP__ or E500V2 |
187 | } |
188 | |
189 | bool LinearScan::is_precolored_fpu_interval(const Interval* i) { |
190 | return i->reg_num() >= LinearScan::nof_cpu_regs && i->reg_num() < LinearScan::nof_regs; |
191 | } |
192 | |
193 | bool LinearScan::is_virtual_fpu_interval(const Interval* i) { |
194 | #if defined(__SOFTFP__) || defined(E500V2) |
195 | return false; |
196 | #else |
197 | return i->reg_num() >= LIR_OprDesc::vreg_base && (i->type() == T_FLOAT || i->type() == T_DOUBLE); |
198 | #endif // __SOFTFP__ or E500V2 |
199 | } |
200 | |
201 | bool LinearScan::is_in_fpu_register(const Interval* i) { |
202 | // fixed intervals not needed for FPU stack allocation |
203 | return i->reg_num() >= nof_regs && pd_first_fpu_reg <= i->assigned_reg() && i->assigned_reg() <= pd_last_fpu_reg; |
204 | } |
205 | |
206 | bool LinearScan::is_oop_interval(const Interval* i) { |
207 | // fixed intervals never contain oops |
208 | return i->reg_num() >= nof_regs && i->type() == T_OBJECT; |
209 | } |
210 | |
211 | |
212 | // ********** General helper functions |
213 | |
214 | // compute next unused stack index that can be used for spilling |
215 | int LinearScan::allocate_spill_slot(bool double_word) { |
216 | int spill_slot; |
217 | if (double_word) { |
218 | if ((_max_spills & 1) == 1) { |
219 | // alignment of double-word values |
220 | // the hole because of the alignment is filled with the next single-word value |
221 | assert(_unused_spill_slot == -1, "wasting a spill slot" ); |
222 | _unused_spill_slot = _max_spills; |
223 | _max_spills++; |
224 | } |
225 | spill_slot = _max_spills; |
226 | _max_spills += 2; |
227 | |
228 | } else if (_unused_spill_slot != -1) { |
229 | // re-use hole that was the result of a previous double-word alignment |
230 | spill_slot = _unused_spill_slot; |
231 | _unused_spill_slot = -1; |
232 | |
233 | } else { |
234 | spill_slot = _max_spills; |
235 | _max_spills++; |
236 | } |
237 | |
238 | int result = spill_slot + LinearScan::nof_regs + frame_map()->argcount(); |
239 | |
240 | // the class OopMapValue uses only 11 bits for storing the name of the |
241 | // oop location. So a stack slot bigger than 2^11 leads to an overflow |
242 | // that is not reported in product builds. Prevent this by checking the |
243 | // spill slot here (altough this value and the later used location name |
244 | // are slightly different) |
245 | if (result > 2000) { |
246 | bailout("too many stack slots used" ); |
247 | } |
248 | |
249 | return result; |
250 | } |
251 | |
252 | void LinearScan::assign_spill_slot(Interval* it) { |
253 | // assign the canonical spill slot of the parent (if a part of the interval |
254 | // is already spilled) or allocate a new spill slot |
255 | if (it->canonical_spill_slot() >= 0) { |
256 | it->assign_reg(it->canonical_spill_slot()); |
257 | } else { |
258 | int spill = allocate_spill_slot(type2spill_size[it->type()] == 2); |
259 | it->set_canonical_spill_slot(spill); |
260 | it->assign_reg(spill); |
261 | } |
262 | } |
263 | |
264 | void LinearScan::propagate_spill_slots() { |
265 | if (!frame_map()->finalize_frame(max_spills())) { |
266 | bailout("frame too large" ); |
267 | } |
268 | } |
269 | |
270 | // create a new interval with a predefined reg_num |
271 | // (only used for parent intervals that are created during the building phase) |
272 | Interval* LinearScan::create_interval(int reg_num) { |
273 | assert(_intervals.at(reg_num) == NULL, "overwriting exisiting interval" ); |
274 | |
275 | Interval* interval = new Interval(reg_num); |
276 | _intervals.at_put(reg_num, interval); |
277 | |
278 | // assign register number for precolored intervals |
279 | if (reg_num < LIR_OprDesc::vreg_base) { |
280 | interval->assign_reg(reg_num); |
281 | } |
282 | return interval; |
283 | } |
284 | |
285 | // assign a new reg_num to the interval and append it to the list of intervals |
286 | // (only used for child intervals that are created during register allocation) |
287 | void LinearScan::append_interval(Interval* it) { |
288 | it->set_reg_num(_intervals.length()); |
289 | _intervals.append(it); |
290 | IntervalList* new_intervals = _new_intervals_from_allocation; |
291 | if (new_intervals == NULL) { |
292 | new_intervals = _new_intervals_from_allocation = new IntervalList(); |
293 | } |
294 | new_intervals->append(it); |
295 | } |
296 | |
297 | // copy the vreg-flags if an interval is split |
298 | void LinearScan::copy_register_flags(Interval* from, Interval* to) { |
299 | if (gen()->is_vreg_flag_set(from->reg_num(), LIRGenerator::byte_reg)) { |
300 | gen()->set_vreg_flag(to->reg_num(), LIRGenerator::byte_reg); |
301 | } |
302 | if (gen()->is_vreg_flag_set(from->reg_num(), LIRGenerator::callee_saved)) { |
303 | gen()->set_vreg_flag(to->reg_num(), LIRGenerator::callee_saved); |
304 | } |
305 | |
306 | // Note: do not copy the must_start_in_memory flag because it is not necessary for child |
307 | // intervals (only the very beginning of the interval must be in memory) |
308 | } |
309 | |
310 | |
311 | // ********** spill move optimization |
312 | // eliminate moves from register to stack if stack slot is known to be correct |
313 | |
314 | // called during building of intervals |
315 | void LinearScan::change_spill_definition_pos(Interval* interval, int def_pos) { |
316 | assert(interval->is_split_parent(), "can only be called for split parents" ); |
317 | |
318 | switch (interval->spill_state()) { |
319 | case noDefinitionFound: |
320 | assert(interval->spill_definition_pos() == -1, "must no be set before" ); |
321 | interval->set_spill_definition_pos(def_pos); |
322 | interval->set_spill_state(oneDefinitionFound); |
323 | break; |
324 | |
325 | case oneDefinitionFound: |
326 | assert(def_pos <= interval->spill_definition_pos(), "positions are processed in reverse order when intervals are created" ); |
327 | if (def_pos < interval->spill_definition_pos() - 2) { |
328 | // second definition found, so no spill optimization possible for this interval |
329 | interval->set_spill_state(noOptimization); |
330 | } else { |
331 | // two consecutive definitions (because of two-operand LIR form) |
332 | assert(block_of_op_with_id(def_pos) == block_of_op_with_id(interval->spill_definition_pos()), "block must be equal" ); |
333 | } |
334 | break; |
335 | |
336 | case noOptimization: |
337 | // nothing to do |
338 | break; |
339 | |
340 | default: |
341 | assert(false, "other states not allowed at this time" ); |
342 | } |
343 | } |
344 | |
345 | // called during register allocation |
346 | void LinearScan::change_spill_state(Interval* interval, int spill_pos) { |
347 | switch (interval->spill_state()) { |
348 | case oneDefinitionFound: { |
349 | int def_loop_depth = block_of_op_with_id(interval->spill_definition_pos())->loop_depth(); |
350 | int spill_loop_depth = block_of_op_with_id(spill_pos)->loop_depth(); |
351 | |
352 | if (def_loop_depth < spill_loop_depth) { |
353 | // the loop depth of the spilling position is higher then the loop depth |
354 | // at the definition of the interval -> move write to memory out of loop |
355 | // by storing at definitin of the interval |
356 | interval->set_spill_state(storeAtDefinition); |
357 | } else { |
358 | // the interval is currently spilled only once, so for now there is no |
359 | // reason to store the interval at the definition |
360 | interval->set_spill_state(oneMoveInserted); |
361 | } |
362 | break; |
363 | } |
364 | |
365 | case oneMoveInserted: { |
366 | // the interval is spilled more then once, so it is better to store it to |
367 | // memory at the definition |
368 | interval->set_spill_state(storeAtDefinition); |
369 | break; |
370 | } |
371 | |
372 | case storeAtDefinition: |
373 | case startInMemory: |
374 | case noOptimization: |
375 | case noDefinitionFound: |
376 | // nothing to do |
377 | break; |
378 | |
379 | default: |
380 | assert(false, "other states not allowed at this time" ); |
381 | } |
382 | } |
383 | |
384 | |
385 | bool LinearScan::must_store_at_definition(const Interval* i) { |
386 | return i->is_split_parent() && i->spill_state() == storeAtDefinition; |
387 | } |
388 | |
389 | // called once before asignment of register numbers |
390 | void LinearScan::eliminate_spill_moves() { |
391 | TIME_LINEAR_SCAN(timer_eliminate_spill_moves); |
392 | TRACE_LINEAR_SCAN(3, tty->print_cr("***** Eliminating unnecessary spill moves" )); |
393 | |
394 | // collect all intervals that must be stored after their definion. |
395 | // the list is sorted by Interval::spill_definition_pos |
396 | Interval* interval; |
397 | Interval* temp_list; |
398 | create_unhandled_lists(&interval, &temp_list, must_store_at_definition, NULL); |
399 | |
400 | #ifdef ASSERT |
401 | Interval* prev = NULL; |
402 | Interval* temp = interval; |
403 | while (temp != Interval::end()) { |
404 | assert(temp->spill_definition_pos() > 0, "invalid spill definition pos" ); |
405 | if (prev != NULL) { |
406 | assert(temp->from() >= prev->from(), "intervals not sorted" ); |
407 | 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" ); |
408 | } |
409 | |
410 | assert(temp->canonical_spill_slot() >= LinearScan::nof_regs, "interval has no spill slot assigned" ); |
411 | assert(temp->spill_definition_pos() >= temp->from(), "invalid order" ); |
412 | assert(temp->spill_definition_pos() <= temp->from() + 2, "only intervals defined once at their start-pos can be optimized" ); |
413 | |
414 | 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())); |
415 | |
416 | temp = temp->next(); |
417 | } |
418 | #endif |
419 | |
420 | LIR_InsertionBuffer insertion_buffer; |
421 | int num_blocks = block_count(); |
422 | for (int i = 0; i < num_blocks; i++) { |
423 | BlockBegin* block = block_at(i); |
424 | LIR_OpList* instructions = block->lir()->instructions_list(); |
425 | int num_inst = instructions->length(); |
426 | bool has_new = false; |
427 | |
428 | // iterate all instructions of the block. skip the first because it is always a label |
429 | for (int j = 1; j < num_inst; j++) { |
430 | LIR_Op* op = instructions->at(j); |
431 | int op_id = op->id(); |
432 | |
433 | if (op_id == -1) { |
434 | // remove move from register to stack if the stack slot is guaranteed to be correct. |
435 | // only moves that have been inserted by LinearScan can be removed. |
436 | assert(op->code() == lir_move, "only moves can have a op_id of -1" ); |
437 | assert(op->as_Op1() != NULL, "move must be LIR_Op1" ); |
438 | assert(op->as_Op1()->result_opr()->is_virtual(), "LinearScan inserts only moves to virtual registers" ); |
439 | |
440 | LIR_Op1* op1 = (LIR_Op1*)op; |
441 | Interval* interval = interval_at(op1->result_opr()->vreg_number()); |
442 | |
443 | if (interval->assigned_reg() >= LinearScan::nof_regs && interval->always_in_memory()) { |
444 | // move target is a stack slot that is always correct, so eliminate instruction |
445 | TRACE_LINEAR_SCAN(4, tty->print_cr("eliminating move from interval %d to %d" , op1->in_opr()->vreg_number(), op1->result_opr()->vreg_number())); |
446 | instructions->at_put(j, NULL); // NULL-instructions are deleted by assign_reg_num |
447 | } |
448 | |
449 | } else { |
450 | // insert move from register to stack just after the beginning of the interval |
451 | assert(interval == Interval::end() || interval->spill_definition_pos() >= op_id, "invalid order" ); |
452 | assert(interval == Interval::end() || (interval->is_split_parent() && interval->spill_state() == storeAtDefinition), "invalid interval" ); |
453 | |
454 | while (interval != Interval::end() && interval->spill_definition_pos() == op_id) { |
455 | if (!has_new) { |
456 | // prepare insertion buffer (appended when all instructions of the block are processed) |
457 | insertion_buffer.init(block->lir()); |
458 | has_new = true; |
459 | } |
460 | |
461 | LIR_Opr from_opr = operand_for_interval(interval); |
462 | LIR_Opr to_opr = canonical_spill_opr(interval); |
463 | assert(from_opr->is_fixed_cpu() || from_opr->is_fixed_fpu(), "from operand must be a register" ); |
464 | assert(to_opr->is_stack(), "to operand must be a stack slot" ); |
465 | |
466 | insertion_buffer.move(j, from_opr, to_opr); |
467 | 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)); |
468 | |
469 | interval = interval->next(); |
470 | } |
471 | } |
472 | } // end of instruction iteration |
473 | |
474 | if (has_new) { |
475 | block->lir()->append(&insertion_buffer); |
476 | } |
477 | } // end of block iteration |
478 | |
479 | assert(interval == Interval::end(), "missed an interval" ); |
480 | } |
481 | |
482 | |
483 | // ********** Phase 1: number all instructions in all blocks |
484 | // Compute depth-first and linear scan block orders, and number LIR_Op nodes for linear scan. |
485 | |
486 | void LinearScan::number_instructions() { |
487 | { |
488 | // dummy-timer to measure the cost of the timer itself |
489 | // (this time is then subtracted from all other timers to get the real value) |
490 | TIME_LINEAR_SCAN(timer_do_nothing); |
491 | } |
492 | TIME_LINEAR_SCAN(timer_number_instructions); |
493 | |
494 | // Assign IDs to LIR nodes and build a mapping, lir_ops, from ID to LIR_Op node. |
495 | int num_blocks = block_count(); |
496 | int num_instructions = 0; |
497 | int i; |
498 | for (i = 0; i < num_blocks; i++) { |
499 | num_instructions += block_at(i)->lir()->instructions_list()->length(); |
500 | } |
501 | |
502 | // initialize with correct length |
503 | _lir_ops = LIR_OpArray(num_instructions, num_instructions, NULL); |
504 | _block_of_op = BlockBeginArray(num_instructions, num_instructions, NULL); |
505 | |
506 | int op_id = 0; |
507 | int idx = 0; |
508 | |
509 | for (i = 0; i < num_blocks; i++) { |
510 | BlockBegin* block = block_at(i); |
511 | block->set_first_lir_instruction_id(op_id); |
512 | LIR_OpList* instructions = block->lir()->instructions_list(); |
513 | |
514 | int num_inst = instructions->length(); |
515 | for (int j = 0; j < num_inst; j++) { |
516 | LIR_Op* op = instructions->at(j); |
517 | op->set_id(op_id); |
518 | |
519 | _lir_ops.at_put(idx, op); |
520 | _block_of_op.at_put(idx, block); |
521 | assert(lir_op_with_id(op_id) == op, "must match" ); |
522 | |
523 | idx++; |
524 | op_id += 2; // numbering of lir_ops by two |
525 | } |
526 | block->set_last_lir_instruction_id(op_id - 2); |
527 | } |
528 | assert(idx == num_instructions, "must match" ); |
529 | assert(idx * 2 == op_id, "must match" ); |
530 | |
531 | _has_call.initialize(num_instructions); |
532 | _has_info.initialize(num_instructions); |
533 | } |
534 | |
535 | |
536 | // ********** Phase 2: compute local live sets separately for each block |
537 | // (sets live_gen and live_kill for each block) |
538 | |
539 | void LinearScan::set_live_gen_kill(Value value, LIR_Op* op, BitMap& live_gen, BitMap& live_kill) { |
540 | LIR_Opr opr = value->operand(); |
541 | Constant* con = value->as_Constant(); |
542 | |
543 | // check some asumptions about debug information |
544 | assert(!value->type()->is_illegal(), "if this local is used by the interpreter it shouldn't be of indeterminate type" ); |
545 | assert(con == NULL || opr->is_virtual() || opr->is_constant() || opr->is_illegal(), "asumption: Constant instructions have only constant operands" ); |
546 | assert(con != NULL || opr->is_virtual(), "asumption: non-Constant instructions have only virtual operands" ); |
547 | |
548 | if ((con == NULL || con->is_pinned()) && opr->is_register()) { |
549 | assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below" ); |
550 | int reg = opr->vreg_number(); |
551 | if (!live_kill.at(reg)) { |
552 | live_gen.set_bit(reg); |
553 | 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)); |
554 | } |
555 | } |
556 | } |
557 | |
558 | |
559 | void LinearScan::compute_local_live_sets() { |
560 | TIME_LINEAR_SCAN(timer_compute_local_live_sets); |
561 | |
562 | int num_blocks = block_count(); |
563 | int live_size = live_set_size(); |
564 | bool local_has_fpu_registers = false; |
565 | int local_num_calls = 0; |
566 | LIR_OpVisitState visitor; |
567 | |
568 | BitMap2D local_interval_in_loop = BitMap2D(_num_virtual_regs, num_loops()); |
569 | |
570 | // iterate all blocks |
571 | for (int i = 0; i < num_blocks; i++) { |
572 | BlockBegin* block = block_at(i); |
573 | |
574 | ResourceBitMap live_gen(live_size); |
575 | ResourceBitMap live_kill(live_size); |
576 | |
577 | if (block->is_set(BlockBegin::exception_entry_flag)) { |
578 | // Phi functions at the begin of an exception handler are |
579 | // implicitly defined (= killed) at the beginning of the block. |
580 | for_each_phi_fun(block, phi, |
581 | if (!phi->is_illegal()) { live_kill.set_bit(phi->operand()->vreg_number()); } |
582 | ); |
583 | } |
584 | |
585 | LIR_OpList* instructions = block->lir()->instructions_list(); |
586 | int num_inst = instructions->length(); |
587 | |
588 | // iterate all instructions of the block. skip the first because it is always a label |
589 | assert(visitor.no_operands(instructions->at(0)), "first operation must always be a label" ); |
590 | for (int j = 1; j < num_inst; j++) { |
591 | LIR_Op* op = instructions->at(j); |
592 | |
593 | // visit operation to collect all operands |
594 | visitor.visit(op); |
595 | |
596 | if (visitor.has_call()) { |
597 | _has_call.set_bit(op->id() >> 1); |
598 | local_num_calls++; |
599 | } |
600 | if (visitor.info_count() > 0) { |
601 | _has_info.set_bit(op->id() >> 1); |
602 | } |
603 | |
604 | // iterate input operands of instruction |
605 | int k, n, reg; |
606 | n = visitor.opr_count(LIR_OpVisitState::inputMode); |
607 | for (k = 0; k < n; k++) { |
608 | LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, k); |
609 | assert(opr->is_register(), "visitor should only return register operands" ); |
610 | |
611 | if (opr->is_virtual_register()) { |
612 | assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below" ); |
613 | reg = opr->vreg_number(); |
614 | if (!live_kill.at(reg)) { |
615 | live_gen.set_bit(reg); |
616 | TRACE_LINEAR_SCAN(4, tty->print_cr(" Setting live_gen for register %d at instruction %d" , reg, op->id())); |
617 | } |
618 | if (block->loop_index() >= 0) { |
619 | local_interval_in_loop.set_bit(reg, block->loop_index()); |
620 | } |
621 | local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu(); |
622 | } |
623 | |
624 | #ifdef ASSERT |
625 | // fixed intervals are never live at block boundaries, so |
626 | // they need not be processed in live sets. |
627 | // this is checked by these assertions to be sure about it. |
628 | // the entry block may have incoming values in registers, which is ok. |
629 | if (!opr->is_virtual_register() && block != ir()->start()) { |
630 | reg = reg_num(opr); |
631 | if (is_processed_reg_num(reg)) { |
632 | assert(live_kill.at(reg), "using fixed register that is not defined in this block" ); |
633 | } |
634 | reg = reg_numHi(opr); |
635 | if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { |
636 | assert(live_kill.at(reg), "using fixed register that is not defined in this block" ); |
637 | } |
638 | } |
639 | #endif |
640 | } |
641 | |
642 | // Add uses of live locals from interpreter's point of view for proper debug information generation |
643 | n = visitor.info_count(); |
644 | for (k = 0; k < n; k++) { |
645 | CodeEmitInfo* info = visitor.info_at(k); |
646 | ValueStack* stack = info->stack(); |
647 | for_each_state_value(stack, value, |
648 | set_live_gen_kill(value, op, live_gen, live_kill) |
649 | ); |
650 | } |
651 | |
652 | // iterate temp operands of instruction |
653 | n = visitor.opr_count(LIR_OpVisitState::tempMode); |
654 | for (k = 0; k < n; k++) { |
655 | LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, k); |
656 | assert(opr->is_register(), "visitor should only return register operands" ); |
657 | |
658 | if (opr->is_virtual_register()) { |
659 | assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below" ); |
660 | reg = opr->vreg_number(); |
661 | live_kill.set_bit(reg); |
662 | if (block->loop_index() >= 0) { |
663 | local_interval_in_loop.set_bit(reg, block->loop_index()); |
664 | } |
665 | local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu(); |
666 | } |
667 | |
668 | #ifdef ASSERT |
669 | // fixed intervals are never live at block boundaries, so |
670 | // they need not be processed in live sets |
671 | // process them only in debug mode so that this can be checked |
672 | if (!opr->is_virtual_register()) { |
673 | reg = reg_num(opr); |
674 | if (is_processed_reg_num(reg)) { |
675 | live_kill.set_bit(reg_num(opr)); |
676 | } |
677 | reg = reg_numHi(opr); |
678 | if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { |
679 | live_kill.set_bit(reg); |
680 | } |
681 | } |
682 | #endif |
683 | } |
684 | |
685 | // iterate output operands of instruction |
686 | n = visitor.opr_count(LIR_OpVisitState::outputMode); |
687 | for (k = 0; k < n; k++) { |
688 | LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, k); |
689 | assert(opr->is_register(), "visitor should only return register operands" ); |
690 | |
691 | if (opr->is_virtual_register()) { |
692 | assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below" ); |
693 | reg = opr->vreg_number(); |
694 | live_kill.set_bit(reg); |
695 | if (block->loop_index() >= 0) { |
696 | local_interval_in_loop.set_bit(reg, block->loop_index()); |
697 | } |
698 | local_has_fpu_registers = local_has_fpu_registers || opr->is_virtual_fpu(); |
699 | } |
700 | |
701 | #ifdef ASSERT |
702 | // fixed intervals are never live at block boundaries, so |
703 | // they need not be processed in live sets |
704 | // process them only in debug mode so that this can be checked |
705 | if (!opr->is_virtual_register()) { |
706 | reg = reg_num(opr); |
707 | if (is_processed_reg_num(reg)) { |
708 | live_kill.set_bit(reg_num(opr)); |
709 | } |
710 | reg = reg_numHi(opr); |
711 | if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { |
712 | live_kill.set_bit(reg); |
713 | } |
714 | } |
715 | #endif |
716 | } |
717 | } // end of instruction iteration |
718 | |
719 | block->set_live_gen (live_gen); |
720 | block->set_live_kill(live_kill); |
721 | block->set_live_in (ResourceBitMap(live_size)); |
722 | block->set_live_out (ResourceBitMap(live_size)); |
723 | |
724 | TRACE_LINEAR_SCAN(4, tty->print("live_gen B%d " , block->block_id()); print_bitmap(block->live_gen())); |
725 | TRACE_LINEAR_SCAN(4, tty->print("live_kill B%d " , block->block_id()); print_bitmap(block->live_kill())); |
726 | } // end of block iteration |
727 | |
728 | // propagate local calculated information into LinearScan object |
729 | _has_fpu_registers = local_has_fpu_registers; |
730 | compilation()->set_has_fpu_code(local_has_fpu_registers); |
731 | |
732 | _num_calls = local_num_calls; |
733 | _interval_in_loop = local_interval_in_loop; |
734 | } |
735 | |
736 | |
737 | // ********** Phase 3: perform a backward dataflow analysis to compute global live sets |
738 | // (sets live_in and live_out for each block) |
739 | |
740 | void LinearScan::compute_global_live_sets() { |
741 | TIME_LINEAR_SCAN(timer_compute_global_live_sets); |
742 | |
743 | int num_blocks = block_count(); |
744 | bool change_occurred; |
745 | bool change_occurred_in_block; |
746 | int iteration_count = 0; |
747 | ResourceBitMap live_out(live_set_size()); // scratch set for calculations |
748 | |
749 | // Perform a backward dataflow analysis to compute live_out and live_in for each block. |
750 | // The loop is executed until a fixpoint is reached (no changes in an iteration) |
751 | // Exception handlers must be processed because not all live values are |
752 | // present in the state array, e.g. because of global value numbering |
753 | do { |
754 | change_occurred = false; |
755 | |
756 | // iterate all blocks in reverse order |
757 | for (int i = num_blocks - 1; i >= 0; i--) { |
758 | BlockBegin* block = block_at(i); |
759 | |
760 | change_occurred_in_block = false; |
761 | |
762 | // live_out(block) is the union of live_in(sux), for successors sux of block |
763 | int n = block->number_of_sux(); |
764 | int e = block->number_of_exception_handlers(); |
765 | if (n + e > 0) { |
766 | // block has successors |
767 | if (n > 0) { |
768 | live_out.set_from(block->sux_at(0)->live_in()); |
769 | for (int j = 1; j < n; j++) { |
770 | live_out.set_union(block->sux_at(j)->live_in()); |
771 | } |
772 | } else { |
773 | live_out.clear(); |
774 | } |
775 | for (int j = 0; j < e; j++) { |
776 | live_out.set_union(block->exception_handler_at(j)->live_in()); |
777 | } |
778 | |
779 | if (!block->live_out().is_same(live_out)) { |
780 | // A change occurred. Swap the old and new live out sets to avoid copying. |
781 | ResourceBitMap temp = block->live_out(); |
782 | block->set_live_out(live_out); |
783 | live_out = temp; |
784 | |
785 | change_occurred = true; |
786 | change_occurred_in_block = true; |
787 | } |
788 | } |
789 | |
790 | if (iteration_count == 0 || change_occurred_in_block) { |
791 | // live_in(block) is the union of live_gen(block) with (live_out(block) & !live_kill(block)) |
792 | // note: live_in has to be computed only in first iteration or if live_out has changed! |
793 | ResourceBitMap live_in = block->live_in(); |
794 | live_in.set_from(block->live_out()); |
795 | live_in.set_difference(block->live_kill()); |
796 | live_in.set_union(block->live_gen()); |
797 | } |
798 | |
799 | #ifndef PRODUCT |
800 | if (TraceLinearScanLevel >= 4) { |
801 | char c = ' '; |
802 | if (iteration_count == 0 || change_occurred_in_block) { |
803 | c = '*'; |
804 | } |
805 | tty->print("(%d) live_in%c B%d " , iteration_count, c, block->block_id()); print_bitmap(block->live_in()); |
806 | tty->print("(%d) live_out%c B%d " , iteration_count, c, block->block_id()); print_bitmap(block->live_out()); |
807 | } |
808 | #endif |
809 | } |
810 | iteration_count++; |
811 | |
812 | if (change_occurred && iteration_count > 50) { |
813 | BAILOUT("too many iterations in compute_global_live_sets" ); |
814 | } |
815 | } while (change_occurred); |
816 | |
817 | |
818 | #ifdef ASSERT |
819 | // check that fixed intervals are not live at block boundaries |
820 | // (live set must be empty at fixed intervals) |
821 | for (int i = 0; i < num_blocks; i++) { |
822 | BlockBegin* block = block_at(i); |
823 | for (int j = 0; j < LIR_OprDesc::vreg_base; j++) { |
824 | assert(block->live_in().at(j) == false, "live_in set of fixed register must be empty" ); |
825 | assert(block->live_out().at(j) == false, "live_out set of fixed register must be empty" ); |
826 | assert(block->live_gen().at(j) == false, "live_gen set of fixed register must be empty" ); |
827 | } |
828 | } |
829 | #endif |
830 | |
831 | // check that the live_in set of the first block is empty |
832 | ResourceBitMap live_in_args(ir()->start()->live_in().size()); |
833 | if (!ir()->start()->live_in().is_same(live_in_args)) { |
834 | #ifdef ASSERT |
835 | tty->print_cr("Error: live_in set of first block must be empty (when this fails, virtual registers are used before they are defined)" ); |
836 | tty->print_cr("affected registers:" ); |
837 | print_bitmap(ir()->start()->live_in()); |
838 | |
839 | // print some additional information to simplify debugging |
840 | for (unsigned int i = 0; i < ir()->start()->live_in().size(); i++) { |
841 | if (ir()->start()->live_in().at(i)) { |
842 | Instruction* instr = gen()->instruction_for_vreg(i); |
843 | tty->print_cr("* vreg %d (HIR instruction %c%d)" , i, instr == NULL ? ' ' : instr->type()->tchar(), instr == NULL ? 0 : instr->id()); |
844 | |
845 | for (int j = 0; j < num_blocks; j++) { |
846 | BlockBegin* block = block_at(j); |
847 | if (block->live_gen().at(i)) { |
848 | tty->print_cr(" used in block B%d" , block->block_id()); |
849 | } |
850 | if (block->live_kill().at(i)) { |
851 | tty->print_cr(" defined in block B%d" , block->block_id()); |
852 | } |
853 | } |
854 | } |
855 | } |
856 | |
857 | #endif |
858 | // when this fails, virtual registers are used before they are defined. |
859 | assert(false, "live_in set of first block must be empty" ); |
860 | // bailout of if this occurs in product mode. |
861 | bailout("live_in set of first block not empty" ); |
862 | } |
863 | } |
864 | |
865 | |
866 | // ********** Phase 4: build intervals |
867 | // (fills the list _intervals) |
868 | |
869 | void LinearScan::add_use(Value value, int from, int to, IntervalUseKind use_kind) { |
870 | assert(!value->type()->is_illegal(), "if this value is used by the interpreter it shouldn't be of indeterminate type" ); |
871 | LIR_Opr opr = value->operand(); |
872 | Constant* con = value->as_Constant(); |
873 | |
874 | if ((con == NULL || con->is_pinned()) && opr->is_register()) { |
875 | assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below" ); |
876 | add_use(opr, from, to, use_kind); |
877 | } |
878 | } |
879 | |
880 | |
881 | void LinearScan::add_def(LIR_Opr opr, int def_pos, IntervalUseKind use_kind) { |
882 | TRACE_LINEAR_SCAN(2, tty->print(" def " ); opr->print(tty); tty->print_cr(" def_pos %d (%d)" , def_pos, use_kind)); |
883 | assert(opr->is_register(), "should not be called otherwise" ); |
884 | |
885 | if (opr->is_virtual_register()) { |
886 | assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below" ); |
887 | add_def(opr->vreg_number(), def_pos, use_kind, opr->type_register()); |
888 | |
889 | } else { |
890 | int reg = reg_num(opr); |
891 | if (is_processed_reg_num(reg)) { |
892 | add_def(reg, def_pos, use_kind, opr->type_register()); |
893 | } |
894 | reg = reg_numHi(opr); |
895 | if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { |
896 | add_def(reg, def_pos, use_kind, opr->type_register()); |
897 | } |
898 | } |
899 | } |
900 | |
901 | void LinearScan::add_use(LIR_Opr opr, int from, int to, IntervalUseKind use_kind) { |
902 | TRACE_LINEAR_SCAN(2, tty->print(" use " ); opr->print(tty); tty->print_cr(" from %d to %d (%d)" , from, to, use_kind)); |
903 | assert(opr->is_register(), "should not be called otherwise" ); |
904 | |
905 | if (opr->is_virtual_register()) { |
906 | assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below" ); |
907 | add_use(opr->vreg_number(), from, to, use_kind, opr->type_register()); |
908 | |
909 | } else { |
910 | int reg = reg_num(opr); |
911 | if (is_processed_reg_num(reg)) { |
912 | add_use(reg, from, to, use_kind, opr->type_register()); |
913 | } |
914 | reg = reg_numHi(opr); |
915 | if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { |
916 | add_use(reg, from, to, use_kind, opr->type_register()); |
917 | } |
918 | } |
919 | } |
920 | |
921 | void LinearScan::add_temp(LIR_Opr opr, int temp_pos, IntervalUseKind use_kind) { |
922 | TRACE_LINEAR_SCAN(2, tty->print(" temp " ); opr->print(tty); tty->print_cr(" temp_pos %d (%d)" , temp_pos, use_kind)); |
923 | assert(opr->is_register(), "should not be called otherwise" ); |
924 | |
925 | if (opr->is_virtual_register()) { |
926 | assert(reg_num(opr) == opr->vreg_number() && !is_valid_reg_num(reg_numHi(opr)), "invalid optimization below" ); |
927 | add_temp(opr->vreg_number(), temp_pos, use_kind, opr->type_register()); |
928 | |
929 | } else { |
930 | int reg = reg_num(opr); |
931 | if (is_processed_reg_num(reg)) { |
932 | add_temp(reg, temp_pos, use_kind, opr->type_register()); |
933 | } |
934 | reg = reg_numHi(opr); |
935 | if (is_valid_reg_num(reg) && is_processed_reg_num(reg)) { |
936 | add_temp(reg, temp_pos, use_kind, opr->type_register()); |
937 | } |
938 | } |
939 | } |
940 | |
941 | |
942 | void LinearScan::add_def(int reg_num, int def_pos, IntervalUseKind use_kind, BasicType type) { |
943 | Interval* interval = interval_at(reg_num); |
944 | if (interval != NULL) { |
945 | assert(interval->reg_num() == reg_num, "wrong interval" ); |
946 | |
947 | if (type != T_ILLEGAL) { |
948 | interval->set_type(type); |
949 | } |
950 | |
951 | Range* r = interval->first(); |
952 | if (r->from() <= def_pos) { |
953 | // Update the starting point (when a range is first created for a use, its |
954 | // start is the beginning of the current block until a def is encountered.) |
955 | r->set_from(def_pos); |
956 | interval->add_use_pos(def_pos, use_kind); |
957 | |
958 | } else { |
959 | // Dead value - make vacuous interval |
960 | // also add use_kind for dead intervals |
961 | interval->add_range(def_pos, def_pos + 1); |
962 | interval->add_use_pos(def_pos, use_kind); |
963 | TRACE_LINEAR_SCAN(2, tty->print_cr("Warning: def of reg %d at %d occurs without use" , reg_num, def_pos)); |
964 | } |
965 | |
966 | } else { |
967 | // Dead value - make vacuous interval |
968 | // also add use_kind for dead intervals |
969 | interval = create_interval(reg_num); |
970 | if (type != T_ILLEGAL) { |
971 | interval->set_type(type); |
972 | } |
973 | |
974 | interval->add_range(def_pos, def_pos + 1); |
975 | interval->add_use_pos(def_pos, use_kind); |
976 | TRACE_LINEAR_SCAN(2, tty->print_cr("Warning: dead value %d at %d in live intervals" , reg_num, def_pos)); |
977 | } |
978 | |
979 | change_spill_definition_pos(interval, def_pos); |
980 | if (use_kind == noUse && interval->spill_state() <= startInMemory) { |
981 | // detection of method-parameters and roundfp-results |
982 | // TODO: move this directly to position where use-kind is computed |
983 | interval->set_spill_state(startInMemory); |
984 | } |
985 | } |
986 | |
987 | void LinearScan::add_use(int reg_num, int from, int to, IntervalUseKind use_kind, BasicType type) { |
988 | Interval* interval = interval_at(reg_num); |
989 | if (interval == NULL) { |
990 | interval = create_interval(reg_num); |
991 | } |
992 | assert(interval->reg_num() == reg_num, "wrong interval" ); |
993 | |
994 | if (type != T_ILLEGAL) { |
995 | interval->set_type(type); |
996 | } |
997 | |
998 | interval->add_range(from, to); |
999 | interval->add_use_pos(to, use_kind); |
1000 | } |
1001 | |
1002 | void LinearScan::add_temp(int reg_num, int temp_pos, IntervalUseKind use_kind, BasicType type) { |
1003 | Interval* interval = interval_at(reg_num); |
1004 | if (interval == NULL) { |
1005 | interval = create_interval(reg_num); |
1006 | } |
1007 | assert(interval->reg_num() == reg_num, "wrong interval" ); |
1008 | |
1009 | if (type != T_ILLEGAL) { |
1010 | interval->set_type(type); |
1011 | } |
1012 | |
1013 | interval->add_range(temp_pos, temp_pos + 1); |
1014 | interval->add_use_pos(temp_pos, use_kind); |
1015 | } |
1016 | |
1017 | |
1018 | // the results of this functions are used for optimizing spilling and reloading |
1019 | // if the functions return shouldHaveRegister and the interval is spilled, |
1020 | // it is not reloaded to a register. |
1021 | IntervalUseKind LinearScan::use_kind_of_output_operand(LIR_Op* op, LIR_Opr opr) { |
1022 | if (op->code() == lir_move) { |
1023 | assert(op->as_Op1() != NULL, "lir_move must be LIR_Op1" ); |
1024 | LIR_Op1* move = (LIR_Op1*)op; |
1025 | LIR_Opr res = move->result_opr(); |
1026 | bool result_in_memory = res->is_virtual() && gen()->is_vreg_flag_set(res->vreg_number(), LIRGenerator::must_start_in_memory); |
1027 | |
1028 | if (result_in_memory) { |
1029 | // Begin of an interval with must_start_in_memory set. |
1030 | // This interval will always get a stack slot first, so return noUse. |
1031 | return noUse; |
1032 | |
1033 | } else if (move->in_opr()->is_stack()) { |
1034 | // method argument (condition must be equal to handle_method_arguments) |
1035 | return noUse; |
1036 | |
1037 | } else if (move->in_opr()->is_register() && move->result_opr()->is_register()) { |
1038 | // Move from register to register |
1039 | if (block_of_op_with_id(op->id())->is_set(BlockBegin::osr_entry_flag)) { |
1040 | // special handling of phi-function moves inside osr-entry blocks |
1041 | // input operand must have a register instead of output operand (leads to better register allocation) |
1042 | return shouldHaveRegister; |
1043 | } |
1044 | } |
1045 | } |
1046 | |
1047 | if (opr->is_virtual() && |
1048 | gen()->is_vreg_flag_set(opr->vreg_number(), LIRGenerator::must_start_in_memory)) { |
1049 | // result is a stack-slot, so prevent immediate reloading |
1050 | return noUse; |
1051 | } |
1052 | |
1053 | // all other operands require a register |
1054 | return mustHaveRegister; |
1055 | } |
1056 | |
1057 | IntervalUseKind LinearScan::use_kind_of_input_operand(LIR_Op* op, LIR_Opr opr) { |
1058 | if (op->code() == lir_move) { |
1059 | assert(op->as_Op1() != NULL, "lir_move must be LIR_Op1" ); |
1060 | LIR_Op1* move = (LIR_Op1*)op; |
1061 | LIR_Opr res = move->result_opr(); |
1062 | bool result_in_memory = res->is_virtual() && gen()->is_vreg_flag_set(res->vreg_number(), LIRGenerator::must_start_in_memory); |
1063 | |
1064 | if (result_in_memory) { |
1065 | // Move to an interval with must_start_in_memory set. |
1066 | // To avoid moves from stack to stack (not allowed) force the input operand to a register |
1067 | return mustHaveRegister; |
1068 | |
1069 | } else if (move->in_opr()->is_register() && move->result_opr()->is_register()) { |
1070 | // Move from register to register |
1071 | if (block_of_op_with_id(op->id())->is_set(BlockBegin::osr_entry_flag)) { |
1072 | // special handling of phi-function moves inside osr-entry blocks |
1073 | // input operand must have a register instead of output operand (leads to better register allocation) |
1074 | return mustHaveRegister; |
1075 | } |
1076 | |
1077 | // The input operand is not forced to a register (moves from stack to register are allowed), |
1078 | // but it is faster if the input operand is in a register |
1079 | return shouldHaveRegister; |
1080 | } |
1081 | } |
1082 | |
1083 | |
1084 | #if defined(X86) || defined(S390) |
1085 | if (op->code() == lir_cmove) { |
1086 | // conditional moves can handle stack operands |
1087 | assert(op->result_opr()->is_register(), "result must always be in a register" ); |
1088 | return shouldHaveRegister; |
1089 | } |
1090 | |
1091 | // optimizations for second input operand of arithmehtic operations on Intel |
1092 | // this operand is allowed to be on the stack in some cases |
1093 | BasicType opr_type = opr->type_register(); |
1094 | if (opr_type == T_FLOAT || opr_type == T_DOUBLE) { |
1095 | if ((UseSSE == 1 && opr_type == T_FLOAT) || UseSSE >= 2 S390_ONLY(|| true)) { |
1096 | // SSE float instruction (T_DOUBLE only supported with SSE2) |
1097 | switch (op->code()) { |
1098 | case lir_cmp: |
1099 | case lir_add: |
1100 | case lir_sub: |
1101 | case lir_mul: |
1102 | case lir_div: |
1103 | { |
1104 | assert(op->as_Op2() != NULL, "must be LIR_Op2" ); |
1105 | LIR_Op2* op2 = (LIR_Op2*)op; |
1106 | if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) { |
1107 | assert((op2->result_opr()->is_register() || op->code() == lir_cmp) && op2->in_opr1()->is_register(), "cannot mark second operand as stack if others are not in register" ); |
1108 | return shouldHaveRegister; |
1109 | } |
1110 | } |
1111 | default: |
1112 | break; |
1113 | } |
1114 | } else { |
1115 | // FPU stack float instruction |
1116 | switch (op->code()) { |
1117 | case lir_add: |
1118 | case lir_sub: |
1119 | case lir_mul: |
1120 | case lir_div: |
1121 | { |
1122 | assert(op->as_Op2() != NULL, "must be LIR_Op2" ); |
1123 | LIR_Op2* op2 = (LIR_Op2*)op; |
1124 | if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) { |
1125 | 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" ); |
1126 | return shouldHaveRegister; |
1127 | } |
1128 | } |
1129 | default: |
1130 | break; |
1131 | } |
1132 | } |
1133 | // We want to sometimes use logical operations on pointers, in particular in GC barriers. |
1134 | // Since 64bit logical operations do not current support operands on stack, we have to make sure |
1135 | // T_OBJECT doesn't get spilled along with T_LONG. |
1136 | } else if (opr_type != T_LONG LP64_ONLY(&& opr_type != T_OBJECT)) { |
1137 | // integer instruction (note: long operands must always be in register) |
1138 | switch (op->code()) { |
1139 | case lir_cmp: |
1140 | case lir_add: |
1141 | case lir_sub: |
1142 | case lir_logic_and: |
1143 | case lir_logic_or: |
1144 | case lir_logic_xor: |
1145 | { |
1146 | assert(op->as_Op2() != NULL, "must be LIR_Op2" ); |
1147 | LIR_Op2* op2 = (LIR_Op2*)op; |
1148 | if (op2->in_opr1() != op2->in_opr2() && op2->in_opr2() == opr) { |
1149 | 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" ); |
1150 | return shouldHaveRegister; |
1151 | } |
1152 | } |
1153 | default: |
1154 | break; |
1155 | } |
1156 | } |
1157 | #endif // X86 S390 |
1158 | |
1159 | // all other operands require a register |
1160 | return mustHaveRegister; |
1161 | } |
1162 | |
1163 | |
1164 | void LinearScan::handle_method_arguments(LIR_Op* op) { |
1165 | // special handling for method arguments (moves from stack to virtual register): |
1166 | // the interval gets no register assigned, but the stack slot. |
1167 | // it is split before the first use by the register allocator. |
1168 | |
1169 | if (op->code() == lir_move) { |
1170 | assert(op->as_Op1() != NULL, "must be LIR_Op1" ); |
1171 | LIR_Op1* move = (LIR_Op1*)op; |
1172 | |
1173 | if (move->in_opr()->is_stack()) { |
1174 | #ifdef ASSERT |
1175 | int arg_size = compilation()->method()->arg_size(); |
1176 | LIR_Opr o = move->in_opr(); |
1177 | if (o->is_single_stack()) { |
1178 | assert(o->single_stack_ix() >= 0 && o->single_stack_ix() < arg_size, "out of range" ); |
1179 | } else if (o->is_double_stack()) { |
1180 | assert(o->double_stack_ix() >= 0 && o->double_stack_ix() < arg_size, "out of range" ); |
1181 | } else { |
1182 | ShouldNotReachHere(); |
1183 | } |
1184 | |
1185 | assert(move->id() > 0, "invalid id" ); |
1186 | assert(block_of_op_with_id(move->id())->number_of_preds() == 0, "move from stack must be in first block" ); |
1187 | assert(move->result_opr()->is_virtual(), "result of move must be a virtual register" ); |
1188 | |
1189 | 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()))); |
1190 | #endif |
1191 | |
1192 | Interval* interval = interval_at(reg_num(move->result_opr())); |
1193 | |
1194 | int stack_slot = LinearScan::nof_regs + (move->in_opr()->is_single_stack() ? move->in_opr()->single_stack_ix() : move->in_opr()->double_stack_ix()); |
1195 | interval->set_canonical_spill_slot(stack_slot); |
1196 | interval->assign_reg(stack_slot); |
1197 | } |
1198 | } |
1199 | } |
1200 | |
1201 | void LinearScan::handle_doubleword_moves(LIR_Op* op) { |
1202 | // special handling for doubleword move from memory to register: |
1203 | // in this case the registers of the input address and the result |
1204 | // registers must not overlap -> add a temp range for the input registers |
1205 | if (op->code() == lir_move) { |
1206 | assert(op->as_Op1() != NULL, "must be LIR_Op1" ); |
1207 | LIR_Op1* move = (LIR_Op1*)op; |
1208 | |
1209 | if (move->result_opr()->is_double_cpu() && move->in_opr()->is_pointer()) { |
1210 | LIR_Address* address = move->in_opr()->as_address_ptr(); |
1211 | if (address != NULL) { |
1212 | if (address->base()->is_valid()) { |
1213 | add_temp(address->base(), op->id(), noUse); |
1214 | } |
1215 | if (address->index()->is_valid()) { |
1216 | add_temp(address->index(), op->id(), noUse); |
1217 | } |
1218 | } |
1219 | } |
1220 | } |
1221 | } |
1222 | |
1223 | void LinearScan::add_register_hints(LIR_Op* op) { |
1224 | switch (op->code()) { |
1225 | case lir_move: // fall through |
1226 | case lir_convert: { |
1227 | assert(op->as_Op1() != NULL, "lir_move, lir_convert must be LIR_Op1" ); |
1228 | LIR_Op1* move = (LIR_Op1*)op; |
1229 | |
1230 | LIR_Opr move_from = move->in_opr(); |
1231 | LIR_Opr move_to = move->result_opr(); |
1232 | |
1233 | if (move_to->is_register() && move_from->is_register()) { |
1234 | Interval* from = interval_at(reg_num(move_from)); |
1235 | Interval* to = interval_at(reg_num(move_to)); |
1236 | if (from != NULL && to != NULL) { |
1237 | to->set_register_hint(from); |
1238 | 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())); |
1239 | } |
1240 | } |
1241 | break; |
1242 | } |
1243 | case lir_cmove: { |
1244 | assert(op->as_Op2() != NULL, "lir_cmove must be LIR_Op2" ); |
1245 | LIR_Op2* cmove = (LIR_Op2*)op; |
1246 | |
1247 | LIR_Opr move_from = cmove->in_opr1(); |
1248 | LIR_Opr move_to = cmove->result_opr(); |
1249 | |
1250 | if (move_to->is_register() && move_from->is_register()) { |
1251 | Interval* from = interval_at(reg_num(move_from)); |
1252 | Interval* to = interval_at(reg_num(move_to)); |
1253 | if (from != NULL && to != NULL) { |
1254 | to->set_register_hint(from); |
1255 | 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())); |
1256 | } |
1257 | } |
1258 | break; |
1259 | } |
1260 | default: |
1261 | break; |
1262 | } |
1263 | } |
1264 | |
1265 | |
1266 | void LinearScan::build_intervals() { |
1267 | TIME_LINEAR_SCAN(timer_build_intervals); |
1268 | |
1269 | // initialize interval list with expected number of intervals |
1270 | // (32 is added to have some space for split children without having to resize the list) |
1271 | _intervals = IntervalList(num_virtual_regs() + 32); |
1272 | // initialize all slots that are used by build_intervals |
1273 | _intervals.at_put_grow(num_virtual_regs() - 1, NULL, NULL); |
1274 | |
1275 | // create a list with all caller-save registers (cpu, fpu, xmm) |
1276 | // when an instruction is a call, a temp range is created for all these registers |
1277 | int num_caller_save_registers = 0; |
1278 | int caller_save_registers[LinearScan::nof_regs]; |
1279 | |
1280 | int i; |
1281 | for (i = 0; i < FrameMap::nof_caller_save_cpu_regs(); i++) { |
1282 | LIR_Opr opr = FrameMap::caller_save_cpu_reg_at(i); |
1283 | assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands" ); |
1284 | assert(reg_numHi(opr) == -1, "missing addition of range for hi-register" ); |
1285 | caller_save_registers[num_caller_save_registers++] = reg_num(opr); |
1286 | } |
1287 | |
1288 | // temp ranges for fpu registers are only created when the method has |
1289 | // virtual fpu operands. Otherwise no allocation for fpu registers is |
1290 | // performed and so the temp ranges would be useless |
1291 | if (has_fpu_registers()) { |
1292 | #ifdef X86 |
1293 | if (UseSSE < 2) { |
1294 | #endif |
1295 | for (i = 0; i < FrameMap::nof_caller_save_fpu_regs; i++) { |
1296 | LIR_Opr opr = FrameMap::caller_save_fpu_reg_at(i); |
1297 | assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands" ); |
1298 | assert(reg_numHi(opr) == -1, "missing addition of range for hi-register" ); |
1299 | caller_save_registers[num_caller_save_registers++] = reg_num(opr); |
1300 | } |
1301 | #ifdef X86 |
1302 | } |
1303 | if (UseSSE > 0) { |
1304 | int num_caller_save_xmm_regs = FrameMap::get_num_caller_save_xmms(); |
1305 | for (i = 0; i < num_caller_save_xmm_regs; i ++) { |
1306 | LIR_Opr opr = FrameMap::caller_save_xmm_reg_at(i); |
1307 | assert(opr->is_valid() && opr->is_register(), "FrameMap should not return invalid operands" ); |
1308 | assert(reg_numHi(opr) == -1, "missing addition of range for hi-register" ); |
1309 | caller_save_registers[num_caller_save_registers++] = reg_num(opr); |
1310 | } |
1311 | } |
1312 | #endif |
1313 | } |
1314 | assert(num_caller_save_registers <= LinearScan::nof_regs, "out of bounds" ); |
1315 | |
1316 | |
1317 | LIR_OpVisitState visitor; |
1318 | |
1319 | // iterate all blocks in reverse order |
1320 | for (i = block_count() - 1; i >= 0; i--) { |
1321 | BlockBegin* block = block_at(i); |
1322 | LIR_OpList* instructions = block->lir()->instructions_list(); |
1323 | int block_from = block->first_lir_instruction_id(); |
1324 | int block_to = block->last_lir_instruction_id(); |
1325 | |
1326 | assert(block_from == instructions->at(0)->id(), "must be" ); |
1327 | assert(block_to == instructions->at(instructions->length() - 1)->id(), "must be" ); |
1328 | |
1329 | // Update intervals for registers live at the end of this block; |
1330 | ResourceBitMap live = block->live_out(); |
1331 | int size = (int)live.size(); |
1332 | for (int number = (int)live.get_next_one_offset(0, size); number < size; number = (int)live.get_next_one_offset(number + 1, size)) { |
1333 | assert(live.at(number), "should not stop here otherwise" ); |
1334 | assert(number >= LIR_OprDesc::vreg_base, "fixed intervals must not be live on block bounds" ); |
1335 | TRACE_LINEAR_SCAN(2, tty->print_cr("live in %d to %d" , number, block_to + 2)); |
1336 | |
1337 | add_use(number, block_from, block_to + 2, noUse, T_ILLEGAL); |
1338 | |
1339 | // add special use positions for loop-end blocks when the |
1340 | // interval is used anywhere inside this loop. It's possible |
1341 | // that the block was part of a non-natural loop, so it might |
1342 | // have an invalid loop index. |
1343 | if (block->is_set(BlockBegin::linear_scan_loop_end_flag) && |
1344 | block->loop_index() != -1 && |
1345 | is_interval_in_loop(number, block->loop_index())) { |
1346 | interval_at(number)->add_use_pos(block_to + 1, loopEndMarker); |
1347 | } |
1348 | } |
1349 | |
1350 | // iterate all instructions of the block in reverse order. |
1351 | // skip the first instruction because it is always a label |
1352 | // definitions of intervals are processed before uses |
1353 | assert(visitor.no_operands(instructions->at(0)), "first operation must always be a label" ); |
1354 | for (int j = instructions->length() - 1; j >= 1; j--) { |
1355 | LIR_Op* op = instructions->at(j); |
1356 | int op_id = op->id(); |
1357 | |
1358 | // visit operation to collect all operands |
1359 | visitor.visit(op); |
1360 | |
1361 | // add a temp range for each register if operation destroys caller-save registers |
1362 | if (visitor.has_call()) { |
1363 | for (int k = 0; k < num_caller_save_registers; k++) { |
1364 | add_temp(caller_save_registers[k], op_id, noUse, T_ILLEGAL); |
1365 | } |
1366 | TRACE_LINEAR_SCAN(4, tty->print_cr("operation destroys all caller-save registers" )); |
1367 | } |
1368 | |
1369 | // Add any platform dependent temps |
1370 | pd_add_temps(op); |
1371 | |
1372 | // visit definitions (output and temp operands) |
1373 | int k, n; |
1374 | n = visitor.opr_count(LIR_OpVisitState::outputMode); |
1375 | for (k = 0; k < n; k++) { |
1376 | LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, k); |
1377 | assert(opr->is_register(), "visitor should only return register operands" ); |
1378 | add_def(opr, op_id, use_kind_of_output_operand(op, opr)); |
1379 | } |
1380 | |
1381 | n = visitor.opr_count(LIR_OpVisitState::tempMode); |
1382 | for (k = 0; k < n; k++) { |
1383 | LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, k); |
1384 | assert(opr->is_register(), "visitor should only return register operands" ); |
1385 | add_temp(opr, op_id, mustHaveRegister); |
1386 | } |
1387 | |
1388 | // visit uses (input operands) |
1389 | n = visitor.opr_count(LIR_OpVisitState::inputMode); |
1390 | for (k = 0; k < n; k++) { |
1391 | LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, k); |
1392 | assert(opr->is_register(), "visitor should only return register operands" ); |
1393 | add_use(opr, block_from, op_id, use_kind_of_input_operand(op, opr)); |
1394 | } |
1395 | |
1396 | // Add uses of live locals from interpreter's point of view for proper |
1397 | // debug information generation |
1398 | // Treat these operands as temp values (if the life range is extended |
1399 | // to a call site, the value would be in a register at the call otherwise) |
1400 | n = visitor.info_count(); |
1401 | for (k = 0; k < n; k++) { |
1402 | CodeEmitInfo* info = visitor.info_at(k); |
1403 | ValueStack* stack = info->stack(); |
1404 | for_each_state_value(stack, value, |
1405 | add_use(value, block_from, op_id + 1, noUse); |
1406 | ); |
1407 | } |
1408 | |
1409 | // special steps for some instructions (especially moves) |
1410 | handle_method_arguments(op); |
1411 | handle_doubleword_moves(op); |
1412 | add_register_hints(op); |
1413 | |
1414 | } // end of instruction iteration |
1415 | } // end of block iteration |
1416 | |
1417 | |
1418 | // add the range [0, 1[ to all fixed intervals |
1419 | // -> the register allocator need not handle unhandled fixed intervals |
1420 | for (int n = 0; n < LinearScan::nof_regs; n++) { |
1421 | Interval* interval = interval_at(n); |
1422 | if (interval != NULL) { |
1423 | interval->add_range(0, 1); |
1424 | } |
1425 | } |
1426 | } |
1427 | |
1428 | |
1429 | // ********** Phase 5: actual register allocation |
1430 | |
1431 | int LinearScan::interval_cmp(Interval** a, Interval** b) { |
1432 | if (*a != NULL) { |
1433 | if (*b != NULL) { |
1434 | return (*a)->from() - (*b)->from(); |
1435 | } else { |
1436 | return -1; |
1437 | } |
1438 | } else { |
1439 | if (*b != NULL) { |
1440 | return 1; |
1441 | } else { |
1442 | return 0; |
1443 | } |
1444 | } |
1445 | } |
1446 | |
1447 | #ifndef PRODUCT |
1448 | int interval_cmp(Interval* const& l, Interval* const& r) { |
1449 | return l->from() - r->from(); |
1450 | } |
1451 | |
1452 | bool find_interval(Interval* interval, IntervalArray* intervals) { |
1453 | bool found; |
1454 | int idx = intervals->find_sorted<Interval*, interval_cmp>(interval, found); |
1455 | |
1456 | if (!found) { |
1457 | return false; |
1458 | } |
1459 | |
1460 | int from = interval->from(); |
1461 | |
1462 | // The index we've found using binary search is pointing to an interval |
1463 | // that is defined in the same place as the interval we were looking for. |
1464 | // So now we have to look around that index and find exact interval. |
1465 | for (int i = idx; i >= 0; i--) { |
1466 | if (intervals->at(i) == interval) { |
1467 | return true; |
1468 | } |
1469 | if (intervals->at(i)->from() != from) { |
1470 | break; |
1471 | } |
1472 | } |
1473 | |
1474 | for (int i = idx + 1; i < intervals->length(); i++) { |
1475 | if (intervals->at(i) == interval) { |
1476 | return true; |
1477 | } |
1478 | if (intervals->at(i)->from() != from) { |
1479 | break; |
1480 | } |
1481 | } |
1482 | |
1483 | return false; |
1484 | } |
1485 | |
1486 | bool LinearScan::is_sorted(IntervalArray* intervals) { |
1487 | int from = -1; |
1488 | int null_count = 0; |
1489 | |
1490 | for (int i = 0; i < intervals->length(); i++) { |
1491 | Interval* it = intervals->at(i); |
1492 | if (it != NULL) { |
1493 | assert(from <= it->from(), "Intervals are unordered" ); |
1494 | from = it->from(); |
1495 | } else { |
1496 | null_count++; |
1497 | } |
1498 | } |
1499 | |
1500 | assert(null_count == 0, "Sorted intervals should not contain nulls" ); |
1501 | |
1502 | null_count = 0; |
1503 | |
1504 | for (int i = 0; i < interval_count(); i++) { |
1505 | Interval* interval = interval_at(i); |
1506 | if (interval != NULL) { |
1507 | assert(find_interval(interval, intervals), "Lists do not contain same intervals" ); |
1508 | } else { |
1509 | null_count++; |
1510 | } |
1511 | } |
1512 | |
1513 | assert(interval_count() - null_count == intervals->length(), |
1514 | "Sorted list should contain the same amount of non-NULL intervals as unsorted list" ); |
1515 | |
1516 | return true; |
1517 | } |
1518 | #endif |
1519 | |
1520 | void LinearScan::add_to_list(Interval** first, Interval** prev, Interval* interval) { |
1521 | if (*prev != NULL) { |
1522 | (*prev)->set_next(interval); |
1523 | } else { |
1524 | *first = interval; |
1525 | } |
1526 | *prev = interval; |
1527 | } |
1528 | |
1529 | void LinearScan::create_unhandled_lists(Interval** list1, Interval** list2, bool (is_list1)(const Interval* i), bool (is_list2)(const Interval* i)) { |
1530 | assert(is_sorted(_sorted_intervals), "interval list is not sorted" ); |
1531 | |
1532 | *list1 = *list2 = Interval::end(); |
1533 | |
1534 | Interval* list1_prev = NULL; |
1535 | Interval* list2_prev = NULL; |
1536 | Interval* v; |
1537 | |
1538 | const int n = _sorted_intervals->length(); |
1539 | for (int i = 0; i < n; i++) { |
1540 | v = _sorted_intervals->at(i); |
1541 | if (v == NULL) continue; |
1542 | |
1543 | if (is_list1(v)) { |
1544 | add_to_list(list1, &list1_prev, v); |
1545 | } else if (is_list2 == NULL || is_list2(v)) { |
1546 | add_to_list(list2, &list2_prev, v); |
1547 | } |
1548 | } |
1549 | |
1550 | if (list1_prev != NULL) list1_prev->set_next(Interval::end()); |
1551 | if (list2_prev != NULL) list2_prev->set_next(Interval::end()); |
1552 | |
1553 | assert(list1_prev == NULL || list1_prev->next() == Interval::end(), "linear list ends not with sentinel" ); |
1554 | assert(list2_prev == NULL || list2_prev->next() == Interval::end(), "linear list ends not with sentinel" ); |
1555 | } |
1556 | |
1557 | |
1558 | void LinearScan::sort_intervals_before_allocation() { |
1559 | TIME_LINEAR_SCAN(timer_sort_intervals_before); |
1560 | |
1561 | if (_needs_full_resort) { |
1562 | // There is no known reason why this should occur but just in case... |
1563 | assert(false, "should never occur" ); |
1564 | // Re-sort existing interval list because an Interval::from() has changed |
1565 | _sorted_intervals->sort(interval_cmp); |
1566 | _needs_full_resort = false; |
1567 | } |
1568 | |
1569 | IntervalList* unsorted_list = &_intervals; |
1570 | int unsorted_len = unsorted_list->length(); |
1571 | int sorted_len = 0; |
1572 | int unsorted_idx; |
1573 | int sorted_idx = 0; |
1574 | int sorted_from_max = -1; |
1575 | |
1576 | // calc number of items for sorted list (sorted list must not contain NULL values) |
1577 | for (unsorted_idx = 0; unsorted_idx < unsorted_len; unsorted_idx++) { |
1578 | if (unsorted_list->at(unsorted_idx) != NULL) { |
1579 | sorted_len++; |
1580 | } |
1581 | } |
1582 | IntervalArray* sorted_list = new IntervalArray(sorted_len, sorted_len, NULL); |
1583 | |
1584 | // special sorting algorithm: the original interval-list is almost sorted, |
1585 | // only some intervals are swapped. So this is much faster than a complete QuickSort |
1586 | for (unsorted_idx = 0; unsorted_idx < unsorted_len; unsorted_idx++) { |
1587 | Interval* cur_interval = unsorted_list->at(unsorted_idx); |
1588 | |
1589 | if (cur_interval != NULL) { |
1590 | int cur_from = cur_interval->from(); |
1591 | |
1592 | if (sorted_from_max <= cur_from) { |
1593 | sorted_list->at_put(sorted_idx++, cur_interval); |
1594 | sorted_from_max = cur_interval->from(); |
1595 | } else { |
1596 | // the asumption that the intervals are already sorted failed, |
1597 | // so this interval must be sorted in manually |
1598 | int j; |
1599 | for (j = sorted_idx - 1; j >= 0 && cur_from < sorted_list->at(j)->from(); j--) { |
1600 | sorted_list->at_put(j + 1, sorted_list->at(j)); |
1601 | } |
1602 | sorted_list->at_put(j + 1, cur_interval); |
1603 | sorted_idx++; |
1604 | } |
1605 | } |
1606 | } |
1607 | _sorted_intervals = sorted_list; |
1608 | assert(is_sorted(_sorted_intervals), "intervals unsorted" ); |
1609 | } |
1610 | |
1611 | void LinearScan::sort_intervals_after_allocation() { |
1612 | TIME_LINEAR_SCAN(timer_sort_intervals_after); |
1613 | |
1614 | if (_needs_full_resort) { |
1615 | // Re-sort existing interval list because an Interval::from() has changed |
1616 | _sorted_intervals->sort(interval_cmp); |
1617 | _needs_full_resort = false; |
1618 | } |
1619 | |
1620 | IntervalArray* old_list = _sorted_intervals; |
1621 | IntervalList* new_list = _new_intervals_from_allocation; |
1622 | int old_len = old_list->length(); |
1623 | int new_len = new_list == NULL ? 0 : new_list->length(); |
1624 | |
1625 | if (new_len == 0) { |
1626 | // no intervals have been added during allocation, so sorted list is already up to date |
1627 | assert(is_sorted(_sorted_intervals), "intervals unsorted" ); |
1628 | return; |
1629 | } |
1630 | |
1631 | // conventional sort-algorithm for new intervals |
1632 | new_list->sort(interval_cmp); |
1633 | |
1634 | // merge old and new list (both already sorted) into one combined list |
1635 | int combined_list_len = old_len + new_len; |
1636 | IntervalArray* combined_list = new IntervalArray(combined_list_len, combined_list_len, NULL); |
1637 | int old_idx = 0; |
1638 | int new_idx = 0; |
1639 | |
1640 | while (old_idx + new_idx < old_len + new_len) { |
1641 | if (new_idx >= new_len || (old_idx < old_len && old_list->at(old_idx)->from() <= new_list->at(new_idx)->from())) { |
1642 | combined_list->at_put(old_idx + new_idx, old_list->at(old_idx)); |
1643 | old_idx++; |
1644 | } else { |
1645 | combined_list->at_put(old_idx + new_idx, new_list->at(new_idx)); |
1646 | new_idx++; |
1647 | } |
1648 | } |
1649 | |
1650 | _sorted_intervals = combined_list; |
1651 | assert(is_sorted(_sorted_intervals), "intervals unsorted" ); |
1652 | } |
1653 | |
1654 | |
1655 | void LinearScan::allocate_registers() { |
1656 | TIME_LINEAR_SCAN(timer_allocate_registers); |
1657 | |
1658 | Interval* precolored_cpu_intervals, *not_precolored_cpu_intervals; |
1659 | Interval* precolored_fpu_intervals, *not_precolored_fpu_intervals; |
1660 | |
1661 | // allocate cpu registers |
1662 | create_unhandled_lists(&precolored_cpu_intervals, ¬_precolored_cpu_intervals, |
1663 | is_precolored_cpu_interval, is_virtual_cpu_interval); |
1664 | |
1665 | // allocate fpu registers |
1666 | create_unhandled_lists(&precolored_fpu_intervals, ¬_precolored_fpu_intervals, |
1667 | is_precolored_fpu_interval, is_virtual_fpu_interval); |
1668 | |
1669 | // the fpu interval allocation cannot be moved down below with the fpu section as |
1670 | // the cpu_lsw.walk() changes interval positions. |
1671 | |
1672 | LinearScanWalker cpu_lsw(this, precolored_cpu_intervals, not_precolored_cpu_intervals); |
1673 | cpu_lsw.walk(); |
1674 | cpu_lsw.finish_allocation(); |
1675 | |
1676 | if (has_fpu_registers()) { |
1677 | LinearScanWalker fpu_lsw(this, precolored_fpu_intervals, not_precolored_fpu_intervals); |
1678 | fpu_lsw.walk(); |
1679 | fpu_lsw.finish_allocation(); |
1680 | } |
1681 | } |
1682 | |
1683 | |
1684 | // ********** Phase 6: resolve data flow |
1685 | // (insert moves at edges between blocks if intervals have been split) |
1686 | |
1687 | // wrapper for Interval::split_child_at_op_id that performs a bailout in product mode |
1688 | // instead of returning NULL |
1689 | Interval* LinearScan::split_child_at_op_id(Interval* interval, int op_id, LIR_OpVisitState::OprMode mode) { |
1690 | Interval* result = interval->split_child_at_op_id(op_id, mode); |
1691 | if (result != NULL) { |
1692 | return result; |
1693 | } |
1694 | |
1695 | assert(false, "must find an interval, but do a clean bailout in product mode" ); |
1696 | result = new Interval(LIR_OprDesc::vreg_base); |
1697 | result->assign_reg(0); |
1698 | result->set_type(T_INT); |
1699 | BAILOUT_("LinearScan: interval is NULL" , result); |
1700 | } |
1701 | |
1702 | |
1703 | Interval* LinearScan::interval_at_block_begin(BlockBegin* block, int reg_num) { |
1704 | assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds" ); |
1705 | assert(interval_at(reg_num) != NULL, "no interval found" ); |
1706 | |
1707 | return split_child_at_op_id(interval_at(reg_num), block->first_lir_instruction_id(), LIR_OpVisitState::outputMode); |
1708 | } |
1709 | |
1710 | Interval* LinearScan::interval_at_block_end(BlockBegin* block, int reg_num) { |
1711 | assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds" ); |
1712 | assert(interval_at(reg_num) != NULL, "no interval found" ); |
1713 | |
1714 | return split_child_at_op_id(interval_at(reg_num), block->last_lir_instruction_id() + 1, LIR_OpVisitState::outputMode); |
1715 | } |
1716 | |
1717 | Interval* LinearScan::interval_at_op_id(int reg_num, int op_id) { |
1718 | assert(LinearScan::nof_regs <= reg_num && reg_num < num_virtual_regs(), "register number out of bounds" ); |
1719 | assert(interval_at(reg_num) != NULL, "no interval found" ); |
1720 | |
1721 | return split_child_at_op_id(interval_at(reg_num), op_id, LIR_OpVisitState::inputMode); |
1722 | } |
1723 | |
1724 | |
1725 | void LinearScan::resolve_collect_mappings(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver) { |
1726 | DEBUG_ONLY(move_resolver.check_empty()); |
1727 | |
1728 | const int size = live_set_size(); |
1729 | const ResourceBitMap live_at_edge = to_block->live_in(); |
1730 | |
1731 | // visit all registers where the live_at_edge bit is set |
1732 | for (int r = (int)live_at_edge.get_next_one_offset(0, size); r < size; r = (int)live_at_edge.get_next_one_offset(r + 1, size)) { |
1733 | assert(r < num_virtual_regs(), "live information set for not exisiting interval" ); |
1734 | assert(from_block->live_out().at(r) && to_block->live_in().at(r), "interval not live at this edge" ); |
1735 | |
1736 | Interval* from_interval = interval_at_block_end(from_block, r); |
1737 | Interval* to_interval = interval_at_block_begin(to_block, r); |
1738 | |
1739 | if (from_interval != to_interval && (from_interval->assigned_reg() != to_interval->assigned_reg() || from_interval->assigned_regHi() != to_interval->assigned_regHi())) { |
1740 | // need to insert move instruction |
1741 | move_resolver.add_mapping(from_interval, to_interval); |
1742 | } |
1743 | } |
1744 | } |
1745 | |
1746 | |
1747 | void LinearScan::resolve_find_insert_pos(BlockBegin* from_block, BlockBegin* to_block, MoveResolver &move_resolver) { |
1748 | if (from_block->number_of_sux() <= 1) { |
1749 | TRACE_LINEAR_SCAN(4, tty->print_cr("inserting moves at end of from_block B%d" , from_block->block_id())); |
1750 | |
1751 | LIR_OpList* instructions = from_block->lir()->instructions_list(); |
1752 | LIR_OpBranch* branch = instructions->last()->as_OpBranch(); |
1753 | if (branch != NULL) { |
1754 | // insert moves before branch |
1755 | assert(branch->cond() == lir_cond_always, "block does not end with an unconditional jump" ); |
1756 | move_resolver.set_insert_position(from_block->lir(), instructions->length() - 2); |
1757 | } else { |
1758 | move_resolver.set_insert_position(from_block->lir(), instructions->length() - 1); |
1759 | } |
1760 | |
1761 | } else { |
1762 | TRACE_LINEAR_SCAN(4, tty->print_cr("inserting moves at beginning of to_block B%d" , to_block->block_id())); |
1763 | #ifdef ASSERT |
1764 | assert(from_block->lir()->instructions_list()->at(0)->as_OpLabel() != NULL, "block does not start with a label" ); |
1765 | |
1766 | // because the number of predecessor edges matches the number of |
1767 | // successor edges, blocks which are reached by switch statements |
1768 | // may have be more than one predecessor but it will be guaranteed |
1769 | // that all predecessors will be the same. |
1770 | for (int i = 0; i < to_block->number_of_preds(); i++) { |
1771 | assert(from_block == to_block->pred_at(i), "all critical edges must be broken" ); |
1772 | } |
1773 | #endif |
1774 | |
1775 | move_resolver.set_insert_position(to_block->lir(), 0); |
1776 | } |
1777 | } |
1778 | |
1779 | |
1780 | // insert necessary moves (spilling or reloading) at edges between blocks if interval has been split |
1781 | void LinearScan::resolve_data_flow() { |
1782 | TIME_LINEAR_SCAN(timer_resolve_data_flow); |
1783 | |
1784 | int num_blocks = block_count(); |
1785 | MoveResolver move_resolver(this); |
1786 | ResourceBitMap block_completed(num_blocks); |
1787 | ResourceBitMap already_resolved(num_blocks); |
1788 | |
1789 | int i; |
1790 | for (i = 0; i < num_blocks; i++) { |
1791 | BlockBegin* block = block_at(i); |
1792 | |
1793 | // check if block has only one predecessor and only one successor |
1794 | if (block->number_of_preds() == 1 && block->number_of_sux() == 1 && block->number_of_exception_handlers() == 0) { |
1795 | LIR_OpList* instructions = block->lir()->instructions_list(); |
1796 | assert(instructions->at(0)->code() == lir_label, "block must start with label" ); |
1797 | assert(instructions->last()->code() == lir_branch, "block with successors must end with branch" ); |
1798 | assert(instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block with successor must end with unconditional branch" ); |
1799 | |
1800 | // check if block is empty (only label and branch) |
1801 | if (instructions->length() == 2) { |
1802 | BlockBegin* pred = block->pred_at(0); |
1803 | BlockBegin* sux = block->sux_at(0); |
1804 | |
1805 | // prevent optimization of two consecutive blocks |
1806 | if (!block_completed.at(pred->linear_scan_number()) && !block_completed.at(sux->linear_scan_number())) { |
1807 | 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())); |
1808 | block_completed.set_bit(block->linear_scan_number()); |
1809 | |
1810 | // directly resolve between pred and sux (without looking at the empty block between) |
1811 | resolve_collect_mappings(pred, sux, move_resolver); |
1812 | if (move_resolver.has_mappings()) { |
1813 | move_resolver.set_insert_position(block->lir(), 0); |
1814 | move_resolver.resolve_and_append_moves(); |
1815 | } |
1816 | } |
1817 | } |
1818 | } |
1819 | } |
1820 | |
1821 | |
1822 | for (i = 0; i < num_blocks; i++) { |
1823 | if (!block_completed.at(i)) { |
1824 | BlockBegin* from_block = block_at(i); |
1825 | already_resolved.set_from(block_completed); |
1826 | |
1827 | int num_sux = from_block->number_of_sux(); |
1828 | for (int s = 0; s < num_sux; s++) { |
1829 | BlockBegin* to_block = from_block->sux_at(s); |
1830 | |
1831 | // check for duplicate edges between the same blocks (can happen with switch blocks) |
1832 | if (!already_resolved.at(to_block->linear_scan_number())) { |
1833 | TRACE_LINEAR_SCAN(3, tty->print_cr("**** processing edge between B%d and B%d" , from_block->block_id(), to_block->block_id())); |
1834 | already_resolved.set_bit(to_block->linear_scan_number()); |
1835 | |
1836 | // collect all intervals that have been split between from_block and to_block |
1837 | resolve_collect_mappings(from_block, to_block, move_resolver); |
1838 | if (move_resolver.has_mappings()) { |
1839 | resolve_find_insert_pos(from_block, to_block, move_resolver); |
1840 | move_resolver.resolve_and_append_moves(); |
1841 | } |
1842 | } |
1843 | } |
1844 | } |
1845 | } |
1846 | } |
1847 | |
1848 | |
1849 | void LinearScan::resolve_exception_entry(BlockBegin* block, int reg_num, MoveResolver &move_resolver) { |
1850 | if (interval_at(reg_num) == NULL) { |
1851 | // if a phi function is never used, no interval is created -> ignore this |
1852 | return; |
1853 | } |
1854 | |
1855 | Interval* interval = interval_at_block_begin(block, reg_num); |
1856 | int reg = interval->assigned_reg(); |
1857 | int regHi = interval->assigned_regHi(); |
1858 | |
1859 | if ((reg < nof_regs && interval->always_in_memory()) || |
1860 | (use_fpu_stack_allocation() && reg >= pd_first_fpu_reg && reg <= pd_last_fpu_reg)) { |
1861 | // the interval is split to get a short range that is located on the stack |
1862 | // in the following two cases: |
1863 | // * the interval started in memory (e.g. method parameter), but is currently in a register |
1864 | // this is an optimization for exception handling that reduces the number of moves that |
1865 | // are necessary for resolving the states when an exception uses this exception handler |
1866 | // * the interval would be on the fpu stack at the begin of the exception handler |
1867 | // this is not allowed because of the complicated fpu stack handling on Intel |
1868 | |
1869 | // range that will be spilled to memory |
1870 | int from_op_id = block->first_lir_instruction_id(); |
1871 | int to_op_id = from_op_id + 1; // short live range of length 1 |
1872 | assert(interval->from() <= from_op_id && interval->to() >= to_op_id, |
1873 | "no split allowed between exception entry and first instruction" ); |
1874 | |
1875 | if (interval->from() != from_op_id) { |
1876 | // the part before from_op_id is unchanged |
1877 | interval = interval->split(from_op_id); |
1878 | interval->assign_reg(reg, regHi); |
1879 | append_interval(interval); |
1880 | } else { |
1881 | _needs_full_resort = true; |
1882 | } |
1883 | assert(interval->from() == from_op_id, "must be true now" ); |
1884 | |
1885 | Interval* spilled_part = interval; |
1886 | if (interval->to() != to_op_id) { |
1887 | // the part after to_op_id is unchanged |
1888 | spilled_part = interval->split_from_start(to_op_id); |
1889 | append_interval(spilled_part); |
1890 | move_resolver.add_mapping(spilled_part, interval); |
1891 | } |
1892 | assign_spill_slot(spilled_part); |
1893 | |
1894 | assert(spilled_part->from() == from_op_id && spilled_part->to() == to_op_id, "just checking" ); |
1895 | } |
1896 | } |
1897 | |
1898 | void LinearScan::resolve_exception_entry(BlockBegin* block, MoveResolver &move_resolver) { |
1899 | assert(block->is_set(BlockBegin::exception_entry_flag), "should not call otherwise" ); |
1900 | DEBUG_ONLY(move_resolver.check_empty()); |
1901 | |
1902 | // visit all registers where the live_in bit is set |
1903 | int size = live_set_size(); |
1904 | for (int r = (int)block->live_in().get_next_one_offset(0, size); r < size; r = (int)block->live_in().get_next_one_offset(r + 1, size)) { |
1905 | resolve_exception_entry(block, r, move_resolver); |
1906 | } |
1907 | |
1908 | // the live_in bits are not set for phi functions of the xhandler entry, so iterate them separately |
1909 | for_each_phi_fun(block, phi, |
1910 | if (!phi->is_illegal()) { resolve_exception_entry(block, phi->operand()->vreg_number(), move_resolver); } |
1911 | ); |
1912 | |
1913 | if (move_resolver.has_mappings()) { |
1914 | // insert moves after first instruction |
1915 | move_resolver.set_insert_position(block->lir(), 0); |
1916 | move_resolver.resolve_and_append_moves(); |
1917 | } |
1918 | } |
1919 | |
1920 | |
1921 | void LinearScan::resolve_exception_edge(XHandler* handler, int throwing_op_id, int reg_num, Phi* phi, MoveResolver &move_resolver) { |
1922 | if (interval_at(reg_num) == NULL) { |
1923 | // if a phi function is never used, no interval is created -> ignore this |
1924 | return; |
1925 | } |
1926 | |
1927 | // the computation of to_interval is equal to resolve_collect_mappings, |
1928 | // but from_interval is more complicated because of phi functions |
1929 | BlockBegin* to_block = handler->entry_block(); |
1930 | Interval* to_interval = interval_at_block_begin(to_block, reg_num); |
1931 | |
1932 | if (phi != NULL) { |
1933 | // phi function of the exception entry block |
1934 | // no moves are created for this phi function in the LIR_Generator, so the |
1935 | // interval at the throwing instruction must be searched using the operands |
1936 | // of the phi function |
1937 | Value from_value = phi->operand_at(handler->phi_operand()); |
1938 | |
1939 | // with phi functions it can happen that the same from_value is used in |
1940 | // multiple mappings, so notify move-resolver that this is allowed |
1941 | move_resolver.set_multiple_reads_allowed(); |
1942 | |
1943 | Constant* con = from_value->as_Constant(); |
1944 | if (con != NULL && !con->is_pinned()) { |
1945 | // unpinned constants may have no register, so add mapping from constant to interval |
1946 | move_resolver.add_mapping(LIR_OprFact::value_type(con->type()), to_interval); |
1947 | } else { |
1948 | // search split child at the throwing op_id |
1949 | Interval* from_interval = interval_at_op_id(from_value->operand()->vreg_number(), throwing_op_id); |
1950 | move_resolver.add_mapping(from_interval, to_interval); |
1951 | } |
1952 | |
1953 | } else { |
1954 | // no phi function, so use reg_num also for from_interval |
1955 | // search split child at the throwing op_id |
1956 | Interval* from_interval = interval_at_op_id(reg_num, throwing_op_id); |
1957 | if (from_interval != to_interval) { |
1958 | // optimization to reduce number of moves: when to_interval is on stack and |
1959 | // the stack slot is known to be always correct, then no move is necessary |
1960 | if (!from_interval->always_in_memory() || from_interval->canonical_spill_slot() != to_interval->assigned_reg()) { |
1961 | move_resolver.add_mapping(from_interval, to_interval); |
1962 | } |
1963 | } |
1964 | } |
1965 | } |
1966 | |
1967 | void LinearScan::resolve_exception_edge(XHandler* handler, int throwing_op_id, MoveResolver &move_resolver) { |
1968 | TRACE_LINEAR_SCAN(4, tty->print_cr("resolving exception handler B%d: throwing_op_id=%d" , handler->entry_block()->block_id(), throwing_op_id)); |
1969 | |
1970 | DEBUG_ONLY(move_resolver.check_empty()); |
1971 | assert(handler->lir_op_id() == -1, "already processed this xhandler" ); |
1972 | DEBUG_ONLY(handler->set_lir_op_id(throwing_op_id)); |
1973 | assert(handler->entry_code() == NULL, "code already present" ); |
1974 | |
1975 | // visit all registers where the live_in bit is set |
1976 | BlockBegin* block = handler->entry_block(); |
1977 | int size = live_set_size(); |
1978 | for (int r = (int)block->live_in().get_next_one_offset(0, size); r < size; r = (int)block->live_in().get_next_one_offset(r + 1, size)) { |
1979 | resolve_exception_edge(handler, throwing_op_id, r, NULL, move_resolver); |
1980 | } |
1981 | |
1982 | // the live_in bits are not set for phi functions of the xhandler entry, so iterate them separately |
1983 | for_each_phi_fun(block, phi, |
1984 | if (!phi->is_illegal()) { resolve_exception_edge(handler, throwing_op_id, phi->operand()->vreg_number(), phi, move_resolver); } |
1985 | ); |
1986 | |
1987 | if (move_resolver.has_mappings()) { |
1988 | LIR_List* entry_code = new LIR_List(compilation()); |
1989 | move_resolver.set_insert_position(entry_code, 0); |
1990 | move_resolver.resolve_and_append_moves(); |
1991 | |
1992 | entry_code->jump(handler->entry_block()); |
1993 | handler->set_entry_code(entry_code); |
1994 | } |
1995 | } |
1996 | |
1997 | |
1998 | void LinearScan::resolve_exception_handlers() { |
1999 | MoveResolver move_resolver(this); |
2000 | LIR_OpVisitState visitor; |
2001 | int num_blocks = block_count(); |
2002 | |
2003 | int i; |
2004 | for (i = 0; i < num_blocks; i++) { |
2005 | BlockBegin* block = block_at(i); |
2006 | if (block->is_set(BlockBegin::exception_entry_flag)) { |
2007 | resolve_exception_entry(block, move_resolver); |
2008 | } |
2009 | } |
2010 | |
2011 | for (i = 0; i < num_blocks; i++) { |
2012 | BlockBegin* block = block_at(i); |
2013 | LIR_List* ops = block->lir(); |
2014 | int num_ops = ops->length(); |
2015 | |
2016 | // iterate all instructions of the block. skip the first because it is always a label |
2017 | assert(visitor.no_operands(ops->at(0)), "first operation must always be a label" ); |
2018 | for (int j = 1; j < num_ops; j++) { |
2019 | LIR_Op* op = ops->at(j); |
2020 | int op_id = op->id(); |
2021 | |
2022 | if (op_id != -1 && has_info(op_id)) { |
2023 | // visit operation to collect all operands |
2024 | visitor.visit(op); |
2025 | assert(visitor.info_count() > 0, "should not visit otherwise" ); |
2026 | |
2027 | XHandlers* xhandlers = visitor.all_xhandler(); |
2028 | int n = xhandlers->length(); |
2029 | for (int k = 0; k < n; k++) { |
2030 | resolve_exception_edge(xhandlers->handler_at(k), op_id, move_resolver); |
2031 | } |
2032 | |
2033 | #ifdef ASSERT |
2034 | } else { |
2035 | visitor.visit(op); |
2036 | assert(visitor.all_xhandler()->length() == 0, "missed exception handler" ); |
2037 | #endif |
2038 | } |
2039 | } |
2040 | } |
2041 | } |
2042 | |
2043 | |
2044 | // ********** Phase 7: assign register numbers back to LIR |
2045 | // (includes computation of debug information and oop maps) |
2046 | |
2047 | VMReg LinearScan::vm_reg_for_interval(Interval* interval) { |
2048 | VMReg reg = interval->cached_vm_reg(); |
2049 | if (!reg->is_valid() ) { |
2050 | reg = vm_reg_for_operand(operand_for_interval(interval)); |
2051 | interval->set_cached_vm_reg(reg); |
2052 | } |
2053 | assert(reg == vm_reg_for_operand(operand_for_interval(interval)), "wrong cached value" ); |
2054 | return reg; |
2055 | } |
2056 | |
2057 | VMReg LinearScan::vm_reg_for_operand(LIR_Opr opr) { |
2058 | assert(opr->is_oop(), "currently only implemented for oop operands" ); |
2059 | return frame_map()->regname(opr); |
2060 | } |
2061 | |
2062 | |
2063 | LIR_Opr LinearScan::operand_for_interval(Interval* interval) { |
2064 | LIR_Opr opr = interval->cached_opr(); |
2065 | if (opr->is_illegal()) { |
2066 | opr = calc_operand_for_interval(interval); |
2067 | interval->set_cached_opr(opr); |
2068 | } |
2069 | |
2070 | assert(opr == calc_operand_for_interval(interval), "wrong cached value" ); |
2071 | return opr; |
2072 | } |
2073 | |
2074 | LIR_Opr LinearScan::calc_operand_for_interval(const Interval* interval) { |
2075 | int assigned_reg = interval->assigned_reg(); |
2076 | BasicType type = interval->type(); |
2077 | |
2078 | if (assigned_reg >= nof_regs) { |
2079 | // stack slot |
2080 | assert(interval->assigned_regHi() == any_reg, "must not have hi register" ); |
2081 | return LIR_OprFact::stack(assigned_reg - nof_regs, type); |
2082 | |
2083 | } else { |
2084 | // register |
2085 | switch (type) { |
2086 | case T_OBJECT: { |
2087 | assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register" ); |
2088 | assert(interval->assigned_regHi() == any_reg, "must not have hi register" ); |
2089 | return LIR_OprFact::single_cpu_oop(assigned_reg); |
2090 | } |
2091 | |
2092 | case T_ADDRESS: { |
2093 | assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register" ); |
2094 | assert(interval->assigned_regHi() == any_reg, "must not have hi register" ); |
2095 | return LIR_OprFact::single_cpu_address(assigned_reg); |
2096 | } |
2097 | |
2098 | case T_METADATA: { |
2099 | assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register" ); |
2100 | assert(interval->assigned_regHi() == any_reg, "must not have hi register" ); |
2101 | return LIR_OprFact::single_cpu_metadata(assigned_reg); |
2102 | } |
2103 | |
2104 | #ifdef __SOFTFP__ |
2105 | case T_FLOAT: // fall through |
2106 | #endif // __SOFTFP__ |
2107 | case T_INT: { |
2108 | assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register" ); |
2109 | assert(interval->assigned_regHi() == any_reg, "must not have hi register" ); |
2110 | return LIR_OprFact::single_cpu(assigned_reg); |
2111 | } |
2112 | |
2113 | #ifdef __SOFTFP__ |
2114 | case T_DOUBLE: // fall through |
2115 | #endif // __SOFTFP__ |
2116 | case T_LONG: { |
2117 | int assigned_regHi = interval->assigned_regHi(); |
2118 | assert(assigned_reg >= pd_first_cpu_reg && assigned_reg <= pd_last_cpu_reg, "no cpu register" ); |
2119 | assert(num_physical_regs(T_LONG) == 1 || |
2120 | (assigned_regHi >= pd_first_cpu_reg && assigned_regHi <= pd_last_cpu_reg), "no cpu register" ); |
2121 | |
2122 | assert(assigned_reg != assigned_regHi, "invalid allocation" ); |
2123 | assert(num_physical_regs(T_LONG) == 1 || assigned_reg < assigned_regHi, |
2124 | "register numbers must be sorted (ensure that e.g. a move from eax,ebx to ebx,eax can not occur)" ); |
2125 | assert((assigned_regHi != any_reg) ^ (num_physical_regs(T_LONG) == 1), "must be match" ); |
2126 | if (requires_adjacent_regs(T_LONG)) { |
2127 | assert(assigned_reg % 2 == 0 && assigned_reg + 1 == assigned_regHi, "must be sequential and even" ); |
2128 | } |
2129 | |
2130 | #ifdef _LP64 |
2131 | return LIR_OprFact::double_cpu(assigned_reg, assigned_reg); |
2132 | #else |
2133 | #if defined(SPARC) || defined(PPC32) |
2134 | return LIR_OprFact::double_cpu(assigned_regHi, assigned_reg); |
2135 | #else |
2136 | return LIR_OprFact::double_cpu(assigned_reg, assigned_regHi); |
2137 | #endif // SPARC |
2138 | #endif // LP64 |
2139 | } |
2140 | |
2141 | #ifndef __SOFTFP__ |
2142 | case T_FLOAT: { |
2143 | #ifdef X86 |
2144 | if (UseSSE >= 1) { |
2145 | int last_xmm_reg = pd_last_xmm_reg; |
2146 | #ifdef _LP64 |
2147 | if (UseAVX < 3) { |
2148 | last_xmm_reg = pd_first_xmm_reg + (pd_nof_xmm_regs_frame_map / 2) - 1; |
2149 | } |
2150 | #endif |
2151 | assert(assigned_reg >= pd_first_xmm_reg && assigned_reg <= last_xmm_reg, "no xmm register" ); |
2152 | assert(interval->assigned_regHi() == any_reg, "must not have hi register" ); |
2153 | return LIR_OprFact::single_xmm(assigned_reg - pd_first_xmm_reg); |
2154 | } |
2155 | #endif |
2156 | |
2157 | assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register" ); |
2158 | assert(interval->assigned_regHi() == any_reg, "must not have hi register" ); |
2159 | return LIR_OprFact::single_fpu(assigned_reg - pd_first_fpu_reg); |
2160 | } |
2161 | |
2162 | case T_DOUBLE: { |
2163 | #ifdef X86 |
2164 | if (UseSSE >= 2) { |
2165 | int last_xmm_reg = pd_last_xmm_reg; |
2166 | #ifdef _LP64 |
2167 | if (UseAVX < 3) { |
2168 | last_xmm_reg = pd_first_xmm_reg + (pd_nof_xmm_regs_frame_map / 2) - 1; |
2169 | } |
2170 | #endif |
2171 | assert(assigned_reg >= pd_first_xmm_reg && assigned_reg <= last_xmm_reg, "no xmm register" ); |
2172 | assert(interval->assigned_regHi() == any_reg, "must not have hi register (double xmm values are stored in one register)" ); |
2173 | return LIR_OprFact::double_xmm(assigned_reg - pd_first_xmm_reg); |
2174 | } |
2175 | #endif |
2176 | |
2177 | #ifdef SPARC |
2178 | assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register" ); |
2179 | assert(interval->assigned_regHi() >= pd_first_fpu_reg && interval->assigned_regHi() <= pd_last_fpu_reg, "no fpu register" ); |
2180 | assert(assigned_reg % 2 == 0 && assigned_reg + 1 == interval->assigned_regHi(), "must be sequential and even" ); |
2181 | LIR_Opr result = LIR_OprFact::double_fpu(interval->assigned_regHi() - pd_first_fpu_reg, assigned_reg - pd_first_fpu_reg); |
2182 | #elif defined(ARM32) |
2183 | assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register" ); |
2184 | assert(interval->assigned_regHi() >= pd_first_fpu_reg && interval->assigned_regHi() <= pd_last_fpu_reg, "no fpu register" ); |
2185 | assert(assigned_reg % 2 == 0 && assigned_reg + 1 == interval->assigned_regHi(), "must be sequential and even" ); |
2186 | LIR_Opr result = LIR_OprFact::double_fpu(assigned_reg - pd_first_fpu_reg, interval->assigned_regHi() - pd_first_fpu_reg); |
2187 | #else |
2188 | assert(assigned_reg >= pd_first_fpu_reg && assigned_reg <= pd_last_fpu_reg, "no fpu register" ); |
2189 | assert(interval->assigned_regHi() == any_reg, "must not have hi register (double fpu values are stored in one register on Intel)" ); |
2190 | LIR_Opr result = LIR_OprFact::double_fpu(assigned_reg - pd_first_fpu_reg); |
2191 | #endif |
2192 | return result; |
2193 | } |
2194 | #endif // __SOFTFP__ |
2195 | |
2196 | default: { |
2197 | ShouldNotReachHere(); |
2198 | return LIR_OprFact::illegalOpr; |
2199 | } |
2200 | } |
2201 | } |
2202 | } |
2203 | |
2204 | LIR_Opr LinearScan::canonical_spill_opr(Interval* interval) { |
2205 | assert(interval->canonical_spill_slot() >= nof_regs, "canonical spill slot not set" ); |
2206 | return LIR_OprFact::stack(interval->canonical_spill_slot() - nof_regs, interval->type()); |
2207 | } |
2208 | |
2209 | LIR_Opr LinearScan::color_lir_opr(LIR_Opr opr, int op_id, LIR_OpVisitState::OprMode mode) { |
2210 | assert(opr->is_virtual(), "should not call this otherwise" ); |
2211 | |
2212 | Interval* interval = interval_at(opr->vreg_number()); |
2213 | assert(interval != NULL, "interval must exist" ); |
2214 | |
2215 | if (op_id != -1) { |
2216 | #ifdef ASSERT |
2217 | BlockBegin* block = block_of_op_with_id(op_id); |
2218 | if (block->number_of_sux() <= 1 && op_id == block->last_lir_instruction_id()) { |
2219 | // check if spill moves could have been appended at the end of this block, but |
2220 | // before the branch instruction. So the split child information for this branch would |
2221 | // be incorrect. |
2222 | LIR_OpBranch* branch = block->lir()->instructions_list()->last()->as_OpBranch(); |
2223 | if (branch != NULL) { |
2224 | if (block->live_out().at(opr->vreg_number())) { |
2225 | assert(branch->cond() == lir_cond_always, "block does not end with an unconditional jump" ); |
2226 | 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)" ); |
2227 | } |
2228 | } |
2229 | } |
2230 | #endif |
2231 | |
2232 | // operands are not changed when an interval is split during allocation, |
2233 | // so search the right interval here |
2234 | interval = split_child_at_op_id(interval, op_id, mode); |
2235 | } |
2236 | |
2237 | LIR_Opr res = operand_for_interval(interval); |
2238 | |
2239 | #ifdef X86 |
2240 | // new semantic for is_last_use: not only set on definite end of interval, |
2241 | // but also before hole |
2242 | // This may still miss some cases (e.g. for dead values), but it is not necessary that the |
2243 | // last use information is completely correct |
2244 | // information is only needed for fpu stack allocation |
2245 | if (res->is_fpu_register()) { |
2246 | if (opr->is_last_use() || op_id == interval->to() || (op_id != -1 && interval->has_hole_between(op_id, op_id + 1))) { |
2247 | assert(op_id == -1 || !is_block_begin(op_id), "holes at begin of block may also result from control flow" ); |
2248 | res = res->make_last_use(); |
2249 | } |
2250 | } |
2251 | #endif |
2252 | |
2253 | assert(!gen()->is_vreg_flag_set(opr->vreg_number(), LIRGenerator::callee_saved) || !FrameMap::is_caller_save_register(res), "bad allocation" ); |
2254 | |
2255 | return res; |
2256 | } |
2257 | |
2258 | |
2259 | #ifdef ASSERT |
2260 | // some methods used to check correctness of debug information |
2261 | |
2262 | void assert_no_register_values(GrowableArray<ScopeValue*>* values) { |
2263 | if (values == NULL) { |
2264 | return; |
2265 | } |
2266 | |
2267 | for (int i = 0; i < values->length(); i++) { |
2268 | ScopeValue* value = values->at(i); |
2269 | |
2270 | if (value->is_location()) { |
2271 | Location location = ((LocationValue*)value)->location(); |
2272 | assert(location.where() == Location::on_stack, "value is in register" ); |
2273 | } |
2274 | } |
2275 | } |
2276 | |
2277 | void assert_no_register_values(GrowableArray<MonitorValue*>* values) { |
2278 | if (values == NULL) { |
2279 | return; |
2280 | } |
2281 | |
2282 | for (int i = 0; i < values->length(); i++) { |
2283 | MonitorValue* value = values->at(i); |
2284 | |
2285 | if (value->owner()->is_location()) { |
2286 | Location location = ((LocationValue*)value->owner())->location(); |
2287 | assert(location.where() == Location::on_stack, "owner is in register" ); |
2288 | } |
2289 | assert(value->basic_lock().where() == Location::on_stack, "basic_lock is in register" ); |
2290 | } |
2291 | } |
2292 | |
2293 | void assert_equal(Location l1, Location l2) { |
2294 | assert(l1.where() == l2.where() && l1.type() == l2.type() && l1.offset() == l2.offset(), "" ); |
2295 | } |
2296 | |
2297 | void assert_equal(ScopeValue* v1, ScopeValue* v2) { |
2298 | if (v1->is_location()) { |
2299 | assert(v2->is_location(), "" ); |
2300 | assert_equal(((LocationValue*)v1)->location(), ((LocationValue*)v2)->location()); |
2301 | } else if (v1->is_constant_int()) { |
2302 | assert(v2->is_constant_int(), "" ); |
2303 | assert(((ConstantIntValue*)v1)->value() == ((ConstantIntValue*)v2)->value(), "" ); |
2304 | } else if (v1->is_constant_double()) { |
2305 | assert(v2->is_constant_double(), "" ); |
2306 | assert(((ConstantDoubleValue*)v1)->value() == ((ConstantDoubleValue*)v2)->value(), "" ); |
2307 | } else if (v1->is_constant_long()) { |
2308 | assert(v2->is_constant_long(), "" ); |
2309 | assert(((ConstantLongValue*)v1)->value() == ((ConstantLongValue*)v2)->value(), "" ); |
2310 | } else if (v1->is_constant_oop()) { |
2311 | assert(v2->is_constant_oop(), "" ); |
2312 | assert(((ConstantOopWriteValue*)v1)->value() == ((ConstantOopWriteValue*)v2)->value(), "" ); |
2313 | } else { |
2314 | ShouldNotReachHere(); |
2315 | } |
2316 | } |
2317 | |
2318 | void assert_equal(MonitorValue* m1, MonitorValue* m2) { |
2319 | assert_equal(m1->owner(), m2->owner()); |
2320 | assert_equal(m1->basic_lock(), m2->basic_lock()); |
2321 | } |
2322 | |
2323 | void assert_equal(IRScopeDebugInfo* d1, IRScopeDebugInfo* d2) { |
2324 | assert(d1->scope() == d2->scope(), "not equal" ); |
2325 | assert(d1->bci() == d2->bci(), "not equal" ); |
2326 | |
2327 | if (d1->locals() != NULL) { |
2328 | assert(d1->locals() != NULL && d2->locals() != NULL, "not equal" ); |
2329 | assert(d1->locals()->length() == d2->locals()->length(), "not equal" ); |
2330 | for (int i = 0; i < d1->locals()->length(); i++) { |
2331 | assert_equal(d1->locals()->at(i), d2->locals()->at(i)); |
2332 | } |
2333 | } else { |
2334 | assert(d1->locals() == NULL && d2->locals() == NULL, "not equal" ); |
2335 | } |
2336 | |
2337 | if (d1->expressions() != NULL) { |
2338 | assert(d1->expressions() != NULL && d2->expressions() != NULL, "not equal" ); |
2339 | assert(d1->expressions()->length() == d2->expressions()->length(), "not equal" ); |
2340 | for (int i = 0; i < d1->expressions()->length(); i++) { |
2341 | assert_equal(d1->expressions()->at(i), d2->expressions()->at(i)); |
2342 | } |
2343 | } else { |
2344 | assert(d1->expressions() == NULL && d2->expressions() == NULL, "not equal" ); |
2345 | } |
2346 | |
2347 | if (d1->monitors() != NULL) { |
2348 | assert(d1->monitors() != NULL && d2->monitors() != NULL, "not equal" ); |
2349 | assert(d1->monitors()->length() == d2->monitors()->length(), "not equal" ); |
2350 | for (int i = 0; i < d1->monitors()->length(); i++) { |
2351 | assert_equal(d1->monitors()->at(i), d2->monitors()->at(i)); |
2352 | } |
2353 | } else { |
2354 | assert(d1->monitors() == NULL && d2->monitors() == NULL, "not equal" ); |
2355 | } |
2356 | |
2357 | if (d1->caller() != NULL) { |
2358 | assert(d1->caller() != NULL && d2->caller() != NULL, "not equal" ); |
2359 | assert_equal(d1->caller(), d2->caller()); |
2360 | } else { |
2361 | assert(d1->caller() == NULL && d2->caller() == NULL, "not equal" ); |
2362 | } |
2363 | } |
2364 | |
2365 | void check_stack_depth(CodeEmitInfo* info, int stack_end) { |
2366 | if (info->stack()->bci() != SynchronizationEntryBCI && !info->scope()->method()->is_native()) { |
2367 | Bytecodes::Code code = info->scope()->method()->java_code_at_bci(info->stack()->bci()); |
2368 | switch (code) { |
2369 | case Bytecodes::_ifnull : // fall through |
2370 | case Bytecodes::_ifnonnull : // fall through |
2371 | case Bytecodes::_ifeq : // fall through |
2372 | case Bytecodes::_ifne : // fall through |
2373 | case Bytecodes::_iflt : // fall through |
2374 | case Bytecodes::_ifge : // fall through |
2375 | case Bytecodes::_ifgt : // fall through |
2376 | case Bytecodes::_ifle : // fall through |
2377 | case Bytecodes::_if_icmpeq : // fall through |
2378 | case Bytecodes::_if_icmpne : // fall through |
2379 | case Bytecodes::_if_icmplt : // fall through |
2380 | case Bytecodes::_if_icmpge : // fall through |
2381 | case Bytecodes::_if_icmpgt : // fall through |
2382 | case Bytecodes::_if_icmple : // fall through |
2383 | case Bytecodes::_if_acmpeq : // fall through |
2384 | case Bytecodes::_if_acmpne : |
2385 | assert(stack_end >= -Bytecodes::depth(code), "must have non-empty expression stack at if bytecode" ); |
2386 | break; |
2387 | default: |
2388 | break; |
2389 | } |
2390 | } |
2391 | } |
2392 | |
2393 | #endif // ASSERT |
2394 | |
2395 | |
2396 | IntervalWalker* LinearScan::init_compute_oop_maps() { |
2397 | // setup lists of potential oops for walking |
2398 | Interval* oop_intervals; |
2399 | Interval* non_oop_intervals; |
2400 | |
2401 | create_unhandled_lists(&oop_intervals, &non_oop_intervals, is_oop_interval, NULL); |
2402 | |
2403 | // intervals that have no oops inside need not to be processed |
2404 | // to ensure a walking until the last instruction id, add a dummy interval |
2405 | // with a high operation id |
2406 | non_oop_intervals = new Interval(any_reg); |
2407 | non_oop_intervals->add_range(max_jint - 2, max_jint - 1); |
2408 | |
2409 | return new IntervalWalker(this, oop_intervals, non_oop_intervals); |
2410 | } |
2411 | |
2412 | |
2413 | OopMap* LinearScan::compute_oop_map(IntervalWalker* iw, LIR_Op* op, CodeEmitInfo* info, bool is_call_site) { |
2414 | TRACE_LINEAR_SCAN(3, tty->print_cr("creating oop map at op_id %d" , op->id())); |
2415 | |
2416 | // walk before the current operation -> intervals that start at |
2417 | // the operation (= output operands of the operation) are not |
2418 | // included in the oop map |
2419 | iw->walk_before(op->id()); |
2420 | |
2421 | int frame_size = frame_map()->framesize(); |
2422 | int arg_count = frame_map()->oop_map_arg_count(); |
2423 | OopMap* map = new OopMap(frame_size, arg_count); |
2424 | |
2425 | // Iterate through active intervals |
2426 | for (Interval* interval = iw->active_first(fixedKind); interval != Interval::end(); interval = interval->next()) { |
2427 | int assigned_reg = interval->assigned_reg(); |
2428 | |
2429 | assert(interval->current_from() <= op->id() && op->id() <= interval->current_to(), "interval should not be active otherwise" ); |
2430 | assert(interval->assigned_regHi() == any_reg, "oop must be single word" ); |
2431 | assert(interval->reg_num() >= LIR_OprDesc::vreg_base, "fixed interval found" ); |
2432 | |
2433 | // Check if this range covers the instruction. Intervals that |
2434 | // start or end at the current operation are not included in the |
2435 | // oop map, except in the case of patching moves. For patching |
2436 | // moves, any intervals which end at this instruction are included |
2437 | // in the oop map since we may safepoint while doing the patch |
2438 | // before we've consumed the inputs. |
2439 | if (op->is_patching() || op->id() < interval->current_to()) { |
2440 | |
2441 | // caller-save registers must not be included into oop-maps at calls |
2442 | 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" ); |
2443 | |
2444 | VMReg name = vm_reg_for_interval(interval); |
2445 | set_oop(map, name); |
2446 | |
2447 | // Spill optimization: when the stack value is guaranteed to be always correct, |
2448 | // then it must be added to the oop map even if the interval is currently in a register |
2449 | if (interval->always_in_memory() && |
2450 | op->id() > interval->spill_definition_pos() && |
2451 | interval->assigned_reg() != interval->canonical_spill_slot()) { |
2452 | assert(interval->spill_definition_pos() > 0, "position not set correctly" ); |
2453 | assert(interval->canonical_spill_slot() >= LinearScan::nof_regs, "no spill slot assigned" ); |
2454 | assert(interval->assigned_reg() < LinearScan::nof_regs, "interval is on stack, so stack slot is registered twice" ); |
2455 | |
2456 | set_oop(map, frame_map()->slot_regname(interval->canonical_spill_slot() - LinearScan::nof_regs)); |
2457 | } |
2458 | } |
2459 | } |
2460 | |
2461 | // add oops from lock stack |
2462 | assert(info->stack() != NULL, "CodeEmitInfo must always have a stack" ); |
2463 | int locks_count = info->stack()->total_locks_size(); |
2464 | for (int i = 0; i < locks_count; i++) { |
2465 | set_oop(map, frame_map()->monitor_object_regname(i)); |
2466 | } |
2467 | |
2468 | return map; |
2469 | } |
2470 | |
2471 | |
2472 | void LinearScan::compute_oop_map(IntervalWalker* iw, const LIR_OpVisitState &visitor, LIR_Op* op) { |
2473 | assert(visitor.info_count() > 0, "no oop map needed" ); |
2474 | |
2475 | // compute oop_map only for first CodeEmitInfo |
2476 | // because it is (in most cases) equal for all other infos of the same operation |
2477 | CodeEmitInfo* first_info = visitor.info_at(0); |
2478 | OopMap* first_oop_map = compute_oop_map(iw, op, first_info, visitor.has_call()); |
2479 | |
2480 | for (int i = 0; i < visitor.info_count(); i++) { |
2481 | CodeEmitInfo* info = visitor.info_at(i); |
2482 | OopMap* oop_map = first_oop_map; |
2483 | |
2484 | // compute worst case interpreter size in case of a deoptimization |
2485 | _compilation->update_interpreter_frame_size(info->interpreter_frame_size()); |
2486 | |
2487 | if (info->stack()->locks_size() != first_info->stack()->locks_size()) { |
2488 | // this info has a different number of locks then the precomputed oop map |
2489 | // (possible for lock and unlock instructions) -> compute oop map with |
2490 | // correct lock information |
2491 | oop_map = compute_oop_map(iw, op, info, visitor.has_call()); |
2492 | } |
2493 | |
2494 | if (info->_oop_map == NULL) { |
2495 | info->_oop_map = oop_map; |
2496 | } else { |
2497 | // a CodeEmitInfo can not be shared between different LIR-instructions |
2498 | // because interval splitting can occur anywhere between two instructions |
2499 | // and so the oop maps must be different |
2500 | // -> check if the already set oop_map is exactly the one calculated for this operation |
2501 | assert(info->_oop_map == oop_map, "same CodeEmitInfo used for multiple LIR instructions" ); |
2502 | } |
2503 | } |
2504 | } |
2505 | |
2506 | |
2507 | // frequently used constants |
2508 | // Allocate them with new so they are never destroyed (otherwise, a |
2509 | // forced exit could destroy these objects while they are still in |
2510 | // use). |
2511 | ConstantOopWriteValue* LinearScan::_oop_null_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantOopWriteValue(NULL); |
2512 | ConstantIntValue* LinearScan::_int_m1_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue(-1); |
2513 | ConstantIntValue* LinearScan::_int_0_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue((jint)0); |
2514 | ConstantIntValue* LinearScan::_int_1_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue(1); |
2515 | ConstantIntValue* LinearScan::_int_2_scope_value = new (ResourceObj::C_HEAP, mtCompiler) ConstantIntValue(2); |
2516 | LocationValue* _illegal_value = new (ResourceObj::C_HEAP, mtCompiler) LocationValue(Location()); |
2517 | |
2518 | void LinearScan::init_compute_debug_info() { |
2519 | // cache for frequently used scope values |
2520 | // (cpu registers and stack slots) |
2521 | int cache_size = (LinearScan::nof_cpu_regs + frame_map()->argcount() + max_spills()) * 2; |
2522 | _scope_value_cache = ScopeValueArray(cache_size, cache_size, NULL); |
2523 | } |
2524 | |
2525 | MonitorValue* LinearScan::location_for_monitor_index(int monitor_index) { |
2526 | Location loc; |
2527 | if (!frame_map()->location_for_monitor_object(monitor_index, &loc)) { |
2528 | bailout("too large frame" ); |
2529 | } |
2530 | ScopeValue* object_scope_value = new LocationValue(loc); |
2531 | |
2532 | if (!frame_map()->location_for_monitor_lock(monitor_index, &loc)) { |
2533 | bailout("too large frame" ); |
2534 | } |
2535 | return new MonitorValue(object_scope_value, loc); |
2536 | } |
2537 | |
2538 | LocationValue* LinearScan::location_for_name(int name, Location::Type loc_type) { |
2539 | Location loc; |
2540 | if (!frame_map()->locations_for_slot(name, loc_type, &loc)) { |
2541 | bailout("too large frame" ); |
2542 | } |
2543 | return new LocationValue(loc); |
2544 | } |
2545 | |
2546 | |
2547 | int LinearScan::append_scope_value_for_constant(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values) { |
2548 | assert(opr->is_constant(), "should not be called otherwise" ); |
2549 | |
2550 | LIR_Const* c = opr->as_constant_ptr(); |
2551 | BasicType t = c->type(); |
2552 | switch (t) { |
2553 | case T_OBJECT: { |
2554 | jobject value = c->as_jobject(); |
2555 | if (value == NULL) { |
2556 | scope_values->append(_oop_null_scope_value); |
2557 | } else { |
2558 | scope_values->append(new ConstantOopWriteValue(c->as_jobject())); |
2559 | } |
2560 | return 1; |
2561 | } |
2562 | |
2563 | case T_INT: // fall through |
2564 | case T_FLOAT: { |
2565 | int value = c->as_jint_bits(); |
2566 | switch (value) { |
2567 | case -1: scope_values->append(_int_m1_scope_value); break; |
2568 | case 0: scope_values->append(_int_0_scope_value); break; |
2569 | case 1: scope_values->append(_int_1_scope_value); break; |
2570 | case 2: scope_values->append(_int_2_scope_value); break; |
2571 | default: scope_values->append(new ConstantIntValue(c->as_jint_bits())); break; |
2572 | } |
2573 | return 1; |
2574 | } |
2575 | |
2576 | case T_LONG: // fall through |
2577 | case T_DOUBLE: { |
2578 | #ifdef _LP64 |
2579 | scope_values->append(_int_0_scope_value); |
2580 | scope_values->append(new ConstantLongValue(c->as_jlong_bits())); |
2581 | #else |
2582 | if (hi_word_offset_in_bytes > lo_word_offset_in_bytes) { |
2583 | scope_values->append(new ConstantIntValue(c->as_jint_hi_bits())); |
2584 | scope_values->append(new ConstantIntValue(c->as_jint_lo_bits())); |
2585 | } else { |
2586 | scope_values->append(new ConstantIntValue(c->as_jint_lo_bits())); |
2587 | scope_values->append(new ConstantIntValue(c->as_jint_hi_bits())); |
2588 | } |
2589 | #endif |
2590 | return 2; |
2591 | } |
2592 | |
2593 | case T_ADDRESS: { |
2594 | #ifdef _LP64 |
2595 | scope_values->append(new ConstantLongValue(c->as_jint())); |
2596 | #else |
2597 | scope_values->append(new ConstantIntValue(c->as_jint())); |
2598 | #endif |
2599 | return 1; |
2600 | } |
2601 | |
2602 | default: |
2603 | ShouldNotReachHere(); |
2604 | return -1; |
2605 | } |
2606 | } |
2607 | |
2608 | int LinearScan::append_scope_value_for_operand(LIR_Opr opr, GrowableArray<ScopeValue*>* scope_values) { |
2609 | if (opr->is_single_stack()) { |
2610 | int stack_idx = opr->single_stack_ix(); |
2611 | bool is_oop = opr->is_oop_register(); |
2612 | int cache_idx = (stack_idx + LinearScan::nof_cpu_regs) * 2 + (is_oop ? 1 : 0); |
2613 | |
2614 | ScopeValue* sv = _scope_value_cache.at(cache_idx); |
2615 | if (sv == NULL) { |
2616 | Location::Type loc_type = is_oop ? Location::oop : Location::normal; |
2617 | sv = location_for_name(stack_idx, loc_type); |
2618 | _scope_value_cache.at_put(cache_idx, sv); |
2619 | } |
2620 | |
2621 | // check if cached value is correct |
2622 | DEBUG_ONLY(assert_equal(sv, location_for_name(stack_idx, is_oop ? Location::oop : Location::normal))); |
2623 | |
2624 | scope_values->append(sv); |
2625 | return 1; |
2626 | |
2627 | } else if (opr->is_single_cpu()) { |
2628 | bool is_oop = opr->is_oop_register(); |
2629 | int cache_idx = opr->cpu_regnr() * 2 + (is_oop ? 1 : 0); |
2630 | Location::Type int_loc_type = NOT_LP64(Location::normal) LP64_ONLY(Location::int_in_long); |
2631 | |
2632 | ScopeValue* sv = _scope_value_cache.at(cache_idx); |
2633 | if (sv == NULL) { |
2634 | Location::Type loc_type = is_oop ? Location::oop : int_loc_type; |
2635 | VMReg rname = frame_map()->regname(opr); |
2636 | sv = new LocationValue(Location::new_reg_loc(loc_type, rname)); |
2637 | _scope_value_cache.at_put(cache_idx, sv); |
2638 | } |
2639 | |
2640 | // check if cached value is correct |
2641 | DEBUG_ONLY(assert_equal(sv, new LocationValue(Location::new_reg_loc(is_oop ? Location::oop : int_loc_type, frame_map()->regname(opr))))); |
2642 | |
2643 | scope_values->append(sv); |
2644 | return 1; |
2645 | |
2646 | #ifdef X86 |
2647 | } else if (opr->is_single_xmm()) { |
2648 | VMReg rname = opr->as_xmm_float_reg()->as_VMReg(); |
2649 | LocationValue* sv = new LocationValue(Location::new_reg_loc(Location::normal, rname)); |
2650 | |
2651 | scope_values->append(sv); |
2652 | return 1; |
2653 | #endif |
2654 | |
2655 | } else if (opr->is_single_fpu()) { |
2656 | #ifdef X86 |
2657 | // the exact location of fpu stack values is only known |
2658 | // during fpu stack allocation, so the stack allocator object |
2659 | // must be present |
2660 | assert(use_fpu_stack_allocation(), "should not have float stack values without fpu stack allocation (all floats must be SSE2)" ); |
2661 | assert(_fpu_stack_allocator != NULL, "must be present" ); |
2662 | opr = _fpu_stack_allocator->to_fpu_stack(opr); |
2663 | #endif |
2664 | |
2665 | Location::Type loc_type = float_saved_as_double ? Location::float_in_dbl : Location::normal; |
2666 | VMReg rname = frame_map()->fpu_regname(opr->fpu_regnr()); |
2667 | #ifndef __SOFTFP__ |
2668 | #ifndef VM_LITTLE_ENDIAN |
2669 | // On S390 a (single precision) float value occupies only the high |
2670 | // word of the full double register. So when the double register is |
2671 | // stored to memory (e.g. by the RegisterSaver), then the float value |
2672 | // is found at offset 0. I.e. the code below is not needed on S390. |
2673 | #ifndef S390 |
2674 | if (! float_saved_as_double) { |
2675 | // On big endian system, we may have an issue if float registers use only |
2676 | // the low half of the (same) double registers. |
2677 | // Both the float and the double could have the same regnr but would correspond |
2678 | // to two different addresses once saved. |
2679 | |
2680 | // get next safely (no assertion checks) |
2681 | VMReg next = VMRegImpl::as_VMReg(1+rname->value()); |
2682 | if (next->is_reg() && |
2683 | (next->as_FloatRegister() == rname->as_FloatRegister())) { |
2684 | // the back-end does use the same numbering for the double and the float |
2685 | rname = next; // VMReg for the low bits, e.g. the real VMReg for the float |
2686 | } |
2687 | } |
2688 | #endif // !S390 |
2689 | #endif |
2690 | #endif |
2691 | LocationValue* sv = new LocationValue(Location::new_reg_loc(loc_type, rname)); |
2692 | |
2693 | scope_values->append(sv); |
2694 | return 1; |
2695 | |
2696 | } else { |
2697 | // double-size operands |
2698 | |
2699 | ScopeValue* first; |
2700 | ScopeValue* second; |
2701 | |
2702 | if (opr->is_double_stack()) { |
2703 | #ifdef _LP64 |
2704 | Location loc1; |
2705 | Location::Type loc_type = opr->type() == T_LONG ? Location::lng : Location::dbl; |
2706 | if (!frame_map()->locations_for_slot(opr->double_stack_ix(), loc_type, &loc1, NULL)) { |
2707 | bailout("too large frame" ); |
2708 | } |
2709 | // Does this reverse on x86 vs. sparc? |
2710 | first = new LocationValue(loc1); |
2711 | second = _int_0_scope_value; |
2712 | #else |
2713 | Location loc1, loc2; |
2714 | if (!frame_map()->locations_for_slot(opr->double_stack_ix(), Location::normal, &loc1, &loc2)) { |
2715 | bailout("too large frame" ); |
2716 | } |
2717 | first = new LocationValue(loc1); |
2718 | second = new LocationValue(loc2); |
2719 | #endif // _LP64 |
2720 | |
2721 | } else if (opr->is_double_cpu()) { |
2722 | #ifdef _LP64 |
2723 | VMReg rname_first = opr->as_register_lo()->as_VMReg(); |
2724 | first = new LocationValue(Location::new_reg_loc(Location::lng, rname_first)); |
2725 | second = _int_0_scope_value; |
2726 | #else |
2727 | VMReg rname_first = opr->as_register_lo()->as_VMReg(); |
2728 | VMReg rname_second = opr->as_register_hi()->as_VMReg(); |
2729 | |
2730 | if (hi_word_offset_in_bytes < lo_word_offset_in_bytes) { |
2731 | // lo/hi and swapped relative to first and second, so swap them |
2732 | VMReg tmp = rname_first; |
2733 | rname_first = rname_second; |
2734 | rname_second = tmp; |
2735 | } |
2736 | |
2737 | first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first)); |
2738 | second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second)); |
2739 | #endif //_LP64 |
2740 | |
2741 | |
2742 | #ifdef X86 |
2743 | } else if (opr->is_double_xmm()) { |
2744 | assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation" ); |
2745 | VMReg rname_first = opr->as_xmm_double_reg()->as_VMReg(); |
2746 | # ifdef _LP64 |
2747 | first = new LocationValue(Location::new_reg_loc(Location::dbl, rname_first)); |
2748 | second = _int_0_scope_value; |
2749 | # else |
2750 | first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first)); |
2751 | // %%% This is probably a waste but we'll keep things as they were for now |
2752 | if (true) { |
2753 | VMReg rname_second = rname_first->next(); |
2754 | second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second)); |
2755 | } |
2756 | # endif |
2757 | #endif |
2758 | |
2759 | } else if (opr->is_double_fpu()) { |
2760 | // On SPARC, fpu_regnrLo/fpu_regnrHi represents the two halves of |
2761 | // the double as float registers in the native ordering. On X86, |
2762 | // fpu_regnrLo is a FPU stack slot whose VMReg represents |
2763 | // the low-order word of the double and fpu_regnrLo + 1 is the |
2764 | // name for the other half. *first and *second must represent the |
2765 | // least and most significant words, respectively. |
2766 | |
2767 | #ifdef X86 |
2768 | // the exact location of fpu stack values is only known |
2769 | // during fpu stack allocation, so the stack allocator object |
2770 | // must be present |
2771 | assert(use_fpu_stack_allocation(), "should not have float stack values without fpu stack allocation (all floats must be SSE2)" ); |
2772 | assert(_fpu_stack_allocator != NULL, "must be present" ); |
2773 | opr = _fpu_stack_allocator->to_fpu_stack(opr); |
2774 | |
2775 | assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation (only fpu_regnrLo is used)" ); |
2776 | #endif |
2777 | #ifdef SPARC |
2778 | assert(opr->fpu_regnrLo() == opr->fpu_regnrHi() + 1, "assumed in calculation (only fpu_regnrHi is used)" ); |
2779 | #endif |
2780 | #ifdef ARM32 |
2781 | assert(opr->fpu_regnrHi() == opr->fpu_regnrLo() + 1, "assumed in calculation (only fpu_regnrLo is used)" ); |
2782 | #endif |
2783 | #ifdef PPC32 |
2784 | assert(opr->fpu_regnrLo() == opr->fpu_regnrHi(), "assumed in calculation (only fpu_regnrHi is used)" ); |
2785 | #endif |
2786 | |
2787 | #ifdef VM_LITTLE_ENDIAN |
2788 | VMReg rname_first = frame_map()->fpu_regname(opr->fpu_regnrLo()); |
2789 | #else |
2790 | VMReg rname_first = frame_map()->fpu_regname(opr->fpu_regnrHi()); |
2791 | #endif |
2792 | |
2793 | #ifdef _LP64 |
2794 | first = new LocationValue(Location::new_reg_loc(Location::dbl, rname_first)); |
2795 | second = _int_0_scope_value; |
2796 | #else |
2797 | first = new LocationValue(Location::new_reg_loc(Location::normal, rname_first)); |
2798 | // %%% This is probably a waste but we'll keep things as they were for now |
2799 | if (true) { |
2800 | VMReg rname_second = rname_first->next(); |
2801 | second = new LocationValue(Location::new_reg_loc(Location::normal, rname_second)); |
2802 | } |
2803 | #endif |
2804 | |
2805 | } else { |
2806 | ShouldNotReachHere(); |
2807 | first = NULL; |
2808 | second = NULL; |
2809 | } |
2810 | |
2811 | assert(first != NULL && second != NULL, "must be set" ); |
2812 | // The convention the interpreter uses is that the second local |
2813 | // holds the first raw word of the native double representation. |
2814 | // This is actually reasonable, since locals and stack arrays |
2815 | // grow downwards in all implementations. |
2816 | // (If, on some machine, the interpreter's Java locals or stack |
2817 | // were to grow upwards, the embedded doubles would be word-swapped.) |
2818 | scope_values->append(second); |
2819 | scope_values->append(first); |
2820 | return 2; |
2821 | } |
2822 | } |
2823 | |
2824 | |
2825 | int LinearScan::append_scope_value(int op_id, Value value, GrowableArray<ScopeValue*>* scope_values) { |
2826 | if (value != NULL) { |
2827 | LIR_Opr opr = value->operand(); |
2828 | Constant* con = value->as_Constant(); |
2829 | |
2830 | assert(con == NULL || opr->is_virtual() || opr->is_constant() || opr->is_illegal(), "asumption: Constant instructions have only constant operands (or illegal if constant is optimized away)" ); |
2831 | assert(con != NULL || opr->is_virtual(), "asumption: non-Constant instructions have only virtual operands" ); |
2832 | |
2833 | if (con != NULL && !con->is_pinned() && !opr->is_constant()) { |
2834 | // Unpinned constants may have a virtual operand for a part of the lifetime |
2835 | // or may be illegal when it was optimized away, |
2836 | // so always use a constant operand |
2837 | opr = LIR_OprFact::value_type(con->type()); |
2838 | } |
2839 | assert(opr->is_virtual() || opr->is_constant(), "other cases not allowed here" ); |
2840 | |
2841 | if (opr->is_virtual()) { |
2842 | LIR_OpVisitState::OprMode mode = LIR_OpVisitState::inputMode; |
2843 | |
2844 | BlockBegin* block = block_of_op_with_id(op_id); |
2845 | if (block->number_of_sux() == 1 && op_id == block->last_lir_instruction_id()) { |
2846 | // generating debug information for the last instruction of a block. |
2847 | // if this instruction is a branch, spill moves are inserted before this branch |
2848 | // and so the wrong operand would be returned (spill moves at block boundaries are not |
2849 | // considered in the live ranges of intervals) |
2850 | // Solution: use the first op_id of the branch target block instead. |
2851 | if (block->lir()->instructions_list()->last()->as_OpBranch() != NULL) { |
2852 | if (block->live_out().at(opr->vreg_number())) { |
2853 | op_id = block->sux_at(0)->first_lir_instruction_id(); |
2854 | mode = LIR_OpVisitState::outputMode; |
2855 | } |
2856 | } |
2857 | } |
2858 | |
2859 | // Get current location of operand |
2860 | // The operand must be live because debug information is considered when building the intervals |
2861 | // if the interval is not live, color_lir_opr will cause an assertion failure |
2862 | opr = color_lir_opr(opr, op_id, mode); |
2863 | assert(!has_call(op_id) || opr->is_stack() || !is_caller_save(reg_num(opr)), "can not have caller-save register operands at calls" ); |
2864 | |
2865 | // Append to ScopeValue array |
2866 | return append_scope_value_for_operand(opr, scope_values); |
2867 | |
2868 | } else { |
2869 | assert(value->as_Constant() != NULL, "all other instructions have only virtual operands" ); |
2870 | assert(opr->is_constant(), "operand must be constant" ); |
2871 | |
2872 | return append_scope_value_for_constant(opr, scope_values); |
2873 | } |
2874 | } else { |
2875 | // append a dummy value because real value not needed |
2876 | scope_values->append(_illegal_value); |
2877 | return 1; |
2878 | } |
2879 | } |
2880 | |
2881 | |
2882 | IRScopeDebugInfo* LinearScan::compute_debug_info_for_scope(int op_id, IRScope* cur_scope, ValueStack* cur_state, ValueStack* innermost_state) { |
2883 | IRScopeDebugInfo* caller_debug_info = NULL; |
2884 | |
2885 | ValueStack* caller_state = cur_state->caller_state(); |
2886 | if (caller_state != NULL) { |
2887 | // process recursively to compute outermost scope first |
2888 | caller_debug_info = compute_debug_info_for_scope(op_id, cur_scope->caller(), caller_state, innermost_state); |
2889 | } |
2890 | |
2891 | // initialize these to null. |
2892 | // If we don't need deopt info or there are no locals, expressions or monitors, |
2893 | // then these get recorded as no information and avoids the allocation of 0 length arrays. |
2894 | GrowableArray<ScopeValue*>* locals = NULL; |
2895 | GrowableArray<ScopeValue*>* expressions = NULL; |
2896 | GrowableArray<MonitorValue*>* monitors = NULL; |
2897 | |
2898 | // describe local variable values |
2899 | int nof_locals = cur_state->locals_size(); |
2900 | if (nof_locals > 0) { |
2901 | locals = new GrowableArray<ScopeValue*>(nof_locals); |
2902 | |
2903 | int pos = 0; |
2904 | while (pos < nof_locals) { |
2905 | assert(pos < cur_state->locals_size(), "why not?" ); |
2906 | |
2907 | Value local = cur_state->local_at(pos); |
2908 | pos += append_scope_value(op_id, local, locals); |
2909 | |
2910 | assert(locals->length() == pos, "must match" ); |
2911 | } |
2912 | assert(locals->length() == cur_scope->method()->max_locals(), "wrong number of locals" ); |
2913 | assert(locals->length() == cur_state->locals_size(), "wrong number of locals" ); |
2914 | } else if (cur_scope->method()->max_locals() > 0) { |
2915 | assert(cur_state->kind() == ValueStack::EmptyExceptionState, "should be" ); |
2916 | nof_locals = cur_scope->method()->max_locals(); |
2917 | locals = new GrowableArray<ScopeValue*>(nof_locals); |
2918 | for(int i = 0; i < nof_locals; i++) { |
2919 | locals->append(_illegal_value); |
2920 | } |
2921 | } |
2922 | |
2923 | // describe expression stack |
2924 | int nof_stack = cur_state->stack_size(); |
2925 | if (nof_stack > 0) { |
2926 | expressions = new GrowableArray<ScopeValue*>(nof_stack); |
2927 | |
2928 | int pos = 0; |
2929 | while (pos < nof_stack) { |
2930 | Value expression = cur_state->stack_at_inc(pos); |
2931 | append_scope_value(op_id, expression, expressions); |
2932 | |
2933 | assert(expressions->length() == pos, "must match" ); |
2934 | } |
2935 | assert(expressions->length() == cur_state->stack_size(), "wrong number of stack entries" ); |
2936 | } |
2937 | |
2938 | // describe monitors |
2939 | int nof_locks = cur_state->locks_size(); |
2940 | if (nof_locks > 0) { |
2941 | int lock_offset = cur_state->caller_state() != NULL ? cur_state->caller_state()->total_locks_size() : 0; |
2942 | monitors = new GrowableArray<MonitorValue*>(nof_locks); |
2943 | for (int i = 0; i < nof_locks; i++) { |
2944 | monitors->append(location_for_monitor_index(lock_offset + i)); |
2945 | } |
2946 | } |
2947 | |
2948 | return new IRScopeDebugInfo(cur_scope, cur_state->bci(), locals, expressions, monitors, caller_debug_info); |
2949 | } |
2950 | |
2951 | |
2952 | void LinearScan::compute_debug_info(CodeEmitInfo* info, int op_id) { |
2953 | TRACE_LINEAR_SCAN(3, tty->print_cr("creating debug information at op_id %d" , op_id)); |
2954 | |
2955 | IRScope* innermost_scope = info->scope(); |
2956 | ValueStack* innermost_state = info->stack(); |
2957 | |
2958 | assert(innermost_scope != NULL && innermost_state != NULL, "why is it missing?" ); |
2959 | |
2960 | DEBUG_ONLY(check_stack_depth(info, innermost_state->stack_size())); |
2961 | |
2962 | if (info->_scope_debug_info == NULL) { |
2963 | // compute debug information |
2964 | info->_scope_debug_info = compute_debug_info_for_scope(op_id, innermost_scope, innermost_state, innermost_state); |
2965 | } else { |
2966 | // debug information already set. Check that it is correct from the current point of view |
2967 | DEBUG_ONLY(assert_equal(info->_scope_debug_info, compute_debug_info_for_scope(op_id, innermost_scope, innermost_state, innermost_state))); |
2968 | } |
2969 | } |
2970 | |
2971 | |
2972 | void LinearScan::assign_reg_num(LIR_OpList* instructions, IntervalWalker* iw) { |
2973 | LIR_OpVisitState visitor; |
2974 | int num_inst = instructions->length(); |
2975 | bool has_dead = false; |
2976 | |
2977 | for (int j = 0; j < num_inst; j++) { |
2978 | LIR_Op* op = instructions->at(j); |
2979 | if (op == NULL) { // this can happen when spill-moves are removed in eliminate_spill_moves |
2980 | has_dead = true; |
2981 | continue; |
2982 | } |
2983 | int op_id = op->id(); |
2984 | |
2985 | // visit instruction to get list of operands |
2986 | visitor.visit(op); |
2987 | |
2988 | // iterate all modes of the visitor and process all virtual operands |
2989 | for_each_visitor_mode(mode) { |
2990 | int n = visitor.opr_count(mode); |
2991 | for (int k = 0; k < n; k++) { |
2992 | LIR_Opr opr = visitor.opr_at(mode, k); |
2993 | if (opr->is_virtual_register()) { |
2994 | visitor.set_opr_at(mode, k, color_lir_opr(opr, op_id, mode)); |
2995 | } |
2996 | } |
2997 | } |
2998 | |
2999 | if (visitor.info_count() > 0) { |
3000 | // exception handling |
3001 | if (compilation()->has_exception_handlers()) { |
3002 | XHandlers* xhandlers = visitor.all_xhandler(); |
3003 | int n = xhandlers->length(); |
3004 | for (int k = 0; k < n; k++) { |
3005 | XHandler* handler = xhandlers->handler_at(k); |
3006 | if (handler->entry_code() != NULL) { |
3007 | assign_reg_num(handler->entry_code()->instructions_list(), NULL); |
3008 | } |
3009 | } |
3010 | } else { |
3011 | assert(visitor.all_xhandler()->length() == 0, "missed exception handler" ); |
3012 | } |
3013 | |
3014 | // compute oop map |
3015 | assert(iw != NULL, "needed for compute_oop_map" ); |
3016 | compute_oop_map(iw, visitor, op); |
3017 | |
3018 | // compute debug information |
3019 | if (!use_fpu_stack_allocation()) { |
3020 | // compute debug information if fpu stack allocation is not needed. |
3021 | // when fpu stack allocation is needed, the debug information can not |
3022 | // be computed here because the exact location of fpu operands is not known |
3023 | // -> debug information is created inside the fpu stack allocator |
3024 | int n = visitor.info_count(); |
3025 | for (int k = 0; k < n; k++) { |
3026 | compute_debug_info(visitor.info_at(k), op_id); |
3027 | } |
3028 | } |
3029 | } |
3030 | |
3031 | #ifdef ASSERT |
3032 | // make sure we haven't made the op invalid. |
3033 | op->verify(); |
3034 | #endif |
3035 | |
3036 | // remove useless moves |
3037 | if (op->code() == lir_move) { |
3038 | assert(op->as_Op1() != NULL, "move must be LIR_Op1" ); |
3039 | LIR_Op1* move = (LIR_Op1*)op; |
3040 | LIR_Opr src = move->in_opr(); |
3041 | LIR_Opr dst = move->result_opr(); |
3042 | if (dst == src || |
3043 | (!dst->is_pointer() && !src->is_pointer() && |
3044 | src->is_same_register(dst))) { |
3045 | instructions->at_put(j, NULL); |
3046 | has_dead = true; |
3047 | } |
3048 | } |
3049 | } |
3050 | |
3051 | if (has_dead) { |
3052 | // iterate all instructions of the block and remove all null-values. |
3053 | int insert_point = 0; |
3054 | for (int j = 0; j < num_inst; j++) { |
3055 | LIR_Op* op = instructions->at(j); |
3056 | if (op != NULL) { |
3057 | if (insert_point != j) { |
3058 | instructions->at_put(insert_point, op); |
3059 | } |
3060 | insert_point++; |
3061 | } |
3062 | } |
3063 | instructions->trunc_to(insert_point); |
3064 | } |
3065 | } |
3066 | |
3067 | void LinearScan::assign_reg_num() { |
3068 | TIME_LINEAR_SCAN(timer_assign_reg_num); |
3069 | |
3070 | init_compute_debug_info(); |
3071 | IntervalWalker* iw = init_compute_oop_maps(); |
3072 | |
3073 | int num_blocks = block_count(); |
3074 | for (int i = 0; i < num_blocks; i++) { |
3075 | BlockBegin* block = block_at(i); |
3076 | assign_reg_num(block->lir()->instructions_list(), iw); |
3077 | } |
3078 | } |
3079 | |
3080 | |
3081 | void LinearScan::do_linear_scan() { |
3082 | NOT_PRODUCT(_total_timer.begin_method()); |
3083 | |
3084 | number_instructions(); |
3085 | |
3086 | NOT_PRODUCT(print_lir(1, "Before Register Allocation" )); |
3087 | |
3088 | compute_local_live_sets(); |
3089 | compute_global_live_sets(); |
3090 | CHECK_BAILOUT(); |
3091 | |
3092 | build_intervals(); |
3093 | CHECK_BAILOUT(); |
3094 | sort_intervals_before_allocation(); |
3095 | |
3096 | NOT_PRODUCT(print_intervals("Before Register Allocation" )); |
3097 | NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_before_alloc)); |
3098 | |
3099 | allocate_registers(); |
3100 | CHECK_BAILOUT(); |
3101 | |
3102 | resolve_data_flow(); |
3103 | if (compilation()->has_exception_handlers()) { |
3104 | resolve_exception_handlers(); |
3105 | } |
3106 | // fill in number of spill slots into frame_map |
3107 | propagate_spill_slots(); |
3108 | CHECK_BAILOUT(); |
3109 | |
3110 | NOT_PRODUCT(print_intervals("After Register Allocation" )); |
3111 | NOT_PRODUCT(print_lir(2, "LIR after register allocation:" )); |
3112 | |
3113 | sort_intervals_after_allocation(); |
3114 | |
3115 | DEBUG_ONLY(verify()); |
3116 | |
3117 | eliminate_spill_moves(); |
3118 | assign_reg_num(); |
3119 | CHECK_BAILOUT(); |
3120 | |
3121 | NOT_PRODUCT(print_lir(2, "LIR after assignment of register numbers:" )); |
3122 | NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_after_asign)); |
3123 | |
3124 | { TIME_LINEAR_SCAN(timer_allocate_fpu_stack); |
3125 | |
3126 | if (use_fpu_stack_allocation()) { |
3127 | allocate_fpu_stack(); // Only has effect on Intel |
3128 | NOT_PRODUCT(print_lir(2, "LIR after FPU stack allocation:" )); |
3129 | } |
3130 | } |
3131 | |
3132 | { TIME_LINEAR_SCAN(timer_optimize_lir); |
3133 | |
3134 | EdgeMoveOptimizer::optimize(ir()->code()); |
3135 | ControlFlowOptimizer::optimize(ir()->code()); |
3136 | // check that cfg is still correct after optimizations |
3137 | ir()->verify(); |
3138 | } |
3139 | |
3140 | NOT_PRODUCT(print_lir(1, "Before Code Generation" , false)); |
3141 | NOT_PRODUCT(LinearScanStatistic::compute(this, _stat_final)); |
3142 | NOT_PRODUCT(_total_timer.end_method(this)); |
3143 | } |
3144 | |
3145 | |
3146 | // ********** Printing functions |
3147 | |
3148 | #ifndef PRODUCT |
3149 | |
3150 | void LinearScan::print_timers(double total) { |
3151 | _total_timer.print(total); |
3152 | } |
3153 | |
3154 | void LinearScan::print_statistics() { |
3155 | _stat_before_alloc.print("before allocation" ); |
3156 | _stat_after_asign.print("after assignment of register" ); |
3157 | _stat_final.print("after optimization" ); |
3158 | } |
3159 | |
3160 | void LinearScan::print_bitmap(BitMap& b) { |
3161 | for (unsigned int i = 0; i < b.size(); i++) { |
3162 | if (b.at(i)) tty->print("%d " , i); |
3163 | } |
3164 | tty->cr(); |
3165 | } |
3166 | |
3167 | void LinearScan::print_intervals(const char* label) { |
3168 | if (TraceLinearScanLevel >= 1) { |
3169 | int i; |
3170 | tty->cr(); |
3171 | tty->print_cr("%s" , label); |
3172 | |
3173 | for (i = 0; i < interval_count(); i++) { |
3174 | Interval* interval = interval_at(i); |
3175 | if (interval != NULL) { |
3176 | interval->print(); |
3177 | } |
3178 | } |
3179 | |
3180 | tty->cr(); |
3181 | tty->print_cr("--- Basic Blocks ---" ); |
3182 | for (i = 0; i < block_count(); i++) { |
3183 | BlockBegin* block = block_at(i); |
3184 | 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()); |
3185 | } |
3186 | tty->cr(); |
3187 | tty->cr(); |
3188 | } |
3189 | |
3190 | if (PrintCFGToFile) { |
3191 | CFGPrinter::print_intervals(&_intervals, label); |
3192 | } |
3193 | } |
3194 | |
3195 | void LinearScan::print_lir(int level, const char* label, bool hir_valid) { |
3196 | if (TraceLinearScanLevel >= level) { |
3197 | tty->cr(); |
3198 | tty->print_cr("%s" , label); |
3199 | print_LIR(ir()->linear_scan_order()); |
3200 | tty->cr(); |
3201 | } |
3202 | |
3203 | if (level == 1 && PrintCFGToFile) { |
3204 | CFGPrinter::print_cfg(ir()->linear_scan_order(), label, hir_valid, true); |
3205 | } |
3206 | } |
3207 | |
3208 | #endif //PRODUCT |
3209 | |
3210 | |
3211 | // ********** verification functions for allocation |
3212 | // (check that all intervals have a correct register and that no registers are overwritten) |
3213 | #ifdef ASSERT |
3214 | |
3215 | void LinearScan::verify() { |
3216 | TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying intervals ******************************************" )); |
3217 | verify_intervals(); |
3218 | |
3219 | TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying that no oops are in fixed intervals ****************" )); |
3220 | verify_no_oops_in_fixed_intervals(); |
3221 | |
3222 | TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying that unpinned constants are not alive across block boundaries" )); |
3223 | verify_constants(); |
3224 | |
3225 | TRACE_LINEAR_SCAN(2, tty->print_cr("********* verifying register allocation ********************************" )); |
3226 | verify_registers(); |
3227 | |
3228 | TRACE_LINEAR_SCAN(2, tty->print_cr("********* no errors found **********************************************" )); |
3229 | } |
3230 | |
3231 | void LinearScan::verify_intervals() { |
3232 | int len = interval_count(); |
3233 | bool has_error = false; |
3234 | |
3235 | for (int i = 0; i < len; i++) { |
3236 | Interval* i1 = interval_at(i); |
3237 | if (i1 == NULL) continue; |
3238 | |
3239 | i1->check_split_children(); |
3240 | |
3241 | if (i1->reg_num() != i) { |
3242 | tty->print_cr("Interval %d is on position %d in list" , i1->reg_num(), i); i1->print(); tty->cr(); |
3243 | has_error = true; |
3244 | } |
3245 | |
3246 | if (i1->reg_num() >= LIR_OprDesc::vreg_base && i1->type() == T_ILLEGAL) { |
3247 | tty->print_cr("Interval %d has no type assigned" , i1->reg_num()); i1->print(); tty->cr(); |
3248 | has_error = true; |
3249 | } |
3250 | |
3251 | if (i1->assigned_reg() == any_reg) { |
3252 | tty->print_cr("Interval %d has no register assigned" , i1->reg_num()); i1->print(); tty->cr(); |
3253 | has_error = true; |
3254 | } |
3255 | |
3256 | if (i1->assigned_reg() == i1->assigned_regHi()) { |
3257 | tty->print_cr("Interval %d: low and high register equal" , i1->reg_num()); i1->print(); tty->cr(); |
3258 | has_error = true; |
3259 | } |
3260 | |
3261 | if (!is_processed_reg_num(i1->assigned_reg())) { |
3262 | tty->print_cr("Can not have an Interval for an ignored register" ); i1->print(); tty->cr(); |
3263 | has_error = true; |
3264 | } |
3265 | |
3266 | // special intervals that are created in MoveResolver |
3267 | // -> ignore them because the range information has no meaning there |
3268 | if (i1->from() == 1 && i1->to() == 2) continue; |
3269 | |
3270 | if (i1->first() == Range::end()) { |
3271 | tty->print_cr("Interval %d has no Range" , i1->reg_num()); i1->print(); tty->cr(); |
3272 | has_error = true; |
3273 | } |
3274 | |
3275 | for (Range* r = i1->first(); r != Range::end(); r = r->next()) { |
3276 | if (r->from() >= r->to()) { |
3277 | tty->print_cr("Interval %d has zero length range" , i1->reg_num()); i1->print(); tty->cr(); |
3278 | has_error = true; |
3279 | } |
3280 | } |
3281 | |
3282 | for (int j = i + 1; j < len; j++) { |
3283 | Interval* i2 = interval_at(j); |
3284 | if (i2 == NULL || (i2->from() == 1 && i2->to() == 2)) continue; |
3285 | |
3286 | int r1 = i1->assigned_reg(); |
3287 | int r1Hi = i1->assigned_regHi(); |
3288 | int r2 = i2->assigned_reg(); |
3289 | int r2Hi = i2->assigned_regHi(); |
3290 | if ((r1 == r2 || r1 == r2Hi || (r1Hi != any_reg && (r1Hi == r2 || r1Hi == r2Hi))) && i1->intersects(i2)) { |
3291 | tty->print_cr("Intervals %d and %d overlap and have the same register assigned" , i1->reg_num(), i2->reg_num()); |
3292 | i1->print(); tty->cr(); |
3293 | i2->print(); tty->cr(); |
3294 | has_error = true; |
3295 | } |
3296 | } |
3297 | } |
3298 | |
3299 | assert(has_error == false, "register allocation invalid" ); |
3300 | } |
3301 | |
3302 | |
3303 | void LinearScan::verify_no_oops_in_fixed_intervals() { |
3304 | Interval* fixed_intervals; |
3305 | Interval* other_intervals; |
3306 | create_unhandled_lists(&fixed_intervals, &other_intervals, is_precolored_cpu_interval, NULL); |
3307 | |
3308 | // to ensure a walking until the last instruction id, add a dummy interval |
3309 | // with a high operation id |
3310 | other_intervals = new Interval(any_reg); |
3311 | other_intervals->add_range(max_jint - 2, max_jint - 1); |
3312 | IntervalWalker* iw = new IntervalWalker(this, fixed_intervals, other_intervals); |
3313 | |
3314 | LIR_OpVisitState visitor; |
3315 | for (int i = 0; i < block_count(); i++) { |
3316 | BlockBegin* block = block_at(i); |
3317 | |
3318 | LIR_OpList* instructions = block->lir()->instructions_list(); |
3319 | |
3320 | for (int j = 0; j < instructions->length(); j++) { |
3321 | LIR_Op* op = instructions->at(j); |
3322 | int op_id = op->id(); |
3323 | |
3324 | visitor.visit(op); |
3325 | |
3326 | if (visitor.info_count() > 0) { |
3327 | iw->walk_before(op->id()); |
3328 | bool check_live = true; |
3329 | if (op->code() == lir_move) { |
3330 | LIR_Op1* move = (LIR_Op1*)op; |
3331 | check_live = (move->patch_code() == lir_patch_none); |
3332 | } |
3333 | LIR_OpBranch* branch = op->as_OpBranch(); |
3334 | if (branch != NULL && branch->stub() != NULL && branch->stub()->is_exception_throw_stub()) { |
3335 | // Don't bother checking the stub in this case since the |
3336 | // exception stub will never return to normal control flow. |
3337 | check_live = false; |
3338 | } |
3339 | |
3340 | // Make sure none of the fixed registers is live across an |
3341 | // oopmap since we can't handle that correctly. |
3342 | if (check_live) { |
3343 | for (Interval* interval = iw->active_first(fixedKind); |
3344 | interval != Interval::end(); |
3345 | interval = interval->next()) { |
3346 | if (interval->current_to() > op->id() + 1) { |
3347 | // This interval is live out of this op so make sure |
3348 | // that this interval represents some value that's |
3349 | // referenced by this op either as an input or output. |
3350 | bool ok = false; |
3351 | for_each_visitor_mode(mode) { |
3352 | int n = visitor.opr_count(mode); |
3353 | for (int k = 0; k < n; k++) { |
3354 | LIR_Opr opr = visitor.opr_at(mode, k); |
3355 | if (opr->is_fixed_cpu()) { |
3356 | if (interval_at(reg_num(opr)) == interval) { |
3357 | ok = true; |
3358 | break; |
3359 | } |
3360 | int hi = reg_numHi(opr); |
3361 | if (hi != -1 && interval_at(hi) == interval) { |
3362 | ok = true; |
3363 | break; |
3364 | } |
3365 | } |
3366 | } |
3367 | } |
3368 | assert(ok, "fixed intervals should never be live across an oopmap point" ); |
3369 | } |
3370 | } |
3371 | } |
3372 | } |
3373 | |
3374 | // oop-maps at calls do not contain registers, so check is not needed |
3375 | if (!visitor.has_call()) { |
3376 | |
3377 | for_each_visitor_mode(mode) { |
3378 | int n = visitor.opr_count(mode); |
3379 | for (int k = 0; k < n; k++) { |
3380 | LIR_Opr opr = visitor.opr_at(mode, k); |
3381 | |
3382 | if (opr->is_fixed_cpu() && opr->is_oop()) { |
3383 | // operand is a non-virtual cpu register and contains an oop |
3384 | TRACE_LINEAR_SCAN(4, op->print_on(tty); tty->print("checking operand " ); opr->print(); tty->cr()); |
3385 | |
3386 | Interval* interval = interval_at(reg_num(opr)); |
3387 | assert(interval != NULL, "no interval" ); |
3388 | |
3389 | if (mode == LIR_OpVisitState::inputMode) { |
3390 | if (interval->to() >= op_id + 1) { |
3391 | assert(interval->to() < op_id + 2 || |
3392 | interval->has_hole_between(op_id, op_id + 2), |
3393 | "oop input operand live after instruction" ); |
3394 | } |
3395 | } else if (mode == LIR_OpVisitState::outputMode) { |
3396 | if (interval->from() <= op_id - 1) { |
3397 | assert(interval->has_hole_between(op_id - 1, op_id), |
3398 | "oop input operand live after instruction" ); |
3399 | } |
3400 | } |
3401 | } |
3402 | } |
3403 | } |
3404 | } |
3405 | } |
3406 | } |
3407 | } |
3408 | |
3409 | |
3410 | void LinearScan::verify_constants() { |
3411 | int num_regs = num_virtual_regs(); |
3412 | int size = live_set_size(); |
3413 | int num_blocks = block_count(); |
3414 | |
3415 | for (int i = 0; i < num_blocks; i++) { |
3416 | BlockBegin* block = block_at(i); |
3417 | ResourceBitMap live_at_edge = block->live_in(); |
3418 | |
3419 | // visit all registers where the live_at_edge bit is set |
3420 | for (int r = (int)live_at_edge.get_next_one_offset(0, size); r < size; r = (int)live_at_edge.get_next_one_offset(r + 1, size)) { |
3421 | TRACE_LINEAR_SCAN(4, tty->print("checking interval %d of block B%d" , r, block->block_id())); |
3422 | |
3423 | Value value = gen()->instruction_for_vreg(r); |
3424 | |
3425 | assert(value != NULL, "all intervals live across block boundaries must have Value" ); |
3426 | assert(value->operand()->is_register() && value->operand()->is_virtual(), "value must have virtual operand" ); |
3427 | assert(value->operand()->vreg_number() == r, "register number must match" ); |
3428 | // TKR assert(value->as_Constant() == NULL || value->is_pinned(), "only pinned constants can be alive accross block boundaries"); |
3429 | } |
3430 | } |
3431 | } |
3432 | |
3433 | |
3434 | class RegisterVerifier: public StackObj { |
3435 | private: |
3436 | LinearScan* _allocator; |
3437 | BlockList _work_list; // all blocks that must be processed |
3438 | IntervalsList _saved_states; // saved information of previous check |
3439 | |
3440 | // simplified access to methods of LinearScan |
3441 | Compilation* compilation() const { return _allocator->compilation(); } |
3442 | Interval* interval_at(int reg_num) const { return _allocator->interval_at(reg_num); } |
3443 | int reg_num(LIR_Opr opr) const { return _allocator->reg_num(opr); } |
3444 | |
3445 | // currently, only registers are processed |
3446 | int state_size() { return LinearScan::nof_regs; } |
3447 | |
3448 | // accessors |
3449 | IntervalList* state_for_block(BlockBegin* block) { return _saved_states.at(block->block_id()); } |
3450 | void set_state_for_block(BlockBegin* block, IntervalList* saved_state) { _saved_states.at_put(block->block_id(), saved_state); } |
3451 | void add_to_work_list(BlockBegin* block) { if (!_work_list.contains(block)) _work_list.append(block); } |
3452 | |
3453 | // helper functions |
3454 | IntervalList* copy(IntervalList* input_state); |
3455 | void state_put(IntervalList* input_state, int reg, Interval* interval); |
3456 | bool check_state(IntervalList* input_state, int reg, Interval* interval); |
3457 | |
3458 | void process_block(BlockBegin* block); |
3459 | void process_xhandler(XHandler* xhandler, IntervalList* input_state); |
3460 | void process_successor(BlockBegin* block, IntervalList* input_state); |
3461 | void process_operations(LIR_List* ops, IntervalList* input_state); |
3462 | |
3463 | public: |
3464 | RegisterVerifier(LinearScan* allocator) |
3465 | : _allocator(allocator) |
3466 | , _work_list(16) |
3467 | , _saved_states(BlockBegin::number_of_blocks(), BlockBegin::number_of_blocks(), NULL) |
3468 | { } |
3469 | |
3470 | void verify(BlockBegin* start); |
3471 | }; |
3472 | |
3473 | |
3474 | // entry function from LinearScan that starts the verification |
3475 | void LinearScan::verify_registers() { |
3476 | RegisterVerifier verifier(this); |
3477 | verifier.verify(block_at(0)); |
3478 | } |
3479 | |
3480 | |
3481 | void RegisterVerifier::verify(BlockBegin* start) { |
3482 | // setup input registers (method arguments) for first block |
3483 | int input_state_len = state_size(); |
3484 | IntervalList* input_state = new IntervalList(input_state_len, input_state_len, NULL); |
3485 | CallingConvention* args = compilation()->frame_map()->incoming_arguments(); |
3486 | for (int n = 0; n < args->length(); n++) { |
3487 | LIR_Opr opr = args->at(n); |
3488 | if (opr->is_register()) { |
3489 | Interval* interval = interval_at(reg_num(opr)); |
3490 | |
3491 | if (interval->assigned_reg() < state_size()) { |
3492 | input_state->at_put(interval->assigned_reg(), interval); |
3493 | } |
3494 | if (interval->assigned_regHi() != LinearScan::any_reg && interval->assigned_regHi() < state_size()) { |
3495 | input_state->at_put(interval->assigned_regHi(), interval); |
3496 | } |
3497 | } |
3498 | } |
3499 | |
3500 | set_state_for_block(start, input_state); |
3501 | add_to_work_list(start); |
3502 | |
3503 | // main loop for verification |
3504 | do { |
3505 | BlockBegin* block = _work_list.at(0); |
3506 | _work_list.remove_at(0); |
3507 | |
3508 | process_block(block); |
3509 | } while (!_work_list.is_empty()); |
3510 | } |
3511 | |
3512 | void RegisterVerifier::process_block(BlockBegin* block) { |
3513 | TRACE_LINEAR_SCAN(2, tty->cr(); tty->print_cr("process_block B%d" , block->block_id())); |
3514 | |
3515 | // must copy state because it is modified |
3516 | IntervalList* input_state = copy(state_for_block(block)); |
3517 | |
3518 | if (TraceLinearScanLevel >= 4) { |
3519 | tty->print_cr("Input-State of intervals:" ); |
3520 | tty->print(" " ); |
3521 | for (int i = 0; i < state_size(); i++) { |
3522 | if (input_state->at(i) != NULL) { |
3523 | tty->print(" %4d" , input_state->at(i)->reg_num()); |
3524 | } else { |
3525 | tty->print(" __" ); |
3526 | } |
3527 | } |
3528 | tty->cr(); |
3529 | tty->cr(); |
3530 | } |
3531 | |
3532 | // process all operations of the block |
3533 | process_operations(block->lir(), input_state); |
3534 | |
3535 | // iterate all successors |
3536 | for (int i = 0; i < block->number_of_sux(); i++) { |
3537 | process_successor(block->sux_at(i), input_state); |
3538 | } |
3539 | } |
3540 | |
3541 | void RegisterVerifier::process_xhandler(XHandler* xhandler, IntervalList* input_state) { |
3542 | TRACE_LINEAR_SCAN(2, tty->print_cr("process_xhandler B%d" , xhandler->entry_block()->block_id())); |
3543 | |
3544 | // must copy state because it is modified |
3545 | input_state = copy(input_state); |
3546 | |
3547 | if (xhandler->entry_code() != NULL) { |
3548 | process_operations(xhandler->entry_code(), input_state); |
3549 | } |
3550 | process_successor(xhandler->entry_block(), input_state); |
3551 | } |
3552 | |
3553 | void RegisterVerifier::process_successor(BlockBegin* block, IntervalList* input_state) { |
3554 | IntervalList* saved_state = state_for_block(block); |
3555 | |
3556 | if (saved_state != NULL) { |
3557 | // this block was already processed before. |
3558 | // check if new input_state is consistent with saved_state |
3559 | |
3560 | bool saved_state_correct = true; |
3561 | for (int i = 0; i < state_size(); i++) { |
3562 | if (input_state->at(i) != saved_state->at(i)) { |
3563 | // current input_state and previous saved_state assume a different |
3564 | // interval in this register -> assume that this register is invalid |
3565 | if (saved_state->at(i) != NULL) { |
3566 | // invalidate old calculation only if it assumed that |
3567 | // register was valid. when the register was already invalid, |
3568 | // then the old calculation was correct. |
3569 | saved_state_correct = false; |
3570 | saved_state->at_put(i, NULL); |
3571 | |
3572 | TRACE_LINEAR_SCAN(4, tty->print_cr("process_successor B%d: invalidating slot %d" , block->block_id(), i)); |
3573 | } |
3574 | } |
3575 | } |
3576 | |
3577 | if (saved_state_correct) { |
3578 | // already processed block with correct input_state |
3579 | TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: previous visit already correct" , block->block_id())); |
3580 | } else { |
3581 | // must re-visit this block |
3582 | TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: must re-visit because input state changed" , block->block_id())); |
3583 | add_to_work_list(block); |
3584 | } |
3585 | |
3586 | } else { |
3587 | // block was not processed before, so set initial input_state |
3588 | TRACE_LINEAR_SCAN(2, tty->print_cr("process_successor B%d: initial visit" , block->block_id())); |
3589 | |
3590 | set_state_for_block(block, copy(input_state)); |
3591 | add_to_work_list(block); |
3592 | } |
3593 | } |
3594 | |
3595 | |
3596 | IntervalList* RegisterVerifier::copy(IntervalList* input_state) { |
3597 | IntervalList* copy_state = new IntervalList(input_state->length()); |
3598 | copy_state->appendAll(input_state); |
3599 | return copy_state; |
3600 | } |
3601 | |
3602 | void RegisterVerifier::state_put(IntervalList* input_state, int reg, Interval* interval) { |
3603 | if (reg != LinearScan::any_reg && reg < state_size()) { |
3604 | if (interval != NULL) { |
3605 | TRACE_LINEAR_SCAN(4, tty->print_cr(" reg[%d] = %d" , reg, interval->reg_num())); |
3606 | } else if (input_state->at(reg) != NULL) { |
3607 | TRACE_LINEAR_SCAN(4, tty->print_cr(" reg[%d] = NULL" , reg)); |
3608 | } |
3609 | |
3610 | input_state->at_put(reg, interval); |
3611 | } |
3612 | } |
3613 | |
3614 | bool RegisterVerifier::check_state(IntervalList* input_state, int reg, Interval* interval) { |
3615 | if (reg != LinearScan::any_reg && reg < state_size()) { |
3616 | if (input_state->at(reg) != interval) { |
3617 | tty->print_cr("!! Error in register allocation: register %d does not contain interval %d" , reg, interval->reg_num()); |
3618 | return true; |
3619 | } |
3620 | } |
3621 | return false; |
3622 | } |
3623 | |
3624 | void RegisterVerifier::process_operations(LIR_List* ops, IntervalList* input_state) { |
3625 | // visit all instructions of the block |
3626 | LIR_OpVisitState visitor; |
3627 | bool has_error = false; |
3628 | |
3629 | for (int i = 0; i < ops->length(); i++) { |
3630 | LIR_Op* op = ops->at(i); |
3631 | visitor.visit(op); |
3632 | |
3633 | TRACE_LINEAR_SCAN(4, op->print_on(tty)); |
3634 | |
3635 | // check if input operands are correct |
3636 | int j; |
3637 | int n = visitor.opr_count(LIR_OpVisitState::inputMode); |
3638 | for (j = 0; j < n; j++) { |
3639 | LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::inputMode, j); |
3640 | if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) { |
3641 | Interval* interval = interval_at(reg_num(opr)); |
3642 | if (op->id() != -1) { |
3643 | interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::inputMode); |
3644 | } |
3645 | |
3646 | has_error |= check_state(input_state, interval->assigned_reg(), interval->split_parent()); |
3647 | has_error |= check_state(input_state, interval->assigned_regHi(), interval->split_parent()); |
3648 | |
3649 | // When an operand is marked with is_last_use, then the fpu stack allocator |
3650 | // removes the register from the fpu stack -> the register contains no value |
3651 | if (opr->is_last_use()) { |
3652 | state_put(input_state, interval->assigned_reg(), NULL); |
3653 | state_put(input_state, interval->assigned_regHi(), NULL); |
3654 | } |
3655 | } |
3656 | } |
3657 | |
3658 | // invalidate all caller save registers at calls |
3659 | if (visitor.has_call()) { |
3660 | for (j = 0; j < FrameMap::nof_caller_save_cpu_regs(); j++) { |
3661 | state_put(input_state, reg_num(FrameMap::caller_save_cpu_reg_at(j)), NULL); |
3662 | } |
3663 | for (j = 0; j < FrameMap::nof_caller_save_fpu_regs; j++) { |
3664 | state_put(input_state, reg_num(FrameMap::caller_save_fpu_reg_at(j)), NULL); |
3665 | } |
3666 | |
3667 | #ifdef X86 |
3668 | int num_caller_save_xmm_regs = FrameMap::get_num_caller_save_xmms(); |
3669 | for (j = 0; j < num_caller_save_xmm_regs; j++) { |
3670 | state_put(input_state, reg_num(FrameMap::caller_save_xmm_reg_at(j)), NULL); |
3671 | } |
3672 | #endif |
3673 | } |
3674 | |
3675 | // process xhandler before output and temp operands |
3676 | XHandlers* xhandlers = visitor.all_xhandler(); |
3677 | n = xhandlers->length(); |
3678 | for (int k = 0; k < n; k++) { |
3679 | process_xhandler(xhandlers->handler_at(k), input_state); |
3680 | } |
3681 | |
3682 | // set temp operands (some operations use temp operands also as output operands, so can't set them NULL) |
3683 | n = visitor.opr_count(LIR_OpVisitState::tempMode); |
3684 | for (j = 0; j < n; j++) { |
3685 | LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::tempMode, j); |
3686 | if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) { |
3687 | Interval* interval = interval_at(reg_num(opr)); |
3688 | if (op->id() != -1) { |
3689 | interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::tempMode); |
3690 | } |
3691 | |
3692 | state_put(input_state, interval->assigned_reg(), interval->split_parent()); |
3693 | state_put(input_state, interval->assigned_regHi(), interval->split_parent()); |
3694 | } |
3695 | } |
3696 | |
3697 | // set output operands |
3698 | n = visitor.opr_count(LIR_OpVisitState::outputMode); |
3699 | for (j = 0; j < n; j++) { |
3700 | LIR_Opr opr = visitor.opr_at(LIR_OpVisitState::outputMode, j); |
3701 | if (opr->is_register() && LinearScan::is_processed_reg_num(reg_num(opr))) { |
3702 | Interval* interval = interval_at(reg_num(opr)); |
3703 | if (op->id() != -1) { |
3704 | interval = interval->split_child_at_op_id(op->id(), LIR_OpVisitState::outputMode); |
3705 | } |
3706 | |
3707 | state_put(input_state, interval->assigned_reg(), interval->split_parent()); |
3708 | state_put(input_state, interval->assigned_regHi(), interval->split_parent()); |
3709 | } |
3710 | } |
3711 | } |
3712 | assert(has_error == false, "Error in register allocation" ); |
3713 | } |
3714 | |
3715 | #endif // ASSERT |
3716 | |
3717 | |
3718 | |
3719 | // **** Implementation of MoveResolver ****************************** |
3720 | |
3721 | MoveResolver::MoveResolver(LinearScan* allocator) : |
3722 | _allocator(allocator), |
3723 | _insert_list(NULL), |
3724 | _insert_idx(-1), |
3725 | _insertion_buffer(), |
3726 | _mapping_from(8), |
3727 | _mapping_from_opr(8), |
3728 | _mapping_to(8), |
3729 | _multiple_reads_allowed(false) |
3730 | { |
3731 | for (int i = 0; i < LinearScan::nof_regs; i++) { |
3732 | _register_blocked[i] = 0; |
3733 | } |
3734 | DEBUG_ONLY(check_empty()); |
3735 | } |
3736 | |
3737 | |
3738 | #ifdef ASSERT |
3739 | |
3740 | void MoveResolver::check_empty() { |
3741 | assert(_mapping_from.length() == 0 && _mapping_from_opr.length() == 0 && _mapping_to.length() == 0, "list must be empty before and after processing" ); |
3742 | for (int i = 0; i < LinearScan::nof_regs; i++) { |
3743 | assert(register_blocked(i) == 0, "register map must be empty before and after processing" ); |
3744 | } |
3745 | assert(_multiple_reads_allowed == false, "must have default value" ); |
3746 | } |
3747 | |
3748 | void MoveResolver::verify_before_resolve() { |
3749 | assert(_mapping_from.length() == _mapping_from_opr.length(), "length must be equal" ); |
3750 | assert(_mapping_from.length() == _mapping_to.length(), "length must be equal" ); |
3751 | assert(_insert_list != NULL && _insert_idx != -1, "insert position not set" ); |
3752 | |
3753 | int i, j; |
3754 | if (!_multiple_reads_allowed) { |
3755 | for (i = 0; i < _mapping_from.length(); i++) { |
3756 | for (j = i + 1; j < _mapping_from.length(); j++) { |
3757 | assert(_mapping_from.at(i) == NULL || _mapping_from.at(i) != _mapping_from.at(j), "cannot read from same interval twice" ); |
3758 | } |
3759 | } |
3760 | } |
3761 | |
3762 | for (i = 0; i < _mapping_to.length(); i++) { |
3763 | for (j = i + 1; j < _mapping_to.length(); j++) { |
3764 | assert(_mapping_to.at(i) != _mapping_to.at(j), "cannot write to same interval twice" ); |
3765 | } |
3766 | } |
3767 | |
3768 | |
3769 | ResourceBitMap used_regs(LinearScan::nof_regs + allocator()->frame_map()->argcount() + allocator()->max_spills()); |
3770 | if (!_multiple_reads_allowed) { |
3771 | for (i = 0; i < _mapping_from.length(); i++) { |
3772 | Interval* it = _mapping_from.at(i); |
3773 | if (it != NULL) { |
3774 | assert(!used_regs.at(it->assigned_reg()), "cannot read from same register twice" ); |
3775 | used_regs.set_bit(it->assigned_reg()); |
3776 | |
3777 | if (it->assigned_regHi() != LinearScan::any_reg) { |
3778 | assert(!used_regs.at(it->assigned_regHi()), "cannot read from same register twice" ); |
3779 | used_regs.set_bit(it->assigned_regHi()); |
3780 | } |
3781 | } |
3782 | } |
3783 | } |
3784 | |
3785 | used_regs.clear(); |
3786 | for (i = 0; i < _mapping_to.length(); i++) { |
3787 | Interval* it = _mapping_to.at(i); |
3788 | assert(!used_regs.at(it->assigned_reg()), "cannot write to same register twice" ); |
3789 | used_regs.set_bit(it->assigned_reg()); |
3790 | |
3791 | if (it->assigned_regHi() != LinearScan::any_reg) { |
3792 | assert(!used_regs.at(it->assigned_regHi()), "cannot write to same register twice" ); |
3793 | used_regs.set_bit(it->assigned_regHi()); |
3794 | } |
3795 | } |
3796 | |
3797 | used_regs.clear(); |
3798 | for (i = 0; i < _mapping_from.length(); i++) { |
3799 | Interval* it = _mapping_from.at(i); |
3800 | if (it != NULL && it->assigned_reg() >= LinearScan::nof_regs) { |
3801 | used_regs.set_bit(it->assigned_reg()); |
3802 | } |
3803 | } |
3804 | for (i = 0; i < _mapping_to.length(); i++) { |
3805 | Interval* it = _mapping_to.at(i); |
3806 | 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" ); |
3807 | } |
3808 | } |
3809 | |
3810 | #endif // ASSERT |
3811 | |
3812 | |
3813 | // mark assigned_reg and assigned_regHi of the interval as blocked |
3814 | void MoveResolver::block_registers(Interval* it) { |
3815 | int reg = it->assigned_reg(); |
3816 | if (reg < LinearScan::nof_regs) { |
3817 | assert(_multiple_reads_allowed || register_blocked(reg) == 0, "register already marked as used" ); |
3818 | set_register_blocked(reg, 1); |
3819 | } |
3820 | reg = it->assigned_regHi(); |
3821 | if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) { |
3822 | assert(_multiple_reads_allowed || register_blocked(reg) == 0, "register already marked as used" ); |
3823 | set_register_blocked(reg, 1); |
3824 | } |
3825 | } |
3826 | |
3827 | // mark assigned_reg and assigned_regHi of the interval as unblocked |
3828 | void MoveResolver::unblock_registers(Interval* it) { |
3829 | int reg = it->assigned_reg(); |
3830 | if (reg < LinearScan::nof_regs) { |
3831 | assert(register_blocked(reg) > 0, "register already marked as unused" ); |
3832 | set_register_blocked(reg, -1); |
3833 | } |
3834 | reg = it->assigned_regHi(); |
3835 | if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) { |
3836 | assert(register_blocked(reg) > 0, "register already marked as unused" ); |
3837 | set_register_blocked(reg, -1); |
3838 | } |
3839 | } |
3840 | |
3841 | // check if assigned_reg and assigned_regHi of the to-interval are not blocked (or only blocked by from) |
3842 | bool MoveResolver::save_to_process_move(Interval* from, Interval* to) { |
3843 | int from_reg = -1; |
3844 | int from_regHi = -1; |
3845 | if (from != NULL) { |
3846 | from_reg = from->assigned_reg(); |
3847 | from_regHi = from->assigned_regHi(); |
3848 | } |
3849 | |
3850 | int reg = to->assigned_reg(); |
3851 | if (reg < LinearScan::nof_regs) { |
3852 | if (register_blocked(reg) > 1 || (register_blocked(reg) == 1 && reg != from_reg && reg != from_regHi)) { |
3853 | return false; |
3854 | } |
3855 | } |
3856 | reg = to->assigned_regHi(); |
3857 | if (reg != LinearScan::any_reg && reg < LinearScan::nof_regs) { |
3858 | if (register_blocked(reg) > 1 || (register_blocked(reg) == 1 && reg != from_reg && reg != from_regHi)) { |
3859 | return false; |
3860 | } |
3861 | } |
3862 | |
3863 | return true; |
3864 | } |
3865 | |
3866 | |
3867 | void MoveResolver::create_insertion_buffer(LIR_List* list) { |
3868 | assert(!_insertion_buffer.initialized(), "overwriting existing buffer" ); |
3869 | _insertion_buffer.init(list); |
3870 | } |
3871 | |
3872 | void MoveResolver::append_insertion_buffer() { |
3873 | if (_insertion_buffer.initialized()) { |
3874 | _insertion_buffer.lir_list()->append(&_insertion_buffer); |
3875 | } |
3876 | assert(!_insertion_buffer.initialized(), "must be uninitialized now" ); |
3877 | |
3878 | _insert_list = NULL; |
3879 | _insert_idx = -1; |
3880 | } |
3881 | |
3882 | void MoveResolver::insert_move(Interval* from_interval, Interval* to_interval) { |
3883 | assert(from_interval->reg_num() != to_interval->reg_num(), "from and to interval equal" ); |
3884 | assert(from_interval->type() == to_interval->type(), "move between different types" ); |
3885 | assert(_insert_list != NULL && _insert_idx != -1, "must setup insert position first" ); |
3886 | assert(_insertion_buffer.lir_list() == _insert_list, "wrong insertion buffer" ); |
3887 | |
3888 | LIR_Opr from_opr = LIR_OprFact::virtual_register(from_interval->reg_num(), from_interval->type()); |
3889 | LIR_Opr to_opr = LIR_OprFact::virtual_register(to_interval->reg_num(), to_interval->type()); |
3890 | |
3891 | if (!_multiple_reads_allowed) { |
3892 | // the last_use flag is an optimization for FPU stack allocation. When the same |
3893 | // input interval is used in more than one move, then it is too difficult to determine |
3894 | // if this move is really the last use. |
3895 | from_opr = from_opr->make_last_use(); |
3896 | } |
3897 | _insertion_buffer.move(_insert_idx, from_opr, to_opr); |
3898 | |
3899 | 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())); |
3900 | } |
3901 | |
3902 | void MoveResolver::insert_move(LIR_Opr from_opr, Interval* to_interval) { |
3903 | assert(from_opr->type() == to_interval->type(), "move between different types" ); |
3904 | assert(_insert_list != NULL && _insert_idx != -1, "must setup insert position first" ); |
3905 | assert(_insertion_buffer.lir_list() == _insert_list, "wrong insertion buffer" ); |
3906 | |
3907 | LIR_Opr to_opr = LIR_OprFact::virtual_register(to_interval->reg_num(), to_interval->type()); |
3908 | _insertion_buffer.move(_insert_idx, from_opr, to_opr); |
3909 | |
3910 | 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())); |
3911 | } |
3912 | |
3913 | |
3914 | void MoveResolver::resolve_mappings() { |
3915 | TRACE_LINEAR_SCAN(4, tty->print_cr("MoveResolver: resolving mappings for Block B%d, index %d" , _insert_list->block() != NULL ? _insert_list->block()->block_id() : -1, _insert_idx)); |
3916 | DEBUG_ONLY(verify_before_resolve()); |
3917 | |
3918 | // Block all registers that are used as input operands of a move. |
3919 | // When a register is blocked, no move to this register is emitted. |
3920 | // This is necessary for detecting cycles in moves. |
3921 | int i; |
3922 | for (i = _mapping_from.length() - 1; i >= 0; i--) { |
3923 | Interval* from_interval = _mapping_from.at(i); |
3924 | if (from_interval != NULL) { |
3925 | block_registers(from_interval); |
3926 | } |
3927 | } |
3928 | |
3929 | int spill_candidate = -1; |
3930 | while (_mapping_from.length() > 0) { |
3931 | bool processed_interval = false; |
3932 | |
3933 | for (i = _mapping_from.length() - 1; i >= 0; i--) { |
3934 | Interval* from_interval = _mapping_from.at(i); |
3935 | Interval* to_interval = _mapping_to.at(i); |
3936 | |
3937 | if (save_to_process_move(from_interval, to_interval)) { |
3938 | // this inverval can be processed because target is free |
3939 | if (from_interval != NULL) { |
3940 | insert_move(from_interval, to_interval); |
3941 | unblock_registers(from_interval); |
3942 | } else { |
3943 | insert_move(_mapping_from_opr.at(i), to_interval); |
3944 | } |
3945 | _mapping_from.remove_at(i); |
3946 | _mapping_from_opr.remove_at(i); |
3947 | _mapping_to.remove_at(i); |
3948 | |
3949 | processed_interval = true; |
3950 | } else if (from_interval != NULL && from_interval->assigned_reg() < LinearScan::nof_regs) { |
3951 | // this interval cannot be processed now because target is not free |
3952 | // it starts in a register, so it is a possible candidate for spilling |
3953 | spill_candidate = i; |
3954 | } |
3955 | } |
3956 | |
3957 | if (!processed_interval) { |
3958 | // no move could be processed because there is a cycle in the move list |
3959 | // (e.g. r1 -> r2, r2 -> r1), so one interval must be spilled to memory |
3960 | guarantee(spill_candidate != -1, "no interval in register for spilling found" ); |
3961 | |
3962 | // create a new spill interval and assign a stack slot to it |
3963 | Interval* from_interval = _mapping_from.at(spill_candidate); |
3964 | Interval* spill_interval = new Interval(-1); |
3965 | spill_interval->set_type(from_interval->type()); |
3966 | |
3967 | // add a dummy range because real position is difficult to calculate |
3968 | // Note: this range is a special case when the integrity of the allocation is checked |
3969 | spill_interval->add_range(1, 2); |
3970 | |
3971 | // do not allocate a new spill slot for temporary interval, but |
3972 | // use spill slot assigned to from_interval. Otherwise moves from |
3973 | // one stack slot to another can happen (not allowed by LIR_Assembler |
3974 | int spill_slot = from_interval->canonical_spill_slot(); |
3975 | if (spill_slot < 0) { |
3976 | spill_slot = allocator()->allocate_spill_slot(type2spill_size[spill_interval->type()] == 2); |
3977 | from_interval->set_canonical_spill_slot(spill_slot); |
3978 | } |
3979 | spill_interval->assign_reg(spill_slot); |
3980 | allocator()->append_interval(spill_interval); |
3981 | |
3982 | TRACE_LINEAR_SCAN(4, tty->print_cr("created new Interval %d for spilling" , spill_interval->reg_num())); |
3983 | |
3984 | // insert a move from register to stack and update the mapping |
3985 | insert_move(from_interval, spill_interval); |
3986 | _mapping_from.at_put(spill_candidate, spill_interval); |
3987 | unblock_registers(from_interval); |
3988 | } |
3989 | } |
3990 | |
3991 | // reset to default value |
3992 | _multiple_reads_allowed = false; |
3993 | |
3994 | // check that all intervals have been processed |
3995 | DEBUG_ONLY(check_empty()); |
3996 | } |
3997 | |
3998 | |
3999 | void MoveResolver::set_insert_position(LIR_List* insert_list, int insert_idx) { |
4000 | TRACE_LINEAR_SCAN(4, tty->print_cr("MoveResolver: setting insert position to Block B%d, index %d" , insert_list->block() != NULL ? insert_list->block()->block_id() : -1, insert_idx)); |
4001 | assert(_insert_list == NULL && _insert_idx == -1, "use move_insert_position instead of set_insert_position when data already set" ); |
4002 | |
4003 | create_insertion_buffer(insert_list); |
4004 | _insert_list = insert_list; |
4005 | _insert_idx = insert_idx; |
4006 | } |
4007 | |
4008 | void MoveResolver::move_insert_position(LIR_List* insert_list, int insert_idx) { |
4009 | TRACE_LINEAR_SCAN(4, tty->print_cr("MoveResolver: moving insert position to Block B%d, index %d" , insert_list->block() != NULL ? insert_list->block()->block_id() : -1, insert_idx)); |
4010 | |
4011 | if (_insert_list != NULL && (insert_list != _insert_list || insert_idx != _insert_idx)) { |
4012 | // insert position changed -> resolve current mappings |
4013 | resolve_mappings(); |
4014 | } |
4015 | |
4016 | if (insert_list != _insert_list) { |
4017 | // block changed -> append insertion_buffer because it is |
4018 | // bound to a specific block and create a new insertion_buffer |
4019 | append_insertion_buffer(); |
4020 | create_insertion_buffer(insert_list); |
4021 | } |
4022 | |
4023 | _insert_list = insert_list; |
4024 | _insert_idx = insert_idx; |
4025 | } |
4026 | |
4027 | void MoveResolver::add_mapping(Interval* from_interval, Interval* to_interval) { |
4028 | 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())); |
4029 | |
4030 | _mapping_from.append(from_interval); |
4031 | _mapping_from_opr.append(LIR_OprFact::illegalOpr); |
4032 | _mapping_to.append(to_interval); |
4033 | } |
4034 | |
4035 | |
4036 | void MoveResolver::add_mapping(LIR_Opr from_opr, Interval* to_interval) { |
4037 | 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())); |
4038 | assert(from_opr->is_constant(), "only for constants" ); |
4039 | |
4040 | _mapping_from.append(NULL); |
4041 | _mapping_from_opr.append(from_opr); |
4042 | _mapping_to.append(to_interval); |
4043 | } |
4044 | |
4045 | void MoveResolver::resolve_and_append_moves() { |
4046 | if (has_mappings()) { |
4047 | resolve_mappings(); |
4048 | } |
4049 | append_insertion_buffer(); |
4050 | } |
4051 | |
4052 | |
4053 | |
4054 | // **** Implementation of Range ************************************* |
4055 | |
4056 | Range::Range(int from, int to, Range* next) : |
4057 | _from(from), |
4058 | _to(to), |
4059 | _next(next) |
4060 | { |
4061 | } |
4062 | |
4063 | // initialize sentinel |
4064 | Range* Range::_end = NULL; |
4065 | void Range::initialize(Arena* arena) { |
4066 | _end = new (arena) Range(max_jint, max_jint, NULL); |
4067 | } |
4068 | |
4069 | int Range::intersects_at(Range* r2) const { |
4070 | const Range* r1 = this; |
4071 | |
4072 | assert(r1 != NULL && r2 != NULL, "null ranges not allowed" ); |
4073 | assert(r1 != _end && r2 != _end, "empty ranges not allowed" ); |
4074 | |
4075 | do { |
4076 | if (r1->from() < r2->from()) { |
4077 | if (r1->to() <= r2->from()) { |
4078 | r1 = r1->next(); if (r1 == _end) return -1; |
4079 | } else { |
4080 | return r2->from(); |
4081 | } |
4082 | } else if (r2->from() < r1->from()) { |
4083 | if (r2->to() <= r1->from()) { |
4084 | r2 = r2->next(); if (r2 == _end) return -1; |
4085 | } else { |
4086 | return r1->from(); |
4087 | } |
4088 | } else { // r1->from() == r2->from() |
4089 | if (r1->from() == r1->to()) { |
4090 | r1 = r1->next(); if (r1 == _end) return -1; |
4091 | } else if (r2->from() == r2->to()) { |
4092 | r2 = r2->next(); if (r2 == _end) return -1; |
4093 | } else { |
4094 | return r1->from(); |
4095 | } |
4096 | } |
4097 | } while (true); |
4098 | } |
4099 | |
4100 | #ifndef PRODUCT |
4101 | void Range::print(outputStream* out) const { |
4102 | out->print("[%d, %d[ " , _from, _to); |
4103 | } |
4104 | #endif |
4105 | |
4106 | |
4107 | |
4108 | // **** Implementation of Interval ********************************** |
4109 | |
4110 | // initialize sentinel |
4111 | Interval* Interval::_end = NULL; |
4112 | void Interval::initialize(Arena* arena) { |
4113 | Range::initialize(arena); |
4114 | _end = new (arena) Interval(-1); |
4115 | } |
4116 | |
4117 | Interval::Interval(int reg_num) : |
4118 | _reg_num(reg_num), |
4119 | _type(T_ILLEGAL), |
4120 | _first(Range::end()), |
4121 | _use_pos_and_kinds(12), |
4122 | _current(Range::end()), |
4123 | _next(_end), |
4124 | _state(invalidState), |
4125 | _assigned_reg(LinearScan::any_reg), |
4126 | _assigned_regHi(LinearScan::any_reg), |
4127 | _cached_to(-1), |
4128 | _cached_opr(LIR_OprFact::illegalOpr), |
4129 | _cached_vm_reg(VMRegImpl::Bad()), |
4130 | _split_children(NULL), |
4131 | _canonical_spill_slot(-1), |
4132 | _insert_move_when_activated(false), |
4133 | _spill_state(noDefinitionFound), |
4134 | _spill_definition_pos(-1), |
4135 | _register_hint(NULL) |
4136 | { |
4137 | _split_parent = this; |
4138 | _current_split_child = this; |
4139 | } |
4140 | |
4141 | int Interval::calc_to() { |
4142 | assert(_first != Range::end(), "interval has no range" ); |
4143 | |
4144 | Range* r = _first; |
4145 | while (r->next() != Range::end()) { |
4146 | r = r->next(); |
4147 | } |
4148 | return r->to(); |
4149 | } |
4150 | |
4151 | |
4152 | #ifdef ASSERT |
4153 | // consistency check of split-children |
4154 | void Interval::check_split_children() { |
4155 | if (_split_children != NULL && _split_children->length() > 0) { |
4156 | assert(is_split_parent(), "only split parents can have children" ); |
4157 | |
4158 | for (int i = 0; i < _split_children->length(); i++) { |
4159 | Interval* i1 = _split_children->at(i); |
4160 | |
4161 | assert(i1->split_parent() == this, "not a split child of this interval" ); |
4162 | assert(i1->type() == type(), "must be equal for all split children" ); |
4163 | assert(i1->canonical_spill_slot() == canonical_spill_slot(), "must be equal for all split children" ); |
4164 | |
4165 | for (int j = i + 1; j < _split_children->length(); j++) { |
4166 | Interval* i2 = _split_children->at(j); |
4167 | |
4168 | assert(i1->reg_num() != i2->reg_num(), "same register number" ); |
4169 | |
4170 | if (i1->from() < i2->from()) { |
4171 | assert(i1->to() <= i2->from() && i1->to() < i2->to(), "intervals overlapping" ); |
4172 | } else { |
4173 | assert(i2->from() < i1->from(), "intervals start at same op_id" ); |
4174 | assert(i2->to() <= i1->from() && i2->to() < i1->to(), "intervals overlapping" ); |
4175 | } |
4176 | } |
4177 | } |
4178 | } |
4179 | } |
4180 | #endif // ASSERT |
4181 | |
4182 | Interval* Interval::register_hint(bool search_split_child) const { |
4183 | if (!search_split_child) { |
4184 | return _register_hint; |
4185 | } |
4186 | |
4187 | if (_register_hint != NULL) { |
4188 | assert(_register_hint->is_split_parent(), "ony split parents are valid hint registers" ); |
4189 | |
4190 | if (_register_hint->assigned_reg() >= 0 && _register_hint->assigned_reg() < LinearScan::nof_regs) { |
4191 | return _register_hint; |
4192 | |
4193 | } else if (_register_hint->_split_children != NULL && _register_hint->_split_children->length() > 0) { |
4194 | // search the first split child that has a register assigned |
4195 | int len = _register_hint->_split_children->length(); |
4196 | for (int i = 0; i < len; i++) { |
4197 | Interval* cur = _register_hint->_split_children->at(i); |
4198 | |
4199 | if (cur->assigned_reg() >= 0 && cur->assigned_reg() < LinearScan::nof_regs) { |
4200 | return cur; |
4201 | } |
4202 | } |
4203 | } |
4204 | } |
4205 | |
4206 | // no hint interval found that has a register assigned |
4207 | return NULL; |
4208 | } |
4209 | |
4210 | |
4211 | Interval* Interval::split_child_at_op_id(int op_id, LIR_OpVisitState::OprMode mode) { |
4212 | assert(is_split_parent(), "can only be called for split parents" ); |
4213 | assert(op_id >= 0, "invalid op_id (method can not be called for spill moves)" ); |
4214 | |
4215 | Interval* result; |
4216 | if (_split_children == NULL || _split_children->length() == 0) { |
4217 | result = this; |
4218 | } else { |
4219 | result = NULL; |
4220 | int len = _split_children->length(); |
4221 | |
4222 | // in outputMode, the end of the interval (op_id == cur->to()) is not valid |
4223 | int to_offset = (mode == LIR_OpVisitState::outputMode ? 0 : 1); |
4224 | |
4225 | int i; |
4226 | for (i = 0; i < len; i++) { |
4227 | Interval* cur = _split_children->at(i); |
4228 | if (cur->from() <= op_id && op_id < cur->to() + to_offset) { |
4229 | if (i > 0) { |
4230 | // exchange current split child to start of list (faster access for next call) |
4231 | _split_children->at_put(i, _split_children->at(0)); |
4232 | _split_children->at_put(0, cur); |
4233 | } |
4234 | |
4235 | // interval found |
4236 | result = cur; |
4237 | break; |
4238 | } |
4239 | } |
4240 | |
4241 | #ifdef ASSERT |
4242 | for (i = 0; i < len; i++) { |
4243 | Interval* tmp = _split_children->at(i); |
4244 | if (tmp != result && tmp->from() <= op_id && op_id < tmp->to() + to_offset) { |
4245 | tty->print_cr("two valid result intervals found for op_id %d: %d and %d" , op_id, result->reg_num(), tmp->reg_num()); |
4246 | result->print(); |
4247 | tmp->print(); |
4248 | assert(false, "two valid result intervals found" ); |
4249 | } |
4250 | } |
4251 | #endif |
4252 | } |
4253 | |
4254 | assert(result != NULL, "no matching interval found" ); |
4255 | assert(result->covers(op_id, mode), "op_id not covered by interval" ); |
4256 | |
4257 | return result; |
4258 | } |
4259 | |
4260 | |
4261 | // returns the last split child that ends before the given op_id |
4262 | Interval* Interval::split_child_before_op_id(int op_id) { |
4263 | assert(op_id >= 0, "invalid op_id" ); |
4264 | |
4265 | Interval* parent = split_parent(); |
4266 | Interval* result = NULL; |
4267 | |
4268 | assert(parent->_split_children != NULL, "no split children available" ); |
4269 | int len = parent->_split_children->length(); |
4270 | assert(len > 0, "no split children available" ); |
4271 | |
4272 | for (int i = len - 1; i >= 0; i--) { |
4273 | Interval* cur = parent->_split_children->at(i); |
4274 | if (cur->to() <= op_id && (result == NULL || result->to() < cur->to())) { |
4275 | result = cur; |
4276 | } |
4277 | } |
4278 | |
4279 | assert(result != NULL, "no split child found" ); |
4280 | return result; |
4281 | } |
4282 | |
4283 | |
4284 | // Note: use positions are sorted descending -> first use has highest index |
4285 | int Interval::first_usage(IntervalUseKind min_use_kind) const { |
4286 | assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals" ); |
4287 | |
4288 | for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { |
4289 | if (_use_pos_and_kinds.at(i + 1) >= min_use_kind) { |
4290 | return _use_pos_and_kinds.at(i); |
4291 | } |
4292 | } |
4293 | return max_jint; |
4294 | } |
4295 | |
4296 | int Interval::next_usage(IntervalUseKind min_use_kind, int from) const { |
4297 | assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals" ); |
4298 | |
4299 | for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { |
4300 | if (_use_pos_and_kinds.at(i) >= from && _use_pos_and_kinds.at(i + 1) >= min_use_kind) { |
4301 | return _use_pos_and_kinds.at(i); |
4302 | } |
4303 | } |
4304 | return max_jint; |
4305 | } |
4306 | |
4307 | int Interval::next_usage_exact(IntervalUseKind exact_use_kind, int from) const { |
4308 | assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals" ); |
4309 | |
4310 | for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { |
4311 | if (_use_pos_and_kinds.at(i) >= from && _use_pos_and_kinds.at(i + 1) == exact_use_kind) { |
4312 | return _use_pos_and_kinds.at(i); |
4313 | } |
4314 | } |
4315 | return max_jint; |
4316 | } |
4317 | |
4318 | int Interval::previous_usage(IntervalUseKind min_use_kind, int from) const { |
4319 | assert(LinearScan::is_virtual_interval(this), "cannot access use positions for fixed intervals" ); |
4320 | |
4321 | int prev = 0; |
4322 | for (int i = _use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { |
4323 | if (_use_pos_and_kinds.at(i) > from) { |
4324 | return prev; |
4325 | } |
4326 | if (_use_pos_and_kinds.at(i + 1) >= min_use_kind) { |
4327 | prev = _use_pos_and_kinds.at(i); |
4328 | } |
4329 | } |
4330 | return prev; |
4331 | } |
4332 | |
4333 | void Interval::add_use_pos(int pos, IntervalUseKind use_kind) { |
4334 | assert(covers(pos, LIR_OpVisitState::inputMode), "use position not covered by live range" ); |
4335 | |
4336 | // do not add use positions for precolored intervals because |
4337 | // they are never used |
4338 | if (use_kind != noUse && reg_num() >= LIR_OprDesc::vreg_base) { |
4339 | #ifdef ASSERT |
4340 | assert(_use_pos_and_kinds.length() % 2 == 0, "must be" ); |
4341 | for (int i = 0; i < _use_pos_and_kinds.length(); i += 2) { |
4342 | assert(pos <= _use_pos_and_kinds.at(i), "already added a use-position with lower position" ); |
4343 | assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind" ); |
4344 | if (i > 0) { |
4345 | assert(_use_pos_and_kinds.at(i) < _use_pos_and_kinds.at(i - 2), "not sorted descending" ); |
4346 | } |
4347 | } |
4348 | #endif |
4349 | |
4350 | // Note: add_use is called in descending order, so list gets sorted |
4351 | // automatically by just appending new use positions |
4352 | int len = _use_pos_and_kinds.length(); |
4353 | if (len == 0 || _use_pos_and_kinds.at(len - 2) > pos) { |
4354 | _use_pos_and_kinds.append(pos); |
4355 | _use_pos_and_kinds.append(use_kind); |
4356 | } else if (_use_pos_and_kinds.at(len - 1) < use_kind) { |
4357 | assert(_use_pos_and_kinds.at(len - 2) == pos, "list not sorted correctly" ); |
4358 | _use_pos_and_kinds.at_put(len - 1, use_kind); |
4359 | } |
4360 | } |
4361 | } |
4362 | |
4363 | void Interval::add_range(int from, int to) { |
4364 | assert(from < to, "invalid range" ); |
4365 | assert(first() == Range::end() || to < first()->next()->from(), "not inserting at begin of interval" ); |
4366 | assert(from <= first()->to(), "not inserting at begin of interval" ); |
4367 | |
4368 | if (first()->from() <= to) { |
4369 | // join intersecting ranges |
4370 | first()->set_from(MIN2(from, first()->from())); |
4371 | first()->set_to (MAX2(to, first()->to())); |
4372 | } else { |
4373 | // insert new range |
4374 | _first = new Range(from, to, first()); |
4375 | } |
4376 | } |
4377 | |
4378 | Interval* Interval::new_split_child() { |
4379 | // allocate new interval |
4380 | Interval* result = new Interval(-1); |
4381 | result->set_type(type()); |
4382 | |
4383 | Interval* parent = split_parent(); |
4384 | result->_split_parent = parent; |
4385 | result->set_register_hint(parent); |
4386 | |
4387 | // insert new interval in children-list of parent |
4388 | if (parent->_split_children == NULL) { |
4389 | assert(is_split_parent(), "list must be initialized at first split" ); |
4390 | |
4391 | parent->_split_children = new IntervalList(4); |
4392 | parent->_split_children->append(this); |
4393 | } |
4394 | parent->_split_children->append(result); |
4395 | |
4396 | return result; |
4397 | } |
4398 | |
4399 | // split this interval at the specified position and return |
4400 | // the remainder as a new interval. |
4401 | // |
4402 | // when an interval is split, a bi-directional link is established between the original interval |
4403 | // (the split parent) and the intervals that are split off this interval (the split children) |
4404 | // When a split child is split again, the new created interval is also a direct child |
4405 | // of the original parent (there is no tree of split children stored, but a flat list) |
4406 | // All split children are spilled to the same stack slot (stored in _canonical_spill_slot) |
4407 | // |
4408 | // Note: The new interval has no valid reg_num |
4409 | Interval* Interval::split(int split_pos) { |
4410 | assert(LinearScan::is_virtual_interval(this), "cannot split fixed intervals" ); |
4411 | |
4412 | // allocate new interval |
4413 | Interval* result = new_split_child(); |
4414 | |
4415 | // split the ranges |
4416 | Range* prev = NULL; |
4417 | Range* cur = _first; |
4418 | while (cur != Range::end() && cur->to() <= split_pos) { |
4419 | prev = cur; |
4420 | cur = cur->next(); |
4421 | } |
4422 | assert(cur != Range::end(), "split interval after end of last range" ); |
4423 | |
4424 | if (cur->from() < split_pos) { |
4425 | result->_first = new Range(split_pos, cur->to(), cur->next()); |
4426 | cur->set_to(split_pos); |
4427 | cur->set_next(Range::end()); |
4428 | |
4429 | } else { |
4430 | assert(prev != NULL, "split before start of first range" ); |
4431 | result->_first = cur; |
4432 | prev->set_next(Range::end()); |
4433 | } |
4434 | result->_current = result->_first; |
4435 | _cached_to = -1; // clear cached value |
4436 | |
4437 | // split list of use positions |
4438 | int total_len = _use_pos_and_kinds.length(); |
4439 | int start_idx = total_len - 2; |
4440 | while (start_idx >= 0 && _use_pos_and_kinds.at(start_idx) < split_pos) { |
4441 | start_idx -= 2; |
4442 | } |
4443 | |
4444 | intStack new_use_pos_and_kinds(total_len - start_idx); |
4445 | int i; |
4446 | for (i = start_idx + 2; i < total_len; i++) { |
4447 | new_use_pos_and_kinds.append(_use_pos_and_kinds.at(i)); |
4448 | } |
4449 | |
4450 | _use_pos_and_kinds.trunc_to(start_idx + 2); |
4451 | result->_use_pos_and_kinds = _use_pos_and_kinds; |
4452 | _use_pos_and_kinds = new_use_pos_and_kinds; |
4453 | |
4454 | #ifdef ASSERT |
4455 | assert(_use_pos_and_kinds.length() % 2 == 0, "must have use kind for each use pos" ); |
4456 | assert(result->_use_pos_and_kinds.length() % 2 == 0, "must have use kind for each use pos" ); |
4457 | assert(_use_pos_and_kinds.length() + result->_use_pos_and_kinds.length() == total_len, "missed some entries" ); |
4458 | |
4459 | for (i = 0; i < _use_pos_and_kinds.length(); i += 2) { |
4460 | assert(_use_pos_and_kinds.at(i) < split_pos, "must be" ); |
4461 | assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind" ); |
4462 | } |
4463 | for (i = 0; i < result->_use_pos_and_kinds.length(); i += 2) { |
4464 | assert(result->_use_pos_and_kinds.at(i) >= split_pos, "must be" ); |
4465 | assert(result->_use_pos_and_kinds.at(i + 1) >= firstValidKind && result->_use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind" ); |
4466 | } |
4467 | #endif |
4468 | |
4469 | return result; |
4470 | } |
4471 | |
4472 | // split this interval at the specified position and return |
4473 | // the head as a new interval (the original interval is the tail) |
4474 | // |
4475 | // Currently, only the first range can be split, and the new interval |
4476 | // must not have split positions |
4477 | Interval* Interval::split_from_start(int split_pos) { |
4478 | assert(LinearScan::is_virtual_interval(this), "cannot split fixed intervals" ); |
4479 | assert(split_pos > from() && split_pos < to(), "can only split inside interval" ); |
4480 | assert(split_pos > _first->from() && split_pos <= _first->to(), "can only split inside first range" ); |
4481 | assert(first_usage(noUse) > split_pos, "can not split when use positions are present" ); |
4482 | |
4483 | // allocate new interval |
4484 | Interval* result = new_split_child(); |
4485 | |
4486 | // the new created interval has only one range (checked by assertion above), |
4487 | // so the splitting of the ranges is very simple |
4488 | result->add_range(_first->from(), split_pos); |
4489 | |
4490 | if (split_pos == _first->to()) { |
4491 | assert(_first->next() != Range::end(), "must not be at end" ); |
4492 | _first = _first->next(); |
4493 | } else { |
4494 | _first->set_from(split_pos); |
4495 | } |
4496 | |
4497 | return result; |
4498 | } |
4499 | |
4500 | |
4501 | // returns true if the op_id is inside the interval |
4502 | bool Interval::covers(int op_id, LIR_OpVisitState::OprMode mode) const { |
4503 | Range* cur = _first; |
4504 | |
4505 | while (cur != Range::end() && cur->to() < op_id) { |
4506 | cur = cur->next(); |
4507 | } |
4508 | if (cur != Range::end()) { |
4509 | assert(cur->to() != cur->next()->from(), "ranges not separated" ); |
4510 | |
4511 | if (mode == LIR_OpVisitState::outputMode) { |
4512 | return cur->from() <= op_id && op_id < cur->to(); |
4513 | } else { |
4514 | return cur->from() <= op_id && op_id <= cur->to(); |
4515 | } |
4516 | } |
4517 | return false; |
4518 | } |
4519 | |
4520 | // returns true if the interval has any hole between hole_from and hole_to |
4521 | // (even if the hole has only the length 1) |
4522 | bool Interval::has_hole_between(int hole_from, int hole_to) { |
4523 | assert(hole_from < hole_to, "check" ); |
4524 | assert(from() <= hole_from && hole_to <= to(), "index out of interval" ); |
4525 | |
4526 | Range* cur = _first; |
4527 | while (cur != Range::end()) { |
4528 | assert(cur->to() < cur->next()->from(), "no space between ranges" ); |
4529 | |
4530 | // hole-range starts before this range -> hole |
4531 | if (hole_from < cur->from()) { |
4532 | return true; |
4533 | |
4534 | // hole-range completely inside this range -> no hole |
4535 | } else if (hole_to <= cur->to()) { |
4536 | return false; |
4537 | |
4538 | // overlapping of hole-range with this range -> hole |
4539 | } else if (hole_from <= cur->to()) { |
4540 | return true; |
4541 | } |
4542 | |
4543 | cur = cur->next(); |
4544 | } |
4545 | |
4546 | return false; |
4547 | } |
4548 | |
4549 | |
4550 | #ifndef PRODUCT |
4551 | void Interval::print(outputStream* out) const { |
4552 | const char* SpillState2Name[] = { "no definition" , "no spill store" , "one spill store" , "store at definition" , "start in memory" , "no optimization" }; |
4553 | const char* UseKind2Name[] = { "N" , "L" , "S" , "M" }; |
4554 | |
4555 | const char* type_name; |
4556 | LIR_Opr opr = LIR_OprFact::illegal(); |
4557 | if (reg_num() < LIR_OprDesc::vreg_base) { |
4558 | type_name = "fixed" ; |
4559 | // need a temporary operand for fixed intervals because type() cannot be called |
4560 | #ifdef X86 |
4561 | int last_xmm_reg = pd_last_xmm_reg; |
4562 | #ifdef _LP64 |
4563 | if (UseAVX < 3) { |
4564 | last_xmm_reg = pd_first_xmm_reg + (pd_nof_xmm_regs_frame_map / 2) - 1; |
4565 | } |
4566 | #endif |
4567 | #endif |
4568 | if (assigned_reg() >= pd_first_cpu_reg && assigned_reg() <= pd_last_cpu_reg) { |
4569 | opr = LIR_OprFact::single_cpu(assigned_reg()); |
4570 | } else if (assigned_reg() >= pd_first_fpu_reg && assigned_reg() <= pd_last_fpu_reg) { |
4571 | opr = LIR_OprFact::single_fpu(assigned_reg() - pd_first_fpu_reg); |
4572 | #ifdef X86 |
4573 | } else if (assigned_reg() >= pd_first_xmm_reg && assigned_reg() <= last_xmm_reg) { |
4574 | opr = LIR_OprFact::single_xmm(assigned_reg() - pd_first_xmm_reg); |
4575 | #endif |
4576 | } else { |
4577 | ShouldNotReachHere(); |
4578 | } |
4579 | } else { |
4580 | type_name = type2name(type()); |
4581 | if (assigned_reg() != -1 && |
4582 | (LinearScan::num_physical_regs(type()) == 1 || assigned_regHi() != -1)) { |
4583 | opr = LinearScan::calc_operand_for_interval(this); |
4584 | } |
4585 | } |
4586 | |
4587 | out->print("%d %s " , reg_num(), type_name); |
4588 | if (opr->is_valid()) { |
4589 | out->print("\"" ); |
4590 | opr->print(out); |
4591 | out->print("\" " ); |
4592 | } |
4593 | out->print("%d %d " , split_parent()->reg_num(), (register_hint(false) != NULL ? register_hint(false)->reg_num() : -1)); |
4594 | |
4595 | // print ranges |
4596 | Range* cur = _first; |
4597 | while (cur != Range::end()) { |
4598 | cur->print(out); |
4599 | cur = cur->next(); |
4600 | assert(cur != NULL, "range list not closed with range sentinel" ); |
4601 | } |
4602 | |
4603 | // print use positions |
4604 | int prev = 0; |
4605 | assert(_use_pos_and_kinds.length() % 2 == 0, "must be" ); |
4606 | for (int i =_use_pos_and_kinds.length() - 2; i >= 0; i -= 2) { |
4607 | assert(_use_pos_and_kinds.at(i + 1) >= firstValidKind && _use_pos_and_kinds.at(i + 1) <= lastValidKind, "invalid use kind" ); |
4608 | assert(prev < _use_pos_and_kinds.at(i), "use positions not sorted" ); |
4609 | |
4610 | out->print("%d %s " , _use_pos_and_kinds.at(i), UseKind2Name[_use_pos_and_kinds.at(i + 1)]); |
4611 | prev = _use_pos_and_kinds.at(i); |
4612 | } |
4613 | |
4614 | out->print(" \"%s\"" , SpillState2Name[spill_state()]); |
4615 | out->cr(); |
4616 | } |
4617 | #endif |
4618 | |
4619 | |
4620 | |
4621 | // **** Implementation of IntervalWalker **************************** |
4622 | |
4623 | IntervalWalker::IntervalWalker(LinearScan* allocator, Interval* unhandled_fixed_first, Interval* unhandled_any_first) |
4624 | : _compilation(allocator->compilation()) |
4625 | , _allocator(allocator) |
4626 | { |
4627 | _unhandled_first[fixedKind] = unhandled_fixed_first; |
4628 | _unhandled_first[anyKind] = unhandled_any_first; |
4629 | _active_first[fixedKind] = Interval::end(); |
4630 | _inactive_first[fixedKind] = Interval::end(); |
4631 | _active_first[anyKind] = Interval::end(); |
4632 | _inactive_first[anyKind] = Interval::end(); |
4633 | _current_position = -1; |
4634 | _current = NULL; |
4635 | next_interval(); |
4636 | } |
4637 | |
4638 | |
4639 | // append interval in order of current range from() |
4640 | void IntervalWalker::append_sorted(Interval** list, Interval* interval) { |
4641 | Interval* prev = NULL; |
4642 | Interval* cur = *list; |
4643 | while (cur->current_from() < interval->current_from()) { |
4644 | prev = cur; cur = cur->next(); |
4645 | } |
4646 | if (prev == NULL) { |
4647 | *list = interval; |
4648 | } else { |
4649 | prev->set_next(interval); |
4650 | } |
4651 | interval->set_next(cur); |
4652 | } |
4653 | |
4654 | void IntervalWalker::append_to_unhandled(Interval** list, Interval* interval) { |
4655 | assert(interval->from() >= current()->current_from(), "cannot append new interval before current walk position" ); |
4656 | |
4657 | Interval* prev = NULL; |
4658 | Interval* cur = *list; |
4659 | while (cur->from() < interval->from() || (cur->from() == interval->from() && cur->first_usage(noUse) < interval->first_usage(noUse))) { |
4660 | prev = cur; cur = cur->next(); |
4661 | } |
4662 | if (prev == NULL) { |
4663 | *list = interval; |
4664 | } else { |
4665 | prev->set_next(interval); |
4666 | } |
4667 | interval->set_next(cur); |
4668 | } |
4669 | |
4670 | |
4671 | inline bool IntervalWalker::remove_from_list(Interval** list, Interval* i) { |
4672 | while (*list != Interval::end() && *list != i) { |
4673 | list = (*list)->next_addr(); |
4674 | } |
4675 | if (*list != Interval::end()) { |
4676 | assert(*list == i, "check" ); |
4677 | *list = (*list)->next(); |
4678 | return true; |
4679 | } else { |
4680 | return false; |
4681 | } |
4682 | } |
4683 | |
4684 | void IntervalWalker::remove_from_list(Interval* i) { |
4685 | bool deleted; |
4686 | |
4687 | if (i->state() == activeState) { |
4688 | deleted = remove_from_list(active_first_addr(anyKind), i); |
4689 | } else { |
4690 | assert(i->state() == inactiveState, "invalid state" ); |
4691 | deleted = remove_from_list(inactive_first_addr(anyKind), i); |
4692 | } |
4693 | |
4694 | assert(deleted, "interval has not been found in list" ); |
4695 | } |
4696 | |
4697 | |
4698 | void IntervalWalker::walk_to(IntervalState state, int from) { |
4699 | assert (state == activeState || state == inactiveState, "wrong state" ); |
4700 | for_each_interval_kind(kind) { |
4701 | Interval** prev = state == activeState ? active_first_addr(kind) : inactive_first_addr(kind); |
4702 | Interval* next = *prev; |
4703 | while (next->current_from() <= from) { |
4704 | Interval* cur = next; |
4705 | next = cur->next(); |
4706 | |
4707 | bool range_has_changed = false; |
4708 | while (cur->current_to() <= from) { |
4709 | cur->next_range(); |
4710 | range_has_changed = true; |
4711 | } |
4712 | |
4713 | // also handle move from inactive list to active list |
4714 | range_has_changed = range_has_changed || (state == inactiveState && cur->current_from() <= from); |
4715 | |
4716 | if (range_has_changed) { |
4717 | // remove cur from list |
4718 | *prev = next; |
4719 | if (cur->current_at_end()) { |
4720 | // move to handled state (not maintained as a list) |
4721 | cur->set_state(handledState); |
4722 | interval_moved(cur, kind, state, handledState); |
4723 | } else if (cur->current_from() <= from){ |
4724 | // sort into active list |
4725 | append_sorted(active_first_addr(kind), cur); |
4726 | cur->set_state(activeState); |
4727 | if (*prev == cur) { |
4728 | assert(state == activeState, "check" ); |
4729 | prev = cur->next_addr(); |
4730 | } |
4731 | interval_moved(cur, kind, state, activeState); |
4732 | } else { |
4733 | // sort into inactive list |
4734 | append_sorted(inactive_first_addr(kind), cur); |
4735 | cur->set_state(inactiveState); |
4736 | if (*prev == cur) { |
4737 | assert(state == inactiveState, "check" ); |
4738 | prev = cur->next_addr(); |
4739 | } |
4740 | interval_moved(cur, kind, state, inactiveState); |
4741 | } |
4742 | } else { |
4743 | prev = cur->next_addr(); |
4744 | continue; |
4745 | } |
4746 | } |
4747 | } |
4748 | } |
4749 | |
4750 | |
4751 | void IntervalWalker::next_interval() { |
4752 | IntervalKind kind; |
4753 | Interval* any = _unhandled_first[anyKind]; |
4754 | Interval* fixed = _unhandled_first[fixedKind]; |
4755 | |
4756 | if (any != Interval::end()) { |
4757 | // intervals may start at same position -> prefer fixed interval |
4758 | kind = fixed != Interval::end() && fixed->from() <= any->from() ? fixedKind : anyKind; |
4759 | |
4760 | assert (kind == fixedKind && fixed->from() <= any->from() || |
4761 | kind == anyKind && any->from() <= fixed->from(), "wrong interval!!!" ); |
4762 | 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" ); |
4763 | |
4764 | } else if (fixed != Interval::end()) { |
4765 | kind = fixedKind; |
4766 | } else { |
4767 | _current = NULL; return; |
4768 | } |
4769 | _current_kind = kind; |
4770 | _current = _unhandled_first[kind]; |
4771 | _unhandled_first[kind] = _current->next(); |
4772 | _current->set_next(Interval::end()); |
4773 | _current->rewind_range(); |
4774 | } |
4775 | |
4776 | |
4777 | void IntervalWalker::walk_to(int lir_op_id) { |
4778 | assert(_current_position <= lir_op_id, "can not walk backwards" ); |
4779 | while (current() != NULL) { |
4780 | bool is_active = current()->from() <= lir_op_id; |
4781 | int id = is_active ? current()->from() : lir_op_id; |
4782 | |
4783 | TRACE_LINEAR_SCAN(2, if (_current_position < id) { tty->cr(); tty->print_cr("walk_to(%d) **************************************************************" , id); }) |
4784 | |
4785 | // set _current_position prior to call of walk_to |
4786 | _current_position = id; |
4787 | |
4788 | // call walk_to even if _current_position == id |
4789 | walk_to(activeState, id); |
4790 | walk_to(inactiveState, id); |
4791 | |
4792 | if (is_active) { |
4793 | current()->set_state(activeState); |
4794 | if (activate_current()) { |
4795 | append_sorted(active_first_addr(current_kind()), current()); |
4796 | interval_moved(current(), current_kind(), unhandledState, activeState); |
4797 | } |
4798 | |
4799 | next_interval(); |
4800 | } else { |
4801 | return; |
4802 | } |
4803 | } |
4804 | } |
4805 | |
4806 | void IntervalWalker::interval_moved(Interval* interval, IntervalKind kind, IntervalState from, IntervalState to) { |
4807 | #ifndef PRODUCT |
4808 | if (TraceLinearScanLevel >= 4) { |
4809 | #define print_state(state) \ |
4810 | switch(state) {\ |
4811 | case unhandledState: tty->print("unhandled"); break;\ |
4812 | case activeState: tty->print("active"); break;\ |
4813 | case inactiveState: tty->print("inactive"); break;\ |
4814 | case handledState: tty->print("handled"); break;\ |
4815 | default: ShouldNotReachHere(); \ |
4816 | } |
4817 | |
4818 | print_state(from); tty->print(" to " ); print_state(to); |
4819 | tty->fill_to(23); |
4820 | interval->print(); |
4821 | |
4822 | #undef print_state |
4823 | } |
4824 | #endif |
4825 | } |
4826 | |
4827 | |
4828 | |
4829 | // **** Implementation of LinearScanWalker ************************** |
4830 | |
4831 | LinearScanWalker::LinearScanWalker(LinearScan* allocator, Interval* unhandled_fixed_first, Interval* unhandled_any_first) |
4832 | : IntervalWalker(allocator, unhandled_fixed_first, unhandled_any_first) |
4833 | , _move_resolver(allocator) |
4834 | { |
4835 | for (int i = 0; i < LinearScan::nof_regs; i++) { |
4836 | _spill_intervals[i] = new IntervalList(2); |
4837 | } |
4838 | } |
4839 | |
4840 | |
4841 | inline void LinearScanWalker::init_use_lists(bool only_process_use_pos) { |
4842 | for (int i = _first_reg; i <= _last_reg; i++) { |
4843 | _use_pos[i] = max_jint; |
4844 | |
4845 | if (!only_process_use_pos) { |
4846 | _block_pos[i] = max_jint; |
4847 | _spill_intervals[i]->clear(); |
4848 | } |
4849 | } |
4850 | } |
4851 | |
4852 | inline void LinearScanWalker::exclude_from_use(int reg) { |
4853 | assert(reg < LinearScan::nof_regs, "interval must have a register assigned (stack slots not allowed)" ); |
4854 | if (reg >= _first_reg && reg <= _last_reg) { |
4855 | _use_pos[reg] = 0; |
4856 | } |
4857 | } |
4858 | inline void LinearScanWalker::exclude_from_use(Interval* i) { |
4859 | assert(i->assigned_reg() != any_reg, "interval has no register assigned" ); |
4860 | |
4861 | exclude_from_use(i->assigned_reg()); |
4862 | exclude_from_use(i->assigned_regHi()); |
4863 | } |
4864 | |
4865 | inline void LinearScanWalker::set_use_pos(int reg, Interval* i, int use_pos, bool only_process_use_pos) { |
4866 | assert(use_pos != 0, "must use exclude_from_use to set use_pos to 0" ); |
4867 | |
4868 | if (reg >= _first_reg && reg <= _last_reg) { |
4869 | if (_use_pos[reg] > use_pos) { |
4870 | _use_pos[reg] = use_pos; |
4871 | } |
4872 | if (!only_process_use_pos) { |
4873 | _spill_intervals[reg]->append(i); |
4874 | } |
4875 | } |
4876 | } |
4877 | inline void LinearScanWalker::set_use_pos(Interval* i, int use_pos, bool only_process_use_pos) { |
4878 | assert(i->assigned_reg() != any_reg, "interval has no register assigned" ); |
4879 | if (use_pos != -1) { |
4880 | set_use_pos(i->assigned_reg(), i, use_pos, only_process_use_pos); |
4881 | set_use_pos(i->assigned_regHi(), i, use_pos, only_process_use_pos); |
4882 | } |
4883 | } |
4884 | |
4885 | inline void LinearScanWalker::set_block_pos(int reg, Interval* i, int block_pos) { |
4886 | if (reg >= _first_reg && reg <= _last_reg) { |
4887 | if (_block_pos[reg] > block_pos) { |
4888 | _block_pos[reg] = block_pos; |
4889 | } |
4890 | if (_use_pos[reg] > block_pos) { |
4891 | _use_pos[reg] = block_pos; |
4892 | } |
4893 | } |
4894 | } |
4895 | inline void LinearScanWalker::set_block_pos(Interval* i, int block_pos) { |
4896 | assert(i->assigned_reg() != any_reg, "interval has no register assigned" ); |
4897 | if (block_pos != -1) { |
4898 | set_block_pos(i->assigned_reg(), i, block_pos); |
4899 | set_block_pos(i->assigned_regHi(), i, block_pos); |
4900 | } |
4901 | } |
4902 | |
4903 | |
4904 | void LinearScanWalker::free_exclude_active_fixed() { |
4905 | Interval* list = active_first(fixedKind); |
4906 | while (list != Interval::end()) { |
4907 | assert(list->assigned_reg() < LinearScan::nof_regs, "active interval must have a register assigned" ); |
4908 | exclude_from_use(list); |
4909 | list = list->next(); |
4910 | } |
4911 | } |
4912 | |
4913 | void LinearScanWalker::free_exclude_active_any() { |
4914 | Interval* list = active_first(anyKind); |
4915 | while (list != Interval::end()) { |
4916 | exclude_from_use(list); |
4917 | list = list->next(); |
4918 | } |
4919 | } |
4920 | |
4921 | void LinearScanWalker::free_collect_inactive_fixed(Interval* cur) { |
4922 | Interval* list = inactive_first(fixedKind); |
4923 | while (list != Interval::end()) { |
4924 | if (cur->to() <= list->current_from()) { |
4925 | assert(list->current_intersects_at(cur) == -1, "must not intersect" ); |
4926 | set_use_pos(list, list->current_from(), true); |
4927 | } else { |
4928 | set_use_pos(list, list->current_intersects_at(cur), true); |
4929 | } |
4930 | list = list->next(); |
4931 | } |
4932 | } |
4933 | |
4934 | void LinearScanWalker::free_collect_inactive_any(Interval* cur) { |
4935 | Interval* list = inactive_first(anyKind); |
4936 | while (list != Interval::end()) { |
4937 | set_use_pos(list, list->current_intersects_at(cur), true); |
4938 | list = list->next(); |
4939 | } |
4940 | } |
4941 | |
4942 | void LinearScanWalker::spill_exclude_active_fixed() { |
4943 | Interval* list = active_first(fixedKind); |
4944 | while (list != Interval::end()) { |
4945 | exclude_from_use(list); |
4946 | list = list->next(); |
4947 | } |
4948 | } |
4949 | |
4950 | void LinearScanWalker::spill_block_inactive_fixed(Interval* cur) { |
4951 | Interval* list = inactive_first(fixedKind); |
4952 | while (list != Interval::end()) { |
4953 | if (cur->to() > list->current_from()) { |
4954 | set_block_pos(list, list->current_intersects_at(cur)); |
4955 | } else { |
4956 | assert(list->current_intersects_at(cur) == -1, "invalid optimization: intervals intersect" ); |
4957 | } |
4958 | |
4959 | list = list->next(); |
4960 | } |
4961 | } |
4962 | |
4963 | void LinearScanWalker::spill_collect_active_any() { |
4964 | Interval* list = active_first(anyKind); |
4965 | while (list != Interval::end()) { |
4966 | set_use_pos(list, MIN2(list->next_usage(loopEndMarker, _current_position), list->to()), false); |
4967 | list = list->next(); |
4968 | } |
4969 | } |
4970 | |
4971 | void LinearScanWalker::spill_collect_inactive_any(Interval* cur) { |
4972 | Interval* list = inactive_first(anyKind); |
4973 | while (list != Interval::end()) { |
4974 | if (list->current_intersects(cur)) { |
4975 | set_use_pos(list, MIN2(list->next_usage(loopEndMarker, _current_position), list->to()), false); |
4976 | } |
4977 | list = list->next(); |
4978 | } |
4979 | } |
4980 | |
4981 | |
4982 | void LinearScanWalker::insert_move(int op_id, Interval* src_it, Interval* dst_it) { |
4983 | // output all moves here. When source and target are equal, the move is |
4984 | // optimized away later in assign_reg_nums |
4985 | |
4986 | op_id = (op_id + 1) & ~1; |
4987 | BlockBegin* op_block = allocator()->block_of_op_with_id(op_id); |
4988 | assert(op_id > 0 && allocator()->block_of_op_with_id(op_id - 2) == op_block, "cannot insert move at block boundary" ); |
4989 | |
4990 | // calculate index of instruction inside instruction list of current block |
4991 | // the minimal index (for a block with no spill moves) can be calculated because the |
4992 | // numbering of instructions is known. |
4993 | // When the block already contains spill moves, the index must be increased until the |
4994 | // correct index is reached. |
4995 | LIR_OpList* list = op_block->lir()->instructions_list(); |
4996 | int index = (op_id - list->at(0)->id()) / 2; |
4997 | assert(list->at(index)->id() <= op_id, "error in calculation" ); |
4998 | |
4999 | while (list->at(index)->id() != op_id) { |
5000 | index++; |
5001 | assert(0 <= index && index < list->length(), "index out of bounds" ); |
5002 | } |
5003 | assert(1 <= index && index < list->length(), "index out of bounds" ); |
5004 | assert(list->at(index)->id() == op_id, "error in calculation" ); |
5005 | |
5006 | // insert new instruction before instruction at position index |
5007 | _move_resolver.move_insert_position(op_block->lir(), index - 1); |
5008 | _move_resolver.add_mapping(src_it, dst_it); |
5009 | } |
5010 | |
5011 | |
5012 | int LinearScanWalker::find_optimal_split_pos(BlockBegin* min_block, BlockBegin* max_block, int max_split_pos) { |
5013 | int from_block_nr = min_block->linear_scan_number(); |
5014 | int to_block_nr = max_block->linear_scan_number(); |
5015 | |
5016 | assert(0 <= from_block_nr && from_block_nr < block_count(), "out of range" ); |
5017 | assert(0 <= to_block_nr && to_block_nr < block_count(), "out of range" ); |
5018 | assert(from_block_nr < to_block_nr, "must cross block boundary" ); |
5019 | |
5020 | // Try to split at end of max_block. If this would be after |
5021 | // max_split_pos, then use the begin of max_block |
5022 | int optimal_split_pos = max_block->last_lir_instruction_id() + 2; |
5023 | if (optimal_split_pos > max_split_pos) { |
5024 | optimal_split_pos = max_block->first_lir_instruction_id(); |
5025 | } |
5026 | |
5027 | int min_loop_depth = max_block->loop_depth(); |
5028 | for (int i = to_block_nr - 1; i >= from_block_nr; i--) { |
5029 | BlockBegin* cur = block_at(i); |
5030 | |
5031 | if (cur->loop_depth() < min_loop_depth) { |
5032 | // block with lower loop-depth found -> split at the end of this block |
5033 | min_loop_depth = cur->loop_depth(); |
5034 | optimal_split_pos = cur->last_lir_instruction_id() + 2; |
5035 | } |
5036 | } |
5037 | assert(optimal_split_pos > allocator()->max_lir_op_id() || allocator()->is_block_begin(optimal_split_pos), "algorithm must move split pos to block boundary" ); |
5038 | |
5039 | return optimal_split_pos; |
5040 | } |
5041 | |
5042 | |
5043 | int LinearScanWalker::find_optimal_split_pos(Interval* it, int min_split_pos, int max_split_pos, bool do_loop_optimization) { |
5044 | int optimal_split_pos = -1; |
5045 | if (min_split_pos == max_split_pos) { |
5046 | // trivial case, no optimization of split position possible |
5047 | TRACE_LINEAR_SCAN(4, tty->print_cr(" min-pos and max-pos are equal, no optimization possible" )); |
5048 | optimal_split_pos = min_split_pos; |
5049 | |
5050 | } else { |
5051 | assert(min_split_pos < max_split_pos, "must be true then" ); |
5052 | assert(min_split_pos > 0, "cannot access min_split_pos - 1 otherwise" ); |
5053 | |
5054 | // reason for using min_split_pos - 1: when the minimal split pos is exactly at the |
5055 | // beginning of a block, then min_split_pos is also a possible split position. |
5056 | // Use the block before as min_block, because then min_block->last_lir_instruction_id() + 2 == min_split_pos |
5057 | BlockBegin* min_block = allocator()->block_of_op_with_id(min_split_pos - 1); |
5058 | |
5059 | // reason for using max_split_pos - 1: otherwise there would be an assertion failure |
5060 | // when an interval ends at the end of the last block of the method |
5061 | // (in this case, max_split_pos == allocator()->max_lir_op_id() + 2, and there is no |
5062 | // block at this op_id) |
5063 | BlockBegin* max_block = allocator()->block_of_op_with_id(max_split_pos - 1); |
5064 | |
5065 | assert(min_block->linear_scan_number() <= max_block->linear_scan_number(), "invalid order" ); |
5066 | if (min_block == max_block) { |
5067 | // split position cannot be moved to block boundary, so split as late as possible |
5068 | TRACE_LINEAR_SCAN(4, tty->print_cr(" cannot move split pos to block boundary because min_pos and max_pos are in same block" )); |
5069 | optimal_split_pos = max_split_pos; |
5070 | |
5071 | } else if (it->has_hole_between(max_split_pos - 1, max_split_pos) && !allocator()->is_block_begin(max_split_pos)) { |
5072 | // Do not move split position if the interval has a hole before max_split_pos. |
5073 | // Intervals resulting from Phi-Functions have more than one definition (marked |
5074 | // as mustHaveRegister) with a hole before each definition. When the register is needed |
5075 | // for the second definition, an earlier reloading is unnecessary. |
5076 | TRACE_LINEAR_SCAN(4, tty->print_cr(" interval has hole just before max_split_pos, so splitting at max_split_pos" )); |
5077 | optimal_split_pos = max_split_pos; |
5078 | |
5079 | } else { |
5080 | // seach optimal block boundary between min_split_pos and max_split_pos |
5081 | 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())); |
5082 | |
5083 | if (do_loop_optimization) { |
5084 | // Loop optimization: if a loop-end marker is found between min- and max-position, |
5085 | // then split before this loop |
5086 | int loop_end_pos = it->next_usage_exact(loopEndMarker, min_block->last_lir_instruction_id() + 2); |
5087 | TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization: loop end found at pos %d" , loop_end_pos)); |
5088 | |
5089 | assert(loop_end_pos > min_split_pos, "invalid order" ); |
5090 | if (loop_end_pos < max_split_pos) { |
5091 | // loop-end marker found between min- and max-position |
5092 | // if it is not the end marker for the same loop as the min-position, then move |
5093 | // the max-position to this loop block. |
5094 | // Desired result: uses tagged as shouldHaveRegister inside a loop cause a reloading |
5095 | // of the interval (normally, only mustHaveRegister causes a reloading) |
5096 | BlockBegin* loop_block = allocator()->block_of_op_with_id(loop_end_pos); |
5097 | |
5098 | 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())); |
5099 | assert(loop_block != min_block, "loop_block and min_block must be different because block boundary is needed between" ); |
5100 | |
5101 | optimal_split_pos = find_optimal_split_pos(min_block, loop_block, loop_block->last_lir_instruction_id() + 2); |
5102 | if (optimal_split_pos == loop_block->last_lir_instruction_id() + 2) { |
5103 | optimal_split_pos = -1; |
5104 | TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization not necessary" )); |
5105 | } else { |
5106 | TRACE_LINEAR_SCAN(4, tty->print_cr(" loop optimization successful" )); |
5107 | } |
5108 | } |
5109 | } |
5110 | |
5111 | if (optimal_split_pos == -1) { |
5112 | // not calculated by loop optimization |
5113 | optimal_split_pos = find_optimal_split_pos(min_block, max_block, max_split_pos); |
5114 | } |
5115 | } |
5116 | } |
5117 | TRACE_LINEAR_SCAN(4, tty->print_cr(" optimal split position: %d" , optimal_split_pos)); |
5118 | |
5119 | return optimal_split_pos; |
5120 | } |
5121 | |
5122 | |
5123 | /* |
5124 | split an interval at the optimal position between min_split_pos and |
5125 | max_split_pos in two parts: |
5126 | 1) the left part has already a location assigned |
5127 | 2) the right part is sorted into to the unhandled-list |
5128 | */ |
5129 | void LinearScanWalker::split_before_usage(Interval* it, int min_split_pos, int max_split_pos) { |
5130 | TRACE_LINEAR_SCAN(2, tty->print ("----- splitting interval: " ); it->print()); |
5131 | TRACE_LINEAR_SCAN(2, tty->print_cr(" between %d and %d" , min_split_pos, max_split_pos)); |
5132 | |
5133 | assert(it->from() < min_split_pos, "cannot split at start of interval" ); |
5134 | assert(current_position() < min_split_pos, "cannot split before current position" ); |
5135 | assert(min_split_pos <= max_split_pos, "invalid order" ); |
5136 | assert(max_split_pos <= it->to(), "cannot split after end of interval" ); |
5137 | |
5138 | int optimal_split_pos = find_optimal_split_pos(it, min_split_pos, max_split_pos, true); |
5139 | |
5140 | assert(min_split_pos <= optimal_split_pos && optimal_split_pos <= max_split_pos, "out of range" ); |
5141 | assert(optimal_split_pos <= it->to(), "cannot split after end of interval" ); |
5142 | assert(optimal_split_pos > it->from(), "cannot split at start of interval" ); |
5143 | |
5144 | if (optimal_split_pos == it->to() && it->next_usage(mustHaveRegister, min_split_pos) == max_jint) { |
5145 | // the split position would be just before the end of the interval |
5146 | // -> no split at all necessary |
5147 | TRACE_LINEAR_SCAN(4, tty->print_cr(" no split necessary because optimal split position is at end of interval" )); |
5148 | return; |
5149 | } |
5150 | |
5151 | // must calculate this before the actual split is performed and before split position is moved to odd op_id |
5152 | bool move_necessary = !allocator()->is_block_begin(optimal_split_pos) && !it->has_hole_between(optimal_split_pos - 1, optimal_split_pos); |
5153 | |
5154 | if (!allocator()->is_block_begin(optimal_split_pos)) { |
5155 | // move position before actual instruction (odd op_id) |
5156 | optimal_split_pos = (optimal_split_pos - 1) | 1; |
5157 | } |
5158 | |
5159 | TRACE_LINEAR_SCAN(4, tty->print_cr(" splitting at position %d" , optimal_split_pos)); |
5160 | assert(allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 1), "split pos must be odd when not on block boundary" ); |
5161 | assert(!allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 0), "split pos must be even on block boundary" ); |
5162 | |
5163 | Interval* split_part = it->split(optimal_split_pos); |
5164 | |
5165 | allocator()->append_interval(split_part); |
5166 | allocator()->copy_register_flags(it, split_part); |
5167 | split_part->set_insert_move_when_activated(move_necessary); |
5168 | append_to_unhandled(unhandled_first_addr(anyKind), split_part); |
5169 | |
5170 | TRACE_LINEAR_SCAN(2, tty->print_cr(" split interval in two parts (insert_move_when_activated: %d)" , move_necessary)); |
5171 | TRACE_LINEAR_SCAN(2, tty->print (" " ); it->print()); |
5172 | TRACE_LINEAR_SCAN(2, tty->print (" " ); split_part->print()); |
5173 | } |
5174 | |
5175 | /* |
5176 | split an interval at the optimal position between min_split_pos and |
5177 | max_split_pos in two parts: |
5178 | 1) the left part has already a location assigned |
5179 | 2) the right part is always on the stack and therefore ignored in further processing |
5180 | */ |
5181 | void LinearScanWalker::split_for_spilling(Interval* it) { |
5182 | // calculate allowed range of splitting position |
5183 | int max_split_pos = current_position(); |
5184 | int min_split_pos = MAX2(it->previous_usage(shouldHaveRegister, max_split_pos) + 1, it->from()); |
5185 | |
5186 | TRACE_LINEAR_SCAN(2, tty->print ("----- splitting and spilling interval: " ); it->print()); |
5187 | TRACE_LINEAR_SCAN(2, tty->print_cr(" between %d and %d" , min_split_pos, max_split_pos)); |
5188 | |
5189 | assert(it->state() == activeState, "why spill interval that is not active?" ); |
5190 | assert(it->from() <= min_split_pos, "cannot split before start of interval" ); |
5191 | assert(min_split_pos <= max_split_pos, "invalid order" ); |
5192 | assert(max_split_pos < it->to(), "cannot split at end end of interval" ); |
5193 | assert(current_position() < it->to(), "interval must not end before current position" ); |
5194 | |
5195 | if (min_split_pos == it->from()) { |
5196 | // the whole interval is never used, so spill it entirely to memory |
5197 | TRACE_LINEAR_SCAN(2, tty->print_cr(" spilling entire interval because split pos is at beginning of interval" )); |
5198 | assert(it->first_usage(shouldHaveRegister) > current_position(), "interval must not have use position before current_position" ); |
5199 | |
5200 | allocator()->assign_spill_slot(it); |
5201 | allocator()->change_spill_state(it, min_split_pos); |
5202 | |
5203 | // Also kick parent intervals out of register to memory when they have no use |
5204 | // position. This avoids short interval in register surrounded by intervals in |
5205 | // memory -> avoid useless moves from memory to register and back |
5206 | Interval* parent = it; |
5207 | while (parent != NULL && parent->is_split_child()) { |
5208 | parent = parent->split_child_before_op_id(parent->from()); |
5209 | |
5210 | if (parent->assigned_reg() < LinearScan::nof_regs) { |
5211 | if (parent->first_usage(shouldHaveRegister) == max_jint) { |
5212 | // parent is never used, so kick it out of its assigned register |
5213 | TRACE_LINEAR_SCAN(4, tty->print_cr(" kicking out interval %d out of its register because it is never used" , parent->reg_num())); |
5214 | allocator()->assign_spill_slot(parent); |
5215 | } else { |
5216 | // do not go further back because the register is actually used by the interval |
5217 | parent = NULL; |
5218 | } |
5219 | } |
5220 | } |
5221 | |
5222 | } else { |
5223 | // search optimal split pos, split interval and spill only the right hand part |
5224 | int optimal_split_pos = find_optimal_split_pos(it, min_split_pos, max_split_pos, false); |
5225 | |
5226 | assert(min_split_pos <= optimal_split_pos && optimal_split_pos <= max_split_pos, "out of range" ); |
5227 | assert(optimal_split_pos < it->to(), "cannot split at end of interval" ); |
5228 | assert(optimal_split_pos >= it->from(), "cannot split before start of interval" ); |
5229 | |
5230 | if (!allocator()->is_block_begin(optimal_split_pos)) { |
5231 | // move position before actual instruction (odd op_id) |
5232 | optimal_split_pos = (optimal_split_pos - 1) | 1; |
5233 | } |
5234 | |
5235 | TRACE_LINEAR_SCAN(4, tty->print_cr(" splitting at position %d" , optimal_split_pos)); |
5236 | assert(allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 1), "split pos must be odd when not on block boundary" ); |
5237 | assert(!allocator()->is_block_begin(optimal_split_pos) || (optimal_split_pos % 2 == 0), "split pos must be even on block boundary" ); |
5238 | |
5239 | Interval* spilled_part = it->split(optimal_split_pos); |
5240 | allocator()->append_interval(spilled_part); |
5241 | allocator()->assign_spill_slot(spilled_part); |
5242 | allocator()->change_spill_state(spilled_part, optimal_split_pos); |
5243 | |
5244 | if (!allocator()->is_block_begin(optimal_split_pos)) { |
5245 | TRACE_LINEAR_SCAN(4, tty->print_cr(" inserting move from interval %d to %d" , it->reg_num(), spilled_part->reg_num())); |
5246 | insert_move(optimal_split_pos, it, spilled_part); |
5247 | } |
5248 | |
5249 | // the current_split_child is needed later when moves are inserted for reloading |
5250 | assert(spilled_part->current_split_child() == it, "overwriting wrong current_split_child" ); |
5251 | spilled_part->make_current_split_child(); |
5252 | |
5253 | TRACE_LINEAR_SCAN(2, tty->print_cr(" split interval in two parts" )); |
5254 | TRACE_LINEAR_SCAN(2, tty->print (" " ); it->print()); |
5255 | TRACE_LINEAR_SCAN(2, tty->print (" " ); spilled_part->print()); |
5256 | } |
5257 | } |
5258 | |
5259 | |
5260 | void LinearScanWalker::split_stack_interval(Interval* it) { |
5261 | int min_split_pos = current_position() + 1; |
5262 | int max_split_pos = MIN2(it->first_usage(shouldHaveRegister), it->to()); |
5263 | |
5264 | split_before_usage(it, min_split_pos, max_split_pos); |
5265 | } |
5266 | |
5267 | void LinearScanWalker::split_when_partial_register_available(Interval* it, int register_available_until) { |
5268 | int min_split_pos = MAX2(it->previous_usage(shouldHaveRegister, register_available_until), it->from() + 1); |
5269 | int max_split_pos = register_available_until; |
5270 | |
5271 | split_before_usage(it, min_split_pos, max_split_pos); |
5272 | } |
5273 | |
5274 | void LinearScanWalker::split_and_spill_interval(Interval* it) { |
5275 | assert(it->state() == activeState || it->state() == inactiveState, "other states not allowed" ); |
5276 | |
5277 | int current_pos = current_position(); |
5278 | if (it->state() == inactiveState) { |
5279 | // the interval is currently inactive, so no spill slot is needed for now. |
5280 | // when the split part is activated, the interval has a new chance to get a register, |
5281 | // so in the best case no stack slot is necessary |
5282 | assert(it->has_hole_between(current_pos - 1, current_pos + 1), "interval can not be inactive otherwise" ); |
5283 | split_before_usage(it, current_pos + 1, current_pos + 1); |
5284 | |
5285 | } else { |
5286 | // search the position where the interval must have a register and split |
5287 | // at the optimal position before. |
5288 | // The new created part is added to the unhandled list and will get a register |
5289 | // when it is activated |
5290 | int min_split_pos = current_pos + 1; |
5291 | int max_split_pos = MIN2(it->next_usage(mustHaveRegister, min_split_pos), it->to()); |
5292 | |
5293 | split_before_usage(it, min_split_pos, max_split_pos); |
5294 | |
5295 | assert(it->next_usage(mustHaveRegister, current_pos) == max_jint, "the remaining part is spilled to stack and therefore has no register" ); |
5296 | split_for_spilling(it); |
5297 | } |
5298 | } |
5299 | |
5300 | |
5301 | int LinearScanWalker::find_free_reg(int reg_needed_until, int interval_to, int hint_reg, int ignore_reg, bool* need_split) { |
5302 | int min_full_reg = any_reg; |
5303 | int max_partial_reg = any_reg; |
5304 | |
5305 | for (int i = _first_reg; i <= _last_reg; i++) { |
5306 | if (i == ignore_reg) { |
5307 | // this register must be ignored |
5308 | |
5309 | } else if (_use_pos[i] >= interval_to) { |
5310 | // this register is free for the full interval |
5311 | if (min_full_reg == any_reg || i == hint_reg || (_use_pos[i] < _use_pos[min_full_reg] && min_full_reg != hint_reg)) { |
5312 | min_full_reg = i; |
5313 | } |
5314 | } else if (_use_pos[i] > reg_needed_until) { |
5315 | // this register is at least free until reg_needed_until |
5316 | if (max_partial_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_partial_reg] && max_partial_reg != hint_reg)) { |
5317 | max_partial_reg = i; |
5318 | } |
5319 | } |
5320 | } |
5321 | |
5322 | if (min_full_reg != any_reg) { |
5323 | return min_full_reg; |
5324 | } else if (max_partial_reg != any_reg) { |
5325 | *need_split = true; |
5326 | return max_partial_reg; |
5327 | } else { |
5328 | return any_reg; |
5329 | } |
5330 | } |
5331 | |
5332 | int LinearScanWalker::find_free_double_reg(int reg_needed_until, int interval_to, int hint_reg, bool* need_split) { |
5333 | assert((_last_reg - _first_reg + 1) % 2 == 0, "adjust algorithm" ); |
5334 | |
5335 | int min_full_reg = any_reg; |
5336 | int max_partial_reg = any_reg; |
5337 | |
5338 | for (int i = _first_reg; i < _last_reg; i+=2) { |
5339 | if (_use_pos[i] >= interval_to && _use_pos[i + 1] >= interval_to) { |
5340 | // this register is free for the full interval |
5341 | if (min_full_reg == any_reg || i == hint_reg || (_use_pos[i] < _use_pos[min_full_reg] && min_full_reg != hint_reg)) { |
5342 | min_full_reg = i; |
5343 | } |
5344 | } else if (_use_pos[i] > reg_needed_until && _use_pos[i + 1] > reg_needed_until) { |
5345 | // this register is at least free until reg_needed_until |
5346 | if (max_partial_reg == any_reg || i == hint_reg || (_use_pos[i] > _use_pos[max_partial_reg] && max_partial_reg != hint_reg)) { |
5347 | max_partial_reg = i; |
5348 | } |
5349 | } |
5350 | } |
5351 | |
5352 | if (min_full_reg != any_reg) { |
5353 | return min_full_reg; |
5354 | } else if (max_partial_reg != any_reg) { |
5355 | *need_split = true; |
5356 | return max_partial_reg; |
5357 | } else { |
5358 | return any_reg; |
5359 | } |
5360 | } |
5361 | |
5362 | |
5363 | bool LinearScanWalker::alloc_free_reg(Interval* cur) { |
5364 | TRACE_LINEAR_SCAN(2, tty->print("trying to find free register for " ); cur->print()); |
5365 | |
5366 | init_use_lists(true); |
5367 | free_exclude_active_fixed(); |
5368 | free_exclude_active_any(); |
5369 | free_collect_inactive_fixed(cur); |
5370 | free_collect_inactive_any(cur); |
5371 | assert(unhandled_first(fixedKind) == Interval::end(), "must not have unhandled fixed intervals because all fixed intervals have a use at position 0" ); |
5372 | |
5373 | // _use_pos contains the start of the next interval that has this register assigned |
5374 | // (either as a fixed register or a normal allocated register in the past) |
5375 | // only intervals overlapping with cur are processed, non-overlapping invervals can be ignored safely |
5376 | TRACE_LINEAR_SCAN(4, tty->print_cr(" state of registers:" )); |
5377 | TRACE_LINEAR_SCAN(4, for (int i = _first_reg; i <= _last_reg; i++) tty->print_cr(" reg %d: use_pos: %d" , i, _use_pos[i])); |
5378 | |
5379 | int hint_reg, hint_regHi; |
5380 | Interval* register_hint = cur->register_hint(); |
5381 | if (register_hint != NULL) { |
5382 | hint_reg = register_hint->assigned_reg(); |
5383 | hint_regHi = register_hint->assigned_regHi(); |
5384 | |
5385 | if (allocator()->is_precolored_cpu_interval(register_hint)) { |
5386 | assert(hint_reg != any_reg && hint_regHi == any_reg, "must be for fixed intervals" ); |
5387 | hint_regHi = hint_reg + 1; // connect e.g. eax-edx |
5388 | } |
5389 | TRACE_LINEAR_SCAN(4, tty->print(" hint registers %d, %d from interval " , hint_reg, hint_regHi); register_hint->print()); |
5390 | |
5391 | } else { |
5392 | hint_reg = any_reg; |
5393 | hint_regHi = any_reg; |
5394 | } |
5395 | assert(hint_reg == any_reg || hint_reg != hint_regHi, "hint reg and regHi equal" ); |
5396 | assert(cur->assigned_reg() == any_reg && cur->assigned_regHi() == any_reg, "register already assigned to interval" ); |
5397 | |
5398 | // the register must be free at least until this position |
5399 | int reg_needed_until = cur->from() + 1; |
5400 | int interval_to = cur->to(); |
5401 | |
5402 | bool need_split = false; |
5403 | int split_pos; |
5404 | int reg; |
5405 | int regHi = any_reg; |
5406 | |
5407 | if (_adjacent_regs) { |
5408 | reg = find_free_double_reg(reg_needed_until, interval_to, hint_reg, &need_split); |
5409 | regHi = reg + 1; |
5410 | if (reg == any_reg) { |
5411 | return false; |
5412 | } |
5413 | split_pos = MIN2(_use_pos[reg], _use_pos[regHi]); |
5414 | |
5415 | } else { |
5416 | reg = find_free_reg(reg_needed_until, interval_to, hint_reg, any_reg, &need_split); |
5417 | if (reg == any_reg) { |
5418 | return false; |
5419 | } |
5420 | split_pos = _use_pos[reg]; |
5421 | |
5422 | if (_num_phys_regs == 2) { |
5423 | regHi = find_free_reg(reg_needed_until, interval_to, hint_regHi, reg, &need_split); |
5424 | |
5425 | if (_use_pos[reg] < interval_to && regHi == any_reg) { |
5426 | // do not split interval if only one register can be assigned until the split pos |
5427 | // (when one register is found for the whole interval, split&spill is only |
5428 | // performed for the hi register) |
5429 | return false; |
5430 | |
5431 | } else if (regHi != any_reg) { |
5432 | split_pos = MIN2(split_pos, _use_pos[regHi]); |
5433 | |
5434 | // sort register numbers to prevent e.g. a move from eax,ebx to ebx,eax |
5435 | if (reg > regHi) { |
5436 | int temp = reg; |
5437 | reg = regHi; |
5438 | regHi = temp; |
5439 | } |
5440 | } |
5441 | } |
5442 | } |
5443 | |
5444 | cur->assign_reg(reg, regHi); |
5445 | TRACE_LINEAR_SCAN(2, tty->print_cr("selected register %d, %d" , reg, regHi)); |
5446 | |
5447 | assert(split_pos > 0, "invalid split_pos" ); |
5448 | if (need_split) { |
5449 | // register not available for full interval, so split it |
5450 | split_when_partial_register_available(cur, split_pos); |
5451 | } |
5452 | |
5453 | // only return true if interval is completely assigned |
5454 | return _num_phys_regs == 1 || regHi != any_reg; |
5455 | } |
5456 | |
5457 | |
5458 | int LinearScanWalker::find_locked_reg(int reg_needed_until, int interval_to, int ignore_reg, bool* need_split) { |
5459 | int max_reg = any_reg; |
5460 | |
5461 | for (int i = _first_reg; i <= _last_reg; i++) { |
5462 | if (i == ignore_reg) { |
5463 | // this register must be ignored |
5464 | |
5465 | } else if (_use_pos[i] > reg_needed_until) { |
5466 | if (max_reg == any_reg || _use_pos[i] > _use_pos[max_reg]) { |
5467 | max_reg = i; |
5468 | } |
5469 | } |
5470 | } |
5471 | |
5472 | if (max_reg != any_reg && _block_pos[max_reg] <= interval_to) { |
5473 | *need_split = true; |
5474 | } |
5475 | |
5476 | return max_reg; |
5477 | } |
5478 | |
5479 | int LinearScanWalker::find_locked_double_reg(int reg_needed_until, int interval_to, bool* need_split) { |
5480 | assert((_last_reg - _first_reg + 1) % 2 == 0, "adjust algorithm" ); |
5481 | |
5482 | int max_reg = any_reg; |
5483 | |
5484 | for (int i = _first_reg; i < _last_reg; i+=2) { |
5485 | if (_use_pos[i] > reg_needed_until && _use_pos[i + 1] > reg_needed_until) { |
5486 | if (max_reg == any_reg || _use_pos[i] > _use_pos[max_reg]) { |
5487 | max_reg = i; |
5488 | } |
5489 | } |
5490 | } |
5491 | |
5492 | if (max_reg != any_reg && |
5493 | (_block_pos[max_reg] <= interval_to || _block_pos[max_reg + 1] <= interval_to)) { |
5494 | *need_split = true; |
5495 | } |
5496 | |
5497 | return max_reg; |
5498 | } |
5499 | |
5500 | void LinearScanWalker::split_and_spill_intersecting_intervals(int reg, int regHi) { |
5501 | assert(reg != any_reg, "no register assigned" ); |
5502 | |
5503 | for (int i = 0; i < _spill_intervals[reg]->length(); i++) { |
5504 | Interval* it = _spill_intervals[reg]->at(i); |
5505 | remove_from_list(it); |
5506 | split_and_spill_interval(it); |
5507 | } |
5508 | |
5509 | if (regHi != any_reg) { |
5510 | IntervalList* processed = _spill_intervals[reg]; |
5511 | for (int i = 0; i < _spill_intervals[regHi]->length(); i++) { |
5512 | Interval* it = _spill_intervals[regHi]->at(i); |
5513 | if (processed->find(it) == -1) { |
5514 | remove_from_list(it); |
5515 | split_and_spill_interval(it); |
5516 | } |
5517 | } |
5518 | } |
5519 | } |
5520 | |
5521 | |
5522 | // Split an Interval and spill it to memory so that cur can be placed in a register |
5523 | void LinearScanWalker::alloc_locked_reg(Interval* cur) { |
5524 | TRACE_LINEAR_SCAN(2, tty->print("need to split and spill to get register for " ); cur->print()); |
5525 | |
5526 | // collect current usage of registers |
5527 | init_use_lists(false); |
5528 | spill_exclude_active_fixed(); |
5529 | assert(unhandled_first(fixedKind) == Interval::end(), "must not have unhandled fixed intervals because all fixed intervals have a use at position 0" ); |
5530 | spill_block_inactive_fixed(cur); |
5531 | spill_collect_active_any(); |
5532 | spill_collect_inactive_any(cur); |
5533 | |
5534 | #ifndef PRODUCT |
5535 | if (TraceLinearScanLevel >= 4) { |
5536 | tty->print_cr(" state of registers:" ); |
5537 | for (int i = _first_reg; i <= _last_reg; i++) { |
5538 | tty->print(" reg %d: use_pos: %d, block_pos: %d, intervals: " , i, _use_pos[i], _block_pos[i]); |
5539 | for (int j = 0; j < _spill_intervals[i]->length(); j++) { |
5540 | tty->print("%d " , _spill_intervals[i]->at(j)->reg_num()); |
5541 | } |
5542 | tty->cr(); |
5543 | } |
5544 | } |
5545 | #endif |
5546 | |
5547 | // the register must be free at least until this position |
5548 | int reg_needed_until = MIN2(cur->first_usage(mustHaveRegister), cur->from() + 1); |
5549 | int interval_to = cur->to(); |
5550 | assert (reg_needed_until > 0 && reg_needed_until < max_jint, "interval has no use" ); |
5551 | |
5552 | int split_pos = 0; |
5553 | int use_pos = 0; |
5554 | bool need_split = false; |
5555 | int reg, regHi; |
5556 | |
5557 | if (_adjacent_regs) { |
5558 | reg = find_locked_double_reg(reg_needed_until, interval_to, &need_split); |
5559 | regHi = reg + 1; |
5560 | |
5561 | if (reg != any_reg) { |
5562 | use_pos = MIN2(_use_pos[reg], _use_pos[regHi]); |
5563 | split_pos = MIN2(_block_pos[reg], _block_pos[regHi]); |
5564 | } |
5565 | } else { |
5566 | reg = find_locked_reg(reg_needed_until, interval_to, cur->assigned_reg(), &need_split); |
5567 | regHi = any_reg; |
5568 | |
5569 | if (reg != any_reg) { |
5570 | use_pos = _use_pos[reg]; |
5571 | split_pos = _block_pos[reg]; |
5572 | |
5573 | if (_num_phys_regs == 2) { |
5574 | if (cur->assigned_reg() != any_reg) { |
5575 | regHi = reg; |
5576 | reg = cur->assigned_reg(); |
5577 | } else { |
5578 | regHi = find_locked_reg(reg_needed_until, interval_to, reg, &need_split); |
5579 | if (regHi != any_reg) { |
5580 | use_pos = MIN2(use_pos, _use_pos[regHi]); |
5581 | split_pos = MIN2(split_pos, _block_pos[regHi]); |
5582 | } |
5583 | } |
5584 | |
5585 | if (regHi != any_reg && reg > regHi) { |
5586 | // sort register numbers to prevent e.g. a move from eax,ebx to ebx,eax |
5587 | int temp = reg; |
5588 | reg = regHi; |
5589 | regHi = temp; |
5590 | } |
5591 | } |
5592 | } |
5593 | } |
5594 | |
5595 | if (reg == any_reg || (_num_phys_regs == 2 && regHi == any_reg) || use_pos <= cur->first_usage(mustHaveRegister)) { |
5596 | // the first use of cur is later than the spilling position -> spill cur |
5597 | 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)); |
5598 | |
5599 | if (cur->first_usage(mustHaveRegister) <= cur->from() + 1) { |
5600 | assert(false, "cannot spill interval that is used in first instruction (possible reason: no register found)" ); |
5601 | // assign a reasonable register and do a bailout in product mode to avoid errors |
5602 | allocator()->assign_spill_slot(cur); |
5603 | BAILOUT("LinearScan: no register found" ); |
5604 | } |
5605 | |
5606 | split_and_spill_interval(cur); |
5607 | } else { |
5608 | TRACE_LINEAR_SCAN(4, tty->print_cr("decided to use register %d, %d" , reg, regHi)); |
5609 | assert(reg != any_reg && (_num_phys_regs == 1 || regHi != any_reg), "no register found" ); |
5610 | assert(split_pos > 0, "invalid split_pos" ); |
5611 | assert(need_split == false || split_pos > cur->from(), "splitting interval at from" ); |
5612 | |
5613 | cur->assign_reg(reg, regHi); |
5614 | if (need_split) { |
5615 | // register not available for full interval, so split it |
5616 | split_when_partial_register_available(cur, split_pos); |
5617 | } |
5618 | |
5619 | // perform splitting and spilling for all affected intervalls |
5620 | split_and_spill_intersecting_intervals(reg, regHi); |
5621 | } |
5622 | } |
5623 | |
5624 | bool LinearScanWalker::no_allocation_possible(Interval* cur) { |
5625 | #ifdef X86 |
5626 | // fast calculation of intervals that can never get a register because the |
5627 | // the next instruction is a call that blocks all registers |
5628 | // Note: this does not work if callee-saved registers are available (e.g. on Sparc) |
5629 | |
5630 | // check if this interval is the result of a split operation |
5631 | // (an interval got a register until this position) |
5632 | int pos = cur->from(); |
5633 | if ((pos & 1) == 1) { |
5634 | // the current instruction is a call that blocks all registers |
5635 | if (pos < allocator()->max_lir_op_id() && allocator()->has_call(pos + 1)) { |
5636 | TRACE_LINEAR_SCAN(4, tty->print_cr(" free register cannot be available because all registers blocked by following call" )); |
5637 | |
5638 | // safety check that there is really no register available |
5639 | assert(alloc_free_reg(cur) == false, "found a register for this interval" ); |
5640 | return true; |
5641 | } |
5642 | |
5643 | } |
5644 | #endif |
5645 | return false; |
5646 | } |
5647 | |
5648 | void LinearScanWalker::init_vars_for_alloc(Interval* cur) { |
5649 | BasicType type = cur->type(); |
5650 | _num_phys_regs = LinearScan::num_physical_regs(type); |
5651 | _adjacent_regs = LinearScan::requires_adjacent_regs(type); |
5652 | |
5653 | if (pd_init_regs_for_alloc(cur)) { |
5654 | // the appropriate register range was selected. |
5655 | } else if (type == T_FLOAT || type == T_DOUBLE) { |
5656 | _first_reg = pd_first_fpu_reg; |
5657 | _last_reg = pd_last_fpu_reg; |
5658 | } else { |
5659 | _first_reg = pd_first_cpu_reg; |
5660 | _last_reg = FrameMap::last_cpu_reg(); |
5661 | } |
5662 | |
5663 | assert(0 <= _first_reg && _first_reg < LinearScan::nof_regs, "out of range" ); |
5664 | assert(0 <= _last_reg && _last_reg < LinearScan::nof_regs, "out of range" ); |
5665 | } |
5666 | |
5667 | |
5668 | bool LinearScanWalker::is_move(LIR_Op* op, Interval* from, Interval* to) { |
5669 | if (op->code() != lir_move) { |
5670 | return false; |
5671 | } |
5672 | assert(op->as_Op1() != NULL, "move must be LIR_Op1" ); |
5673 | |
5674 | LIR_Opr in = ((LIR_Op1*)op)->in_opr(); |
5675 | LIR_Opr res = ((LIR_Op1*)op)->result_opr(); |
5676 | return in->is_virtual() && res->is_virtual() && in->vreg_number() == from->reg_num() && res->vreg_number() == to->reg_num(); |
5677 | } |
5678 | |
5679 | // optimization (especially for phi functions of nested loops): |
5680 | // assign same spill slot to non-intersecting intervals |
5681 | void LinearScanWalker::combine_spilled_intervals(Interval* cur) { |
5682 | if (cur->is_split_child()) { |
5683 | // optimization is only suitable for split parents |
5684 | return; |
5685 | } |
5686 | |
5687 | Interval* register_hint = cur->register_hint(false); |
5688 | if (register_hint == NULL) { |
5689 | // cur is not the target of a move, otherwise register_hint would be set |
5690 | return; |
5691 | } |
5692 | assert(register_hint->is_split_parent(), "register hint must be split parent" ); |
5693 | |
5694 | if (cur->spill_state() != noOptimization || register_hint->spill_state() != noOptimization) { |
5695 | // combining the stack slots for intervals where spill move optimization is applied |
5696 | // is not benefitial and would cause problems |
5697 | return; |
5698 | } |
5699 | |
5700 | int begin_pos = cur->from(); |
5701 | int end_pos = cur->to(); |
5702 | if (end_pos > allocator()->max_lir_op_id() || (begin_pos & 1) != 0 || (end_pos & 1) != 0) { |
5703 | // safety check that lir_op_with_id is allowed |
5704 | return; |
5705 | } |
5706 | |
5707 | 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)) { |
5708 | // cur and register_hint are not connected with two moves |
5709 | return; |
5710 | } |
5711 | |
5712 | Interval* begin_hint = register_hint->split_child_at_op_id(begin_pos, LIR_OpVisitState::inputMode); |
5713 | Interval* end_hint = register_hint->split_child_at_op_id(end_pos, LIR_OpVisitState::outputMode); |
5714 | if (begin_hint == end_hint || begin_hint->to() != begin_pos || end_hint->from() != end_pos) { |
5715 | // register_hint must be split, otherwise the re-writing of use positions does not work |
5716 | return; |
5717 | } |
5718 | |
5719 | assert(begin_hint->assigned_reg() != any_reg, "must have register assigned" ); |
5720 | assert(end_hint->assigned_reg() == any_reg, "must not have register assigned" ); |
5721 | assert(cur->first_usage(mustHaveRegister) == begin_pos, "must have use position at begin of interval because of move" ); |
5722 | assert(end_hint->first_usage(mustHaveRegister) == end_pos, "must have use position at begin of interval because of move" ); |
5723 | |
5724 | if (begin_hint->assigned_reg() < LinearScan::nof_regs) { |
5725 | // register_hint is not spilled at begin_pos, so it would not be benefitial to immediately spill cur |
5726 | return; |
5727 | } |
5728 | assert(register_hint->canonical_spill_slot() != -1, "must be set when part of interval was spilled" ); |
5729 | |
5730 | // modify intervals such that cur gets the same stack slot as register_hint |
5731 | // delete use positions to prevent the intervals to get a register at beginning |
5732 | cur->set_canonical_spill_slot(register_hint->canonical_spill_slot()); |
5733 | cur->remove_first_use_pos(); |
5734 | end_hint->remove_first_use_pos(); |
5735 | } |
5736 | |
5737 | |
5738 | // allocate a physical register or memory location to an interval |
5739 | bool LinearScanWalker::activate_current() { |
5740 | Interval* cur = current(); |
5741 | bool result = true; |
5742 | |
5743 | TRACE_LINEAR_SCAN(2, tty->print ("+++++ activating interval " ); cur->print()); |
5744 | 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())); |
5745 | |
5746 | if (cur->assigned_reg() >= LinearScan::nof_regs) { |
5747 | // activating an interval that has a stack slot assigned -> split it at first use position |
5748 | // used for method parameters |
5749 | TRACE_LINEAR_SCAN(4, tty->print_cr(" interval has spill slot assigned (method parameter) -> split it before first use" )); |
5750 | |
5751 | split_stack_interval(cur); |
5752 | result = false; |
5753 | |
5754 | } else if (allocator()->gen()->is_vreg_flag_set(cur->reg_num(), LIRGenerator::must_start_in_memory)) { |
5755 | // activating an interval that must start in a stack slot, but may get a register later |
5756 | // used for lir_roundfp: rounding is done by store to stack and reload later |
5757 | TRACE_LINEAR_SCAN(4, tty->print_cr(" interval must start in stack slot -> split it before first use" )); |
5758 | assert(cur->assigned_reg() == any_reg && cur->assigned_regHi() == any_reg, "register already assigned" ); |
5759 | |
5760 | allocator()->assign_spill_slot(cur); |
5761 | split_stack_interval(cur); |
5762 | result = false; |
5763 | |
5764 | } else if (cur->assigned_reg() == any_reg) { |
5765 | // interval has not assigned register -> normal allocation |
5766 | // (this is the normal case for most intervals) |
5767 | TRACE_LINEAR_SCAN(4, tty->print_cr(" normal allocation of register" )); |
5768 | |
5769 | // assign same spill slot to non-intersecting intervals |
5770 | combine_spilled_intervals(cur); |
5771 | |
5772 | init_vars_for_alloc(cur); |
5773 | if (no_allocation_possible(cur) || !alloc_free_reg(cur)) { |
5774 | // no empty register available. |
5775 | // split and spill another interval so that this interval gets a register |
5776 | alloc_locked_reg(cur); |
5777 | } |
5778 | |
5779 | // spilled intervals need not be move to active-list |
5780 | if (cur->assigned_reg() >= LinearScan::nof_regs) { |
5781 | result = false; |
5782 | } |
5783 | } |
5784 | |
5785 | // load spilled values that become active from stack slot to register |
5786 | if (cur->insert_move_when_activated()) { |
5787 | assert(cur->is_split_child(), "must be" ); |
5788 | assert(cur->current_split_child() != NULL, "must be" ); |
5789 | assert(cur->current_split_child()->reg_num() != cur->reg_num(), "cannot insert move between same interval" ); |
5790 | 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())); |
5791 | |
5792 | insert_move(cur->from(), cur->current_split_child(), cur); |
5793 | } |
5794 | cur->make_current_split_child(); |
5795 | |
5796 | return result; // true = interval is moved to active list |
5797 | } |
5798 | |
5799 | |
5800 | // Implementation of EdgeMoveOptimizer |
5801 | |
5802 | EdgeMoveOptimizer::EdgeMoveOptimizer() : |
5803 | _edge_instructions(4), |
5804 | _edge_instructions_idx(4) |
5805 | { |
5806 | } |
5807 | |
5808 | void EdgeMoveOptimizer::optimize(BlockList* code) { |
5809 | EdgeMoveOptimizer optimizer = EdgeMoveOptimizer(); |
5810 | |
5811 | // ignore the first block in the list (index 0 is not processed) |
5812 | for (int i = code->length() - 1; i >= 1; i--) { |
5813 | BlockBegin* block = code->at(i); |
5814 | |
5815 | if (block->number_of_preds() > 1 && !block->is_set(BlockBegin::exception_entry_flag)) { |
5816 | optimizer.optimize_moves_at_block_end(block); |
5817 | } |
5818 | if (block->number_of_sux() == 2) { |
5819 | optimizer.optimize_moves_at_block_begin(block); |
5820 | } |
5821 | } |
5822 | } |
5823 | |
5824 | |
5825 | // clear all internal data structures |
5826 | void EdgeMoveOptimizer::init_instructions() { |
5827 | _edge_instructions.clear(); |
5828 | _edge_instructions_idx.clear(); |
5829 | } |
5830 | |
5831 | // append a lir-instruction-list and the index of the current operation in to the list |
5832 | void EdgeMoveOptimizer::append_instructions(LIR_OpList* instructions, int instructions_idx) { |
5833 | _edge_instructions.append(instructions); |
5834 | _edge_instructions_idx.append(instructions_idx); |
5835 | } |
5836 | |
5837 | // return the current operation of the given edge (predecessor or successor) |
5838 | LIR_Op* EdgeMoveOptimizer::instruction_at(int edge) { |
5839 | LIR_OpList* instructions = _edge_instructions.at(edge); |
5840 | int idx = _edge_instructions_idx.at(edge); |
5841 | |
5842 | if (idx < instructions->length()) { |
5843 | return instructions->at(idx); |
5844 | } else { |
5845 | return NULL; |
5846 | } |
5847 | } |
5848 | |
5849 | // removes the current operation of the given edge (predecessor or successor) |
5850 | void EdgeMoveOptimizer::remove_cur_instruction(int edge, bool decrement_index) { |
5851 | LIR_OpList* instructions = _edge_instructions.at(edge); |
5852 | int idx = _edge_instructions_idx.at(edge); |
5853 | instructions->remove_at(idx); |
5854 | |
5855 | if (decrement_index) { |
5856 | _edge_instructions_idx.at_put(edge, idx - 1); |
5857 | } |
5858 | } |
5859 | |
5860 | |
5861 | bool EdgeMoveOptimizer::operations_different(LIR_Op* op1, LIR_Op* op2) { |
5862 | if (op1 == NULL || op2 == NULL) { |
5863 | // at least one block is already empty -> no optimization possible |
5864 | return true; |
5865 | } |
5866 | |
5867 | if (op1->code() == lir_move && op2->code() == lir_move) { |
5868 | assert(op1->as_Op1() != NULL, "move must be LIR_Op1" ); |
5869 | assert(op2->as_Op1() != NULL, "move must be LIR_Op1" ); |
5870 | LIR_Op1* move1 = (LIR_Op1*)op1; |
5871 | LIR_Op1* move2 = (LIR_Op1*)op2; |
5872 | if (move1->info() == move2->info() && move1->in_opr() == move2->in_opr() && move1->result_opr() == move2->result_opr()) { |
5873 | // these moves are exactly equal and can be optimized |
5874 | return false; |
5875 | } |
5876 | |
5877 | } else if (op1->code() == lir_fxch && op2->code() == lir_fxch) { |
5878 | assert(op1->as_Op1() != NULL, "fxch must be LIR_Op1" ); |
5879 | assert(op2->as_Op1() != NULL, "fxch must be LIR_Op1" ); |
5880 | LIR_Op1* fxch1 = (LIR_Op1*)op1; |
5881 | LIR_Op1* fxch2 = (LIR_Op1*)op2; |
5882 | if (fxch1->in_opr()->as_jint() == fxch2->in_opr()->as_jint()) { |
5883 | // equal FPU stack operations can be optimized |
5884 | return false; |
5885 | } |
5886 | |
5887 | } else if (op1->code() == lir_fpop_raw && op2->code() == lir_fpop_raw) { |
5888 | // equal FPU stack operations can be optimized |
5889 | return false; |
5890 | } |
5891 | |
5892 | // no optimization possible |
5893 | return true; |
5894 | } |
5895 | |
5896 | void EdgeMoveOptimizer::optimize_moves_at_block_end(BlockBegin* block) { |
5897 | TRACE_LINEAR_SCAN(4, tty->print_cr("optimizing moves at end of block B%d" , block->block_id())); |
5898 | |
5899 | if (block->is_predecessor(block)) { |
5900 | // currently we can't handle this correctly. |
5901 | return; |
5902 | } |
5903 | |
5904 | init_instructions(); |
5905 | int num_preds = block->number_of_preds(); |
5906 | assert(num_preds > 1, "do not call otherwise" ); |
5907 | assert(!block->is_set(BlockBegin::exception_entry_flag), "exception handlers not allowed" ); |
5908 | |
5909 | // setup a list with the lir-instructions of all predecessors |
5910 | int i; |
5911 | for (i = 0; i < num_preds; i++) { |
5912 | BlockBegin* pred = block->pred_at(i); |
5913 | LIR_OpList* pred_instructions = pred->lir()->instructions_list(); |
5914 | |
5915 | if (pred->number_of_sux() != 1) { |
5916 | // this can happen with switch-statements where multiple edges are between |
5917 | // the same blocks. |
5918 | return; |
5919 | } |
5920 | |
5921 | assert(pred->number_of_sux() == 1, "can handle only one successor" ); |
5922 | assert(pred->sux_at(0) == block, "invalid control flow" ); |
5923 | assert(pred_instructions->last()->code() == lir_branch, "block with successor must end with branch" ); |
5924 | assert(pred_instructions->last()->as_OpBranch() != NULL, "branch must be LIR_OpBranch" ); |
5925 | assert(pred_instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block must end with unconditional branch" ); |
5926 | |
5927 | if (pred_instructions->last()->info() != NULL) { |
5928 | // can not optimize instructions when debug info is needed |
5929 | return; |
5930 | } |
5931 | |
5932 | // ignore the unconditional branch at the end of the block |
5933 | append_instructions(pred_instructions, pred_instructions->length() - 2); |
5934 | } |
5935 | |
5936 | |
5937 | // process lir-instructions while all predecessors end with the same instruction |
5938 | while (true) { |
5939 | LIR_Op* op = instruction_at(0); |
5940 | for (i = 1; i < num_preds; i++) { |
5941 | if (operations_different(op, instruction_at(i))) { |
5942 | // these instructions are different and cannot be optimized -> |
5943 | // no further optimization possible |
5944 | return; |
5945 | } |
5946 | } |
5947 | |
5948 | TRACE_LINEAR_SCAN(4, tty->print("found instruction that is equal in all %d predecessors: " , num_preds); op->print()); |
5949 | |
5950 | // insert the instruction at the beginning of the current block |
5951 | block->lir()->insert_before(1, op); |
5952 | |
5953 | // delete the instruction at the end of all predecessors |
5954 | for (i = 0; i < num_preds; i++) { |
5955 | remove_cur_instruction(i, true); |
5956 | } |
5957 | } |
5958 | } |
5959 | |
5960 | |
5961 | void EdgeMoveOptimizer::optimize_moves_at_block_begin(BlockBegin* block) { |
5962 | TRACE_LINEAR_SCAN(4, tty->print_cr("optimization moves at begin of block B%d" , block->block_id())); |
5963 | |
5964 | init_instructions(); |
5965 | int num_sux = block->number_of_sux(); |
5966 | |
5967 | LIR_OpList* cur_instructions = block->lir()->instructions_list(); |
5968 | |
5969 | assert(num_sux == 2, "method should not be called otherwise" ); |
5970 | assert(cur_instructions->last()->code() == lir_branch, "block with successor must end with branch" ); |
5971 | assert(cur_instructions->last()->as_OpBranch() != NULL, "branch must be LIR_OpBranch" ); |
5972 | assert(cur_instructions->last()->as_OpBranch()->cond() == lir_cond_always, "block must end with unconditional branch" ); |
5973 | |
5974 | if (cur_instructions->last()->info() != NULL) { |
5975 | // can no optimize instructions when debug info is needed |
5976 | return; |
5977 | } |
5978 | |
5979 | LIR_Op* branch = cur_instructions->at(cur_instructions->length() - 2); |
5980 | if (branch->info() != NULL || (branch->code() != lir_branch && branch->code() != lir_cond_float_branch)) { |
5981 | // not a valid case for optimization |
5982 | // currently, only blocks that end with two branches (conditional branch followed |
5983 | // by unconditional branch) are optimized |
5984 | return; |
5985 | } |
5986 | |
5987 | // now it is guaranteed that the block ends with two branch instructions. |
5988 | // the instructions are inserted at the end of the block before these two branches |
5989 | int insert_idx = cur_instructions->length() - 2; |
5990 | |
5991 | int i; |
5992 | #ifdef ASSERT |
5993 | for (i = insert_idx - 1; i >= 0; i--) { |
5994 | LIR_Op* op = cur_instructions->at(i); |
5995 | if ((op->code() == lir_branch || op->code() == lir_cond_float_branch) && ((LIR_OpBranch*)op)->block() != NULL) { |
5996 | assert(false, "block with two successors can have only two branch instructions" ); |
5997 | } |
5998 | } |
5999 | #endif |
6000 | |
6001 | // setup a list with the lir-instructions of all successors |
6002 | for (i = 0; i < num_sux; i++) { |
6003 | BlockBegin* sux = block->sux_at(i); |
6004 | LIR_OpList* sux_instructions = sux->lir()->instructions_list(); |
6005 | |
6006 | assert(sux_instructions->at(0)->code() == lir_label, "block must start with label" ); |
6007 | |
6008 | if (sux->number_of_preds() != 1) { |
6009 | // this can happen with switch-statements where multiple edges are between |
6010 | // the same blocks. |
6011 | return; |
6012 | } |
6013 | assert(sux->pred_at(0) == block, "invalid control flow" ); |
6014 | assert(!sux->is_set(BlockBegin::exception_entry_flag), "exception handlers not allowed" ); |
6015 | |
6016 | // ignore the label at the beginning of the block |
6017 | append_instructions(sux_instructions, 1); |
6018 | } |
6019 | |
6020 | // process lir-instructions while all successors begin with the same instruction |
6021 | while (true) { |
6022 | LIR_Op* op = instruction_at(0); |
6023 | for (i = 1; i < num_sux; i++) { |
6024 | if (operations_different(op, instruction_at(i))) { |
6025 | // these instructions are different and cannot be optimized -> |
6026 | // no further optimization possible |
6027 | return; |
6028 | } |
6029 | } |
6030 | |
6031 | TRACE_LINEAR_SCAN(4, tty->print("----- found instruction that is equal in all %d successors: " , num_sux); op->print()); |
6032 | |
6033 | // insert instruction at end of current block |
6034 | block->lir()->insert_before(insert_idx, op); |
6035 | insert_idx++; |
6036 | |
6037 | // delete the instructions at the beginning of all successors |
6038 | for (i = 0; i < num_sux; i++) { |
6039 | remove_cur_instruction(i, false); |
6040 | } |
6041 | } |
6042 | } |
6043 | |
6044 | |
6045 | // Implementation of ControlFlowOptimizer |
6046 | |
6047 | ControlFlowOptimizer::ControlFlowOptimizer() : |
6048 | _original_preds(4) |
6049 | { |
6050 | } |
6051 | |
6052 | void ControlFlowOptimizer::optimize(BlockList* code) { |
6053 | ControlFlowOptimizer optimizer = ControlFlowOptimizer(); |
6054 | |
6055 | // push the OSR entry block to the end so that we're not jumping over it. |
6056 | BlockBegin* osr_entry = code->at(0)->end()->as_Base()->osr_entry(); |
6057 | if (osr_entry) { |
6058 | int index = osr_entry->linear_scan_number(); |
6059 | assert(code->at(index) == osr_entry, "wrong index" ); |
6060 | code->remove_at(index); |
6061 | code->append(osr_entry); |
6062 | } |
6063 | |
6064 | optimizer.reorder_short_loops(code); |
6065 | optimizer.delete_empty_blocks(code); |
6066 | optimizer.delete_unnecessary_jumps(code); |
6067 | optimizer.delete_jumps_to_return(code); |
6068 | } |
6069 | |
6070 | void ControlFlowOptimizer::reorder_short_loop(BlockList* code, BlockBegin* , int ) { |
6071 | int i = header_idx + 1; |
6072 | int max_end = MIN2(header_idx + ShortLoopSize, code->length()); |
6073 | while (i < max_end && code->at(i)->loop_depth() >= header_block->loop_depth()) { |
6074 | i++; |
6075 | } |
6076 | |
6077 | if (i == code->length() || code->at(i)->loop_depth() < header_block->loop_depth()) { |
6078 | int end_idx = i - 1; |
6079 | BlockBegin* end_block = code->at(end_idx); |
6080 | |
6081 | if (end_block->number_of_sux() == 1 && end_block->sux_at(0) == header_block) { |
6082 | // short loop from header_idx to end_idx found -> reorder blocks such that |
6083 | // the header_block is the last block instead of the first block of the loop |
6084 | TRACE_LINEAR_SCAN(1, tty->print_cr("Reordering short loop: length %d, header B%d, end B%d" , |
6085 | end_idx - header_idx + 1, |
6086 | header_block->block_id(), end_block->block_id())); |
6087 | |
6088 | for (int j = header_idx; j < end_idx; j++) { |
6089 | code->at_put(j, code->at(j + 1)); |
6090 | } |
6091 | code->at_put(end_idx, header_block); |
6092 | |
6093 | // correct the flags so that any loop alignment occurs in the right place. |
6094 | assert(code->at(end_idx)->is_set(BlockBegin::backward_branch_target_flag), "must be backward branch target" ); |
6095 | code->at(end_idx)->clear(BlockBegin::backward_branch_target_flag); |
6096 | code->at(header_idx)->set(BlockBegin::backward_branch_target_flag); |
6097 | } |
6098 | } |
6099 | } |
6100 | |
6101 | void ControlFlowOptimizer::reorder_short_loops(BlockList* code) { |
6102 | for (int i = code->length() - 1; i >= 0; i--) { |
6103 | BlockBegin* block = code->at(i); |
6104 | |
6105 | if (block->is_set(BlockBegin::linear_scan_loop_header_flag)) { |
6106 | reorder_short_loop(code, block, i); |
6107 | } |
6108 | } |
6109 | |
6110 | DEBUG_ONLY(verify(code)); |
6111 | } |
6112 | |
6113 | // only blocks with exactly one successor can be deleted. Such blocks |
6114 | // must always end with an unconditional branch to this successor |
6115 | bool ControlFlowOptimizer::can_delete_block(BlockBegin* block) { |
6116 | if (block->number_of_sux() != 1 || block->number_of_exception_handlers() != 0 || block->is_entry_block()) { |
6117 | return false; |
6118 | } |
6119 | |
6120 | LIR_OpList* instructions = block->lir()->instructions_list(); |
6121 | |
6122 | assert(instructions->length() >= 2, "block must have label and branch" ); |
6123 | assert(instructions->at(0)->code() == lir_label, "first instruction must always be a label" ); |
6124 | assert(instructions->last()->as_OpBranch() != NULL, "last instrcution must always be a branch" ); |
6125 | assert(instructions->last()->as_OpBranch()->cond() == lir_cond_always, "branch must be unconditional" ); |
6126 | assert(instructions->last()->as_OpBranch()->block() == block->sux_at(0), "branch target must be the successor" ); |
6127 | |
6128 | // block must have exactly one successor |
6129 | |
6130 | if (instructions->length() == 2 && instructions->last()->info() == NULL) { |
6131 | return true; |
6132 | } |
6133 | return false; |
6134 | } |
6135 | |
6136 | // substitute branch targets in all branch-instructions of this blocks |
6137 | void ControlFlowOptimizer::substitute_branch_target(BlockBegin* block, BlockBegin* target_from, BlockBegin* target_to) { |
6138 | 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())); |
6139 | |
6140 | LIR_OpList* instructions = block->lir()->instructions_list(); |
6141 | |
6142 | assert(instructions->at(0)->code() == lir_label, "first instruction must always be a label" ); |
6143 | for (int i = instructions->length() - 1; i >= 1; i--) { |
6144 | LIR_Op* op = instructions->at(i); |
6145 | |
6146 | if (op->code() == lir_branch || op->code() == lir_cond_float_branch) { |
6147 | assert(op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch" ); |
6148 | LIR_OpBranch* branch = (LIR_OpBranch*)op; |
6149 | |
6150 | if (branch->block() == target_from) { |
6151 | branch->change_block(target_to); |
6152 | } |
6153 | if (branch->ublock() == target_from) { |
6154 | branch->change_ublock(target_to); |
6155 | } |
6156 | } |
6157 | } |
6158 | } |
6159 | |
6160 | void ControlFlowOptimizer::delete_empty_blocks(BlockList* code) { |
6161 | int old_pos = 0; |
6162 | int new_pos = 0; |
6163 | int num_blocks = code->length(); |
6164 | |
6165 | while (old_pos < num_blocks) { |
6166 | BlockBegin* block = code->at(old_pos); |
6167 | |
6168 | if (can_delete_block(block)) { |
6169 | BlockBegin* new_target = block->sux_at(0); |
6170 | |
6171 | // propagate backward branch target flag for correct code alignment |
6172 | if (block->is_set(BlockBegin::backward_branch_target_flag)) { |
6173 | new_target->set(BlockBegin::backward_branch_target_flag); |
6174 | } |
6175 | |
6176 | // collect a list with all predecessors that contains each predecessor only once |
6177 | // the predecessors of cur are changed during the substitution, so a copy of the |
6178 | // predecessor list is necessary |
6179 | int j; |
6180 | _original_preds.clear(); |
6181 | for (j = block->number_of_preds() - 1; j >= 0; j--) { |
6182 | BlockBegin* pred = block->pred_at(j); |
6183 | if (_original_preds.find(pred) == -1) { |
6184 | _original_preds.append(pred); |
6185 | } |
6186 | } |
6187 | |
6188 | for (j = _original_preds.length() - 1; j >= 0; j--) { |
6189 | BlockBegin* pred = _original_preds.at(j); |
6190 | substitute_branch_target(pred, block, new_target); |
6191 | pred->substitute_sux(block, new_target); |
6192 | } |
6193 | } else { |
6194 | // adjust position of this block in the block list if blocks before |
6195 | // have been deleted |
6196 | if (new_pos != old_pos) { |
6197 | code->at_put(new_pos, code->at(old_pos)); |
6198 | } |
6199 | new_pos++; |
6200 | } |
6201 | old_pos++; |
6202 | } |
6203 | code->trunc_to(new_pos); |
6204 | |
6205 | DEBUG_ONLY(verify(code)); |
6206 | } |
6207 | |
6208 | void ControlFlowOptimizer::delete_unnecessary_jumps(BlockList* code) { |
6209 | // skip the last block because there a branch is always necessary |
6210 | for (int i = code->length() - 2; i >= 0; i--) { |
6211 | BlockBegin* block = code->at(i); |
6212 | LIR_OpList* instructions = block->lir()->instructions_list(); |
6213 | |
6214 | LIR_Op* last_op = instructions->last(); |
6215 | if (last_op->code() == lir_branch) { |
6216 | assert(last_op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch" ); |
6217 | LIR_OpBranch* last_branch = (LIR_OpBranch*)last_op; |
6218 | |
6219 | assert(last_branch->block() != NULL, "last branch must always have a block as target" ); |
6220 | assert(last_branch->label() == last_branch->block()->label(), "must be equal" ); |
6221 | |
6222 | if (last_branch->info() == NULL) { |
6223 | if (last_branch->block() == code->at(i + 1)) { |
6224 | |
6225 | TRACE_LINEAR_SCAN(3, tty->print_cr("Deleting unconditional branch at end of block B%d" , block->block_id())); |
6226 | |
6227 | // delete last branch instruction |
6228 | instructions->trunc_to(instructions->length() - 1); |
6229 | |
6230 | } else { |
6231 | LIR_Op* prev_op = instructions->at(instructions->length() - 2); |
6232 | if (prev_op->code() == lir_branch || prev_op->code() == lir_cond_float_branch) { |
6233 | assert(prev_op->as_OpBranch() != NULL, "branch must be of type LIR_OpBranch" ); |
6234 | LIR_OpBranch* prev_branch = (LIR_OpBranch*)prev_op; |
6235 | |
6236 | if (prev_branch->stub() == NULL) { |
6237 | |
6238 | LIR_Op2* prev_cmp = NULL; |
6239 | // There might be a cmove inserted for profiling which depends on the same |
6240 | // compare. If we change the condition of the respective compare, we have |
6241 | // to take care of this cmove as well. |
6242 | LIR_Op2* prev_cmove = NULL; |
6243 | |
6244 | for(int j = instructions->length() - 3; j >= 0 && prev_cmp == NULL; j--) { |
6245 | prev_op = instructions->at(j); |
6246 | // check for the cmove |
6247 | if (prev_op->code() == lir_cmove) { |
6248 | assert(prev_op->as_Op2() != NULL, "cmove must be of type LIR_Op2" ); |
6249 | prev_cmove = (LIR_Op2*)prev_op; |
6250 | assert(prev_branch->cond() == prev_cmove->condition(), "should be the same" ); |
6251 | } |
6252 | if (prev_op->code() == lir_cmp) { |
6253 | assert(prev_op->as_Op2() != NULL, "branch must be of type LIR_Op2" ); |
6254 | prev_cmp = (LIR_Op2*)prev_op; |
6255 | assert(prev_branch->cond() == prev_cmp->condition(), "should be the same" ); |
6256 | } |
6257 | } |
6258 | // Guarantee because it is dereferenced below. |
6259 | guarantee(prev_cmp != NULL, "should have found comp instruction for branch" ); |
6260 | if (prev_branch->block() == code->at(i + 1) && prev_branch->info() == NULL) { |
6261 | |
6262 | TRACE_LINEAR_SCAN(3, tty->print_cr("Negating conditional branch and deleting unconditional branch at end of block B%d" , block->block_id())); |
6263 | |
6264 | // eliminate a conditional branch to the immediate successor |
6265 | prev_branch->change_block(last_branch->block()); |
6266 | prev_branch->negate_cond(); |
6267 | prev_cmp->set_condition(prev_branch->cond()); |
6268 | instructions->trunc_to(instructions->length() - 1); |
6269 | // if we do change the condition, we have to change the cmove as well |
6270 | if (prev_cmove != NULL) { |
6271 | prev_cmove->set_condition(prev_branch->cond()); |
6272 | LIR_Opr t = prev_cmove->in_opr1(); |
6273 | prev_cmove->set_in_opr1(prev_cmove->in_opr2()); |
6274 | prev_cmove->set_in_opr2(t); |
6275 | } |
6276 | } |
6277 | } |
6278 | } |
6279 | } |
6280 | } |
6281 | } |
6282 | } |
6283 | |
6284 | DEBUG_ONLY(verify(code)); |
6285 | } |
6286 | |
6287 | void ControlFlowOptimizer::delete_jumps_to_return(BlockList* code) { |
6288 | #ifdef ASSERT |
6289 | ResourceBitMap return_converted(BlockBegin::number_of_blocks()); |
6290 | #endif |
6291 | |
6292 | for (int i = code->length() - 1; i >= 0; i--) { |
6293 | BlockBegin* block = code->at(i); |
6294 | LIR_OpList* cur_instructions = block->lir()->instructions_list(); |
6295 | LIR_Op* cur_last_op = cur_instructions->last(); |
6296 | |
6297 | assert(cur_instructions->at(0)->code() == lir_label, "first instruction must always be a label" ); |
6298 | if (cur_instructions->length() == 2 && cur_last_op->code() == lir_return) { |
6299 | // the block contains only a label and a return |
6300 | // if a predecessor ends with an unconditional jump to this block, then the jump |
6301 | // can be replaced with a return instruction |
6302 | // |
6303 | // Note: the original block with only a return statement cannot be deleted completely |
6304 | // because the predecessors might have other (conditional) jumps to this block |
6305 | // -> this may lead to unnecesary return instructions in the final code |
6306 | |
6307 | assert(cur_last_op->info() == NULL, "return instructions do not have debug information" ); |
6308 | assert(block->number_of_sux() == 0 || |
6309 | (return_converted.at(block->block_id()) && block->number_of_sux() == 1), |
6310 | "blocks that end with return must not have successors" ); |
6311 | |
6312 | assert(cur_last_op->as_Op1() != NULL, "return must be LIR_Op1" ); |
6313 | LIR_Opr return_opr = ((LIR_Op1*)cur_last_op)->in_opr(); |
6314 | |
6315 | for (int j = block->number_of_preds() - 1; j >= 0; j--) { |
6316 | BlockBegin* pred = block->pred_at(j); |
6317 | LIR_OpList* pred_instructions = pred->lir()->instructions_list(); |
6318 | LIR_Op* pred_last_op = pred_instructions->last(); |
6319 | |
6320 | if (pred_last_op->code() == lir_branch) { |
6321 | assert(pred_last_op->as_OpBranch() != NULL, "branch must be LIR_OpBranch" ); |
6322 | LIR_OpBranch* pred_last_branch = (LIR_OpBranch*)pred_last_op; |
6323 | |
6324 | if (pred_last_branch->block() == block && pred_last_branch->cond() == lir_cond_always && pred_last_branch->info() == NULL) { |
6325 | // replace the jump to a return with a direct return |
6326 | // Note: currently the edge between the blocks is not deleted |
6327 | pred_instructions->at_put(pred_instructions->length() - 1, new LIR_Op1(lir_return, return_opr)); |
6328 | #ifdef ASSERT |
6329 | return_converted.set_bit(pred->block_id()); |
6330 | #endif |
6331 | } |
6332 | } |
6333 | } |
6334 | } |
6335 | } |
6336 | } |
6337 | |
6338 | |
6339 | #ifdef ASSERT |
6340 | void ControlFlowOptimizer::verify(BlockList* code) { |
6341 | for (int i = 0; i < code->length(); i++) { |
6342 | BlockBegin* block = code->at(i); |
6343 | LIR_OpList* instructions = block->lir()->instructions_list(); |
6344 | |
6345 | int j; |
6346 | for (j = 0; j < instructions->length(); j++) { |
6347 | LIR_OpBranch* op_branch = instructions->at(j)->as_OpBranch(); |
6348 | |
6349 | if (op_branch != NULL) { |
6350 | assert(op_branch->block() == NULL || code->find(op_branch->block()) != -1, "branch target not valid" ); |
6351 | assert(op_branch->ublock() == NULL || code->find(op_branch->ublock()) != -1, "branch target not valid" ); |
6352 | } |
6353 | } |
6354 | |
6355 | for (j = 0; j < block->number_of_sux() - 1; j++) { |
6356 | BlockBegin* sux = block->sux_at(j); |
6357 | assert(code->find(sux) != -1, "successor not valid" ); |
6358 | } |
6359 | |
6360 | for (j = 0; j < block->number_of_preds() - 1; j++) { |
6361 | BlockBegin* pred = block->pred_at(j); |
6362 | assert(code->find(pred) != -1, "successor not valid" ); |
6363 | } |
6364 | } |
6365 | } |
6366 | #endif |
6367 | |
6368 | |
6369 | #ifndef PRODUCT |
6370 | |
6371 | // Implementation of LinearStatistic |
6372 | |
6373 | const char* LinearScanStatistic::counter_name(int counter_idx) { |
6374 | switch (counter_idx) { |
6375 | case counter_method: return "compiled methods" ; |
6376 | case counter_fpu_method: return "methods using fpu" ; |
6377 | case counter_loop_method: return "methods with loops" ; |
6378 | case counter_exception_method:return "methods with xhandler" ; |
6379 | |
6380 | case counter_loop: return "loops" ; |
6381 | case counter_block: return "blocks" ; |
6382 | case counter_loop_block: return "blocks inside loop" ; |
6383 | case counter_exception_block: return "exception handler entries" ; |
6384 | case counter_interval: return "intervals" ; |
6385 | case counter_fixed_interval: return "fixed intervals" ; |
6386 | case counter_range: return "ranges" ; |
6387 | case counter_fixed_range: return "fixed ranges" ; |
6388 | case counter_use_pos: return "use positions" ; |
6389 | case counter_fixed_use_pos: return "fixed use positions" ; |
6390 | case counter_spill_slots: return "spill slots" ; |
6391 | |
6392 | // counter for classes of lir instructions |
6393 | case counter_instruction: return "total instructions" ; |
6394 | case counter_label: return "labels" ; |
6395 | case counter_entry: return "method entries" ; |
6396 | case counter_return: return "method returns" ; |
6397 | case counter_call: return "method calls" ; |
6398 | case counter_move: return "moves" ; |
6399 | case counter_cmp: return "compare" ; |
6400 | case counter_cond_branch: return "conditional branches" ; |
6401 | case counter_uncond_branch: return "unconditional branches" ; |
6402 | case counter_stub_branch: return "branches to stub" ; |
6403 | case counter_alu: return "artithmetic + logic" ; |
6404 | case counter_alloc: return "allocations" ; |
6405 | case counter_sync: return "synchronisation" ; |
6406 | case counter_throw: return "throw" ; |
6407 | case counter_unwind: return "unwind" ; |
6408 | case counter_typecheck: return "type+null-checks" ; |
6409 | case counter_fpu_stack: return "fpu-stack" ; |
6410 | case counter_misc_inst: return "other instructions" ; |
6411 | case counter_other_inst: return "misc. instructions" ; |
6412 | |
6413 | // counter for different types of moves |
6414 | case counter_move_total: return "total moves" ; |
6415 | case counter_move_reg_reg: return "register->register" ; |
6416 | case counter_move_reg_stack: return "register->stack" ; |
6417 | case counter_move_stack_reg: return "stack->register" ; |
6418 | case counter_move_stack_stack:return "stack->stack" ; |
6419 | case counter_move_reg_mem: return "register->memory" ; |
6420 | case counter_move_mem_reg: return "memory->register" ; |
6421 | case counter_move_const_any: return "constant->any" ; |
6422 | |
6423 | case blank_line_1: return "" ; |
6424 | case blank_line_2: return "" ; |
6425 | |
6426 | default: ShouldNotReachHere(); return "" ; |
6427 | } |
6428 | } |
6429 | |
6430 | LinearScanStatistic::Counter LinearScanStatistic::base_counter(int counter_idx) { |
6431 | if (counter_idx == counter_fpu_method || counter_idx == counter_loop_method || counter_idx == counter_exception_method) { |
6432 | return counter_method; |
6433 | } else if (counter_idx == counter_loop_block || counter_idx == counter_exception_block) { |
6434 | return counter_block; |
6435 | } else if (counter_idx >= counter_instruction && counter_idx <= counter_other_inst) { |
6436 | return counter_instruction; |
6437 | } else if (counter_idx >= counter_move_total && counter_idx <= counter_move_const_any) { |
6438 | return counter_move_total; |
6439 | } |
6440 | return invalid_counter; |
6441 | } |
6442 | |
6443 | LinearScanStatistic::LinearScanStatistic() { |
6444 | for (int i = 0; i < number_of_counters; i++) { |
6445 | _counters_sum[i] = 0; |
6446 | _counters_max[i] = -1; |
6447 | } |
6448 | |
6449 | } |
6450 | |
6451 | // add the method-local numbers to the total sum |
6452 | void LinearScanStatistic::sum_up(LinearScanStatistic &method_statistic) { |
6453 | for (int i = 0; i < number_of_counters; i++) { |
6454 | _counters_sum[i] += method_statistic._counters_sum[i]; |
6455 | _counters_max[i] = MAX2(_counters_max[i], method_statistic._counters_sum[i]); |
6456 | } |
6457 | } |
6458 | |
6459 | void LinearScanStatistic::print(const char* title) { |
6460 | if (CountLinearScan || TraceLinearScanLevel > 0) { |
6461 | tty->cr(); |
6462 | tty->print_cr("***** LinearScan statistic - %s *****" , title); |
6463 | |
6464 | for (int i = 0; i < number_of_counters; i++) { |
6465 | if (_counters_sum[i] > 0 || _counters_max[i] >= 0) { |
6466 | tty->print("%25s: %8d" , counter_name(i), _counters_sum[i]); |
6467 | |
6468 | LinearScanStatistic::Counter cntr = base_counter(i); |
6469 | if (cntr != invalid_counter) { |
6470 | tty->print(" (%5.1f%%) " , _counters_sum[i] * 100.0 / _counters_sum[cntr]); |
6471 | } else { |
6472 | tty->print(" " ); |
6473 | } |
6474 | |
6475 | if (_counters_max[i] >= 0) { |
6476 | tty->print("%8d" , _counters_max[i]); |
6477 | } |
6478 | } |
6479 | tty->cr(); |
6480 | } |
6481 | } |
6482 | } |
6483 | |
6484 | void LinearScanStatistic::collect(LinearScan* allocator) { |
6485 | inc_counter(counter_method); |
6486 | if (allocator->has_fpu_registers()) { |
6487 | inc_counter(counter_fpu_method); |
6488 | } |
6489 | if (allocator->num_loops() > 0) { |
6490 | inc_counter(counter_loop_method); |
6491 | } |
6492 | inc_counter(counter_loop, allocator->num_loops()); |
6493 | inc_counter(counter_spill_slots, allocator->max_spills()); |
6494 | |
6495 | int i; |
6496 | for (i = 0; i < allocator->interval_count(); i++) { |
6497 | Interval* cur = allocator->interval_at(i); |
6498 | |
6499 | if (cur != NULL) { |
6500 | inc_counter(counter_interval); |
6501 | inc_counter(counter_use_pos, cur->num_use_positions()); |
6502 | if (LinearScan::is_precolored_interval(cur)) { |
6503 | inc_counter(counter_fixed_interval); |
6504 | inc_counter(counter_fixed_use_pos, cur->num_use_positions()); |
6505 | } |
6506 | |
6507 | Range* range = cur->first(); |
6508 | while (range != Range::end()) { |
6509 | inc_counter(counter_range); |
6510 | if (LinearScan::is_precolored_interval(cur)) { |
6511 | inc_counter(counter_fixed_range); |
6512 | } |
6513 | range = range->next(); |
6514 | } |
6515 | } |
6516 | } |
6517 | |
6518 | bool has_xhandlers = false; |
6519 | // Note: only count blocks that are in code-emit order |
6520 | for (i = 0; i < allocator->ir()->code()->length(); i++) { |
6521 | BlockBegin* cur = allocator->ir()->code()->at(i); |
6522 | |
6523 | inc_counter(counter_block); |
6524 | if (cur->loop_depth() > 0) { |
6525 | inc_counter(counter_loop_block); |
6526 | } |
6527 | if (cur->is_set(BlockBegin::exception_entry_flag)) { |
6528 | inc_counter(counter_exception_block); |
6529 | has_xhandlers = true; |
6530 | } |
6531 | |
6532 | LIR_OpList* instructions = cur->lir()->instructions_list(); |
6533 | for (int j = 0; j < instructions->length(); j++) { |
6534 | LIR_Op* op = instructions->at(j); |
6535 | |
6536 | inc_counter(counter_instruction); |
6537 | |
6538 | switch (op->code()) { |
6539 | case lir_label: inc_counter(counter_label); break; |
6540 | case lir_std_entry: |
6541 | case lir_osr_entry: inc_counter(counter_entry); break; |
6542 | case lir_return: inc_counter(counter_return); break; |
6543 | |
6544 | case lir_rtcall: |
6545 | case lir_static_call: |
6546 | case lir_optvirtual_call: |
6547 | case lir_virtual_call: inc_counter(counter_call); break; |
6548 | |
6549 | case lir_move: { |
6550 | inc_counter(counter_move); |
6551 | inc_counter(counter_move_total); |
6552 | |
6553 | LIR_Opr in = op->as_Op1()->in_opr(); |
6554 | LIR_Opr res = op->as_Op1()->result_opr(); |
6555 | if (in->is_register()) { |
6556 | if (res->is_register()) { |
6557 | inc_counter(counter_move_reg_reg); |
6558 | } else if (res->is_stack()) { |
6559 | inc_counter(counter_move_reg_stack); |
6560 | } else if (res->is_address()) { |
6561 | inc_counter(counter_move_reg_mem); |
6562 | } else { |
6563 | ShouldNotReachHere(); |
6564 | } |
6565 | } else if (in->is_stack()) { |
6566 | if (res->is_register()) { |
6567 | inc_counter(counter_move_stack_reg); |
6568 | } else { |
6569 | inc_counter(counter_move_stack_stack); |
6570 | } |
6571 | } else if (in->is_address()) { |
6572 | assert(res->is_register(), "must be" ); |
6573 | inc_counter(counter_move_mem_reg); |
6574 | } else if (in->is_constant()) { |
6575 | inc_counter(counter_move_const_any); |
6576 | } else { |
6577 | ShouldNotReachHere(); |
6578 | } |
6579 | break; |
6580 | } |
6581 | |
6582 | case lir_cmp: inc_counter(counter_cmp); break; |
6583 | |
6584 | case lir_branch: |
6585 | case lir_cond_float_branch: { |
6586 | LIR_OpBranch* branch = op->as_OpBranch(); |
6587 | if (branch->block() == NULL) { |
6588 | inc_counter(counter_stub_branch); |
6589 | } else if (branch->cond() == lir_cond_always) { |
6590 | inc_counter(counter_uncond_branch); |
6591 | } else { |
6592 | inc_counter(counter_cond_branch); |
6593 | } |
6594 | break; |
6595 | } |
6596 | |
6597 | case lir_neg: |
6598 | case lir_add: |
6599 | case lir_sub: |
6600 | case lir_mul: |
6601 | case lir_mul_strictfp: |
6602 | case lir_div: |
6603 | case lir_div_strictfp: |
6604 | case lir_rem: |
6605 | case lir_sqrt: |
6606 | case lir_abs: |
6607 | case lir_log10: |
6608 | case lir_logic_and: |
6609 | case lir_logic_or: |
6610 | case lir_logic_xor: |
6611 | case lir_shl: |
6612 | case lir_shr: |
6613 | case lir_ushr: inc_counter(counter_alu); break; |
6614 | |
6615 | case lir_alloc_object: |
6616 | case lir_alloc_array: inc_counter(counter_alloc); break; |
6617 | |
6618 | case lir_monaddr: |
6619 | case lir_lock: |
6620 | case lir_unlock: inc_counter(counter_sync); break; |
6621 | |
6622 | case lir_throw: inc_counter(counter_throw); break; |
6623 | |
6624 | case lir_unwind: inc_counter(counter_unwind); break; |
6625 | |
6626 | case lir_null_check: |
6627 | case lir_leal: |
6628 | case lir_instanceof: |
6629 | case lir_checkcast: |
6630 | case lir_store_check: inc_counter(counter_typecheck); break; |
6631 | |
6632 | case lir_fpop_raw: |
6633 | case lir_fxch: |
6634 | case lir_fld: inc_counter(counter_fpu_stack); break; |
6635 | |
6636 | case lir_nop: |
6637 | case lir_push: |
6638 | case lir_pop: |
6639 | case lir_convert: |
6640 | case lir_roundfp: |
6641 | case lir_cmove: inc_counter(counter_misc_inst); break; |
6642 | |
6643 | default: inc_counter(counter_other_inst); break; |
6644 | } |
6645 | } |
6646 | } |
6647 | |
6648 | if (has_xhandlers) { |
6649 | inc_counter(counter_exception_method); |
6650 | } |
6651 | } |
6652 | |
6653 | void LinearScanStatistic::compute(LinearScan* allocator, LinearScanStatistic &global_statistic) { |
6654 | if (CountLinearScan || TraceLinearScanLevel > 0) { |
6655 | |
6656 | LinearScanStatistic local_statistic = LinearScanStatistic(); |
6657 | |
6658 | local_statistic.collect(allocator); |
6659 | global_statistic.sum_up(local_statistic); |
6660 | |
6661 | if (TraceLinearScanLevel > 2) { |
6662 | local_statistic.print("current local statistic" ); |
6663 | } |
6664 | } |
6665 | } |
6666 | |
6667 | |
6668 | // Implementation of LinearTimers |
6669 | |
6670 | LinearScanTimers::LinearScanTimers() { |
6671 | for (int i = 0; i < number_of_timers; i++) { |
6672 | timer(i)->reset(); |
6673 | } |
6674 | } |
6675 | |
6676 | const char* LinearScanTimers::timer_name(int idx) { |
6677 | switch (idx) { |
6678 | case timer_do_nothing: return "Nothing (Time Check)" ; |
6679 | case timer_number_instructions: return "Number Instructions" ; |
6680 | case timer_compute_local_live_sets: return "Local Live Sets" ; |
6681 | case timer_compute_global_live_sets: return "Global Live Sets" ; |
6682 | case timer_build_intervals: return "Build Intervals" ; |
6683 | case timer_sort_intervals_before: return "Sort Intervals Before" ; |
6684 | case timer_allocate_registers: return "Allocate Registers" ; |
6685 | case timer_resolve_data_flow: return "Resolve Data Flow" ; |
6686 | case timer_sort_intervals_after: return "Sort Intervals After" ; |
6687 | case timer_eliminate_spill_moves: return "Spill optimization" ; |
6688 | case timer_assign_reg_num: return "Assign Reg Num" ; |
6689 | case timer_allocate_fpu_stack: return "Allocate FPU Stack" ; |
6690 | case timer_optimize_lir: return "Optimize LIR" ; |
6691 | default: ShouldNotReachHere(); return "" ; |
6692 | } |
6693 | } |
6694 | |
6695 | void LinearScanTimers::begin_method() { |
6696 | if (TimeEachLinearScan) { |
6697 | // reset all timers to measure only current method |
6698 | for (int i = 0; i < number_of_timers; i++) { |
6699 | timer(i)->reset(); |
6700 | } |
6701 | } |
6702 | } |
6703 | |
6704 | void LinearScanTimers::end_method(LinearScan* allocator) { |
6705 | if (TimeEachLinearScan) { |
6706 | |
6707 | double c = timer(timer_do_nothing)->seconds(); |
6708 | double total = 0; |
6709 | for (int i = 1; i < number_of_timers; i++) { |
6710 | total += timer(i)->seconds() - c; |
6711 | } |
6712 | |
6713 | if (total >= 0.0005) { |
6714 | // print all information in one line for automatic processing |
6715 | tty->print("@" ); allocator->compilation()->method()->print_name(); |
6716 | |
6717 | tty->print("@ %d " , allocator->compilation()->method()->code_size()); |
6718 | tty->print("@ %d " , allocator->block_at(allocator->block_count() - 1)->last_lir_instruction_id() / 2); |
6719 | tty->print("@ %d " , allocator->block_count()); |
6720 | tty->print("@ %d " , allocator->num_virtual_regs()); |
6721 | tty->print("@ %d " , allocator->interval_count()); |
6722 | tty->print("@ %d " , allocator->_num_calls); |
6723 | tty->print("@ %d " , allocator->num_loops()); |
6724 | |
6725 | tty->print("@ %6.6f " , total); |
6726 | for (int i = 1; i < number_of_timers; i++) { |
6727 | tty->print("@ %4.1f " , ((timer(i)->seconds() - c) / total) * 100); |
6728 | } |
6729 | tty->cr(); |
6730 | } |
6731 | } |
6732 | } |
6733 | |
6734 | void LinearScanTimers::print(double total_time) { |
6735 | if (TimeLinearScan) { |
6736 | // correction value: sum of dummy-timer that only measures the time that |
6737 | // is necesary to start and stop itself |
6738 | double c = timer(timer_do_nothing)->seconds(); |
6739 | |
6740 | for (int i = 0; i < number_of_timers; i++) { |
6741 | double t = timer(i)->seconds(); |
6742 | 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); |
6743 | } |
6744 | } |
6745 | } |
6746 | |
6747 | #endif // #ifndef PRODUCT |
6748 | |