1 /*
2 * Copyright (c) 2003, 2015, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2014, Red Hat Inc. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "precompiled.hpp"
27 #include "interp_masm_aarch64.hpp"
28 #include "interpreter/interpreter.hpp"
29 #include "interpreter/interpreterRuntime.hpp"
30 #include "logging/log.hpp"
31 #include "oops/arrayOop.hpp"
32 #include "oops/markOop.hpp"
33 #include "oops/methodData.hpp"
34 #include "oops/method.hpp"
35 #include "prims/jvmtiExport.hpp"
36 #include "prims/jvmtiThreadState.hpp"
37 #include "runtime/basicLock.hpp"
38 #include "runtime/biasedLocking.hpp"
39 #include "runtime/sharedRuntime.hpp"
40 #include "runtime/thread.inline.hpp"
41
42
43 void InterpreterMacroAssembler::narrow(Register result) {
44
45 // Get method->_constMethod->_result_type
46 ldr(rscratch1, Address(rfp, frame::interpreter_frame_method_offset * wordSize));
47 ldr(rscratch1, Address(rscratch1, Method::const_offset()));
48 ldrb(rscratch1, Address(rscratch1, ConstMethod::result_type_offset()));
49
50 Label done, notBool, notByte, notChar;
51
52 // common case first
53 cmpw(rscratch1, T_INT);
54 br(Assembler::EQ, done);
55
56 // mask integer result to narrower return type.
57 cmpw(rscratch1, T_BOOLEAN);
58 br(Assembler::NE, notBool);
59 andw(result, result, 0x1);
60 b(done);
61
62 bind(notBool);
63 cmpw(rscratch1, T_BYTE);
64 br(Assembler::NE, notByte);
65 sbfx(result, result, 0, 8);
66 b(done);
67
68 bind(notByte);
69 cmpw(rscratch1, T_CHAR);
70 br(Assembler::NE, notChar);
71 ubfx(result, result, 0, 16); // truncate upper 16 bits
72 b(done);
73
74 bind(notChar);
75 sbfx(result, result, 0, 16); // sign-extend short
76
77 // Nothing to do for T_INT
78 bind(done);
79 }
80
81 void InterpreterMacroAssembler::jump_to_entry(address entry) {
82 assert(entry, "Entry must have been generated by now");
83 b(entry);
84 }
85
86 void InterpreterMacroAssembler::check_and_handle_popframe(Register java_thread) {
87 if (JvmtiExport::can_pop_frame()) {
88 Label L;
89 // Initiate popframe handling only if it is not already being
90 // processed. If the flag has the popframe_processing bit set, it
91 // means that this code is called *during* popframe handling - we
92 // don't want to reenter.
93 // This method is only called just after the call into the vm in
94 // call_VM_base, so the arg registers are available.
95 ldrw(rscratch1, Address(rthread, JavaThread::popframe_condition_offset()));
96 tbz(rscratch1, exact_log2(JavaThread::popframe_pending_bit), L);
97 tbnz(rscratch1, exact_log2(JavaThread::popframe_processing_bit), L);
98 // Call Interpreter::remove_activation_preserving_args_entry() to get the
99 // address of the same-named entrypoint in the generated interpreter code.
100 call_VM_leaf(CAST_FROM_FN_PTR(address, Interpreter::remove_activation_preserving_args_entry));
101 br(r0);
102 bind(L);
103 }
104 }
105
106
107 void InterpreterMacroAssembler::load_earlyret_value(TosState state) {
108 ldr(r2, Address(rthread, JavaThread::jvmti_thread_state_offset()));
109 const Address tos_addr(r2, JvmtiThreadState::earlyret_tos_offset());
110 const Address oop_addr(r2, JvmtiThreadState::earlyret_oop_offset());
111 const Address val_addr(r2, JvmtiThreadState::earlyret_value_offset());
112 switch (state) {
113 case atos: ldr(r0, oop_addr);
114 str(zr, oop_addr);
115 verify_oop(r0, state); break;
116 case ltos: ldr(r0, val_addr); break;
117 case btos: // fall through
118 case ztos: // fall through
119 case ctos: // fall through
120 case stos: // fall through
121 case itos: ldrw(r0, val_addr); break;
122 case ftos: ldrs(v0, val_addr); break;
123 case dtos: ldrd(v0, val_addr); break;
124 case vtos: /* nothing to do */ break;
125 default : ShouldNotReachHere();
126 }
127 // Clean up tos value in the thread object
128 movw(rscratch1, (int) ilgl);
129 strw(rscratch1, tos_addr);
130 strw(zr, val_addr);
131 }
132
133
134 void InterpreterMacroAssembler::check_and_handle_earlyret(Register java_thread) {
135 if (JvmtiExport::can_force_early_return()) {
136 Label L;
137 ldr(rscratch1, Address(rthread, JavaThread::jvmti_thread_state_offset()));
138 cbz(rscratch1, L); // if (thread->jvmti_thread_state() == NULL) exit;
139
140 // Initiate earlyret handling only if it is not already being processed.
141 // If the flag has the earlyret_processing bit set, it means that this code
142 // is called *during* earlyret handling - we don't want to reenter.
143 ldrw(rscratch1, Address(rscratch1, JvmtiThreadState::earlyret_state_offset()));
144 cmpw(rscratch1, JvmtiThreadState::earlyret_pending);
145 br(Assembler::NE, L);
146
147 // Call Interpreter::remove_activation_early_entry() to get the address of the
148 // same-named entrypoint in the generated interpreter code.
149 ldr(rscratch1, Address(rthread, JavaThread::jvmti_thread_state_offset()));
150 ldrw(rscratch1, Address(rscratch1, JvmtiThreadState::earlyret_tos_offset()));
151 call_VM_leaf(CAST_FROM_FN_PTR(address, Interpreter::remove_activation_early_entry), rscratch1);
152 br(r0);
153 bind(L);
154 }
155 }
156
157 void InterpreterMacroAssembler::get_unsigned_2_byte_index_at_bcp(
158 Register reg,
159 int bcp_offset) {
160 assert(bcp_offset >= 0, "bcp is still pointing to start of bytecode");
161 ldrh(reg, Address(rbcp, bcp_offset));
162 rev16(reg, reg);
163 }
164
165 void InterpreterMacroAssembler::get_dispatch() {
166 unsigned long offset;
167 adrp(rdispatch, ExternalAddress((address)Interpreter::dispatch_table()), offset);
168 lea(rdispatch, Address(rdispatch, offset));
169 }
170
171 void InterpreterMacroAssembler::get_cache_index_at_bcp(Register index,
172 int bcp_offset,
173 size_t index_size) {
174 assert(bcp_offset > 0, "bcp is still pointing to start of bytecode");
175 if (index_size == sizeof(u2)) {
176 load_unsigned_short(index, Address(rbcp, bcp_offset));
177 } else if (index_size == sizeof(u4)) {
178 // assert(EnableInvokeDynamic, "giant index used only for JSR 292");
179 ldrw(index, Address(rbcp, bcp_offset));
180 // Check if the secondary index definition is still ~x, otherwise
181 // we have to change the following assembler code to calculate the
182 // plain index.
183 assert(ConstantPool::decode_invokedynamic_index(~123) == 123, "else change next line");
184 eonw(index, index, zr); // convert to plain index
185 } else if (index_size == sizeof(u1)) {
186 load_unsigned_byte(index, Address(rbcp, bcp_offset));
187 } else {
188 ShouldNotReachHere();
189 }
190 }
191
192 // Return
193 // Rindex: index into constant pool
194 // Rcache: address of cache entry - ConstantPoolCache::base_offset()
195 //
196 // A caller must add ConstantPoolCache::base_offset() to Rcache to get
197 // the true address of the cache entry.
198 //
199 void InterpreterMacroAssembler::get_cache_and_index_at_bcp(Register cache,
200 Register index,
201 int bcp_offset,
202 size_t index_size) {
203 assert_different_registers(cache, index);
204 assert_different_registers(cache, rcpool);
205 get_cache_index_at_bcp(index, bcp_offset, index_size);
206 assert(sizeof(ConstantPoolCacheEntry) == 4 * wordSize, "adjust code below");
207 // convert from field index to ConstantPoolCacheEntry
208 // aarch64 already has the cache in rcpool so there is no need to
209 // install it in cache. instead we pre-add the indexed offset to
210 // rcpool and return it in cache. All clients of this method need to
211 // be modified accordingly.
212 add(cache, rcpool, index, Assembler::LSL, 5);
213 }
214
215
216 void InterpreterMacroAssembler::get_cache_and_index_and_bytecode_at_bcp(Register cache,
217 Register index,
218 Register bytecode,
219 int byte_no,
220 int bcp_offset,
221 size_t index_size) {
222 get_cache_and_index_at_bcp(cache, index, bcp_offset, index_size);
223 // We use a 32-bit load here since the layout of 64-bit words on
224 // little-endian machines allow us that.
225 // n.b. unlike x86 cache already includes the index offset
226 lea(bytecode, Address(cache,
227 ConstantPoolCache::base_offset()
228 + ConstantPoolCacheEntry::indices_offset()));
229 ldarw(bytecode, bytecode);
230 const int shift_count = (1 + byte_no) * BitsPerByte;
231 ubfx(bytecode, bytecode, shift_count, BitsPerByte);
232 }
233
234 void InterpreterMacroAssembler::get_cache_entry_pointer_at_bcp(Register cache,
235 Register tmp,
236 int bcp_offset,
237 size_t index_size) {
238 assert(cache != tmp, "must use different register");
239 get_cache_index_at_bcp(tmp, bcp_offset, index_size);
240 assert(sizeof(ConstantPoolCacheEntry) == 4 * wordSize, "adjust code below");
241 // convert from field index to ConstantPoolCacheEntry index
242 // and from word offset to byte offset
243 assert(exact_log2(in_bytes(ConstantPoolCacheEntry::size_in_bytes())) == 2 + LogBytesPerWord, "else change next line");
244 ldr(cache, Address(rfp, frame::interpreter_frame_cache_offset * wordSize));
245 // skip past the header
246 add(cache, cache, in_bytes(ConstantPoolCache::base_offset()));
247 add(cache, cache, tmp, Assembler::LSL, 2 + LogBytesPerWord); // construct pointer to cache entry
248 }
249
250 void InterpreterMacroAssembler::get_method_counters(Register method,
251 Register mcs, Label& skip) {
252 Label has_counters;
253 ldr(mcs, Address(method, Method::method_counters_offset()));
254 cbnz(mcs, has_counters);
255 call_VM(noreg, CAST_FROM_FN_PTR(address,
256 InterpreterRuntime::build_method_counters), method);
257 ldr(mcs, Address(method, Method::method_counters_offset()));
258 cbz(mcs, skip); // No MethodCounters allocated, OutOfMemory
259 bind(has_counters);
260 }
261
262 // Load object from cpool->resolved_references(index)
263 void InterpreterMacroAssembler::load_resolved_reference_at_index(
264 Register result, Register index) {
265 assert_different_registers(result, index);
266 // convert from field index to resolved_references() index and from
267 // word index to byte offset. Since this is a java object, it can be compressed
268 Register tmp = index; // reuse
269 lslw(tmp, tmp, LogBytesPerHeapOop);
270
271 get_constant_pool(result);
272 // load pointer for resolved_references[] objArray
273 ldr(result, Address(result, ConstantPool::resolved_references_offset_in_bytes()));
274 // JNIHandles::resolve(obj);
275 ldr(result, Address(result, 0));
276 // Add in the index
277 add(result, result, tmp);
278 load_heap_oop(result, Address(result, arrayOopDesc::base_offset_in_bytes(T_OBJECT)));
279 }
280
281 // Generate a subtype check: branch to ok_is_subtype if sub_klass is a
282 // subtype of super_klass.
283 //
284 // Args:
285 // r0: superklass
286 // Rsub_klass: subklass
287 //
288 // Kills:
289 // r2, r5
290 void InterpreterMacroAssembler::gen_subtype_check(Register Rsub_klass,
291 Label& ok_is_subtype) {
292 assert(Rsub_klass != r0, "r0 holds superklass");
293 assert(Rsub_klass != r2, "r2 holds 2ndary super array length");
294 assert(Rsub_klass != r5, "r5 holds 2ndary super array scan ptr");
295
296 // Profile the not-null value's klass.
297 profile_typecheck(r2, Rsub_klass, r5); // blows r2, reloads r5
298
299 // Do the check.
300 check_klass_subtype(Rsub_klass, r0, r2, ok_is_subtype); // blows r2
301
302 // Profile the failure of the check.
303 profile_typecheck_failed(r2); // blows r2
304 }
305
306 // Java Expression Stack
307
308 void InterpreterMacroAssembler::pop_ptr(Register r) {
309 ldr(r, post(esp, wordSize));
310 }
311
312 void InterpreterMacroAssembler::pop_i(Register r) {
313 ldrw(r, post(esp, wordSize));
314 }
315
316 void InterpreterMacroAssembler::pop_l(Register r) {
317 ldr(r, post(esp, 2 * Interpreter::stackElementSize));
318 }
319
320 void InterpreterMacroAssembler::push_ptr(Register r) {
321 str(r, pre(esp, -wordSize));
322 }
323
324 void InterpreterMacroAssembler::push_i(Register r) {
325 str(r, pre(esp, -wordSize));
326 }
327
328 void InterpreterMacroAssembler::push_l(Register r) {
329 str(zr, pre(esp, -wordSize));
330 str(r, pre(esp, -wordsize));
331 }
332
333 void InterpreterMacroAssembler::pop_f(FloatRegister r) {
334 ldrs(r, post(esp, wordSize));
335 }
336
337 void InterpreterMacroAssembler::pop_d(FloatRegister r) {
338 ldrd(r, post(esp, 2 * Interpreter::stackElementSize));
339 }
340
341 void InterpreterMacroAssembler::push_f(FloatRegister r) {
342 strs(r, pre(esp, -wordSize));
343 }
344
345 void InterpreterMacroAssembler::push_d(FloatRegister r) {
346 strd(r, pre(esp, 2* -wordSize));
347 }
348
349 void InterpreterMacroAssembler::pop(TosState state) {
350 switch (state) {
351 case atos: pop_ptr(); break;
352 case btos:
353 case ztos:
354 case ctos:
355 case stos:
356 case itos: pop_i(); break;
357 case ltos: pop_l(); break;
358 case ftos: pop_f(); break;
359 case dtos: pop_d(); break;
360 case vtos: /* nothing to do */ break;
361 default: ShouldNotReachHere();
362 }
363 verify_oop(r0, state);
364 }
365
366 void InterpreterMacroAssembler::push(TosState state) {
367 verify_oop(r0, state);
368 switch (state) {
369 case atos: push_ptr(); break;
370 case btos:
371 case ztos:
372 case ctos:
373 case stos:
374 case itos: push_i(); break;
375 case ltos: push_l(); break;
376 case ftos: push_f(); break;
377 case dtos: push_d(); break;
378 case vtos: /* nothing to do */ break;
379 default : ShouldNotReachHere();
380 }
381 }
382
383 // Helpers for swap and dup
384 void InterpreterMacroAssembler::load_ptr(int n, Register val) {
385 ldr(val, Address(esp, Interpreter::expr_offset_in_bytes(n)));
386 }
387
388 void InterpreterMacroAssembler::store_ptr(int n, Register val) {
389 str(val, Address(esp, Interpreter::expr_offset_in_bytes(n)));
390 }
391
392
393 void InterpreterMacroAssembler::prepare_to_jump_from_interpreted() {
394 // set sender sp
395 mov(r13, sp);
396 // record last_sp
397 str(esp, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize));
398 }
399
400 // Jump to from_interpreted entry of a call unless single stepping is possible
401 // in this thread in which case we must call the i2i entry
402 void InterpreterMacroAssembler::jump_from_interpreted(Register method, Register temp) {
403 prepare_to_jump_from_interpreted();
404
405 if (JvmtiExport::can_post_interpreter_events()) {
406 Label run_compiled_code;
407 // JVMTI events, such as single-stepping, are implemented partly by avoiding running
408 // compiled code in threads for which the event is enabled. Check here for
409 // interp_only_mode if these events CAN be enabled.
410 // interp_only is an int, on little endian it is sufficient to test the byte only
411 // Is a cmpl faster?
412 ldr(rscratch1, Address(rthread, JavaThread::interp_only_mode_offset()));
413 cbz(rscratch1, run_compiled_code);
414 ldr(rscratch1, Address(method, Method::interpreter_entry_offset()));
415 br(rscratch1);
416 bind(run_compiled_code);
417 }
418
419 ldr(rscratch1, Address(method, Method::from_interpreted_offset()));
420 br(rscratch1);
421 }
422
423 // The following two routines provide a hook so that an implementation
424 // can schedule the dispatch in two parts. amd64 does not do this.
425 void InterpreterMacroAssembler::dispatch_prolog(TosState state, int step) {
426 }
427
428 void InterpreterMacroAssembler::dispatch_epilog(TosState state, int step) {
429 dispatch_next(state, step);
430 }
431
432 void InterpreterMacroAssembler::dispatch_base(TosState state,
433 address* table,
434 bool verifyoop) {
435 if (VerifyActivationFrameSize) {
436 Unimplemented();
437 }
438 if (verifyoop) {
439 verify_oop(r0, state);
440 }
441 if (table == Interpreter::dispatch_table(state)) {
442 addw(rscratch2, rscratch1, Interpreter::distance_from_dispatch_table(state));
443 ldr(rscratch2, Address(rdispatch, rscratch2, Address::uxtw(3)));
444 } else {
445 mov(rscratch2, (address)table);
446 ldr(rscratch2, Address(rscratch2, rscratch1, Address::uxtw(3)));
447 }
448 br(rscratch2);
449 }
450
451 void InterpreterMacroAssembler::dispatch_only(TosState state) {
452 dispatch_base(state, Interpreter::dispatch_table(state));
453 }
454
455 void InterpreterMacroAssembler::dispatch_only_normal(TosState state) {
456 dispatch_base(state, Interpreter::normal_table(state));
457 }
458
459 void InterpreterMacroAssembler::dispatch_only_noverify(TosState state) {
460 dispatch_base(state, Interpreter::normal_table(state), false);
461 }
462
463
464 void InterpreterMacroAssembler::dispatch_next(TosState state, int step) {
465 // load next bytecode
466 ldrb(rscratch1, Address(pre(rbcp, step)));
467 dispatch_base(state, Interpreter::dispatch_table(state));
468 }
469
470 void InterpreterMacroAssembler::dispatch_via(TosState state, address* table) {
471 // load current bytecode
472 ldrb(rscratch1, Address(rbcp, 0));
473 dispatch_base(state, table);
474 }
475
476 // remove activation
477 //
478 // Unlock the receiver if this is a synchronized method.
479 // Unlock any Java monitors from syncronized blocks.
480 // Remove the activation from the stack.
481 //
482 // If there are locked Java monitors
483 // If throw_monitor_exception
484 // throws IllegalMonitorStateException
485 // Else if install_monitor_exception
486 // installs IllegalMonitorStateException
487 // Else
488 // no error processing
489 void InterpreterMacroAssembler::remove_activation(
490 TosState state,
491 bool throw_monitor_exception,
492 bool install_monitor_exception,
493 bool notify_jvmdi) {
494 // Note: Registers r3 xmm0 may be in use for the
495 // result check if synchronized method
496 Label unlocked, unlock, no_unlock;
497
498 // get the value of _do_not_unlock_if_synchronized into r3
499 const Address do_not_unlock_if_synchronized(rthread,
500 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
501 ldrb(r3, do_not_unlock_if_synchronized);
502 strb(zr, do_not_unlock_if_synchronized); // reset the flag
503
504 // get method access flags
505 ldr(r1, Address(rfp, frame::interpreter_frame_method_offset * wordSize));
506 ldr(r2, Address(r1, Method::access_flags_offset()));
507 tbz(r2, exact_log2(JVM_ACC_SYNCHRONIZED), unlocked);
508
509 // Don't unlock anything if the _do_not_unlock_if_synchronized flag
510 // is set.
511 cbnz(r3, no_unlock);
512
513 // unlock monitor
514 push(state); // save result
515
516 // BasicObjectLock will be first in list, since this is a
517 // synchronized method. However, need to check that the object has
518 // not been unlocked by an explicit monitorexit bytecode.
519 const Address monitor(rfp, frame::interpreter_frame_initial_sp_offset *
520 wordSize - (int) sizeof(BasicObjectLock));
521 // We use c_rarg1 so that if we go slow path it will be the correct
522 // register for unlock_object to pass to VM directly
523 lea(c_rarg1, monitor); // address of first monitor
524
525 ldr(r0, Address(c_rarg1, BasicObjectLock::obj_offset_in_bytes()));
526 cbnz(r0, unlock);
527
528 pop(state);
529 if (throw_monitor_exception) {
530 // Entry already unlocked, need to throw exception
531 call_VM(noreg, CAST_FROM_FN_PTR(address,
532 InterpreterRuntime::throw_illegal_monitor_state_exception));
533 should_not_reach_here();
534 } else {
535 // Monitor already unlocked during a stack unroll. If requested,
536 // install an illegal_monitor_state_exception. Continue with
537 // stack unrolling.
538 if (install_monitor_exception) {
539 call_VM(noreg, CAST_FROM_FN_PTR(address,
540 InterpreterRuntime::new_illegal_monitor_state_exception));
541 }
542 b(unlocked);
543 }
544
545 bind(unlock);
546 unlock_object(c_rarg1);
547 pop(state);
548
549 // Check that for block-structured locking (i.e., that all locked
550 // objects has been unlocked)
551 bind(unlocked);
552
553 // r0: Might contain return value
554
555 // Check that all monitors are unlocked
556 {
557 Label loop, exception, entry, restart;
558 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
559 const Address monitor_block_top(
560 rfp, frame::interpreter_frame_monitor_block_top_offset * wordSize);
561 const Address monitor_block_bot(
562 rfp, frame::interpreter_frame_initial_sp_offset * wordSize);
563
564 bind(restart);
565 // We use c_rarg1 so that if we go slow path it will be the correct
566 // register for unlock_object to pass to VM directly
567 ldr(c_rarg1, monitor_block_top); // points to current entry, starting
568 // with top-most entry
569 lea(r19, monitor_block_bot); // points to word before bottom of
570 // monitor block
571 b(entry);
572
573 // Entry already locked, need to throw exception
574 bind(exception);
575
576 if (throw_monitor_exception) {
577 // Throw exception
578 MacroAssembler::call_VM(noreg,
579 CAST_FROM_FN_PTR(address, InterpreterRuntime::
580 throw_illegal_monitor_state_exception));
581 should_not_reach_here();
582 } else {
583 // Stack unrolling. Unlock object and install illegal_monitor_exception.
584 // Unlock does not block, so don't have to worry about the frame.
585 // We don't have to preserve c_rarg1 since we are going to throw an exception.
586
587 push(state);
588 unlock_object(c_rarg1);
589 pop(state);
590
591 if (install_monitor_exception) {
592 call_VM(noreg, CAST_FROM_FN_PTR(address,
593 InterpreterRuntime::
594 new_illegal_monitor_state_exception));
595 }
596
597 b(restart);
598 }
599
600 bind(loop);
601 // check if current entry is used
602 ldr(rscratch1, Address(c_rarg1, BasicObjectLock::obj_offset_in_bytes()));
603 cbnz(rscratch1, exception);
604
605 add(c_rarg1, c_rarg1, entry_size); // otherwise advance to next entry
606 bind(entry);
607 cmp(c_rarg1, r19); // check if bottom reached
608 br(Assembler::NE, loop); // if not at bottom then check this entry
609 }
610
611 bind(no_unlock);
612
613 // jvmti support
614 if (notify_jvmdi) {
615 notify_method_exit(state, NotifyJVMTI); // preserve TOSCA
616 } else {
617 notify_method_exit(state, SkipNotifyJVMTI); // preserve TOSCA
618 }
619
620 // remove activation
621 // get sender esp
622 ldr(esp,
623 Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize));
624 // remove frame anchor
625 leave();
626 // If we're returning to interpreted code we will shortly be
627 // adjusting SP to allow some space for ESP. If we're returning to
628 // compiled code the saved sender SP was saved in sender_sp, so this
629 // restores it.
630 andr(sp, esp, -16);
631 }
632
633 // Lock object
634 //
635 // Args:
636 // c_rarg1: BasicObjectLock to be used for locking
637 //
638 // Kills:
639 // r0
640 // c_rarg0, c_rarg1, c_rarg2, c_rarg3, .. (param regs)
641 // rscratch1, rscratch2 (scratch regs)
642 void InterpreterMacroAssembler::lock_object(Register lock_reg)
643 {
644 assert(lock_reg == c_rarg1, "The argument is only for looks. It must be c_rarg1");
645 if (UseHeavyMonitors) {
646 call_VM(noreg,
647 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter),
648 lock_reg);
649 } else {
650 Label done;
651
652 const Register swap_reg = r0;
653 const Register tmp = c_rarg2;
654 const Register obj_reg = c_rarg3; // Will contain the oop
655
656 const int obj_offset = BasicObjectLock::obj_offset_in_bytes();
657 const int lock_offset = BasicObjectLock::lock_offset_in_bytes ();
658 const int mark_offset = lock_offset +
659 BasicLock::displaced_header_offset_in_bytes();
660
661 Label slow_case;
662
663 // Load object pointer into obj_reg %c_rarg3
664 ldr(obj_reg, Address(lock_reg, obj_offset));
665
666 if (UseBiasedLocking) {
667 biased_locking_enter(lock_reg, obj_reg, swap_reg, tmp, false, done, &slow_case);
668 }
669
670 // Load (object->mark() | 1) into swap_reg
671 ldr(rscratch1, Address(obj_reg, 0));
672 orr(swap_reg, rscratch1, 1);
673
674 // Save (object->mark() | 1) into BasicLock's displaced header
675 str(swap_reg, Address(lock_reg, mark_offset));
676
677 assert(lock_offset == 0,
678 "displached header must be first word in BasicObjectLock");
679
680 Label fail;
681 if (PrintBiasedLockingStatistics) {
682 Label fast;
683 cmpxchgptr(swap_reg, lock_reg, obj_reg, rscratch1, fast, &fail);
684 bind(fast);
685 atomic_incw(Address((address)BiasedLocking::fast_path_entry_count_addr()),
686 rscratch2, rscratch1, tmp);
687 b(done);
688 bind(fail);
689 } else {
690 cmpxchgptr(swap_reg, lock_reg, obj_reg, rscratch1, done, /*fallthrough*/NULL);
691 }
692
693 // Test if the oopMark is an obvious stack pointer, i.e.,
694 // 1) (mark & 7) == 0, and
695 // 2) rsp <= mark < mark + os::pagesize()
696 //
697 // These 3 tests can be done by evaluating the following
698 // expression: ((mark - rsp) & (7 - os::vm_page_size())),
699 // assuming both stack pointer and pagesize have their
700 // least significant 3 bits clear.
701 // NOTE: the oopMark is in swap_reg %r0 as the result of cmpxchg
702 // NOTE2: aarch64 does not like to subtract sp from rn so take a
703 // copy
704 mov(rscratch1, sp);
705 sub(swap_reg, swap_reg, rscratch1);
706 ands(swap_reg, swap_reg, (unsigned long)(7 - os::vm_page_size()));
707
708 // Save the test result, for recursive case, the result is zero
709 str(swap_reg, Address(lock_reg, mark_offset));
710
711 if (PrintBiasedLockingStatistics) {
712 br(Assembler::NE, slow_case);
713 atomic_incw(Address((address)BiasedLocking::fast_path_entry_count_addr()),
714 rscratch2, rscratch1, tmp);
715 }
716 br(Assembler::EQ, done);
717
718 bind(slow_case);
719
720 // Call the runtime routine for slow case
721 call_VM(noreg,
722 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter),
723 lock_reg);
724
725 bind(done);
726 }
727 }
728
729
730 // Unlocks an object. Used in monitorexit bytecode and
731 // remove_activation. Throws an IllegalMonitorException if object is
732 // not locked by current thread.
733 //
734 // Args:
735 // c_rarg1: BasicObjectLock for lock
736 //
737 // Kills:
738 // r0
739 // c_rarg0, c_rarg1, c_rarg2, c_rarg3, ... (param regs)
740 // rscratch1, rscratch2 (scratch regs)
741 void InterpreterMacroAssembler::unlock_object(Register lock_reg)
742 {
743 assert(lock_reg == c_rarg1, "The argument is only for looks. It must be rarg1");
744
745 if (UseHeavyMonitors) {
746 call_VM(noreg,
747 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorexit),
748 lock_reg);
749 } else {
750 Label done;
751
752 const Register swap_reg = r0;
753 const Register header_reg = c_rarg2; // Will contain the old oopMark
754 const Register obj_reg = c_rarg3; // Will contain the oop
755
756 save_bcp(); // Save in case of exception
757
758 // Convert from BasicObjectLock structure to object and BasicLock
759 // structure Store the BasicLock address into %r0
760 lea(swap_reg, Address(lock_reg, BasicObjectLock::lock_offset_in_bytes()));
761
762 // Load oop into obj_reg(%c_rarg3)
763 ldr(obj_reg, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes()));
764
765 // Free entry
766 str(zr, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes()));
767
768 if (UseBiasedLocking) {
769 biased_locking_exit(obj_reg, header_reg, done);
770 }
771
772 // Load the old header from BasicLock structure
773 ldr(header_reg, Address(swap_reg,
774 BasicLock::displaced_header_offset_in_bytes()));
775
776 // Test for recursion
777 cbz(header_reg, done);
778
779 // Atomic swap back the old header
780 cmpxchgptr(swap_reg, header_reg, obj_reg, rscratch1, done, /*fallthrough*/NULL);
781
782 // Call the runtime routine for slow case.
783 str(obj_reg, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes())); // restore obj
784 call_VM(noreg,
785 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorexit),
786 lock_reg);
787
788 bind(done);
789
790 restore_bcp();
791 }
792 }
793
794 void InterpreterMacroAssembler::test_method_data_pointer(Register mdp,
795 Label& zero_continue) {
796 assert(ProfileInterpreter, "must be profiling interpreter");
797 ldr(mdp, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize));
798 cbz(mdp, zero_continue);
799 }
800
801 // Set the method data pointer for the current bcp.
802 void InterpreterMacroAssembler::set_method_data_pointer_for_bcp() {
803 assert(ProfileInterpreter, "must be profiling interpreter");
804 Label set_mdp;
805 stp(r0, r1, Address(pre(sp, -2 * wordSize)));
806
807 // Test MDO to avoid the call if it is NULL.
808 ldr(r0, Address(rmethod, in_bytes(Method::method_data_offset())));
809 cbz(r0, set_mdp);
810 call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::bcp_to_di), rmethod, rbcp);
811 // r0: mdi
812 // mdo is guaranteed to be non-zero here, we checked for it before the call.
813 ldr(r1, Address(rmethod, in_bytes(Method::method_data_offset())));
814 lea(r1, Address(r1, in_bytes(MethodData::data_offset())));
815 add(r0, r1, r0);
816 str(r0, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize));
817 bind(set_mdp);
818 ldp(r0, r1, Address(post(sp, 2 * wordSize)));
819 }
820
821 void InterpreterMacroAssembler::verify_method_data_pointer() {
822 assert(ProfileInterpreter, "must be profiling interpreter");
823 #ifdef ASSERT
824 Label verify_continue;
825 stp(r0, r1, Address(pre(sp, -2 * wordSize)));
826 stp(r2, r3, Address(pre(sp, -2 * wordSize)));
827 test_method_data_pointer(r3, verify_continue); // If mdp is zero, continue
828 get_method(r1);
829
830 // If the mdp is valid, it will point to a DataLayout header which is
831 // consistent with the bcp. The converse is highly probable also.
832 ldrsh(r2, Address(r3, in_bytes(DataLayout::bci_offset())));
833 ldr(rscratch1, Address(r1, Method::const_offset()));
834 add(r2, r2, rscratch1, Assembler::LSL);
835 lea(r2, Address(r2, ConstMethod::codes_offset()));
836 cmp(r2, rbcp);
837 br(Assembler::EQ, verify_continue);
838 // r1: method
839 // rbcp: bcp // rbcp == 22
840 // r3: mdp
841 call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::verify_mdp),
842 r1, rbcp, r3);
843 bind(verify_continue);
844 ldp(r2, r3, Address(post(sp, 2 * wordSize)));
845 ldp(r0, r1, Address(post(sp, 2 * wordSize)));
846 #endif // ASSERT
847 }
848
849
850 void InterpreterMacroAssembler::set_mdp_data_at(Register mdp_in,
851 int constant,
852 Register value) {
853 assert(ProfileInterpreter, "must be profiling interpreter");
854 Address data(mdp_in, constant);
855 str(value, data);
856 }
857
858
859 void InterpreterMacroAssembler::increment_mdp_data_at(Register mdp_in,
860 int constant,
861 bool decrement) {
862 increment_mdp_data_at(mdp_in, noreg, constant, decrement);
863 }
864
865 void InterpreterMacroAssembler::increment_mdp_data_at(Register mdp_in,
866 Register reg,
867 int constant,
868 bool decrement) {
869 assert(ProfileInterpreter, "must be profiling interpreter");
870 // %%% this does 64bit counters at best it is wasting space
871 // at worst it is a rare bug when counters overflow
872
873 assert_different_registers(rscratch2, rscratch1, mdp_in, reg);
874
875 Address addr1(mdp_in, constant);
876 Address addr2(rscratch2, reg, Address::lsl(0));
877 Address &addr = addr1;
878 if (reg != noreg) {
879 lea(rscratch2, addr1);
880 addr = addr2;
881 }
882
883 if (decrement) {
884 // Decrement the register. Set condition codes.
885 // Intel does this
886 // addptr(data, (int32_t) -DataLayout::counter_increment);
887 // If the decrement causes the counter to overflow, stay negative
888 // Label L;
889 // jcc(Assembler::negative, L);
890 // addptr(data, (int32_t) DataLayout::counter_increment);
891 // so we do this
892 ldr(rscratch1, addr);
893 subs(rscratch1, rscratch1, (unsigned)DataLayout::counter_increment);
894 Label L;
895 br(Assembler::LO, L); // skip store if counter underflow
896 str(rscratch1, addr);
897 bind(L);
898 } else {
899 assert(DataLayout::counter_increment == 1,
900 "flow-free idiom only works with 1");
901 // Intel does this
902 // Increment the register. Set carry flag.
903 // addptr(data, DataLayout::counter_increment);
904 // If the increment causes the counter to overflow, pull back by 1.
905 // sbbptr(data, (int32_t)0);
906 // so we do this
907 ldr(rscratch1, addr);
908 adds(rscratch1, rscratch1, DataLayout::counter_increment);
909 Label L;
910 br(Assembler::CS, L); // skip store if counter overflow
911 str(rscratch1, addr);
912 bind(L);
913 }
914 }
915
916 void InterpreterMacroAssembler::set_mdp_flag_at(Register mdp_in,
917 int flag_byte_constant) {
918 assert(ProfileInterpreter, "must be profiling interpreter");
919 int header_offset = in_bytes(DataLayout::header_offset());
920 int header_bits = DataLayout::flag_mask_to_header_mask(flag_byte_constant);
921 // Set the flag
922 ldr(rscratch1, Address(mdp_in, header_offset));
923 orr(rscratch1, rscratch1, header_bits);
924 str(rscratch1, Address(mdp_in, header_offset));
925 }
926
927
928 void InterpreterMacroAssembler::test_mdp_data_at(Register mdp_in,
929 int offset,
930 Register value,
931 Register test_value_out,
932 Label& not_equal_continue) {
933 assert(ProfileInterpreter, "must be profiling interpreter");
934 if (test_value_out == noreg) {
935 ldr(rscratch1, Address(mdp_in, offset));
936 cmp(value, rscratch1);
937 } else {
938 // Put the test value into a register, so caller can use it:
939 ldr(test_value_out, Address(mdp_in, offset));
940 cmp(value, test_value_out);
941 }
942 br(Assembler::NE, not_equal_continue);
943 }
944
945
946 void InterpreterMacroAssembler::update_mdp_by_offset(Register mdp_in,
947 int offset_of_disp) {
948 assert(ProfileInterpreter, "must be profiling interpreter");
949 ldr(rscratch1, Address(mdp_in, offset_of_disp));
950 add(mdp_in, mdp_in, rscratch1, LSL);
951 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize));
952 }
953
954
955 void InterpreterMacroAssembler::update_mdp_by_offset(Register mdp_in,
956 Register reg,
957 int offset_of_disp) {
958 assert(ProfileInterpreter, "must be profiling interpreter");
959 lea(rscratch1, Address(mdp_in, offset_of_disp));
960 ldr(rscratch1, Address(rscratch1, reg, Address::lsl(0)));
961 add(mdp_in, mdp_in, rscratch1, LSL);
962 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize));
963 }
964
965
966 void InterpreterMacroAssembler::update_mdp_by_constant(Register mdp_in,
967 int constant) {
968 assert(ProfileInterpreter, "must be profiling interpreter");
969 add(mdp_in, mdp_in, (unsigned)constant);
970 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize));
971 }
972
973
974 void InterpreterMacroAssembler::update_mdp_for_ret(Register return_bci) {
975 assert(ProfileInterpreter, "must be profiling interpreter");
976 // save/restore across call_VM
977 stp(zr, return_bci, Address(pre(sp, -2 * wordSize)));
978 call_VM(noreg,
979 CAST_FROM_FN_PTR(address, InterpreterRuntime::update_mdp_for_ret),
980 return_bci);
981 ldp(zr, return_bci, Address(post(sp, 2 * wordSize)));
982 }
983
984
985 void InterpreterMacroAssembler::profile_taken_branch(Register mdp,
986 Register bumped_count) {
987 if (ProfileInterpreter) {
988 Label profile_continue;
989
990 // If no method data exists, go to profile_continue.
991 // Otherwise, assign to mdp
992 test_method_data_pointer(mdp, profile_continue);
993
994 // We are taking a branch. Increment the taken count.
995 // We inline increment_mdp_data_at to return bumped_count in a register
996 //increment_mdp_data_at(mdp, in_bytes(JumpData::taken_offset()));
997 Address data(mdp, in_bytes(JumpData::taken_offset()));
998 ldr(bumped_count, data);
999 assert(DataLayout::counter_increment == 1,
1000 "flow-free idiom only works with 1");
1001 // Intel does this to catch overflow
1002 // addptr(bumped_count, DataLayout::counter_increment);
1003 // sbbptr(bumped_count, 0);
1004 // so we do this
1005 adds(bumped_count, bumped_count, DataLayout::counter_increment);
1006 Label L;
1007 br(Assembler::CS, L); // skip store if counter overflow
1008 str(bumped_count, data);
1009 bind(L);
1010 // The method data pointer needs to be updated to reflect the new target.
1011 update_mdp_by_offset(mdp, in_bytes(JumpData::displacement_offset()));
1012 bind(profile_continue);
1013 }
1014 }
1015
1016
1017 void InterpreterMacroAssembler::profile_not_taken_branch(Register mdp) {
1018 if (ProfileInterpreter) {
1019 Label profile_continue;
1020
1021 // If no method data exists, go to profile_continue.
1022 test_method_data_pointer(mdp, profile_continue);
1023
1024 // We are taking a branch. Increment the not taken count.
1025 increment_mdp_data_at(mdp, in_bytes(BranchData::not_taken_offset()));
1026
1027 // The method data pointer needs to be updated to correspond to
1028 // the next bytecode
1029 update_mdp_by_constant(mdp, in_bytes(BranchData::branch_data_size()));
1030 bind(profile_continue);
1031 }
1032 }
1033
1034
1035 void InterpreterMacroAssembler::profile_call(Register mdp) {
1036 if (ProfileInterpreter) {
1037 Label profile_continue;
1038
1039 // If no method data exists, go to profile_continue.
1040 test_method_data_pointer(mdp, profile_continue);
1041
1042 // We are making a call. Increment the count.
1043 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
1044
1045 // The method data pointer needs to be updated to reflect the new target.
1046 update_mdp_by_constant(mdp, in_bytes(CounterData::counter_data_size()));
1047 bind(profile_continue);
1048 }
1049 }
1050
1051 void InterpreterMacroAssembler::profile_final_call(Register mdp) {
1052 if (ProfileInterpreter) {
1053 Label profile_continue;
1054
1055 // If no method data exists, go to profile_continue.
1056 test_method_data_pointer(mdp, profile_continue);
1057
1058 // We are making a call. Increment the count.
1059 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
1060
1061 // The method data pointer needs to be updated to reflect the new target.
1062 update_mdp_by_constant(mdp,
1063 in_bytes(VirtualCallData::
1064 virtual_call_data_size()));
1065 bind(profile_continue);
1066 }
1067 }
1068
1069
1070 void InterpreterMacroAssembler::profile_virtual_call(Register receiver,
1071 Register mdp,
1072 Register reg2,
1073 bool receiver_can_be_null) {
1074 if (ProfileInterpreter) {
1075 Label profile_continue;
1076
1077 // If no method data exists, go to profile_continue.
1078 test_method_data_pointer(mdp, profile_continue);
1079
1080 Label skip_receiver_profile;
1081 if (receiver_can_be_null) {
1082 Label not_null;
1083 // We are making a call. Increment the count for null receiver.
1084 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
1085 b(skip_receiver_profile);
1086 bind(not_null);
1087 }
1088
1089 // Record the receiver type.
1090 record_klass_in_profile(receiver, mdp, reg2, true);
1091 bind(skip_receiver_profile);
1092
1093 // The method data pointer needs to be updated to reflect the new target.
1094 #if INCLUDE_JVMCI
1095 if (MethodProfileWidth == 0) {
1096 update_mdp_by_constant(mdp, in_bytes(VirtualCallData::virtual_call_data_size()));
1097 }
1098 #else // INCLUDE_JVMCI
1099 update_mdp_by_constant(mdp,
1100 in_bytes(VirtualCallData::
1101 virtual_call_data_size()));
1102 #endif // INCLUDE_JVMCI
1103 bind(profile_continue);
1104 }
1105 }
1106
1107 #if INCLUDE_JVMCI
1108 void InterpreterMacroAssembler::profile_called_method(Register method, Register mdp, Register reg2) {
1109 assert_different_registers(method, mdp, reg2);
1110 if (ProfileInterpreter && MethodProfileWidth > 0) {
1111 Label profile_continue;
1112
1113 // If no method data exists, go to profile_continue.
1114 test_method_data_pointer(mdp, profile_continue);
1115
1116 Label done;
1117 record_item_in_profile_helper(method, mdp, reg2, 0, done, MethodProfileWidth,
1118 &VirtualCallData::method_offset, &VirtualCallData::method_count_offset, in_bytes(VirtualCallData::nonprofiled_receiver_count_offset()));
1119 bind(done);
1120
1121 update_mdp_by_constant(mdp, in_bytes(VirtualCallData::virtual_call_data_size()));
1122 bind(profile_continue);
1123 }
1124 }
1125 #endif // INCLUDE_JVMCI
1126
1127 // This routine creates a state machine for updating the multi-row
1128 // type profile at a virtual call site (or other type-sensitive bytecode).
1129 // The machine visits each row (of receiver/count) until the receiver type
1130 // is found, or until it runs out of rows. At the same time, it remembers
1131 // the location of the first empty row. (An empty row records null for its
1132 // receiver, and can be allocated for a newly-observed receiver type.)
1133 // Because there are two degrees of freedom in the state, a simple linear
1134 // search will not work; it must be a decision tree. Hence this helper
1135 // function is recursive, to generate the required tree structured code.
1136 // It's the interpreter, so we are trading off code space for speed.
1137 // See below for example code.
1138 void InterpreterMacroAssembler::record_klass_in_profile_helper(
1139 Register receiver, Register mdp,
1140 Register reg2, int start_row,
1141 Label& done, bool is_virtual_call) {
1142 if (TypeProfileWidth == 0) {
1143 if (is_virtual_call) {
1144 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
1145 }
1146 #if INCLUDE_JVMCI
1147 else if (EnableJVMCI) {
1148 increment_mdp_data_at(mdp, in_bytes(ReceiverTypeData::nonprofiled_receiver_count_offset()));
1149 }
1150 #endif // INCLUDE_JVMCI
1151 } else {
1152 int non_profiled_offset = -1;
1153 if (is_virtual_call) {
1154 non_profiled_offset = in_bytes(CounterData::count_offset());
1155 }
1156 #if INCLUDE_JVMCI
1157 else if (EnableJVMCI) {
1158 non_profiled_offset = in_bytes(ReceiverTypeData::nonprofiled_receiver_count_offset());
1159 }
1160 #endif // INCLUDE_JVMCI
1161
1162 record_item_in_profile_helper(receiver, mdp, reg2, 0, done, TypeProfileWidth,
1163 &VirtualCallData::receiver_offset, &VirtualCallData::receiver_count_offset, non_profiled_offset);
1164 }
1165 }
1166
1167 void InterpreterMacroAssembler::record_item_in_profile_helper(Register item, Register mdp,
1168 Register reg2, int start_row, Label& done, int total_rows,
1169 OffsetFunction item_offset_fn, OffsetFunction item_count_offset_fn,
1170 int non_profiled_offset) {
1171 int last_row = total_rows - 1;
1172 assert(start_row <= last_row, "must be work left to do");
1173 // Test this row for both the item and for null.
1174 // Take any of three different outcomes:
1175 // 1. found item => increment count and goto done
1176 // 2. found null => keep looking for case 1, maybe allocate this cell
1177 // 3. found something else => keep looking for cases 1 and 2
1178 // Case 3 is handled by a recursive call.
1179 for (int row = start_row; row <= last_row; row++) {
1180 Label next_test;
1181 bool test_for_null_also = (row == start_row);
1182
1183 // See if the item is item[n].
1184 int item_offset = in_bytes(item_offset_fn(row));
1185 test_mdp_data_at(mdp, item_offset, item,
1186 (test_for_null_also ? reg2 : noreg),
1187 next_test);
1188 // (Reg2 now contains the item from the CallData.)
1189
1190 // The item is item[n]. Increment count[n].
1191 int count_offset = in_bytes(item_count_offset_fn(row));
1192 increment_mdp_data_at(mdp, count_offset);
1193 b(done);
1194 bind(next_test);
1195
1196 if (test_for_null_also) {
1197 Label found_null;
1198 // Failed the equality check on item[n]... Test for null.
1199 if (start_row == last_row) {
1200 // The only thing left to do is handle the null case.
1201 if (non_profiled_offset >= 0) {
1202 cbz(reg2, found_null);
1203 // Item did not match any saved item and there is no empty row for it.
1204 // Increment total counter to indicate polymorphic case.
1205 increment_mdp_data_at(mdp, non_profiled_offset);
1206 b(done);
1207 bind(found_null);
1208 } else {
1209 cbnz(reg2, done);
1210 }
1211 break;
1212 }
1213 // Since null is rare, make it be the branch-taken case.
1214 cbz(reg2, found_null);
1215
1216 // Put all the "Case 3" tests here.
1217 record_item_in_profile_helper(item, mdp, reg2, start_row + 1, done, total_rows,
1218 item_offset_fn, item_count_offset_fn, non_profiled_offset);
1219
1220 // Found a null. Keep searching for a matching item,
1221 // but remember that this is an empty (unused) slot.
1222 bind(found_null);
1223 }
1224 }
1225
1226 // In the fall-through case, we found no matching item, but we
1227 // observed the item[start_row] is NULL.
1228
1229 // Fill in the item field and increment the count.
1230 int item_offset = in_bytes(item_offset_fn(start_row));
1231 set_mdp_data_at(mdp, item_offset, item);
1232 int count_offset = in_bytes(item_count_offset_fn(start_row));
1233 mov(reg2, DataLayout::counter_increment);
1234 set_mdp_data_at(mdp, count_offset, reg2);
1235 if (start_row > 0) {
1236 b(done);
1237 }
1238 }
1239
1240 // Example state machine code for three profile rows:
1241 // // main copy of decision tree, rooted at row[1]
1242 // if (row[0].rec == rec) { row[0].incr(); goto done; }
1243 // if (row[0].rec != NULL) {
1244 // // inner copy of decision tree, rooted at row[1]
1245 // if (row[1].rec == rec) { row[1].incr(); goto done; }
1246 // if (row[1].rec != NULL) {
1247 // // degenerate decision tree, rooted at row[2]
1248 // if (row[2].rec == rec) { row[2].incr(); goto done; }
1249 // if (row[2].rec != NULL) { count.incr(); goto done; } // overflow
1250 // row[2].init(rec); goto done;
1251 // } else {
1252 // // remember row[1] is empty
1253 // if (row[2].rec == rec) { row[2].incr(); goto done; }
1254 // row[1].init(rec); goto done;
1255 // }
1256 // } else {
1257 // // remember row[0] is empty
1258 // if (row[1].rec == rec) { row[1].incr(); goto done; }
1259 // if (row[2].rec == rec) { row[2].incr(); goto done; }
1260 // row[0].init(rec); goto done;
1261 // }
1262 // done:
1263
1264 void InterpreterMacroAssembler::record_klass_in_profile(Register receiver,
1265 Register mdp, Register reg2,
1266 bool is_virtual_call) {
1267 assert(ProfileInterpreter, "must be profiling");
1268 Label done;
1269
1270 record_klass_in_profile_helper(receiver, mdp, reg2, 0, done, is_virtual_call);
1271
1272 bind (done);
1273 }
1274
1275 void InterpreterMacroAssembler::profile_ret(Register return_bci,
1276 Register mdp) {
1277 if (ProfileInterpreter) {
1278 Label profile_continue;
1279 uint row;
1280
1281 // If no method data exists, go to profile_continue.
1282 test_method_data_pointer(mdp, profile_continue);
1283
1284 // Update the total ret count.
1285 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
1286
1287 for (row = 0; row < RetData::row_limit(); row++) {
1288 Label next_test;
1289
1290 // See if return_bci is equal to bci[n]:
1291 test_mdp_data_at(mdp,
1292 in_bytes(RetData::bci_offset(row)),
1293 return_bci, noreg,
1294 next_test);
1295
1296 // return_bci is equal to bci[n]. Increment the count.
1297 increment_mdp_data_at(mdp, in_bytes(RetData::bci_count_offset(row)));
1298
1299 // The method data pointer needs to be updated to reflect the new target.
1300 update_mdp_by_offset(mdp,
1301 in_bytes(RetData::bci_displacement_offset(row)));
1302 b(profile_continue);
1303 bind(next_test);
1304 }
1305
1306 update_mdp_for_ret(return_bci);
1307
1308 bind(profile_continue);
1309 }
1310 }
1311
1312 void InterpreterMacroAssembler::profile_null_seen(Register mdp) {
1313 if (ProfileInterpreter) {
1314 Label profile_continue;
1315
1316 // If no method data exists, go to profile_continue.
1317 test_method_data_pointer(mdp, profile_continue);
1318
1319 set_mdp_flag_at(mdp, BitData::null_seen_byte_constant());
1320
1321 // The method data pointer needs to be updated.
1322 int mdp_delta = in_bytes(BitData::bit_data_size());
1323 if (TypeProfileCasts) {
1324 mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size());
1325 }
1326 update_mdp_by_constant(mdp, mdp_delta);
1327
1328 bind(profile_continue);
1329 }
1330 }
1331
1332 void InterpreterMacroAssembler::profile_typecheck_failed(Register mdp) {
1333 if (ProfileInterpreter && TypeProfileCasts) {
1334 Label profile_continue;
1335
1336 // If no method data exists, go to profile_continue.
1337 test_method_data_pointer(mdp, profile_continue);
1338
1339 int count_offset = in_bytes(CounterData::count_offset());
1340 // Back up the address, since we have already bumped the mdp.
1341 count_offset -= in_bytes(VirtualCallData::virtual_call_data_size());
1342
1343 // *Decrement* the counter. We expect to see zero or small negatives.
1344 increment_mdp_data_at(mdp, count_offset, true);
1345
1346 bind (profile_continue);
1347 }
1348 }
1349
1350 void InterpreterMacroAssembler::profile_typecheck(Register mdp, Register klass, Register reg2) {
1351 if (ProfileInterpreter) {
1352 Label profile_continue;
1353
1354 // If no method data exists, go to profile_continue.
1355 test_method_data_pointer(mdp, profile_continue);
1356
1357 // The method data pointer needs to be updated.
1358 int mdp_delta = in_bytes(BitData::bit_data_size());
1359 if (TypeProfileCasts) {
1360 mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size());
1361
1362 // Record the object type.
1363 record_klass_in_profile(klass, mdp, reg2, false);
1364 }
1365 update_mdp_by_constant(mdp, mdp_delta);
1366
1367 bind(profile_continue);
1368 }
1369 }
1370
1371 void InterpreterMacroAssembler::profile_switch_default(Register mdp) {
1372 if (ProfileInterpreter) {
1373 Label profile_continue;
1374
1375 // If no method data exists, go to profile_continue.
1376 test_method_data_pointer(mdp, profile_continue);
1377
1378 // Update the default case count
1379 increment_mdp_data_at(mdp,
1380 in_bytes(MultiBranchData::default_count_offset()));
1381
1382 // The method data pointer needs to be updated.
1383 update_mdp_by_offset(mdp,
1384 in_bytes(MultiBranchData::
1385 default_displacement_offset()));
1386
1387 bind(profile_continue);
1388 }
1389 }
1390
1391 void InterpreterMacroAssembler::profile_switch_case(Register index,
1392 Register mdp,
1393 Register reg2) {
1394 if (ProfileInterpreter) {
1395 Label profile_continue;
1396
1397 // If no method data exists, go to profile_continue.
1398 test_method_data_pointer(mdp, profile_continue);
1399
1400 // Build the base (index * per_case_size_in_bytes()) +
1401 // case_array_offset_in_bytes()
1402 movw(reg2, in_bytes(MultiBranchData::per_case_size()));
1403 movw(rscratch1, in_bytes(MultiBranchData::case_array_offset()));
1404 Assembler::maddw(index, index, reg2, rscratch1);
1405
1406 // Update the case count
1407 increment_mdp_data_at(mdp,
1408 index,
1409 in_bytes(MultiBranchData::relative_count_offset()));
1410
1411 // The method data pointer needs to be updated.
1412 update_mdp_by_offset(mdp,
1413 index,
1414 in_bytes(MultiBranchData::
1415 relative_displacement_offset()));
1416
1417 bind(profile_continue);
1418 }
1419 }
1420
1421 void InterpreterMacroAssembler::verify_oop(Register reg, TosState state) {
1422 if (state == atos) {
1423 MacroAssembler::verify_oop(reg);
1424 }
1425 }
1426
1427 void InterpreterMacroAssembler::verify_FPU(int stack_depth, TosState state) { ; }
1428
1429
1430 void InterpreterMacroAssembler::notify_method_entry() {
1431 // Whenever JVMTI is interp_only_mode, method entry/exit events are sent to
1432 // track stack depth. If it is possible to enter interp_only_mode we add
1433 // the code to check if the event should be sent.
1434 if (JvmtiExport::can_post_interpreter_events()) {
1435 Label L;
1436 ldrw(r3, Address(rthread, JavaThread::interp_only_mode_offset()));
1437 cbzw(r3, L);
1438 call_VM(noreg, CAST_FROM_FN_PTR(address,
1439 InterpreterRuntime::post_method_entry));
1440 bind(L);
1441 }
1442
1443 {
1444 SkipIfEqual skip(this, &DTraceMethodProbes, false);
1445 get_method(c_rarg1);
1446 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
1447 rthread, c_rarg1);
1448 }
1449
1450 // RedefineClasses() tracing support for obsolete method entry
1451 if (log_is_enabled(Trace, redefine, class, obsolete)) {
1452 get_method(c_rarg1);
1453 call_VM_leaf(
1454 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
1455 rthread, c_rarg1);
1456 }
1457
1458 }
1459
1460
1461 void InterpreterMacroAssembler::notify_method_exit(
1462 TosState state, NotifyMethodExitMode mode) {
1463 // Whenever JVMTI is interp_only_mode, method entry/exit events are sent to
1464 // track stack depth. If it is possible to enter interp_only_mode we add
1465 // the code to check if the event should be sent.
1466 if (mode == NotifyJVMTI && JvmtiExport::can_post_interpreter_events()) {
1467 Label L;
1468 // Note: frame::interpreter_frame_result has a dependency on how the
1469 // method result is saved across the call to post_method_exit. If this
1470 // is changed then the interpreter_frame_result implementation will
1471 // need to be updated too.
1472
1473 // template interpreter will leave the result on the top of the stack.
1474 push(state);
1475 ldrw(r3, Address(rthread, JavaThread::interp_only_mode_offset()));
1476 cbz(r3, L);
1477 call_VM(noreg,
1478 CAST_FROM_FN_PTR(address, InterpreterRuntime::post_method_exit));
1479 bind(L);
1480 pop(state);
1481 }
1482
1483 {
1484 SkipIfEqual skip(this, &DTraceMethodProbes, false);
1485 push(state);
1486 get_method(c_rarg1);
1487 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
1488 rthread, c_rarg1);
1489 pop(state);
1490 }
1491 }
1492
1493
1494 // Jump if ((*counter_addr += increment) & mask) satisfies the condition.
1495 void InterpreterMacroAssembler::increment_mask_and_jump(Address counter_addr,
1496 int increment, Address mask,
1497 Register scratch, Register scratch2,
1498 bool preloaded, Condition cond,
1499 Label* where) {
1500 if (!preloaded) {
1501 ldrw(scratch, counter_addr);
1502 }
1503 add(scratch, scratch, increment);
1504 strw(scratch, counter_addr);
1505 ldrw(scratch2, mask);
1506 ands(scratch, scratch, scratch2);
1507 br(cond, *where);
1508 }
1509
1510 void InterpreterMacroAssembler::call_VM_leaf_base(address entry_point,
1511 int number_of_arguments) {
1512 // interpreter specific
1513 //
1514 // Note: No need to save/restore rbcp & rlocals pointer since these
1515 // are callee saved registers and no blocking/ GC can happen
1516 // in leaf calls.
1517 #ifdef ASSERT
1518 {
1519 Label L;
1520 ldr(rscratch1, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize));
1521 cbz(rscratch1, L);
1522 stop("InterpreterMacroAssembler::call_VM_leaf_base:"
1523 " last_sp != NULL");
1524 bind(L);
1525 }
1526 #endif /* ASSERT */
1527 // super call
1528 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments);
1529 }
1530
1531 void InterpreterMacroAssembler::call_VM_base(Register oop_result,
1532 Register java_thread,
1533 Register last_java_sp,
1534 address entry_point,
1535 int number_of_arguments,
1536 bool check_exceptions) {
1537 // interpreter specific
1538 //
1539 // Note: Could avoid restoring locals ptr (callee saved) - however doesn't
1540 // really make a difference for these runtime calls, since they are
1541 // slow anyway. Btw., bcp must be saved/restored since it may change
1542 // due to GC.
1543 // assert(java_thread == noreg , "not expecting a precomputed java thread");
1544 save_bcp();
1545 #ifdef ASSERT
1546 {
1547 Label L;
1548 ldr(rscratch1, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize));
1549 cbz(rscratch1, L);
1550 stop("InterpreterMacroAssembler::call_VM_leaf_base:"
1551 " last_sp != NULL");
1552 bind(L);
1553 }
1554 #endif /* ASSERT */
1555 // super call
1556 MacroAssembler::call_VM_base(oop_result, noreg, last_java_sp,
1557 entry_point, number_of_arguments,
1558 check_exceptions);
1559 // interpreter specific
1560 restore_bcp();
1561 restore_locals();
1562 }
1563
1564 void InterpreterMacroAssembler::profile_obj_type(Register obj, const Address& mdo_addr) {
1565 Label update, next, none;
1566
1567 verify_oop(obj);
1568
1569 cbnz(obj, update);
1570 orptr(mdo_addr, TypeEntries::null_seen);
1571 b(next);
1572
1573 bind(update);
1574 load_klass(obj, obj);
1575
1576 ldr(rscratch1, mdo_addr);
1577 eor(obj, obj, rscratch1);
1578 tst(obj, TypeEntries::type_klass_mask);
1579 br(Assembler::EQ, next); // klass seen before, nothing to
1580 // do. The unknown bit may have been
1581 // set already but no need to check.
1582
1583 tbnz(obj, exact_log2(TypeEntries::type_unknown), next);
1584 // already unknown. Nothing to do anymore.
1585
1586 ldr(rscratch1, mdo_addr);
1587 cbz(rscratch1, none);
1588 cmp(rscratch1, TypeEntries::null_seen);
1589 br(Assembler::EQ, none);
1590 // There is a chance that the checks above (re-reading profiling
1591 // data from memory) fail if another thread has just set the
1592 // profiling to this obj's klass
1593 ldr(rscratch1, mdo_addr);
1594 eor(obj, obj, rscratch1);
1595 tst(obj, TypeEntries::type_klass_mask);
1596 br(Assembler::EQ, next);
1597
1598 // different than before. Cannot keep accurate profile.
1599 orptr(mdo_addr, TypeEntries::type_unknown);
1600 b(next);
1601
1602 bind(none);
1603 // first time here. Set profile type.
1604 str(obj, mdo_addr);
1605
1606 bind(next);
1607 }
1608
1609 void InterpreterMacroAssembler::profile_arguments_type(Register mdp, Register callee, Register tmp, bool is_virtual) {
1610 if (!ProfileInterpreter) {
1611 return;
1612 }
1613
1614 if (MethodData::profile_arguments() || MethodData::profile_return()) {
1615 Label profile_continue;
1616
1617 test_method_data_pointer(mdp, profile_continue);
1618
1619 int off_to_start = is_virtual ? in_bytes(VirtualCallData::virtual_call_data_size()) : in_bytes(CounterData::counter_data_size());
1620
1621 ldrb(rscratch1, Address(mdp, in_bytes(DataLayout::tag_offset()) - off_to_start));
1622 cmp(rscratch1, is_virtual ? DataLayout::virtual_call_type_data_tag : DataLayout::call_type_data_tag);
1623 br(Assembler::NE, profile_continue);
1624
1625 if (MethodData::profile_arguments()) {
1626 Label done;
1627 int off_to_args = in_bytes(TypeEntriesAtCall::args_data_offset());
1628
1629 for (int i = 0; i < TypeProfileArgsLimit; i++) {
1630 if (i > 0 || MethodData::profile_return()) {
1631 // If return value type is profiled we may have no argument to profile
1632 ldr(tmp, Address(mdp, in_bytes(TypeEntriesAtCall::cell_count_offset())));
1633 sub(tmp, tmp, i*TypeStackSlotEntries::per_arg_count());
1634 cmp(tmp, TypeStackSlotEntries::per_arg_count());
1635 add(rscratch1, mdp, off_to_args);
1636 br(Assembler::LT, done);
1637 }
1638 ldr(tmp, Address(callee, Method::const_offset()));
1639 load_unsigned_short(tmp, Address(tmp, ConstMethod::size_of_parameters_offset()));
1640 // stack offset o (zero based) from the start of the argument
1641 // list, for n arguments translates into offset n - o - 1 from
1642 // the end of the argument list
1643 ldr(rscratch1, Address(mdp, in_bytes(TypeEntriesAtCall::stack_slot_offset(i))));
1644 sub(tmp, tmp, rscratch1);
1645 sub(tmp, tmp, 1);
1646 Address arg_addr = argument_address(tmp);
1647 ldr(tmp, arg_addr);
1648
1649 Address mdo_arg_addr(mdp, in_bytes(TypeEntriesAtCall::argument_type_offset(i)));
1650 profile_obj_type(tmp, mdo_arg_addr);
1651
1652 int to_add = in_bytes(TypeStackSlotEntries::per_arg_size());
1653 off_to_args += to_add;
1654 }
1655
1656 if (MethodData::profile_return()) {
1657 ldr(tmp, Address(mdp, in_bytes(TypeEntriesAtCall::cell_count_offset())));
1658 sub(tmp, tmp, TypeProfileArgsLimit*TypeStackSlotEntries::per_arg_count());
1659 }
1660
1661 add(rscratch1, mdp, off_to_args);
1662 bind(done);
1663 mov(mdp, rscratch1);
1664
1665 if (MethodData::profile_return()) {
1666 // We're right after the type profile for the last
1667 // argument. tmp is the number of cells left in the
1668 // CallTypeData/VirtualCallTypeData to reach its end. Non null
1669 // if there's a return to profile.
1670 assert(ReturnTypeEntry::static_cell_count() < TypeStackSlotEntries::per_arg_count(), "can't move past ret type");
1671 add(mdp, mdp, tmp, LSL, exact_log2(DataLayout::cell_size));
1672 }
1673 str(mdp, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize));
1674 } else {
1675 assert(MethodData::profile_return(), "either profile call args or call ret");
1676 update_mdp_by_constant(mdp, in_bytes(TypeEntriesAtCall::return_only_size()));
1677 }
1678
1679 // mdp points right after the end of the
1680 // CallTypeData/VirtualCallTypeData, right after the cells for the
1681 // return value type if there's one
1682
1683 bind(profile_continue);
1684 }
1685 }
1686
1687 void InterpreterMacroAssembler::profile_return_type(Register mdp, Register ret, Register tmp) {
1688 assert_different_registers(mdp, ret, tmp, rbcp);
1689 if (ProfileInterpreter && MethodData::profile_return()) {
1690 Label profile_continue, done;
1691
1692 test_method_data_pointer(mdp, profile_continue);
1693
1694 if (MethodData::profile_return_jsr292_only()) {
1695 assert(Method::intrinsic_id_size_in_bytes() == 2, "assuming Method::_intrinsic_id is u2");
1696
1697 // If we don't profile all invoke bytecodes we must make sure
1698 // it's a bytecode we indeed profile. We can't go back to the
1699 // begining of the ProfileData we intend to update to check its
1700 // type because we're right after it and we don't known its
1701 // length
1702 Label do_profile;
1703 ldrb(rscratch1, Address(rbcp, 0));
1704 cmp(rscratch1, Bytecodes::_invokedynamic);
1705 br(Assembler::EQ, do_profile);
1706 cmp(rscratch1, Bytecodes::_invokehandle);
1707 br(Assembler::EQ, do_profile);
1708 get_method(tmp);
1709 ldrh(rscratch1, Address(tmp, Method::intrinsic_id_offset_in_bytes()));
1710 cmp(rscratch1, vmIntrinsics::_compiledLambdaForm);
1711 br(Assembler::NE, profile_continue);
1712
1713 bind(do_profile);
1714 }
1715
1716 Address mdo_ret_addr(mdp, -in_bytes(ReturnTypeEntry::size()));
1717 mov(tmp, ret);
1718 profile_obj_type(tmp, mdo_ret_addr);
1719
1720 bind(profile_continue);
1721 }
1722 }
1723
1724 void InterpreterMacroAssembler::profile_parameters_type(Register mdp, Register tmp1, Register tmp2) {
1725 if (ProfileInterpreter && MethodData::profile_parameters()) {
1726 Label profile_continue, done;
1727
1728 test_method_data_pointer(mdp, profile_continue);
1729
1730 // Load the offset of the area within the MDO used for
1731 // parameters. If it's negative we're not profiling any parameters
1732 ldr(tmp1, Address(mdp, in_bytes(MethodData::parameters_type_data_di_offset()) - in_bytes(MethodData::data_offset())));
1733 cmp(tmp1, 0u);
1734 br(Assembler::LT, profile_continue);
1735
1736 // Compute a pointer to the area for parameters from the offset
1737 // and move the pointer to the slot for the last
1738 // parameters. Collect profiling from last parameter down.
1739 // mdo start + parameters offset + array length - 1
1740 add(mdp, mdp, tmp1);
1741 ldr(tmp1, Address(mdp, ArrayData::array_len_offset()));
1742 sub(tmp1, tmp1, TypeStackSlotEntries::per_arg_count());
1743
1744 Label loop;
1745 bind(loop);
1746
1747 int off_base = in_bytes(ParametersTypeData::stack_slot_offset(0));
1748 int type_base = in_bytes(ParametersTypeData::type_offset(0));
1749 int per_arg_scale = exact_log2(DataLayout::cell_size);
1750 add(rscratch1, mdp, off_base);
1751 add(rscratch2, mdp, type_base);
1752
1753 Address arg_off(rscratch1, tmp1, Address::lsl(per_arg_scale));
1754 Address arg_type(rscratch2, tmp1, Address::lsl(per_arg_scale));
1755
1756 // load offset on the stack from the slot for this parameter
1757 ldr(tmp2, arg_off);
1758 neg(tmp2, tmp2);
1759 // read the parameter from the local area
1760 ldr(tmp2, Address(rlocals, tmp2, Address::lsl(Interpreter::logStackElementSize)));
1761
1762 // profile the parameter
1763 profile_obj_type(tmp2, arg_type);
1764
1765 // go to next parameter
1766 subs(tmp1, tmp1, TypeStackSlotEntries::per_arg_count());
1767 br(Assembler::GE, loop);
1768
1769 bind(profile_continue);
1770 }
1771 }