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 oopDesc::bs()->interpreter_read_barrier_not_null(this, result);
277 // Add in the index
278 add(result, result, tmp);
279 load_heap_oop(result, Address(result, arrayOopDesc::base_offset_in_bytes(T_OBJECT)));
280 }
281
282 // Generate a subtype check: branch to ok_is_subtype if sub_klass is a
283 // subtype of super_klass.
284 //
285 // Args:
286 // r0: superklass
287 // Rsub_klass: subklass
288 //
289 // Kills:
290 // r2, r5
291 void InterpreterMacroAssembler::gen_subtype_check(Register Rsub_klass,
292 Label& ok_is_subtype) {
293 assert(Rsub_klass != r0, "r0 holds superklass");
294 assert(Rsub_klass != r2, "r2 holds 2ndary super array length");
295 assert(Rsub_klass != r5, "r5 holds 2ndary super array scan ptr");
296
297 // Profile the not-null value's klass.
298 profile_typecheck(r2, Rsub_klass, r5); // blows r2, reloads r5
299
300 // Do the check.
301 check_klass_subtype(Rsub_klass, r0, r2, ok_is_subtype); // blows r2
302
303 // Profile the failure of the check.
304 profile_typecheck_failed(r2); // blows r2
305 }
306
307 // Java Expression Stack
308
309 void InterpreterMacroAssembler::pop_ptr(Register r) {
310 ldr(r, post(esp, wordSize));
311 }
312
313 void InterpreterMacroAssembler::pop_i(Register r) {
314 ldrw(r, post(esp, wordSize));
315 }
316
317 void InterpreterMacroAssembler::pop_l(Register r) {
318 ldr(r, post(esp, 2 * Interpreter::stackElementSize));
319 }
320
321 void InterpreterMacroAssembler::push_ptr(Register r) {
322 str(r, pre(esp, -wordSize));
323 }
324
325 void InterpreterMacroAssembler::push_i(Register r) {
326 str(r, pre(esp, -wordSize));
327 }
328
329 void InterpreterMacroAssembler::push_l(Register r) {
330 str(r, pre(esp, 2 * -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 shenandoah_store_check(obj_reg);
667
668 if (UseBiasedLocking) {
669 biased_locking_enter(lock_reg, obj_reg, swap_reg, tmp, false, done, &slow_case);
670 }
671
672 // Load (object->mark() | 1) into swap_reg
673 ldr(rscratch1, Address(obj_reg, 0));
674 orr(swap_reg, rscratch1, 1);
675
676 // Save (object->mark() | 1) into BasicLock's displaced header
677 str(swap_reg, Address(lock_reg, mark_offset));
678
679 assert(lock_offset == 0,
680 "displached header must be first word in BasicObjectLock");
681
682 Label fail;
683 if (PrintBiasedLockingStatistics) {
684 Label fast;
685 cmpxchgptr(swap_reg, lock_reg, obj_reg, rscratch1, fast, &fail);
686 bind(fast);
687 atomic_incw(Address((address)BiasedLocking::fast_path_entry_count_addr()),
688 rscratch2, rscratch1, tmp);
689 b(done);
690 bind(fail);
691 } else {
692 cmpxchgptr(swap_reg, lock_reg, obj_reg, rscratch1, done, /*fallthrough*/NULL);
693 }
694
695 // Test if the oopMark is an obvious stack pointer, i.e.,
696 // 1) (mark & 7) == 0, and
697 // 2) rsp <= mark < mark + os::pagesize()
698 //
699 // These 3 tests can be done by evaluating the following
700 // expression: ((mark - rsp) & (7 - os::vm_page_size())),
701 // assuming both stack pointer and pagesize have their
702 // least significant 3 bits clear.
703 // NOTE: the oopMark is in swap_reg %r0 as the result of cmpxchg
704 // NOTE2: aarch64 does not like to subtract sp from rn so take a
705 // copy
706 mov(rscratch1, sp);
707 sub(swap_reg, swap_reg, rscratch1);
708 ands(swap_reg, swap_reg, (unsigned long)(7 - os::vm_page_size()));
709
710 // Save the test result, for recursive case, the result is zero
711 str(swap_reg, Address(lock_reg, mark_offset));
712
713 if (PrintBiasedLockingStatistics) {
714 br(Assembler::NE, slow_case);
715 atomic_incw(Address((address)BiasedLocking::fast_path_entry_count_addr()),
716 rscratch2, rscratch1, tmp);
717 }
718 br(Assembler::EQ, done);
719
720 bind(slow_case);
721
722 // Call the runtime routine for slow case
723 call_VM(noreg,
724 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter),
725 lock_reg);
726
727 bind(done);
728 }
729 }
730
731
732 // Unlocks an object. Used in monitorexit bytecode and
733 // remove_activation. Throws an IllegalMonitorException if object is
734 // not locked by current thread.
735 //
736 // Args:
737 // c_rarg1: BasicObjectLock for lock
738 //
739 // Kills:
740 // r0
741 // c_rarg0, c_rarg1, c_rarg2, c_rarg3, ... (param regs)
742 // rscratch1, rscratch2 (scratch regs)
743 void InterpreterMacroAssembler::unlock_object(Register lock_reg)
744 {
745 assert(lock_reg == c_rarg1, "The argument is only for looks. It must be rarg1");
746
747 if (UseHeavyMonitors) {
748 call_VM(noreg,
749 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorexit),
750 lock_reg);
751 } else {
752 Label done;
753
754 const Register swap_reg = r0;
755 const Register header_reg = c_rarg2; // Will contain the old oopMark
756 const Register obj_reg = c_rarg3; // Will contain the oop
757
758 save_bcp(); // Save in case of exception
759
760 // Convert from BasicObjectLock structure to object and BasicLock
761 // structure Store the BasicLock address into %r0
762 lea(swap_reg, Address(lock_reg, BasicObjectLock::lock_offset_in_bytes()));
763
764 // Load oop into obj_reg(%c_rarg3)
765 ldr(obj_reg, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes()));
766
767 shenandoah_store_check(obj_reg);
768
769 // Free entry
770 str(zr, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes()));
771
772 if (UseBiasedLocking) {
773 biased_locking_exit(obj_reg, header_reg, done);
774 }
775
776 // Load the old header from BasicLock structure
777 ldr(header_reg, Address(swap_reg,
778 BasicLock::displaced_header_offset_in_bytes()));
779
780 // Test for recursion
781 cbz(header_reg, done);
782
783 // Atomic swap back the old header
784 cmpxchgptr(swap_reg, header_reg, obj_reg, rscratch1, done, /*fallthrough*/NULL);
785
786 // Call the runtime routine for slow case.
787 str(obj_reg, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes())); // restore obj
788 call_VM(noreg,
789 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorexit),
790 lock_reg);
791
792 bind(done);
793
794 restore_bcp();
795 }
796 }
797
798 void InterpreterMacroAssembler::test_method_data_pointer(Register mdp,
799 Label& zero_continue) {
800 assert(ProfileInterpreter, "must be profiling interpreter");
801 ldr(mdp, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize));
802 cbz(mdp, zero_continue);
803 }
804
805 // Set the method data pointer for the current bcp.
806 void InterpreterMacroAssembler::set_method_data_pointer_for_bcp() {
807 assert(ProfileInterpreter, "must be profiling interpreter");
808 Label set_mdp;
809 stp(r0, r1, Address(pre(sp, -2 * wordSize)));
810
811 // Test MDO to avoid the call if it is NULL.
812 ldr(r0, Address(rmethod, in_bytes(Method::method_data_offset())));
813 cbz(r0, set_mdp);
814 call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::bcp_to_di), rmethod, rbcp);
815 // r0: mdi
816 // mdo is guaranteed to be non-zero here, we checked for it before the call.
817 ldr(r1, Address(rmethod, in_bytes(Method::method_data_offset())));
818 lea(r1, Address(r1, in_bytes(MethodData::data_offset())));
819 add(r0, r1, r0);
820 str(r0, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize));
821 bind(set_mdp);
822 ldp(r0, r1, Address(post(sp, 2 * wordSize)));
823 }
824
825 void InterpreterMacroAssembler::verify_method_data_pointer() {
826 assert(ProfileInterpreter, "must be profiling interpreter");
827 #ifdef ASSERT
828 Label verify_continue;
829 stp(r0, r1, Address(pre(sp, -2 * wordSize)));
830 stp(r2, r3, Address(pre(sp, -2 * wordSize)));
831 test_method_data_pointer(r3, verify_continue); // If mdp is zero, continue
832 get_method(r1);
833
834 // If the mdp is valid, it will point to a DataLayout header which is
835 // consistent with the bcp. The converse is highly probable also.
836 ldrsh(r2, Address(r3, in_bytes(DataLayout::bci_offset())));
837 ldr(rscratch1, Address(r1, Method::const_offset()));
838 add(r2, r2, rscratch1, Assembler::LSL);
839 lea(r2, Address(r2, ConstMethod::codes_offset()));
840 cmp(r2, rbcp);
841 br(Assembler::EQ, verify_continue);
842 // r1: method
843 // rbcp: bcp // rbcp == 22
844 // r3: mdp
845 call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::verify_mdp),
846 r1, rbcp, r3);
847 bind(verify_continue);
848 ldp(r2, r3, Address(post(sp, 2 * wordSize)));
849 ldp(r0, r1, Address(post(sp, 2 * wordSize)));
850 #endif // ASSERT
851 }
852
853
854 void InterpreterMacroAssembler::set_mdp_data_at(Register mdp_in,
855 int constant,
856 Register value) {
857 assert(ProfileInterpreter, "must be profiling interpreter");
858 Address data(mdp_in, constant);
859 str(value, data);
860 }
861
862
863 void InterpreterMacroAssembler::increment_mdp_data_at(Register mdp_in,
864 int constant,
865 bool decrement) {
866 increment_mdp_data_at(mdp_in, noreg, constant, decrement);
867 }
868
869 void InterpreterMacroAssembler::increment_mdp_data_at(Register mdp_in,
870 Register reg,
871 int constant,
872 bool decrement) {
873 assert(ProfileInterpreter, "must be profiling interpreter");
874 // %%% this does 64bit counters at best it is wasting space
875 // at worst it is a rare bug when counters overflow
876
877 assert_different_registers(rscratch2, rscratch1, mdp_in, reg);
878
879 Address addr1(mdp_in, constant);
880 Address addr2(rscratch2, reg, Address::lsl(0));
881 Address &addr = addr1;
882 if (reg != noreg) {
883 lea(rscratch2, addr1);
884 addr = addr2;
885 }
886
887 if (decrement) {
888 // Decrement the register. Set condition codes.
889 // Intel does this
890 // addptr(data, (int32_t) -DataLayout::counter_increment);
891 // If the decrement causes the counter to overflow, stay negative
892 // Label L;
893 // jcc(Assembler::negative, L);
894 // addptr(data, (int32_t) DataLayout::counter_increment);
895 // so we do this
896 ldr(rscratch1, addr);
897 subs(rscratch1, rscratch1, (unsigned)DataLayout::counter_increment);
898 Label L;
899 br(Assembler::LO, L); // skip store if counter underflow
900 str(rscratch1, addr);
901 bind(L);
902 } else {
903 assert(DataLayout::counter_increment == 1,
904 "flow-free idiom only works with 1");
905 // Intel does this
906 // Increment the register. Set carry flag.
907 // addptr(data, DataLayout::counter_increment);
908 // If the increment causes the counter to overflow, pull back by 1.
909 // sbbptr(data, (int32_t)0);
910 // so we do this
911 ldr(rscratch1, addr);
912 adds(rscratch1, rscratch1, DataLayout::counter_increment);
913 Label L;
914 br(Assembler::CS, L); // skip store if counter overflow
915 str(rscratch1, addr);
916 bind(L);
917 }
918 }
919
920 void InterpreterMacroAssembler::set_mdp_flag_at(Register mdp_in,
921 int flag_byte_constant) {
922 assert(ProfileInterpreter, "must be profiling interpreter");
923 int header_offset = in_bytes(DataLayout::header_offset());
924 int header_bits = DataLayout::flag_mask_to_header_mask(flag_byte_constant);
925 // Set the flag
926 ldr(rscratch1, Address(mdp_in, header_offset));
927 orr(rscratch1, rscratch1, header_bits);
928 str(rscratch1, Address(mdp_in, header_offset));
929 }
930
931
932 void InterpreterMacroAssembler::test_mdp_data_at(Register mdp_in,
933 int offset,
934 Register value,
935 Register test_value_out,
936 Label& not_equal_continue) {
937 assert(ProfileInterpreter, "must be profiling interpreter");
938 if (test_value_out == noreg) {
939 ldr(rscratch1, Address(mdp_in, offset));
940 cmp(value, rscratch1);
941 } else {
942 // Put the test value into a register, so caller can use it:
943 ldr(test_value_out, Address(mdp_in, offset));
944 cmp(value, test_value_out);
945 }
946 br(Assembler::NE, not_equal_continue);
947 }
948
949
950 void InterpreterMacroAssembler::update_mdp_by_offset(Register mdp_in,
951 int offset_of_disp) {
952 assert(ProfileInterpreter, "must be profiling interpreter");
953 ldr(rscratch1, Address(mdp_in, offset_of_disp));
954 add(mdp_in, mdp_in, rscratch1, LSL);
955 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize));
956 }
957
958
959 void InterpreterMacroAssembler::update_mdp_by_offset(Register mdp_in,
960 Register reg,
961 int offset_of_disp) {
962 assert(ProfileInterpreter, "must be profiling interpreter");
963 lea(rscratch1, Address(mdp_in, offset_of_disp));
964 ldr(rscratch1, Address(rscratch1, reg, Address::lsl(0)));
965 add(mdp_in, mdp_in, rscratch1, LSL);
966 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize));
967 }
968
969
970 void InterpreterMacroAssembler::update_mdp_by_constant(Register mdp_in,
971 int constant) {
972 assert(ProfileInterpreter, "must be profiling interpreter");
973 add(mdp_in, mdp_in, (unsigned)constant);
974 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize));
975 }
976
977
978 void InterpreterMacroAssembler::update_mdp_for_ret(Register return_bci) {
979 assert(ProfileInterpreter, "must be profiling interpreter");
980 // save/restore across call_VM
981 stp(zr, return_bci, Address(pre(sp, -2 * wordSize)));
982 call_VM(noreg,
983 CAST_FROM_FN_PTR(address, InterpreterRuntime::update_mdp_for_ret),
984 return_bci);
985 ldp(zr, return_bci, Address(post(sp, 2 * wordSize)));
986 }
987
988
989 void InterpreterMacroAssembler::profile_taken_branch(Register mdp,
990 Register bumped_count) {
991 if (ProfileInterpreter) {
992 Label profile_continue;
993
994 // If no method data exists, go to profile_continue.
995 // Otherwise, assign to mdp
996 test_method_data_pointer(mdp, profile_continue);
997
998 // We are taking a branch. Increment the taken count.
999 // We inline increment_mdp_data_at to return bumped_count in a register
1000 //increment_mdp_data_at(mdp, in_bytes(JumpData::taken_offset()));
1001 Address data(mdp, in_bytes(JumpData::taken_offset()));
1002 ldr(bumped_count, data);
1003 assert(DataLayout::counter_increment == 1,
1004 "flow-free idiom only works with 1");
1005 // Intel does this to catch overflow
1006 // addptr(bumped_count, DataLayout::counter_increment);
1007 // sbbptr(bumped_count, 0);
1008 // so we do this
1009 adds(bumped_count, bumped_count, DataLayout::counter_increment);
1010 Label L;
1011 br(Assembler::CS, L); // skip store if counter overflow
1012 str(bumped_count, data);
1013 bind(L);
1014 // The method data pointer needs to be updated to reflect the new target.
1015 update_mdp_by_offset(mdp, in_bytes(JumpData::displacement_offset()));
1016 bind(profile_continue);
1017 }
1018 }
1019
1020
1021 void InterpreterMacroAssembler::profile_not_taken_branch(Register mdp) {
1022 if (ProfileInterpreter) {
1023 Label profile_continue;
1024
1025 // If no method data exists, go to profile_continue.
1026 test_method_data_pointer(mdp, profile_continue);
1027
1028 // We are taking a branch. Increment the not taken count.
1029 increment_mdp_data_at(mdp, in_bytes(BranchData::not_taken_offset()));
1030
1031 // The method data pointer needs to be updated to correspond to
1032 // the next bytecode
1033 update_mdp_by_constant(mdp, in_bytes(BranchData::branch_data_size()));
1034 bind(profile_continue);
1035 }
1036 }
1037
1038
1039 void InterpreterMacroAssembler::profile_call(Register mdp) {
1040 if (ProfileInterpreter) {
1041 Label profile_continue;
1042
1043 // If no method data exists, go to profile_continue.
1044 test_method_data_pointer(mdp, profile_continue);
1045
1046 // We are making a call. Increment the count.
1047 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
1048
1049 // The method data pointer needs to be updated to reflect the new target.
1050 update_mdp_by_constant(mdp, in_bytes(CounterData::counter_data_size()));
1051 bind(profile_continue);
1052 }
1053 }
1054
1055 void InterpreterMacroAssembler::profile_final_call(Register mdp) {
1056 if (ProfileInterpreter) {
1057 Label profile_continue;
1058
1059 // If no method data exists, go to profile_continue.
1060 test_method_data_pointer(mdp, profile_continue);
1061
1062 // We are making a call. Increment the count.
1063 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
1064
1065 // The method data pointer needs to be updated to reflect the new target.
1066 update_mdp_by_constant(mdp,
1067 in_bytes(VirtualCallData::
1068 virtual_call_data_size()));
1069 bind(profile_continue);
1070 }
1071 }
1072
1073
1074 void InterpreterMacroAssembler::profile_virtual_call(Register receiver,
1075 Register mdp,
1076 Register reg2,
1077 bool receiver_can_be_null) {
1078 if (ProfileInterpreter) {
1079 Label profile_continue;
1080
1081 // If no method data exists, go to profile_continue.
1082 test_method_data_pointer(mdp, profile_continue);
1083
1084 Label skip_receiver_profile;
1085 if (receiver_can_be_null) {
1086 Label not_null;
1087 // We are making a call. Increment the count for null receiver.
1088 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
1089 b(skip_receiver_profile);
1090 bind(not_null);
1091 }
1092
1093 // Record the receiver type.
1094 record_klass_in_profile(receiver, mdp, reg2, true);
1095 bind(skip_receiver_profile);
1096
1097 // The method data pointer needs to be updated to reflect the new target.
1098 #if INCLUDE_JVMCI
1099 if (MethodProfileWidth == 0) {
1100 update_mdp_by_constant(mdp, in_bytes(VirtualCallData::virtual_call_data_size()));
1101 }
1102 #else // INCLUDE_JVMCI
1103 update_mdp_by_constant(mdp,
1104 in_bytes(VirtualCallData::
1105 virtual_call_data_size()));
1106 #endif // INCLUDE_JVMCI
1107 bind(profile_continue);
1108 }
1109 }
1110
1111 #if INCLUDE_JVMCI
1112 void InterpreterMacroAssembler::profile_called_method(Register method, Register mdp, Register reg2) {
1113 assert_different_registers(method, mdp, reg2);
1114 if (ProfileInterpreter && MethodProfileWidth > 0) {
1115 Label profile_continue;
1116
1117 // If no method data exists, go to profile_continue.
1118 test_method_data_pointer(mdp, profile_continue);
1119
1120 Label done;
1121 record_item_in_profile_helper(method, mdp, reg2, 0, done, MethodProfileWidth,
1122 &VirtualCallData::method_offset, &VirtualCallData::method_count_offset, in_bytes(VirtualCallData::nonprofiled_receiver_count_offset()));
1123 bind(done);
1124
1125 update_mdp_by_constant(mdp, in_bytes(VirtualCallData::virtual_call_data_size()));
1126 bind(profile_continue);
1127 }
1128 }
1129 #endif // INCLUDE_JVMCI
1130
1131 // This routine creates a state machine for updating the multi-row
1132 // type profile at a virtual call site (or other type-sensitive bytecode).
1133 // The machine visits each row (of receiver/count) until the receiver type
1134 // is found, or until it runs out of rows. At the same time, it remembers
1135 // the location of the first empty row. (An empty row records null for its
1136 // receiver, and can be allocated for a newly-observed receiver type.)
1137 // Because there are two degrees of freedom in the state, a simple linear
1138 // search will not work; it must be a decision tree. Hence this helper
1139 // function is recursive, to generate the required tree structured code.
1140 // It's the interpreter, so we are trading off code space for speed.
1141 // See below for example code.
1142 void InterpreterMacroAssembler::record_klass_in_profile_helper(
1143 Register receiver, Register mdp,
1144 Register reg2, int start_row,
1145 Label& done, bool is_virtual_call) {
1146 if (TypeProfileWidth == 0) {
1147 if (is_virtual_call) {
1148 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
1149 }
1150 #if INCLUDE_JVMCI
1151 else if (EnableJVMCI) {
1152 increment_mdp_data_at(mdp, in_bytes(ReceiverTypeData::nonprofiled_receiver_count_offset()));
1153 }
1154 #endif // INCLUDE_JVMCI
1155 } else {
1156 int non_profiled_offset = -1;
1157 if (is_virtual_call) {
1158 non_profiled_offset = in_bytes(CounterData::count_offset());
1159 }
1160 #if INCLUDE_JVMCI
1161 else if (EnableJVMCI) {
1162 non_profiled_offset = in_bytes(ReceiverTypeData::nonprofiled_receiver_count_offset());
1163 }
1164 #endif // INCLUDE_JVMCI
1165
1166 record_item_in_profile_helper(receiver, mdp, reg2, 0, done, TypeProfileWidth,
1167 &VirtualCallData::receiver_offset, &VirtualCallData::receiver_count_offset, non_profiled_offset);
1168 }
1169 }
1170
1171 void InterpreterMacroAssembler::record_item_in_profile_helper(Register item, Register mdp,
1172 Register reg2, int start_row, Label& done, int total_rows,
1173 OffsetFunction item_offset_fn, OffsetFunction item_count_offset_fn,
1174 int non_profiled_offset) {
1175 int last_row = total_rows - 1;
1176 assert(start_row <= last_row, "must be work left to do");
1177 // Test this row for both the item and for null.
1178 // Take any of three different outcomes:
1179 // 1. found item => increment count and goto done
1180 // 2. found null => keep looking for case 1, maybe allocate this cell
1181 // 3. found something else => keep looking for cases 1 and 2
1182 // Case 3 is handled by a recursive call.
1183 for (int row = start_row; row <= last_row; row++) {
1184 Label next_test;
1185 bool test_for_null_also = (row == start_row);
1186
1187 // See if the item is item[n].
1188 int item_offset = in_bytes(item_offset_fn(row));
1189 test_mdp_data_at(mdp, item_offset, item,
1190 (test_for_null_also ? reg2 : noreg),
1191 next_test);
1192 // (Reg2 now contains the item from the CallData.)
1193
1194 // The item is item[n]. Increment count[n].
1195 int count_offset = in_bytes(item_count_offset_fn(row));
1196 increment_mdp_data_at(mdp, count_offset);
1197 b(done);
1198 bind(next_test);
1199
1200 if (test_for_null_also) {
1201 Label found_null;
1202 // Failed the equality check on item[n]... Test for null.
1203 if (start_row == last_row) {
1204 // The only thing left to do is handle the null case.
1205 if (non_profiled_offset >= 0) {
1206 cbz(reg2, found_null);
1207 // Item did not match any saved item and there is no empty row for it.
1208 // Increment total counter to indicate polymorphic case.
1209 increment_mdp_data_at(mdp, non_profiled_offset);
1210 b(done);
1211 bind(found_null);
1212 } else {
1213 cbnz(reg2, done);
1214 }
1215 break;
1216 }
1217 // Since null is rare, make it be the branch-taken case.
1218 cbz(reg2, found_null);
1219
1220 // Put all the "Case 3" tests here.
1221 record_item_in_profile_helper(item, mdp, reg2, start_row + 1, done, total_rows,
1222 item_offset_fn, item_count_offset_fn, non_profiled_offset);
1223
1224 // Found a null. Keep searching for a matching item,
1225 // but remember that this is an empty (unused) slot.
1226 bind(found_null);
1227 }
1228 }
1229
1230 // In the fall-through case, we found no matching item, but we
1231 // observed the item[start_row] is NULL.
1232
1233 // Fill in the item field and increment the count.
1234 int item_offset = in_bytes(item_offset_fn(start_row));
1235 set_mdp_data_at(mdp, item_offset, item);
1236 int count_offset = in_bytes(item_count_offset_fn(start_row));
1237 mov(reg2, DataLayout::counter_increment);
1238 set_mdp_data_at(mdp, count_offset, reg2);
1239 if (start_row > 0) {
1240 b(done);
1241 }
1242 }
1243
1244 // Example state machine code for three profile rows:
1245 // // main copy of decision tree, rooted at row[1]
1246 // if (row[0].rec == rec) { row[0].incr(); goto done; }
1247 // if (row[0].rec != NULL) {
1248 // // inner copy of decision tree, rooted at row[1]
1249 // if (row[1].rec == rec) { row[1].incr(); goto done; }
1250 // if (row[1].rec != NULL) {
1251 // // degenerate decision tree, rooted at row[2]
1252 // if (row[2].rec == rec) { row[2].incr(); goto done; }
1253 // if (row[2].rec != NULL) { count.incr(); goto done; } // overflow
1254 // row[2].init(rec); goto done;
1255 // } else {
1256 // // remember row[1] is empty
1257 // if (row[2].rec == rec) { row[2].incr(); goto done; }
1258 // row[1].init(rec); goto done;
1259 // }
1260 // } else {
1261 // // remember row[0] is empty
1262 // if (row[1].rec == rec) { row[1].incr(); goto done; }
1263 // if (row[2].rec == rec) { row[2].incr(); goto done; }
1264 // row[0].init(rec); goto done;
1265 // }
1266 // done:
1267
1268 void InterpreterMacroAssembler::record_klass_in_profile(Register receiver,
1269 Register mdp, Register reg2,
1270 bool is_virtual_call) {
1271 assert(ProfileInterpreter, "must be profiling");
1272 Label done;
1273
1274 record_klass_in_profile_helper(receiver, mdp, reg2, 0, done, is_virtual_call);
1275
1276 bind (done);
1277 }
1278
1279 void InterpreterMacroAssembler::profile_ret(Register return_bci,
1280 Register mdp) {
1281 if (ProfileInterpreter) {
1282 Label profile_continue;
1283 uint row;
1284
1285 // If no method data exists, go to profile_continue.
1286 test_method_data_pointer(mdp, profile_continue);
1287
1288 // Update the total ret count.
1289 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
1290
1291 for (row = 0; row < RetData::row_limit(); row++) {
1292 Label next_test;
1293
1294 // See if return_bci is equal to bci[n]:
1295 test_mdp_data_at(mdp,
1296 in_bytes(RetData::bci_offset(row)),
1297 return_bci, noreg,
1298 next_test);
1299
1300 // return_bci is equal to bci[n]. Increment the count.
1301 increment_mdp_data_at(mdp, in_bytes(RetData::bci_count_offset(row)));
1302
1303 // The method data pointer needs to be updated to reflect the new target.
1304 update_mdp_by_offset(mdp,
1305 in_bytes(RetData::bci_displacement_offset(row)));
1306 b(profile_continue);
1307 bind(next_test);
1308 }
1309
1310 update_mdp_for_ret(return_bci);
1311
1312 bind(profile_continue);
1313 }
1314 }
1315
1316 void InterpreterMacroAssembler::profile_null_seen(Register mdp) {
1317 if (ProfileInterpreter) {
1318 Label profile_continue;
1319
1320 // If no method data exists, go to profile_continue.
1321 test_method_data_pointer(mdp, profile_continue);
1322
1323 set_mdp_flag_at(mdp, BitData::null_seen_byte_constant());
1324
1325 // The method data pointer needs to be updated.
1326 int mdp_delta = in_bytes(BitData::bit_data_size());
1327 if (TypeProfileCasts) {
1328 mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size());
1329 }
1330 update_mdp_by_constant(mdp, mdp_delta);
1331
1332 bind(profile_continue);
1333 }
1334 }
1335
1336 void InterpreterMacroAssembler::profile_typecheck_failed(Register mdp) {
1337 if (ProfileInterpreter && TypeProfileCasts) {
1338 Label profile_continue;
1339
1340 // If no method data exists, go to profile_continue.
1341 test_method_data_pointer(mdp, profile_continue);
1342
1343 int count_offset = in_bytes(CounterData::count_offset());
1344 // Back up the address, since we have already bumped the mdp.
1345 count_offset -= in_bytes(VirtualCallData::virtual_call_data_size());
1346
1347 // *Decrement* the counter. We expect to see zero or small negatives.
1348 increment_mdp_data_at(mdp, count_offset, true);
1349
1350 bind (profile_continue);
1351 }
1352 }
1353
1354 void InterpreterMacroAssembler::profile_typecheck(Register mdp, Register klass, Register reg2) {
1355 if (ProfileInterpreter) {
1356 Label profile_continue;
1357
1358 // If no method data exists, go to profile_continue.
1359 test_method_data_pointer(mdp, profile_continue);
1360
1361 // The method data pointer needs to be updated.
1362 int mdp_delta = in_bytes(BitData::bit_data_size());
1363 if (TypeProfileCasts) {
1364 mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size());
1365
1366 // Record the object type.
1367 record_klass_in_profile(klass, mdp, reg2, false);
1368 }
1369 update_mdp_by_constant(mdp, mdp_delta);
1370
1371 bind(profile_continue);
1372 }
1373 }
1374
1375 void InterpreterMacroAssembler::profile_switch_default(Register mdp) {
1376 if (ProfileInterpreter) {
1377 Label profile_continue;
1378
1379 // If no method data exists, go to profile_continue.
1380 test_method_data_pointer(mdp, profile_continue);
1381
1382 // Update the default case count
1383 increment_mdp_data_at(mdp,
1384 in_bytes(MultiBranchData::default_count_offset()));
1385
1386 // The method data pointer needs to be updated.
1387 update_mdp_by_offset(mdp,
1388 in_bytes(MultiBranchData::
1389 default_displacement_offset()));
1390
1391 bind(profile_continue);
1392 }
1393 }
1394
1395 void InterpreterMacroAssembler::profile_switch_case(Register index,
1396 Register mdp,
1397 Register reg2) {
1398 if (ProfileInterpreter) {
1399 Label profile_continue;
1400
1401 // If no method data exists, go to profile_continue.
1402 test_method_data_pointer(mdp, profile_continue);
1403
1404 // Build the base (index * per_case_size_in_bytes()) +
1405 // case_array_offset_in_bytes()
1406 movw(reg2, in_bytes(MultiBranchData::per_case_size()));
1407 movw(rscratch1, in_bytes(MultiBranchData::case_array_offset()));
1408 Assembler::maddw(index, index, reg2, rscratch1);
1409
1410 // Update the case count
1411 increment_mdp_data_at(mdp,
1412 index,
1413 in_bytes(MultiBranchData::relative_count_offset()));
1414
1415 // The method data pointer needs to be updated.
1416 update_mdp_by_offset(mdp,
1417 index,
1418 in_bytes(MultiBranchData::
1419 relative_displacement_offset()));
1420
1421 bind(profile_continue);
1422 }
1423 }
1424
1425 void InterpreterMacroAssembler::verify_oop(Register reg, TosState state) {
1426 if (state == atos) {
1427 MacroAssembler::verify_oop(reg);
1428 }
1429 }
1430
1431 void InterpreterMacroAssembler::verify_FPU(int stack_depth, TosState state) { ; }
1432
1433
1434 void InterpreterMacroAssembler::notify_method_entry() {
1435 // Whenever JVMTI is interp_only_mode, method entry/exit events are sent to
1436 // track stack depth. If it is possible to enter interp_only_mode we add
1437 // the code to check if the event should be sent.
1438 if (JvmtiExport::can_post_interpreter_events()) {
1439 Label L;
1440 ldrw(r3, Address(rthread, JavaThread::interp_only_mode_offset()));
1441 cbzw(r3, L);
1442 call_VM(noreg, CAST_FROM_FN_PTR(address,
1443 InterpreterRuntime::post_method_entry));
1444 bind(L);
1445 }
1446
1447 {
1448 SkipIfEqual skip(this, &DTraceMethodProbes, false);
1449 get_method(c_rarg1);
1450 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
1451 rthread, c_rarg1);
1452 }
1453
1454 // RedefineClasses() tracing support for obsolete method entry
1455 if (log_is_enabled(Trace, redefine, class, obsolete)) {
1456 get_method(c_rarg1);
1457 call_VM_leaf(
1458 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
1459 rthread, c_rarg1);
1460 }
1461
1462 }
1463
1464
1465 void InterpreterMacroAssembler::notify_method_exit(
1466 TosState state, NotifyMethodExitMode mode) {
1467 // Whenever JVMTI is interp_only_mode, method entry/exit events are sent to
1468 // track stack depth. If it is possible to enter interp_only_mode we add
1469 // the code to check if the event should be sent.
1470 if (mode == NotifyJVMTI && JvmtiExport::can_post_interpreter_events()) {
1471 Label L;
1472 // Note: frame::interpreter_frame_result has a dependency on how the
1473 // method result is saved across the call to post_method_exit. If this
1474 // is changed then the interpreter_frame_result implementation will
1475 // need to be updated too.
1476
1477 // template interpreter will leave the result on the top of the stack.
1478 push(state);
1479 ldrw(r3, Address(rthread, JavaThread::interp_only_mode_offset()));
1480 cbz(r3, L);
1481 call_VM(noreg,
1482 CAST_FROM_FN_PTR(address, InterpreterRuntime::post_method_exit));
1483 bind(L);
1484 pop(state);
1485 }
1486
1487 {
1488 SkipIfEqual skip(this, &DTraceMethodProbes, false);
1489 push(state);
1490 get_method(c_rarg1);
1491 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
1492 rthread, c_rarg1);
1493 pop(state);
1494 }
1495 }
1496
1497
1498 // Jump if ((*counter_addr += increment) & mask) satisfies the condition.
1499 void InterpreterMacroAssembler::increment_mask_and_jump(Address counter_addr,
1500 int increment, Address mask,
1501 Register scratch, Register scratch2,
1502 bool preloaded, Condition cond,
1503 Label* where) {
1504 if (!preloaded) {
1505 ldrw(scratch, counter_addr);
1506 }
1507 add(scratch, scratch, increment);
1508 strw(scratch, counter_addr);
1509 ldrw(scratch2, mask);
1510 ands(scratch, scratch, scratch2);
1511 br(cond, *where);
1512 }
1513
1514 void InterpreterMacroAssembler::call_VM_leaf_base(address entry_point,
1515 int number_of_arguments) {
1516 // interpreter specific
1517 //
1518 // Note: No need to save/restore rbcp & rlocals pointer since these
1519 // are callee saved registers and no blocking/ GC can happen
1520 // in leaf calls.
1521 #ifdef ASSERT
1522 {
1523 Label L;
1524 ldr(rscratch1, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize));
1525 cbz(rscratch1, L);
1526 stop("InterpreterMacroAssembler::call_VM_leaf_base:"
1527 " last_sp != NULL");
1528 bind(L);
1529 }
1530 #endif /* ASSERT */
1531 // super call
1532 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments);
1533 }
1534
1535 void InterpreterMacroAssembler::call_VM_base(Register oop_result,
1536 Register java_thread,
1537 Register last_java_sp,
1538 address entry_point,
1539 int number_of_arguments,
1540 bool check_exceptions) {
1541 // interpreter specific
1542 //
1543 // Note: Could avoid restoring locals ptr (callee saved) - however doesn't
1544 // really make a difference for these runtime calls, since they are
1545 // slow anyway. Btw., bcp must be saved/restored since it may change
1546 // due to GC.
1547 // assert(java_thread == noreg , "not expecting a precomputed java thread");
1548 save_bcp();
1549 #ifdef ASSERT
1550 {
1551 Label L;
1552 ldr(rscratch1, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize));
1553 cbz(rscratch1, L);
1554 stop("InterpreterMacroAssembler::call_VM_leaf_base:"
1555 " last_sp != NULL");
1556 bind(L);
1557 }
1558 #endif /* ASSERT */
1559 // super call
1560 MacroAssembler::call_VM_base(oop_result, noreg, last_java_sp,
1561 entry_point, number_of_arguments,
1562 check_exceptions);
1563 // interpreter specific
1564 restore_bcp();
1565 restore_locals();
1566 }
1567
1568 void InterpreterMacroAssembler::profile_obj_type(Register obj, const Address& mdo_addr) {
1569 Label update, next, none;
1570
1571 verify_oop(obj);
1572
1573 cbnz(obj, update);
1574 orptr(mdo_addr, TypeEntries::null_seen);
1575 b(next);
1576
1577 bind(update);
1578 load_klass(obj, obj);
1579
1580 ldr(rscratch1, mdo_addr);
1581 eor(obj, obj, rscratch1);
1582 tst(obj, TypeEntries::type_klass_mask);
1583 br(Assembler::EQ, next); // klass seen before, nothing to
1584 // do. The unknown bit may have been
1585 // set already but no need to check.
1586
1587 tbnz(obj, exact_log2(TypeEntries::type_unknown), next);
1588 // already unknown. Nothing to do anymore.
1589
1590 ldr(rscratch1, mdo_addr);
1591 cbz(rscratch1, none);
1592 cmp(rscratch1, TypeEntries::null_seen);
1593 br(Assembler::EQ, none);
1594 // There is a chance that the checks above (re-reading profiling
1595 // data from memory) fail if another thread has just set the
1596 // profiling to this obj's klass
1597 ldr(rscratch1, mdo_addr);
1598 eor(obj, obj, rscratch1);
1599 tst(obj, TypeEntries::type_klass_mask);
1600 br(Assembler::EQ, next);
1601
1602 // different than before. Cannot keep accurate profile.
1603 orptr(mdo_addr, TypeEntries::type_unknown);
1604 b(next);
1605
1606 bind(none);
1607 // first time here. Set profile type.
1608 str(obj, mdo_addr);
1609
1610 bind(next);
1611 }
1612
1613 void InterpreterMacroAssembler::profile_arguments_type(Register mdp, Register callee, Register tmp, bool is_virtual) {
1614 if (!ProfileInterpreter) {
1615 return;
1616 }
1617
1618 if (MethodData::profile_arguments() || MethodData::profile_return()) {
1619 Label profile_continue;
1620
1621 test_method_data_pointer(mdp, profile_continue);
1622
1623 int off_to_start = is_virtual ? in_bytes(VirtualCallData::virtual_call_data_size()) : in_bytes(CounterData::counter_data_size());
1624
1625 ldrb(rscratch1, Address(mdp, in_bytes(DataLayout::tag_offset()) - off_to_start));
1626 cmp(rscratch1, is_virtual ? DataLayout::virtual_call_type_data_tag : DataLayout::call_type_data_tag);
1627 br(Assembler::NE, profile_continue);
1628
1629 if (MethodData::profile_arguments()) {
1630 Label done;
1631 int off_to_args = in_bytes(TypeEntriesAtCall::args_data_offset());
1632
1633 for (int i = 0; i < TypeProfileArgsLimit; i++) {
1634 if (i > 0 || MethodData::profile_return()) {
1635 // If return value type is profiled we may have no argument to profile
1636 ldr(tmp, Address(mdp, in_bytes(TypeEntriesAtCall::cell_count_offset())));
1637 sub(tmp, tmp, i*TypeStackSlotEntries::per_arg_count());
1638 cmp(tmp, TypeStackSlotEntries::per_arg_count());
1639 add(rscratch1, mdp, off_to_args);
1640 br(Assembler::LT, done);
1641 }
1642 ldr(tmp, Address(callee, Method::const_offset()));
1643 load_unsigned_short(tmp, Address(tmp, ConstMethod::size_of_parameters_offset()));
1644 // stack offset o (zero based) from the start of the argument
1645 // list, for n arguments translates into offset n - o - 1 from
1646 // the end of the argument list
1647 ldr(rscratch1, Address(mdp, in_bytes(TypeEntriesAtCall::stack_slot_offset(i))));
1648 sub(tmp, tmp, rscratch1);
1649 sub(tmp, tmp, 1);
1650 Address arg_addr = argument_address(tmp);
1651 ldr(tmp, arg_addr);
1652
1653 Address mdo_arg_addr(mdp, in_bytes(TypeEntriesAtCall::argument_type_offset(i)));
1654 profile_obj_type(tmp, mdo_arg_addr);
1655
1656 int to_add = in_bytes(TypeStackSlotEntries::per_arg_size());
1657 off_to_args += to_add;
1658 }
1659
1660 if (MethodData::profile_return()) {
1661 ldr(tmp, Address(mdp, in_bytes(TypeEntriesAtCall::cell_count_offset())));
1662 sub(tmp, tmp, TypeProfileArgsLimit*TypeStackSlotEntries::per_arg_count());
1663 }
1664
1665 add(rscratch1, mdp, off_to_args);
1666 bind(done);
1667 mov(mdp, rscratch1);
1668
1669 if (MethodData::profile_return()) {
1670 // We're right after the type profile for the last
1671 // argument. tmp is the number of cells left in the
1672 // CallTypeData/VirtualCallTypeData to reach its end. Non null
1673 // if there's a return to profile.
1674 assert(ReturnTypeEntry::static_cell_count() < TypeStackSlotEntries::per_arg_count(), "can't move past ret type");
1675 add(mdp, mdp, tmp, LSL, exact_log2(DataLayout::cell_size));
1676 }
1677 str(mdp, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize));
1678 } else {
1679 assert(MethodData::profile_return(), "either profile call args or call ret");
1680 update_mdp_by_constant(mdp, in_bytes(TypeEntriesAtCall::return_only_size()));
1681 }
1682
1683 // mdp points right after the end of the
1684 // CallTypeData/VirtualCallTypeData, right after the cells for the
1685 // return value type if there's one
1686
1687 bind(profile_continue);
1688 }
1689 }
1690
1691 void InterpreterMacroAssembler::profile_return_type(Register mdp, Register ret, Register tmp) {
1692 assert_different_registers(mdp, ret, tmp, rbcp);
1693 if (ProfileInterpreter && MethodData::profile_return()) {
1694 Label profile_continue, done;
1695
1696 test_method_data_pointer(mdp, profile_continue);
1697
1698 if (MethodData::profile_return_jsr292_only()) {
1699 assert(Method::intrinsic_id_size_in_bytes() == 2, "assuming Method::_intrinsic_id is u2");
1700
1701 // If we don't profile all invoke bytecodes we must make sure
1702 // it's a bytecode we indeed profile. We can't go back to the
1703 // begining of the ProfileData we intend to update to check its
1704 // type because we're right after it and we don't known its
1705 // length
1706 Label do_profile;
1707 ldrb(rscratch1, Address(rbcp, 0));
1708 cmp(rscratch1, Bytecodes::_invokedynamic);
1709 br(Assembler::EQ, do_profile);
1710 cmp(rscratch1, Bytecodes::_invokehandle);
1711 br(Assembler::EQ, do_profile);
1712 get_method(tmp);
1713 ldrh(rscratch1, Address(tmp, Method::intrinsic_id_offset_in_bytes()));
1714 cmp(rscratch1, vmIntrinsics::_compiledLambdaForm);
1715 br(Assembler::NE, profile_continue);
1716
1717 bind(do_profile);
1718 }
1719
1720 Address mdo_ret_addr(mdp, -in_bytes(ReturnTypeEntry::size()));
1721 mov(tmp, ret);
1722 profile_obj_type(tmp, mdo_ret_addr);
1723
1724 bind(profile_continue);
1725 }
1726 }
1727
1728 void InterpreterMacroAssembler::profile_parameters_type(Register mdp, Register tmp1, Register tmp2) {
1729 if (ProfileInterpreter && MethodData::profile_parameters()) {
1730 Label profile_continue, done;
1731
1732 test_method_data_pointer(mdp, profile_continue);
1733
1734 // Load the offset of the area within the MDO used for
1735 // parameters. If it's negative we're not profiling any parameters
1736 ldr(tmp1, Address(mdp, in_bytes(MethodData::parameters_type_data_di_offset()) - in_bytes(MethodData::data_offset())));
1737 cmp(tmp1, 0u);
1738 br(Assembler::LT, profile_continue);
1739
1740 // Compute a pointer to the area for parameters from the offset
1741 // and move the pointer to the slot for the last
1742 // parameters. Collect profiling from last parameter down.
1743 // mdo start + parameters offset + array length - 1
1744 add(mdp, mdp, tmp1);
1745 ldr(tmp1, Address(mdp, ArrayData::array_len_offset()));
1746 sub(tmp1, tmp1, TypeStackSlotEntries::per_arg_count());
1747
1748 Label loop;
1749 bind(loop);
1750
1751 int off_base = in_bytes(ParametersTypeData::stack_slot_offset(0));
1752 int type_base = in_bytes(ParametersTypeData::type_offset(0));
1753 int per_arg_scale = exact_log2(DataLayout::cell_size);
1754 add(rscratch1, mdp, off_base);
1755 add(rscratch2, mdp, type_base);
1756
1757 Address arg_off(rscratch1, tmp1, Address::lsl(per_arg_scale));
1758 Address arg_type(rscratch2, tmp1, Address::lsl(per_arg_scale));
1759
1760 // load offset on the stack from the slot for this parameter
1761 ldr(tmp2, arg_off);
1762 neg(tmp2, tmp2);
1763 // read the parameter from the local area
1764 ldr(tmp2, Address(rlocals, tmp2, Address::lsl(Interpreter::logStackElementSize)));
1765
1766 // profile the parameter
1767 profile_obj_type(tmp2, arg_type);
1768
1769 // go to next parameter
1770 subs(tmp1, tmp1, TypeStackSlotEntries::per_arg_count());
1771 br(Assembler::GE, loop);
1772
1773 bind(profile_continue);
1774 }
1775 }