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 "oops/arrayOop.hpp"
31 #include "oops/markOop.hpp"
32 #include "oops/methodData.hpp"
33 #include "oops/method.hpp"
34 #include "prims/jvmtiExport.hpp"
35 #include "prims/jvmtiRedefineClassesTrace.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 // Implementation of InterpreterMacroAssembler
44
45 void InterpreterMacroAssembler::jump_to_entry(address entry) {
46 assert(entry, "Entry must have been generated by now");
47 b(entry);
48 }
49
50 #ifndef CC_INTERP
51
52 void InterpreterMacroAssembler::check_and_handle_popframe(Register java_thread) {
53 if (JvmtiExport::can_pop_frame()) {
54 Label L;
55 // Initiate popframe handling only if it is not already being
56 // processed. If the flag has the popframe_processing bit set, it
57 // means that this code is called *during* popframe handling - we
58 // don't want to reenter.
59 // This method is only called just after the call into the vm in
60 // call_VM_base, so the arg registers are available.
61 ldrw(rscratch1, Address(rthread, JavaThread::popframe_condition_offset()));
62 tstw(rscratch1, JavaThread::popframe_pending_bit);
63 br(Assembler::EQ, L);
64 tstw(rscratch1, JavaThread::popframe_processing_bit);
65 br(Assembler::NE, L);
66 // Call Interpreter::remove_activation_preserving_args_entry() to get the
67 // address of the same-named entrypoint in the generated interpreter code.
68 call_VM_leaf(CAST_FROM_FN_PTR(address, Interpreter::remove_activation_preserving_args_entry));
69 br(r0);
70 bind(L);
71 }
72 }
73
74
75 void InterpreterMacroAssembler::load_earlyret_value(TosState state) {
76 ldr(r2, Address(rthread, JavaThread::jvmti_thread_state_offset()));
77 const Address tos_addr(r2, JvmtiThreadState::earlyret_tos_offset());
78 const Address oop_addr(r2, JvmtiThreadState::earlyret_oop_offset());
79 const Address val_addr(r2, JvmtiThreadState::earlyret_value_offset());
80 switch (state) {
81 case atos: ldr(r0, oop_addr);
82 str(zr, oop_addr);
83 verify_oop(r0, state); break;
84 case ltos: ldr(r0, val_addr); break;
85 case btos: // fall through
86 case ctos: // fall through
87 case stos: // fall through
88 case itos: ldrw(r0, val_addr); break;
89 case ftos: ldrs(v0, val_addr); break;
90 case dtos: ldrd(v0, val_addr); break;
91 case vtos: /* nothing to do */ break;
92 default : ShouldNotReachHere();
93 }
94 // Clean up tos value in the thread object
95 movw(rscratch1, (int) ilgl);
96 strw(rscratch1, tos_addr);
97 strw(zr, val_addr);
98 }
99
100
101 void InterpreterMacroAssembler::check_and_handle_earlyret(Register java_thread) {
102 if (JvmtiExport::can_force_early_return()) {
103 Label L;
104 ldr(rscratch1, Address(rthread, JavaThread::jvmti_thread_state_offset()));
105 cbz(rscratch1, L); // if (thread->jvmti_thread_state() == NULL) exit;
106
107 // Initiate earlyret handling only if it is not already being processed.
108 // If the flag has the earlyret_processing bit set, it means that this code
109 // is called *during* earlyret handling - we don't want to reenter.
110 ldrw(rscratch1, Address(rscratch1, JvmtiThreadState::earlyret_state_offset()));
111 cmpw(rscratch1, JvmtiThreadState::earlyret_pending);
112 br(Assembler::NE, L);
113
114 // Call Interpreter::remove_activation_early_entry() to get the address of the
115 // same-named entrypoint in the generated interpreter code.
116 ldr(rscratch1, Address(rthread, JavaThread::jvmti_thread_state_offset()));
117 ldrw(rscratch1, Address(rscratch1, JvmtiThreadState::earlyret_tos_offset()));
118 call_VM_leaf(CAST_FROM_FN_PTR(address, Interpreter::remove_activation_early_entry), rscratch1);
119 br(r0);
120 bind(L);
121 }
122 }
123
124 void InterpreterMacroAssembler::get_unsigned_2_byte_index_at_bcp(
125 Register reg,
126 int bcp_offset) {
127 assert(bcp_offset >= 0, "bcp is still pointing to start of bytecode");
128 ldrh(reg, Address(rbcp, bcp_offset));
129 rev16(reg, reg);
130 }
131
132 void InterpreterMacroAssembler::get_dispatch() {
133 unsigned long offset;
134 adrp(rdispatch, ExternalAddress((address)Interpreter::dispatch_table()), offset);
135 lea(rdispatch, Address(rdispatch, offset));
136 }
137
138 void InterpreterMacroAssembler::get_cache_index_at_bcp(Register index,
139 int bcp_offset,
140 size_t index_size) {
141 assert(bcp_offset > 0, "bcp is still pointing to start of bytecode");
142 if (index_size == sizeof(u2)) {
143 load_unsigned_short(index, Address(rbcp, bcp_offset));
144 } else if (index_size == sizeof(u4)) {
145 // assert(EnableInvokeDynamic, "giant index used only for JSR 292");
146 ldrw(index, Address(rbcp, bcp_offset));
147 // Check if the secondary index definition is still ~x, otherwise
148 // we have to change the following assembler code to calculate the
149 // plain index.
150 assert(ConstantPool::decode_invokedynamic_index(~123) == 123, "else change next line");
151 eonw(index, index, zr); // convert to plain index
152 } else if (index_size == sizeof(u1)) {
153 load_unsigned_byte(index, Address(rbcp, bcp_offset));
154 } else {
155 ShouldNotReachHere();
156 }
157 }
158
159 // Return
160 // Rindex: index into constant pool
161 // Rcache: address of cache entry - ConstantPoolCache::base_offset()
162 //
163 // A caller must add ConstantPoolCache::base_offset() to Rcache to get
164 // the true address of the cache entry.
165 //
166 void InterpreterMacroAssembler::get_cache_and_index_at_bcp(Register cache,
167 Register index,
168 int bcp_offset,
169 size_t index_size) {
170 assert_different_registers(cache, index);
171 assert_different_registers(cache, rcpool);
172 get_cache_index_at_bcp(index, bcp_offset, index_size);
173 assert(sizeof(ConstantPoolCacheEntry) == 4 * wordSize, "adjust code below");
174 // convert from field index to ConstantPoolCacheEntry
175 // aarch64 already has the cache in rcpool so there is no need to
176 // install it in cache. instead we pre-add the indexed offset to
177 // rcpool and return it in cache. All clients of this method need to
178 // be modified accordingly.
179 add(cache, rcpool, index, Assembler::LSL, 5);
180 }
181
182
183 void InterpreterMacroAssembler::get_cache_and_index_and_bytecode_at_bcp(Register cache,
184 Register index,
185 Register bytecode,
186 int byte_no,
187 int bcp_offset,
188 size_t index_size) {
189 get_cache_and_index_at_bcp(cache, index, bcp_offset, index_size);
190 // We use a 32-bit load here since the layout of 64-bit words on
191 // little-endian machines allow us that.
192 // n.b. unlike x86 cache alreeady includes the index offset
193 ldrw(bytecode, Address(cache,
194 ConstantPoolCache::base_offset()
195 + ConstantPoolCacheEntry::indices_offset()));
196 const int shift_count = (1 + byte_no) * BitsPerByte;
197 ubfx(bytecode, bytecode, shift_count, BitsPerByte);
198 }
199
200 void InterpreterMacroAssembler::get_cache_entry_pointer_at_bcp(Register cache,
201 Register tmp,
202 int bcp_offset,
203 size_t index_size) {
204 assert(cache != tmp, "must use different register");
205 get_cache_index_at_bcp(tmp, bcp_offset, index_size);
206 assert(sizeof(ConstantPoolCacheEntry) == 4 * wordSize, "adjust code below");
207 // convert from field index to ConstantPoolCacheEntry index
208 // and from word offset to byte offset
209 assert(exact_log2(in_bytes(ConstantPoolCacheEntry::size_in_bytes())) == 2 + LogBytesPerWord, "else change next line");
210 ldr(cache, Address(rfp, frame::interpreter_frame_cache_offset * wordSize));
211 // skip past the header
212 add(cache, cache, in_bytes(ConstantPoolCache::base_offset()));
213 add(cache, cache, tmp, Assembler::LSL, 2 + LogBytesPerWord); // construct pointer to cache entry
214 }
215
216 void InterpreterMacroAssembler::get_method_counters(Register method,
217 Register mcs, Label& skip) {
218 Label has_counters;
219 ldr(mcs, Address(method, Method::method_counters_offset()));
220 cbnz(mcs, has_counters);
221 call_VM(noreg, CAST_FROM_FN_PTR(address,
222 InterpreterRuntime::build_method_counters), method);
223 ldr(mcs, Address(method, Method::method_counters_offset()));
224 cbz(mcs, skip); // No MethodCounters allocated, OutOfMemory
225 bind(has_counters);
226 }
227
228 // Load object from cpool->resolved_references(index)
229 void InterpreterMacroAssembler::load_resolved_reference_at_index(
230 Register result, Register index) {
231 assert_different_registers(result, index);
232 // convert from field index to resolved_references() index and from
233 // word index to byte offset. Since this is a java object, it can be compressed
234 Register tmp = index; // reuse
235 lslw(tmp, tmp, LogBytesPerHeapOop);
236
237 get_constant_pool(result);
238 // load pointer for resolved_references[] objArray
239 ldr(result, Address(result, ConstantPool::resolved_references_offset_in_bytes()));
240 // JNIHandles::resolve(obj);
241 ldr(result, Address(result, 0));
242 // Add in the index
243 add(result, result, tmp);
244 load_heap_oop(result, Address(result, arrayOopDesc::base_offset_in_bytes(T_OBJECT)));
245 }
246
247 // Generate a subtype check: branch to ok_is_subtype if sub_klass is a
248 // subtype of super_klass.
249 //
250 // Args:
251 // r0: superklass
252 // Rsub_klass: subklass
253 //
254 // Kills:
255 // r2, r5
256 void InterpreterMacroAssembler::gen_subtype_check(Register Rsub_klass,
257 Label& ok_is_subtype) {
258 assert(Rsub_klass != r0, "r0 holds superklass");
259 assert(Rsub_klass != r2, "r2 holds 2ndary super array length");
260 assert(Rsub_klass != r5, "r5 holds 2ndary super array scan ptr");
261
262 // Profile the not-null value's klass.
263 profile_typecheck(r2, Rsub_klass, r5); // blows r2, reloads r5
264
265 // Do the check.
266 check_klass_subtype(Rsub_klass, r0, r2, ok_is_subtype); // blows r2
267
268 // Profile the failure of the check.
269 profile_typecheck_failed(r2); // blows r2
270 }
271
272 // Java Expression Stack
273
274 void InterpreterMacroAssembler::pop_ptr(Register r) {
275 ldr(r, post(esp, wordSize));
276 }
277
278 void InterpreterMacroAssembler::pop_i(Register r) {
279 ldrw(r, post(esp, wordSize));
280 }
281
282 void InterpreterMacroAssembler::pop_l(Register r) {
283 ldr(r, post(esp, 2 * Interpreter::stackElementSize));
284 }
285
286 void InterpreterMacroAssembler::push_ptr(Register r) {
287 str(r, pre(esp, -wordSize));
288 }
289
290 void InterpreterMacroAssembler::push_i(Register r) {
291 str(r, pre(esp, -wordSize));
292 }
293
294 void InterpreterMacroAssembler::push_l(Register r) {
295 str(r, pre(esp, 2 * -wordSize));
296 }
297
298 void InterpreterMacroAssembler::pop_f(FloatRegister r) {
299 ldrs(r, post(esp, wordSize));
300 }
301
302 void InterpreterMacroAssembler::pop_d(FloatRegister r) {
303 ldrd(r, post(esp, 2 * Interpreter::stackElementSize));
304 }
305
306 void InterpreterMacroAssembler::push_f(FloatRegister r) {
307 strs(r, pre(esp, -wordSize));
308 }
309
310 void InterpreterMacroAssembler::push_d(FloatRegister r) {
311 strd(r, pre(esp, 2* -wordSize));
312 }
313
314 void InterpreterMacroAssembler::pop(TosState state) {
315 switch (state) {
316 case atos: pop_ptr(); break;
317 case btos:
318 case ctos:
319 case stos:
320 case itos: pop_i(); break;
321 case ltos: pop_l(); break;
322 case ftos: pop_f(); break;
323 case dtos: pop_d(); break;
324 case vtos: /* nothing to do */ break;
325 default: ShouldNotReachHere();
326 }
327 verify_oop(r0, state);
328 }
329
330 void InterpreterMacroAssembler::push(TosState state) {
331 verify_oop(r0, state);
332 switch (state) {
333 case atos: push_ptr(); break;
334 case btos:
335 case ctos:
336 case stos:
337 case itos: push_i(); break;
338 case ltos: push_l(); break;
339 case ftos: push_f(); break;
340 case dtos: push_d(); break;
341 case vtos: /* nothing to do */ break;
342 default : ShouldNotReachHere();
343 }
344 }
345
346 // Helpers for swap and dup
347 void InterpreterMacroAssembler::load_ptr(int n, Register val) {
348 ldr(val, Address(esp, Interpreter::expr_offset_in_bytes(n)));
349 }
350
351 void InterpreterMacroAssembler::store_ptr(int n, Register val) {
352 str(val, Address(esp, Interpreter::expr_offset_in_bytes(n)));
353 }
354
355
356 void InterpreterMacroAssembler::prepare_to_jump_from_interpreted() {
357 // set sender sp
358 mov(r13, sp);
359 // record last_sp
360 str(esp, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize));
361 }
362
363 // Jump to from_interpreted entry of a call unless single stepping is possible
364 // in this thread in which case we must call the i2i entry
365 void InterpreterMacroAssembler::jump_from_interpreted(Register method, Register temp) {
366 prepare_to_jump_from_interpreted();
367
368 if (JvmtiExport::can_post_interpreter_events()) {
369 Label run_compiled_code;
370 // JVMTI events, such as single-stepping, are implemented partly by avoiding running
371 // compiled code in threads for which the event is enabled. Check here for
372 // interp_only_mode if these events CAN be enabled.
373 // interp_only is an int, on little endian it is sufficient to test the byte only
374 // Is a cmpl faster?
375 ldr(rscratch1, Address(rthread, JavaThread::interp_only_mode_offset()));
376 cbz(rscratch1, run_compiled_code);
377 ldr(rscratch1, Address(method, Method::interpreter_entry_offset()));
378 br(rscratch1);
379 bind(run_compiled_code);
380 }
381
382 ldr(rscratch1, Address(method, Method::from_interpreted_offset()));
383 br(rscratch1);
384 }
385
386 // The following two routines provide a hook so that an implementation
387 // can schedule the dispatch in two parts. amd64 does not do this.
388 void InterpreterMacroAssembler::dispatch_prolog(TosState state, int step) {
389 }
390
391 void InterpreterMacroAssembler::dispatch_epilog(TosState state, int step) {
392 dispatch_next(state, step);
393 }
394
395 void InterpreterMacroAssembler::dispatch_base(TosState state,
396 address* table,
397 bool verifyoop) {
398 if (VerifyActivationFrameSize) {
399 Unimplemented();
400 }
401 if (verifyoop) {
402 verify_oop(r0, state);
403 }
404 if (table == Interpreter::dispatch_table(state)) {
405 addw(rscratch2, rscratch1, Interpreter::distance_from_dispatch_table(state));
406 ldr(rscratch2, Address(rdispatch, rscratch2, Address::uxtw(3)));
407 } else {
408 mov(rscratch2, (address)table);
409 ldr(rscratch2, Address(rscratch2, rscratch1, Address::uxtw(3)));
410 }
411 br(rscratch2);
412 }
413
414 void InterpreterMacroAssembler::dispatch_only(TosState state) {
415 dispatch_base(state, Interpreter::dispatch_table(state));
416 }
417
418 void InterpreterMacroAssembler::dispatch_only_normal(TosState state) {
419 dispatch_base(state, Interpreter::normal_table(state));
420 }
421
422 void InterpreterMacroAssembler::dispatch_only_noverify(TosState state) {
423 dispatch_base(state, Interpreter::normal_table(state), false);
424 }
425
426
427 void InterpreterMacroAssembler::dispatch_next(TosState state, int step) {
428 // load next bytecode
429 ldrb(rscratch1, Address(pre(rbcp, step)));
430 dispatch_base(state, Interpreter::dispatch_table(state));
431 }
432
433 void InterpreterMacroAssembler::dispatch_via(TosState state, address* table) {
434 // load current bytecode
435 ldrb(rscratch1, Address(rbcp, 0));
436 dispatch_base(state, table);
437 }
438
439 // remove activation
440 //
441 // Unlock the receiver if this is a synchronized method.
442 // Unlock any Java monitors from syncronized blocks.
443 // Remove the activation from the stack.
444 //
445 // If there are locked Java monitors
446 // If throw_monitor_exception
447 // throws IllegalMonitorStateException
448 // Else if install_monitor_exception
449 // installs IllegalMonitorStateException
450 // Else
451 // no error processing
452 void InterpreterMacroAssembler::remove_activation(
453 TosState state,
454 bool throw_monitor_exception,
455 bool install_monitor_exception,
456 bool notify_jvmdi) {
457 // Note: Registers r3 xmm0 may be in use for the
458 // result check if synchronized method
459 Label unlocked, unlock, no_unlock;
460
461 // get the value of _do_not_unlock_if_synchronized into r3
462 const Address do_not_unlock_if_synchronized(rthread,
463 in_bytes(JavaThread::do_not_unlock_if_synchronized_offset()));
464 ldrb(r3, do_not_unlock_if_synchronized);
465 strb(zr, do_not_unlock_if_synchronized); // reset the flag
466
467 // get method access flags
468 ldr(r1, Address(rfp, frame::interpreter_frame_method_offset * wordSize));
469 ldr(r2, Address(r1, Method::access_flags_offset()));
470 tst(r2, JVM_ACC_SYNCHRONIZED);
471 br(Assembler::EQ, unlocked);
472
473 // Don't unlock anything if the _do_not_unlock_if_synchronized flag
474 // is set.
475 cbnz(r3, no_unlock);
476
477 // unlock monitor
478 push(state); // save result
479
480 // BasicObjectLock will be first in list, since this is a
481 // synchronized method. However, need to check that the object has
482 // not been unlocked by an explicit monitorexit bytecode.
483 const Address monitor(rfp, frame::interpreter_frame_initial_sp_offset *
484 wordSize - (int) sizeof(BasicObjectLock));
485 // We use c_rarg1 so that if we go slow path it will be the correct
486 // register for unlock_object to pass to VM directly
487 lea(c_rarg1, monitor); // address of first monitor
488
489 ldr(r0, Address(c_rarg1, BasicObjectLock::obj_offset_in_bytes()));
490 cbnz(r0, unlock);
491
492 pop(state);
493 if (throw_monitor_exception) {
494 // Entry already unlocked, need to throw exception
495 call_VM(noreg, CAST_FROM_FN_PTR(address,
496 InterpreterRuntime::throw_illegal_monitor_state_exception));
497 should_not_reach_here();
498 } else {
499 // Monitor already unlocked during a stack unroll. If requested,
500 // install an illegal_monitor_state_exception. Continue with
501 // stack unrolling.
502 if (install_monitor_exception) {
503 call_VM(noreg, CAST_FROM_FN_PTR(address,
504 InterpreterRuntime::new_illegal_monitor_state_exception));
505 }
506 b(unlocked);
507 }
508
509 bind(unlock);
510 unlock_object(c_rarg1);
511 pop(state);
512
513 // Check that for block-structured locking (i.e., that all locked
514 // objects has been unlocked)
515 bind(unlocked);
516
517 // r0: Might contain return value
518
519 // Check that all monitors are unlocked
520 {
521 Label loop, exception, entry, restart;
522 const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
523 const Address monitor_block_top(
524 rfp, frame::interpreter_frame_monitor_block_top_offset * wordSize);
525 const Address monitor_block_bot(
526 rfp, frame::interpreter_frame_initial_sp_offset * wordSize);
527
528 bind(restart);
529 // We use c_rarg1 so that if we go slow path it will be the correct
530 // register for unlock_object to pass to VM directly
531 ldr(c_rarg1, monitor_block_top); // points to current entry, starting
532 // with top-most entry
533 lea(r19, monitor_block_bot); // points to word before bottom of
534 // monitor block
535 b(entry);
536
537 // Entry already locked, need to throw exception
538 bind(exception);
539
540 if (throw_monitor_exception) {
541 // Throw exception
542 MacroAssembler::call_VM(noreg,
543 CAST_FROM_FN_PTR(address, InterpreterRuntime::
544 throw_illegal_monitor_state_exception));
545 should_not_reach_here();
546 } else {
547 // Stack unrolling. Unlock object and install illegal_monitor_exception.
548 // Unlock does not block, so don't have to worry about the frame.
549 // We don't have to preserve c_rarg1 since we are going to throw an exception.
550
551 push(state);
552 unlock_object(c_rarg1);
553 pop(state);
554
555 if (install_monitor_exception) {
556 call_VM(noreg, CAST_FROM_FN_PTR(address,
557 InterpreterRuntime::
558 new_illegal_monitor_state_exception));
559 }
560
561 b(restart);
562 }
563
564 bind(loop);
565 // check if current entry is used
566 ldr(rscratch1, Address(c_rarg1, BasicObjectLock::obj_offset_in_bytes()));
567 cbnz(rscratch1, exception);
568
569 add(c_rarg1, c_rarg1, entry_size); // otherwise advance to next entry
570 bind(entry);
571 cmp(c_rarg1, r19); // check if bottom reached
572 br(Assembler::NE, loop); // if not at bottom then check this entry
573 }
574
575 bind(no_unlock);
576
577 // jvmti support
578 if (notify_jvmdi) {
579 notify_method_exit(state, NotifyJVMTI); // preserve TOSCA
580 } else {
581 notify_method_exit(state, SkipNotifyJVMTI); // preserve TOSCA
582 }
583
584 // remove activation
585 // get sender esp
586 ldr(esp,
587 Address(rfp, frame::interpreter_frame_sender_sp_offset * wordSize));
588 // remove frame anchor
589 leave();
590 // If we're returning to interpreted code we will shortly be
591 // adjusting SP to allow some space for ESP. If we're returning to
592 // compiled code the saved sender SP was saved in sender_sp, so this
593 // restores it.
594 andr(sp, esp, -16);
595 }
596
597 #endif // C_INTERP
598
599 // Lock object
600 //
601 // Args:
602 // c_rarg1: BasicObjectLock to be used for locking
603 //
604 // Kills:
605 // r0
606 // c_rarg0, c_rarg1, c_rarg2, c_rarg3, .. (param regs)
607 // rscratch1, rscratch2 (scratch regs)
608 void InterpreterMacroAssembler::lock_object(Register lock_reg)
609 {
610 assert(lock_reg == c_rarg1, "The argument is only for looks. It must be c_rarg1");
611 if (UseHeavyMonitors) {
612 call_VM(noreg,
613 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter),
614 lock_reg);
615 } else {
616 Label done;
617
618 const Register swap_reg = r0;
619 const Register tmp = c_rarg2;
620 const Register obj_reg = c_rarg3; // Will contain the oop
621
622 const int obj_offset = BasicObjectLock::obj_offset_in_bytes();
623 const int lock_offset = BasicObjectLock::lock_offset_in_bytes ();
624 const int mark_offset = lock_offset +
625 BasicLock::displaced_header_offset_in_bytes();
626
627 Label slow_case;
628
629 // Load object pointer into obj_reg %c_rarg3
630 ldr(obj_reg, Address(lock_reg, obj_offset));
631
632 if (UseBiasedLocking) {
633 biased_locking_enter(lock_reg, obj_reg, swap_reg, tmp, false, done, &slow_case);
634 }
635
636 // Load (object->mark() | 1) into swap_reg
637 ldr(rscratch1, Address(obj_reg, 0));
638 orr(swap_reg, rscratch1, 1);
639
640 // Save (object->mark() | 1) into BasicLock's displaced header
641 str(swap_reg, Address(lock_reg, mark_offset));
642
643 assert(lock_offset == 0,
644 "displached header must be first word in BasicObjectLock");
645
646 Label fail;
647 if (PrintBiasedLockingStatistics) {
648 Label fast;
649 cmpxchgptr(swap_reg, lock_reg, obj_reg, rscratch1, fast, &fail);
650 bind(fast);
651 atomic_incw(Address((address)BiasedLocking::fast_path_entry_count_addr()),
652 rscratch2, rscratch1, tmp);
653 b(done);
654 bind(fail);
655 } else {
656 cmpxchgptr(swap_reg, lock_reg, obj_reg, rscratch1, done, /*fallthrough*/NULL);
657 }
658
659 // Test if the oopMark is an obvious stack pointer, i.e.,
660 // 1) (mark & 7) == 0, and
661 // 2) rsp <= mark < mark + os::pagesize()
662 //
663 // These 3 tests can be done by evaluating the following
664 // expression: ((mark - rsp) & (7 - os::vm_page_size())),
665 // assuming both stack pointer and pagesize have their
666 // least significant 3 bits clear.
667 // NOTE: the oopMark is in swap_reg %r0 as the result of cmpxchg
668 // NOTE2: aarch64 does not like to subtract sp from rn so take a
669 // copy
670 mov(rscratch1, sp);
671 sub(swap_reg, swap_reg, rscratch1);
672 ands(swap_reg, swap_reg, (unsigned long)(7 - os::vm_page_size()));
673
674 // Save the test result, for recursive case, the result is zero
675 str(swap_reg, Address(lock_reg, mark_offset));
676
677 if (PrintBiasedLockingStatistics) {
678 br(Assembler::NE, slow_case);
679 atomic_incw(Address((address)BiasedLocking::fast_path_entry_count_addr()),
680 rscratch2, rscratch1, tmp);
681 }
682 br(Assembler::EQ, done);
683
684 bind(slow_case);
685
686 // Call the runtime routine for slow case
687 call_VM(noreg,
688 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter),
689 lock_reg);
690
691 bind(done);
692 }
693 }
694
695
696 // Unlocks an object. Used in monitorexit bytecode and
697 // remove_activation. Throws an IllegalMonitorException if object is
698 // not locked by current thread.
699 //
700 // Args:
701 // c_rarg1: BasicObjectLock for lock
702 //
703 // Kills:
704 // r0
705 // c_rarg0, c_rarg1, c_rarg2, c_rarg3, ... (param regs)
706 // rscratch1, rscratch2 (scratch regs)
707 void InterpreterMacroAssembler::unlock_object(Register lock_reg)
708 {
709 assert(lock_reg == c_rarg1, "The argument is only for looks. It must be rarg1");
710
711 if (UseHeavyMonitors) {
712 call_VM(noreg,
713 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorexit),
714 lock_reg);
715 } else {
716 Label done;
717
718 const Register swap_reg = r0;
719 const Register header_reg = c_rarg2; // Will contain the old oopMark
720 const Register obj_reg = c_rarg3; // Will contain the oop
721
722 save_bcp(); // Save in case of exception
723
724 // Convert from BasicObjectLock structure to object and BasicLock
725 // structure Store the BasicLock address into %r0
726 lea(swap_reg, Address(lock_reg, BasicObjectLock::lock_offset_in_bytes()));
727
728 // Load oop into obj_reg(%c_rarg3)
729 ldr(obj_reg, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes()));
730
731 // Free entry
732 str(zr, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes()));
733
734 if (UseBiasedLocking) {
735 biased_locking_exit(obj_reg, header_reg, done);
736 }
737
738 // Load the old header from BasicLock structure
739 ldr(header_reg, Address(swap_reg,
740 BasicLock::displaced_header_offset_in_bytes()));
741
742 // Test for recursion
743 cbz(header_reg, done);
744
745 // Atomic swap back the old header
746 cmpxchgptr(swap_reg, header_reg, obj_reg, rscratch1, done, /*fallthrough*/NULL);
747
748 // Call the runtime routine for slow case.
749 str(obj_reg, Address(lock_reg, BasicObjectLock::obj_offset_in_bytes())); // restore obj
750 call_VM(noreg,
751 CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorexit),
752 lock_reg);
753
754 bind(done);
755
756 restore_bcp();
757 }
758 }
759
760 #ifndef CC_INTERP
761
762 void InterpreterMacroAssembler::test_method_data_pointer(Register mdp,
763 Label& zero_continue) {
764 assert(ProfileInterpreter, "must be profiling interpreter");
765 ldr(mdp, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize));
766 cbz(mdp, zero_continue);
767 }
768
769 // Set the method data pointer for the current bcp.
770 void InterpreterMacroAssembler::set_method_data_pointer_for_bcp() {
771 assert(ProfileInterpreter, "must be profiling interpreter");
772 Label set_mdp;
773 stp(r0, r1, Address(pre(sp, -2 * wordSize)));
774
775 // Test MDO to avoid the call if it is NULL.
776 ldr(r0, Address(rmethod, in_bytes(Method::method_data_offset())));
777 cbz(r0, set_mdp);
778 call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::bcp_to_di), rmethod, rbcp);
779 // r0: mdi
780 // mdo is guaranteed to be non-zero here, we checked for it before the call.
781 ldr(r1, Address(rmethod, in_bytes(Method::method_data_offset())));
782 lea(r1, Address(r1, in_bytes(MethodData::data_offset())));
783 add(r0, r1, r0);
784 str(r0, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize));
785 bind(set_mdp);
786 ldp(r0, r1, Address(post(sp, 2 * wordSize)));
787 }
788
789 void InterpreterMacroAssembler::verify_method_data_pointer() {
790 assert(ProfileInterpreter, "must be profiling interpreter");
791 #ifdef ASSERT
792 Label verify_continue;
793 stp(r0, r1, Address(pre(sp, -2 * wordSize)));
794 stp(r2, r3, Address(pre(sp, -2 * wordSize)));
795 test_method_data_pointer(r3, verify_continue); // If mdp is zero, continue
796 get_method(r1);
797
798 // If the mdp is valid, it will point to a DataLayout header which is
799 // consistent with the bcp. The converse is highly probable also.
800 ldrsh(r2, Address(r3, in_bytes(DataLayout::bci_offset())));
801 ldr(rscratch1, Address(r1, Method::const_offset()));
802 add(r2, r2, rscratch1, Assembler::LSL);
803 lea(r2, Address(r2, ConstMethod::codes_offset()));
804 cmp(r2, rbcp);
805 br(Assembler::EQ, verify_continue);
806 // r1: method
807 // rbcp: bcp // rbcp == 22
808 // r3: mdp
809 call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::verify_mdp),
810 r1, rbcp, r3);
811 bind(verify_continue);
812 ldp(r2, r3, Address(post(sp, 2 * wordSize)));
813 ldp(r0, r1, Address(post(sp, 2 * wordSize)));
814 #endif // ASSERT
815 }
816
817
818 void InterpreterMacroAssembler::set_mdp_data_at(Register mdp_in,
819 int constant,
820 Register value) {
821 assert(ProfileInterpreter, "must be profiling interpreter");
822 Address data(mdp_in, constant);
823 str(value, data);
824 }
825
826
827 void InterpreterMacroAssembler::increment_mdp_data_at(Register mdp_in,
828 int constant,
829 bool decrement) {
830 increment_mdp_data_at(mdp_in, noreg, constant, decrement);
831 }
832
833 void InterpreterMacroAssembler::increment_mdp_data_at(Register mdp_in,
834 Register reg,
835 int constant,
836 bool decrement) {
837 assert(ProfileInterpreter, "must be profiling interpreter");
838 // %%% this does 64bit counters at best it is wasting space
839 // at worst it is a rare bug when counters overflow
840
841 assert_different_registers(rscratch2, rscratch1, mdp_in, reg);
842
843 Address addr1(mdp_in, constant);
844 Address addr2(rscratch2, reg, Address::lsl(0));
845 Address &addr = addr1;
846 if (reg != noreg) {
847 lea(rscratch2, addr1);
848 addr = addr2;
849 }
850
851 if (decrement) {
852 // Decrement the register. Set condition codes.
853 // Intel does this
854 // addptr(data, (int32_t) -DataLayout::counter_increment);
855 // If the decrement causes the counter to overflow, stay negative
856 // Label L;
857 // jcc(Assembler::negative, L);
858 // addptr(data, (int32_t) DataLayout::counter_increment);
859 // so we do this
860 ldr(rscratch1, addr);
861 subs(rscratch1, rscratch1, (unsigned)DataLayout::counter_increment);
862 Label L;
863 br(Assembler::LO, L); // skip store if counter underflow
864 str(rscratch1, addr);
865 bind(L);
866 } else {
867 assert(DataLayout::counter_increment == 1,
868 "flow-free idiom only works with 1");
869 // Intel does this
870 // Increment the register. Set carry flag.
871 // addptr(data, DataLayout::counter_increment);
872 // If the increment causes the counter to overflow, pull back by 1.
873 // sbbptr(data, (int32_t)0);
874 // so we do this
875 ldr(rscratch1, addr);
876 adds(rscratch1, rscratch1, DataLayout::counter_increment);
877 Label L;
878 br(Assembler::CS, L); // skip store if counter overflow
879 str(rscratch1, addr);
880 bind(L);
881 }
882 }
883
884 void InterpreterMacroAssembler::set_mdp_flag_at(Register mdp_in,
885 int flag_byte_constant) {
886 assert(ProfileInterpreter, "must be profiling interpreter");
887 int header_offset = in_bytes(DataLayout::header_offset());
888 int header_bits = DataLayout::flag_mask_to_header_mask(flag_byte_constant);
889 // Set the flag
890 ldr(rscratch1, Address(mdp_in, header_offset));
891 orr(rscratch1, rscratch1, header_bits);
892 str(rscratch1, Address(mdp_in, header_offset));
893 }
894
895
896 void InterpreterMacroAssembler::test_mdp_data_at(Register mdp_in,
897 int offset,
898 Register value,
899 Register test_value_out,
900 Label& not_equal_continue) {
901 assert(ProfileInterpreter, "must be profiling interpreter");
902 if (test_value_out == noreg) {
903 ldr(rscratch1, Address(mdp_in, offset));
904 cmp(value, rscratch1);
905 } else {
906 // Put the test value into a register, so caller can use it:
907 ldr(test_value_out, Address(mdp_in, offset));
908 cmp(value, test_value_out);
909 }
910 br(Assembler::NE, not_equal_continue);
911 }
912
913
914 void InterpreterMacroAssembler::update_mdp_by_offset(Register mdp_in,
915 int offset_of_disp) {
916 assert(ProfileInterpreter, "must be profiling interpreter");
917 ldr(rscratch1, Address(mdp_in, offset_of_disp));
918 add(mdp_in, mdp_in, rscratch1, LSL);
919 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize));
920 }
921
922
923 void InterpreterMacroAssembler::update_mdp_by_offset(Register mdp_in,
924 Register reg,
925 int offset_of_disp) {
926 assert(ProfileInterpreter, "must be profiling interpreter");
927 lea(rscratch1, Address(mdp_in, offset_of_disp));
928 ldr(rscratch1, Address(rscratch1, reg, Address::lsl(0)));
929 add(mdp_in, mdp_in, rscratch1, LSL);
930 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize));
931 }
932
933
934 void InterpreterMacroAssembler::update_mdp_by_constant(Register mdp_in,
935 int constant) {
936 assert(ProfileInterpreter, "must be profiling interpreter");
937 add(mdp_in, mdp_in, (unsigned)constant);
938 str(mdp_in, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize));
939 }
940
941
942 void InterpreterMacroAssembler::update_mdp_for_ret(Register return_bci) {
943 assert(ProfileInterpreter, "must be profiling interpreter");
944 // save/restore across call_VM
945 stp(zr, return_bci, Address(pre(sp, -2 * wordSize)));
946 call_VM(noreg,
947 CAST_FROM_FN_PTR(address, InterpreterRuntime::update_mdp_for_ret),
948 return_bci);
949 ldp(zr, return_bci, Address(post(sp, 2 * wordSize)));
950 }
951
952
953 void InterpreterMacroAssembler::profile_taken_branch(Register mdp,
954 Register bumped_count) {
955 if (ProfileInterpreter) {
956 Label profile_continue;
957
958 // If no method data exists, go to profile_continue.
959 // Otherwise, assign to mdp
960 test_method_data_pointer(mdp, profile_continue);
961
962 // We are taking a branch. Increment the taken count.
963 // We inline increment_mdp_data_at to return bumped_count in a register
964 //increment_mdp_data_at(mdp, in_bytes(JumpData::taken_offset()));
965 Address data(mdp, in_bytes(JumpData::taken_offset()));
966 ldr(bumped_count, data);
967 assert(DataLayout::counter_increment == 1,
968 "flow-free idiom only works with 1");
969 // Intel does this to catch overflow
970 // addptr(bumped_count, DataLayout::counter_increment);
971 // sbbptr(bumped_count, 0);
972 // so we do this
973 adds(bumped_count, bumped_count, DataLayout::counter_increment);
974 Label L;
975 br(Assembler::CS, L); // skip store if counter overflow
976 str(bumped_count, data);
977 bind(L);
978 // The method data pointer needs to be updated to reflect the new target.
979 update_mdp_by_offset(mdp, in_bytes(JumpData::displacement_offset()));
980 bind(profile_continue);
981 }
982 }
983
984
985 void InterpreterMacroAssembler::profile_not_taken_branch(Register mdp) {
986 if (ProfileInterpreter) {
987 Label profile_continue;
988
989 // If no method data exists, go to profile_continue.
990 test_method_data_pointer(mdp, profile_continue);
991
992 // We are taking a branch. Increment the not taken count.
993 increment_mdp_data_at(mdp, in_bytes(BranchData::not_taken_offset()));
994
995 // The method data pointer needs to be updated to correspond to
996 // the next bytecode
997 update_mdp_by_constant(mdp, in_bytes(BranchData::branch_data_size()));
998 bind(profile_continue);
999 }
1000 }
1001
1002
1003 void InterpreterMacroAssembler::profile_call(Register mdp) {
1004 if (ProfileInterpreter) {
1005 Label profile_continue;
1006
1007 // If no method data exists, go to profile_continue.
1008 test_method_data_pointer(mdp, profile_continue);
1009
1010 // We are making a call. Increment the count.
1011 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
1012
1013 // The method data pointer needs to be updated to reflect the new target.
1014 update_mdp_by_constant(mdp, in_bytes(CounterData::counter_data_size()));
1015 bind(profile_continue);
1016 }
1017 }
1018
1019 void InterpreterMacroAssembler::profile_final_call(Register mdp) {
1020 if (ProfileInterpreter) {
1021 Label profile_continue;
1022
1023 // If no method data exists, go to profile_continue.
1024 test_method_data_pointer(mdp, profile_continue);
1025
1026 // We are making a call. Increment the count.
1027 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
1028
1029 // The method data pointer needs to be updated to reflect the new target.
1030 update_mdp_by_constant(mdp,
1031 in_bytes(VirtualCallData::
1032 virtual_call_data_size()));
1033 bind(profile_continue);
1034 }
1035 }
1036
1037
1038 void InterpreterMacroAssembler::profile_virtual_call(Register receiver,
1039 Register mdp,
1040 Register reg2,
1041 bool receiver_can_be_null) {
1042 if (ProfileInterpreter) {
1043 Label profile_continue;
1044
1045 // If no method data exists, go to profile_continue.
1046 test_method_data_pointer(mdp, profile_continue);
1047
1048 Label skip_receiver_profile;
1049 if (receiver_can_be_null) {
1050 Label not_null;
1051 // We are making a call. Increment the count for null receiver.
1052 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
1053 b(skip_receiver_profile);
1054 bind(not_null);
1055 }
1056
1057 // Record the receiver type.
1058 record_klass_in_profile(receiver, mdp, reg2, true);
1059 bind(skip_receiver_profile);
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 // This routine creates a state machine for updating the multi-row
1070 // type profile at a virtual call site (or other type-sensitive bytecode).
1071 // The machine visits each row (of receiver/count) until the receiver type
1072 // is found, or until it runs out of rows. At the same time, it remembers
1073 // the location of the first empty row. (An empty row records null for its
1074 // receiver, and can be allocated for a newly-observed receiver type.)
1075 // Because there are two degrees of freedom in the state, a simple linear
1076 // search will not work; it must be a decision tree. Hence this helper
1077 // function is recursive, to generate the required tree structured code.
1078 // It's the interpreter, so we are trading off code space for speed.
1079 // See below for example code.
1080 void InterpreterMacroAssembler::record_klass_in_profile_helper(
1081 Register receiver, Register mdp,
1082 Register reg2, int start_row,
1083 Label& done, bool is_virtual_call) {
1084 if (TypeProfileWidth == 0) {
1085 if (is_virtual_call) {
1086 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
1087 }
1088 return;
1089 }
1090
1091 int last_row = VirtualCallData::row_limit() - 1;
1092 assert(start_row <= last_row, "must be work left to do");
1093 // Test this row for both the receiver and for null.
1094 // Take any of three different outcomes:
1095 // 1. found receiver => increment count and goto done
1096 // 2. found null => keep looking for case 1, maybe allocate this cell
1097 // 3. found something else => keep looking for cases 1 and 2
1098 // Case 3 is handled by a recursive call.
1099 for (int row = start_row; row <= last_row; row++) {
1100 Label next_test;
1101 bool test_for_null_also = (row == start_row);
1102
1103 // See if the receiver is receiver[n].
1104 int recvr_offset = in_bytes(VirtualCallData::receiver_offset(row));
1105 test_mdp_data_at(mdp, recvr_offset, receiver,
1106 (test_for_null_also ? reg2 : noreg),
1107 next_test);
1108 // (Reg2 now contains the receiver from the CallData.)
1109
1110 // The receiver is receiver[n]. Increment count[n].
1111 int count_offset = in_bytes(VirtualCallData::receiver_count_offset(row));
1112 increment_mdp_data_at(mdp, count_offset);
1113 b(done);
1114 bind(next_test);
1115
1116 if (test_for_null_also) {
1117 Label found_null;
1118 // Failed the equality check on receiver[n]... Test for null.
1119 if (start_row == last_row) {
1120 // The only thing left to do is handle the null case.
1121 if (is_virtual_call) {
1122 cbz(reg2, found_null);
1123 // Receiver did not match any saved receiver and there is no empty row for it.
1124 // Increment total counter to indicate polymorphic case.
1125 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
1126 b(done);
1127 bind(found_null);
1128 } else {
1129 cbz(reg2, done);
1130 }
1131 break;
1132 }
1133 // Since null is rare, make it be the branch-taken case.
1134 cbz(reg2,found_null);
1135
1136 // Put all the "Case 3" tests here.
1137 record_klass_in_profile_helper(receiver, mdp, reg2, start_row + 1, done, is_virtual_call);
1138
1139 // Found a null. Keep searching for a matching receiver,
1140 // but remember that this is an empty (unused) slot.
1141 bind(found_null);
1142 }
1143 }
1144
1145 // In the fall-through case, we found no matching receiver, but we
1146 // observed the receiver[start_row] is NULL.
1147
1148 // Fill in the receiver field and increment the count.
1149 int recvr_offset = in_bytes(VirtualCallData::receiver_offset(start_row));
1150 set_mdp_data_at(mdp, recvr_offset, receiver);
1151 int count_offset = in_bytes(VirtualCallData::receiver_count_offset(start_row));
1152 mov(reg2, DataLayout::counter_increment);
1153 set_mdp_data_at(mdp, count_offset, reg2);
1154 if (start_row > 0) {
1155 b(done);
1156 }
1157 }
1158
1159 // Example state machine code for three profile rows:
1160 // // main copy of decision tree, rooted at row[1]
1161 // if (row[0].rec == rec) { row[0].incr(); goto done; }
1162 // if (row[0].rec != NULL) {
1163 // // inner copy of decision tree, rooted at row[1]
1164 // if (row[1].rec == rec) { row[1].incr(); goto done; }
1165 // if (row[1].rec != NULL) {
1166 // // degenerate decision tree, rooted at row[2]
1167 // if (row[2].rec == rec) { row[2].incr(); goto done; }
1168 // if (row[2].rec != NULL) { count.incr(); goto done; } // overflow
1169 // row[2].init(rec); goto done;
1170 // } else {
1171 // // remember row[1] is empty
1172 // if (row[2].rec == rec) { row[2].incr(); goto done; }
1173 // row[1].init(rec); goto done;
1174 // }
1175 // } else {
1176 // // remember row[0] is empty
1177 // if (row[1].rec == rec) { row[1].incr(); goto done; }
1178 // if (row[2].rec == rec) { row[2].incr(); goto done; }
1179 // row[0].init(rec); goto done;
1180 // }
1181 // done:
1182
1183 void InterpreterMacroAssembler::record_klass_in_profile(Register receiver,
1184 Register mdp, Register reg2,
1185 bool is_virtual_call) {
1186 assert(ProfileInterpreter, "must be profiling");
1187 Label done;
1188
1189 record_klass_in_profile_helper(receiver, mdp, reg2, 0, done, is_virtual_call);
1190
1191 bind (done);
1192 }
1193
1194 void InterpreterMacroAssembler::profile_ret(Register return_bci,
1195 Register mdp) {
1196 if (ProfileInterpreter) {
1197 Label profile_continue;
1198 uint row;
1199
1200 // If no method data exists, go to profile_continue.
1201 test_method_data_pointer(mdp, profile_continue);
1202
1203 // Update the total ret count.
1204 increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()));
1205
1206 for (row = 0; row < RetData::row_limit(); row++) {
1207 Label next_test;
1208
1209 // See if return_bci is equal to bci[n]:
1210 test_mdp_data_at(mdp,
1211 in_bytes(RetData::bci_offset(row)),
1212 return_bci, noreg,
1213 next_test);
1214
1215 // return_bci is equal to bci[n]. Increment the count.
1216 increment_mdp_data_at(mdp, in_bytes(RetData::bci_count_offset(row)));
1217
1218 // The method data pointer needs to be updated to reflect the new target.
1219 update_mdp_by_offset(mdp,
1220 in_bytes(RetData::bci_displacement_offset(row)));
1221 b(profile_continue);
1222 bind(next_test);
1223 }
1224
1225 update_mdp_for_ret(return_bci);
1226
1227 bind(profile_continue);
1228 }
1229 }
1230
1231 void InterpreterMacroAssembler::profile_null_seen(Register mdp) {
1232 if (ProfileInterpreter) {
1233 Label profile_continue;
1234
1235 // If no method data exists, go to profile_continue.
1236 test_method_data_pointer(mdp, profile_continue);
1237
1238 set_mdp_flag_at(mdp, BitData::null_seen_byte_constant());
1239
1240 // The method data pointer needs to be updated.
1241 int mdp_delta = in_bytes(BitData::bit_data_size());
1242 if (TypeProfileCasts) {
1243 mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size());
1244 }
1245 update_mdp_by_constant(mdp, mdp_delta);
1246
1247 bind(profile_continue);
1248 }
1249 }
1250
1251 void InterpreterMacroAssembler::profile_typecheck_failed(Register mdp) {
1252 if (ProfileInterpreter && TypeProfileCasts) {
1253 Label profile_continue;
1254
1255 // If no method data exists, go to profile_continue.
1256 test_method_data_pointer(mdp, profile_continue);
1257
1258 int count_offset = in_bytes(CounterData::count_offset());
1259 // Back up the address, since we have already bumped the mdp.
1260 count_offset -= in_bytes(VirtualCallData::virtual_call_data_size());
1261
1262 // *Decrement* the counter. We expect to see zero or small negatives.
1263 increment_mdp_data_at(mdp, count_offset, true);
1264
1265 bind (profile_continue);
1266 }
1267 }
1268
1269 void InterpreterMacroAssembler::profile_typecheck(Register mdp, Register klass, Register reg2) {
1270 if (ProfileInterpreter) {
1271 Label profile_continue;
1272
1273 // If no method data exists, go to profile_continue.
1274 test_method_data_pointer(mdp, profile_continue);
1275
1276 // The method data pointer needs to be updated.
1277 int mdp_delta = in_bytes(BitData::bit_data_size());
1278 if (TypeProfileCasts) {
1279 mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size());
1280
1281 // Record the object type.
1282 record_klass_in_profile(klass, mdp, reg2, false);
1283 }
1284 update_mdp_by_constant(mdp, mdp_delta);
1285
1286 bind(profile_continue);
1287 }
1288 }
1289
1290 void InterpreterMacroAssembler::profile_switch_default(Register mdp) {
1291 if (ProfileInterpreter) {
1292 Label profile_continue;
1293
1294 // If no method data exists, go to profile_continue.
1295 test_method_data_pointer(mdp, profile_continue);
1296
1297 // Update the default case count
1298 increment_mdp_data_at(mdp,
1299 in_bytes(MultiBranchData::default_count_offset()));
1300
1301 // The method data pointer needs to be updated.
1302 update_mdp_by_offset(mdp,
1303 in_bytes(MultiBranchData::
1304 default_displacement_offset()));
1305
1306 bind(profile_continue);
1307 }
1308 }
1309
1310 void InterpreterMacroAssembler::profile_switch_case(Register index,
1311 Register mdp,
1312 Register reg2) {
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 // Build the base (index * per_case_size_in_bytes()) +
1320 // case_array_offset_in_bytes()
1321 movw(reg2, in_bytes(MultiBranchData::per_case_size()));
1322 movw(rscratch1, in_bytes(MultiBranchData::case_array_offset()));
1323 Assembler::maddw(index, index, reg2, rscratch1);
1324
1325 // Update the case count
1326 increment_mdp_data_at(mdp,
1327 index,
1328 in_bytes(MultiBranchData::relative_count_offset()));
1329
1330 // The method data pointer needs to be updated.
1331 update_mdp_by_offset(mdp,
1332 index,
1333 in_bytes(MultiBranchData::
1334 relative_displacement_offset()));
1335
1336 bind(profile_continue);
1337 }
1338 }
1339
1340 void InterpreterMacroAssembler::verify_oop(Register reg, TosState state) {
1341 if (state == atos) {
1342 MacroAssembler::verify_oop(reg);
1343 }
1344 }
1345
1346 void InterpreterMacroAssembler::verify_FPU(int stack_depth, TosState state) { ; }
1347 #endif // !CC_INTERP
1348
1349
1350 void InterpreterMacroAssembler::notify_method_entry() {
1351 // Whenever JVMTI is interp_only_mode, method entry/exit events are sent to
1352 // track stack depth. If it is possible to enter interp_only_mode we add
1353 // the code to check if the event should be sent.
1354 if (JvmtiExport::can_post_interpreter_events()) {
1355 Label L;
1356 ldr(r3, Address(rthread, JavaThread::interp_only_mode_offset()));
1357 tst(r3, ~0);
1358 br(Assembler::EQ, L);
1359 call_VM(noreg, CAST_FROM_FN_PTR(address,
1360 InterpreterRuntime::post_method_entry));
1361 bind(L);
1362 }
1363
1364 {
1365 SkipIfEqual skip(this, &DTraceMethodProbes, false);
1366 get_method(c_rarg1);
1367 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
1368 rthread, c_rarg1);
1369 }
1370
1371 // RedefineClasses() tracing support for obsolete method entry
1372 if (RC_TRACE_IN_RANGE(0x00001000, 0x00002000)) {
1373 get_method(c_rarg1);
1374 call_VM_leaf(
1375 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
1376 rthread, c_rarg1);
1377 }
1378
1379 }
1380
1381
1382 void InterpreterMacroAssembler::notify_method_exit(
1383 TosState state, NotifyMethodExitMode mode) {
1384 // Whenever JVMTI is interp_only_mode, method entry/exit events are sent to
1385 // track stack depth. If it is possible to enter interp_only_mode we add
1386 // the code to check if the event should be sent.
1387 if (mode == NotifyJVMTI && JvmtiExport::can_post_interpreter_events()) {
1388 Label L;
1389 // Note: frame::interpreter_frame_result has a dependency on how the
1390 // method result is saved across the call to post_method_exit. If this
1391 // is changed then the interpreter_frame_result implementation will
1392 // need to be updated too.
1393
1394 // For c++ interpreter the result is always stored at a known location in the frame
1395 // template interpreter will leave it on the top of the stack.
1396 NOT_CC_INTERP(push(state);)
1397 ldrw(r3, Address(rthread, JavaThread::interp_only_mode_offset()));
1398 cbz(r3, L);
1399 call_VM(noreg,
1400 CAST_FROM_FN_PTR(address, InterpreterRuntime::post_method_exit));
1401 bind(L);
1402 NOT_CC_INTERP(pop(state));
1403 }
1404
1405 {
1406 SkipIfEqual skip(this, &DTraceMethodProbes, false);
1407 NOT_CC_INTERP(push(state));
1408 get_method(c_rarg1);
1409 call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
1410 rthread, c_rarg1);
1411 NOT_CC_INTERP(pop(state));
1412 }
1413 }
1414
1415
1416 // Jump if ((*counter_addr += increment) & mask) satisfies the condition.
1417 void InterpreterMacroAssembler::increment_mask_and_jump(Address counter_addr,
1418 int increment, Address mask,
1419 Register scratch, Register scratch2,
1420 bool preloaded, Condition cond,
1421 Label* where) {
1422 if (!preloaded) {
1423 ldrw(scratch, counter_addr);
1424 }
1425 add(scratch, scratch, increment);
1426 strw(scratch, counter_addr);
1427 ldrw(scratch2, mask);
1428 ands(scratch, scratch, scratch2);
1429 br(cond, *where);
1430 }
1431
1432 void InterpreterMacroAssembler::call_VM_leaf_base(address entry_point,
1433 int number_of_arguments) {
1434 // interpreter specific
1435 //
1436 // Note: No need to save/restore rbcp & rlocals pointer since these
1437 // are callee saved registers and no blocking/ GC can happen
1438 // in leaf calls.
1439 #ifdef ASSERT
1440 {
1441 Label L;
1442 ldr(rscratch1, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize));
1443 cbz(rscratch1, L);
1444 stop("InterpreterMacroAssembler::call_VM_leaf_base:"
1445 " last_sp != NULL");
1446 bind(L);
1447 }
1448 #endif /* ASSERT */
1449 // super call
1450 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments);
1451 }
1452
1453 void InterpreterMacroAssembler::call_VM_base(Register oop_result,
1454 Register java_thread,
1455 Register last_java_sp,
1456 address entry_point,
1457 int number_of_arguments,
1458 bool check_exceptions) {
1459 // interpreter specific
1460 //
1461 // Note: Could avoid restoring locals ptr (callee saved) - however doesn't
1462 // really make a difference for these runtime calls, since they are
1463 // slow anyway. Btw., bcp must be saved/restored since it may change
1464 // due to GC.
1465 // assert(java_thread == noreg , "not expecting a precomputed java thread");
1466 save_bcp();
1467 #ifdef ASSERT
1468 {
1469 Label L;
1470 ldr(rscratch1, Address(rfp, frame::interpreter_frame_last_sp_offset * wordSize));
1471 cbz(rscratch1, L);
1472 stop("InterpreterMacroAssembler::call_VM_leaf_base:"
1473 " last_sp != NULL");
1474 bind(L);
1475 }
1476 #endif /* ASSERT */
1477 // super call
1478 MacroAssembler::call_VM_base(oop_result, noreg, last_java_sp,
1479 entry_point, number_of_arguments,
1480 check_exceptions);
1481 // interpreter specific
1482 restore_bcp();
1483 restore_locals();
1484 }
1485
1486 void InterpreterMacroAssembler::profile_obj_type(Register obj, const Address& mdo_addr) {
1487 Label update, next, none;
1488
1489 verify_oop(obj);
1490
1491 cbnz(obj, update);
1492 orptr(mdo_addr, TypeEntries::null_seen);
1493 b(next);
1494
1495 bind(update);
1496 load_klass(obj, obj);
1497
1498 ldr(rscratch1, mdo_addr);
1499 eor(obj, obj, rscratch1);
1500 tst(obj, TypeEntries::type_klass_mask);
1501 br(Assembler::EQ, next); // klass seen before, nothing to
1502 // do. The unknown bit may have been
1503 // set already but no need to check.
1504
1505 tst(obj, TypeEntries::type_unknown);
1506 br(Assembler::NE, next); // already unknown. Nothing to do anymore.
1507
1508 ldr(rscratch1, mdo_addr);
1509 cbz(rscratch1, none);
1510 cmp(rscratch1, TypeEntries::null_seen);
1511 br(Assembler::EQ, none);
1512 // There is a chance that the checks above (re-reading profiling
1513 // data from memory) fail if another thread has just set the
1514 // profiling to this obj's klass
1515 ldr(rscratch1, mdo_addr);
1516 eor(obj, obj, rscratch1);
1517 tst(obj, TypeEntries::type_klass_mask);
1518 br(Assembler::EQ, next);
1519
1520 // different than before. Cannot keep accurate profile.
1521 orptr(mdo_addr, TypeEntries::type_unknown);
1522 b(next);
1523
1524 bind(none);
1525 // first time here. Set profile type.
1526 str(obj, mdo_addr);
1527
1528 bind(next);
1529 }
1530
1531 void InterpreterMacroAssembler::profile_arguments_type(Register mdp, Register callee, Register tmp, bool is_virtual) {
1532 if (!ProfileInterpreter) {
1533 return;
1534 }
1535
1536 if (MethodData::profile_arguments() || MethodData::profile_return()) {
1537 Label profile_continue;
1538
1539 test_method_data_pointer(mdp, profile_continue);
1540
1541 int off_to_start = is_virtual ? in_bytes(VirtualCallData::virtual_call_data_size()) : in_bytes(CounterData::counter_data_size());
1542
1543 ldrb(rscratch1, Address(mdp, in_bytes(DataLayout::tag_offset()) - off_to_start));
1544 cmp(rscratch1, is_virtual ? DataLayout::virtual_call_type_data_tag : DataLayout::call_type_data_tag);
1545 br(Assembler::NE, profile_continue);
1546
1547 if (MethodData::profile_arguments()) {
1548 Label done;
1549 int off_to_args = in_bytes(TypeEntriesAtCall::args_data_offset());
1550
1551 for (int i = 0; i < TypeProfileArgsLimit; i++) {
1552 if (i > 0 || MethodData::profile_return()) {
1553 // If return value type is profiled we may have no argument to profile
1554 ldr(tmp, Address(mdp, in_bytes(TypeEntriesAtCall::cell_count_offset())));
1555 sub(tmp, tmp, i*TypeStackSlotEntries::per_arg_count());
1556 cmp(tmp, TypeStackSlotEntries::per_arg_count());
1557 add(rscratch1, mdp, off_to_args);
1558 br(Assembler::LT, done);
1559 }
1560 ldr(tmp, Address(callee, Method::const_offset()));
1561 load_unsigned_short(tmp, Address(tmp, ConstMethod::size_of_parameters_offset()));
1562 // stack offset o (zero based) from the start of the argument
1563 // list, for n arguments translates into offset n - o - 1 from
1564 // the end of the argument list
1565 ldr(rscratch1, Address(mdp, in_bytes(TypeEntriesAtCall::stack_slot_offset(i))));
1566 sub(tmp, tmp, rscratch1);
1567 sub(tmp, tmp, 1);
1568 Address arg_addr = argument_address(tmp);
1569 ldr(tmp, arg_addr);
1570
1571 Address mdo_arg_addr(mdp, in_bytes(TypeEntriesAtCall::argument_type_offset(i)));
1572 profile_obj_type(tmp, mdo_arg_addr);
1573
1574 int to_add = in_bytes(TypeStackSlotEntries::per_arg_size());
1575 off_to_args += to_add;
1576 }
1577
1578 if (MethodData::profile_return()) {
1579 ldr(tmp, Address(mdp, in_bytes(TypeEntriesAtCall::cell_count_offset())));
1580 sub(tmp, tmp, TypeProfileArgsLimit*TypeStackSlotEntries::per_arg_count());
1581 }
1582
1583 add(rscratch1, mdp, off_to_args);
1584 bind(done);
1585 mov(mdp, rscratch1);
1586
1587 if (MethodData::profile_return()) {
1588 // We're right after the type profile for the last
1589 // argument. tmp is the number of cells left in the
1590 // CallTypeData/VirtualCallTypeData to reach its end. Non null
1591 // if there's a return to profile.
1592 assert(ReturnTypeEntry::static_cell_count() < TypeStackSlotEntries::per_arg_count(), "can't move past ret type");
1593 add(mdp, mdp, tmp, LSL, exact_log2(DataLayout::cell_size));
1594 }
1595 str(mdp, Address(rfp, frame::interpreter_frame_mdp_offset * wordSize));
1596 } else {
1597 assert(MethodData::profile_return(), "either profile call args or call ret");
1598 update_mdp_by_constant(mdp, in_bytes(TypeEntriesAtCall::return_only_size()));
1599 }
1600
1601 // mdp points right after the end of the
1602 // CallTypeData/VirtualCallTypeData, right after the cells for the
1603 // return value type if there's one
1604
1605 bind(profile_continue);
1606 }
1607 }
1608
1609 void InterpreterMacroAssembler::profile_return_type(Register mdp, Register ret, Register tmp) {
1610 assert_different_registers(mdp, ret, tmp, rbcp);
1611 if (ProfileInterpreter && MethodData::profile_return()) {
1612 Label profile_continue, done;
1613
1614 test_method_data_pointer(mdp, profile_continue);
1615
1616 if (MethodData::profile_return_jsr292_only()) {
1617 assert(Method::intrinsic_id_size_in_bytes() == 2, "assuming Method::_intrinsic_id is u2");
1618
1619 // If we don't profile all invoke bytecodes we must make sure
1620 // it's a bytecode we indeed profile. We can't go back to the
1621 // begining of the ProfileData we intend to update to check its
1622 // type because we're right after it and we don't known its
1623 // length
1624 Label do_profile;
1625 ldrb(rscratch1, Address(rbcp, 0));
1626 cmp(rscratch1, Bytecodes::_invokedynamic);
1627 br(Assembler::EQ, do_profile);
1628 cmp(rscratch1, Bytecodes::_invokehandle);
1629 br(Assembler::EQ, do_profile);
1630 get_method(tmp);
1631 ldrh(rscratch1, Address(tmp, Method::intrinsic_id_offset_in_bytes()));
1632 cmp(rscratch1, vmIntrinsics::_compiledLambdaForm);
1633 br(Assembler::NE, profile_continue);
1634
1635 bind(do_profile);
1636 }
1637
1638 Address mdo_ret_addr(mdp, -in_bytes(ReturnTypeEntry::size()));
1639 mov(tmp, ret);
1640 profile_obj_type(tmp, mdo_ret_addr);
1641
1642 bind(profile_continue);
1643 }
1644 }
1645
1646 void InterpreterMacroAssembler::profile_parameters_type(Register mdp, Register tmp1, Register tmp2) {
1647 if (ProfileInterpreter && MethodData::profile_parameters()) {
1648 Label profile_continue, done;
1649
1650 test_method_data_pointer(mdp, profile_continue);
1651
1652 // Load the offset of the area within the MDO used for
1653 // parameters. If it's negative we're not profiling any parameters
1654 ldr(tmp1, Address(mdp, in_bytes(MethodData::parameters_type_data_di_offset()) - in_bytes(MethodData::data_offset())));
1655 cmp(tmp1, 0u);
1656 br(Assembler::LT, profile_continue);
1657
1658 // Compute a pointer to the area for parameters from the offset
1659 // and move the pointer to the slot for the last
1660 // parameters. Collect profiling from last parameter down.
1661 // mdo start + parameters offset + array length - 1
1662 add(mdp, mdp, tmp1);
1663 ldr(tmp1, Address(mdp, ArrayData::array_len_offset()));
1664 sub(tmp1, tmp1, TypeStackSlotEntries::per_arg_count());
1665
1666 Label loop;
1667 bind(loop);
1668
1669 int off_base = in_bytes(ParametersTypeData::stack_slot_offset(0));
1670 int type_base = in_bytes(ParametersTypeData::type_offset(0));
1671 int per_arg_scale = exact_log2(DataLayout::cell_size);
1672 add(rscratch1, mdp, off_base);
1673 add(rscratch2, mdp, type_base);
1674
1675 Address arg_off(rscratch1, tmp1, Address::lsl(per_arg_scale));
1676 Address arg_type(rscratch2, tmp1, Address::lsl(per_arg_scale));
1677
1678 // load offset on the stack from the slot for this parameter
1679 ldr(tmp2, arg_off);
1680 neg(tmp2, tmp2);
1681 // read the parameter from the local area
1682 ldr(tmp2, Address(rlocals, tmp2, Address::lsl(Interpreter::logStackElementSize)));
1683
1684 // profile the parameter
1685 profile_obj_type(tmp2, arg_type);
1686
1687 // go to next parameter
1688 subs(tmp1, tmp1, TypeStackSlotEntries::per_arg_count());
1689 br(Assembler::GE, loop);
1690
1691 bind(profile_continue);
1692 }
1693 }