1 /*
2 * Copyright (c) 2018, 2019, Red Hat, Inc. All rights reserved.
3 *
4 * This code is free software; you can redistribute it and/or modify it
5 * under the terms of the GNU General Public License version 2 only, as
6 * published by the Free Software Foundation.
7 *
8 * This code is distributed in the hope that it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
11 * version 2 for more details (a copy is included in the LICENSE file that
12 * accompanied this code).
13 *
14 * You should have received a copy of the GNU General Public License version
15 * 2 along with this work; if not, write to the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
17 *
18 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
19 * or visit www.oracle.com if you need additional information or have any
20 * questions.
21 *
22 */
23
24 #include "precompiled.hpp"
25 #include "gc/shenandoah/shenandoahBarrierSet.hpp"
26 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
27 #include "gc/shenandoah/shenandoahForwarding.hpp"
28 #include "gc/shenandoah/shenandoahHeap.hpp"
29 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
30 #include "gc/shenandoah/shenandoahRuntime.hpp"
31 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
32 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
33 #include "interpreter/interpreter.hpp"
34 #include "interpreter/interp_masm.hpp"
35 #include "runtime/sharedRuntime.hpp"
36 #include "runtime/thread.hpp"
37 #include "utilities/macros.hpp"
38 #ifdef COMPILER1
39 #include "c1/c1_LIRAssembler.hpp"
40 #include "c1/c1_MacroAssembler.hpp"
41 #include "gc/shenandoah/c1/shenandoahBarrierSetC1.hpp"
42 #endif
43
44 #define __ masm->
45
46 address ShenandoahBarrierSetAssembler::_shenandoah_lrb = NULL;
47
48 void ShenandoahBarrierSetAssembler::arraycopy_prologue(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
49 Register src, Register dst, Register count) {
50
51 bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0;
52
53 if (type == T_OBJECT || type == T_ARRAY) {
54
55 if ((ShenandoahSATBBarrier && !dest_uninitialized) || ShenandoahStoreValEnqueueBarrier || ShenandoahLoadRefBarrier) {
56 #ifdef _LP64
57 Register thread = r15_thread;
58 #else
59 Register thread = rax;
60 if (thread == src || thread == dst || thread == count) {
61 thread = rbx;
62 }
63 if (thread == src || thread == dst || thread == count) {
64 thread = rcx;
65 }
66 if (thread == src || thread == dst || thread == count) {
67 thread = rdx;
68 }
69 __ push(thread);
70 __ get_thread(thread);
71 #endif
72 assert_different_registers(src, dst, count, thread);
73
74 Label done;
75 // Short-circuit if count == 0.
76 __ testptr(count, count);
77 __ jcc(Assembler::zero, done);
78
79 // Avoid runtime call when not active.
80 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
81 int flags;
82 if (ShenandoahSATBBarrier && dest_uninitialized) {
83 flags = ShenandoahHeap::HAS_FORWARDED;
84 } else {
85 flags = ShenandoahHeap::HAS_FORWARDED | ShenandoahHeap::MARKING;
86 }
87 __ testb(gc_state, flags);
88 __ jcc(Assembler::zero, done);
89
90 __ pusha(); // push registers
91
92 #ifdef _LP64
93 assert(src == rdi, "expected");
94 assert(dst == rsi, "expected");
95 assert(count == rdx, "expected");
96 if (UseCompressedOops) {
97 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::arraycopy_barrier_narrow_oop_entry),
98 src, dst, count);
99 } else
100 #endif
101 {
102 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::arraycopy_barrier_oop_entry),
103 src, dst, count);
104 }
105
106 __ popa();
107 __ bind(done);
108 NOT_LP64(__ pop(thread);)
109 }
110 }
111
112 }
113
114 void ShenandoahBarrierSetAssembler::shenandoah_write_barrier_pre(MacroAssembler* masm,
115 Register obj,
116 Register pre_val,
117 Register thread,
118 Register tmp,
119 bool tosca_live,
120 bool expand_call) {
121
122 if (ShenandoahSATBBarrier) {
123 satb_write_barrier_pre(masm, obj, pre_val, thread, tmp, tosca_live, expand_call);
124 }
125 }
126
127 void ShenandoahBarrierSetAssembler::satb_write_barrier_pre(MacroAssembler* masm,
128 Register obj,
129 Register pre_val,
130 Register thread,
131 Register tmp,
132 bool tosca_live,
133 bool expand_call) {
134 // If expand_call is true then we expand the call_VM_leaf macro
135 // directly to skip generating the check by
136 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
137
138 #ifdef _LP64
139 assert(thread == r15_thread, "must be");
140 #endif // _LP64
141
142 Label done;
143 Label runtime;
144
145 assert(pre_val != noreg, "check this code");
146
147 if (obj != noreg) {
148 assert_different_registers(obj, pre_val, tmp);
149 assert(pre_val != rax, "check this code");
150 }
151
152 Address in_progress(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()));
153 Address index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
154 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
155
156 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
157 __ testb(gc_state, ShenandoahHeap::MARKING);
158 __ jcc(Assembler::zero, done);
159
160 // Do we need to load the previous value?
161 if (obj != noreg) {
162 __ load_heap_oop(pre_val, Address(obj, 0), noreg, noreg, AS_RAW);
163 }
164
165 // Is the previous value null?
166 __ cmpptr(pre_val, (int32_t) NULL_WORD);
167 __ jcc(Assembler::equal, done);
168
169 // Can we store original value in the thread's buffer?
170 // Is index == 0?
171 // (The index field is typed as size_t.)
172
173 __ movptr(tmp, index); // tmp := *index_adr
174 __ cmpptr(tmp, 0); // tmp == 0?
175 __ jcc(Assembler::equal, runtime); // If yes, goto runtime
176
177 __ subptr(tmp, wordSize); // tmp := tmp - wordSize
178 __ movptr(index, tmp); // *index_adr := tmp
179 __ addptr(tmp, buffer); // tmp := tmp + *buffer_adr
180
181 // Record the previous value
182 __ movptr(Address(tmp, 0), pre_val);
183 __ jmp(done);
184
185 __ bind(runtime);
186 // save the live input values
187 if(tosca_live) __ push(rax);
188
189 if (obj != noreg && obj != rax)
190 __ push(obj);
191
192 if (pre_val != rax)
193 __ push(pre_val);
194
195 // Calling the runtime using the regular call_VM_leaf mechanism generates
196 // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
197 // that checks that the *(ebp+frame::interpreter_frame_last_sp) == NULL.
198 //
199 // If we care generating the pre-barrier without a frame (e.g. in the
200 // intrinsified Reference.get() routine) then ebp might be pointing to
201 // the caller frame and so this check will most likely fail at runtime.
202 //
203 // Expanding the call directly bypasses the generation of the check.
204 // So when we do not have have a full interpreter frame on the stack
205 // expand_call should be passed true.
206
207 NOT_LP64( __ push(thread); )
208
209 #ifdef _LP64
210 // We move pre_val into c_rarg0 early, in order to avoid smashing it, should
211 // pre_val be c_rarg1 (where the call prologue would copy thread argument).
212 // Note: this should not accidentally smash thread, because thread is always r15.
213 assert(thread != c_rarg0, "smashed arg");
214 if (c_rarg0 != pre_val) {
215 __ mov(c_rarg0, pre_val);
216 }
217 #endif
218
219 if (expand_call) {
220 LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); )
221 #ifdef _LP64
222 if (c_rarg1 != thread) {
223 __ mov(c_rarg1, thread);
224 }
225 // Already moved pre_val into c_rarg0 above
226 #else
227 __ push(thread);
228 __ push(pre_val);
229 #endif
230 __ MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), 2);
231 } else {
232 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), LP64_ONLY(c_rarg0) NOT_LP64(pre_val), thread);
233 }
234
235 NOT_LP64( __ pop(thread); )
236
237 // save the live input values
238 if (pre_val != rax)
239 __ pop(pre_val);
240
241 if (obj != noreg && obj != rax)
242 __ pop(obj);
243
244 if(tosca_live) __ pop(rax);
245
246 __ bind(done);
247 }
248
249 void ShenandoahBarrierSetAssembler::load_reference_barrier_not_null(MacroAssembler* masm, Register dst, Address src) {
250 assert(ShenandoahLoadRefBarrier, "Should be enabled");
251
252 Label done;
253
254 #ifdef _LP64
255 Register thread = r15_thread;
256 #else
257 Register thread = rcx;
258 if (thread == dst) {
259 thread = rbx;
260 }
261 __ push(thread);
262 __ get_thread(thread);
263 #endif
264
265 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
266 __ testb(gc_state, ShenandoahHeap::HAS_FORWARDED);
267 __ jccb(Assembler::zero, done);
268
269 // Use rsi for src address
270 const Register src_addr = rsi;
271 // Setup address parameter first, if it does not clobber oop in dst
272 bool need_addr_setup = (src_addr != dst);
273
274 if (need_addr_setup) {
275 __ push(src_addr);
276 __ lea(src_addr, src);
277
278 if (dst != rax) {
279 // Move obj into rax and save rax
280 __ push(rax);
281 __ movptr(rax, dst);
282 }
283 } else {
284 // dst == rsi
285 __ push(rax);
286 __ movptr(rax, dst);
287
288 // we can clobber it, since it is outgoing register
289 __ lea(src_addr, src);
290 }
291
292 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, ShenandoahBarrierSetAssembler::shenandoah_lrb())));
293
294 if (need_addr_setup) {
295 if (dst != rax) {
296 __ movptr(dst, rax);
297 __ pop(rax);
298 }
299 __ pop(src_addr);
300 } else {
301 __ movptr(dst, rax);
302 __ pop(rax);
303 }
304
305 __ bind(done);
306
307 #ifndef _LP64
308 __ pop(thread);
309 #endif
310 }
311
312 void ShenandoahBarrierSetAssembler::storeval_barrier(MacroAssembler* masm, Register dst, Register tmp) {
313 if (ShenandoahStoreValEnqueueBarrier) {
314 storeval_barrier_impl(masm, dst, tmp);
315 }
316 }
317
318 void ShenandoahBarrierSetAssembler::storeval_barrier_impl(MacroAssembler* masm, Register dst, Register tmp) {
319 assert(ShenandoahStoreValEnqueueBarrier, "should be enabled");
320
321 if (dst == noreg) return;
322
323 if (ShenandoahStoreValEnqueueBarrier) {
324 // The set of registers to be saved+restored is the same as in the write-barrier above.
325 // Those are the commonly used registers in the interpreter.
326 __ pusha();
327 // __ push_callee_saved_registers();
328 __ subptr(rsp, 2 * Interpreter::stackElementSize);
329 __ movdbl(Address(rsp, 0), xmm0);
330
331 #ifdef _LP64
332 Register thread = r15_thread;
333 #else
334 Register thread = rcx;
335 if (thread == dst || thread == tmp) {
336 thread = rdi;
337 }
338 if (thread == dst || thread == tmp) {
339 thread = rbx;
340 }
341 __ get_thread(thread);
342 #endif
343 assert_different_registers(dst, tmp, thread);
344
345 satb_write_barrier_pre(masm, noreg, dst, thread, tmp, true, false);
346 __ movdbl(xmm0, Address(rsp, 0));
347 __ addptr(rsp, 2 * Interpreter::stackElementSize);
348 //__ pop_callee_saved_registers();
349 __ popa();
350 }
351 }
352
353 void ShenandoahBarrierSetAssembler::load_reference_barrier(MacroAssembler* masm, Register dst, Address src) {
354 if (ShenandoahLoadRefBarrier) {
355 Label done;
356 __ testptr(dst, dst);
357 __ jcc(Assembler::zero, done);
358 load_reference_barrier_not_null(masm, dst, src);
359 __ bind(done);
360 }
361 }
362
363 //
364 // Arguments:
365 //
366 // Inputs:
367 // src: oop location, might be clobbered
368 // tmp1: scratch register, might not be valid.
369 //
370 // Output:
371 // dst: oop loaded from src location
372 //
373 // Kill:
374 // tmp1 (if it is valid)
375 //
376 void ShenandoahBarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
377 Register dst, Address src, Register tmp1, Register tmp_thread) {
378 // 1: non-reference load, no additional barrier is needed
379 if (!is_reference_type(type)) {
380 BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp_thread);
381 return;
382 }
383
384 assert((decorators & ON_UNKNOWN_OOP_REF) == 0, "Not expected");
385
386 // 2: load a reference from src location and apply LRB if needed
387 if (ShenandoahBarrierSet::need_load_reference_barrier(decorators, type)) {
388 Register result_dst = dst;
389 bool use_tmp1_for_dst = false;
390
391 // Preserve src location for LRB
392 if (dst == src.base() || dst == src.index()) {
393 // Use tmp1 for dst if possible, as it is not used in BarrierAssembler::load_at()
394 if (tmp1->is_valid() && tmp1 != src.base() && tmp1 != src.index()) {
395 dst = tmp1;
396 use_tmp1_for_dst = true;
397 } else {
398 dst = rdi;
399 __ push(dst);
400 }
401 assert_different_registers(dst, src.base(), src.index());
402 }
403
404 BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp_thread);
405
406 load_reference_barrier(masm, dst, src);
407
408 // Move loaded oop to final destination
409 if (dst != result_dst) {
410 __ movptr(result_dst, dst);
411
412 if (!use_tmp1_for_dst) {
413 __ pop(dst);
414 }
415
416 dst = result_dst;
417 }
418 } else {
419 BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp_thread);
420 }
421
422 // 3: apply keep-alive barrier if needed
423 if (ShenandoahBarrierSet::need_keep_alive_barrier(decorators, type)) {
424 __ push_IU_state();
425 // That path can be reached from the c2i adapter with live fp
426 // arguments in registers.
427 LP64_ONLY(assert(Argument::n_float_register_parameters_j == 8, "8 fp registers to save at java call"));
428 __ subptr(rsp, 64);
429 __ movdbl(Address(rsp, 0), xmm0);
430 __ movdbl(Address(rsp, 8), xmm1);
431 __ movdbl(Address(rsp, 16), xmm2);
432 __ movdbl(Address(rsp, 24), xmm3);
433 __ movdbl(Address(rsp, 32), xmm4);
434 __ movdbl(Address(rsp, 40), xmm5);
435 __ movdbl(Address(rsp, 48), xmm6);
436 __ movdbl(Address(rsp, 56), xmm7);
437
438 Register thread = NOT_LP64(tmp_thread) LP64_ONLY(r15_thread);
439 assert_different_registers(dst, tmp1, tmp_thread);
440 if (!thread->is_valid()) {
441 thread = rdx;
442 }
443 NOT_LP64(__ get_thread(thread));
444 // Generate the SATB pre-barrier code to log the value of
445 // the referent field in an SATB buffer.
446 shenandoah_write_barrier_pre(masm /* masm */,
447 noreg /* obj */,
448 dst /* pre_val */,
449 thread /* thread */,
450 tmp1 /* tmp */,
451 true /* tosca_live */,
452 true /* expand_call */);
453 __ movdbl(xmm0, Address(rsp, 0));
454 __ movdbl(xmm1, Address(rsp, 8));
455 __ movdbl(xmm2, Address(rsp, 16));
456 __ movdbl(xmm3, Address(rsp, 24));
457 __ movdbl(xmm4, Address(rsp, 32));
458 __ movdbl(xmm5, Address(rsp, 40));
459 __ movdbl(xmm6, Address(rsp, 48));
460 __ movdbl(xmm7, Address(rsp, 56));
461 __ addptr(rsp, 64);
462 __ pop_IU_state();
463 }
464 }
465
466 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
467 Address dst, Register val, Register tmp1, Register tmp2) {
468
469 bool on_oop = type == T_OBJECT || type == T_ARRAY;
470 bool in_heap = (decorators & IN_HEAP) != 0;
471 bool as_normal = (decorators & AS_NORMAL) != 0;
472 if (on_oop && in_heap) {
473 bool needs_pre_barrier = as_normal;
474
475 Register tmp3 = LP64_ONLY(r8) NOT_LP64(rsi);
476 Register rthread = LP64_ONLY(r15_thread) NOT_LP64(rcx);
477 // flatten object address if needed
478 // We do it regardless of precise because we need the registers
479 if (dst.index() == noreg && dst.disp() == 0) {
480 if (dst.base() != tmp1) {
481 __ movptr(tmp1, dst.base());
482 }
483 } else {
484 __ lea(tmp1, dst);
485 }
486
487 assert_different_registers(val, tmp1, tmp2, tmp3, rthread);
488
489 #ifndef _LP64
490 __ get_thread(rthread);
491 InterpreterMacroAssembler *imasm = static_cast<InterpreterMacroAssembler*>(masm);
492 imasm->save_bcp();
493 #endif
494
495 if (needs_pre_barrier) {
496 shenandoah_write_barrier_pre(masm /*masm*/,
497 tmp1 /* obj */,
498 tmp2 /* pre_val */,
499 rthread /* thread */,
500 tmp3 /* tmp */,
501 val != noreg /* tosca_live */,
502 false /* expand_call */);
503 }
504 if (val == noreg) {
505 BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
506 } else {
507 storeval_barrier(masm, val, tmp3);
508 BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
509 }
510 NOT_LP64(imasm->restore_bcp());
511 } else {
512 BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2);
513 }
514 }
515
516 void ShenandoahBarrierSetAssembler::try_resolve_jobject_in_native(MacroAssembler* masm, Register jni_env,
517 Register obj, Register tmp, Label& slowpath) {
518 Label done;
519 // Resolve jobject
520 BarrierSetAssembler::try_resolve_jobject_in_native(masm, jni_env, obj, tmp, slowpath);
521
522 // Check for null.
523 __ testptr(obj, obj);
524 __ jcc(Assembler::zero, done);
525
526 Address gc_state(jni_env, ShenandoahThreadLocalData::gc_state_offset() - JavaThread::jni_environment_offset());
527 __ testb(gc_state, ShenandoahHeap::EVACUATION);
528 __ jccb(Assembler::notZero, slowpath);
529 __ bind(done);
530 }
531
532 // Special Shenandoah CAS implementation that handles false negatives
533 // due to concurrent evacuation.
534 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm,
535 Register res, Address addr, Register oldval, Register newval,
536 bool exchange, Register tmp1, Register tmp2) {
537 assert(ShenandoahCASBarrier, "Should only be used when CAS barrier is enabled");
538 assert(oldval == rax, "must be in rax for implicit use in cmpxchg");
539 assert_different_registers(oldval, newval, tmp1, tmp2);
540
541 Label L_success, L_failure;
542
543 // Remember oldval for retry logic below
544 #ifdef _LP64
545 if (UseCompressedOops) {
546 __ movl(tmp1, oldval);
547 } else
548 #endif
549 {
550 __ movptr(tmp1, oldval);
551 }
552
553 // Step 1. Fast-path.
554 //
555 // Try to CAS with given arguments. If successful, then we are done.
556
557 if (os::is_MP()) __ lock();
558 #ifdef _LP64
559 if (UseCompressedOops) {
560 __ cmpxchgl(newval, addr);
561 } else
562 #endif
563 {
564 __ cmpxchgptr(newval, addr);
565 }
566 __ jcc(Assembler::equal, L_success);
567
568 // Step 2. CAS had failed. This may be a false negative.
569 //
570 // The trouble comes when we compare the to-space pointer with the from-space
571 // pointer to the same object. To resolve this, it will suffice to resolve
572 // the value from memory -- this will give both to-space pointers.
573 // If they mismatch, then it was a legitimate failure.
574 //
575 // Before reaching to resolve sequence, see if we can avoid the whole shebang
576 // with filters.
577
578 // Filter: when offending in-memory value is NULL, the failure is definitely legitimate
579 __ testptr(oldval, oldval);
580 __ jcc(Assembler::zero, L_failure);
581
582 // Filter: when heap is stable, the failure is definitely legitimate
583 #ifdef _LP64
584 const Register thread = r15_thread;
585 #else
586 const Register thread = tmp2;
587 __ get_thread(thread);
588 #endif
589 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
590 __ testb(gc_state, ShenandoahHeap::HAS_FORWARDED);
591 __ jcc(Assembler::zero, L_failure);
592
593 #ifdef _LP64
594 if (UseCompressedOops) {
595 __ movl(tmp2, oldval);
596 __ decode_heap_oop(tmp2);
597 } else
598 #endif
599 {
600 __ movptr(tmp2, oldval);
601 }
602
603 // Decode offending in-memory value.
604 // Test if-forwarded
605 __ testb(Address(tmp2, oopDesc::mark_offset_in_bytes()), markOopDesc::marked_value);
606 __ jcc(Assembler::noParity, L_failure); // When odd number of bits, then not forwarded
607 __ jcc(Assembler::zero, L_failure); // When it is 00, then also not forwarded
608
609 // Load and mask forwarding pointer
610 __ movptr(tmp2, Address(tmp2, oopDesc::mark_offset_in_bytes()));
611 __ shrptr(tmp2, 2);
612 __ shlptr(tmp2, 2);
613
614 #ifdef _LP64
615 if (UseCompressedOops) {
616 __ decode_heap_oop(tmp1); // decode for comparison
617 }
618 #endif
619
620 // Now we have the forwarded offender in tmp2.
621 // Compare and if they don't match, we have legitimate failure
622 __ cmpptr(tmp1, tmp2);
623 __ jcc(Assembler::notEqual, L_failure);
624
625 // Step 3. Need to fix the memory ptr before continuing.
626 //
627 // At this point, we have from-space oldval in the register, and its to-space
628 // address is in tmp2. Let's try to update it into memory. We don't care if it
629 // succeeds or not. If it does, then the retrying CAS would see it and succeed.
630 // If this fixup fails, this means somebody else beat us to it, and necessarily
631 // with to-space ptr store. We still have to do the retry, because the GC might
632 // have updated the reference for us.
633
634 #ifdef _LP64
635 if (UseCompressedOops) {
636 __ encode_heap_oop(tmp2); // previously decoded at step 2.
637 }
638 #endif
639
640 if (os::is_MP()) __ lock();
641 #ifdef _LP64
642 if (UseCompressedOops) {
643 __ cmpxchgl(tmp2, addr);
644 } else
645 #endif
646 {
647 __ cmpxchgptr(tmp2, addr);
648 }
649
650 // Step 4. Try to CAS again.
651 //
652 // This is guaranteed not to have false negatives, because oldval is definitely
653 // to-space, and memory pointer is to-space as well. Nothing is able to store
654 // from-space ptr into memory anymore. Make sure oldval is restored, after being
655 // garbled during retries.
656 //
657 #ifdef _LP64
658 if (UseCompressedOops) {
659 __ movl(oldval, tmp2);
660 } else
661 #endif
662 {
663 __ movptr(oldval, tmp2);
664 }
665
666 if (os::is_MP()) __ lock();
667 #ifdef _LP64
668 if (UseCompressedOops) {
669 __ cmpxchgl(newval, addr);
670 } else
671 #endif
672 {
673 __ cmpxchgptr(newval, addr);
674 }
675 if (!exchange) {
676 __ jccb(Assembler::equal, L_success); // fastpath, peeking into Step 5, no need to jump
677 }
678
679 // Step 5. If we need a boolean result out of CAS, set the flag appropriately.
680 // and promote the result. Note that we handle the flag from both the 1st and 2nd CAS.
681 // Otherwise, failure witness for CAE is in oldval on all paths, and we can return.
682
683 if (exchange) {
684 __ bind(L_failure);
685 __ bind(L_success);
686 } else {
687 assert(res != NULL, "need result register");
688
689 Label exit;
690 __ bind(L_failure);
691 __ xorptr(res, res);
692 __ jmpb(exit);
693
694 __ bind(L_success);
695 __ movptr(res, 1);
696 __ bind(exit);
697 }
698 }
699
700 #undef __
701
702 #ifdef COMPILER1
703
704 #define __ ce->masm()->
705
706 void ShenandoahBarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, ShenandoahPreBarrierStub* stub) {
707 ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
708 // At this point we know that marking is in progress.
709 // If do_load() is true then we have to emit the
710 // load of the previous value; otherwise it has already
711 // been loaded into _pre_val.
712
713 __ bind(*stub->entry());
714 assert(stub->pre_val()->is_register(), "Precondition.");
715
716 Register pre_val_reg = stub->pre_val()->as_register();
717
718 if (stub->do_load()) {
719 ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/, false /*unaligned*/);
720 }
721
722 __ cmpptr(pre_val_reg, (int32_t)NULL_WORD);
723 __ jcc(Assembler::equal, *stub->continuation());
724 ce->store_parameter(stub->pre_val()->as_register(), 0);
725 __ call(RuntimeAddress(bs->pre_barrier_c1_runtime_code_blob()->code_begin()));
726 __ jmp(*stub->continuation());
727
728 }
729
730 void ShenandoahBarrierSetAssembler::gen_load_reference_barrier_stub(LIR_Assembler* ce, ShenandoahLoadReferenceBarrierStub* stub) {
731 ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
732 __ bind(*stub->entry());
733
734 Register obj = stub->obj()->as_register();
735 Register res = stub->result()->as_register();
736 Register addr = stub->addr()->as_pointer_register();
737 Register tmp1 = stub->tmp1()->as_register();
738 Register tmp2 = stub->tmp2()->as_register();
739 assert_different_registers(obj, res, addr, tmp1, tmp2);
740
741 Label slow_path;
742
743 assert(res == rax, "result must arrive in rax");
744
745 if (res != obj) {
746 __ mov(res, obj);
747 }
748
749 // Check for null.
750 __ testptr(res, res);
751 __ jcc(Assembler::zero, *stub->continuation());
752
753 // Check for object being in the collection set.
754 __ mov(tmp1, res);
755 __ shrptr(tmp1, ShenandoahHeapRegion::region_size_bytes_shift_jint());
756 __ movptr(tmp2, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr());
757 #ifdef _LP64
758 __ movbool(tmp2, Address(tmp2, tmp1, Address::times_1));
759 __ testbool(tmp2);
760 #else
761 // On x86_32, C1 register allocator can give us the register without 8-bit support.
762 // Do the full-register access and test to avoid compilation failures.
763 __ movptr(tmp2, Address(tmp2, tmp1, Address::times_1));
764 __ testptr(tmp2, 0xFF);
765 #endif
766 __ jcc(Assembler::zero, *stub->continuation());
767
768 __ bind(slow_path);
769 ce->store_parameter(res, 0);
770 ce->store_parameter(addr, 1);
771 __ call(RuntimeAddress(bs->load_reference_barrier_rt_code_blob()->code_begin()));
772
773 __ jmp(*stub->continuation());
774 }
775
776 #undef __
777
778 #define __ sasm->
779
780 void ShenandoahBarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) {
781 __ prologue("shenandoah_pre_barrier", false);
782 // arg0 : previous value of memory
783
784 __ push(rax);
785 __ push(rdx);
786
787 const Register pre_val = rax;
788 const Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
789 const Register tmp = rdx;
790
791 NOT_LP64(__ get_thread(thread);)
792
793 Address queue_index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
794 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
795
796 Label done;
797 Label runtime;
798
799 // Is SATB still active?
800 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
801 __ testb(gc_state, ShenandoahHeap::MARKING);
802 __ jcc(Assembler::zero, done);
803
804 // Can we store original value in the thread's buffer?
805
806 __ movptr(tmp, queue_index);
807 __ testptr(tmp, tmp);
808 __ jcc(Assembler::zero, runtime);
809 __ subptr(tmp, wordSize);
810 __ movptr(queue_index, tmp);
811 __ addptr(tmp, buffer);
812
813 // prev_val (rax)
814 __ load_parameter(0, pre_val);
815 __ movptr(Address(tmp, 0), pre_val);
816 __ jmp(done);
817
818 __ bind(runtime);
819
820 __ save_live_registers_no_oop_map(true);
821
822 // load the pre-value
823 __ load_parameter(0, rcx);
824 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), rcx, thread);
825
826 __ restore_live_registers(true);
827
828 __ bind(done);
829
830 __ pop(rdx);
831 __ pop(rax);
832
833 __ epilogue();
834 }
835
836 void ShenandoahBarrierSetAssembler::generate_c1_load_reference_barrier_runtime_stub(StubAssembler* sasm) {
837 __ prologue("shenandoah_load_reference_barrier", false);
838 // arg0 : object to be resolved
839
840 __ save_live_registers_no_oop_map(true);
841
842 #ifdef _LP64
843 __ load_parameter(0, c_rarg0);
844 __ load_parameter(1, c_rarg1);
845 if (UseCompressedOops) {
846 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_narrow), c_rarg0, c_rarg1);
847 } else {
848 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier), c_rarg0, c_rarg1);
849 }
850 #else
851 __ load_parameter(0, rax);
852 __ load_parameter(1, rbx);
853 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier), rax, rbx);
854 #endif
855
856 __ restore_live_registers_except_rax(true);
857
858 __ epilogue();
859 }
860
861 #undef __
862
863 #endif // COMPILER1
864
865 address ShenandoahBarrierSetAssembler::shenandoah_lrb() {
866 assert(_shenandoah_lrb != NULL, "need load reference barrier stub");
867 return _shenandoah_lrb;
868 }
869
870 #define __ cgen->assembler()->
871
872 /*
873 * Incoming parameters:
874 * rax: oop
875 * rsi: load address
876 */
877 address ShenandoahBarrierSetAssembler::generate_shenandoah_lrb(StubCodeGenerator* cgen) {
878 __ align(CodeEntryAlignment);
879 StubCodeMark mark(cgen, "StubRoutines", "shenandoah_lrb");
880 address start = __ pc();
881
882 Label resolve_oop, slow_path;
883
884 // We use RDI, which also serves as argument register for slow call.
885 // RAX always holds the src object ptr, except after the slow call,
886 // then it holds the result. R8/RBX is used as temporary register.
887
888 Register tmp1 = rdi;
889 Register tmp2 = LP64_ONLY(r8) NOT_LP64(rbx);
890
891 __ push(tmp1);
892 __ push(tmp2);
893
894 // Check for object being in the collection set.
895 __ mov(tmp1, rax);
896 __ shrptr(tmp1, ShenandoahHeapRegion::region_size_bytes_shift_jint());
897 __ movptr(tmp2, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr());
898 __ movbool(tmp2, Address(tmp2, tmp1, Address::times_1));
899 __ testbool(tmp2);
900 __ jccb(Assembler::notZero, resolve_oop);
901 __ pop(tmp2);
902 __ pop(tmp1);
903 __ ret(0);
904
905 // Test if object is already resolved.
906 __ bind(resolve_oop);
907 __ movptr(tmp2, Address(rax, oopDesc::mark_offset_in_bytes()));
908 // Test if both lowest bits are set. We trick it by negating the bits
909 // then test for both bits clear.
910 __ notptr(tmp2);
911 __ testb(tmp2, markOopDesc::marked_value);
912 __ jccb(Assembler::notZero, slow_path);
913 // Clear both lower bits. It's still inverted, so set them, and then invert back.
914 __ orptr(tmp2, markOopDesc::marked_value);
915 __ notptr(tmp2);
916 // At this point, tmp2 contains the decoded forwarding pointer.
917 __ mov(rax, tmp2);
918
919 __ pop(tmp2);
920 __ pop(tmp1);
921 __ ret(0);
922
923 __ bind(slow_path);
924
925 __ push(rcx);
926 __ push(rdx);
927 __ push(rdi);
928 #ifdef _LP64
929 __ push(r8);
930 __ push(r9);
931 __ push(r10);
932 __ push(r11);
933 __ push(r12);
934 __ push(r13);
935 __ push(r14);
936 __ push(r15);
937 #endif
938 __ push(rbp);
939 __ movptr(rbp, rsp);
940 __ andptr(rsp, -StackAlignmentInBytes);
941 __ push_FPU_state();
942 if (UseCompressedOops) {
943 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_narrow), rax, rsi);
944 } else {
945 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier), rax, rsi);
946 }
947 __ pop_FPU_state();
948 __ movptr(rsp, rbp);
949 __ pop(rbp);
950 #ifdef _LP64
951 __ pop(r15);
952 __ pop(r14);
953 __ pop(r13);
954 __ pop(r12);
955 __ pop(r11);
956 __ pop(r10);
957 __ pop(r9);
958 __ pop(r8);
959 #endif
960 __ pop(rdi);
961 __ pop(rdx);
962 __ pop(rcx);
963
964 __ pop(tmp2);
965 __ pop(tmp1);
966 __ ret(0);
967
968 return start;
969 }
970
971 #undef __
972
973 void ShenandoahBarrierSetAssembler::barrier_stubs_init() {
974 if (ShenandoahLoadRefBarrier) {
975 int stub_code_size = 4096;
976 ResourceMark rm;
977 BufferBlob* bb = BufferBlob::create("shenandoah_barrier_stubs", stub_code_size);
978 CodeBuffer buf(bb);
979 StubCodeGenerator cgen(&buf);
980 _shenandoah_lrb = generate_shenandoah_lrb(&cgen);
981 }
982 }