1 /*
2 * Copyright (c) 2018, Red Hat, Inc. and/or its affiliates.
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/brooksPointer.hpp"
26 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
27 #include "gc/shenandoah/shenandoahConnectionMatrix.hpp"
28 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
29 #include "gc/shenandoah/shenandoahRuntime.hpp"
30 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
31 #include "interpreter/interpreter.hpp"
32 #include "interpreter/interp_masm.hpp"
33 #include "runtime/sharedRuntime.hpp"
34 #include "runtime/thread.hpp"
35 #include "utilities/macros.hpp"
36 #ifdef COMPILER1
37 #include "c1/c1_LIRAssembler.hpp"
38 #include "c1/c1_MacroAssembler.hpp"
39 #include "gc/shenandoah/c1/shenandoahBarrierSetC1.hpp"
40 #endif
41
42 #define __ masm->
43
44 void ShenandoahBarrierSetAssembler::arraycopy_prologue(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
45 Register src, Register dst, Register count) {
46
47 bool checkcast = (decorators & ARRAYCOPY_CHECKCAST) != 0;
48 bool disjoint = (decorators & ARRAYCOPY_DISJOINT) != 0;
49 bool obj_int = type == T_OBJECT LP64_ONLY(&& UseCompressedOops);
50 bool dest_uninitialized = (decorators & AS_DEST_NOT_INITIALIZED) != 0;
51
52 if (type == T_OBJECT || type == T_ARRAY) {
53 #ifdef _LP64
54 if (!checkcast && !obj_int) {
55 // Save count for barrier
56 __ movptr(r11, count);
57 } else if (disjoint && obj_int) {
58 // Save dst in r11 in the disjoint case
59 __ movq(r11, dst);
60 }
61 #else
62 if (disjoint) {
63 __ mov(rdx, dst); // save 'to'
64 }
65 #endif
66
67 if (!dest_uninitialized) {
68 Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
69 #ifndef _LP64
70 __ push(thread);
71 __ get_thread(thread);
72 #endif
73
74 Label filtered;
75 Address in_progress(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()));
76 // Is marking active?
77 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
78 __ cmpl(in_progress, 0);
79 } else {
80 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");
81 __ cmpb(in_progress, 0);
82 }
83
84 NOT_LP64(__ pop(thread);)
85
86 __ jcc(Assembler::equal, filtered);
87
88 __ pusha(); // push registers
89 #ifdef _LP64
90 if (count == c_rarg0) {
91 if (dst == c_rarg1) {
92 // exactly backwards!!
93 __ xchgptr(c_rarg1, c_rarg0);
94 } else {
95 __ movptr(c_rarg1, count);
96 __ movptr(c_rarg0, dst);
97 }
98 } else {
99 __ movptr(c_rarg0, dst);
100 __ movptr(c_rarg1, count);
101 }
102 if (UseCompressedOops) {
103 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_narrow_oop_entry), 2);
104 } else {
105 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry), 2);
106 }
107 #else
108 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry),
109 dst, count);
110 #endif
111 __ popa();
112 __ bind(filtered);
113 }
114 }
115
116 }
117
118 void ShenandoahBarrierSetAssembler::arraycopy_epilogue(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
119 Register src, Register dst, Register count) {
120 bool checkcast = (decorators & ARRAYCOPY_CHECKCAST) != 0;
121 bool disjoint = (decorators & ARRAYCOPY_DISJOINT) != 0;
122 bool obj_int = type == T_OBJECT LP64_ONLY(&& UseCompressedOops);
123 Register tmp = rax;
124
125 if (type == T_OBJECT || type == T_ARRAY) {
126 #ifdef _LP64
127 if (!checkcast && !obj_int) {
128 // Save count for barrier
129 count = r11;
130 } else if (disjoint && obj_int) {
131 // Use the saved dst in the disjoint case
132 dst = r11;
133 } else if (checkcast) {
134 tmp = rscratch1;
135 }
136 #else
137 if (disjoint) {
138 __ mov(dst, rdx); // restore 'to'
139 }
140 #endif
141
142 __ pusha(); // push registers (overkill)
143 #ifdef _LP64
144 if (c_rarg0 == count) { // On win64 c_rarg0 == rcx
145 assert_different_registers(c_rarg1, dst);
146 __ mov(c_rarg1, count);
147 __ mov(c_rarg0, dst);
148 } else {
149 assert_different_registers(c_rarg0, count);
150 __ mov(c_rarg0, dst);
151 __ mov(c_rarg1, count);
152 }
153 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry), 2);
154 #else
155 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry),
156 dst, count);
157 #endif
158 __ popa();
159 }
160 }
161
162 void ShenandoahBarrierSetAssembler::shenandoah_write_barrier_pre(MacroAssembler* masm,
163 Register obj,
164 Register pre_val,
165 Register thread,
166 Register tmp,
167 bool tosca_live,
168 bool expand_call) {
169
170 if (ShenandoahSATBBarrier) {
171 satb_write_barrier_pre(masm, obj, pre_val, thread, tmp, tosca_live, expand_call);
172 }
173 }
174
175 void ShenandoahBarrierSetAssembler::satb_write_barrier_pre(MacroAssembler* masm,
176 Register obj,
177 Register pre_val,
178 Register thread,
179 Register tmp,
180 bool tosca_live,
181 bool expand_call) {
182 // If expand_call is true then we expand the call_VM_leaf macro
183 // directly to skip generating the check by
184 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
185
186 #ifdef _LP64
187 assert(thread == r15_thread, "must be");
188 #endif // _LP64
189
190 Label done;
191 Label runtime;
192
193 assert(pre_val != noreg, "check this code");
194
195 if (obj != noreg) {
196 assert_different_registers(obj, pre_val, tmp);
197 assert(pre_val != rax, "check this code");
198 }
199
200 Address in_progress(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()));
201 Address index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
202 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
203
204 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
205 __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL);
206 __ jcc(Assembler::zero, done);
207
208 // Do we need to load the previous value?
209 if (obj != noreg) {
210 __ load_heap_oop(pre_val, Address(obj, 0), noreg, noreg, AS_RAW);
211 }
212
213 // Is the previous value null?
214 __ cmpptr(pre_val, (int32_t) NULL_WORD);
215 __ jcc(Assembler::equal, done);
216
217 // Can we store original value in the thread's buffer?
218 // Is index == 0?
219 // (The index field is typed as size_t.)
220
221 __ movptr(tmp, index); // tmp := *index_adr
222 __ cmpptr(tmp, 0); // tmp == 0?
223 __ jcc(Assembler::equal, runtime); // If yes, goto runtime
224
225 __ subptr(tmp, wordSize); // tmp := tmp - wordSize
226 __ movptr(index, tmp); // *index_adr := tmp
227 __ addptr(tmp, buffer); // tmp := tmp + *buffer_adr
228
229 // Record the previous value
230 __ movptr(Address(tmp, 0), pre_val);
231 __ jmp(done);
232
233 __ bind(runtime);
234 // save the live input values
235 if(tosca_live) __ push(rax);
236
237 if (obj != noreg && obj != rax)
238 __ push(obj);
239
240 if (pre_val != rax)
241 __ push(pre_val);
242
243 // Calling the runtime using the regular call_VM_leaf mechanism generates
244 // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
245 // that checks that the *(ebp+frame::interpreter_frame_last_sp) == NULL.
246 //
247 // If we care generating the pre-barrier without a frame (e.g. in the
248 // intrinsified Reference.get() routine) then ebp might be pointing to
249 // the caller frame and so this check will most likely fail at runtime.
250 //
251 // Expanding the call directly bypasses the generation of the check.
252 // So when we do not have have a full interpreter frame on the stack
253 // expand_call should be passed true.
254
255 NOT_LP64( __ push(thread); )
256
257 if (expand_call) {
258 LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); )
259 #ifdef _LP64
260 if (c_rarg1 != thread) {
261 __ mov(c_rarg1, thread);
262 }
263 if (c_rarg0 != pre_val) {
264 __ mov(c_rarg0, pre_val);
265 }
266 #else
267 __ push(thread);
268 __ push(pre_val);
269 #endif
270 __ MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), 2);
271 } else {
272 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), pre_val, thread);
273 }
274
275 NOT_LP64( __ pop(thread); )
276
277 // save the live input values
278 if (pre_val != rax)
279 __ pop(pre_val);
280
281 if (obj != noreg && obj != rax)
282 __ pop(obj);
283
284 if(tosca_live) __ pop(rax);
285
286 __ bind(done);
287 }
288
289 void ShenandoahBarrierSetAssembler::shenandoah_write_barrier_post(MacroAssembler* masm,
290 Register store_addr,
291 Register new_val,
292 Register thread,
293 Register tmp,
294 Register tmp2) {
295 assert(UseShenandoahGC, "why else should we be here?");
296
297 if (! UseShenandoahMatrix) {
298 // No need for that barrier if not using matrix.
299 return;
300 }
301
302 Label done;
303 __ testptr(new_val, new_val);
304 __ jcc(Assembler::zero, done);
305 ShenandoahConnectionMatrix* matrix = ShenandoahHeap::heap()->connection_matrix();
306 address matrix_addr = matrix->matrix_addr();
307 __ movptr(rscratch1, (intptr_t) ShenandoahHeap::heap()->base());
308 // Compute to-region index
309 __ movptr(tmp, new_val);
310 __ subptr(tmp, rscratch1);
311 __ shrptr(tmp, ShenandoahHeapRegion::region_size_bytes_shift_jint());
312 // Compute from-region index
313 __ movptr(tmp2, store_addr);
314 __ subptr(tmp2, rscratch1);
315 __ shrptr(tmp2, ShenandoahHeapRegion::region_size_bytes_shift_jint());
316 // Compute matrix index
317 __ imulptr(tmp, tmp, matrix->stride_jint());
318 __ addptr(tmp, tmp2);
319 // Address is _matrix[to * stride + from]
320 __ movptr(rscratch1, (intptr_t) matrix_addr);
321 // Test if the element is already set.
322 __ cmpb(Address(rscratch1, tmp, Address::times_1), 0);
323 __ jcc(Assembler::notEqual, done);
324 // Store true, if not yet set.
325 __ movb(Address(rscratch1, tmp, Address::times_1), 1);
326 __ bind(done);
327 }
328
329 void ShenandoahBarrierSetAssembler::read_barrier(MacroAssembler* masm, Register dst) {
330 if (ShenandoahReadBarrier) {
331 read_barrier_impl(masm, dst);
332 }
333 }
334
335 void ShenandoahBarrierSetAssembler::read_barrier_impl(MacroAssembler* masm, Register dst) {
336 assert(UseShenandoahGC && (ShenandoahReadBarrier || ShenandoahStoreValReadBarrier), "should be enabled");
337 Label is_null;
338 __ testptr(dst, dst);
339 __ jcc(Assembler::zero, is_null);
340 read_barrier_not_null_impl(masm, dst);
341 __ bind(is_null);
342 }
343
344 void ShenandoahBarrierSetAssembler::read_barrier_not_null(MacroAssembler* masm, Register dst) {
345 if (ShenandoahReadBarrier) {
346 read_barrier_not_null_impl(masm, dst);
347 }
348 }
349
350 void ShenandoahBarrierSetAssembler::read_barrier_not_null_impl(MacroAssembler* masm, Register dst) {
351 assert(UseShenandoahGC && (ShenandoahReadBarrier || ShenandoahStoreValReadBarrier), "should be enabled");
352 __ movptr(dst, Address(dst, BrooksPointer::byte_offset()));
353 }
354
355
356 void ShenandoahBarrierSetAssembler::write_barrier(MacroAssembler* masm, Register dst) {
357 if (ShenandoahWriteBarrier) {
358 write_barrier_impl(masm, dst);
359 }
360 }
361
362 void ShenandoahBarrierSetAssembler::write_barrier_impl(MacroAssembler* masm, Register dst) {
363 assert(UseShenandoahGC && (ShenandoahWriteBarrier || ShenandoahStoreValEnqueueBarrier), "should be enabled");
364 #ifdef _LP64
365 assert(dst != rscratch1, "different regs");
366
367 Label done;
368
369 Address gc_state(r15_thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
370 __ testb(gc_state, ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL);
371
372 // Now check if evacuation is in progress.
373 read_barrier_not_null(masm, dst);
374
375 __ jcc(Assembler::zero, done);
376 __ push(rscratch1);
377 __ push(rscratch2);
378
379 __ movptr(rscratch1, dst);
380 __ shrptr(rscratch1, ShenandoahHeapRegion::region_size_bytes_shift_jint());
381 __ movptr(rscratch2, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr());
382 __ movbool(rscratch2, Address(rscratch2, rscratch1, Address::times_1));
383 __ testb(rscratch2, 0x1);
384
385 __ pop(rscratch2);
386 __ pop(rscratch1);
387
388 __ jcc(Assembler::zero, done);
389
390 __ push(rscratch1);
391
392 // Save possibly live regs.
393 if (dst != rax) {
394 __ push(rax);
395 }
396 if (dst != rbx) {
397 __ push(rbx);
398 }
399 if (dst != rcx) {
400 __ push(rcx);
401 }
402 if (dst != rdx) {
403 __ push(rdx);
404 }
405 if (dst != c_rarg1) {
406 __ push(c_rarg1);
407 }
408
409 __ subptr(rsp, 2 * Interpreter::stackElementSize);
410 __ movdbl(Address(rsp, 0), xmm0);
411
412 // Call into runtime
413 __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_IRT), dst);
414 __ mov(rscratch1, rax);
415
416 // Restore possibly live regs.
417 __ movdbl(xmm0, Address(rsp, 0));
418 __ addptr(rsp, 2 * Interpreter::stackElementSize);
419
420 if (dst != c_rarg1) {
421 __ pop(c_rarg1);
422 }
423 if (dst != rdx) {
424 __ pop(rdx);
425 }
426 if (dst != rcx) {
427 __ pop(rcx);
428 }
429 if (dst != rbx) {
430 __ pop(rbx);
431 }
432 if (dst != rax) {
433 __ pop(rax);
434 }
435
436 // Move result into dst reg.
437 __ mov(dst, rscratch1);
438
439 __ pop(rscratch1);
440
441 __ bind(done);
442 #else
443 Unimplemented();
444 #endif
445 }
446
447 void ShenandoahBarrierSetAssembler::storeval_barrier(MacroAssembler* masm, Register dst, Register tmp) {
448 if (ShenandoahStoreValReadBarrier || ShenandoahStoreValEnqueueBarrier) {
449 storeval_barrier_impl(masm, dst, tmp);
450 }
451 }
452
453 void ShenandoahBarrierSetAssembler::storeval_barrier_impl(MacroAssembler* masm, Register dst, Register tmp) {
454 assert(UseShenandoahGC && (ShenandoahStoreValReadBarrier || ShenandoahStoreValEnqueueBarrier), "should be enabled");
455
456 if (dst == noreg) return;
457
458 #ifdef _LP64
459 if (ShenandoahStoreValEnqueueBarrier) {
460 Label is_null;
461 __ testptr(dst, dst);
462 __ jcc(Assembler::zero, is_null);
463 write_barrier_impl(masm, dst);
464 __ bind(is_null);
465
466 // The set of registers to be saved+restored is the same as in the write-barrier above.
467 // Those are the commonly used registers in the interpreter.
468 __ pusha();
469 // __ push_callee_saved_registers();
470 __ subptr(rsp, 2 * Interpreter::stackElementSize);
471 __ movdbl(Address(rsp, 0), xmm0);
472
473 satb_write_barrier_pre(masm, noreg, dst, r15_thread, tmp, true, false);
474 __ movdbl(xmm0, Address(rsp, 0));
475 __ addptr(rsp, 2 * Interpreter::stackElementSize);
476 //__ pop_callee_saved_registers();
477 __ popa();
478 }
479 if (ShenandoahStoreValReadBarrier) {
480 read_barrier_impl(masm, dst);
481 }
482 #else
483 Unimplemented();
484 #endif
485 }
486
487 void ShenandoahBarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
488 Register dst, Address src, Register tmp1, Register tmp_thread) {
489 bool on_oop = type == T_OBJECT || type == T_ARRAY;
490 bool in_heap = (decorators & IN_HEAP) != 0;
491 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
492 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
493 bool on_reference = on_weak || on_phantom;
494 if (in_heap) {
495 read_barrier_not_null(masm, src.base());
496 }
497 BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp_thread);
498 if (ShenandoahKeepAliveBarrier && on_oop && on_reference) {
499 const Register thread = NOT_LP64(tmp_thread) LP64_ONLY(r15_thread);
500 NOT_LP64(__ get_thread(thread));
501
502 // Generate the SATB pre-barrier code to log the value of
503 // the referent field in an SATB buffer.
504 shenandoah_write_barrier_pre(masm /* masm */,
505 noreg /* obj */,
506 dst /* pre_val */,
507 thread /* thread */,
508 tmp1 /* tmp */,
509 true /* tosca_live */,
510 true /* expand_call */);
511 }
512 }
513
514 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
515 Address dst, Register val, Register tmp1, Register tmp2) {
516
517 bool in_heap = (decorators & IN_HEAP) != 0;
518 bool in_concurrent_root = (decorators & IN_CONCURRENT_ROOT) != 0;
519 if (in_heap) {
520 write_barrier(masm, dst.base());
521 }
522 if (type == T_OBJECT || type == T_ARRAY) {
523 bool needs_pre_barrier = in_heap || in_concurrent_root;
524 bool needs_post_barrier = val != noreg && in_heap && UseShenandoahMatrix;
525
526 Register tmp3 = LP64_ONLY(r8) NOT_LP64(rsi);
527 Register rthread = LP64_ONLY(r15_thread) NOT_LP64(rcx);
528 // flatten object address if needed
529 // We do it regardless of precise because we need the registers
530 if (dst.index() == noreg && dst.disp() == 0) {
531 if (dst.base() != tmp1) {
532 __ movptr(tmp1, dst.base());
533 }
534 } else {
535 __ lea(tmp1, dst);
536 }
537
538 #ifndef _LP64
539 InterpreterMacroAssembler *imasm = static_cast<InterpreterMacroAssembler*>(masm);
540 #endif
541
542 NOT_LP64(__ get_thread(rcx));
543 NOT_LP64(imasm->save_bcp());
544
545 if (needs_pre_barrier) {
546 shenandoah_write_barrier_pre(masm /*masm*/,
547 tmp1 /* obj */,
548 tmp2 /* pre_val */,
549 rthread /* thread */,
550 tmp3 /* tmp */,
551 val != noreg /* tosca_live */,
552 false /* expand_call */);
553 }
554 if (val == noreg) {
555 BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
556 } else {
557 storeval_barrier(masm, val, tmp3);
558 Register new_val = val;
559 if (needs_post_barrier) {
560 if (UseCompressedOops) {
561 new_val = tmp2;
562 __ movptr(new_val, val);
563 }
564 }
565 BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
566
567 if (needs_post_barrier) {
568 shenandoah_write_barrier_post(masm /*masm*/,
569 tmp1 /* store_adr */,
570 new_val /* new_val */,
571 rthread /* thread */,
572 tmp3 /* tmp */,
573 tmp2 /* tmp2 */);
574 }
575 }
576 NOT_LP64(imasm->restore_bcp());
577
578 } else {
579 BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2);
580 }
581 }
582
583 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm, DecoratorSet decorators, Register op1, Register op2) {
584 __ cmpptr(op1, op2);
585 if (ShenandoahAcmpBarrier) {
586 Label done;
587 __ jccb(Assembler::equal, done);
588 read_barrier(masm, op1);
589 read_barrier(masm, op2);
590 __ cmpptr(op1, op2);
591 __ bind(done);
592 }
593 }
594
595 void ShenandoahBarrierSetAssembler::obj_equals_addr(MacroAssembler* masm, DecoratorSet decorators, Register src1, Address src2) {
596 __ cmpptr(src1, src2);
597 if (ShenandoahAcmpBarrier) {
598 Label done;
599 __ jccb(Assembler::equal, done);
600 __ movptr(rscratch2, src2);
601 read_barrier(masm, src1);
602 read_barrier(masm, rscratch2);
603 __ cmpptr(src1, rscratch2);
604 __ bind(done);
605 }
606 }
607
608 void ShenandoahBarrierSetAssembler::resolve_for_read(MacroAssembler* masm, DecoratorSet decorators, Register obj) {
609 bool oop_not_null = (decorators & OOP_NOT_NULL) != 0;
610 if (oop_not_null) {
611 read_barrier_not_null(masm, obj);
612 } else {
613 read_barrier(masm, obj);
614 }
615 }
616
617 void ShenandoahBarrierSetAssembler::resolve_for_write(MacroAssembler* masm, DecoratorSet decorators, Register obj) {
618 write_barrier(masm, obj);
619 }
620
621 // Special Shenandoah CAS implementation that handles false negatives
622 // due to concurrent evacuation.
623 #ifndef _LP64
624 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm, DecoratorSet decorators,
625 Register res, Address addr, Register oldval, Register newval,
626 bool exchange, bool encode, Register tmp1, Register tmp2) {
627 // Shenandoah has no 32-bit version for this.
628 Unimplemented();
629 }
630 #else
631 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm, DecoratorSet decorators,
632 Register res, Address addr, Register oldval, Register newval,
633 bool exchange, bool encode, Register tmp1, Register tmp2) {
634
635 if (!ShenandoahCASBarrier) {
636 BarrierSetAssembler::cmpxchg_oop(masm, decorators, res, addr, oldval, newval, exchange, encode, tmp1, tmp2);
637 return;
638 }
639
640 assert(ShenandoahCASBarrier, "Should only be used when CAS barrier is enabled");
641 assert(oldval == rax, "must be in rax for implicit use in cmpxchg");
642
643 Label retry, done;
644
645 // Apply storeval barrier to newval.
646 if (encode) {
647 if (newval == c_rarg1 && ShenandoahStoreValEnqueueBarrier) {
648 __ mov(tmp2, newval);
649 storeval_barrier(masm, tmp2, tmp1);
650 } else {
651 storeval_barrier(masm, newval, tmp1);
652 }
653 }
654
655 if (UseCompressedOops) {
656 if (encode) {
657 __ encode_heap_oop(oldval);
658 __ mov(rscratch1, newval);
659 __ encode_heap_oop(rscratch1);
660 newval = rscratch1;
661 }
662 }
663
664 // Remember oldval for retry logic below
665 if (UseCompressedOops) {
666 __ movl(tmp1, oldval);
667 } else {
668 __ movptr(tmp1, oldval);
669 }
670
671 // Step 1. Try to CAS with given arguments. If successful, then we are done,
672 // and can safely return.
673 if (os::is_MP()) __ lock();
674 if (UseCompressedOops) {
675 __ cmpxchgl(newval, addr);
676 } else {
677 __ cmpxchgptr(newval, addr);
678 }
679 __ jcc(Assembler::equal, done, true);
680
681 // Step 2. CAS had failed. This may be a false negative.
682 //
683 // The trouble comes when we compare the to-space pointer with the from-space
684 // pointer to the same object. To resolve this, it will suffice to read both
685 // oldval and the value from memory through the read barriers -- this will give
686 // both to-space pointers. If they mismatch, then it was a legitimate failure.
687 //
688 if (UseCompressedOops) {
689 __ decode_heap_oop(tmp1);
690 }
691 __ resolve_for_read(0, tmp1);
692
693 if (UseCompressedOops) {
694 __ movl(tmp2, oldval);
695 __ decode_heap_oop(tmp2);
696 } else {
697 __ movptr(tmp2, oldval);
698 }
699 __ resolve_for_read(0, tmp2);
700
701 __ cmpptr(tmp1, tmp2);
702 __ jcc(Assembler::notEqual, done, true);
703
704 // Step 3. Try to CAS again with resolved to-space pointers.
705 //
706 // Corner case: it may happen that somebody stored the from-space pointer
707 // to memory while we were preparing for retry. Therefore, we can fail again
708 // on retry, and so need to do this in loop, always re-reading the failure
709 // witness through the read barrier.
710 __ bind(retry);
711 if (os::is_MP()) __ lock();
712 if (UseCompressedOops) {
713 __ cmpxchgl(newval, addr);
714 } else {
715 __ cmpxchgptr(newval, addr);
716 }
717 __ jcc(Assembler::equal, done, true);
718
719 if (UseCompressedOops) {
720 __ movl(tmp2, oldval);
721 __ decode_heap_oop(tmp2);
722 } else {
723 __ movptr(tmp2, oldval);
724 }
725 __ resolve_for_read(0, tmp2);
726
727 __ cmpptr(tmp1, tmp2);
728 __ jcc(Assembler::equal, retry, true);
729
730 // Step 4. If we need a boolean result out of CAS, check the flag again,
731 // and promote the result. Note that we handle the flag from both the CAS
732 // itself and from the retry loop.
733 __ bind(done);
734 if (!exchange) {
735 assert(res != NULL, "need result register");
736 __ setb(Assembler::equal, res);
737 __ movzbl(res, res);
738 }
739 }
740 #endif // LP64
741
742 void ShenandoahBarrierSetAssembler::xchg_oop(MacroAssembler* masm, DecoratorSet decorators,
743 Register obj, Address addr, Register tmp) {
744 storeval_barrier(masm, obj, tmp);
745 BarrierSetAssembler::xchg_oop(masm, decorators, obj, addr, tmp);
746 }
747
748 #ifdef COMPILER1
749
750 #undef __
751 #define __ ce->masm()->
752
753 void ShenandoahBarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, ShenandoahPreBarrierStub* stub) {
754 ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
755 // At this point we know that marking is in progress.
756 // If do_load() is true then we have to emit the
757 // load of the previous value; otherwise it has already
758 // been loaded into _pre_val.
759
760 __ bind(*stub->entry());
761 assert(stub->pre_val()->is_register(), "Precondition.");
762
763 Register pre_val_reg = stub->pre_val()->as_register();
764
765 if (stub->do_load()) {
766 ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/, false /*unaligned*/);
767 }
768
769 __ cmpptr(pre_val_reg, (int32_t)NULL_WORD);
770 __ jcc(Assembler::equal, *stub->continuation());
771 ce->store_parameter(stub->pre_val()->as_register(), 0);
772 __ call(RuntimeAddress(bs->pre_barrier_c1_runtime_code_blob()->code_begin()));
773 __ jmp(*stub->continuation());
774
775 }
776
777 #undef __
778
779 #define __ sasm->
780
781 void ShenandoahBarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) {
782 __ prologue("shenandoah_pre_barrier", false);
783 // arg0 : previous value of memory
784
785 __ push(rax);
786 __ push(rdx);
787
788 const Register pre_val = rax;
789 const Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
790 const Register tmp = rdx;
791
792 NOT_LP64(__ get_thread(thread);)
793
794 Address queue_index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
795 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
796
797 Label done;
798 Label runtime;
799
800 // Is SATB still active?
801 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
802 __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL);
803 __ jcc(Assembler::zero, done);
804
805 // Can we store original value in the thread's buffer?
806
807 __ movptr(tmp, queue_index);
808 __ testptr(tmp, tmp);
809 __ jcc(Assembler::zero, runtime);
810 __ subptr(tmp, wordSize);
811 __ movptr(queue_index, tmp);
812 __ addptr(tmp, buffer);
813
814 // prev_val (rax)
815 __ load_parameter(0, pre_val);
816 __ movptr(Address(tmp, 0), pre_val);
817 __ jmp(done);
818
819 __ bind(runtime);
820
821 __ save_live_registers_no_oop_map(true);
822
823 // load the pre-value
824 __ load_parameter(0, rcx);
825 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), rcx, thread);
826
827 __ restore_live_registers(true);
828
829 __ bind(done);
830
831 __ pop(rdx);
832 __ pop(rax);
833
834 __ epilogue();
835 }
836
837 #undef __
838
839 #endif // COMPILER1
840