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/shenandoahHeap.hpp"
29 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
30 #include "gc/shenandoah/shenandoahRuntime.hpp"
31 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
32 #include "interpreter/interpreter.hpp"
33 #include "interpreter/interp_masm.hpp"
34 #include "runtime/sharedRuntime.hpp"
35 #include "runtime/thread.hpp"
36 #include "utilities/macros.hpp"
37 #ifdef COMPILER1
38 #include "c1/c1_LIRAssembler.hpp"
39 #include "c1/c1_MacroAssembler.hpp"
40 #include "gc/shenandoah/c1/shenandoahBarrierSetC1.hpp"
41 #endif
42
43 #define __ masm->
44
45 address ShenandoahBarrierSetAssembler::_shenandoah_wb = NULL;
46 address ShenandoahBarrierSetAssembler::_shenandoah_wb_C = NULL;
47
48 void ShenandoahBarrierSetAssembler::arraycopy_prologue(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
49 Register src, Register dst, Register count) {
50
51 bool checkcast = (decorators & ARRAYCOPY_CHECKCAST) != 0;
52 bool disjoint = (decorators & ARRAYCOPY_DISJOINT) != 0;
53 bool obj_int = type == T_OBJECT LP64_ONLY(&& UseCompressedOops);
54 bool dest_uninitialized = (decorators & AS_DEST_NOT_INITIALIZED) != 0;
55
56 if (type == T_OBJECT || type == T_ARRAY) {
57 #ifdef _LP64
58 if (!checkcast && !obj_int) {
59 // Save count for barrier
60 __ movptr(r11, count);
61 } else if (disjoint && obj_int) {
62 // Save dst in r11 in the disjoint case
63 __ movq(r11, dst);
64 }
65 #else
66 if (disjoint) {
67 __ mov(rdx, dst); // save 'to'
68 }
69 #endif
70
71 if (!dest_uninitialized) {
72 Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
73 #ifndef _LP64
74 __ push(thread);
75 __ get_thread(thread);
76 #endif
77
78 Label filtered;
79 Address in_progress(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()));
80 // Is marking active?
81 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
82 __ cmpl(in_progress, 0);
83 } else {
84 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");
85 __ cmpb(in_progress, 0);
86 }
87
88 NOT_LP64(__ pop(thread);)
89
90 __ jcc(Assembler::equal, filtered);
91
92 __ pusha(); // push registers
93 #ifdef _LP64
94 if (count == c_rarg0) {
95 if (dst == c_rarg1) {
96 // exactly backwards!!
97 __ xchgptr(c_rarg1, c_rarg0);
98 } else {
99 __ movptr(c_rarg1, count);
100 __ movptr(c_rarg0, dst);
101 }
102 } else {
103 __ movptr(c_rarg0, dst);
104 __ movptr(c_rarg1, count);
105 }
106 if (UseCompressedOops) {
107 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_narrow_oop_entry), 2);
108 } else {
109 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry), 2);
110 }
111 #else
112 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry),
113 dst, count);
114 #endif
115 __ popa();
116 __ bind(filtered);
117 }
118 }
119
120 }
121
122 void ShenandoahBarrierSetAssembler::arraycopy_epilogue(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
123 Register src, Register dst, Register count) {
124 bool checkcast = (decorators & ARRAYCOPY_CHECKCAST) != 0;
125 bool disjoint = (decorators & ARRAYCOPY_DISJOINT) != 0;
126 bool obj_int = type == T_OBJECT LP64_ONLY(&& UseCompressedOops);
127 Register tmp = rax;
128
129 if (type == T_OBJECT || type == T_ARRAY) {
130 #ifdef _LP64
131 if (!checkcast && !obj_int) {
132 // Save count for barrier
133 count = r11;
134 } else if (disjoint && obj_int) {
135 // Use the saved dst in the disjoint case
136 dst = r11;
137 } else if (checkcast) {
138 tmp = rscratch1;
139 }
140 #else
141 if (disjoint) {
142 __ mov(dst, rdx); // restore 'to'
143 }
144 #endif
145
146 __ pusha(); // push registers (overkill)
147 #ifdef _LP64
148 if (c_rarg0 == count) { // On win64 c_rarg0 == rcx
149 assert_different_registers(c_rarg1, dst);
150 __ mov(c_rarg1, count);
151 __ mov(c_rarg0, dst);
152 } else {
153 assert_different_registers(c_rarg0, count);
154 __ mov(c_rarg0, dst);
155 __ mov(c_rarg1, count);
156 }
157 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry), 2);
158 #else
159 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry),
160 dst, count);
161 #endif
162 __ popa();
163 }
164 }
165
166 void ShenandoahBarrierSetAssembler::shenandoah_write_barrier_pre(MacroAssembler* masm,
167 Register obj,
168 Register pre_val,
169 Register thread,
170 Register tmp,
171 bool tosca_live,
172 bool expand_call) {
173
174 if (ShenandoahSATBBarrier) {
175 satb_write_barrier_pre(masm, obj, pre_val, thread, tmp, tosca_live, expand_call);
176 }
177 }
178
179 void ShenandoahBarrierSetAssembler::satb_write_barrier_pre(MacroAssembler* masm,
180 Register obj,
181 Register pre_val,
182 Register thread,
183 Register tmp,
184 bool tosca_live,
185 bool expand_call) {
186 // If expand_call is true then we expand the call_VM_leaf macro
187 // directly to skip generating the check by
188 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
189
190 #ifdef _LP64
191 assert(thread == r15_thread, "must be");
192 #endif // _LP64
193
194 Label done;
195 Label runtime;
196
197 assert(pre_val != noreg, "check this code");
198
199 if (obj != noreg) {
200 assert_different_registers(obj, pre_val, tmp);
201 assert(pre_val != rax, "check this code");
202 }
203
204 Address in_progress(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()));
205 Address index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
206 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
207
208 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
209 __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL);
210 __ jcc(Assembler::zero, done);
211
212 // Do we need to load the previous value?
213 if (obj != noreg) {
214 __ load_heap_oop(pre_val, Address(obj, 0), noreg, noreg, AS_RAW);
215 }
216
217 // Is the previous value null?
218 __ cmpptr(pre_val, (int32_t) NULL_WORD);
219 __ jcc(Assembler::equal, done);
220
221 // Can we store original value in the thread's buffer?
222 // Is index == 0?
223 // (The index field is typed as size_t.)
224
225 __ movptr(tmp, index); // tmp := *index_adr
226 __ cmpptr(tmp, 0); // tmp == 0?
227 __ jcc(Assembler::equal, runtime); // If yes, goto runtime
228
229 __ subptr(tmp, wordSize); // tmp := tmp - wordSize
230 __ movptr(index, tmp); // *index_adr := tmp
231 __ addptr(tmp, buffer); // tmp := tmp + *buffer_adr
232
233 // Record the previous value
234 __ movptr(Address(tmp, 0), pre_val);
235 __ jmp(done);
236
237 __ bind(runtime);
238 // save the live input values
239 if(tosca_live) __ push(rax);
240
241 if (obj != noreg && obj != rax)
242 __ push(obj);
243
244 if (pre_val != rax)
245 __ push(pre_val);
246
247 // Calling the runtime using the regular call_VM_leaf mechanism generates
248 // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
249 // that checks that the *(ebp+frame::interpreter_frame_last_sp) == NULL.
250 //
251 // If we care generating the pre-barrier without a frame (e.g. in the
252 // intrinsified Reference.get() routine) then ebp might be pointing to
253 // the caller frame and so this check will most likely fail at runtime.
254 //
255 // Expanding the call directly bypasses the generation of the check.
256 // So when we do not have have a full interpreter frame on the stack
257 // expand_call should be passed true.
258
259 NOT_LP64( __ push(thread); )
260
261 if (expand_call) {
262 LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); )
263 #ifdef _LP64
264 if (c_rarg1 != thread) {
265 __ mov(c_rarg1, thread);
266 }
267 if (c_rarg0 != pre_val) {
268 __ mov(c_rarg0, pre_val);
269 }
270 #else
271 __ push(thread);
272 __ push(pre_val);
273 #endif
274 __ MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), 2);
275 } else {
276 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), pre_val, thread);
277 }
278
279 NOT_LP64( __ pop(thread); )
280
281 // save the live input values
282 if (pre_val != rax)
283 __ pop(pre_val);
284
285 if (obj != noreg && obj != rax)
286 __ pop(obj);
287
288 if(tosca_live) __ pop(rax);
289
290 __ bind(done);
291 }
292
293 void ShenandoahBarrierSetAssembler::shenandoah_write_barrier_post(MacroAssembler* masm,
294 Register store_addr,
295 Register new_val,
296 Register thread,
297 Register tmp,
298 Register tmp2) {
299 assert(UseShenandoahGC, "why else should we be here?");
300
301 if (! UseShenandoahMatrix) {
302 // No need for that barrier if not using matrix.
303 return;
304 }
305
306 Label done;
307 __ testptr(new_val, new_val);
308 __ jcc(Assembler::zero, done);
309 ShenandoahConnectionMatrix* matrix = ShenandoahHeap::heap()->connection_matrix();
310 address matrix_addr = matrix->matrix_addr();
311 __ movptr(rscratch1, (intptr_t) ShenandoahHeap::heap()->base());
312 // Compute to-region index
313 __ movptr(tmp, new_val);
314 __ subptr(tmp, rscratch1);
315 __ shrptr(tmp, ShenandoahHeapRegion::region_size_bytes_shift_jint());
316 // Compute from-region index
317 __ movptr(tmp2, store_addr);
318 __ subptr(tmp2, rscratch1);
319 __ shrptr(tmp2, ShenandoahHeapRegion::region_size_bytes_shift_jint());
320 // Compute matrix index
321 __ imulptr(tmp, tmp, matrix->stride_jint());
322 __ addptr(tmp, tmp2);
323 // Address is _matrix[to * stride + from]
324 __ movptr(rscratch1, (intptr_t) matrix_addr);
325 // Test if the element is already set.
326 __ cmpb(Address(rscratch1, tmp, Address::times_1), 0);
327 __ jcc(Assembler::notEqual, done);
328 // Store true, if not yet set.
329 __ movb(Address(rscratch1, tmp, Address::times_1), 1);
330 __ bind(done);
331 }
332
333 void ShenandoahBarrierSetAssembler::read_barrier(MacroAssembler* masm, Register dst) {
334 if (ShenandoahReadBarrier) {
335 read_barrier_impl(masm, dst);
336 }
337 }
338
339 void ShenandoahBarrierSetAssembler::read_barrier_impl(MacroAssembler* masm, Register dst) {
340 assert(UseShenandoahGC && (ShenandoahReadBarrier || ShenandoahStoreValReadBarrier), "should be enabled");
341 Label is_null;
342 __ testptr(dst, dst);
343 __ jcc(Assembler::zero, is_null);
344 read_barrier_not_null_impl(masm, dst);
345 __ bind(is_null);
346 }
347
348 void ShenandoahBarrierSetAssembler::read_barrier_not_null(MacroAssembler* masm, Register dst) {
349 if (ShenandoahReadBarrier) {
350 read_barrier_not_null_impl(masm, dst);
351 }
352 }
353
354 void ShenandoahBarrierSetAssembler::read_barrier_not_null_impl(MacroAssembler* masm, Register dst) {
355 assert(UseShenandoahGC && (ShenandoahReadBarrier || ShenandoahStoreValReadBarrier), "should be enabled");
356 __ movptr(dst, Address(dst, BrooksPointer::byte_offset()));
357 }
358
359
360 void ShenandoahBarrierSetAssembler::write_barrier(MacroAssembler* masm, Register dst) {
361 if (ShenandoahWriteBarrier) {
362 write_barrier_impl(masm, dst);
363 }
364 }
365
366 void ShenandoahBarrierSetAssembler::write_barrier_impl(MacroAssembler* masm, Register dst) {
367 assert(UseShenandoahGC && (ShenandoahWriteBarrier || ShenandoahStoreValEnqueueBarrier), "should be enabled");
368 #ifdef _LP64
369 assert(dst != rscratch1, "different regs");
370
371 Label done;
372
373 Address gc_state(r15_thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
374 __ testb(gc_state, ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL);
375
376 // Now check if evacuation is in progress.
377 read_barrier_not_null(masm, dst);
378
379 __ jcc(Assembler::zero, done);
380 __ push(rscratch1);
381 __ push(rscratch2);
382
383 __ movptr(rscratch1, dst);
384 __ shrptr(rscratch1, ShenandoahHeapRegion::region_size_bytes_shift_jint());
385 __ movptr(rscratch2, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr());
386 __ movbool(rscratch2, Address(rscratch2, rscratch1, Address::times_1));
387 __ testb(rscratch2, 0x1);
388
389 __ pop(rscratch2);
390 __ pop(rscratch1);
391
392 __ jcc(Assembler::zero, done);
393
394 __ push(rscratch1);
395
396 // Save possibly live regs.
397 if (dst != rax) {
398 __ push(rax);
399 }
400 if (dst != rbx) {
401 __ push(rbx);
402 }
403 if (dst != rcx) {
404 __ push(rcx);
405 }
406 if (dst != rdx) {
407 __ push(rdx);
408 }
409 if (dst != c_rarg1) {
410 __ push(c_rarg1);
411 }
412
413 __ subptr(rsp, 2 * Interpreter::stackElementSize);
414 __ movdbl(Address(rsp, 0), xmm0);
415
416 // Call into runtime
417 __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_IRT), dst);
418 __ mov(rscratch1, rax);
419
420 // Restore possibly live regs.
421 __ movdbl(xmm0, Address(rsp, 0));
422 __ addptr(rsp, 2 * Interpreter::stackElementSize);
423
424 if (dst != c_rarg1) {
425 __ pop(c_rarg1);
426 }
427 if (dst != rdx) {
428 __ pop(rdx);
429 }
430 if (dst != rcx) {
431 __ pop(rcx);
432 }
433 if (dst != rbx) {
434 __ pop(rbx);
435 }
436 if (dst != rax) {
437 __ pop(rax);
438 }
439
440 // Move result into dst reg.
441 __ mov(dst, rscratch1);
442
443 __ pop(rscratch1);
444
445 __ bind(done);
446 #else
447 Unimplemented();
448 #endif
449 }
450
451 void ShenandoahBarrierSetAssembler::storeval_barrier(MacroAssembler* masm, Register dst, Register tmp) {
452 if (ShenandoahStoreValReadBarrier || ShenandoahStoreValEnqueueBarrier) {
453 storeval_barrier_impl(masm, dst, tmp);
454 }
455 }
456
457 void ShenandoahBarrierSetAssembler::storeval_barrier_impl(MacroAssembler* masm, Register dst, Register tmp) {
458 assert(UseShenandoahGC && (ShenandoahStoreValReadBarrier || ShenandoahStoreValEnqueueBarrier), "should be enabled");
459
460 if (dst == noreg) return;
461
462 #ifdef _LP64
463 if (ShenandoahStoreValEnqueueBarrier) {
464 Label is_null;
465 __ testptr(dst, dst);
466 __ jcc(Assembler::zero, is_null);
467 write_barrier_impl(masm, dst);
468 __ bind(is_null);
469
470 // The set of registers to be saved+restored is the same as in the write-barrier above.
471 // Those are the commonly used registers in the interpreter.
472 __ pusha();
473 // __ push_callee_saved_registers();
474 __ subptr(rsp, 2 * Interpreter::stackElementSize);
475 __ movdbl(Address(rsp, 0), xmm0);
476
477 satb_write_barrier_pre(masm, noreg, dst, r15_thread, tmp, true, false);
478 __ movdbl(xmm0, Address(rsp, 0));
479 __ addptr(rsp, 2 * Interpreter::stackElementSize);
480 //__ pop_callee_saved_registers();
481 __ popa();
482 }
483 if (ShenandoahStoreValReadBarrier) {
484 read_barrier_impl(masm, dst);
485 }
486 #else
487 Unimplemented();
488 #endif
489 }
490
491 void ShenandoahBarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
492 Register dst, Address src, Register tmp1, Register tmp_thread) {
493 bool on_oop = type == T_OBJECT || type == T_ARRAY;
494 bool in_heap = (decorators & IN_HEAP) != 0;
495 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
496 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
497 bool on_reference = on_weak || on_phantom;
498 if (in_heap) {
499 read_barrier_not_null(masm, src.base());
500 }
501 BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp_thread);
502 if (ShenandoahKeepAliveBarrier && on_oop && on_reference) {
503 const Register thread = NOT_LP64(tmp_thread) LP64_ONLY(r15_thread);
504 NOT_LP64(__ get_thread(thread));
505
506 // Generate the SATB pre-barrier code to log the value of
507 // the referent field in an SATB buffer.
508 shenandoah_write_barrier_pre(masm /* masm */,
509 noreg /* obj */,
510 dst /* pre_val */,
511 thread /* thread */,
512 tmp1 /* tmp */,
513 true /* tosca_live */,
514 true /* expand_call */);
515 }
516 }
517
518 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
519 Address dst, Register val, Register tmp1, Register tmp2) {
520
521 bool in_heap = (decorators & IN_HEAP) != 0;
522 bool in_concurrent_root = (decorators & IN_CONCURRENT_ROOT) != 0;
523 if (in_heap) {
524 write_barrier(masm, dst.base());
525 }
526 if (type == T_OBJECT || type == T_ARRAY) {
527 bool needs_pre_barrier = in_heap || in_concurrent_root;
528 bool needs_post_barrier = val != noreg && in_heap && UseShenandoahMatrix;
529
530 Register tmp3 = LP64_ONLY(r8) NOT_LP64(rsi);
531 Register rthread = LP64_ONLY(r15_thread) NOT_LP64(rcx);
532 // flatten object address if needed
533 // We do it regardless of precise because we need the registers
534 if (dst.index() == noreg && dst.disp() == 0) {
535 if (dst.base() != tmp1) {
536 __ movptr(tmp1, dst.base());
537 }
538 } else {
539 __ lea(tmp1, dst);
540 }
541
542 #ifndef _LP64
543 InterpreterMacroAssembler *imasm = static_cast<InterpreterMacroAssembler*>(masm);
544 #endif
545
546 NOT_LP64(__ get_thread(rcx));
547 NOT_LP64(imasm->save_bcp());
548
549 if (needs_pre_barrier) {
550 shenandoah_write_barrier_pre(masm /*masm*/,
551 tmp1 /* obj */,
552 tmp2 /* pre_val */,
553 rthread /* thread */,
554 tmp3 /* tmp */,
555 val != noreg /* tosca_live */,
556 false /* expand_call */);
557 }
558 if (val == noreg) {
559 BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
560 } else {
561 storeval_barrier(masm, val, tmp3);
562 Register new_val = val;
563 if (needs_post_barrier) {
564 if (UseCompressedOops) {
565 new_val = tmp2;
566 __ movptr(new_val, val);
567 }
568 }
569 BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
570
571 if (needs_post_barrier) {
572 shenandoah_write_barrier_post(masm /*masm*/,
573 tmp1 /* store_adr */,
574 new_val /* new_val */,
575 rthread /* thread */,
576 tmp3 /* tmp */,
577 tmp2 /* tmp2 */);
578 }
579 }
580 NOT_LP64(imasm->restore_bcp());
581
582 } else {
583 BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2);
584 }
585 }
586
587 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm, DecoratorSet decorators, Register op1, Register op2) {
588 __ cmpptr(op1, op2);
589 if (ShenandoahAcmpBarrier) {
590 Label done;
591 __ jccb(Assembler::equal, done);
592 read_barrier(masm, op1);
593 read_barrier(masm, op2);
594 __ cmpptr(op1, op2);
595 __ bind(done);
596 }
597 }
598
599 void ShenandoahBarrierSetAssembler::obj_equals_addr(MacroAssembler* masm, DecoratorSet decorators, Register src1, Address src2) {
600 __ cmpptr(src1, src2);
601 if (ShenandoahAcmpBarrier) {
602 Label done;
603 __ jccb(Assembler::equal, done);
604 __ movptr(rscratch2, src2);
605 read_barrier(masm, src1);
606 read_barrier(masm, rscratch2);
607 __ cmpptr(src1, rscratch2);
608 __ bind(done);
609 }
610 }
611
612 void ShenandoahBarrierSetAssembler::resolve_for_read(MacroAssembler* masm, DecoratorSet decorators, Register obj) {
613 bool oop_not_null = (decorators & OOP_NOT_NULL) != 0;
614 if (oop_not_null) {
615 read_barrier_not_null(masm, obj);
616 } else {
617 read_barrier(masm, obj);
618 }
619 }
620
621 void ShenandoahBarrierSetAssembler::resolve_for_write(MacroAssembler* masm, DecoratorSet decorators, Register obj) {
622 write_barrier(masm, obj);
623 }
624
625 // Special Shenandoah CAS implementation that handles false negatives
626 // due to concurrent evacuation.
627 #ifndef _LP64
628 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm, DecoratorSet decorators,
629 Register res, Address addr, Register oldval, Register newval,
630 bool exchange, bool encode, Register tmp1, Register tmp2) {
631 // Shenandoah has no 32-bit version for this.
632 Unimplemented();
633 }
634 #else
635 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm, DecoratorSet decorators,
636 Register res, Address addr, Register oldval, Register newval,
637 bool exchange, bool encode, Register tmp1, Register tmp2) {
638
639 if (!ShenandoahCASBarrier) {
640 BarrierSetAssembler::cmpxchg_oop(masm, decorators, res, addr, oldval, newval, exchange, encode, tmp1, tmp2);
641 return;
642 }
643
644 assert(ShenandoahCASBarrier, "Should only be used when CAS barrier is enabled");
645 assert(oldval == rax, "must be in rax for implicit use in cmpxchg");
646
647 Label retry, done;
648
649 // Apply storeval barrier to newval.
650 if (encode) {
651 if (newval == c_rarg1 && ShenandoahStoreValEnqueueBarrier) {
652 __ mov(tmp2, newval);
653 storeval_barrier(masm, tmp2, tmp1);
654 } else {
655 storeval_barrier(masm, newval, tmp1);
656 }
657 }
658
659 if (UseCompressedOops) {
660 if (encode) {
661 __ encode_heap_oop(oldval);
662 __ mov(rscratch1, newval);
663 __ encode_heap_oop(rscratch1);
664 newval = rscratch1;
665 }
666 }
667
668 // Remember oldval for retry logic below
669 if (UseCompressedOops) {
670 __ movl(tmp1, oldval);
671 } else {
672 __ movptr(tmp1, oldval);
673 }
674
675 // Step 1. Try to CAS with given arguments. If successful, then we are done,
676 // and can safely return.
677 if (os::is_MP()) __ lock();
678 if (UseCompressedOops) {
679 __ cmpxchgl(newval, addr);
680 } else {
681 __ cmpxchgptr(newval, addr);
682 }
683 __ jcc(Assembler::equal, done, true);
684
685 // Step 2. CAS had failed. This may be a false negative.
686 //
687 // The trouble comes when we compare the to-space pointer with the from-space
688 // pointer to the same object. To resolve this, it will suffice to read both
689 // oldval and the value from memory through the read barriers -- this will give
690 // both to-space pointers. If they mismatch, then it was a legitimate failure.
691 //
692 if (UseCompressedOops) {
693 __ decode_heap_oop(tmp1);
694 }
695 __ resolve_for_read(0, tmp1);
696
697 if (UseCompressedOops) {
698 __ movl(tmp2, oldval);
699 __ decode_heap_oop(tmp2);
700 } else {
701 __ movptr(tmp2, oldval);
702 }
703 __ resolve_for_read(0, tmp2);
704
705 __ cmpptr(tmp1, tmp2);
706 __ jcc(Assembler::notEqual, done, true);
707
708 // Step 3. Try to CAS again with resolved to-space pointers.
709 //
710 // Corner case: it may happen that somebody stored the from-space pointer
711 // to memory while we were preparing for retry. Therefore, we can fail again
712 // on retry, and so need to do this in loop, always re-reading the failure
713 // witness through the read barrier.
714 __ bind(retry);
715 if (os::is_MP()) __ lock();
716 if (UseCompressedOops) {
717 __ cmpxchgl(newval, addr);
718 } else {
719 __ cmpxchgptr(newval, addr);
720 }
721 __ jcc(Assembler::equal, done, true);
722
723 if (UseCompressedOops) {
724 __ movl(tmp2, oldval);
725 __ decode_heap_oop(tmp2);
726 } else {
727 __ movptr(tmp2, oldval);
728 }
729 __ resolve_for_read(0, tmp2);
730
731 __ cmpptr(tmp1, tmp2);
732 __ jcc(Assembler::equal, retry, true);
733
734 // Step 4. If we need a boolean result out of CAS, check the flag again,
735 // and promote the result. Note that we handle the flag from both the CAS
736 // itself and from the retry loop.
737 __ bind(done);
738 if (!exchange) {
739 assert(res != NULL, "need result register");
740 __ setb(Assembler::equal, res);
741 __ movzbl(res, res);
742 }
743 }
744 #endif // LP64
745
746 void ShenandoahBarrierSetAssembler::xchg_oop(MacroAssembler* masm, DecoratorSet decorators,
747 Register obj, Address addr, Register tmp) {
748 storeval_barrier(masm, obj, tmp);
749 BarrierSetAssembler::xchg_oop(masm, decorators, obj, addr, tmp);
750 }
751
752 #ifdef COMPILER1
753
754 #undef __
755 #define __ ce->masm()->
756
757 void ShenandoahBarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, ShenandoahPreBarrierStub* stub) {
758 ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
759 // At this point we know that marking is in progress.
760 // If do_load() is true then we have to emit the
761 // load of the previous value; otherwise it has already
762 // been loaded into _pre_val.
763
764 __ bind(*stub->entry());
765 assert(stub->pre_val()->is_register(), "Precondition.");
766
767 Register pre_val_reg = stub->pre_val()->as_register();
768
769 if (stub->do_load()) {
770 ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/, false /*unaligned*/);
771 }
772
773 __ cmpptr(pre_val_reg, (int32_t)NULL_WORD);
774 __ jcc(Assembler::equal, *stub->continuation());
775 ce->store_parameter(stub->pre_val()->as_register(), 0);
776 __ call(RuntimeAddress(bs->pre_barrier_c1_runtime_code_blob()->code_begin()));
777 __ jmp(*stub->continuation());
778
779 }
780
781 #undef __
782
783 #define __ sasm->
784
785 void ShenandoahBarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) {
786 __ prologue("shenandoah_pre_barrier", false);
787 // arg0 : previous value of memory
788
789 __ push(rax);
790 __ push(rdx);
791
792 const Register pre_val = rax;
793 const Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
794 const Register tmp = rdx;
795
796 NOT_LP64(__ get_thread(thread);)
797
798 Address queue_index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
799 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
800
801 Label done;
802 Label runtime;
803
804 // Is SATB still active?
805 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
806 __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL);
807 __ jcc(Assembler::zero, done);
808
809 // Can we store original value in the thread's buffer?
810
811 __ movptr(tmp, queue_index);
812 __ testptr(tmp, tmp);
813 __ jcc(Assembler::zero, runtime);
814 __ subptr(tmp, wordSize);
815 __ movptr(queue_index, tmp);
816 __ addptr(tmp, buffer);
817
818 // prev_val (rax)
819 __ load_parameter(0, pre_val);
820 __ movptr(Address(tmp, 0), pre_val);
821 __ jmp(done);
822
823 __ bind(runtime);
824
825 __ save_live_registers_no_oop_map(true);
826
827 // load the pre-value
828 __ load_parameter(0, rcx);
829 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), rcx, thread);
830
831 __ restore_live_registers(true);
832
833 __ bind(done);
834
835 __ pop(rdx);
836 __ pop(rax);
837
838 __ epilogue();
839 }
840
841 #undef __
842
843 #endif // COMPILER1
844
845 address ShenandoahBarrierSetAssembler::shenandoah_wb() {
846 assert(_shenandoah_wb != NULL, "need write barrier stub");
847 return _shenandoah_wb;
848 }
849
850 address ShenandoahBarrierSetAssembler::shenandoah_wb_C() {
851 assert(_shenandoah_wb_C != NULL, "need write barrier stub");
852 return _shenandoah_wb_C;
853 }
854
855 #define __ cgen->assembler()->
856
857 address ShenandoahBarrierSetAssembler::generate_shenandoah_wb(StubCodeGenerator* cgen, bool c_abi, bool do_cset_test) {
858 __ align(CodeEntryAlignment);
859 StubCodeMark mark(cgen, "StubRoutines", "shenandoah_wb");
860 address start = __ pc();
861
862 Label not_done;
863
864 // We use RDI, which also serves as argument register for slow call.
865 // RAX always holds the src object ptr, except after the slow call and
866 // the cmpxchg, then it holds the result.
867 // R8 and RCX are used as temporary registers.
868 if (!c_abi) {
869 __ push(rdi);
870 __ push(r8);
871 }
872
873 // Check for object beeing in the collection set.
874 // TODO: Can we use only 1 register here?
875 // The source object arrives here in rax.
876 // live: rax
877 // live: rdi
878 if (!c_abi) {
879 __ mov(rdi, rax);
880 } else {
881 if (rax != c_rarg0) {
882 __ mov(rax, c_rarg0);
883 }
884 }
885 if (do_cset_test) {
886 __ shrptr(rdi, ShenandoahHeapRegion::region_size_bytes_shift_jint());
887 // live: r8
888 __ movptr(r8, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr());
889 __ movbool(r8, Address(r8, rdi, Address::times_1));
890 // unlive: rdi
891 __ testbool(r8);
892 // unlive: r8
893 __ jccb(Assembler::notZero, not_done);
894
895 if (!c_abi) {
896 __ pop(r8);
897 __ pop(rdi);
898 }
899 __ ret(0);
900
901 __ bind(not_done);
902 }
903
904 if (!c_abi) {
905 __ push(rcx);
906 }
907
908 if (!c_abi) {
909 __ push(rdx);
910 __ push(rdi);
911 __ push(rsi);
912 __ push(r8);
913 __ push(r9);
914 __ push(r10);
915 __ push(r11);
916 __ push(r12);
917 __ push(r13);
918 __ push(r14);
919 __ push(r15);
920 }
921 __ save_vector_registers();
922 __ movptr(rdi, rax);
923 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_JRT), rdi);
924 __ restore_vector_registers();
925 if (!c_abi) {
926 __ pop(r15);
927 __ pop(r14);
928 __ pop(r13);
929 __ pop(r12);
930 __ pop(r11);
931 __ pop(r10);
932 __ pop(r9);
933 __ pop(r8);
934 __ pop(rsi);
935 __ pop(rdi);
936 __ pop(rdx);
937
938 __ pop(rcx);
939 __ pop(r8);
940 __ pop(rdi);
941 }
942 __ ret(0);
943
944 return start;
945 }
946
947 #undef __
948
949 void ShenandoahBarrierSetAssembler::barrier_stubs_init() {
950 if (ShenandoahWriteBarrier || ShenandoahStoreValEnqueueBarrier) {
951 int stub_code_size = 1536;
952 ResourceMark rm;
953 BufferBlob* bb = BufferBlob::create("shenandoah_barrier_stubs", stub_code_size);
954 CodeBuffer buf(bb);
955 StubCodeGenerator cgen(&buf);
956 _shenandoah_wb = generate_shenandoah_wb(&cgen, false, true);
957 _shenandoah_wb_C = generate_shenandoah_wb(&cgen, true, !ShenandoahWriteBarrierCsetTestInIR);
958 }
959 }