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/shenandoahHeap.hpp"
28 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
29 #include "gc/shenandoah/shenandoahHeuristics.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 & IS_DEST_UNINITIALIZED) != 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 && !ShenandoahHeap::heap()->heuristics()->can_do_traversal_gc()) {
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 #ifdef _LP64
262 // We move pre_val into c_rarg0 early, in order to avoid smashing it, should
263 // pre_val be c_rarg1 (where the call prologue would copy thread argument).
264 // Note: this should not accidentally smash thread, because thread is always r15.
265 assert(thread != c_rarg0, "smashed arg");
266 if (c_rarg0 != pre_val) {
267 __ mov(c_rarg0, pre_val);
268 }
269 #endif
270
271 if (expand_call) {
272 LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); )
273 #ifdef _LP64
274 if (c_rarg1 != thread) {
275 __ mov(c_rarg1, thread);
276 }
277 // Already moved pre_val into c_rarg0 above
278 #else
279 __ push(thread);
280 __ push(pre_val);
281 #endif
282 __ MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), 2);
283 } else {
284 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), LP64_ONLY(c_rarg0) NOT_LP64(pre_val), thread);
285 }
286
287 NOT_LP64( __ pop(thread); )
288
289 // save the live input values
290 if (pre_val != rax)
291 __ pop(pre_val);
292
293 if (obj != noreg && obj != rax)
294 __ pop(obj);
295
296 if(tosca_live) __ pop(rax);
297
298 __ bind(done);
299 }
300
301 void ShenandoahBarrierSetAssembler::read_barrier(MacroAssembler* masm, Register dst) {
302 if (ShenandoahReadBarrier) {
303 read_barrier_impl(masm, dst);
304 }
305 }
306
307 void ShenandoahBarrierSetAssembler::read_barrier_impl(MacroAssembler* masm, Register dst) {
308 assert(UseShenandoahGC && (ShenandoahReadBarrier || ShenandoahStoreValReadBarrier || ShenandoahCASBarrier), "should be enabled");
309 Label is_null;
310 __ testptr(dst, dst);
311 __ jcc(Assembler::zero, is_null);
312 read_barrier_not_null_impl(masm, dst);
313 __ bind(is_null);
314 }
315
316 void ShenandoahBarrierSetAssembler::read_barrier_not_null(MacroAssembler* masm, Register dst) {
317 if (ShenandoahReadBarrier) {
318 read_barrier_not_null_impl(masm, dst);
319 }
320 }
321
322 void ShenandoahBarrierSetAssembler::read_barrier_not_null_impl(MacroAssembler* masm, Register dst) {
323 assert(UseShenandoahGC && (ShenandoahReadBarrier || ShenandoahStoreValReadBarrier || ShenandoahCASBarrier), "should be enabled");
324 __ movptr(dst, Address(dst, BrooksPointer::byte_offset()));
325 }
326
327
328 void ShenandoahBarrierSetAssembler::write_barrier(MacroAssembler* masm, Register dst) {
329 if (ShenandoahWriteBarrier) {
330 write_barrier_impl(masm, dst);
331 }
332 }
333
334 void ShenandoahBarrierSetAssembler::write_barrier_impl(MacroAssembler* masm, Register dst) {
335 assert(UseShenandoahGC && (ShenandoahWriteBarrier || ShenandoahStoreValEnqueueBarrier), "Should be enabled");
336 #ifdef _LP64
337 Label done;
338
339 Address gc_state(r15_thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
340 __ testb(gc_state, ShenandoahHeap::HAS_FORWARDED | ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL);
341 __ jccb(Assembler::zero, done);
342
343 // Heap is unstable, need to perform the read-barrier even if WB is inactive
344 if (ShenandoahWriteBarrierRB) {
345 read_barrier_not_null(masm, dst);
346 }
347
348 __ testb(gc_state, ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL);
349 __ jccb(Assembler::zero, done);
350
351 if (dst != rax) {
352 __ xchgptr(dst, rax); // Move obj into rax and save rax into obj.
353 }
354
355 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, ShenandoahBarrierSetAssembler::shenandoah_wb())));
356
357 if (dst != rax) {
358 __ xchgptr(rax, dst); // Swap back obj with rax.
359 }
360
361 __ bind(done);
362 #else
363 Unimplemented();
364 #endif
365 }
366
367 void ShenandoahBarrierSetAssembler::storeval_barrier(MacroAssembler* masm, Register dst, Register tmp) {
368 if (ShenandoahStoreValReadBarrier || ShenandoahStoreValEnqueueBarrier) {
369 storeval_barrier_impl(masm, dst, tmp);
370 }
371 }
372
373 void ShenandoahBarrierSetAssembler::storeval_barrier_impl(MacroAssembler* masm, Register dst, Register tmp) {
374 assert(UseShenandoahGC && (ShenandoahStoreValReadBarrier || ShenandoahStoreValEnqueueBarrier), "should be enabled");
375
376 if (dst == noreg) return;
377
378 #ifdef _LP64
379 if (ShenandoahStoreValEnqueueBarrier) {
380 Label is_null;
381 __ testptr(dst, dst);
382 __ jcc(Assembler::zero, is_null);
383 write_barrier_impl(masm, dst);
384 __ bind(is_null);
385
386 // The set of registers to be saved+restored is the same as in the write-barrier above.
387 // Those are the commonly used registers in the interpreter.
388 __ pusha();
389 // __ push_callee_saved_registers();
390 __ subptr(rsp, 2 * Interpreter::stackElementSize);
391 __ movdbl(Address(rsp, 0), xmm0);
392
393 satb_write_barrier_pre(masm, noreg, dst, r15_thread, tmp, true, false);
394 __ movdbl(xmm0, Address(rsp, 0));
395 __ addptr(rsp, 2 * Interpreter::stackElementSize);
396 //__ pop_callee_saved_registers();
397 __ popa();
398 }
399 if (ShenandoahStoreValReadBarrier) {
400 read_barrier_impl(masm, dst);
401 }
402 #else
403 Unimplemented();
404 #endif
405 }
406
407 void ShenandoahBarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
408 Register dst, Address src, Register tmp1, Register tmp_thread) {
409 bool on_oop = type == T_OBJECT || type == T_ARRAY;
410 bool in_heap = (decorators & IN_HEAP) != 0;
411 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
412 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
413 bool on_reference = on_weak || on_phantom;
414 if (in_heap) {
415 read_barrier_not_null(masm, src.base());
416 }
417 BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp_thread);
418 if (ShenandoahKeepAliveBarrier && on_oop && on_reference) {
419 const Register thread = NOT_LP64(tmp_thread) LP64_ONLY(r15_thread);
420 NOT_LP64(__ get_thread(thread));
421
422 // Generate the SATB pre-barrier code to log the value of
423 // the referent field in an SATB buffer.
424 shenandoah_write_barrier_pre(masm /* masm */,
425 noreg /* obj */,
426 dst /* pre_val */,
427 thread /* thread */,
428 tmp1 /* tmp */,
429 true /* tosca_live */,
430 true /* expand_call */);
431 }
432 }
433
434 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
435 Address dst, Register val, Register tmp1, Register tmp2) {
436
437 bool in_heap = (decorators & IN_HEAP) != 0;
438 bool as_normal = (decorators & AS_NORMAL) != 0;
439 if (in_heap) {
440 write_barrier(masm, dst.base());
441 }
442 if (type == T_OBJECT || type == T_ARRAY) {
443 bool needs_pre_barrier = as_normal;
444
445 Register tmp3 = LP64_ONLY(r8) NOT_LP64(rsi);
446 Register rthread = LP64_ONLY(r15_thread) NOT_LP64(rcx);
447 // flatten object address if needed
448 // We do it regardless of precise because we need the registers
449 if (dst.index() == noreg && dst.disp() == 0) {
450 if (dst.base() != tmp1) {
451 __ movptr(tmp1, dst.base());
452 }
453 } else {
454 __ lea(tmp1, dst);
455 }
456
457 #ifndef _LP64
458 InterpreterMacroAssembler *imasm = static_cast<InterpreterMacroAssembler*>(masm);
459 #endif
460
461 NOT_LP64(__ get_thread(rcx));
462 NOT_LP64(imasm->save_bcp());
463
464 if (needs_pre_barrier) {
465 shenandoah_write_barrier_pre(masm /*masm*/,
466 tmp1 /* obj */,
467 tmp2 /* pre_val */,
468 rthread /* thread */,
469 tmp3 /* tmp */,
470 val != noreg /* tosca_live */,
471 false /* expand_call */);
472 }
473 if (val == noreg) {
474 BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
475 } else {
476 storeval_barrier(masm, val, tmp3);
477 BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
478 }
479 NOT_LP64(imasm->restore_bcp());
480 } else {
481 BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2);
482 }
483 }
484
485 #ifndef _LP64
486 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm,
487 Address obj1, jobject obj2) {
488 Unimplemented();
489 }
490
491 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm,
492 Register obj1, jobject obj2) {
493 Unimplemented();
494 }
495 #endif
496
497
498 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm, Register op1, Register op2) {
499 __ cmpptr(op1, op2);
500 if (ShenandoahAcmpBarrier) {
501 Label done;
502 __ jccb(Assembler::equal, done);
503 read_barrier(masm, op1);
504 read_barrier(masm, op2);
505 __ cmpptr(op1, op2);
506 __ bind(done);
507 }
508 }
509
510 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm, Register src1, Address src2) {
511 __ cmpptr(src1, src2);
512 if (ShenandoahAcmpBarrier) {
513 Label done;
514 __ jccb(Assembler::equal, done);
515 __ movptr(rscratch2, src2);
516 read_barrier(masm, src1);
517 read_barrier(masm, rscratch2);
518 __ cmpptr(src1, rscratch2);
519 __ bind(done);
520 }
521 }
522
523 void ShenandoahBarrierSetAssembler::tlab_allocate(MacroAssembler* masm,
524 Register thread, Register obj,
525 Register var_size_in_bytes,
526 int con_size_in_bytes,
527 Register t1, Register t2,
528 Label& slow_case) {
529 assert_different_registers(obj, t1, t2);
530 assert_different_registers(obj, var_size_in_bytes, t1);
531 Register end = t2;
532 if (!thread->is_valid()) {
533 #ifdef _LP64
534 thread = r15_thread;
535 #else
536 assert(t1->is_valid(), "need temp reg");
537 thread = t1;
538 __ get_thread(thread);
539 #endif
540 }
541
542 __ verify_tlab();
543
544 __ movptr(obj, Address(thread, JavaThread::tlab_top_offset()));
545 if (var_size_in_bytes == noreg) {
546 __ lea(end, Address(obj, con_size_in_bytes + BrooksPointer::byte_size()));
547 } else {
548 __ addptr(var_size_in_bytes, BrooksPointer::byte_size());
549 __ lea(end, Address(obj, var_size_in_bytes, Address::times_1));
550 }
551 __ cmpptr(end, Address(thread, JavaThread::tlab_end_offset()));
552 __ jcc(Assembler::above, slow_case);
553
554 // update the tlab top pointer
555 __ movptr(Address(thread, JavaThread::tlab_top_offset()), end);
556
557 // Initialize brooks pointer
558 #ifdef _LP64
559 __ incrementq(obj, BrooksPointer::byte_size());
560 #else
561 __ incrementl(obj, BrooksPointer::byte_size());
562 #endif
563 __ movptr(Address(obj, BrooksPointer::byte_offset()), obj);
564
565 // recover var_size_in_bytes if necessary
566 if (var_size_in_bytes == end) {
567 __ subptr(var_size_in_bytes, obj);
568 }
569 __ verify_tlab();
570 }
571
572 void ShenandoahBarrierSetAssembler::resolve(MacroAssembler* masm, DecoratorSet decorators, Register obj) {
573 bool oop_not_null = (decorators & IS_NOT_NULL) != 0;
574 bool is_write = (decorators & ACCESS_WRITE) != 0;
575 if (is_write) {
576 if (oop_not_null) {
577 write_barrier(masm, obj);
578 } else {
579 Label done;
580 __ testptr(obj, obj);
581 __ jcc(Assembler::zero, done);
582 write_barrier(masm, obj);
583 __ bind(done);
584 }
585 } else {
586 if (oop_not_null) {
587 read_barrier_not_null(masm, obj);
588 } else {
589 read_barrier(masm, obj);
590 }
591 }
592 }
593
594 // Special Shenandoah CAS implementation that handles false negatives
595 // due to concurrent evacuation.
596 #ifndef _LP64
597 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm,
598 Register res, Address addr, Register oldval, Register newval,
599 bool exchange, bool encode, Register tmp1, Register tmp2) {
600 // Shenandoah has no 32-bit version for this.
601 Unimplemented();
602 }
603 #else
604 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm,
605 Register res, Address addr, Register oldval, Register newval,
606 bool exchange, bool encode, Register tmp1, Register tmp2) {
607 if (!ShenandoahCASBarrier) {
608 #ifdef _LP64
609 if (UseCompressedOops) {
610 if (encode) {
611 __ encode_heap_oop(oldval);
612 __ mov(rscratch1, newval);
613 __ encode_heap_oop(rscratch1);
614 newval = rscratch1;
615 }
616 if (os::is_MP()) {
617 __ lock();
618 }
619 // oldval (rax) is implicitly used by this instruction
620 __ cmpxchgl(newval, addr);
621 } else
622 #endif
623 {
624 if (os::is_MP()) {
625 __ lock();
626 }
627 __ cmpxchgptr(newval, addr);
628 }
629
630 if (!exchange) {
631 assert(res != NULL, "need result register");
632 __ setb(Assembler::equal, res);
633 __ movzbl(res, res);
634 }
635 return;
636 }
637
638 assert(ShenandoahCASBarrier, "Should only be used when CAS barrier is enabled");
639 assert(oldval == rax, "must be in rax for implicit use in cmpxchg");
640
641 Label retry, done;
642
643 // Apply storeval barrier to newval.
644 if (encode) {
645 storeval_barrier(masm, newval, tmp1);
646 }
647
648 if (UseCompressedOops) {
649 if (encode) {
650 __ encode_heap_oop(oldval);
651 __ mov(rscratch1, newval);
652 __ encode_heap_oop(rscratch1);
653 newval = rscratch1;
654 }
655 }
656
657 // Remember oldval for retry logic below
658 if (UseCompressedOops) {
659 __ movl(tmp1, oldval);
660 } else {
661 __ movptr(tmp1, oldval);
662 }
663
664 // Step 1. Try to CAS with given arguments. If successful, then we are done,
665 // and can safely return.
666 if (os::is_MP()) __ lock();
667 if (UseCompressedOops) {
668 __ cmpxchgl(newval, addr);
669 } else {
670 __ cmpxchgptr(newval, addr);
671 }
672 __ jcc(Assembler::equal, done, true);
673
674 // Step 2. CAS had failed. This may be a false negative.
675 //
676 // The trouble comes when we compare the to-space pointer with the from-space
677 // pointer to the same object. To resolve this, it will suffice to read both
678 // oldval and the value from memory through the read barriers -- this will give
679 // both to-space pointers. If they mismatch, then it was a legitimate failure.
680 //
681 if (UseCompressedOops) {
682 __ decode_heap_oop(tmp1);
683 }
684 read_barrier_impl(masm, tmp1);
685
686 if (UseCompressedOops) {
687 __ movl(tmp2, oldval);
688 __ decode_heap_oop(tmp2);
689 } else {
690 __ movptr(tmp2, oldval);
691 }
692 read_barrier_impl(masm, tmp2);
693
694 __ cmpptr(tmp1, tmp2);
695 __ jcc(Assembler::notEqual, done, true);
696
697 // Step 3. Try to CAS again with resolved to-space pointers.
698 //
699 // Corner case: it may happen that somebody stored the from-space pointer
700 // to memory while we were preparing for retry. Therefore, we can fail again
701 // on retry, and so need to do this in loop, always re-reading the failure
702 // witness through the read barrier.
703 __ bind(retry);
704 if (os::is_MP()) __ lock();
705 if (UseCompressedOops) {
706 __ cmpxchgl(newval, addr);
707 } else {
708 __ cmpxchgptr(newval, addr);
709 }
710 __ jcc(Assembler::equal, done, true);
711
712 if (UseCompressedOops) {
713 __ movl(tmp2, oldval);
714 __ decode_heap_oop(tmp2);
715 } else {
716 __ movptr(tmp2, oldval);
717 }
718 read_barrier_impl(masm, tmp2);
719
720 __ cmpptr(tmp1, tmp2);
721 __ jcc(Assembler::equal, retry, true);
722
723 // Step 4. If we need a boolean result out of CAS, check the flag again,
724 // and promote the result. Note that we handle the flag from both the CAS
725 // itself and from the retry loop.
726 __ bind(done);
727 if (!exchange) {
728 assert(res != NULL, "need result register");
729 __ setb(Assembler::equal, res);
730 __ movzbl(res, res);
731 }
732 }
733 #endif // LP64
734
735 #ifdef COMPILER1
736
737 #undef __
738 #define __ ce->masm()->
739
740 void ShenandoahBarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, ShenandoahPreBarrierStub* stub) {
741 ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
742 // At this point we know that marking is in progress.
743 // If do_load() is true then we have to emit the
744 // load of the previous value; otherwise it has already
745 // been loaded into _pre_val.
746
747 __ bind(*stub->entry());
748 assert(stub->pre_val()->is_register(), "Precondition.");
749
750 Register pre_val_reg = stub->pre_val()->as_register();
751
752 if (stub->do_load()) {
753 ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/, false /*unaligned*/);
754 }
755
756 __ cmpptr(pre_val_reg, (int32_t)NULL_WORD);
757 __ jcc(Assembler::equal, *stub->continuation());
758 ce->store_parameter(stub->pre_val()->as_register(), 0);
759 __ call(RuntimeAddress(bs->pre_barrier_c1_runtime_code_blob()->code_begin()));
760 __ jmp(*stub->continuation());
761
762 }
763
764 void ShenandoahBarrierSetAssembler::gen_write_barrier_stub(LIR_Assembler* ce, ShenandoahWriteBarrierStub* stub) {
765 __ bind(*stub->entry());
766
767 Label done;
768 Register obj = stub->obj()->as_register();
769 Register res = stub->result()->as_register();
770
771 if (res != obj) {
772 __ mov(res, obj);
773 }
774
775 // Check for null.
776 if (stub->needs_null_check()) {
777 __ testptr(res, res);
778 __ jcc(Assembler::zero, done);
779 }
780
781 write_barrier(ce->masm(), res);
782
783 __ bind(done);
784 __ jmp(*stub->continuation());
785 }
786
787 #undef __
788
789 #define __ sasm->
790
791 void ShenandoahBarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) {
792 __ prologue("shenandoah_pre_barrier", false);
793 // arg0 : previous value of memory
794
795 __ push(rax);
796 __ push(rdx);
797
798 const Register pre_val = rax;
799 const Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
800 const Register tmp = rdx;
801
802 NOT_LP64(__ get_thread(thread);)
803
804 Address queue_index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
805 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
806
807 Label done;
808 Label runtime;
809
810 // Is SATB still active?
811 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
812 __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL);
813 __ jcc(Assembler::zero, done);
814
815 // Can we store original value in the thread's buffer?
816
817 __ movptr(tmp, queue_index);
818 __ testptr(tmp, tmp);
819 __ jcc(Assembler::zero, runtime);
820 __ subptr(tmp, wordSize);
821 __ movptr(queue_index, tmp);
822 __ addptr(tmp, buffer);
823
824 // prev_val (rax)
825 __ load_parameter(0, pre_val);
826 __ movptr(Address(tmp, 0), pre_val);
827 __ jmp(done);
828
829 __ bind(runtime);
830
831 __ save_live_registers_no_oop_map(true);
832
833 // load the pre-value
834 __ load_parameter(0, rcx);
835 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), rcx, thread);
836
837 __ restore_live_registers(true);
838
839 __ bind(done);
840
841 __ pop(rdx);
842 __ pop(rax);
843
844 __ epilogue();
845 }
846
847 #undef __
848
849 #endif // COMPILER1
850
851 address ShenandoahBarrierSetAssembler::shenandoah_wb() {
852 assert(_shenandoah_wb != NULL, "need write barrier stub");
853 return _shenandoah_wb;
854 }
855
856 address ShenandoahBarrierSetAssembler::shenandoah_wb_C() {
857 assert(_shenandoah_wb_C != NULL, "need write barrier stub");
858 return _shenandoah_wb_C;
859 }
860
861 #define __ cgen->assembler()->
862
863 address ShenandoahBarrierSetAssembler::generate_shenandoah_wb(StubCodeGenerator* cgen, bool c_abi, bool do_cset_test) {
864 __ align(CodeEntryAlignment);
865 StubCodeMark mark(cgen, "StubRoutines", "shenandoah_wb");
866 address start = __ pc();
867
868 #ifdef _LP64
869 Label not_done;
870
871 // We use RDI, which also serves as argument register for slow call.
872 // RAX always holds the src object ptr, except after the slow call and
873 // the cmpxchg, then it holds the result.
874 // R8 and RCX are used as temporary registers.
875 if (!c_abi) {
876 __ push(rdi);
877 __ push(r8);
878 }
879
880 // Check for object beeing in the collection set.
881 // TODO: Can we use only 1 register here?
882 // The source object arrives here in rax.
883 // live: rax
884 // live: rdi
885 if (!c_abi) {
886 __ mov(rdi, rax);
887 } else {
888 if (rax != c_rarg0) {
889 __ mov(rax, c_rarg0);
890 }
891 }
892 if (do_cset_test) {
893 __ shrptr(rdi, ShenandoahHeapRegion::region_size_bytes_shift_jint());
894 // live: r8
895 __ movptr(r8, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr());
896 __ movbool(r8, Address(r8, rdi, Address::times_1));
897 // unlive: rdi
898 __ testbool(r8);
899 // unlive: r8
900 __ jccb(Assembler::notZero, not_done);
901
902 if (!c_abi) {
903 __ pop(r8);
904 __ pop(rdi);
905 }
906 __ ret(0);
907
908 __ bind(not_done);
909 }
910
911 if (!c_abi) {
912 __ push(rcx);
913 }
914
915 if (!c_abi) {
916 __ push(rdx);
917 __ push(rdi);
918 __ push(rsi);
919 __ push(r8);
920 __ push(r9);
921 __ push(r10);
922 __ push(r11);
923 __ push(r12);
924 __ push(r13);
925 __ push(r14);
926 __ push(r15);
927 }
928 __ save_vector_registers();
929 __ movptr(rdi, rax);
930 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_JRT), rdi);
931 __ restore_vector_registers();
932 if (!c_abi) {
933 __ pop(r15);
934 __ pop(r14);
935 __ pop(r13);
936 __ pop(r12);
937 __ pop(r11);
938 __ pop(r10);
939 __ pop(r9);
940 __ pop(r8);
941 __ pop(rsi);
942 __ pop(rdi);
943 __ pop(rdx);
944
945 __ pop(rcx);
946 __ pop(r8);
947 __ pop(rdi);
948 }
949 __ ret(0);
950 #else
951 ShouldNotReachHere();
952 #endif
953 return start;
954 }
955
956 #undef __
957
958 void ShenandoahBarrierSetAssembler::barrier_stubs_init() {
959 if (ShenandoahWriteBarrier || ShenandoahStoreValEnqueueBarrier) {
960 int stub_code_size = 4096;
961 ResourceMark rm;
962 BufferBlob* bb = BufferBlob::create("shenandoah_barrier_stubs", stub_code_size);
963 CodeBuffer buf(bb);
964 StubCodeGenerator cgen(&buf);
965 _shenandoah_wb = generate_shenandoah_wb(&cgen, false, true);
966 _shenandoah_wb_C = generate_shenandoah_wb(&cgen, true, !ShenandoahWriteBarrierCsetTestInIR);
967 }
968 }