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/shenandoahBarrierSetAssembler.hpp"
26 #include "gc/shenandoah/shenandoahHeap.hpp"
27 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
28 #include "gc/shenandoah/shenandoahHeuristics.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 address ShenandoahBarrierSetAssembler::_shenandoah_wb = NULL;
45 address ShenandoahBarrierSetAssembler::_shenandoah_wb_C = NULL;
46
47 void ShenandoahBarrierSetAssembler::arraycopy_prologue(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
48 Register src, Register dst, Register count) {
49
50 bool checkcast = (decorators & ARRAYCOPY_CHECKCAST) != 0;
51 bool disjoint = (decorators & ARRAYCOPY_DISJOINT) != 0;
52 bool obj_int = type == T_OBJECT LP64_ONLY(&& UseCompressedOops);
53 bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0;
54
55 if (type == T_OBJECT || type == T_ARRAY) {
56 #ifdef _LP64
57 if (!checkcast && !obj_int) {
58 // Save count for barrier
59 __ movptr(r11, count);
60 } else if (disjoint && obj_int) {
61 // Save dst in r11 in the disjoint case
62 __ movq(r11, dst);
63 }
64 #else
65 if (disjoint) {
66 __ mov(rdx, dst); // save 'to'
67 }
68 #endif
69
70 if (!dest_uninitialized && !ShenandoahHeap::heap()->heuristics()->can_do_traversal_gc()) {
71 Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
72 #ifndef _LP64
73 __ push(thread);
74 __ get_thread(thread);
75 #endif
76
77 Label filtered;
78 Address in_progress(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()));
79 // Is marking active?
80 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
81 __ cmpl(in_progress, 0);
82 } else {
83 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");
84 __ cmpb(in_progress, 0);
85 }
86
87 NOT_LP64(__ pop(thread);)
88
89 __ jcc(Assembler::equal, filtered);
90
91 __ pusha(); // push registers
92 #ifdef _LP64
93 if (count == c_rarg0) {
94 if (dst == c_rarg1) {
95 // exactly backwards!!
96 __ xchgptr(c_rarg1, c_rarg0);
97 } else {
98 __ movptr(c_rarg1, count);
99 __ movptr(c_rarg0, dst);
100 }
101 } else {
102 __ movptr(c_rarg0, dst);
103 __ movptr(c_rarg1, count);
104 }
105 if (UseCompressedOops) {
106 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_narrow_oop_entry), 2);
107 } else {
108 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry), 2);
109 }
110 #else
111 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry),
112 dst, count);
113 #endif
114 __ popa();
115 __ bind(filtered);
116 }
117 }
118
119 }
120
121 void ShenandoahBarrierSetAssembler::arraycopy_epilogue(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
122 Register src, Register dst, Register count) {
123 bool checkcast = (decorators & ARRAYCOPY_CHECKCAST) != 0;
124 bool disjoint = (decorators & ARRAYCOPY_DISJOINT) != 0;
125 bool obj_int = type == T_OBJECT LP64_ONLY(&& UseCompressedOops);
126 Register tmp = rax;
127
128 if (type == T_OBJECT || type == T_ARRAY) {
129 #ifdef _LP64
130 if (!checkcast && !obj_int) {
131 // Save count for barrier
132 count = r11;
133 } else if (disjoint && obj_int) {
134 // Use the saved dst in the disjoint case
135 dst = r11;
136 } else if (checkcast) {
137 tmp = rscratch1;
138 }
139 #else
140 if (disjoint) {
141 __ mov(dst, rdx); // restore 'to'
142 }
143 #endif
144
145 __ pusha(); // push registers (overkill)
146 #ifdef _LP64
147 if (c_rarg0 == count) { // On win64 c_rarg0 == rcx
148 assert_different_registers(c_rarg1, dst);
149 __ mov(c_rarg1, count);
150 __ mov(c_rarg0, dst);
151 } else {
152 assert_different_registers(c_rarg0, count);
153 __ mov(c_rarg0, dst);
154 __ mov(c_rarg1, count);
155 }
156 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry), 2);
157 #else
158 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry),
159 dst, count);
160 #endif
161 __ popa();
162 }
163 }
164
165 void ShenandoahBarrierSetAssembler::shenandoah_write_barrier_pre(MacroAssembler* masm,
166 Register obj,
167 Register pre_val,
168 Register thread,
169 Register tmp,
170 bool tosca_live,
171 bool expand_call) {
172
173 if (ShenandoahSATBBarrier) {
174 satb_write_barrier_pre(masm, obj, pre_val, thread, tmp, tosca_live, expand_call);
175 }
176 }
177
178 void ShenandoahBarrierSetAssembler::satb_write_barrier_pre(MacroAssembler* masm,
179 Register obj,
180 Register pre_val,
181 Register thread,
182 Register tmp,
183 bool tosca_live,
184 bool expand_call) {
185 // If expand_call is true then we expand the call_VM_leaf macro
186 // directly to skip generating the check by
187 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
188
189 #ifdef _LP64
190 assert(thread == r15_thread, "must be");
191 #endif // _LP64
192
193 Label done;
194 Label runtime;
195
196 assert(pre_val != noreg, "check this code");
197
198 if (obj != noreg) {
199 assert_different_registers(obj, pre_val, tmp);
200 assert(pre_val != rax, "check this code");
201 }
202
203 Address in_progress(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()));
204 Address index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
205 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
206
207 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
208 __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL);
209 __ jcc(Assembler::zero, done);
210
211 // Do we need to load the previous value?
212 if (obj != noreg) {
213 __ load_heap_oop(pre_val, Address(obj, 0), noreg, noreg, AS_RAW);
214 }
215
216 // Is the previous value null?
217 __ cmpptr(pre_val, (int32_t) NULL_WORD);
218 __ jcc(Assembler::equal, done);
219
220 // Can we store original value in the thread's buffer?
221 // Is index == 0?
222 // (The index field is typed as size_t.)
223
224 __ movptr(tmp, index); // tmp := *index_adr
225 __ cmpptr(tmp, 0); // tmp == 0?
226 __ jcc(Assembler::equal, runtime); // If yes, goto runtime
227
228 __ subptr(tmp, wordSize); // tmp := tmp - wordSize
229 __ movptr(index, tmp); // *index_adr := tmp
230 __ addptr(tmp, buffer); // tmp := tmp + *buffer_adr
231
232 // Record the previous value
233 __ movptr(Address(tmp, 0), pre_val);
234 __ jmp(done);
235
236 __ bind(runtime);
237 // save the live input values
238 if(tosca_live) __ push(rax);
239
240 if (obj != noreg && obj != rax)
241 __ push(obj);
242
243 if (pre_val != rax)
244 __ push(pre_val);
245
246 // Calling the runtime using the regular call_VM_leaf mechanism generates
247 // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
248 // that checks that the *(ebp+frame::interpreter_frame_last_sp) == NULL.
249 //
250 // If we care generating the pre-barrier without a frame (e.g. in the
251 // intrinsified Reference.get() routine) then ebp might be pointing to
252 // the caller frame and so this check will most likely fail at runtime.
253 //
254 // Expanding the call directly bypasses the generation of the check.
255 // So when we do not have have a full interpreter frame on the stack
256 // expand_call should be passed true.
257
258 NOT_LP64( __ push(thread); )
259
260 #ifdef _LP64
261 // We move pre_val into c_rarg0 early, in order to avoid smashing it, should
262 // pre_val be c_rarg1 (where the call prologue would copy thread argument).
263 // Note: this should not accidentally smash thread, because thread is always r15.
264 assert(thread != c_rarg0, "smashed arg");
265 if (c_rarg0 != pre_val) {
266 __ mov(c_rarg0, pre_val);
267 }
268 #endif
269
270 if (expand_call) {
271 LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); )
272 #ifdef _LP64
273 if (c_rarg1 != thread) {
274 __ mov(c_rarg1, thread);
275 }
276 // Already moved pre_val into c_rarg0 above
277 #else
278 __ push(thread);
279 __ push(pre_val);
280 #endif
281 __ MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), 2);
282 } else {
283 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), LP64_ONLY(c_rarg0) NOT_LP64(pre_val), thread);
284 }
285
286 NOT_LP64( __ pop(thread); )
287
288 // save the live input values
289 if (pre_val != rax)
290 __ pop(pre_val);
291
292 if (obj != noreg && obj != rax)
293 __ pop(obj);
294
295 if(tosca_live) __ pop(rax);
296
297 __ bind(done);
298 }
299
300 void ShenandoahBarrierSetAssembler::read_barrier(MacroAssembler* masm, Register dst) {
301 if (ShenandoahReadBarrier) {
302 read_barrier_impl(masm, dst);
303 }
304 }
305
306 void ShenandoahBarrierSetAssembler::read_barrier_impl(MacroAssembler* masm, Register dst) {
307 assert(UseShenandoahGC && (ShenandoahReadBarrier || ShenandoahStoreValReadBarrier || ShenandoahCASBarrier), "should be enabled");
308 Label is_null;
309 __ testptr(dst, dst);
310 __ jcc(Assembler::zero, is_null);
311 read_barrier_not_null_impl(masm, dst);
312 __ bind(is_null);
313 }
314
315 void ShenandoahBarrierSetAssembler::read_barrier_not_null(MacroAssembler* masm, Register dst) {
316 if (ShenandoahReadBarrier) {
317 read_barrier_not_null_impl(masm, dst);
318 }
319 }
320
321 void ShenandoahBarrierSetAssembler::read_barrier_not_null_impl(MacroAssembler* masm, Register dst) {
322 assert(UseShenandoahGC && (ShenandoahReadBarrier || ShenandoahStoreValReadBarrier || ShenandoahCASBarrier), "should be enabled");
323 __ movptr(dst, Address(dst, ShenandoahBrooksPointer::byte_offset()));
324 }
325
326
327 void ShenandoahBarrierSetAssembler::write_barrier(MacroAssembler* masm, Register dst) {
328 if (ShenandoahWriteBarrier) {
329 write_barrier_impl(masm, dst);
330 }
331 }
332
333 void ShenandoahBarrierSetAssembler::write_barrier_impl(MacroAssembler* masm, Register dst) {
334 assert(UseShenandoahGC && (ShenandoahWriteBarrier || ShenandoahStoreValEnqueueBarrier), "Should be enabled");
335 #ifdef _LP64
336 Label done;
337
338 Address gc_state(r15_thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
339 __ testb(gc_state, ShenandoahHeap::HAS_FORWARDED | ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL);
340 __ jccb(Assembler::zero, done);
341
342 // Heap is unstable, need to perform the read-barrier even if WB is inactive
343 if (ShenandoahWriteBarrierRB) {
344 read_barrier_not_null(masm, dst);
345 }
346
347 __ testb(gc_state, ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL);
348 __ jccb(Assembler::zero, done);
349
350 if (dst != rax) {
351 __ xchgptr(dst, rax); // Move obj into rax and save rax into obj.
352 }
353
354 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, ShenandoahBarrierSetAssembler::shenandoah_wb())));
355
356 if (dst != rax) {
357 __ xchgptr(rax, dst); // Swap back obj with rax.
358 }
359
360 __ bind(done);
361 #else
362 Unimplemented();
363 #endif
364 }
365
366 void ShenandoahBarrierSetAssembler::storeval_barrier(MacroAssembler* masm, Register dst, Register tmp) {
367 if (ShenandoahStoreValReadBarrier || ShenandoahStoreValEnqueueBarrier) {
368 storeval_barrier_impl(masm, dst, tmp);
369 }
370 }
371
372 void ShenandoahBarrierSetAssembler::storeval_barrier_impl(MacroAssembler* masm, Register dst, Register tmp) {
373 assert(UseShenandoahGC && (ShenandoahStoreValReadBarrier || ShenandoahStoreValEnqueueBarrier), "should be enabled");
374
375 if (dst == noreg) return;
376
377 #ifdef _LP64
378 if (ShenandoahStoreValEnqueueBarrier) {
379 Label is_null;
380 __ testptr(dst, dst);
381 __ jcc(Assembler::zero, is_null);
382 write_barrier_impl(masm, dst);
383 __ bind(is_null);
384
385 // The set of registers to be saved+restored is the same as in the write-barrier above.
386 // Those are the commonly used registers in the interpreter.
387 __ pusha();
388 // __ push_callee_saved_registers();
389 __ subptr(rsp, 2 * Interpreter::stackElementSize);
390 __ movdbl(Address(rsp, 0), xmm0);
391
392 satb_write_barrier_pre(masm, noreg, dst, r15_thread, tmp, true, false);
393 __ movdbl(xmm0, Address(rsp, 0));
394 __ addptr(rsp, 2 * Interpreter::stackElementSize);
395 //__ pop_callee_saved_registers();
396 __ popa();
397 }
398 if (ShenandoahStoreValReadBarrier) {
399 read_barrier_impl(masm, dst);
400 }
401 #else
402 Unimplemented();
403 #endif
404 }
405
406 void ShenandoahBarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
407 Register dst, Address src, Register tmp1, Register tmp_thread) {
408 bool on_oop = type == T_OBJECT || type == T_ARRAY;
409 bool in_heap = (decorators & IN_HEAP) != 0;
410 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
411 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
412 bool on_reference = on_weak || on_phantom;
413 if (in_heap) {
414 read_barrier_not_null(masm, src.base());
415 }
416 BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp_thread);
417 if (ShenandoahKeepAliveBarrier && on_oop && on_reference) {
418 const Register thread = NOT_LP64(tmp_thread) LP64_ONLY(r15_thread);
419 NOT_LP64(__ get_thread(thread));
420
421 // Generate the SATB pre-barrier code to log the value of
422 // the referent field in an SATB buffer.
423 shenandoah_write_barrier_pre(masm /* masm */,
424 noreg /* obj */,
425 dst /* pre_val */,
426 thread /* thread */,
427 tmp1 /* tmp */,
428 true /* tosca_live */,
429 true /* expand_call */);
430 }
431 }
432
433 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
434 Address dst, Register val, Register tmp1, Register tmp2) {
435
436 bool in_heap = (decorators & IN_HEAP) != 0;
437 bool as_normal = (decorators & AS_NORMAL) != 0;
438 if (in_heap) {
439 write_barrier(masm, dst.base());
440 }
441 if (type == T_OBJECT || type == T_ARRAY) {
442 bool needs_pre_barrier = as_normal;
443
444 Register tmp3 = LP64_ONLY(r8) NOT_LP64(rsi);
445 Register rthread = LP64_ONLY(r15_thread) NOT_LP64(rcx);
446 // flatten object address if needed
447 // We do it regardless of precise because we need the registers
448 if (dst.index() == noreg && dst.disp() == 0) {
449 if (dst.base() != tmp1) {
450 __ movptr(tmp1, dst.base());
451 }
452 } else {
453 __ lea(tmp1, dst);
454 }
455
456 #ifndef _LP64
457 InterpreterMacroAssembler *imasm = static_cast<InterpreterMacroAssembler*>(masm);
458 #endif
459
460 NOT_LP64(__ get_thread(rcx));
461 NOT_LP64(imasm->save_bcp());
462
463 if (needs_pre_barrier) {
464 shenandoah_write_barrier_pre(masm /*masm*/,
465 tmp1 /* obj */,
466 tmp2 /* pre_val */,
467 rthread /* thread */,
468 tmp3 /* tmp */,
469 val != noreg /* tosca_live */,
470 false /* expand_call */);
471 }
472 if (val == noreg) {
473 BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
474 } else {
475 storeval_barrier(masm, val, tmp3);
476 BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
477 }
478 NOT_LP64(imasm->restore_bcp());
479 } else {
480 BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2);
481 }
482 }
483
484 #ifndef _LP64
485 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm,
486 Address obj1, jobject obj2) {
487 Unimplemented();
488 }
489
490 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm,
491 Register obj1, jobject obj2) {
492 Unimplemented();
493 }
494 #endif
495
496
497 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm, Register op1, Register op2) {
498 __ cmpptr(op1, op2);
499 if (ShenandoahAcmpBarrier) {
500 Label done;
501 __ jccb(Assembler::equal, done);
502 read_barrier(masm, op1);
503 read_barrier(masm, op2);
504 __ cmpptr(op1, op2);
505 __ bind(done);
506 }
507 }
508
509 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm, Register src1, Address src2) {
510 __ cmpptr(src1, src2);
511 if (ShenandoahAcmpBarrier) {
512 Label done;
513 __ jccb(Assembler::equal, done);
514 __ movptr(rscratch2, src2);
515 read_barrier(masm, src1);
516 read_barrier(masm, rscratch2);
517 __ cmpptr(src1, rscratch2);
518 __ bind(done);
519 }
520 }
521
522 void ShenandoahBarrierSetAssembler::tlab_allocate(MacroAssembler* masm,
523 Register thread, Register obj,
524 Register var_size_in_bytes,
525 int con_size_in_bytes,
526 Register t1, Register t2,
527 Label& slow_case) {
528 assert_different_registers(obj, t1, t2);
529 assert_different_registers(obj, var_size_in_bytes, t1);
530 Register end = t2;
531 if (!thread->is_valid()) {
532 #ifdef _LP64
533 thread = r15_thread;
534 #else
535 assert(t1->is_valid(), "need temp reg");
536 thread = t1;
537 __ get_thread(thread);
538 #endif
539 }
540
541 __ verify_tlab();
542
543 __ movptr(obj, Address(thread, JavaThread::tlab_top_offset()));
544 if (var_size_in_bytes == noreg) {
545 __ lea(end, Address(obj, con_size_in_bytes + ShenandoahBrooksPointer::byte_size()));
546 } else {
547 __ addptr(var_size_in_bytes, ShenandoahBrooksPointer::byte_size());
548 __ lea(end, Address(obj, var_size_in_bytes, Address::times_1));
549 }
550 __ cmpptr(end, Address(thread, JavaThread::tlab_end_offset()));
551 __ jcc(Assembler::above, slow_case);
552
553 // update the tlab top pointer
554 __ movptr(Address(thread, JavaThread::tlab_top_offset()), end);
555
556 // Initialize brooks pointer
557 #ifdef _LP64
558 __ incrementq(obj, ShenandoahBrooksPointer::byte_size());
559 #else
560 __ incrementl(obj, ShenandoahBrooksPointer::byte_size());
561 #endif
562 __ movptr(Address(obj, ShenandoahBrooksPointer::byte_offset()), obj);
563
564 // recover var_size_in_bytes if necessary
565 if (var_size_in_bytes == end) {
566 __ subptr(var_size_in_bytes, obj);
567 }
568 __ verify_tlab();
569 }
570
571 void ShenandoahBarrierSetAssembler::resolve(MacroAssembler* masm, DecoratorSet decorators, Register obj) {
572 bool oop_not_null = (decorators & IS_NOT_NULL) != 0;
573 bool is_write = (decorators & ACCESS_WRITE) != 0;
574 if (is_write) {
575 if (oop_not_null) {
576 write_barrier(masm, obj);
577 } else {
578 Label done;
579 __ testptr(obj, obj);
580 __ jcc(Assembler::zero, done);
581 write_barrier(masm, obj);
582 __ bind(done);
583 }
584 } else {
585 if (oop_not_null) {
586 read_barrier_not_null(masm, obj);
587 } else {
588 read_barrier(masm, obj);
589 }
590 }
591 }
592
593 // Special Shenandoah CAS implementation that handles false negatives
594 // due to concurrent evacuation.
595 #ifndef _LP64
596 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm,
597 Register res, Address addr, Register oldval, Register newval,
598 bool exchange, bool encode, Register tmp1, Register tmp2) {
599 // Shenandoah has no 32-bit version for this.
600 Unimplemented();
601 }
602 #else
603 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm,
604 Register res, Address addr, Register oldval, Register newval,
605 bool exchange, bool encode, Register tmp1, Register tmp2) {
606 if (!ShenandoahCASBarrier) {
607 #ifdef _LP64
608 if (UseCompressedOops) {
609 if (encode) {
610 __ encode_heap_oop(oldval);
611 __ mov(rscratch1, newval);
612 __ encode_heap_oop(rscratch1);
613 newval = rscratch1;
614 }
615 if (os::is_MP()) {
616 __ lock();
617 }
618 // oldval (rax) is implicitly used by this instruction
619 __ cmpxchgl(newval, addr);
620 } else
621 #endif
622 {
623 if (os::is_MP()) {
624 __ lock();
625 }
626 __ cmpxchgptr(newval, addr);
627 }
628
629 if (!exchange) {
630 assert(res != NULL, "need result register");
631 __ setb(Assembler::equal, res);
632 __ movzbl(res, res);
633 }
634 return;
635 }
636
637 assert(ShenandoahCASBarrier, "Should only be used when CAS barrier is enabled");
638 assert(oldval == rax, "must be in rax for implicit use in cmpxchg");
639
640 Label retry, done;
641
642 // Apply storeval barrier to newval.
643 if (encode) {
644 storeval_barrier(masm, newval, tmp1);
645 }
646
647 if (UseCompressedOops) {
648 if (encode) {
649 __ encode_heap_oop(oldval);
650 __ mov(rscratch1, newval);
651 __ encode_heap_oop(rscratch1);
652 newval = rscratch1;
653 }
654 }
655
656 // Remember oldval for retry logic below
657 if (UseCompressedOops) {
658 __ movl(tmp1, oldval);
659 } else {
660 __ movptr(tmp1, oldval);
661 }
662
663 // Step 1. Try to CAS with given arguments. If successful, then we are done,
664 // and can safely return.
665 if (os::is_MP()) __ lock();
666 if (UseCompressedOops) {
667 __ cmpxchgl(newval, addr);
668 } else {
669 __ cmpxchgptr(newval, addr);
670 }
671 __ jcc(Assembler::equal, done, true);
672
673 // Step 2. CAS had failed. This may be a false negative.
674 //
675 // The trouble comes when we compare the to-space pointer with the from-space
676 // pointer to the same object. To resolve this, it will suffice to read both
677 // oldval and the value from memory through the read barriers -- this will give
678 // both to-space pointers. If they mismatch, then it was a legitimate failure.
679 //
680 if (UseCompressedOops) {
681 __ decode_heap_oop(tmp1);
682 }
683 read_barrier_impl(masm, tmp1);
684
685 if (UseCompressedOops) {
686 __ movl(tmp2, oldval);
687 __ decode_heap_oop(tmp2);
688 } else {
689 __ movptr(tmp2, oldval);
690 }
691 read_barrier_impl(masm, tmp2);
692
693 __ cmpptr(tmp1, tmp2);
694 __ jcc(Assembler::notEqual, done, true);
695
696 // Step 3. Try to CAS again with resolved to-space pointers.
697 //
698 // Corner case: it may happen that somebody stored the from-space pointer
699 // to memory while we were preparing for retry. Therefore, we can fail again
700 // on retry, and so need to do this in loop, always re-reading the failure
701 // witness through the read barrier.
702 __ bind(retry);
703 if (os::is_MP()) __ lock();
704 if (UseCompressedOops) {
705 __ cmpxchgl(newval, addr);
706 } else {
707 __ cmpxchgptr(newval, addr);
708 }
709 __ jcc(Assembler::equal, done, true);
710
711 if (UseCompressedOops) {
712 __ movl(tmp2, oldval);
713 __ decode_heap_oop(tmp2);
714 } else {
715 __ movptr(tmp2, oldval);
716 }
717 read_barrier_impl(masm, tmp2);
718
719 __ cmpptr(tmp1, tmp2);
720 __ jcc(Assembler::equal, retry, true);
721
722 // Step 4. If we need a boolean result out of CAS, check the flag again,
723 // and promote the result. Note that we handle the flag from both the CAS
724 // itself and from the retry loop.
725 __ bind(done);
726 if (!exchange) {
727 assert(res != NULL, "need result register");
728 __ setb(Assembler::equal, res);
729 __ movzbl(res, res);
730 }
731 }
732 #endif // LP64
733
734 void ShenandoahBarrierSetAssembler::save_vector_registers(MacroAssembler* masm) {
735 int num_xmm_regs = LP64_ONLY(16) NOT_LP64(8);
736 if (UseAVX > 2) {
737 num_xmm_regs = LP64_ONLY(32) NOT_LP64(8);
738 }
739
740 if (UseSSE == 1) {
741 __ subptr(rsp, sizeof(jdouble)*8);
742 for (int n = 0; n < 8; n++) {
743 __ movflt(Address(rsp, n*sizeof(jdouble)), as_XMMRegister(n));
744 }
745 } else if (UseSSE >= 2) {
746 if (UseAVX > 2) {
747 __ push(rbx);
748 __ movl(rbx, 0xffff);
749 __ kmovwl(k1, rbx);
750 __ pop(rbx);
751 }
752 #ifdef COMPILER2
753 if (MaxVectorSize > 16) {
754 if(UseAVX > 2) {
755 // Save upper half of ZMM registers
756 __ subptr(rsp, 32*num_xmm_regs);
757 for (int n = 0; n < num_xmm_regs; n++) {
758 __ vextractf64x4_high(Address(rsp, n*32), as_XMMRegister(n));
759 }
760 }
761 assert(UseAVX > 0, "256 bit vectors are supported only with AVX");
762 // Save upper half of YMM registers
763 __ subptr(rsp, 16*num_xmm_regs);
764 for (int n = 0; n < num_xmm_regs; n++) {
765 __ vextractf128_high(Address(rsp, n*16), as_XMMRegister(n));
766 }
767 }
768 #endif
769 // Save whole 128bit (16 bytes) XMM registers
770 __ subptr(rsp, 16*num_xmm_regs);
771 #ifdef _LP64
772 if (VM_Version::supports_evex()) {
773 for (int n = 0; n < num_xmm_regs; n++) {
774 __ vextractf32x4(Address(rsp, n*16), as_XMMRegister(n), 0);
775 }
776 } else {
777 for (int n = 0; n < num_xmm_regs; n++) {
778 __ movdqu(Address(rsp, n*16), as_XMMRegister(n));
779 }
780 }
781 #else
782 for (int n = 0; n < num_xmm_regs; n++) {
783 __ movdqu(Address(rsp, n*16), as_XMMRegister(n));
784 }
785 #endif
786 }
787 }
788
789 void ShenandoahBarrierSetAssembler::restore_vector_registers(MacroAssembler* masm) {
790 int num_xmm_regs = LP64_ONLY(16) NOT_LP64(8);
791 if (UseAVX > 2) {
792 num_xmm_regs = LP64_ONLY(32) NOT_LP64(8);
793 }
794 if (UseSSE == 1) {
795 for (int n = 0; n < 8; n++) {
796 __ movflt(as_XMMRegister(n), Address(rsp, n*sizeof(jdouble)));
797 }
798 __ addptr(rsp, sizeof(jdouble)*8);
799 } else if (UseSSE >= 2) {
800 // Restore whole 128bit (16 bytes) XMM registers
801 #ifdef _LP64
802 if (VM_Version::supports_evex()) {
803 for (int n = 0; n < num_xmm_regs; n++) {
804 __ vinsertf32x4(as_XMMRegister(n), as_XMMRegister(n), Address(rsp, n*16), 0);
805 }
806 } else {
807 for (int n = 0; n < num_xmm_regs; n++) {
808 __ movdqu(as_XMMRegister(n), Address(rsp, n*16));
809 }
810 }
811 #else
812 for (int n = 0; n < num_xmm_regs; n++) {
813 __ movdqu(as_XMMRegister(n), Address(rsp, n*16));
814 }
815 #endif
816 __ addptr(rsp, 16*num_xmm_regs);
817
818 #ifdef COMPILER2
819 if (MaxVectorSize > 16) {
820 // Restore upper half of YMM registers.
821 for (int n = 0; n < num_xmm_regs; n++) {
822 __ vinsertf128_high(as_XMMRegister(n), Address(rsp, n*16));
823 }
824 __ addptr(rsp, 16*num_xmm_regs);
825 if (UseAVX > 2) {
826 for (int n = 0; n < num_xmm_regs; n++) {
827 __ vinsertf64x4_high(as_XMMRegister(n), Address(rsp, n*32));
828 }
829 __ addptr(rsp, 32*num_xmm_regs);
830 }
831 }
832 #endif
833 }
834 }
835
836 #ifdef COMPILER1
837
838 #undef __
839 #define __ ce->masm()->
840
841 void ShenandoahBarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, ShenandoahPreBarrierStub* stub) {
842 ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
843 // At this point we know that marking is in progress.
844 // If do_load() is true then we have to emit the
845 // load of the previous value; otherwise it has already
846 // been loaded into _pre_val.
847
848 __ bind(*stub->entry());
849 assert(stub->pre_val()->is_register(), "Precondition.");
850
851 Register pre_val_reg = stub->pre_val()->as_register();
852
853 if (stub->do_load()) {
854 ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/, false /*unaligned*/);
855 }
856
857 __ cmpptr(pre_val_reg, (int32_t)NULL_WORD);
858 __ jcc(Assembler::equal, *stub->continuation());
859 ce->store_parameter(stub->pre_val()->as_register(), 0);
860 __ call(RuntimeAddress(bs->pre_barrier_c1_runtime_code_blob()->code_begin()));
861 __ jmp(*stub->continuation());
862
863 }
864
865 void ShenandoahBarrierSetAssembler::gen_write_barrier_stub(LIR_Assembler* ce, ShenandoahWriteBarrierStub* stub) {
866 __ bind(*stub->entry());
867
868 Label done;
869 Register obj = stub->obj()->as_register();
870 Register res = stub->result()->as_register();
871
872 if (res != obj) {
873 __ mov(res, obj);
874 }
875
876 // Check for null.
877 if (stub->needs_null_check()) {
878 __ testptr(res, res);
879 __ jcc(Assembler::zero, done);
880 }
881
882 write_barrier(ce->masm(), res);
883
884 __ bind(done);
885 __ jmp(*stub->continuation());
886 }
887
888 #undef __
889
890 #define __ sasm->
891
892 void ShenandoahBarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) {
893 __ prologue("shenandoah_pre_barrier", false);
894 // arg0 : previous value of memory
895
896 __ push(rax);
897 __ push(rdx);
898
899 const Register pre_val = rax;
900 const Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
901 const Register tmp = rdx;
902
903 NOT_LP64(__ get_thread(thread);)
904
905 Address queue_index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
906 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
907
908 Label done;
909 Label runtime;
910
911 // Is SATB still active?
912 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
913 __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL);
914 __ jcc(Assembler::zero, done);
915
916 // Can we store original value in the thread's buffer?
917
918 __ movptr(tmp, queue_index);
919 __ testptr(tmp, tmp);
920 __ jcc(Assembler::zero, runtime);
921 __ subptr(tmp, wordSize);
922 __ movptr(queue_index, tmp);
923 __ addptr(tmp, buffer);
924
925 // prev_val (rax)
926 __ load_parameter(0, pre_val);
927 __ movptr(Address(tmp, 0), pre_val);
928 __ jmp(done);
929
930 __ bind(runtime);
931
932 __ save_live_registers_no_oop_map(true);
933
934 // load the pre-value
935 __ load_parameter(0, rcx);
936 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), rcx, thread);
937
938 __ restore_live_registers(true);
939
940 __ bind(done);
941
942 __ pop(rdx);
943 __ pop(rax);
944
945 __ epilogue();
946 }
947
948 #undef __
949
950 #endif // COMPILER1
951
952 address ShenandoahBarrierSetAssembler::shenandoah_wb() {
953 assert(_shenandoah_wb != NULL, "need write barrier stub");
954 return _shenandoah_wb;
955 }
956
957 address ShenandoahBarrierSetAssembler::shenandoah_wb_C() {
958 assert(_shenandoah_wb_C != NULL, "need write barrier stub");
959 return _shenandoah_wb_C;
960 }
961
962 #define __ cgen->assembler()->
963
964 address ShenandoahBarrierSetAssembler::generate_shenandoah_wb(StubCodeGenerator* cgen, bool c_abi, bool do_cset_test) {
965 __ align(CodeEntryAlignment);
966 StubCodeMark mark(cgen, "StubRoutines", "shenandoah_wb");
967 address start = __ pc();
968
969 #ifdef _LP64
970 Label not_done;
971
972 // We use RDI, which also serves as argument register for slow call.
973 // RAX always holds the src object ptr, except after the slow call and
974 // the cmpxchg, then it holds the result.
975 // R8 and RCX are used as temporary registers.
976 if (!c_abi) {
977 __ push(rdi);
978 __ push(r8);
979 }
980
981 // Check for object beeing in the collection set.
982 // TODO: Can we use only 1 register here?
983 // The source object arrives here in rax.
984 // live: rax
985 // live: rdi
986 if (!c_abi) {
987 __ mov(rdi, rax);
988 } else {
989 if (rax != c_rarg0) {
990 __ mov(rax, c_rarg0);
991 }
992 }
993 if (do_cset_test) {
994 __ shrptr(rdi, ShenandoahHeapRegion::region_size_bytes_shift_jint());
995 // live: r8
996 __ movptr(r8, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr());
997 __ movbool(r8, Address(r8, rdi, Address::times_1));
998 // unlive: rdi
999 __ testbool(r8);
1000 // unlive: r8
1001 __ jccb(Assembler::notZero, not_done);
1002
1003 if (!c_abi) {
1004 __ pop(r8);
1005 __ pop(rdi);
1006 }
1007 __ ret(0);
1008
1009 __ bind(not_done);
1010 }
1011
1012 if (!c_abi) {
1013 __ push(rcx);
1014 }
1015
1016 if (!c_abi) {
1017 __ push(rdx);
1018 __ push(rdi);
1019 __ push(rsi);
1020 __ push(r8);
1021 __ push(r9);
1022 __ push(r10);
1023 __ push(r11);
1024 __ push(r12);
1025 __ push(r13);
1026 __ push(r14);
1027 __ push(r15);
1028 }
1029 save_vector_registers(cgen->assembler());
1030 __ movptr(rdi, rax);
1031 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_JRT), rdi);
1032 restore_vector_registers(cgen->assembler());
1033 if (!c_abi) {
1034 __ pop(r15);
1035 __ pop(r14);
1036 __ pop(r13);
1037 __ pop(r12);
1038 __ pop(r11);
1039 __ pop(r10);
1040 __ pop(r9);
1041 __ pop(r8);
1042 __ pop(rsi);
1043 __ pop(rdi);
1044 __ pop(rdx);
1045
1046 __ pop(rcx);
1047 __ pop(r8);
1048 __ pop(rdi);
1049 }
1050 __ ret(0);
1051 #else
1052 ShouldNotReachHere();
1053 #endif
1054 return start;
1055 }
1056
1057 #undef __
1058
1059 void ShenandoahBarrierSetAssembler::barrier_stubs_init() {
1060 if (ShenandoahWriteBarrier || ShenandoahStoreValEnqueueBarrier) {
1061 int stub_code_size = 4096;
1062 ResourceMark rm;
1063 BufferBlob* bb = BufferBlob::create("shenandoah_barrier_stubs", stub_code_size);
1064 CodeBuffer buf(bb);
1065 StubCodeGenerator cgen(&buf);
1066 _shenandoah_wb = generate_shenandoah_wb(&cgen, false, true);
1067 _shenandoah_wb_C = generate_shenandoah_wb(&cgen, true, false);
1068 }
1069 }