1 /*
2 * Copyright (c) 2018, Red Hat, Inc. All rights reserved.
3 *
4 * This code is free software; you can redistribute it and/or modify it
5 * under the terms of the GNU General Public License version 2 only, as
6 * published by the Free Software Foundation.
7 *
8 * This code is distributed in the hope that it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
11 * version 2 for more details (a copy is included in the LICENSE file that
12 * accompanied this code).
13 *
14 * You should have received a copy of the GNU General Public License version
15 * 2 along with this work; if not, write to the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
17 *
18 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
19 * or visit www.oracle.com if you need additional information or have any
20 * questions.
21 *
22 */
23
24 #include "precompiled.hpp"
25 #include "gc/shenandoah/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
46 void ShenandoahBarrierSetAssembler::arraycopy_prologue(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
47 Register src, Register dst, Register count) {
48
49 bool checkcast = (decorators & ARRAYCOPY_CHECKCAST) != 0;
50 bool disjoint = (decorators & ARRAYCOPY_DISJOINT) != 0;
51 bool obj_int = type == T_OBJECT LP64_ONLY(&& UseCompressedOops);
52 bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0;
53
54 if (type == T_OBJECT || type == T_ARRAY) {
55 #ifdef _LP64
56 if (!checkcast && !obj_int) {
57 // Save count for barrier
58 __ movptr(r11, count);
59 } else if (disjoint && obj_int) {
60 // Save dst in r11 in the disjoint case
61 __ movq(r11, dst);
62 }
63 #else
64 if (disjoint) {
65 __ mov(rdx, dst); // save 'to'
66 }
67 #endif
68
69 if (ShenandoahSATBBarrier && !dest_uninitialized && !ShenandoahHeap::heap()->heuristics()->can_do_traversal_gc()) {
70 Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
71 #ifndef _LP64
72 __ push(thread);
73 __ get_thread(thread);
74 #endif
75
76 Label done;
77 // Short-circuit if count == 0.
78 __ testptr(count, count);
79 __ jcc(Assembler::zero, done);
80
81 // Avoid runtime call when not marking.
82 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
83 __ testb(gc_state, ShenandoahHeap::MARKING);
84 __ jcc(Assembler::zero, done);
85
86 __ pusha(); // push registers
87 #ifdef _LP64
88 if (count == c_rarg0) {
89 if (dst == c_rarg1) {
90 // exactly backwards!!
91 __ xchgptr(c_rarg1, c_rarg0);
92 } else {
93 __ movptr(c_rarg1, count);
94 __ movptr(c_rarg0, dst);
95 }
96 } else {
97 __ movptr(c_rarg0, dst);
98 __ movptr(c_rarg1, count);
99 }
100 if (UseCompressedOops) {
101 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_narrow_oop_entry), 2);
102 } else {
103 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry), 2);
104 }
105 #else
106 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry),
107 dst, count);
108 #endif
109 __ popa();
110 __ bind(done);
111 NOT_LP64(__ pop(thread);)
112 }
113 }
114
115 }
116
117 void ShenandoahBarrierSetAssembler::arraycopy_epilogue(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
118 Register src, Register dst, Register count) {
119 bool checkcast = (decorators & ARRAYCOPY_CHECKCAST) != 0;
120 bool disjoint = (decorators & ARRAYCOPY_DISJOINT) != 0;
121 bool obj_int = type == T_OBJECT LP64_ONLY(&& UseCompressedOops);
122 Register tmp = rax;
123
124 if (type == T_OBJECT || type == T_ARRAY) {
125 #ifdef _LP64
126 if (!checkcast && !obj_int) {
127 // Save count for barrier
128 count = r11;
129 } else if (disjoint && obj_int) {
130 // Use the saved dst in the disjoint case
131 dst = r11;
132 } else if (checkcast) {
133 tmp = rscratch1;
134 }
135 #else
136 if (disjoint) {
137 __ mov(dst, rdx); // restore 'to'
138 }
139 #endif
140
141 __ pusha(); // push registers (overkill)
142 #ifdef _LP64
143 if (c_rarg0 == count) { // On win64 c_rarg0 == rcx
144 assert_different_registers(c_rarg1, dst);
145 __ mov(c_rarg1, count);
146 __ mov(c_rarg0, dst);
147 } else {
148 assert_different_registers(c_rarg0, count);
149 __ mov(c_rarg0, dst);
150 __ mov(c_rarg1, count);
151 }
152 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry), 2);
153 #else
154 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry),
155 dst, count);
156 #endif
157 __ popa();
158 }
159 }
160
161 void ShenandoahBarrierSetAssembler::shenandoah_write_barrier_pre(MacroAssembler* masm,
162 Register obj,
163 Register pre_val,
164 Register thread,
165 Register tmp,
166 bool tosca_live,
167 bool expand_call) {
168
169 if (ShenandoahSATBBarrier) {
170 satb_write_barrier_pre(masm, obj, pre_val, thread, tmp, tosca_live, expand_call);
171 }
172 }
173
174 void ShenandoahBarrierSetAssembler::satb_write_barrier_pre(MacroAssembler* masm,
175 Register obj,
176 Register pre_val,
177 Register thread,
178 Register tmp,
179 bool tosca_live,
180 bool expand_call) {
181 // If expand_call is true then we expand the call_VM_leaf macro
182 // directly to skip generating the check by
183 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
184
185 #ifdef _LP64
186 assert(thread == r15_thread, "must be");
187 #endif // _LP64
188
189 Label done;
190 Label runtime;
191
192 assert(pre_val != noreg, "check this code");
193
194 if (obj != noreg) {
195 assert_different_registers(obj, pre_val, tmp);
196 assert(pre_val != rax, "check this code");
197 }
198
199 Address in_progress(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()));
200 Address index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
201 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
202
203 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
204 __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL);
205 __ jcc(Assembler::zero, done);
206
207 // Do we need to load the previous value?
208 if (obj != noreg) {
209 __ load_heap_oop(pre_val, Address(obj, 0), noreg, noreg, AS_RAW);
210 }
211
212 // Is the previous value null?
213 __ cmpptr(pre_val, (int32_t) NULL_WORD);
214 __ jcc(Assembler::equal, done);
215
216 // Can we store original value in the thread's buffer?
217 // Is index == 0?
218 // (The index field is typed as size_t.)
219
220 __ movptr(tmp, index); // tmp := *index_adr
221 __ cmpptr(tmp, 0); // tmp == 0?
222 __ jcc(Assembler::equal, runtime); // If yes, goto runtime
223
224 __ subptr(tmp, wordSize); // tmp := tmp - wordSize
225 __ movptr(index, tmp); // *index_adr := tmp
226 __ addptr(tmp, buffer); // tmp := tmp + *buffer_adr
227
228 // Record the previous value
229 __ movptr(Address(tmp, 0), pre_val);
230 __ jmp(done);
231
232 __ bind(runtime);
233 // save the live input values
234 if(tosca_live) __ push(rax);
235
236 if (obj != noreg && obj != rax)
237 __ push(obj);
238
239 if (pre_val != rax)
240 __ push(pre_val);
241
242 // Calling the runtime using the regular call_VM_leaf mechanism generates
243 // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
244 // that checks that the *(ebp+frame::interpreter_frame_last_sp) == NULL.
245 //
246 // If we care generating the pre-barrier without a frame (e.g. in the
247 // intrinsified Reference.get() routine) then ebp might be pointing to
248 // the caller frame and so this check will most likely fail at runtime.
249 //
250 // Expanding the call directly bypasses the generation of the check.
251 // So when we do not have have a full interpreter frame on the stack
252 // expand_call should be passed true.
253
254 NOT_LP64( __ push(thread); )
255
256 #ifdef _LP64
257 // We move pre_val into c_rarg0 early, in order to avoid smashing it, should
258 // pre_val be c_rarg1 (where the call prologue would copy thread argument).
259 // Note: this should not accidentally smash thread, because thread is always r15.
260 assert(thread != c_rarg0, "smashed arg");
261 if (c_rarg0 != pre_val) {
262 __ mov(c_rarg0, pre_val);
263 }
264 #endif
265
266 if (expand_call) {
267 LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); )
268 #ifdef _LP64
269 if (c_rarg1 != thread) {
270 __ mov(c_rarg1, thread);
271 }
272 // Already moved pre_val into c_rarg0 above
273 #else
274 __ push(thread);
275 __ push(pre_val);
276 #endif
277 __ MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), 2);
278 } else {
279 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), LP64_ONLY(c_rarg0) NOT_LP64(pre_val), thread);
280 }
281
282 NOT_LP64( __ pop(thread); )
283
284 // save the live input values
285 if (pre_val != rax)
286 __ pop(pre_val);
287
288 if (obj != noreg && obj != rax)
289 __ pop(obj);
290
291 if(tosca_live) __ pop(rax);
292
293 __ bind(done);
294 }
295
296 void ShenandoahBarrierSetAssembler::read_barrier(MacroAssembler* masm, Register dst) {
297 if (ShenandoahReadBarrier) {
298 read_barrier_impl(masm, dst);
299 }
300 }
301
302 void ShenandoahBarrierSetAssembler::read_barrier_impl(MacroAssembler* masm, Register dst) {
303 assert(UseShenandoahGC && (ShenandoahReadBarrier || ShenandoahStoreValReadBarrier || ShenandoahCASBarrier), "should be enabled");
304 Label is_null;
305 __ testptr(dst, dst);
306 __ jcc(Assembler::zero, is_null);
307 read_barrier_not_null_impl(masm, dst);
308 __ bind(is_null);
309 }
310
311 void ShenandoahBarrierSetAssembler::read_barrier_not_null(MacroAssembler* masm, Register dst) {
312 if (ShenandoahReadBarrier) {
313 read_barrier_not_null_impl(masm, dst);
314 }
315 }
316
317 void ShenandoahBarrierSetAssembler::read_barrier_not_null_impl(MacroAssembler* masm, Register dst) {
318 assert(UseShenandoahGC && (ShenandoahReadBarrier || ShenandoahStoreValReadBarrier || ShenandoahCASBarrier), "should be enabled");
319 __ movptr(dst, Address(dst, ShenandoahBrooksPointer::byte_offset()));
320 }
321
322
323 void ShenandoahBarrierSetAssembler::write_barrier(MacroAssembler* masm, Register dst) {
324 if (ShenandoahWriteBarrier) {
325 write_barrier_impl(masm, dst);
326 }
327 }
328
329 void ShenandoahBarrierSetAssembler::write_barrier_impl(MacroAssembler* masm, Register dst) {
330 assert(UseShenandoahGC && (ShenandoahWriteBarrier || ShenandoahStoreValEnqueueBarrier), "Should be enabled");
331 #ifdef _LP64
332 Label done;
333
334 Address gc_state(r15_thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
335 __ testb(gc_state, ShenandoahHeap::HAS_FORWARDED | ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL);
336 __ jccb(Assembler::zero, done);
337
338 // Heap is unstable, need to perform the read-barrier even if WB is inactive
339 read_barrier_not_null(masm, dst);
340
341 __ testb(gc_state, ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL);
342 __ jccb(Assembler::zero, done);
343
344 if (dst != rax) {
345 __ xchgptr(dst, rax); // Move obj into rax and save rax into obj.
346 }
347
348 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, ShenandoahBarrierSetAssembler::shenandoah_wb())));
349
350 if (dst != rax) {
351 __ xchgptr(rax, dst); // Swap back obj with rax.
352 }
353
354 __ bind(done);
355 #else
356 Unimplemented();
357 #endif
358 }
359
360 void ShenandoahBarrierSetAssembler::storeval_barrier(MacroAssembler* masm, Register dst, Register tmp) {
361 if (ShenandoahStoreValReadBarrier || ShenandoahStoreValEnqueueBarrier) {
362 storeval_barrier_impl(masm, dst, tmp);
363 }
364 }
365
366 void ShenandoahBarrierSetAssembler::storeval_barrier_impl(MacroAssembler* masm, Register dst, Register tmp) {
367 assert(UseShenandoahGC && (ShenandoahStoreValReadBarrier || ShenandoahStoreValEnqueueBarrier), "should be enabled");
368
369 if (dst == noreg) return;
370
371 #ifdef _LP64
372 if (ShenandoahStoreValEnqueueBarrier) {
373 Label is_null;
374 __ testptr(dst, dst);
375 __ jcc(Assembler::zero, is_null);
376 write_barrier_impl(masm, dst);
377 __ bind(is_null);
378
379 // The set of registers to be saved+restored is the same as in the write-barrier above.
380 // Those are the commonly used registers in the interpreter.
381 __ pusha();
382 // __ push_callee_saved_registers();
383 __ subptr(rsp, 2 * Interpreter::stackElementSize);
384 __ movdbl(Address(rsp, 0), xmm0);
385
386 satb_write_barrier_pre(masm, noreg, dst, r15_thread, tmp, true, false);
387 __ movdbl(xmm0, Address(rsp, 0));
388 __ addptr(rsp, 2 * Interpreter::stackElementSize);
389 //__ pop_callee_saved_registers();
390 __ popa();
391 }
392 if (ShenandoahStoreValReadBarrier) {
393 read_barrier_impl(masm, dst);
394 }
395 #else
396 Unimplemented();
397 #endif
398 }
399
400 void ShenandoahBarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
401 Register dst, Address src, Register tmp1, Register tmp_thread) {
402 bool on_oop = type == T_OBJECT || type == T_ARRAY;
403 bool in_heap = (decorators & IN_HEAP) != 0;
404 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
405 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
406 bool on_reference = on_weak || on_phantom;
407 if (in_heap) {
408 read_barrier_not_null(masm, src.base());
409 }
410 BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp_thread);
411 if (ShenandoahKeepAliveBarrier && on_oop && on_reference) {
412 const Register thread = NOT_LP64(tmp_thread) LP64_ONLY(r15_thread);
413 NOT_LP64(__ get_thread(thread));
414
415 // Generate the SATB pre-barrier code to log the value of
416 // the referent field in an SATB buffer.
417 shenandoah_write_barrier_pre(masm /* masm */,
418 noreg /* obj */,
419 dst /* pre_val */,
420 thread /* thread */,
421 tmp1 /* tmp */,
422 true /* tosca_live */,
423 true /* expand_call */);
424 }
425 }
426
427 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
428 Address dst, Register val, Register tmp1, Register tmp2) {
429
430 bool in_heap = (decorators & IN_HEAP) != 0;
431 bool as_normal = (decorators & AS_NORMAL) != 0;
432 if (in_heap) {
433 write_barrier(masm, dst.base());
434 }
435 if (type == T_OBJECT || type == T_ARRAY) {
436 bool needs_pre_barrier = as_normal;
437
438 Register tmp3 = LP64_ONLY(r8) NOT_LP64(rsi);
439 Register rthread = LP64_ONLY(r15_thread) NOT_LP64(rcx);
440 // flatten object address if needed
441 // We do it regardless of precise because we need the registers
442 if (dst.index() == noreg && dst.disp() == 0) {
443 if (dst.base() != tmp1) {
444 __ movptr(tmp1, dst.base());
445 }
446 } else {
447 __ lea(tmp1, dst);
448 }
449
450 #ifndef _LP64
451 InterpreterMacroAssembler *imasm = static_cast<InterpreterMacroAssembler*>(masm);
452 #endif
453
454 NOT_LP64(__ get_thread(rcx));
455 NOT_LP64(imasm->save_bcp());
456
457 if (needs_pre_barrier) {
458 shenandoah_write_barrier_pre(masm /*masm*/,
459 tmp1 /* obj */,
460 tmp2 /* pre_val */,
461 rthread /* thread */,
462 tmp3 /* tmp */,
463 val != noreg /* tosca_live */,
464 false /* expand_call */);
465 }
466 if (val == noreg) {
467 BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
468 } else {
469 storeval_barrier(masm, val, tmp3);
470 BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
471 }
472 NOT_LP64(imasm->restore_bcp());
473 } else {
474 BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2);
475 }
476 }
477
478 #ifndef _LP64
479 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm,
480 Address obj1, jobject obj2) {
481 Unimplemented();
482 }
483
484 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm,
485 Register obj1, jobject obj2) {
486 Unimplemented();
487 }
488 #endif
489
490
491 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm, Register op1, Register op2) {
492 __ cmpptr(op1, op2);
493 if (ShenandoahAcmpBarrier) {
494 Label done;
495 __ jccb(Assembler::equal, done);
496 read_barrier(masm, op1);
497 read_barrier(masm, op2);
498 __ cmpptr(op1, op2);
499 __ bind(done);
500 }
501 }
502
503 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm, Register src1, Address src2) {
504 __ cmpptr(src1, src2);
505 if (ShenandoahAcmpBarrier) {
506 Label done;
507 __ jccb(Assembler::equal, done);
508 __ movptr(rscratch2, src2);
509 read_barrier(masm, src1);
510 read_barrier(masm, rscratch2);
511 __ cmpptr(src1, rscratch2);
512 __ bind(done);
513 }
514 }
515
516 void ShenandoahBarrierSetAssembler::tlab_allocate(MacroAssembler* masm,
517 Register thread, Register obj,
518 Register var_size_in_bytes,
519 int con_size_in_bytes,
520 Register t1, Register t2,
521 Label& slow_case) {
522 assert_different_registers(obj, t1, t2);
523 assert_different_registers(obj, var_size_in_bytes, t1);
524 Register end = t2;
525 if (!thread->is_valid()) {
526 #ifdef _LP64
527 thread = r15_thread;
528 #else
529 assert(t1->is_valid(), "need temp reg");
530 thread = t1;
531 __ get_thread(thread);
532 #endif
533 }
534
535 __ verify_tlab();
536
537 __ movptr(obj, Address(thread, JavaThread::tlab_top_offset()));
538 if (var_size_in_bytes == noreg) {
539 __ lea(end, Address(obj, con_size_in_bytes + ShenandoahBrooksPointer::byte_size()));
540 } else {
541 __ addptr(var_size_in_bytes, ShenandoahBrooksPointer::byte_size());
542 __ lea(end, Address(obj, var_size_in_bytes, Address::times_1));
543 }
544 __ cmpptr(end, Address(thread, JavaThread::tlab_end_offset()));
545 __ jcc(Assembler::above, slow_case);
546
547 // update the tlab top pointer
548 __ movptr(Address(thread, JavaThread::tlab_top_offset()), end);
549
550 // Initialize brooks pointer
551 #ifdef _LP64
552 __ incrementq(obj, ShenandoahBrooksPointer::byte_size());
553 #else
554 __ incrementl(obj, ShenandoahBrooksPointer::byte_size());
555 #endif
556 __ movptr(Address(obj, ShenandoahBrooksPointer::byte_offset()), obj);
557
558 // recover var_size_in_bytes if necessary
559 if (var_size_in_bytes == end) {
560 __ subptr(var_size_in_bytes, obj);
561 }
562 __ verify_tlab();
563 }
564
565 void ShenandoahBarrierSetAssembler::resolve(MacroAssembler* masm, DecoratorSet decorators, Register obj) {
566 bool oop_not_null = (decorators & IS_NOT_NULL) != 0;
567 bool is_write = (decorators & ACCESS_WRITE) != 0;
568 if (is_write) {
569 if (oop_not_null) {
570 write_barrier(masm, obj);
571 } else {
572 Label done;
573 __ testptr(obj, obj);
574 __ jcc(Assembler::zero, done);
575 write_barrier(masm, obj);
576 __ bind(done);
577 }
578 } else {
579 if (oop_not_null) {
580 read_barrier_not_null(masm, obj);
581 } else {
582 read_barrier(masm, obj);
583 }
584 }
585 }
586
587 // Special Shenandoah CAS implementation that handles false negatives
588 // due to concurrent evacuation.
589 #ifndef _LP64
590 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm,
591 Register res, Address addr, Register oldval, Register newval,
592 bool exchange, Register tmp1, Register tmp2) {
593 // Shenandoah has no 32-bit version for this.
594 Unimplemented();
595 }
596 #else
597 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm,
598 Register res, Address addr, Register oldval, Register newval,
599 bool exchange, Register tmp1, Register tmp2) {
600 assert(ShenandoahCASBarrier, "Should only be used when CAS barrier is enabled");
601 assert(oldval == rax, "must be in rax for implicit use in cmpxchg");
602
603 Label retry, done;
604
605 // Remember oldval for retry logic below
606 if (UseCompressedOops) {
607 __ movl(tmp1, oldval);
608 } else {
609 __ movptr(tmp1, oldval);
610 }
611
612 // Step 1. Try to CAS with given arguments. If successful, then we are done,
613 // and can safely return.
614 if (os::is_MP()) __ lock();
615 if (UseCompressedOops) {
616 __ cmpxchgl(newval, addr);
617 } else {
618 __ cmpxchgptr(newval, addr);
619 }
620 __ jcc(Assembler::equal, done, true);
621
622 // Step 2. CAS had failed. This may be a false negative.
623 //
624 // The trouble comes when we compare the to-space pointer with the from-space
625 // pointer to the same object. To resolve this, it will suffice to read both
626 // oldval and the value from memory through the read barriers -- this will give
627 // both to-space pointers. If they mismatch, then it was a legitimate failure.
628 //
629 if (UseCompressedOops) {
630 __ decode_heap_oop(tmp1);
631 }
632 read_barrier_impl(masm, tmp1);
633
634 if (UseCompressedOops) {
635 __ movl(tmp2, oldval);
636 __ decode_heap_oop(tmp2);
637 } else {
638 __ movptr(tmp2, oldval);
639 }
640 read_barrier_impl(masm, tmp2);
641
642 __ cmpptr(tmp1, tmp2);
643 __ jcc(Assembler::notEqual, done, true);
644
645 // Step 3. Try to CAS again with resolved to-space pointers.
646 //
647 // Corner case: it may happen that somebody stored the from-space pointer
648 // to memory while we were preparing for retry. Therefore, we can fail again
649 // on retry, and so need to do this in loop, always re-reading the failure
650 // witness through the read barrier.
651 __ bind(retry);
652 if (os::is_MP()) __ lock();
653 if (UseCompressedOops) {
654 __ cmpxchgl(newval, addr);
655 } else {
656 __ cmpxchgptr(newval, addr);
657 }
658 __ jcc(Assembler::equal, done, true);
659
660 if (UseCompressedOops) {
661 __ movl(tmp2, oldval);
662 __ decode_heap_oop(tmp2);
663 } else {
664 __ movptr(tmp2, oldval);
665 }
666 read_barrier_impl(masm, tmp2);
667
668 __ cmpptr(tmp1, tmp2);
669 __ jcc(Assembler::equal, retry, true);
670
671 // Step 4. If we need a boolean result out of CAS, check the flag again,
672 // and promote the result. Note that we handle the flag from both the CAS
673 // itself and from the retry loop.
674 __ bind(done);
675 if (!exchange) {
676 assert(res != NULL, "need result register");
677 __ setb(Assembler::equal, res);
678 __ movzbl(res, res);
679 }
680 }
681 #endif // LP64
682
683 void ShenandoahBarrierSetAssembler::save_vector_registers(MacroAssembler* masm) {
684 int num_xmm_regs = LP64_ONLY(16) NOT_LP64(8);
685 if (UseAVX > 2) {
686 num_xmm_regs = LP64_ONLY(32) NOT_LP64(8);
687 }
688
689 if (UseSSE == 1) {
690 __ subptr(rsp, sizeof(jdouble)*8);
691 for (int n = 0; n < 8; n++) {
692 __ movflt(Address(rsp, n*sizeof(jdouble)), as_XMMRegister(n));
693 }
694 } else if (UseSSE >= 2) {
695 if (UseAVX > 2) {
696 __ push(rbx);
697 __ movl(rbx, 0xffff);
698 __ kmovwl(k1, rbx);
699 __ pop(rbx);
700 }
701 #ifdef COMPILER2
702 if (MaxVectorSize > 16) {
703 if(UseAVX > 2) {
704 // Save upper half of ZMM registers
705 __ subptr(rsp, 32*num_xmm_regs);
706 for (int n = 0; n < num_xmm_regs; n++) {
707 __ vextractf64x4_high(Address(rsp, n*32), as_XMMRegister(n));
708 }
709 }
710 assert(UseAVX > 0, "256 bit vectors are supported only with AVX");
711 // Save upper half of YMM registers
712 __ subptr(rsp, 16*num_xmm_regs);
713 for (int n = 0; n < num_xmm_regs; n++) {
714 __ vextractf128_high(Address(rsp, n*16), as_XMMRegister(n));
715 }
716 }
717 #endif
718 // Save whole 128bit (16 bytes) XMM registers
719 __ subptr(rsp, 16*num_xmm_regs);
720 #ifdef _LP64
721 if (VM_Version::supports_evex()) {
722 for (int n = 0; n < num_xmm_regs; n++) {
723 __ vextractf32x4(Address(rsp, n*16), as_XMMRegister(n), 0);
724 }
725 } else {
726 for (int n = 0; n < num_xmm_regs; n++) {
727 __ movdqu(Address(rsp, n*16), as_XMMRegister(n));
728 }
729 }
730 #else
731 for (int n = 0; n < num_xmm_regs; n++) {
732 __ movdqu(Address(rsp, n*16), as_XMMRegister(n));
733 }
734 #endif
735 }
736 }
737
738 void ShenandoahBarrierSetAssembler::restore_vector_registers(MacroAssembler* masm) {
739 int num_xmm_regs = LP64_ONLY(16) NOT_LP64(8);
740 if (UseAVX > 2) {
741 num_xmm_regs = LP64_ONLY(32) NOT_LP64(8);
742 }
743 if (UseSSE == 1) {
744 for (int n = 0; n < 8; n++) {
745 __ movflt(as_XMMRegister(n), Address(rsp, n*sizeof(jdouble)));
746 }
747 __ addptr(rsp, sizeof(jdouble)*8);
748 } else if (UseSSE >= 2) {
749 // Restore whole 128bit (16 bytes) XMM registers
750 #ifdef _LP64
751 if (VM_Version::supports_evex()) {
752 for (int n = 0; n < num_xmm_regs; n++) {
753 __ vinsertf32x4(as_XMMRegister(n), as_XMMRegister(n), Address(rsp, n*16), 0);
754 }
755 } else {
756 for (int n = 0; n < num_xmm_regs; n++) {
757 __ movdqu(as_XMMRegister(n), Address(rsp, n*16));
758 }
759 }
760 #else
761 for (int n = 0; n < num_xmm_regs; n++) {
762 __ movdqu(as_XMMRegister(n), Address(rsp, n*16));
763 }
764 #endif
765 __ addptr(rsp, 16*num_xmm_regs);
766
767 #ifdef COMPILER2
768 if (MaxVectorSize > 16) {
769 // Restore upper half of YMM registers.
770 for (int n = 0; n < num_xmm_regs; n++) {
771 __ vinsertf128_high(as_XMMRegister(n), Address(rsp, n*16));
772 }
773 __ addptr(rsp, 16*num_xmm_regs);
774 if (UseAVX > 2) {
775 for (int n = 0; n < num_xmm_regs; n++) {
776 __ vinsertf64x4_high(as_XMMRegister(n), Address(rsp, n*32));
777 }
778 __ addptr(rsp, 32*num_xmm_regs);
779 }
780 }
781 #endif
782 }
783 }
784
785 #ifdef COMPILER1
786
787 #undef __
788 #define __ ce->masm()->
789
790 void ShenandoahBarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, ShenandoahPreBarrierStub* stub) {
791 ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
792 // At this point we know that marking is in progress.
793 // If do_load() is true then we have to emit the
794 // load of the previous value; otherwise it has already
795 // been loaded into _pre_val.
796
797 __ bind(*stub->entry());
798 assert(stub->pre_val()->is_register(), "Precondition.");
799
800 Register pre_val_reg = stub->pre_val()->as_register();
801
802 if (stub->do_load()) {
803 ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/, false /*unaligned*/);
804 }
805
806 __ cmpptr(pre_val_reg, (int32_t)NULL_WORD);
807 __ jcc(Assembler::equal, *stub->continuation());
808 ce->store_parameter(stub->pre_val()->as_register(), 0);
809 __ call(RuntimeAddress(bs->pre_barrier_c1_runtime_code_blob()->code_begin()));
810 __ jmp(*stub->continuation());
811
812 }
813
814 void ShenandoahBarrierSetAssembler::gen_write_barrier_stub(LIR_Assembler* ce, ShenandoahWriteBarrierStub* stub) {
815 __ bind(*stub->entry());
816
817 Label done;
818 Register obj = stub->obj()->as_register();
819 Register res = stub->result()->as_register();
820
821 if (res != obj) {
822 __ mov(res, obj);
823 }
824
825 // Check for null.
826 if (stub->needs_null_check()) {
827 __ testptr(res, res);
828 __ jcc(Assembler::zero, done);
829 }
830
831 write_barrier(ce->masm(), res);
832
833 __ bind(done);
834 __ jmp(*stub->continuation());
835 }
836
837 #undef __
838
839 #define __ sasm->
840
841 void ShenandoahBarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) {
842 __ prologue("shenandoah_pre_barrier", false);
843 // arg0 : previous value of memory
844
845 __ push(rax);
846 __ push(rdx);
847
848 const Register pre_val = rax;
849 const Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
850 const Register tmp = rdx;
851
852 NOT_LP64(__ get_thread(thread);)
853
854 Address queue_index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
855 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
856
857 Label done;
858 Label runtime;
859
860 // Is SATB still active?
861 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
862 __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL);
863 __ jcc(Assembler::zero, done);
864
865 // Can we store original value in the thread's buffer?
866
867 __ movptr(tmp, queue_index);
868 __ testptr(tmp, tmp);
869 __ jcc(Assembler::zero, runtime);
870 __ subptr(tmp, wordSize);
871 __ movptr(queue_index, tmp);
872 __ addptr(tmp, buffer);
873
874 // prev_val (rax)
875 __ load_parameter(0, pre_val);
876 __ movptr(Address(tmp, 0), pre_val);
877 __ jmp(done);
878
879 __ bind(runtime);
880
881 __ save_live_registers_no_oop_map(true);
882
883 // load the pre-value
884 __ load_parameter(0, rcx);
885 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), rcx, thread);
886
887 __ restore_live_registers(true);
888
889 __ bind(done);
890
891 __ pop(rdx);
892 __ pop(rax);
893
894 __ epilogue();
895 }
896
897 #undef __
898
899 #endif // COMPILER1
900
901 address ShenandoahBarrierSetAssembler::shenandoah_wb() {
902 assert(_shenandoah_wb != NULL, "need write barrier stub");
903 return _shenandoah_wb;
904 }
905
906 #define __ cgen->assembler()->
907
908 address ShenandoahBarrierSetAssembler::generate_shenandoah_wb(StubCodeGenerator* cgen) {
909 __ align(CodeEntryAlignment);
910 StubCodeMark mark(cgen, "StubRoutines", "shenandoah_wb");
911 address start = __ pc();
912
913 #ifdef _LP64
914 Label not_done;
915
916 // We use RDI, which also serves as argument register for slow call.
917 // RAX always holds the src object ptr, except after the slow call and
918 // the cmpxchg, then it holds the result.
919 // R8 and RCX are used as temporary registers.
920 __ push(rdi);
921 __ push(r8);
922
923 // Check for object beeing in the collection set.
924 // TODO: Can we use only 1 register here?
925 // The source object arrives here in rax.
926 // live: rax
927 // live: rdi
928 __ mov(rdi, rax);
929 __ shrptr(rdi, ShenandoahHeapRegion::region_size_bytes_shift_jint());
930 // live: r8
931 __ movptr(r8, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr());
932 __ movbool(r8, Address(r8, rdi, Address::times_1));
933 // unlive: rdi
934 __ testbool(r8);
935 // unlive: r8
936 __ jccb(Assembler::notZero, not_done);
937
938 __ pop(r8);
939 __ pop(rdi);
940 __ ret(0);
941
942 __ bind(not_done);
943
944 __ push(rcx);
945 __ push(rdx);
946 __ push(rdi);
947 __ push(rsi);
948 __ push(r8);
949 __ push(r9);
950 __ push(r10);
951 __ push(r11);
952 __ push(r12);
953 __ push(r13);
954 __ push(r14);
955 __ push(r15);
956 save_vector_registers(cgen->assembler());
957 __ movptr(rdi, rax);
958 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_JRT), rdi);
959 restore_vector_registers(cgen->assembler());
960 __ pop(r15);
961 __ pop(r14);
962 __ pop(r13);
963 __ pop(r12);
964 __ pop(r11);
965 __ pop(r10);
966 __ pop(r9);
967 __ pop(r8);
968 __ pop(rsi);
969 __ pop(rdi);
970 __ pop(rdx);
971 __ pop(rcx);
972
973 __ pop(r8);
974 __ pop(rdi);
975 __ ret(0);
976 #else
977 ShouldNotReachHere();
978 #endif
979 return start;
980 }
981
982 #undef __
983
984 void ShenandoahBarrierSetAssembler::barrier_stubs_init() {
985 if (ShenandoahWriteBarrier || ShenandoahStoreValEnqueueBarrier) {
986 int stub_code_size = 4096;
987 ResourceMark rm;
988 BufferBlob* bb = BufferBlob::create("shenandoah_barrier_stubs", stub_code_size);
989 CodeBuffer buf(bb);
990 StubCodeGenerator cgen(&buf);
991 _shenandoah_wb = generate_shenandoah_wb(&cgen);
992 }
993 }