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_lrb = 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 Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
142 #ifndef _LP64
143 __ push(thread);
144 __ get_thread(thread);
145 #endif
146
147 // Short-circuit if count == 0.
148 Label done;
149 __ testptr(count, count);
150 __ jcc(Assembler::zero, done);
151
152 // Skip runtime call if no forwarded objects.
153 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
154 __ testb(gc_state, ShenandoahHeap::UPDATEREFS);
155 __ jcc(Assembler::zero, done);
156
157 __ pusha(); // push registers (overkill)
158 #ifdef _LP64
159 if (c_rarg0 == count) { // On win64 c_rarg0 == rcx
160 assert_different_registers(c_rarg1, dst);
161 __ mov(c_rarg1, count);
162 __ mov(c_rarg0, dst);
163 } else {
164 assert_different_registers(c_rarg0, count);
165 __ mov(c_rarg0, dst);
166 __ mov(c_rarg1, count);
167 }
168 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry), 2);
169 #else
170 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry),
171 dst, count);
172 #endif
173 __ popa();
174
175 __ bind(done);
176 NOT_LP64(__ pop(thread);)
177 }
178 }
179
180 void ShenandoahBarrierSetAssembler::shenandoah_write_barrier_pre(MacroAssembler* masm,
181 Register obj,
182 Register pre_val,
183 Register thread,
184 Register tmp,
185 bool tosca_live,
186 bool expand_call) {
187
188 if (ShenandoahSATBBarrier) {
189 satb_write_barrier_pre(masm, obj, pre_val, thread, tmp, tosca_live, expand_call);
190 }
191 }
192
193 void ShenandoahBarrierSetAssembler::satb_write_barrier_pre(MacroAssembler* masm,
194 Register obj,
195 Register pre_val,
196 Register thread,
197 Register tmp,
198 bool tosca_live,
199 bool expand_call) {
200 // If expand_call is true then we expand the call_VM_leaf macro
201 // directly to skip generating the check by
202 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
203
204 #ifdef _LP64
205 assert(thread == r15_thread, "must be");
206 #endif // _LP64
207
208 Label done;
209 Label runtime;
210
211 assert(pre_val != noreg, "check this code");
212
213 if (obj != noreg) {
214 assert_different_registers(obj, pre_val, tmp);
215 assert(pre_val != rax, "check this code");
216 }
217
218 Address in_progress(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()));
219 Address index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
220 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
221
222 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
223 __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL);
224 __ jcc(Assembler::zero, done);
225
226 // Do we need to load the previous value?
227 if (obj != noreg) {
228 __ load_heap_oop(pre_val, Address(obj, 0), noreg, noreg, AS_RAW);
229 }
230
231 // Is the previous value null?
232 __ cmpptr(pre_val, (int32_t) NULL_WORD);
233 __ jcc(Assembler::equal, done);
234
235 // Can we store original value in the thread's buffer?
236 // Is index == 0?
237 // (The index field is typed as size_t.)
238
239 __ movptr(tmp, index); // tmp := *index_adr
240 __ cmpptr(tmp, 0); // tmp == 0?
241 __ jcc(Assembler::equal, runtime); // If yes, goto runtime
242
243 __ subptr(tmp, wordSize); // tmp := tmp - wordSize
244 __ movptr(index, tmp); // *index_adr := tmp
245 __ addptr(tmp, buffer); // tmp := tmp + *buffer_adr
246
247 // Record the previous value
248 __ movptr(Address(tmp, 0), pre_val);
249 __ jmp(done);
250
251 __ bind(runtime);
252 // save the live input values
253 if(tosca_live) __ push(rax);
254
255 if (obj != noreg && obj != rax)
256 __ push(obj);
257
258 if (pre_val != rax)
259 __ push(pre_val);
260
261 // Calling the runtime using the regular call_VM_leaf mechanism generates
262 // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
263 // that checks that the *(ebp+frame::interpreter_frame_last_sp) == NULL.
264 //
265 // If we care generating the pre-barrier without a frame (e.g. in the
266 // intrinsified Reference.get() routine) then ebp might be pointing to
267 // the caller frame and so this check will most likely fail at runtime.
268 //
269 // Expanding the call directly bypasses the generation of the check.
270 // So when we do not have have a full interpreter frame on the stack
271 // expand_call should be passed true.
272
273 NOT_LP64( __ push(thread); )
274
275 #ifdef _LP64
276 // We move pre_val into c_rarg0 early, in order to avoid smashing it, should
277 // pre_val be c_rarg1 (where the call prologue would copy thread argument).
278 // Note: this should not accidentally smash thread, because thread is always r15.
279 assert(thread != c_rarg0, "smashed arg");
280 if (c_rarg0 != pre_val) {
281 __ mov(c_rarg0, pre_val);
282 }
283 #endif
284
285 if (expand_call) {
286 LP64_ONLY( assert(pre_val != c_rarg1, "smashed arg"); )
287 #ifdef _LP64
288 if (c_rarg1 != thread) {
289 __ mov(c_rarg1, thread);
290 }
291 // Already moved pre_val into c_rarg0 above
292 #else
293 __ push(thread);
294 __ push(pre_val);
295 #endif
296 __ MacroAssembler::call_VM_leaf_base(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), 2);
297 } else {
298 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), LP64_ONLY(c_rarg0) NOT_LP64(pre_val), thread);
299 }
300
301 NOT_LP64( __ pop(thread); )
302
303 // save the live input values
304 if (pre_val != rax)
305 __ pop(pre_val);
306
307 if (obj != noreg && obj != rax)
308 __ pop(obj);
309
310 if(tosca_live) __ pop(rax);
311
312 __ bind(done);
313 }
314
315 void ShenandoahBarrierSetAssembler::resolve_forward_pointer(MacroAssembler* masm, Register dst) {
316 assert(ShenandoahCASBarrier, "should be enabled");
317 Label is_null;
318 __ testptr(dst, dst);
319 __ jcc(Assembler::zero, is_null);
320 resolve_forward_pointer_not_null(masm, dst);
321 __ bind(is_null);
322 }
323
324 void ShenandoahBarrierSetAssembler::resolve_forward_pointer_not_null(MacroAssembler* masm, Register dst) {
325 assert(ShenandoahCASBarrier || ShenandoahLoadRefBarrier, "should be enabled");
326 __ movptr(dst, Address(dst, ShenandoahBrooksPointer::byte_offset()));
327 }
328
329
330 void ShenandoahBarrierSetAssembler::load_reference_barrier_not_null(MacroAssembler* masm, Register dst) {
331 assert(ShenandoahLoadRefBarrier, "Should be enabled");
332 #ifdef _LP64
333 Label done;
334
335 Address gc_state(r15_thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
336 __ testb(gc_state, ShenandoahHeap::HAS_FORWARDED | ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL);
337 __ jccb(Assembler::zero, done);
338
339 // Heap is unstable, need to perform the resolve even if LRB is inactive
340 resolve_forward_pointer_not_null(masm, dst);
341
342 __ testb(gc_state, ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL);
343 __ jccb(Assembler::zero, done);
344
345 if (dst != rax) {
346 __ xchgptr(dst, rax); // Move obj into rax and save rax into obj.
347 }
348
349 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, ShenandoahBarrierSetAssembler::shenandoah_lrb())));
350
351 if (dst != rax) {
352 __ xchgptr(rax, dst); // Swap back obj with rax.
353 }
354
355 __ bind(done);
356 #else
357 Unimplemented();
358 #endif
359 }
360
361 void ShenandoahBarrierSetAssembler::storeval_barrier(MacroAssembler* masm, Register dst, Register tmp) {
362 if (ShenandoahStoreValEnqueueBarrier) {
363 storeval_barrier_impl(masm, dst, tmp);
364 }
365 }
366
367 void ShenandoahBarrierSetAssembler::storeval_barrier_impl(MacroAssembler* masm, Register dst, Register tmp) {
368 assert(ShenandoahStoreValEnqueueBarrier, "should be enabled");
369
370 if (dst == noreg) return;
371
372 #ifdef _LP64
373 if (ShenandoahStoreValEnqueueBarrier) {
374 // The set of registers to be saved+restored is the same as in the write-barrier above.
375 // Those are the commonly used registers in the interpreter.
376 __ pusha();
377 // __ push_callee_saved_registers();
378 __ subptr(rsp, 2 * Interpreter::stackElementSize);
379 __ movdbl(Address(rsp, 0), xmm0);
380
381 satb_write_barrier_pre(masm, noreg, dst, r15_thread, tmp, true, false);
382 __ movdbl(xmm0, Address(rsp, 0));
383 __ addptr(rsp, 2 * Interpreter::stackElementSize);
384 //__ pop_callee_saved_registers();
385 __ popa();
386 }
387 #else
388 Unimplemented();
389 #endif
390 }
391
392 void ShenandoahBarrierSetAssembler::load_reference_barrier(MacroAssembler* masm, Register dst) {
393 if (ShenandoahLoadRefBarrier) {
394 Label done;
395 __ testptr(dst, dst);
396 __ jcc(Assembler::zero, done);
397 load_reference_barrier_not_null(masm, dst);
398 __ bind(done);
399 }
400 }
401
402 void ShenandoahBarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
403 Register dst, Address src, Register tmp1, Register tmp_thread) {
404 bool on_oop = type == T_OBJECT || type == T_ARRAY;
405 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
406 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
407 bool on_reference = on_weak || on_phantom;
408 BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp_thread);
409 if (on_oop) {
410 load_reference_barrier(masm, dst);
411
412 if (ShenandoahKeepAliveBarrier && on_reference) {
413 const Register thread = NOT_LP64(tmp_thread) LP64_ONLY(r15_thread);
414 NOT_LP64(__ get_thread(thread));
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
428 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
429 Address dst, Register val, Register tmp1, Register tmp2) {
430
431 bool on_oop = type == T_OBJECT || type == T_ARRAY;
432 bool in_heap = (decorators & IN_HEAP) != 0;
433 bool as_normal = (decorators & AS_NORMAL) != 0;
434 if (on_oop && in_heap) {
435 bool needs_pre_barrier = as_normal;
436
437 Register tmp3 = LP64_ONLY(r8) NOT_LP64(rsi);
438 Register rthread = LP64_ONLY(r15_thread) NOT_LP64(rcx);
439 // flatten object address if needed
440 // We do it regardless of precise because we need the registers
441 if (dst.index() == noreg && dst.disp() == 0) {
442 if (dst.base() != tmp1) {
443 __ movptr(tmp1, dst.base());
444 }
445 } else {
446 __ lea(tmp1, dst);
447 }
448
449 #ifndef _LP64
450 InterpreterMacroAssembler *imasm = static_cast<InterpreterMacroAssembler*>(masm);
451 #endif
452
453 NOT_LP64(__ get_thread(rcx));
454 NOT_LP64(imasm->save_bcp());
455
456 if (needs_pre_barrier) {
457 shenandoah_write_barrier_pre(masm /*masm*/,
458 tmp1 /* obj */,
459 tmp2 /* pre_val */,
460 rthread /* thread */,
461 tmp3 /* tmp */,
462 val != noreg /* tosca_live */,
463 false /* expand_call */);
464 }
465 if (val == noreg) {
466 BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
467 } else {
468 storeval_barrier(masm, val, tmp3);
469 BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
470 }
471 NOT_LP64(imasm->restore_bcp());
472 } else {
473 BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2);
474 }
475 }
476
477 void ShenandoahBarrierSetAssembler::tlab_allocate(MacroAssembler* masm,
478 Register thread, Register obj,
479 Register var_size_in_bytes,
480 int con_size_in_bytes,
481 Register t1, Register t2,
482 Label& slow_case) {
483 assert_different_registers(obj, t1, t2);
484 assert_different_registers(obj, var_size_in_bytes, t1);
485 Register end = t2;
486 if (!thread->is_valid()) {
487 #ifdef _LP64
488 thread = r15_thread;
489 #else
490 assert(t1->is_valid(), "need temp reg");
491 thread = t1;
492 __ get_thread(thread);
493 #endif
494 }
495
496 __ verify_tlab();
497
498 __ movptr(obj, Address(thread, JavaThread::tlab_top_offset()));
499 if (var_size_in_bytes == noreg) {
500 __ lea(end, Address(obj, con_size_in_bytes + ShenandoahBrooksPointer::byte_size()));
501 } else {
502 __ addptr(var_size_in_bytes, ShenandoahBrooksPointer::byte_size());
503 __ lea(end, Address(obj, var_size_in_bytes, Address::times_1));
504 }
505 __ cmpptr(end, Address(thread, JavaThread::tlab_end_offset()));
506 __ jcc(Assembler::above, slow_case);
507
508 // update the tlab top pointer
509 __ movptr(Address(thread, JavaThread::tlab_top_offset()), end);
510
511 // Initialize brooks pointer
512 #ifdef _LP64
513 __ incrementq(obj, ShenandoahBrooksPointer::byte_size());
514 #else
515 __ incrementl(obj, ShenandoahBrooksPointer::byte_size());
516 #endif
517 __ movptr(Address(obj, ShenandoahBrooksPointer::byte_offset()), obj);
518
519 // recover var_size_in_bytes if necessary
520 if (var_size_in_bytes == end) {
521 __ subptr(var_size_in_bytes, obj);
522 }
523 __ verify_tlab();
524 }
525
526 // Special Shenandoah CAS implementation that handles false negatives
527 // due to concurrent evacuation.
528 #ifndef _LP64
529 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm,
530 Register res, Address addr, Register oldval, Register newval,
531 bool exchange, Register tmp1, Register tmp2) {
532 // Shenandoah has no 32-bit version for this.
533 Unimplemented();
534 }
535 #else
536 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm,
537 Register res, Address addr, Register oldval, Register newval,
538 bool exchange, Register tmp1, Register tmp2) {
539 assert(ShenandoahCASBarrier, "Should only be used when CAS barrier is enabled");
540 assert(oldval == rax, "must be in rax for implicit use in cmpxchg");
541
542 Label retry, done;
543
544 // Remember oldval for retry logic below
545 if (UseCompressedOops) {
546 __ movl(tmp1, oldval);
547 } else {
548 __ movptr(tmp1, oldval);
549 }
550
551 // Step 1. Try to CAS with given arguments. If successful, then we are done,
552 // and can safely return.
553 if (os::is_MP()) __ lock();
554 if (UseCompressedOops) {
555 __ cmpxchgl(newval, addr);
556 } else {
557 __ cmpxchgptr(newval, addr);
558 }
559 __ jcc(Assembler::equal, done, true);
560
561 // Step 2. CAS had failed. This may be a false negative.
562 //
563 // The trouble comes when we compare the to-space pointer with the from-space
564 // pointer to the same object. To resolve this, it will suffice to resolve both
565 // oldval and the value from memory -- this will give both to-space pointers.
566 // If they mismatch, then it was a legitimate failure.
567 //
568 if (UseCompressedOops) {
569 __ decode_heap_oop(tmp1);
570 }
571 resolve_forward_pointer(masm, tmp1);
572
573 if (UseCompressedOops) {
574 __ movl(tmp2, oldval);
575 __ decode_heap_oop(tmp2);
576 } else {
577 __ movptr(tmp2, oldval);
578 }
579 resolve_forward_pointer(masm, tmp2);
580
581 __ cmpptr(tmp1, tmp2);
582 __ jcc(Assembler::notEqual, done, true);
583
584 // Step 3. Try to CAS again with resolved to-space pointers.
585 //
586 // Corner case: it may happen that somebody stored the from-space pointer
587 // to memory while we were preparing for retry. Therefore, we can fail again
588 // on retry, and so need to do this in loop, always resolving the failure
589 // witness.
590 __ bind(retry);
591 if (os::is_MP()) __ lock();
592 if (UseCompressedOops) {
593 __ cmpxchgl(newval, addr);
594 } else {
595 __ cmpxchgptr(newval, addr);
596 }
597 __ jcc(Assembler::equal, done, true);
598
599 if (UseCompressedOops) {
600 __ movl(tmp2, oldval);
601 __ decode_heap_oop(tmp2);
602 } else {
603 __ movptr(tmp2, oldval);
604 }
605 resolve_forward_pointer(masm, tmp2);
606
607 __ cmpptr(tmp1, tmp2);
608 __ jcc(Assembler::equal, retry, true);
609
610 // Step 4. If we need a boolean result out of CAS, check the flag again,
611 // and promote the result. Note that we handle the flag from both the CAS
612 // itself and from the retry loop.
613 __ bind(done);
614 if (!exchange) {
615 assert(res != NULL, "need result register");
616 __ setb(Assembler::equal, res);
617 __ movzbl(res, res);
618 }
619 }
620 #endif // LP64
621
622 void ShenandoahBarrierSetAssembler::save_vector_registers(MacroAssembler* masm) {
623 int num_xmm_regs = LP64_ONLY(16) NOT_LP64(8);
624 if (UseAVX > 2) {
625 num_xmm_regs = LP64_ONLY(32) NOT_LP64(8);
626 }
627
628 if (UseSSE == 1) {
629 __ subptr(rsp, sizeof(jdouble)*8);
630 for (int n = 0; n < 8; n++) {
631 __ movflt(Address(rsp, n*sizeof(jdouble)), as_XMMRegister(n));
632 }
633 } else if (UseSSE >= 2) {
634 if (UseAVX > 2) {
635 __ push(rbx);
636 __ movl(rbx, 0xffff);
637 __ kmovwl(k1, rbx);
638 __ pop(rbx);
639 }
640 #ifdef COMPILER2
641 if (MaxVectorSize > 16) {
642 if(UseAVX > 2) {
643 // Save upper half of ZMM registers
644 __ subptr(rsp, 32*num_xmm_regs);
645 for (int n = 0; n < num_xmm_regs; n++) {
646 __ vextractf64x4_high(Address(rsp, n*32), as_XMMRegister(n));
647 }
648 }
649 assert(UseAVX > 0, "256 bit vectors are supported only with AVX");
650 // Save upper half of YMM registers
651 __ subptr(rsp, 16*num_xmm_regs);
652 for (int n = 0; n < num_xmm_regs; n++) {
653 __ vextractf128_high(Address(rsp, n*16), as_XMMRegister(n));
654 }
655 }
656 #endif
657 // Save whole 128bit (16 bytes) XMM registers
658 __ subptr(rsp, 16*num_xmm_regs);
659 #ifdef _LP64
660 if (VM_Version::supports_evex()) {
661 for (int n = 0; n < num_xmm_regs; n++) {
662 __ vextractf32x4(Address(rsp, n*16), as_XMMRegister(n), 0);
663 }
664 } else {
665 for (int n = 0; n < num_xmm_regs; n++) {
666 __ movdqu(Address(rsp, n*16), as_XMMRegister(n));
667 }
668 }
669 #else
670 for (int n = 0; n < num_xmm_regs; n++) {
671 __ movdqu(Address(rsp, n*16), as_XMMRegister(n));
672 }
673 #endif
674 }
675 }
676
677 void ShenandoahBarrierSetAssembler::restore_vector_registers(MacroAssembler* masm) {
678 int num_xmm_regs = LP64_ONLY(16) NOT_LP64(8);
679 if (UseAVX > 2) {
680 num_xmm_regs = LP64_ONLY(32) NOT_LP64(8);
681 }
682 if (UseSSE == 1) {
683 for (int n = 0; n < 8; n++) {
684 __ movflt(as_XMMRegister(n), Address(rsp, n*sizeof(jdouble)));
685 }
686 __ addptr(rsp, sizeof(jdouble)*8);
687 } else if (UseSSE >= 2) {
688 // Restore whole 128bit (16 bytes) XMM registers
689 #ifdef _LP64
690 if (VM_Version::supports_evex()) {
691 for (int n = 0; n < num_xmm_regs; n++) {
692 __ vinsertf32x4(as_XMMRegister(n), as_XMMRegister(n), Address(rsp, n*16), 0);
693 }
694 } else {
695 for (int n = 0; n < num_xmm_regs; n++) {
696 __ movdqu(as_XMMRegister(n), Address(rsp, n*16));
697 }
698 }
699 #else
700 for (int n = 0; n < num_xmm_regs; n++) {
701 __ movdqu(as_XMMRegister(n), Address(rsp, n*16));
702 }
703 #endif
704 __ addptr(rsp, 16*num_xmm_regs);
705
706 #ifdef COMPILER2
707 if (MaxVectorSize > 16) {
708 // Restore upper half of YMM registers.
709 for (int n = 0; n < num_xmm_regs; n++) {
710 __ vinsertf128_high(as_XMMRegister(n), Address(rsp, n*16));
711 }
712 __ addptr(rsp, 16*num_xmm_regs);
713 if (UseAVX > 2) {
714 for (int n = 0; n < num_xmm_regs; n++) {
715 __ vinsertf64x4_high(as_XMMRegister(n), Address(rsp, n*32));
716 }
717 __ addptr(rsp, 32*num_xmm_regs);
718 }
719 }
720 #endif
721 }
722 }
723
724 #ifdef COMPILER1
725
726 #undef __
727 #define __ ce->masm()->
728
729 void ShenandoahBarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, ShenandoahPreBarrierStub* stub) {
730 ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
731 // At this point we know that marking is in progress.
732 // If do_load() is true then we have to emit the
733 // load of the previous value; otherwise it has already
734 // been loaded into _pre_val.
735
736 __ bind(*stub->entry());
737 assert(stub->pre_val()->is_register(), "Precondition.");
738
739 Register pre_val_reg = stub->pre_val()->as_register();
740
741 if (stub->do_load()) {
742 ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/, false /*unaligned*/);
743 }
744
745 __ cmpptr(pre_val_reg, (int32_t)NULL_WORD);
746 __ jcc(Assembler::equal, *stub->continuation());
747 ce->store_parameter(stub->pre_val()->as_register(), 0);
748 __ call(RuntimeAddress(bs->pre_barrier_c1_runtime_code_blob()->code_begin()));
749 __ jmp(*stub->continuation());
750
751 }
752
753 void ShenandoahBarrierSetAssembler::gen_load_reference_barrier_stub(LIR_Assembler* ce, ShenandoahLoadReferenceBarrierStub* stub) {
754 __ bind(*stub->entry());
755
756 Label done;
757 Register obj = stub->obj()->as_register();
758 Register res = stub->result()->as_register();
759
760 if (res != obj) {
761 __ mov(res, obj);
762 }
763
764 // Check for null.
765 if (stub->needs_null_check()) {
766 __ testptr(res, res);
767 __ jcc(Assembler::zero, done);
768 }
769
770 load_reference_barrier_not_null(ce->masm(), res);
771
772 __ bind(done);
773 __ jmp(*stub->continuation());
774 }
775
776 #undef __
777
778 #define __ sasm->
779
780 void ShenandoahBarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) {
781 __ prologue("shenandoah_pre_barrier", false);
782 // arg0 : previous value of memory
783
784 __ push(rax);
785 __ push(rdx);
786
787 const Register pre_val = rax;
788 const Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
789 const Register tmp = rdx;
790
791 NOT_LP64(__ get_thread(thread);)
792
793 Address queue_index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
794 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
795
796 Label done;
797 Label runtime;
798
799 // Is SATB still active?
800 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
801 __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL);
802 __ jcc(Assembler::zero, done);
803
804 // Can we store original value in the thread's buffer?
805
806 __ movptr(tmp, queue_index);
807 __ testptr(tmp, tmp);
808 __ jcc(Assembler::zero, runtime);
809 __ subptr(tmp, wordSize);
810 __ movptr(queue_index, tmp);
811 __ addptr(tmp, buffer);
812
813 // prev_val (rax)
814 __ load_parameter(0, pre_val);
815 __ movptr(Address(tmp, 0), pre_val);
816 __ jmp(done);
817
818 __ bind(runtime);
819
820 __ save_live_registers_no_oop_map(true);
821
822 // load the pre-value
823 __ load_parameter(0, rcx);
824 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), rcx, thread);
825
826 __ restore_live_registers(true);
827
828 __ bind(done);
829
830 __ pop(rdx);
831 __ pop(rax);
832
833 __ epilogue();
834 }
835
836 #undef __
837
838 #endif // COMPILER1
839
840 address ShenandoahBarrierSetAssembler::shenandoah_lrb() {
841 assert(_shenandoah_lrb != NULL, "need load reference barrier stub");
842 return _shenandoah_lrb;
843 }
844
845 #define __ cgen->assembler()->
846
847 address ShenandoahBarrierSetAssembler::generate_shenandoah_lrb(StubCodeGenerator* cgen) {
848 __ align(CodeEntryAlignment);
849 StubCodeMark mark(cgen, "StubRoutines", "shenandoah_lrb");
850 address start = __ pc();
851
852 #ifdef _LP64
853 Label not_done;
854
855 // We use RDI, which also serves as argument register for slow call.
856 // RAX always holds the src object ptr, except after the slow call and
857 // the cmpxchg, then it holds the result.
858 // R8 and RCX are used as temporary registers.
859 __ push(rdi);
860 __ push(r8);
861
862 // Check for object beeing in the collection set.
863 // TODO: Can we use only 1 register here?
864 // The source object arrives here in rax.
865 // live: rax
866 // live: rdi
867 __ mov(rdi, rax);
868 __ shrptr(rdi, ShenandoahHeapRegion::region_size_bytes_shift_jint());
869 // live: r8
870 __ movptr(r8, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr());
871 __ movbool(r8, Address(r8, rdi, Address::times_1));
872 // unlive: rdi
873 __ testbool(r8);
874 // unlive: r8
875 __ jccb(Assembler::notZero, not_done);
876
877 __ pop(r8);
878 __ pop(rdi);
879 __ ret(0);
880
881 __ bind(not_done);
882
883 __ push(rcx);
884 __ push(rdx);
885 __ push(rdi);
886 __ push(rsi);
887 __ push(r8);
888 __ push(r9);
889 __ push(r10);
890 __ push(r11);
891 __ push(r12);
892 __ push(r13);
893 __ push(r14);
894 __ push(r15);
895 save_vector_registers(cgen->assembler());
896 __ movptr(rdi, rax);
897 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_JRT), rdi);
898 restore_vector_registers(cgen->assembler());
899 __ pop(r15);
900 __ pop(r14);
901 __ pop(r13);
902 __ pop(r12);
903 __ pop(r11);
904 __ pop(r10);
905 __ pop(r9);
906 __ pop(r8);
907 __ pop(rsi);
908 __ pop(rdi);
909 __ pop(rdx);
910 __ pop(rcx);
911
912 __ pop(r8);
913 __ pop(rdi);
914 __ ret(0);
915 #else
916 ShouldNotReachHere();
917 #endif
918 return start;
919 }
920
921 #undef __
922
923 void ShenandoahBarrierSetAssembler::barrier_stubs_init() {
924 if (ShenandoahLoadRefBarrier) {
925 int stub_code_size = 4096;
926 ResourceMark rm;
927 BufferBlob* bb = BufferBlob::create("shenandoah_barrier_stubs", stub_code_size);
928 CodeBuffer buf(bb);
929 StubCodeGenerator cgen(&buf);
930 _shenandoah_lrb = generate_shenandoah_lrb(&cgen);
931 }
932 }