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