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 __ 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::resolve_forward_pointer(MacroAssembler* masm, Register dst) {
297 assert(ShenandoahCASBarrier, "should be enabled");
298 Label is_null;
299 __ testptr(dst, dst);
300 __ jcc(Assembler::zero, is_null);
301 resolve_forward_pointer_not_null(masm, dst);
302 __ bind(is_null);
303 }
304
305 void ShenandoahBarrierSetAssembler::resolve_forward_pointer_not_null(MacroAssembler* masm, Register dst) {
306 assert(ShenandoahCASBarrier || ShenandoahLoadRefBarrier, "should be enabled");
307 __ movptr(dst, Address(dst, ShenandoahBrooksPointer::byte_offset()));
308 }
309
310
311 void ShenandoahBarrierSetAssembler::load_reference_barrier_not_null(MacroAssembler* masm, Register dst) {
312 assert(ShenandoahLoadRefBarrier, "Should be enabled");
313 #ifdef _LP64
314 Label done;
315
316 Address gc_state(r15_thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
317 __ testb(gc_state, ShenandoahHeap::HAS_FORWARDED | ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL);
318 __ jccb(Assembler::zero, done);
319
320 // Heap is unstable, need to perform the resolve even if LRB is inactive
321 resolve_forward_pointer_not_null(masm, dst);
322
323 __ testb(gc_state, ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL);
324 __ jccb(Assembler::zero, done);
325
326 if (dst != rax) {
327 __ xchgptr(dst, rax); // Move obj into rax and save rax into obj.
328 }
329
330 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, ShenandoahBarrierSetAssembler::shenandoah_lrb())));
331
332 if (dst != rax) {
333 __ xchgptr(rax, dst); // Swap back obj with rax.
334 }
335
336 __ bind(done);
337 #else
338 Unimplemented();
339 #endif
340 }
341
342 void ShenandoahBarrierSetAssembler::storeval_barrier(MacroAssembler* masm, Register dst, Register tmp) {
343 if (ShenandoahStoreValEnqueueBarrier) {
344 storeval_barrier_impl(masm, dst, tmp);
345 }
346 }
347
348 void ShenandoahBarrierSetAssembler::storeval_barrier_impl(MacroAssembler* masm, Register dst, Register tmp) {
349 assert(ShenandoahStoreValEnqueueBarrier, "should be enabled");
350
351 if (dst == noreg) return;
352
353 #ifdef _LP64
354 if (ShenandoahStoreValEnqueueBarrier) {
355 // The set of registers to be saved+restored is the same as in the write-barrier above.
356 // Those are the commonly used registers in the interpreter.
357 __ pusha();
358 // __ push_callee_saved_registers();
359 __ subptr(rsp, 2 * Interpreter::stackElementSize);
360 __ movdbl(Address(rsp, 0), xmm0);
361
362 satb_write_barrier_pre(masm, noreg, dst, r15_thread, tmp, true, false);
363 __ movdbl(xmm0, Address(rsp, 0));
364 __ addptr(rsp, 2 * Interpreter::stackElementSize);
365 //__ pop_callee_saved_registers();
366 __ popa();
367 }
368 #else
369 Unimplemented();
370 #endif
371 }
372
373 void ShenandoahBarrierSetAssembler::load_reference_barrier(MacroAssembler* masm, Register dst) {
374 if (ShenandoahLoadRefBarrier) {
375 Label done;
376 __ testptr(dst, dst);
377 __ jcc(Assembler::zero, done);
378 load_reference_barrier_not_null(masm, dst);
379 __ bind(done);
380 }
381 }
382
383 void ShenandoahBarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
384 Register dst, Address src, Register tmp1, Register tmp_thread) {
385 bool on_oop = type == T_OBJECT || type == T_ARRAY;
386 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
387 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
388 bool on_reference = on_weak || on_phantom;
389 BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp_thread);
390 if (on_oop) {
391 load_reference_barrier(masm, dst);
392
393 if (ShenandoahKeepAliveBarrier && on_reference) {
394 const Register thread = NOT_LP64(tmp_thread) LP64_ONLY(r15_thread);
395 NOT_LP64(__ get_thread(thread));
396 // Generate the SATB pre-barrier code to log the value of
397 // the referent field in an SATB buffer.
398 shenandoah_write_barrier_pre(masm /* masm */,
399 noreg /* obj */,
400 dst /* pre_val */,
401 thread /* thread */,
402 tmp1 /* tmp */,
403 true /* tosca_live */,
404 true /* expand_call */);
405 }
406 }
407 }
408
409 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
410 Address dst, Register val, Register tmp1, Register tmp2) {
411
412 bool on_oop = type == T_OBJECT || type == T_ARRAY;
413 bool in_heap = (decorators & IN_HEAP) != 0;
414 bool as_normal = (decorators & AS_NORMAL) != 0;
415 if (on_oop && in_heap) {
416 bool needs_pre_barrier = as_normal;
417
418 Register tmp3 = LP64_ONLY(r8) NOT_LP64(rsi);
419 Register rthread = LP64_ONLY(r15_thread) NOT_LP64(rcx);
420 // flatten object address if needed
421 // We do it regardless of precise because we need the registers
422 if (dst.index() == noreg && dst.disp() == 0) {
423 if (dst.base() != tmp1) {
424 __ movptr(tmp1, dst.base());
425 }
426 } else {
427 __ lea(tmp1, dst);
428 }
429
430 #ifndef _LP64
431 InterpreterMacroAssembler *imasm = static_cast<InterpreterMacroAssembler*>(masm);
432 #endif
433
434 NOT_LP64(__ get_thread(rcx));
435 NOT_LP64(imasm->save_bcp());
436
437 if (needs_pre_barrier) {
438 shenandoah_write_barrier_pre(masm /*masm*/,
439 tmp1 /* obj */,
440 tmp2 /* pre_val */,
441 rthread /* thread */,
442 tmp3 /* tmp */,
443 val != noreg /* tosca_live */,
444 false /* expand_call */);
445 }
446 if (val == noreg) {
447 BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
448 } else {
449 storeval_barrier(masm, val, tmp3);
450 BarrierSetAssembler::store_at(masm, decorators, type, Address(tmp1, 0), val, noreg, noreg);
451 }
452 NOT_LP64(imasm->restore_bcp());
453 } else {
454 BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2);
455 }
456 }
457
458 void ShenandoahBarrierSetAssembler::tlab_allocate(MacroAssembler* masm,
459 Register thread, Register obj,
460 Register var_size_in_bytes,
461 int con_size_in_bytes,
462 Register t1, Register t2,
463 Label& slow_case) {
464 assert_different_registers(obj, t1, t2);
465 assert_different_registers(obj, var_size_in_bytes, t1);
466 Register end = t2;
467 if (!thread->is_valid()) {
468 #ifdef _LP64
469 thread = r15_thread;
470 #else
471 assert(t1->is_valid(), "need temp reg");
472 thread = t1;
473 __ get_thread(thread);
474 #endif
475 }
476
477 __ verify_tlab();
478
479 __ movptr(obj, Address(thread, JavaThread::tlab_top_offset()));
480 if (var_size_in_bytes == noreg) {
481 __ lea(end, Address(obj, con_size_in_bytes + ShenandoahBrooksPointer::byte_size()));
482 } else {
483 __ addptr(var_size_in_bytes, ShenandoahBrooksPointer::byte_size());
484 __ lea(end, Address(obj, var_size_in_bytes, Address::times_1));
485 }
486 __ cmpptr(end, Address(thread, JavaThread::tlab_end_offset()));
487 __ jcc(Assembler::above, slow_case);
488
489 // update the tlab top pointer
490 __ movptr(Address(thread, JavaThread::tlab_top_offset()), end);
491
492 // Initialize brooks pointer
493 #ifdef _LP64
494 __ incrementq(obj, ShenandoahBrooksPointer::byte_size());
495 #else
496 __ incrementl(obj, ShenandoahBrooksPointer::byte_size());
497 #endif
498 __ movptr(Address(obj, ShenandoahBrooksPointer::byte_offset()), obj);
499
500 // recover var_size_in_bytes if necessary
501 if (var_size_in_bytes == end) {
502 __ subptr(var_size_in_bytes, obj);
503 }
504 __ verify_tlab();
505 }
506
507 // Special Shenandoah CAS implementation that handles false negatives
508 // due to concurrent evacuation.
509 #ifndef _LP64
510 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm,
511 Register res, Address addr, Register oldval, Register newval,
512 bool exchange, Register tmp1, Register tmp2) {
513 // Shenandoah has no 32-bit version for this.
514 Unimplemented();
515 }
516 #else
517 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm,
518 Register res, Address addr, Register oldval, Register newval,
519 bool exchange, Register tmp1, Register tmp2) {
520 assert(ShenandoahCASBarrier, "Should only be used when CAS barrier is enabled");
521 assert(oldval == rax, "must be in rax for implicit use in cmpxchg");
522
523 Label retry, done;
524
525 // Remember oldval for retry logic below
526 if (UseCompressedOops) {
527 __ movl(tmp1, oldval);
528 } else {
529 __ movptr(tmp1, oldval);
530 }
531
532 // Step 1. Try to CAS with given arguments. If successful, then we are done,
533 // and can safely return.
534 if (os::is_MP()) __ lock();
535 if (UseCompressedOops) {
536 __ cmpxchgl(newval, addr);
537 } else {
538 __ cmpxchgptr(newval, addr);
539 }
540 __ jcc(Assembler::equal, done, true);
541
542 // Step 2. CAS had failed. This may be a false negative.
543 //
544 // The trouble comes when we compare the to-space pointer with the from-space
545 // pointer to the same object. To resolve this, it will suffice to resolve both
546 // oldval and the value from memory -- this will give both to-space pointers.
547 // If they mismatch, then it was a legitimate failure.
548 //
549 if (UseCompressedOops) {
550 __ decode_heap_oop(tmp1);
551 }
552 resolve_forward_pointer(masm, tmp1);
553
554 if (UseCompressedOops) {
555 __ movl(tmp2, oldval);
556 __ decode_heap_oop(tmp2);
557 } else {
558 __ movptr(tmp2, oldval);
559 }
560 resolve_forward_pointer(masm, tmp2);
561
562 __ cmpptr(tmp1, tmp2);
563 __ jcc(Assembler::notEqual, done, true);
564
565 // Step 3. Try to CAS again with resolved to-space pointers.
566 //
567 // Corner case: it may happen that somebody stored the from-space pointer
568 // to memory while we were preparing for retry. Therefore, we can fail again
569 // on retry, and so need to do this in loop, always resolving the failure
570 // witness.
571 __ bind(retry);
572 if (os::is_MP()) __ lock();
573 if (UseCompressedOops) {
574 __ cmpxchgl(newval, addr);
575 } else {
576 __ cmpxchgptr(newval, addr);
577 }
578 __ jcc(Assembler::equal, done, true);
579
580 if (UseCompressedOops) {
581 __ movl(tmp2, oldval);
582 __ decode_heap_oop(tmp2);
583 } else {
584 __ movptr(tmp2, oldval);
585 }
586 resolve_forward_pointer(masm, tmp2);
587
588 __ cmpptr(tmp1, tmp2);
589 __ jcc(Assembler::equal, retry, true);
590
591 // Step 4. If we need a boolean result out of CAS, check the flag again,
592 // and promote the result. Note that we handle the flag from both the CAS
593 // itself and from the retry loop.
594 __ bind(done);
595 if (!exchange) {
596 assert(res != NULL, "need result register");
597 __ setb(Assembler::equal, res);
598 __ movzbl(res, res);
599 }
600 }
601 #endif // LP64
602
603 void ShenandoahBarrierSetAssembler::save_vector_registers(MacroAssembler* masm) {
604 int num_xmm_regs = LP64_ONLY(16) NOT_LP64(8);
605 if (UseAVX > 2) {
606 num_xmm_regs = LP64_ONLY(32) NOT_LP64(8);
607 }
608
609 if (UseSSE == 1) {
610 __ subptr(rsp, sizeof(jdouble)*8);
611 for (int n = 0; n < 8; n++) {
612 __ movflt(Address(rsp, n*sizeof(jdouble)), as_XMMRegister(n));
613 }
614 } else if (UseSSE >= 2) {
615 if (UseAVX > 2) {
616 __ push(rbx);
617 __ movl(rbx, 0xffff);
618 __ kmovwl(k1, rbx);
619 __ pop(rbx);
620 }
621 #ifdef COMPILER2
622 if (MaxVectorSize > 16) {
623 if(UseAVX > 2) {
624 // Save upper half of ZMM registers
625 __ subptr(rsp, 32*num_xmm_regs);
626 for (int n = 0; n < num_xmm_regs; n++) {
627 __ vextractf64x4_high(Address(rsp, n*32), as_XMMRegister(n));
628 }
629 }
630 assert(UseAVX > 0, "256 bit vectors are supported only with AVX");
631 // Save upper half of YMM registers
632 __ subptr(rsp, 16*num_xmm_regs);
633 for (int n = 0; n < num_xmm_regs; n++) {
634 __ vextractf128_high(Address(rsp, n*16), as_XMMRegister(n));
635 }
636 }
637 #endif
638 // Save whole 128bit (16 bytes) XMM registers
639 __ subptr(rsp, 16*num_xmm_regs);
640 #ifdef _LP64
641 if (VM_Version::supports_evex()) {
642 for (int n = 0; n < num_xmm_regs; n++) {
643 __ vextractf32x4(Address(rsp, n*16), as_XMMRegister(n), 0);
644 }
645 } else {
646 for (int n = 0; n < num_xmm_regs; n++) {
647 __ movdqu(Address(rsp, n*16), as_XMMRegister(n));
648 }
649 }
650 #else
651 for (int n = 0; n < num_xmm_regs; n++) {
652 __ movdqu(Address(rsp, n*16), as_XMMRegister(n));
653 }
654 #endif
655 }
656 }
657
658 void ShenandoahBarrierSetAssembler::restore_vector_registers(MacroAssembler* masm) {
659 int num_xmm_regs = LP64_ONLY(16) NOT_LP64(8);
660 if (UseAVX > 2) {
661 num_xmm_regs = LP64_ONLY(32) NOT_LP64(8);
662 }
663 if (UseSSE == 1) {
664 for (int n = 0; n < 8; n++) {
665 __ movflt(as_XMMRegister(n), Address(rsp, n*sizeof(jdouble)));
666 }
667 __ addptr(rsp, sizeof(jdouble)*8);
668 } else if (UseSSE >= 2) {
669 // Restore whole 128bit (16 bytes) XMM registers
670 #ifdef _LP64
671 if (VM_Version::supports_evex()) {
672 for (int n = 0; n < num_xmm_regs; n++) {
673 __ vinsertf32x4(as_XMMRegister(n), as_XMMRegister(n), Address(rsp, n*16), 0);
674 }
675 } else {
676 for (int n = 0; n < num_xmm_regs; n++) {
677 __ movdqu(as_XMMRegister(n), Address(rsp, n*16));
678 }
679 }
680 #else
681 for (int n = 0; n < num_xmm_regs; n++) {
682 __ movdqu(as_XMMRegister(n), Address(rsp, n*16));
683 }
684 #endif
685 __ addptr(rsp, 16*num_xmm_regs);
686
687 #ifdef COMPILER2
688 if (MaxVectorSize > 16) {
689 // Restore upper half of YMM registers.
690 for (int n = 0; n < num_xmm_regs; n++) {
691 __ vinsertf128_high(as_XMMRegister(n), Address(rsp, n*16));
692 }
693 __ addptr(rsp, 16*num_xmm_regs);
694 if (UseAVX > 2) {
695 for (int n = 0; n < num_xmm_regs; n++) {
696 __ vinsertf64x4_high(as_XMMRegister(n), Address(rsp, n*32));
697 }
698 __ addptr(rsp, 32*num_xmm_regs);
699 }
700 }
701 #endif
702 }
703 }
704
705 #ifdef COMPILER1
706
707 #undef __
708 #define __ ce->masm()->
709
710 void ShenandoahBarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, ShenandoahPreBarrierStub* stub) {
711 ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
712 // At this point we know that marking is in progress.
713 // If do_load() is true then we have to emit the
714 // load of the previous value; otherwise it has already
715 // been loaded into _pre_val.
716
717 __ bind(*stub->entry());
718 assert(stub->pre_val()->is_register(), "Precondition.");
719
720 Register pre_val_reg = stub->pre_val()->as_register();
721
722 if (stub->do_load()) {
723 ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/, false /*unaligned*/);
724 }
725
726 __ cmpptr(pre_val_reg, (int32_t)NULL_WORD);
727 __ jcc(Assembler::equal, *stub->continuation());
728 ce->store_parameter(stub->pre_val()->as_register(), 0);
729 __ call(RuntimeAddress(bs->pre_barrier_c1_runtime_code_blob()->code_begin()));
730 __ jmp(*stub->continuation());
731
732 }
733
734 void ShenandoahBarrierSetAssembler::gen_load_reference_barrier_stub(LIR_Assembler* ce, ShenandoahLoadReferenceBarrierStub* stub) {
735 __ bind(*stub->entry());
736
737 Label done;
738 Register obj = stub->obj()->as_register();
739 Register res = stub->result()->as_register();
740
741 if (res != obj) {
742 __ mov(res, obj);
743 }
744
745 // Check for null.
746 if (stub->needs_null_check()) {
747 __ testptr(res, res);
748 __ jcc(Assembler::zero, done);
749 }
750
751 load_reference_barrier_not_null(ce->masm(), res);
752
753 __ bind(done);
754 __ jmp(*stub->continuation());
755 }
756
757 #undef __
758
759 #define __ sasm->
760
761 void ShenandoahBarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) {
762 __ prologue("shenandoah_pre_barrier", false);
763 // arg0 : previous value of memory
764
765 __ push(rax);
766 __ push(rdx);
767
768 const Register pre_val = rax;
769 const Register thread = NOT_LP64(rax) LP64_ONLY(r15_thread);
770 const Register tmp = rdx;
771
772 NOT_LP64(__ get_thread(thread);)
773
774 Address queue_index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
775 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
776
777 Label done;
778 Label runtime;
779
780 // Is SATB still active?
781 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
782 __ testb(gc_state, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL);
783 __ jcc(Assembler::zero, done);
784
785 // Can we store original value in the thread's buffer?
786
787 __ movptr(tmp, queue_index);
788 __ testptr(tmp, tmp);
789 __ jcc(Assembler::zero, runtime);
790 __ subptr(tmp, wordSize);
791 __ movptr(queue_index, tmp);
792 __ addptr(tmp, buffer);
793
794 // prev_val (rax)
795 __ load_parameter(0, pre_val);
796 __ movptr(Address(tmp, 0), pre_val);
797 __ jmp(done);
798
799 __ bind(runtime);
800
801 __ save_live_registers_no_oop_map(true);
802
803 // load the pre-value
804 __ load_parameter(0, rcx);
805 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), rcx, thread);
806
807 __ restore_live_registers(true);
808
809 __ bind(done);
810
811 __ pop(rdx);
812 __ pop(rax);
813
814 __ epilogue();
815 }
816
817 #undef __
818
819 #endif // COMPILER1
820
821 address ShenandoahBarrierSetAssembler::shenandoah_lrb() {
822 assert(_shenandoah_lrb != NULL, "need load reference barrier stub");
823 return _shenandoah_lrb;
824 }
825
826 #define __ cgen->assembler()->
827
828 address ShenandoahBarrierSetAssembler::generate_shenandoah_lrb(StubCodeGenerator* cgen) {
829 __ align(CodeEntryAlignment);
830 StubCodeMark mark(cgen, "StubRoutines", "shenandoah_lrb");
831 address start = __ pc();
832
833 #ifdef _LP64
834 Label not_done;
835
836 // We use RDI, which also serves as argument register for slow call.
837 // RAX always holds the src object ptr, except after the slow call and
838 // the cmpxchg, then it holds the result.
839 // R8 and RCX are used as temporary registers.
840 __ push(rdi);
841 __ push(r8);
842
843 // Check for object beeing in the collection set.
844 // TODO: Can we use only 1 register here?
845 // The source object arrives here in rax.
846 // live: rax
847 // live: rdi
848 __ mov(rdi, rax);
849 __ shrptr(rdi, ShenandoahHeapRegion::region_size_bytes_shift_jint());
850 // live: r8
851 __ movptr(r8, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr());
852 __ movbool(r8, Address(r8, rdi, Address::times_1));
853 // unlive: rdi
854 __ testbool(r8);
855 // unlive: r8
856 __ jccb(Assembler::notZero, not_done);
857
858 __ pop(r8);
859 __ pop(rdi);
860 __ ret(0);
861
862 __ bind(not_done);
863
864 __ push(rcx);
865 __ push(rdx);
866 __ push(rdi);
867 __ push(rsi);
868 __ push(r8);
869 __ push(r9);
870 __ push(r10);
871 __ push(r11);
872 __ push(r12);
873 __ push(r13);
874 __ push(r14);
875 __ push(r15);
876 save_vector_registers(cgen->assembler());
877 __ movptr(rdi, rax);
878 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_JRT), rdi);
879 restore_vector_registers(cgen->assembler());
880 __ pop(r15);
881 __ pop(r14);
882 __ pop(r13);
883 __ pop(r12);
884 __ pop(r11);
885 __ pop(r10);
886 __ pop(r9);
887 __ pop(r8);
888 __ pop(rsi);
889 __ pop(rdi);
890 __ pop(rdx);
891 __ pop(rcx);
892
893 __ pop(r8);
894 __ pop(rdi);
895 __ ret(0);
896 #else
897 ShouldNotReachHere();
898 #endif
899 return start;
900 }
901
902 #undef __
903
904 void ShenandoahBarrierSetAssembler::barrier_stubs_init() {
905 if (ShenandoahLoadRefBarrier) {
906 int stub_code_size = 4096;
907 ResourceMark rm;
908 BufferBlob* bb = BufferBlob::create("shenandoah_barrier_stubs", stub_code_size);
909 CodeBuffer buf(bb);
910 StubCodeGenerator cgen(&buf);
911 _shenandoah_lrb = generate_shenandoah_lrb(&cgen);
912 }
913 }