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