1 /*
2 * Copyright (c) 2018, Red Hat, Inc. and/or its affiliates.
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/brooksPointer.hpp"
26 #include "gc/shenandoah/shenandoahBarrierSetAssembler.hpp"
27 #include "gc/shenandoah/shenandoahConnectionMatrix.hpp"
28 #include "gc/shenandoah/shenandoahHeapRegion.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 #ifdef COMPILER1
36 #include "c1/c1_LIRAssembler.hpp"
37 #include "c1/c1_MacroAssembler.hpp"
38 #include "gc/shenandoah/c1/shenandoahBarrierSetC1.hpp"
39 #endif
40
41 #define __ masm->
42
43 void ShenandoahBarrierSetAssembler::arraycopy_prologue(MacroAssembler* masm, DecoratorSet decorators, bool is_oop,
44 Register addr, Register count, RegSet saved_regs) {
45 if (is_oop) {
46 bool dest_uninitialized = (decorators & AS_DEST_NOT_INITIALIZED) != 0;
47 if (!dest_uninitialized) {
48 __ push(saved_regs, sp);
49 if (count == c_rarg0) {
50 if (addr == c_rarg1) {
51 // exactly backwards!!
52 __ mov(rscratch1, c_rarg0);
53 __ mov(c_rarg0, c_rarg1);
54 __ mov(c_rarg1, rscratch1);
55 } else {
56 __ mov(c_rarg1, count);
57 __ mov(c_rarg0, addr);
58 }
59 } else {
60 __ mov(c_rarg0, addr);
61 __ mov(c_rarg1, count);
62 }
63 if (UseCompressedOops) {
64 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_narrow_oop_entry), 2);
65 } else {
66 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry), 2);
67 }
68 __ pop(saved_regs, sp);
69 }
70 }
71 }
72
73 void ShenandoahBarrierSetAssembler::arraycopy_epilogue(MacroAssembler* masm, DecoratorSet decorators, bool is_oop,
74 Register start, Register end, Register scratch, RegSet saved_regs) {
75 if (is_oop) {
76 __ push(saved_regs, sp);
77 // must compute element count unless barrier set interface is changed (other platforms supply count)
78 assert_different_registers(start, end, scratch);
79 __ lea(scratch, Address(end, BytesPerHeapOop));
80 __ sub(scratch, scratch, start); // subtract start to get #bytes
81 __ lsr(scratch, scratch, LogBytesPerHeapOop); // convert to element count
82 __ mov(c_rarg0, start);
83 __ mov(c_rarg1, scratch);
84 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry), 2);
85 __ pop(saved_regs, sp);
86 }
87 }
88
89 void ShenandoahBarrierSetAssembler::shenandoah_write_barrier_pre(MacroAssembler* masm,
90 Register obj,
91 Register pre_val,
92 Register thread,
93 Register tmp,
94 bool tosca_live,
95 bool expand_call) {
96 if (ShenandoahSATBBarrier) {
97 satb_write_barrier_pre(masm, obj, pre_val, thread, tmp, tosca_live, expand_call);
98 }
99 }
100
101 void ShenandoahBarrierSetAssembler::satb_write_barrier_pre(MacroAssembler* masm,
102 Register obj,
103 Register pre_val,
104 Register thread,
105 Register tmp,
106 bool tosca_live,
107 bool expand_call) {
108 // If expand_call is true then we expand the call_VM_leaf macro
109 // directly to skip generating the check by
110 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
111
112 assert(thread == rthread, "must be");
113
114 Label done;
115 Label runtime;
116
117 assert_different_registers(obj, pre_val, tmp, rscratch1);
118 assert(pre_val != noreg && tmp != noreg, "expecting a register");
119
120 Address in_progress(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()));
121 Address index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
122 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
123
124 // Is marking active?
125 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
126 __ ldrw(tmp, in_progress);
127 } else {
128 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");
129 __ ldrb(tmp, in_progress);
130 }
131 __ cbzw(tmp, done);
132
133 // Do we need to load the previous value?
134 if (obj != noreg) {
135 __ load_heap_oop(pre_val, Address(obj, 0), noreg, noreg, AS_RAW);
136 }
137
138 // Is the previous value null?
139 __ cbz(pre_val, done);
140
141 // Can we store original value in the thread's buffer?
142 // Is index == 0?
143 // (The index field is typed as size_t.)
144
145 __ ldr(tmp, index); // tmp := *index_adr
146 __ cbz(tmp, runtime); // tmp == 0?
147 // If yes, goto runtime
148
149 __ sub(tmp, tmp, wordSize); // tmp := tmp - wordSize
150 __ str(tmp, index); // *index_adr := tmp
151 __ ldr(rscratch1, buffer);
152 __ add(tmp, tmp, rscratch1); // tmp := tmp + *buffer_adr
153
154 // Record the previous value
155 __ str(pre_val, Address(tmp, 0));
156 __ b(done);
157
158 __ bind(runtime);
159 // save the live input values
160 RegSet saved = RegSet::of(pre_val);
161 if (tosca_live) saved += RegSet::of(r0);
162 if (obj != noreg) saved += RegSet::of(obj);
163
164 __ push(saved, sp);
165
166 // Calling the runtime using the regular call_VM_leaf mechanism generates
167 // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
168 // that checks that the *(rfp+frame::interpreter_frame_last_sp) == NULL.
169 //
170 // If we care generating the pre-barrier without a frame (e.g. in the
171 // intrinsified Reference.get() routine) then ebp might be pointing to
172 // the caller frame and so this check will most likely fail at runtime.
173 //
174 // Expanding the call directly bypasses the generation of the check.
175 // So when we do not have have a full interpreter frame on the stack
176 // expand_call should be passed true.
177
178 if (expand_call) {
179 assert(pre_val != c_rarg1, "smashed arg");
180 __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), pre_val, thread);
181 } else {
182 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), pre_val, thread);
183 }
184
185 __ pop(saved, sp);
186
187 __ bind(done);
188 }
189
190 void ShenandoahBarrierSetAssembler::shenandoah_write_barrier_post(MacroAssembler* masm,
191 Register store_addr,
192 Register new_val,
193 Register thread,
194 Register tmp,
195 Register tmp2) {
196 assert(thread == rthread, "must be");
197 assert(UseShenandoahGC, "expect Shenandoah GC");
198
199 if (! UseShenandoahMatrix) {
200 // No need for that barrier if not using matrix.
201 return;
202 }
203
204 assert_different_registers(store_addr, new_val, thread, tmp, tmp2, rscratch1);
205
206 Label done;
207 __ cbz(new_val, done);
208
209 ShenandoahConnectionMatrix* matrix = ShenandoahHeap::heap()->connection_matrix();
210
211 // Compute to-region index
212 __ lsr(tmp, new_val, ShenandoahHeapRegion::region_size_bytes_shift_jint());
213
214 // Compute from-region index
215 __ lsr(tmp2, store_addr, ShenandoahHeapRegion::region_size_bytes_shift_jint());
216
217 // Compute matrix index
218 __ mov(rscratch1, matrix->stride_jint());
219 // Address is _matrix[to * stride + from]
220 __ madd(tmp, tmp, rscratch1, tmp2);
221 __ mov(rscratch1, matrix->magic_offset());
222 Address loc(tmp, rscratch1);
223
224 __ ldrb(tmp2, loc);
225 __ cbnz(tmp2, done);
226 __ mov(tmp2, 1);
227 __ strb(tmp2, loc);
228 __ bind(done);
229 }
230
231 void ShenandoahBarrierSetAssembler::read_barrier(MacroAssembler* masm, Register dst) {
232 if (ShenandoahReadBarrier) {
233 read_barrier_impl(masm, dst);
234 }
235 }
236
237 void ShenandoahBarrierSetAssembler::read_barrier_impl(MacroAssembler* masm, Register dst) {
238 assert(UseShenandoahGC && (ShenandoahReadBarrier || ShenandoahStoreValReadBarrier), "should be enabled");
239 Label is_null;
240 __ cbz(dst, is_null);
241 read_barrier_not_null_impl(masm, dst);
242 __ bind(is_null);
243 }
244
245 void ShenandoahBarrierSetAssembler::read_barrier_not_null(MacroAssembler* masm, Register dst) {
246 if (ShenandoahReadBarrier) {
247 read_barrier_not_null_impl(masm, dst);
248 }
249 }
250
251
252 void ShenandoahBarrierSetAssembler::read_barrier_not_null_impl(MacroAssembler* masm, Register dst) {
253 assert(UseShenandoahGC && (ShenandoahReadBarrier || ShenandoahStoreValReadBarrier), "should be enabled");
254 __ ldr(dst, Address(dst, BrooksPointer::byte_offset()));
255 }
256
257 void ShenandoahBarrierSetAssembler::write_barrier(MacroAssembler* masm, Register dst) {
258 if (ShenandoahWriteBarrier) {
259 write_barrier_impl(masm, dst);
260 }
261 }
262
263 void ShenandoahBarrierSetAssembler::write_barrier_impl(MacroAssembler* masm, Register dst) {
264 assert(UseShenandoahGC && (ShenandoahWriteBarrier || ShenandoahStoreValEnqueueBarrier), "should be enabled");
265 assert(dst != rscratch1, "different regs");
266 assert(dst != rscratch2, "Need rscratch2");
267
268 Label done;
269
270 Address gc_state(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
271 __ ldrb(rscratch1, gc_state);
272 __ membar(Assembler::LoadLoad);
273
274 // Now check if evacuation is in progress.
275 read_barrier_not_null(masm, dst);
276
277 __ mov(rscratch2, ShenandoahHeap::EVACUATION | ShenandoahHeap::TRAVERSAL);
278 __ tst(rscratch1, rscratch2);
279 __ br(Assembler::EQ, done);
280
281 __ lsr(rscratch1, dst, ShenandoahHeapRegion::region_size_bytes_shift_jint());
282 __ mov(rscratch2, ShenandoahHeap::in_cset_fast_test_addr());
283 __ ldrb(rscratch2, Address(rscratch2, rscratch1));
284 __ tst(rscratch2, 0x1);
285 __ br(Assembler::EQ, done);
286
287 // Save possibly live regs.
288 RegSet live_regs = RegSet::range(r0, r4) - dst;
289 __ push(live_regs, sp);
290 __ strd(v0, __ pre(sp, 2 * -wordSize));
291
292 // Call into runtime
293 __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_barrier_IRT), dst);
294
295 // Move result into dst reg.
296 __ mov(dst, r0);
297
298 // Restore possibly live regs.
299 __ ldrd(v0, __ post(sp, 2 * wordSize));
300 __ pop(live_regs, sp);
301
302 __ bind(done);
303 }
304
305 void ShenandoahBarrierSetAssembler::storeval_barrier(MacroAssembler* masm, Register dst, Register tmp) {
306 if (ShenandoahStoreValEnqueueBarrier) {
307 Label is_null;
308 __ cbz(dst, is_null);
309 write_barrier_impl(masm, dst);
310 __ bind(is_null);
311 // Save possibly live regs.
312 RegSet live_regs = RegSet::range(r0, r4) - dst;
313 __ push(live_regs, sp);
314 __ strd(v0, __ pre(sp, 2 * -wordSize));
315
316 satb_write_barrier_pre(masm, noreg, dst, rthread, tmp, true, false);
317
318 // Restore possibly live regs.
319 __ ldrd(v0, __ post(sp, 2 * wordSize));
320 __ pop(live_regs, sp);
321 }
322 if (ShenandoahStoreValReadBarrier) {
323 read_barrier_impl(masm, dst);
324 }
325 }
326
327 void ShenandoahBarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
328 Register dst, Address src, Register tmp1, Register tmp_thread) {
329 bool on_oop = type == T_OBJECT || type == T_ARRAY;
330 bool in_heap = (decorators & IN_HEAP) != 0;
331 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
332 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
333 bool on_reference = on_weak || on_phantom;
334
335 if (in_heap) {
336 read_barrier_not_null(masm, src.base());
337 }
338
339 BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp_thread);
340 if (ShenandoahKeepAliveBarrier && on_oop && on_reference) {
341 __ enter();
342 satb_write_barrier_pre(masm /* masm */,
343 noreg /* obj */,
344 dst /* pre_val */,
345 rthread /* thread */,
346 tmp1 /* tmp */,
347 true /* tosca_live */,
348 true /* expand_call */);
349 __ leave();
350 }
351 }
352
353 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
354 Address dst, Register val, Register tmp1, Register tmp2) {
355 bool on_oop = type == T_OBJECT || type == T_ARRAY;
356 bool in_heap = (decorators & IN_HEAP) != 0;
357 if (in_heap) {
358 write_barrier(masm, dst.base());
359 }
360 if (!on_oop) {
361 BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2);
362 return;
363 }
364
365 // flatten object address if needed
366 if (dst.index() == noreg && dst.offset() == 0) {
367 if (dst.base() != r3) {
368 __ mov(r3, dst.base());
369 }
370 } else {
371 __ lea(r3, dst);
372 }
373
374 shenandoah_write_barrier_pre(masm,
375 r3 /* obj */,
376 tmp2 /* pre_val */,
377 rthread /* thread */,
378 tmp1 /* tmp */,
379 val != noreg /* tosca_live */,
380 false /* expand_call */);
381
382 if (val == noreg) {
383 BarrierSetAssembler::store_at(masm, decorators, type, Address(r3, 0), noreg, noreg, noreg);
384 } else {
385 storeval_barrier(masm, val, tmp1);
386 // G1 barrier needs uncompressed oop for region cross check.
387 Register new_val = val;
388 if (UseCompressedOops) {
389 new_val = rscratch2;
390 __ mov(new_val, val);
391 }
392 BarrierSetAssembler::store_at(masm, decorators, type, Address(r3, 0), val, noreg, noreg);
393 shenandoah_write_barrier_post(masm,
394 r3 /* store_adr */,
395 new_val /* new_val */,
396 rthread /* thread */,
397 tmp1 /* tmp */,
398 tmp2 /* tmp2 */);
399 }
400
401 }
402
403 void ShenandoahBarrierSetAssembler::obj_equals(MacroAssembler* masm, DecoratorSet decorators, Register op1, Register op2) {
404 __ cmp(op1, op2);
405 if (ShenandoahAcmpBarrier) {
406 Label done;
407 __ br(Assembler::EQ, done);
408 // The object may have been evacuated, but we won't see it without a
409 // membar here.
410 __ membar(Assembler::LoadStore| Assembler::LoadLoad);
411 read_barrier(masm, op1);
412 read_barrier(masm, op2);
413 __ cmp(op1, op2);
414 __ bind(done);
415 }
416 }
417
418 void ShenandoahBarrierSetAssembler::resolve_for_read(MacroAssembler* masm, DecoratorSet decorators, Register obj) {
419 bool oop_not_null = (decorators & OOP_NOT_NULL) != 0;
420 if (oop_not_null) {
421 read_barrier_not_null(masm, obj);
422 } else {
423 read_barrier(masm, obj);
424 }
425 }
426
427 void ShenandoahBarrierSetAssembler::resolve_for_write(MacroAssembler* masm, DecoratorSet decorators, Register obj) {
428 write_barrier(masm, obj);
429 }
430
431 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm, Register addr, Register expected, Register new_val,
432 bool acquire, bool release, bool weak, bool encode,
433 Register tmp1, Register tmp2, Register tmp3,
434 Register result) {
435
436 if (encode) {
437 storeval_barrier(masm, new_val, tmp3);
438 }
439
440 if (UseCompressedOops) {
441 if (encode) {
442 __ encode_heap_oop(tmp1, expected);
443 expected = tmp1;
444 __ encode_heap_oop(tmp2, new_val);
445 new_val = tmp2;
446 }
447 }
448 bool is_cae = (result != noreg);
449 bool is_narrow = UseCompressedOops;
450 Assembler::operand_size size = is_narrow ? Assembler::word : Assembler::xword;
451 if (! is_cae) result = rscratch1;
452
453 assert_different_registers(addr, expected, new_val, result, tmp3);
454
455 Label retry, done, fail;
456
457 // CAS, using LL/SC pair.
458 __ bind(retry);
459 __ load_exclusive(result, addr, size, acquire);
460 if (is_narrow) {
461 __ cmpw(result, expected);
462 } else {
463 __ cmp(result, expected);
464 }
465 __ br(Assembler::NE, fail);
466 __ store_exclusive(tmp3, new_val, addr, size, release);
467 if (weak) {
468 __ cmpw(tmp3, 0u); // If the store fails, return NE to our caller
469 } else {
470 __ cbnzw(tmp3, retry);
471 }
472 __ b(done);
473
474 __ bind(fail);
475 // Check if rb(expected)==rb(result)
476 // Shuffle registers so that we have memory value ready for next expected.
477 __ mov(tmp3, expected);
478 __ mov(expected, result);
479 if (is_narrow) {
480 __ decode_heap_oop(result, result);
481 __ decode_heap_oop(tmp3, tmp3);
482 }
483 __ resolve_for_read(0, result);
484 __ resolve_for_read(0, tmp3);
485 __ cmp(result, tmp3);
486 // Retry with expected now being the value we just loaded from addr.
487 __ br(Assembler::EQ, retry);
488 if (is_narrow && is_cae) {
489 // For cmp-and-exchange and narrow oops, we need to restore
490 // the compressed old-value. We moved it to 'expected' a few lines up.
491 __ mov(result, expected);
492 }
493 __ bind(done);
494
495 }
496
497 #ifdef COMPILER1
498
499 #undef __
500 #define __ ce->masm()->
501
502 void ShenandoahBarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, ShenandoahPreBarrierStub* stub) {
503 ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
504 // At this point we know that marking is in progress.
505 // If do_load() is true then we have to emit the
506 // load of the previous value; otherwise it has already
507 // been loaded into _pre_val.
508
509 __ bind(*stub->entry());
510
511 assert(stub->pre_val()->is_register(), "Precondition.");
512
513 Register pre_val_reg = stub->pre_val()->as_register();
514
515 if (stub->do_load()) {
516 ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/, false /*unaligned*/);
517 }
518 __ cbz(pre_val_reg, *stub->continuation());
519 ce->store_parameter(stub->pre_val()->as_register(), 0);
520 __ far_call(RuntimeAddress(bs->pre_barrier_c1_runtime_code_blob()->code_begin()));
521 __ b(*stub->continuation());
522 }
523
524 #undef __
525
526 #define __ sasm->
527
528 void ShenandoahBarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) {
529 __ prologue("shenandoah_pre_barrier", false);
530
531 // arg0 : previous value of memory
532
533 BarrierSet* bs = BarrierSet::barrier_set();
534
535 const Register pre_val = r0;
536 const Register thread = rthread;
537 const Register tmp = rscratch1;
538
539 Address queue_index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
540 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
541
542 Label done;
543 Label runtime;
544
545 // Is marking still active?
546 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
547 __ ldrb(tmp, gc_state);
548 __ mov(rscratch2, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL);
549 __ tst(tmp, rscratch2);
550 __ br(Assembler::EQ, done);
551
552 // Can we store original value in the thread's buffer?
553 __ ldr(tmp, queue_index);
554 __ cbz(tmp, runtime);
555
556 __ sub(tmp, tmp, wordSize);
557 __ str(tmp, queue_index);
558 __ ldr(rscratch2, buffer);
559 __ add(tmp, tmp, rscratch2);
560 __ load_parameter(0, rscratch2);
561 __ str(rscratch2, Address(tmp, 0));
562 __ b(done);
563
564 __ bind(runtime);
565 __ push_call_clobbered_registers();
566 __ load_parameter(0, pre_val);
567 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), pre_val, thread);
568 __ pop_call_clobbered_registers();
569 __ bind(done);
570
571 __ epilogue();
572 }
573
574 #undef __
575
576 #endif // COMPILER1