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/shenandoahForwarding.hpp"
27 #include "gc/shenandoah/shenandoahHeap.hpp"
28 #include "gc/shenandoah/shenandoahHeapRegion.hpp"
29 #include "gc/shenandoah/shenandoahHeuristics.hpp"
30 #include "gc/shenandoah/shenandoahRuntime.hpp"
31 #include "gc/shenandoah/shenandoahThreadLocalData.hpp"
32 #include "interpreter/interpreter.hpp"
33 #include "interpreter/interp_masm.hpp"
34 #include "runtime/sharedRuntime.hpp"
35 #include "runtime/thread.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, bool is_oop,
47 Register addr, Register count, RegSet saved_regs) {
48 if (is_oop) {
49 bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0;
50 if (ShenandoahSATBBarrier && !dest_uninitialized && !ShenandoahHeap::heap()->heuristics()->can_do_traversal_gc()) {
51
52 Label done;
53
54 // Avoid calling runtime if count == 0
55 __ cbz(count, done);
56
57 // Is marking active?
58 Address gc_state(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
59 __ ldrb(rscratch1, gc_state);
60 __ tbz(rscratch1, ShenandoahHeap::MARKING_BITPOS, done);
61
62 __ push(saved_regs, sp);
63 if (count == c_rarg0) {
64 if (addr == c_rarg1) {
65 // exactly backwards!!
66 __ mov(rscratch1, c_rarg0);
67 __ mov(c_rarg0, c_rarg1);
68 __ mov(c_rarg1, rscratch1);
69 } else {
70 __ mov(c_rarg1, count);
71 __ mov(c_rarg0, addr);
72 }
73 } else {
74 __ mov(c_rarg0, addr);
75 __ mov(c_rarg1, count);
76 }
77 if (UseCompressedOops) {
78 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_narrow_oop_entry), 2);
79 } else {
80 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_pre_oop_entry), 2);
81 }
82 __ pop(saved_regs, sp);
83 __ bind(done);
84 }
85 }
86 }
87
88 void ShenandoahBarrierSetAssembler::arraycopy_epilogue(MacroAssembler* masm, DecoratorSet decorators, bool is_oop,
89 Register start, Register count, Register scratch, RegSet saved_regs) {
90 if (is_oop) {
91 Label done;
92
93 // Avoid calling runtime if count == 0
94 __ cbz(count, done);
95
96 // Is updating references?
97 Address gc_state(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
98 __ ldrb(rscratch1, gc_state);
99 __ tbz(rscratch1, ShenandoahHeap::UPDATEREFS_BITPOS, done);
100
101 __ push(saved_regs, sp);
102 assert_different_registers(start, count, scratch);
103 assert_different_registers(c_rarg0, count);
104 __ mov(c_rarg0, start);
105 __ mov(c_rarg1, count);
106 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_array_post_entry), 2);
107 __ pop(saved_regs, sp);
108
109 __ bind(done);
110 }
111 }
112
113 void ShenandoahBarrierSetAssembler::shenandoah_write_barrier_pre(MacroAssembler* masm,
114 Register obj,
115 Register pre_val,
116 Register thread,
117 Register tmp,
118 bool tosca_live,
119 bool expand_call) {
120 if (ShenandoahSATBBarrier) {
121 satb_write_barrier_pre(masm, obj, pre_val, thread, tmp, tosca_live, expand_call);
122 }
123 }
124
125 void ShenandoahBarrierSetAssembler::satb_write_barrier_pre(MacroAssembler* masm,
126 Register obj,
127 Register pre_val,
128 Register thread,
129 Register tmp,
130 bool tosca_live,
131 bool expand_call) {
132 // If expand_call is true then we expand the call_VM_leaf macro
133 // directly to skip generating the check by
134 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
135
136 assert(thread == rthread, "must be");
137
138 Label done;
139 Label runtime;
140
141 assert_different_registers(obj, pre_val, tmp, rscratch1);
142 assert(pre_val != noreg && tmp != noreg, "expecting a register");
143
144 Address in_progress(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_active_offset()));
145 Address index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
146 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
147
148 // Is marking active?
149 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
150 __ ldrw(tmp, in_progress);
151 } else {
152 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");
153 __ ldrb(tmp, in_progress);
154 }
155 __ cbzw(tmp, done);
156
157 // Do we need to load the previous value?
158 if (obj != noreg) {
159 __ load_heap_oop(pre_val, Address(obj, 0), noreg, noreg, AS_RAW);
160 }
161
162 // Is the previous value null?
163 __ cbz(pre_val, done);
164
165 // Can we store original value in the thread's buffer?
166 // Is index == 0?
167 // (The index field is typed as size_t.)
168
169 __ ldr(tmp, index); // tmp := *index_adr
170 __ cbz(tmp, runtime); // tmp == 0?
171 // If yes, goto runtime
172
173 __ sub(tmp, tmp, wordSize); // tmp := tmp - wordSize
174 __ str(tmp, index); // *index_adr := tmp
175 __ ldr(rscratch1, buffer);
176 __ add(tmp, tmp, rscratch1); // tmp := tmp + *buffer_adr
177
178 // Record the previous value
179 __ str(pre_val, Address(tmp, 0));
180 __ b(done);
181
182 __ bind(runtime);
183 // save the live input values
184 RegSet saved = RegSet::of(pre_val);
185 if (tosca_live) saved += RegSet::of(r0);
186 if (obj != noreg) saved += RegSet::of(obj);
187
188 __ push(saved, sp);
189
190 // Calling the runtime using the regular call_VM_leaf mechanism generates
191 // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
192 // that checks that the *(rfp+frame::interpreter_frame_last_sp) == NULL.
193 //
194 // If we care generating the pre-barrier without a frame (e.g. in the
195 // intrinsified Reference.get() routine) then ebp might be pointing to
196 // the caller frame and so this check will most likely fail at runtime.
197 //
198 // Expanding the call directly bypasses the generation of the check.
199 // So when we do not have have a full interpreter frame on the stack
200 // expand_call should be passed true.
201
202 if (expand_call) {
203 assert(pre_val != c_rarg1, "smashed arg");
204 __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), pre_val, thread);
205 } else {
206 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), pre_val, thread);
207 }
208
209 __ pop(saved, sp);
210
211 __ bind(done);
212 }
213
214 void ShenandoahBarrierSetAssembler::resolve_forward_pointer(MacroAssembler* masm, Register dst, Register tmp) {
215 assert(ShenandoahLoadRefBarrier || ShenandoahCASBarrier, "Should be enabled");
216 Label is_null;
217 __ cbz(dst, is_null);
218 resolve_forward_pointer_not_null(masm, dst, tmp);
219 __ bind(is_null);
220 }
221
222 // IMPORTANT: This must preserve all registers, even rscratch1 and rscratch2, except those explicitely
223 // passed in.
224 void ShenandoahBarrierSetAssembler::resolve_forward_pointer_not_null(MacroAssembler* masm, Register dst, Register tmp) {
225 assert(ShenandoahLoadRefBarrier || ShenandoahCASBarrier, "Should be enabled");
226 // The below loads the mark word, checks if the lowest two bits are
227 // set, and if so, clear the lowest two bits and copy the result
228 // to dst. Otherwise it leaves dst alone.
229 // Implementing this is surprisingly awkward. I do it here by:
230 // - Inverting the mark word
231 // - Test lowest two bits == 0
232 // - If so, set the lowest two bits
233 // - Invert the result back, and copy to dst
234 Label done;
235 __ ldr(tmp, Address(dst, oopDesc::mark_offset_in_bytes()));
236 __ eon(tmp, tmp, zr);
237 __ ands(zr, tmp, markOopDesc::lock_mask_in_place);
238 __ br(Assembler::NE, done);
239 __ orr(tmp, tmp, markOopDesc::marked_value);
240 __ eon(dst, tmp, zr);
241 __ bind(done);
242 }
243
244 void ShenandoahBarrierSetAssembler::load_reference_barrier_not_null(MacroAssembler* masm, Register dst, Register tmp) {
245 assert(ShenandoahLoadRefBarrier, "Should be enabled");
246 assert(dst != rscratch2, "need rscratch2");
247
248 Label done;
249 __ enter();
250 Address gc_state(rthread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
251 __ ldrb(rscratch2, gc_state);
252
253 // Check for heap stability
254 __ tbz(rscratch2, ShenandoahHeap::HAS_FORWARDED_BITPOS, done);
255
256 RegSet to_save = RegSet::of(r0);
257 if (dst != r0) {
258 __ push(to_save, sp);
259 __ mov(r0, dst);
260 }
261
262 __ far_call(RuntimeAddress(CAST_FROM_FN_PTR(address, ShenandoahBarrierSetAssembler::shenandoah_lrb())));
263
264 if (dst != r0) {
265 __ mov(dst, r0);
266 __ pop(to_save, sp);
267 }
268
269 __ bind(done);
270 __ leave();
271 }
272
273 void ShenandoahBarrierSetAssembler::storeval_barrier(MacroAssembler* masm, Register dst, Register tmp) {
274 if (ShenandoahStoreValEnqueueBarrier) {
275 // Save possibly live regs.
276 RegSet live_regs = RegSet::range(r0, r4) - dst;
277 __ push(live_regs, sp);
278 __ strd(v0, __ pre(sp, 2 * -wordSize));
279
280 satb_write_barrier_pre(masm, noreg, dst, rthread, tmp, true, false);
281
282 // Restore possibly live regs.
283 __ ldrd(v0, __ post(sp, 2 * wordSize));
284 __ pop(live_regs, sp);
285 }
286 }
287
288 void ShenandoahBarrierSetAssembler::load_reference_barrier(MacroAssembler* masm, Register dst, Register tmp) {
289 if (ShenandoahLoadRefBarrier) {
290 Label is_null;
291 __ cbz(dst, is_null);
292 load_reference_barrier_not_null(masm, dst, tmp);
293 __ bind(is_null);
294 }
295 }
296
297 void ShenandoahBarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
298 Register dst, Address src, Register tmp1, Register tmp_thread) {
299 bool on_oop = type == T_OBJECT || type == T_ARRAY;
300 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
301 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
302 bool on_reference = on_weak || on_phantom;
303
304 BarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp_thread);
305 if (on_oop) {
306 load_reference_barrier(masm, dst, tmp1);
307
308 if (ShenandoahKeepAliveBarrier && on_reference) {
309 __ enter();
310 satb_write_barrier_pre(masm /* masm */,
311 noreg /* obj */,
312 dst /* pre_val */,
313 rthread /* thread */,
314 tmp1 /* tmp */,
315 true /* tosca_live */,
316 true /* expand_call */);
317 __ leave();
318 }
319 }
320 }
321
322 void ShenandoahBarrierSetAssembler::store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
323 Address dst, Register val, Register tmp1, Register tmp2) {
324 bool on_oop = type == T_OBJECT || type == T_ARRAY;
325 if (!on_oop) {
326 BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2);
327 return;
328 }
329
330 // flatten object address if needed
331 if (dst.index() == noreg && dst.offset() == 0) {
332 if (dst.base() != r3) {
333 __ mov(r3, dst.base());
334 }
335 } else {
336 __ lea(r3, dst);
337 }
338
339 shenandoah_write_barrier_pre(masm,
340 r3 /* obj */,
341 tmp2 /* pre_val */,
342 rthread /* thread */,
343 tmp1 /* tmp */,
344 val != noreg /* tosca_live */,
345 false /* expand_call */);
346
347 if (val == noreg) {
348 BarrierSetAssembler::store_at(masm, decorators, type, Address(r3, 0), noreg, noreg, noreg);
349 } else {
350 storeval_barrier(masm, val, tmp1);
351 // G1 barrier needs uncompressed oop for region cross check.
352 Register new_val = val;
353 if (UseCompressedOops) {
354 new_val = rscratch2;
355 __ mov(new_val, val);
356 }
357 BarrierSetAssembler::store_at(masm, decorators, type, Address(r3, 0), val, noreg, noreg);
358 }
359
360 }
361
362 void ShenandoahBarrierSetAssembler::cmpxchg_oop(MacroAssembler* masm, Register addr, Register expected, Register new_val,
363 bool acquire, bool release, bool weak, bool is_cae,
364 Register tmp, Register result) {
365 Register tmp1 = rscratch1;
366 Register tmp2 = rscratch2;
367 bool is_narrow = UseCompressedOops;
368 Assembler::operand_size size = is_narrow ? Assembler::word : Assembler::xword;
369
370 assert_different_registers(addr, expected, new_val, tmp1, tmp2);
371
372 Label retry, done, fail;
373
374 // CAS, using LL/SC pair.
375 __ bind(retry);
376 __ load_exclusive(tmp1, addr, size, acquire);
377 if (is_narrow) {
378 __ cmpw(tmp1, expected);
379 } else {
380 __ cmp(tmp1, expected);
381 }
382 __ br(Assembler::NE, fail);
383 __ store_exclusive(tmp2, new_val, addr, size, release);
384 if (weak) {
385 __ cmpw(tmp2, 0u); // If the store fails, return NE to our caller
386 } else {
387 __ cbnzw(tmp2, retry);
388 }
389 __ b(done);
390
391 __ bind(fail);
392 // Check if rb(expected)==rb(tmp1)
393 // Shuffle registers so that we have memory value ready for next expected.
394 __ mov(tmp2, expected);
395 __ mov(expected, tmp1);
396 if (is_narrow) {
397 __ decode_heap_oop(tmp1, tmp1);
398 __ decode_heap_oop(tmp2, tmp2);
399 }
400 resolve_forward_pointer(masm, tmp1, tmp);
401 resolve_forward_pointer(masm, tmp2, tmp);
402 __ cmp(tmp1, tmp2);
403 // Retry with expected now being the value we just loaded from addr.
404 __ br(Assembler::EQ, retry);
405 if (is_cae && is_narrow) {
406 // For cmp-and-exchange and narrow oops, we need to restore
407 // the compressed old-value. We moved it to 'expected' a few lines up.
408 __ mov(tmp1, expected);
409 }
410 __ bind(done);
411
412 if (is_cae) {
413 __ mov(result, tmp1);
414 } else {
415 __ cset(result, Assembler::EQ);
416 }
417 }
418
419 #ifdef COMPILER1
420
421 #undef __
422 #define __ ce->masm()->
423
424 void ShenandoahBarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, ShenandoahPreBarrierStub* stub) {
425 ShenandoahBarrierSetC1* bs = (ShenandoahBarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
426 // At this point we know that marking is in progress.
427 // If do_load() is true then we have to emit the
428 // load of the previous value; otherwise it has already
429 // been loaded into _pre_val.
430
431 __ bind(*stub->entry());
432
433 assert(stub->pre_val()->is_register(), "Precondition.");
434
435 Register pre_val_reg = stub->pre_val()->as_register();
436
437 if (stub->do_load()) {
438 ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/, false /*unaligned*/);
439 }
440 __ cbz(pre_val_reg, *stub->continuation());
441 ce->store_parameter(stub->pre_val()->as_register(), 0);
442 __ far_call(RuntimeAddress(bs->pre_barrier_c1_runtime_code_blob()->code_begin()));
443 __ b(*stub->continuation());
444 }
445
446 void ShenandoahBarrierSetAssembler::gen_load_reference_barrier_stub(LIR_Assembler* ce, ShenandoahLoadReferenceBarrierStub* stub) {
447
448 Register obj = stub->obj()->as_register();
449 Register res = stub->result()->as_register();
450
451 Label done;
452
453 __ bind(*stub->entry());
454
455 if (res != obj) {
456 __ mov(res, obj);
457 }
458 // Check for null.
459 if (stub->needs_null_check()) {
460 __ cbz(res, done);
461 }
462
463 load_reference_barrier_not_null(ce->masm(), res, rscratch1);
464
465 __ bind(done);
466 __ b(*stub->continuation());
467 }
468
469 #undef __
470
471 #define __ sasm->
472
473 void ShenandoahBarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) {
474 __ prologue("shenandoah_pre_barrier", false);
475
476 // arg0 : previous value of memory
477
478 BarrierSet* bs = BarrierSet::barrier_set();
479
480 const Register pre_val = r0;
481 const Register thread = rthread;
482 const Register tmp = rscratch1;
483
484 Address queue_index(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_index_offset()));
485 Address buffer(thread, in_bytes(ShenandoahThreadLocalData::satb_mark_queue_buffer_offset()));
486
487 Label done;
488 Label runtime;
489
490 // Is marking still active?
491 Address gc_state(thread, in_bytes(ShenandoahThreadLocalData::gc_state_offset()));
492 __ ldrb(tmp, gc_state);
493 __ mov(rscratch2, ShenandoahHeap::MARKING | ShenandoahHeap::TRAVERSAL);
494 __ tst(tmp, rscratch2);
495 __ br(Assembler::EQ, done);
496
497 // Can we store original value in the thread's buffer?
498 __ ldr(tmp, queue_index);
499 __ cbz(tmp, runtime);
500
501 __ sub(tmp, tmp, wordSize);
502 __ str(tmp, queue_index);
503 __ ldr(rscratch2, buffer);
504 __ add(tmp, tmp, rscratch2);
505 __ load_parameter(0, rscratch2);
506 __ str(rscratch2, Address(tmp, 0));
507 __ b(done);
508
509 __ bind(runtime);
510 __ push_call_clobbered_registers();
511 __ load_parameter(0, pre_val);
512 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahRuntime::write_ref_field_pre_entry), pre_val, thread);
513 __ pop_call_clobbered_registers();
514 __ bind(done);
515
516 __ epilogue();
517 }
518
519 #undef __
520
521 #endif // COMPILER1
522
523 address ShenandoahBarrierSetAssembler::shenandoah_lrb() {
524 assert(_shenandoah_lrb != NULL, "need load reference barrier stub");
525 return _shenandoah_lrb;
526 }
527
528 #define __ cgen->assembler()->
529
530 // Shenandoah load reference barrier.
531 //
532 // Input:
533 // r0: OOP to evacuate. Not null.
534 //
535 // Output:
536 // r0: Pointer to evacuated OOP.
537 //
538 // Trash rscratch1, rscratch2. Preserve everything else.
539 address ShenandoahBarrierSetAssembler::generate_shenandoah_lrb(StubCodeGenerator* cgen) {
540
541 __ align(6);
542 StubCodeMark mark(cgen, "StubRoutines", "shenandoah_lrb");
543 address start = __ pc();
544
545 Label work, done;
546 __ mov(rscratch2, ShenandoahHeap::in_cset_fast_test_addr());
547 __ lsr(rscratch1, r0, ShenandoahHeapRegion::region_size_bytes_shift_jint());
548 __ ldrb(rscratch2, Address(rscratch2, rscratch1));
549 __ tbnz(rscratch2, 0, work);
550 __ ret(lr);
551 __ bind(work);
552
553 __ mov(rscratch2, r0);
554 resolve_forward_pointer_not_null(cgen->assembler(), r0, rscratch1);
555 __ cmp(rscratch2, r0);
556 __ br(Assembler::NE, done);
557
558 __ enter(); // required for proper stackwalking of RuntimeStub frame
559
560 __ push_call_clobbered_registers();
561
562 __ mov(lr, CAST_FROM_FN_PTR(address, ShenandoahRuntime::load_reference_barrier_JRT));
563 __ blrt(lr, 1, 0, MacroAssembler::ret_type_integral);
564 __ mov(rscratch1, r0);
565 __ pop_call_clobbered_registers();
566 __ mov(r0, rscratch1);
567
568 __ leave(); // required for proper stackwalking of RuntimeStub frame
569 __ bind(done);
570 __ ret(lr);
571
572 return start;
573 }
574
575 #undef __
576
577 void ShenandoahBarrierSetAssembler::barrier_stubs_init() {
578 if (ShenandoahLoadRefBarrier) {
579 int stub_code_size = 2048;
580 ResourceMark rm;
581 BufferBlob* bb = BufferBlob::create("shenandoah_barrier_stubs", stub_code_size);
582 CodeBuffer buf(bb);
583 StubCodeGenerator cgen(&buf);
584 _shenandoah_lrb = generate_shenandoah_lrb(&cgen);
585 }
586 }