1 /*
2 * Copyright (c) 2018, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2015-2018, Azul Systems, Inc. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "precompiled.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "c1/c1_LIRAssembler.hpp"
29 #include "c1/c1_MacroAssembler.hpp"
30 #include "gc/g1/c1/g1BarrierSetC1.hpp"
31 #include "gc/g1/g1BarrierSet.hpp"
32 #include "gc/g1/g1BarrierSetAssembler.hpp"
33 #include "gc/g1/g1BarrierSetRuntime.hpp"
34 #include "gc/g1/g1CardTable.hpp"
35 #include "gc/g1/g1ThreadLocalData.hpp"
36 #include "gc/g1/heapRegion.hpp"
37 #include "gc/shared/collectedHeap.hpp"
38 #include "runtime/sharedRuntime.hpp"
39 #include "runtime/thread.hpp"
40 #include "interpreter/interp_masm.hpp"
41 #include "runtime/sharedRuntime.hpp"
42
43 #define __ masm->
44
45 void G1BarrierSetAssembler::gen_write_ref_array_pre_barrier(MacroAssembler* masm, DecoratorSet decorators,
46 Register addr, Register count) {
47 bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0;
48 // With G1, don't generate the call if we statically know that the target in uninitialized
49 if (!dest_uninitialized) {
50 __ push(RegSet::range(r0, r3), sp);
51 if (count == c_rarg0) {
52 if (addr == c_rarg1) {
53 // exactly backwards!!
54 __ eor(c_rarg0, c_rarg0, c_rarg1);
55 __ eor(c_rarg1, c_rarg0, c_rarg1);
56 __ eor(c_rarg0, c_rarg0, c_rarg1);
57 } else {
58 __ mov(c_rarg1, count);
59 __ mov(c_rarg0, addr);
60 }
61 } else {
62 __ mov(c_rarg0, addr);
63 __ mov(c_rarg1, count);
64 }
65 __ call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_array_pre_oop_entry), 2);
66 __ pop(RegSet::range(r0, r3), sp);
67 }
68 }
69
70 void G1BarrierSetAssembler::gen_write_ref_array_post_barrier(MacroAssembler* masm, DecoratorSet decorators,
71 Register start, Register end, Register scratch) {
72 // must compute element count unless barrier set interface is changed (other platforms supply count)
73 assert_different_registers(start, end, scratch);
74 __ lea(scratch, Address(end, BytesPerHeapOop));
75 __ sub(scratch, scratch, start); // subtract start to get #bytes
76 __ lsr(scratch, scratch, LogBytesPerHeapOop); // convert to element count
77 __ mov(c_rarg0, start);
78 __ mov(c_rarg1, scratch);
79 __ call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_array_post_entry), 2);
80 }
81
82
83 void G1BarrierSetAssembler::g1_write_barrier_pre(MacroAssembler* masm,
84 Address obj,
85 Register pre_val,
86 Register thread,
87 Register tmp,
88 bool tosca_live,
89 bool expand_call) {
90 // If expand_call is true then we expand the call_VM_leaf macro
91 // directly to skip generating the check by
92 // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.
93
94 assert(thread == rthread, "must be");
95
96 Label done;
97 Label runtime;
98
99 assert(pre_val != noreg, "check this code");
100
101 assert_different_registers(pre_val, tmp);
102 if (obj.get_mode() != Address::no_mode)
103 assert(!obj.uses(pre_val) && !obj.uses(tmp), "destroys register");
104
105 Address in_progress(thread, in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()));
106 Address index(thread, in_bytes(G1ThreadLocalData::satb_mark_queue_index_offset()));
107 Address buffer(thread, in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset()));
108
109
110 // Is marking active?
111 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
112 __ ldr(tmp, in_progress);
113 } else {
114 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");
115 __ ldrb(tmp, in_progress);
116 }
117 __ cbz(tmp, done);
118
119 // Do we need to load the previous value?
120 if (obj.get_mode() != Address::no_mode) {
121 __ load_heap_oop(pre_val, obj, noreg, noreg, AS_RAW);
122 }
123
124 // Is the previous value null?
125 __ cbz(pre_val, done);
126
127 // Can we store original value in the thread's buffer?
128 // Is index == 0?
129 // (The index field is typed as size_t.)
130
131 __ ldr(tmp, index); // tmp := *index_adr
132 __ cbz(tmp, runtime); // tmp == 0?
133 // If yes, goto runtime
134
135 __ sub(tmp, tmp, wordSize); // tmp := tmp - wordSize
136 __ str(tmp, index); // *index_adr := tmp
137 __ ldr(rscratch1, buffer);
138 __ add(tmp, tmp, rscratch1); // tmp := tmp + *buffer_adr
139
140 // Record the previous value
141 __ str(pre_val, Address(tmp));
142 __ b(done);
143
144 __ bind(runtime);
145 // save the live input values
146 __ push(r0->bit(tosca_live) | obj.reg_bits() | pre_val->bit(true) | lr->bit(true), sp);
147
148 // Calling the runtime using the regular call_VM_leaf mechanism generates
149 // code (generated by InterpreterMacroAssember::call_VM_leaf_base)
150 // that checks that the *(rfp+frame::interpreter_frame_last_sp) == NULL.
151 //
152 // If we care generating the pre-barrier without a frame (e.g. in the
153 // intrinsified Reference.get() routine) then ebp might be pointing to
154 // the caller frame and so this check will most likely fail at runtime.
155 //
156 // Expanding the call directly bypasses the generation of the check.
157 // So when we do not have have a full interpreter frame on the stack
158 // expand_call should be passed true.
159
160 if (expand_call) {
161 assert(pre_val != c_rarg1, "smashed arg");
162 __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), pre_val, thread);
163 } else {
164 __ call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), pre_val, thread);
165 }
166
167 __ pop(r0->bit(tosca_live) | obj.reg_bits() | pre_val->bit(true) | lr->bit(true), sp);
168
169 __ bind(done);
170 }
171
172 void G1BarrierSetAssembler::g1_write_barrier_post(MacroAssembler* masm,
173 Address store_addr,
174 Register new_val,
175 Register thread,
176 Register tmp,
177 Register tmp2) {
178 assert(thread == rthread, "must be");
179
180 Address queue_index(thread, in_bytes(G1ThreadLocalData::dirty_card_queue_index_offset()));
181 Address buffer(thread, in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset()));
182
183 BarrierSet* bs = BarrierSet::barrier_set();
184 CardTableBarrierSet* ctbs = barrier_set_cast<CardTableBarrierSet>(bs);
185 CardTable* ct = ctbs->card_table();
186 assert(sizeof(*ct->byte_map_base()) == sizeof(jbyte), "adjust this code");
187
188 Label done;
189 Label runtime;
190
191 // Does store cross heap regions?
192
193 __ lea(tmp2, store_addr);
194 __ eor(tmp, tmp2, new_val);
195 __ lsrs(tmp, tmp, HeapRegion::LogOfHRGrainBytes);
196 __ b(done, Assembler::EQ);
197
198 // crosses regions, storing NULL?
199
200 __ cbz(new_val, done);
201
202 // storing region crossing non-NULL, is card already dirty?
203
204 assert(sizeof(*ct->byte_map_base()) == sizeof(jbyte), "adjust this code");
205 const Register card_addr = tmp;
206
207 __ lsr(card_addr, tmp2, CardTable::card_shift);
208
209 //ExternalAddress cardtable((address) ct->byte_map_base());
210 __ mov(tmp2, (unsigned)ct->byte_map_base());
211
212 // get the address of the card
213 __ add(card_addr, card_addr, tmp2);
214 __ ldrb(tmp2, Address(card_addr));
215 __ cmp(tmp2, (int)G1CardTable::g1_young_card_val());
216 __ b(done, Assembler::EQ);
217
218 assert((int)CardTable::dirty_card_val() == 0, "must be 0");
219
220 __ membar(Assembler::StoreLoad);
221
222 __ ldrb(tmp2, Address(card_addr));
223 __ cbz(tmp2, done);
224
225 // storing a region crossing, non-NULL oop, card is clean.
226 // dirty card and log.
227 __ mov(rscratch1, 0);
228 __ strb(rscratch1, Address(card_addr));
229
230 __ ldr(rscratch1, queue_index);
231 __ cbz(rscratch1, runtime);
232 __ sub(rscratch1, rscratch1, wordSize);
233 __ str(rscratch1, queue_index);
234
235 __ ldr(tmp2, buffer);
236 __ str(card_addr, Address(tmp2, rscratch1));
237 __ b(done);
238
239 __ bind(runtime);
240 // save the live input values
241 __ push(store_addr.reg_bits() | new_val->bit(true), sp);
242 __ call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_post_entry), card_addr, thread);
243 __ pop(store_addr.reg_bits() | new_val->bit(true), sp);
244
245 __ bind(done);
246 }
247
248 void G1BarrierSetAssembler::load_word_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
249 Register dst, Address src, Register tmp1, Register tmp_thread) {
250 bool on_oop = type == T_OBJECT || type == T_ARRAY;
251 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
252 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
253 bool on_reference = on_weak || on_phantom;
254 ModRefBarrierSetAssembler::load_word_at(masm, decorators, type, dst, src, tmp1, tmp_thread);
255 if (on_oop && on_reference) {
256 // LR is live. It must be saved around calls.
257 // Generate the G1 pre-barrier code to log the value of
258 // the referent field in an SATB buffer.
259 g1_write_barrier_pre(masm /* masm */,
260 Address() /* obj */,
261 dst /* pre_val */,
262 rthread /* thread */,
263 tmp1 /* tmp */,
264 true /* tosca_live */,
265 true /* expand_call */);
266 }
267 }
268
269 void G1BarrierSetAssembler::load_tos_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
270 Address src, Register tmp1, Register tmp_thread) {
271 bool on_oop = type == T_OBJECT || type == T_ARRAY;
272 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
273 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
274 bool on_reference = on_weak || on_phantom;
275 ModRefBarrierSetAssembler::load_tos_at(masm, decorators, type, src, tmp1, tmp_thread);
276 if (on_oop && on_reference) {
277 // Generate the G1 pre-barrier code to log the value of
278 // the referent field in an SATB buffer.
279 g1_write_barrier_pre(masm /* masm */,
280 Address() /* obj */,
281 r0 /* pre_val */, // atos is in r0
282 rthread /* thread */,
283 tmp1 /* tmp */,
284 true /* tosca_live */,
285 true /* expand_call */);
286 }
287 }
288
289 void G1BarrierSetAssembler::oop_store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
290 Address dst, Register val, Register tmp1, Register tmp2) {
291 g1_write_barrier_pre(masm,
292 dst /* obj */,
293 tmp2 /* pre_val */,
294 rthread /* thread */,
295 tmp1 /* tmp */,
296 val != noreg /* tosca_live */,
297 false /* expand_call */);
298
299 if (val == noreg) {
300 BarrierSetAssembler::store_word_at(masm, decorators, type, dst, noreg, tmp1, noreg);
301 } else {
302 BarrierSetAssembler::store_word_at(masm, decorators, type, dst, val, noreg, noreg);
303 g1_write_barrier_post(masm,
304 dst /* store_adr */,
305 val /* new_val */,
306 rthread /* thread */,
307 tmp1 /* tmp */,
308 tmp2 /* tmp2 */);
309 }
310
311 }
312
313 #ifdef COMPILER1
314
315 #undef __
316 #define __ ce->masm()->
317
318 void G1BarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, G1PreBarrierStub* stub) {
319 G1BarrierSetC1* bs = (G1BarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
320 // At this point we know that marking is in progress.
321 // If do_load() is true then we have to emit the
322 // load of the previous value; otherwise it has already
323 // been loaded into _pre_val.
324
325 __ bind(*stub->entry());
326
327 assert(stub->pre_val()->is_register(), "Precondition.");
328
329 Register pre_val_reg = stub->pre_val()->as_register();
330
331 if (stub->do_load()) {
332 ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/, false /*unaligned*/);
333 }
334 __ cbz(pre_val_reg, *stub->continuation());
335 ce->store_parameter(stub->pre_val()->as_register(), 0);
336 __ far_call(RuntimeAddress(bs->pre_barrier_c1_runtime_code_blob()->code_begin()));
337 __ b(*stub->continuation());
338 }
339
340 void G1BarrierSetAssembler::gen_post_barrier_stub(LIR_Assembler* ce, G1PostBarrierStub* stub) {
341 G1BarrierSetC1* bs = (G1BarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
342 __ bind(*stub->entry());
343 assert(stub->addr()->is_register(), "Precondition.");
344 assert(stub->new_val()->is_register(), "Precondition.");
345 Register new_val_reg = stub->new_val()->as_register();
346 __ cbz(new_val_reg, *stub->continuation());
347 ce->store_parameter(stub->addr()->as_pointer_register(), 0);
348 __ far_call(RuntimeAddress(bs->post_barrier_c1_runtime_code_blob()->code_begin()));
349 __ b(*stub->continuation());
350 }
351
352 #undef __
353
354 #define __ sasm->
355
356 void G1BarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) {
357 __ prologue("g1_pre_barrier", false);
358
359 // arg0 : previous value of memory
360
361 BarrierSet* bs = BarrierSet::barrier_set();
362
363 const Register pre_val = r0;
364 const Register thread = rthread;
365 const Register tmp = rscratch1;
366
367 Address in_progress(thread, in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset()));
368 Address queue_index(thread, in_bytes(G1ThreadLocalData::satb_mark_queue_index_offset()));
369 Address buffer(thread, in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset()));
370
371 Label done;
372 Label runtime;
373
374 // Is marking still active?
375 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
376 __ ldr(tmp, in_progress);
377 } else {
378 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");
379 __ ldrb(tmp, in_progress);
380 }
381 __ cbz(tmp, done);
382
383 // Can we store original value in the thread's buffer?
384 __ ldr(tmp, queue_index);
385 __ cbz(tmp, runtime);
386
387 __ sub(tmp, tmp, wordSize);
388 __ str(tmp, queue_index);
389 __ ldr(rscratch2, buffer);
390 __ add(tmp, tmp, rscratch2);
391 __ load_parameter(0, rscratch2);
392 __ str(rscratch2, Address(tmp, 0));
393 __ b(done);
394
395 __ bind(runtime);
396 __ push_call_clobbered_registers();
397 __ load_parameter(0, pre_val);
398 __ call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_pre_entry), pre_val, thread);
399 __ pop_call_clobbered_registers();
400 __ bind(done);
401
402 __ epilogue();
403 }
404
405 void G1BarrierSetAssembler::generate_c1_post_barrier_runtime_stub(StubAssembler* sasm) {
406 __ prologue("g1_post_barrier", false);
407
408 // arg0: store_address
409 Address store_addr(rfp, 2*BytesPerWord);
410
411 BarrierSet* bs = BarrierSet::barrier_set();
412 CardTableBarrierSet* ctbs = barrier_set_cast<CardTableBarrierSet>(bs);
413 CardTable* ct = ctbs->card_table();
414 assert(sizeof(*ct->byte_map_base()) == sizeof(jbyte), "adjust this code");
415
416 Label done;
417 Label runtime;
418
419 // At this point we know new_value is non-NULL and the new_value crosses regions.
420 // Must check to see if card is already dirty
421
422 const Register thread = rthread;
423
424 Address queue_index(thread, in_bytes(G1ThreadLocalData::dirty_card_queue_index_offset()));
425 Address buffer(thread, in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset()));
426
427 const Register card_addr = rscratch2;
428 ExternalAddress cardtable((address) ct->byte_map_base());
429
430 __ load_parameter(0, card_addr);
431 __ lsr(card_addr, card_addr, CardTable::card_shift);
432 __ mov(rscratch1, cardtable);
433 __ add(card_addr, card_addr, rscratch1);
434 __ ldrb(rscratch1, Address(card_addr));
435 __ cmp(rscratch1, (int)G1CardTable::g1_young_card_val());
436 __ b(done, Assembler::EQ);
437
438 assert((int)CardTable::dirty_card_val() == 0, "must be 0");
439
440 __ membar(Assembler::StoreLoad);
441 __ ldrb(rscratch1, Address(card_addr));
442 __ cbz(rscratch1, done);
443
444 // storing region crossing non-NULL, card is clean.
445 // dirty card and log.
446 __ mov(rscratch1, 0);
447 __ strb(rscratch1, Address(card_addr));
448
449 __ ldr(rscratch1, queue_index);
450 __ cbz(rscratch1, runtime);
451 __ sub(rscratch1, rscratch1, wordSize);
452 __ str(rscratch1, queue_index);
453
454 // Reuse LR to hold buffer_addr
455 const Register buffer_addr = lr;
456
457 __ ldr(buffer_addr, buffer);
458 __ str(card_addr, Address(buffer_addr, rscratch1));
459 __ b(done);
460
461 __ bind(runtime);
462 __ push_call_clobbered_registers();
463 __ call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSetRuntime::write_ref_field_post_entry), card_addr, thread);
464 __ pop_call_clobbered_registers();
465 __ bind(done);
466 __ epilogue();
467 }
468
469 #undef __
470
471 #endif // COMPILER1