1 /*
2 * Copyright (c) 2018, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2018, SAP SE. 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 "registerSaver_s390.hpp"
29 #include "gc/g1/g1CardTable.hpp"
30 #include "gc/g1/g1BarrierSet.hpp"
31 #include "gc/g1/g1BarrierSetAssembler.hpp"
32 #include "gc/g1/g1ThreadLocalData.hpp"
33 #include "gc/g1/heapRegion.hpp"
34 #include "interpreter/interp_masm.hpp"
35 #include "runtime/sharedRuntime.hpp"
36 #ifdef COMPILER1
37 #include "c1/c1_LIRAssembler.hpp"
38 #include "c1/c1_MacroAssembler.hpp"
39 #include "gc/g1/c1/g1BarrierSetC1.hpp"
40 #endif
41
42 #define __ masm->
43
44 #define BLOCK_COMMENT(str) if (PrintAssembly) __ block_comment(str)
45
46 void G1BarrierSetAssembler::gen_write_ref_array_pre_barrier(MacroAssembler* masm, DecoratorSet decorators,
47 Register addr, Register count) {
48 bool dest_uninitialized = (decorators & AS_DEST_NOT_INITIALIZED) != 0;
49
50 // With G1, don't generate the call if we statically know that the target is uninitialized.
51 if (!dest_uninitialized) {
52 // Is marking active?
53 Label filtered;
54 assert_different_registers(addr, Z_R0_scratch); // would be destroyed by push_frame()
55 assert_different_registers(count, Z_R0_scratch); // would be destroyed by push_frame()
56 Register Rtmp1 = Z_R0_scratch;
57 const int active_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
58 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
59 __ load_and_test_int(Rtmp1, Address(Z_thread, active_offset));
60 } else {
61 guarantee(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");
62 __ load_and_test_byte(Rtmp1, Address(Z_thread, active_offset));
63 }
64 __ z_bre(filtered); // Activity indicator is zero, so there is no marking going on currently.
65
66 RegisterSaver::save_live_registers(masm, RegisterSaver::arg_registers); // Creates frame.
67
68 if (UseCompressedOops) {
69 __ call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSet::write_ref_array_pre_narrow_oop_entry), addr, count);
70 } else {
71 __ call_VM_leaf(CAST_FROM_FN_PTR(address, G1BarrierSet::write_ref_array_pre_oop_entry), addr, count);
72 }
73
74 RegisterSaver::restore_live_registers(masm, RegisterSaver::arg_registers);
75
76 __ bind(filtered);
77 }
78 }
79
80 void G1BarrierSetAssembler::gen_write_ref_array_post_barrier(MacroAssembler* masm, DecoratorSet decorators,
81 Register addr, Register count, bool do_return) {
82 address entry_point = CAST_FROM_FN_PTR(address, G1BarrierSet::write_ref_array_post_entry);
83 if (!do_return) {
84 assert_different_registers(addr, Z_R0_scratch); // would be destroyed by push_frame()
85 assert_different_registers(count, Z_R0_scratch); // would be destroyed by push_frame()
86 RegisterSaver::save_live_registers(masm, RegisterSaver::arg_registers); // Creates frame.
87 __ call_VM_leaf(entry_point, addr, count);
88 RegisterSaver::restore_live_registers(masm, RegisterSaver::arg_registers);
89 } else {
90 // Tail call: call c and return to stub caller.
91 __ lgr_if_needed(Z_ARG1, addr);
92 __ lgr_if_needed(Z_ARG2, count);
93 __ load_const(Z_R1, entry_point);
94 __ z_br(Z_R1); // Branch without linking, callee will return to stub caller.
95 }
96 }
97
98 void G1BarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
99 const Address& src, Register dst, Register tmp1, Register tmp2, Label *L_handle_null) {
100 bool on_oop = type == T_OBJECT || type == T_ARRAY;
101 bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;
102 bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;
103 bool on_reference = on_weak || on_phantom;
104 Label done;
105 if (on_oop && on_reference && L_handle_null == NULL) { L_handle_null = &done; }
106 ModRefBarrierSetAssembler::load_at(masm, decorators, type, src, dst, tmp1, tmp2, L_handle_null);
107 if (on_oop && on_reference) {
108 // Generate the G1 pre-barrier code to log the value of
109 // the referent field in an SATB buffer.
110 g1_write_barrier_pre(masm, decorators | OOP_NOT_NULL,
111 NULL /* obj */,
112 dst /* pre_val */,
113 noreg/* preserve */ ,
114 tmp1, tmp2 /* tmp */,
115 true /* pre_val_needed */);
116 }
117 __ bind(done);
118 }
119
120 void G1BarrierSetAssembler::g1_write_barrier_pre(MacroAssembler* masm, DecoratorSet decorators,
121 const Address* obj,
122 Register Rpre_val, // Ideally, this is a non-volatile register.
123 Register Rval, // Will be preserved.
124 Register Rtmp1, // If Rpre_val is volatile, either Rtmp1
125 Register Rtmp2, // or Rtmp2 has to be non-volatile.
126 bool pre_val_needed // Save Rpre_val across runtime call, caller uses it.
127 ) {
128
129 bool not_null = (decorators & OOP_NOT_NULL) != 0,
130 preloaded = obj == NULL;
131
132 const Register Robj = obj ? obj->base() : noreg,
133 Roff = obj ? obj->index() : noreg;
134 const int active_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
135 const int buffer_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset());
136 const int index_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_index_offset());
137 assert_different_registers(Rtmp1, Rtmp2, Z_R0_scratch); // None of the Rtmp<i> must be Z_R0!!
138 assert_different_registers(Robj, Z_R0_scratch); // Used for addressing. Furthermore, push_frame destroys Z_R0!!
139 assert_different_registers(Rval, Z_R0_scratch); // push_frame destroys Z_R0!!
140
141 Label callRuntime, filtered;
142
143 BLOCK_COMMENT("g1_write_barrier_pre {");
144
145 // Is marking active?
146 // Note: value is loaded for test purposes only. No further use here.
147 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
148 __ load_and_test_int(Rtmp1, Address(Z_thread, active_offset));
149 } else {
150 guarantee(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");
151 __ load_and_test_byte(Rtmp1, Address(Z_thread, active_offset));
152 }
153 __ z_bre(filtered); // Activity indicator is zero, so there is no marking going on currently.
154
155 assert(Rpre_val != noreg, "must have a real register");
156
157
158 // If an object is given, we need to load the previous value into Rpre_val.
159 if (obj) {
160 // Load the previous value...
161 if (UseCompressedOops) {
162 __ z_llgf(Rpre_val, *obj);
163 } else {
164 __ z_lg(Rpre_val, *obj);
165 }
166 }
167
168 // Is the previous value NULL?
169 // If so, we don't need to record it and we're done.
170 // Note: pre_val is loaded, decompressed and stored (directly or via runtime call).
171 // Register contents is preserved across runtime call if caller requests to do so.
172 if (preloaded && not_null) {
173 #ifdef ASSERT
174 __ z_ltgr(Rpre_val, Rpre_val);
175 __ asm_assert_ne("null oop not allowed (G1 pre)", 0x321); // Checked by caller.
176 #endif
177 } else {
178 __ z_ltgr(Rpre_val, Rpre_val);
179 __ z_bre(filtered); // previous value is NULL, so we don't need to record it.
180 }
181
182 // Decode the oop now. We know it's not NULL.
183 if (Robj != noreg && UseCompressedOops) {
184 __ oop_decoder(Rpre_val, Rpre_val, /*maybeNULL=*/false);
185 }
186
187 // OK, it's not filtered, so we'll need to call enqueue.
188
189 // We can store the original value in the thread's buffer
190 // only if index > 0. Otherwise, we need runtime to handle.
191 // (The index field is typed as size_t.)
192 Register Rbuffer = Rtmp1, Rindex = Rtmp2;
193 assert_different_registers(Rbuffer, Rindex, Rpre_val);
194
195 __ z_lg(Rbuffer, buffer_offset, Z_thread);
196
197 __ load_and_test_long(Rindex, Address(Z_thread, index_offset));
198 __ z_bre(callRuntime); // If index == 0, goto runtime.
199
200 __ add2reg(Rindex, -wordSize); // Decrement index.
201 __ z_stg(Rindex, index_offset, Z_thread);
202
203 // Record the previous value.
204 __ z_stg(Rpre_val, 0, Rbuffer, Rindex);
205 __ z_bru(filtered); // We are done.
206
207 Rbuffer = noreg; // end of life
208 Rindex = noreg; // end of life
209
210 __ bind(callRuntime);
211
212 // Save some registers (inputs and result) over runtime call
213 // by spilling them into the top frame.
214 if (Robj != noreg && Robj->is_volatile()) {
215 __ z_stg(Robj, Robj->encoding()*BytesPerWord, Z_SP);
216 }
217 if (Roff != noreg && Roff->is_volatile()) {
218 __ z_stg(Roff, Roff->encoding()*BytesPerWord, Z_SP);
219 }
220 if (Rval != noreg && Rval->is_volatile()) {
221 __ z_stg(Rval, Rval->encoding()*BytesPerWord, Z_SP);
222 }
223
224 // Save Rpre_val (result) over runtime call.
225 Register Rpre_save = Rpre_val;
226 if ((Rpre_val == Z_R0_scratch) || (pre_val_needed && Rpre_val->is_volatile())) {
227 guarantee(!Rtmp1->is_volatile() || !Rtmp2->is_volatile(), "oops!");
228 Rpre_save = !Rtmp1->is_volatile() ? Rtmp1 : Rtmp2;
229 }
230 __ lgr_if_needed(Rpre_save, Rpre_val);
231
232 // Push frame to protect top frame with return pc and spilled register values.
233 __ save_return_pc();
234 __ push_frame_abi160(0); // Will use Z_R0 as tmp.
235
236 // Rpre_val may be destroyed by push_frame().
237 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), Rpre_save, Z_thread);
238
239 __ pop_frame();
240 __ restore_return_pc();
241
242 // Restore spilled values.
243 if (Robj != noreg && Robj->is_volatile()) {
244 __ z_lg(Robj, Robj->encoding()*BytesPerWord, Z_SP);
245 }
246 if (Roff != noreg && Roff->is_volatile()) {
247 __ z_lg(Roff, Roff->encoding()*BytesPerWord, Z_SP);
248 }
249 if (Rval != noreg && Rval->is_volatile()) {
250 __ z_lg(Rval, Rval->encoding()*BytesPerWord, Z_SP);
251 }
252 if (pre_val_needed && Rpre_val->is_volatile()) {
253 __ lgr_if_needed(Rpre_val, Rpre_save);
254 }
255
256 __ bind(filtered);
257 BLOCK_COMMENT("} g1_write_barrier_pre");
258 }
259
260 void G1BarrierSetAssembler::g1_write_barrier_post(MacroAssembler* masm, DecoratorSet decorators, Register Rstore_addr, Register Rnew_val,
261 Register Rtmp1, Register Rtmp2, Register Rtmp3) {
262 bool not_null = (decorators & OOP_NOT_NULL) != 0;
263
264 assert_different_registers(Rstore_addr, Rnew_val, Rtmp1, Rtmp2); // Most probably, Rnew_val == Rtmp3.
265
266 Label callRuntime, filtered;
267
268 CardTableBarrierSet* ct = barrier_set_cast<CardTableBarrierSet>(BarrierSet::barrier_set());
269 assert(sizeof(*ct->card_table()->byte_map_base()) == sizeof(jbyte), "adjust this code");
270
271 BLOCK_COMMENT("g1_write_barrier_post {");
272
273 // Does store cross heap regions?
274 // It does if the two addresses specify different grain addresses.
275 if (G1RSBarrierRegionFilter) {
276 if (VM_Version::has_DistinctOpnds()) {
277 __ z_xgrk(Rtmp1, Rstore_addr, Rnew_val);
278 } else {
279 __ z_lgr(Rtmp1, Rstore_addr);
280 __ z_xgr(Rtmp1, Rnew_val);
281 }
282 __ z_srag(Rtmp1, Rtmp1, HeapRegion::LogOfHRGrainBytes);
283 __ z_bre(filtered);
284 }
285
286 // Crosses regions, storing NULL?
287 if (not_null) {
288 #ifdef ASSERT
289 __ z_ltgr(Rnew_val, Rnew_val);
290 __ asm_assert_ne("null oop not allowed (G1 post)", 0x322); // Checked by caller.
291 #endif
292 } else {
293 __ z_ltgr(Rnew_val, Rnew_val);
294 __ z_bre(filtered);
295 }
296
297 Rnew_val = noreg; // end of lifetime
298
299 // Storing region crossing non-NULL, is card already dirty?
300 assert(sizeof(*ct->card_table()->byte_map_base()) == sizeof(jbyte), "adjust this code");
301 assert_different_registers(Rtmp1, Rtmp2, Rtmp3);
302 // Make sure not to use Z_R0 for any of these registers.
303 Register Rcard_addr = (Rtmp1 != Z_R0_scratch) ? Rtmp1 : Rtmp3;
304 Register Rbase = (Rtmp2 != Z_R0_scratch) ? Rtmp2 : Rtmp3;
305
306 // calculate address of card
307 __ load_const_optimized(Rbase, (address)ct->card_table()->byte_map_base()); // Card table base.
308 __ z_srlg(Rcard_addr, Rstore_addr, CardTable::card_shift); // Index into card table.
309 __ z_algr(Rcard_addr, Rbase); // Explicit calculation needed for cli.
310 Rbase = noreg; // end of lifetime
311
312 // Filter young.
313 assert((unsigned int)G1CardTable::g1_young_card_val() <= 255, "otherwise check this code");
314 __ z_cli(0, Rcard_addr, G1CardTable::g1_young_card_val());
315 __ z_bre(filtered);
316
317 // Check the card value. If dirty, we're done.
318 // This also avoids false sharing of the (already dirty) card.
319 __ z_sync(); // Required to support concurrent cleaning.
320 assert((unsigned int)G1CardTable::dirty_card_val() <= 255, "otherwise check this code");
321 __ z_cli(0, Rcard_addr, G1CardTable::dirty_card_val()); // Reload after membar.
322 __ z_bre(filtered);
323
324 // Storing a region crossing, non-NULL oop, card is clean.
325 // Dirty card and log.
326 __ z_mvi(0, Rcard_addr, G1CardTable::dirty_card_val());
327
328 Register Rcard_addr_x = Rcard_addr;
329 Register Rqueue_index = (Rtmp2 != Z_R0_scratch) ? Rtmp2 : Rtmp1;
330 Register Rqueue_buf = (Rtmp3 != Z_R0_scratch) ? Rtmp3 : Rtmp1;
331 const int qidx_off = in_bytes(G1ThreadLocalData::dirty_card_queue_index_offset());
332 const int qbuf_off = in_bytes(G1ThreadLocalData::dirty_card_queue_buffer_offset());
333 if ((Rcard_addr == Rqueue_buf) || (Rcard_addr == Rqueue_index)) {
334 Rcard_addr_x = Z_R0_scratch; // Register shortage. We have to use Z_R0.
335 }
336 __ lgr_if_needed(Rcard_addr_x, Rcard_addr);
337
338 __ load_and_test_long(Rqueue_index, Address(Z_thread, qidx_off));
339 __ z_bre(callRuntime); // Index == 0 then jump to runtime.
340
341 __ z_lg(Rqueue_buf, qbuf_off, Z_thread);
342
343 __ add2reg(Rqueue_index, -wordSize); // Decrement index.
344 __ z_stg(Rqueue_index, qidx_off, Z_thread);
345
346 __ z_stg(Rcard_addr_x, 0, Rqueue_index, Rqueue_buf); // Store card.
347 __ z_bru(filtered);
348
349 __ bind(callRuntime);
350
351 // TODO: do we need a frame? Introduced to be on the safe side.
352 bool needs_frame = true;
353 __ lgr_if_needed(Rcard_addr, Rcard_addr_x); // copy back asap. push_frame will destroy Z_R0_scratch!
354
355 // VM call need frame to access(write) O register.
356 if (needs_frame) {
357 __ save_return_pc();
358 __ push_frame_abi160(0); // Will use Z_R0 as tmp on old CPUs.
359 }
360
361 // Save the live input values.
362 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), Rcard_addr, Z_thread);
363
364 if (needs_frame) {
365 __ pop_frame();
366 __ restore_return_pc();
367 }
368
369 __ bind(filtered);
370
371 BLOCK_COMMENT("} g1_write_barrier_post");
372 }
373
374 void G1BarrierSetAssembler::oop_store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,
375 const Address& dst, Register val, Register tmp1, Register tmp2, Register tmp3) {
376 bool on_array = (decorators & IN_HEAP_ARRAY) != 0;
377 bool on_anonymous = (decorators & ON_UNKNOWN_OOP_REF) != 0;
378 bool precise = on_array || on_anonymous;
379 // Load and record the previous value.
380 g1_write_barrier_pre(masm, decorators, &dst, tmp3, val, tmp1, tmp2, false);
381
382 BarrierSetAssembler::store_at(masm, decorators, type, dst, val, tmp1, tmp2, tmp3);
383
384 // No need for post barrier if storing NULL
385 if (val != noreg) {
386 const Register base = dst.base(),
387 idx = dst.index();
388 const intptr_t disp = dst.disp();
389 if (precise && (disp != 0 || idx != noreg)) {
390 __ add2reg_with_index(base, disp, idx, base);
391 }
392 g1_write_barrier_post(masm, decorators, base, val, tmp1, tmp2, tmp3);
393 }
394 }
395
396 void G1BarrierSetAssembler::resolve_jobject(MacroAssembler* masm, Register value, Register tmp1, Register tmp2) {
397 NearLabel Ldone, Lnot_weak;
398 __ z_ltgr(tmp1, value);
399 __ z_bre(Ldone); // Use NULL result as-is.
400
401 __ z_nill(value, ~JNIHandles::weak_tag_mask);
402 __ z_lg(value, 0, value); // Resolve (untagged) jobject.
403
404 __ z_tmll(tmp1, JNIHandles::weak_tag_mask); // Test for jweak tag.
405 __ z_braz(Lnot_weak);
406 __ verify_oop(value);
407 DecoratorSet decorators = IN_ROOT | ON_PHANTOM_OOP_REF;
408 g1_write_barrier_pre(masm, decorators, (const Address*)NULL, value, noreg, tmp1, tmp2, true);
409 __ bind(Lnot_weak);
410 __ verify_oop(value);
411 __ bind(Ldone);
412 }
413
414 #ifdef COMPILER1
415
416 #undef __
417 #define __ ce->masm()->
418
419 void G1BarrierSetAssembler::gen_pre_barrier_stub(LIR_Assembler* ce, G1PreBarrierStub* stub) {
420 G1BarrierSetC1* bs = (G1BarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
421 // At this point we know that marking is in progress.
422 // If do_load() is true then we have to emit the
423 // load of the previous value; otherwise it has already
424 // been loaded into _pre_val.
425 __ bind(*stub->entry());
426 ce->check_reserved_argument_area(16); // RT stub needs 2 spill slots.
427 assert(stub->pre_val()->is_register(), "Precondition.");
428
429 Register pre_val_reg = stub->pre_val()->as_register();
430
431 if (stub->do_load()) {
432 ce->mem2reg(stub->addr(), stub->pre_val(), T_OBJECT, stub->patch_code(), stub->info(), false /*wide*/, false /*unaligned*/);
433 }
434
435 __ z_ltgr(Z_R1_scratch, pre_val_reg); // Pass oop in Z_R1_scratch to Runtime1::g1_pre_barrier_slow_id.
436 __ branch_optimized(Assembler::bcondZero, *stub->continuation());
437 ce->emit_call_c(bs->pre_barrier_c1_runtime_code_blob()->code_begin());
438 __ branch_optimized(Assembler::bcondAlways, *stub->continuation());
439 }
440
441 void G1BarrierSetAssembler::gen_post_barrier_stub(LIR_Assembler* ce, G1PostBarrierStub* stub) {
442 G1BarrierSetC1* bs = (G1BarrierSetC1*)BarrierSet::barrier_set()->barrier_set_c1();
443 __ bind(*stub->entry());
444 ce->check_reserved_argument_area(16); // RT stub needs 2 spill slots.
445 assert(stub->addr()->is_register(), "Precondition.");
446 assert(stub->new_val()->is_register(), "Precondition.");
447 Register new_val_reg = stub->new_val()->as_register();
448 __ z_ltgr(new_val_reg, new_val_reg);
449 __ branch_optimized(Assembler::bcondZero, *stub->continuation());
450 __ z_lgr(Z_R1_scratch, stub->addr()->as_pointer_register());
451 ce->emit_call_c(bs->post_barrier_c1_runtime_code_blob()->code_begin());
452 __ branch_optimized(Assembler::bcondAlways, *stub->continuation());
453 }
454
455 #undef __
456
457 #define __ sasm->
458
459 static OopMap* save_volatile_registers(StubAssembler* sasm, Register return_pc = Z_R14) {
460 __ block_comment("save_volatile_registers");
461 RegisterSaver::RegisterSet reg_set = RegisterSaver::all_volatile_registers;
462 int frame_size_in_slots = RegisterSaver::live_reg_frame_size(reg_set) / VMRegImpl::stack_slot_size;
463 sasm->set_frame_size(frame_size_in_slots / VMRegImpl::slots_per_word);
464 return RegisterSaver::save_live_registers(sasm, reg_set, return_pc);
465 }
466
467 static void restore_volatile_registers(StubAssembler* sasm) {
468 __ block_comment("restore_volatile_registers");
469 RegisterSaver::RegisterSet reg_set = RegisterSaver::all_volatile_registers;
470 RegisterSaver::restore_live_registers(sasm, reg_set);
471 }
472
473 void G1BarrierSetAssembler::generate_c1_pre_barrier_runtime_stub(StubAssembler* sasm) {
474 // Z_R1_scratch: previous value of memory
475
476 BarrierSet* bs = BarrierSet::barrier_set();
477 __ set_info("g1_pre_barrier_slow_id", false);
478
479 Register pre_val = Z_R1_scratch;
480 Register tmp = Z_R6; // Must be non-volatile because it is used to save pre_val.
481 Register tmp2 = Z_R7;
482
483 Label refill, restart, marking_not_active;
484 int satb_q_active_byte_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
485 int satb_q_index_byte_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_index_offset());
486 int satb_q_buf_byte_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_buffer_offset());
487
488 // Save tmp registers (see assertion in G1PreBarrierStub::emit_code()).
489 __ z_stg(tmp, 0*BytesPerWord + FrameMap::first_available_sp_in_frame, Z_SP);
490 __ z_stg(tmp2, 1*BytesPerWord + FrameMap::first_available_sp_in_frame, Z_SP);
491
492 // Is marking still active?
493 if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {
494 __ load_and_test_int(tmp, Address(Z_thread, satb_q_active_byte_offset));
495 } else {
496 assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");
497 __ load_and_test_byte(tmp, Address(Z_thread, satb_q_active_byte_offset));
498 }
499 __ z_bre(marking_not_active); // Activity indicator is zero, so there is no marking going on currently.
500
501 __ bind(restart);
502 // Load the index into the SATB buffer. SATBMarkQueue::_index is a
503 // size_t so ld_ptr is appropriate.
504 __ z_ltg(tmp, satb_q_index_byte_offset, Z_R0, Z_thread);
505
506 // index == 0?
507 __ z_brz(refill);
508
509 __ z_lg(tmp2, satb_q_buf_byte_offset, Z_thread);
510 __ add2reg(tmp, -oopSize);
511
512 __ z_stg(pre_val, 0, tmp, tmp2); // [_buf + index] := <address_of_card>
513 __ z_stg(tmp, satb_q_index_byte_offset, Z_thread);
514
515 __ bind(marking_not_active);
516 // Restore tmp registers (see assertion in G1PreBarrierStub::emit_code()).
517 __ z_lg(tmp, 0*BytesPerWord + FrameMap::first_available_sp_in_frame, Z_SP);
518 __ z_lg(tmp2, 1*BytesPerWord + FrameMap::first_available_sp_in_frame, Z_SP);
519 __ z_br(Z_R14);
520
521 __ bind(refill);
522 save_volatile_registers(sasm);
523 __ z_lgr(tmp, pre_val); // save pre_val
524 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SATBMarkQueueSet::handle_zero_index_for_thread),
525 Z_thread);
526 __ z_lgr(pre_val, tmp); // restore pre_val
527 restore_volatile_registers(sasm);
528 __ z_bru(restart);
529 }
530
531 void G1BarrierSetAssembler::generate_c1_post_barrier_runtime_stub(StubAssembler* sasm) {
532 // Z_R1_scratch: oop address, address of updated memory slot
533
534 BarrierSet* bs = BarrierSet::barrier_set();
535 __ set_info("g1_post_barrier_slow_id", false);
536
537 Register addr_oop = Z_R1_scratch;
538 Register addr_card = Z_R1_scratch;
539 Register r1 = Z_R6; // Must be saved/restored.
540 Register r2 = Z_R7; // Must be saved/restored.
541 Register cardtable = r1; // Must be non-volatile, because it is used to save addr_card.
542 CardTableBarrierSet* ctbs = barrier_set_cast<CardTableBarrierSet>(bs);
543 CardTable* ct = ctbs->card_table();
544 jbyte* byte_map_base = ct->byte_map_base();
545
546 // Save registers used below (see assertion in G1PreBarrierStub::emit_code()).
547 __ z_stg(r1, 0*BytesPerWord + FrameMap::first_available_sp_in_frame, Z_SP);
548
549 Label not_already_dirty, restart, refill, young_card;
550
551 // Calculate address of card corresponding to the updated oop slot.
552 AddressLiteral rs(byte_map_base);
553 __ z_srlg(addr_card, addr_oop, CardTable::card_shift);
554 addr_oop = noreg; // dead now
555 __ load_const_optimized(cardtable, rs); // cardtable := <card table base>
556 __ z_agr(addr_card, cardtable); // addr_card := addr_oop>>card_shift + cardtable
557
558 __ z_cli(0, addr_card, (int)G1CardTable::g1_young_card_val());
559 __ z_bre(young_card);
560
561 __ z_sync(); // Required to support concurrent cleaning.
562
563 __ z_cli(0, addr_card, (int)CardTable::dirty_card_val());
564 __ z_brne(not_already_dirty);
565
566 __ bind(young_card);
567 // We didn't take the branch, so we're already dirty: restore
568 // used registers and return.
569 __ z_lg(r1, 0*BytesPerWord + FrameMap::first_available_sp_in_frame, Z_SP);
570 __ z_br(Z_R14);
571
572 // Not dirty.
573 __ bind(not_already_dirty);
574
575 // First, dirty it: [addr_card] := 0
576 __ z_mvi(0, addr_card, CardTable::dirty_card_val());
577
578 Register idx = cardtable; // Must be non-volatile, because it is used to save addr_card.
579 Register buf = r2;
580 cardtable = noreg; // now dead
581
582 // Save registers used below (see assertion in G1PreBarrierStub::emit_code()).
583 __ z_stg(r2, 1*BytesPerWord + FrameMap::first_available_sp_in_frame, Z_SP);
584
585 ByteSize dirty_card_q_index_byte_offset = G1ThreadLocalData::dirty_card_queue_index_offset();
586 ByteSize dirty_card_q_buf_byte_offset = G1ThreadLocalData::dirty_card_queue_buffer_offset();
587
588 __ bind(restart);
589
590 // Get the index into the update buffer. DirtyCardQueue::_index is
591 // a size_t so z_ltg is appropriate here.
592 __ z_ltg(idx, Address(Z_thread, dirty_card_q_index_byte_offset));
593
594 // index == 0?
595 __ z_brz(refill);
596
597 __ z_lg(buf, Address(Z_thread, dirty_card_q_buf_byte_offset));
598 __ add2reg(idx, -oopSize);
599
600 __ z_stg(addr_card, 0, idx, buf); // [_buf + index] := <address_of_card>
601 __ z_stg(idx, Address(Z_thread, dirty_card_q_index_byte_offset));
602 // Restore killed registers and return.
603 __ z_lg(r1, 0*BytesPerWord + FrameMap::first_available_sp_in_frame, Z_SP);
604 __ z_lg(r2, 1*BytesPerWord + FrameMap::first_available_sp_in_frame, Z_SP);
605 __ z_br(Z_R14);
606
607 __ bind(refill);
608 save_volatile_registers(sasm);
609 __ z_lgr(idx, addr_card); // Save addr_card, tmp3 must be non-volatile.
610 __ call_VM_leaf(CAST_FROM_FN_PTR(address, DirtyCardQueueSet::handle_zero_index_for_thread),
611 Z_thread);
612 __ z_lgr(addr_card, idx);
613 restore_volatile_registers(sasm); // Restore addr_card.
614 __ z_bru(restart);
615 }
616
617 #undef __
618
619 #endif // COMPILER1