1 /*
2 * Copyright (c) 1997, 2020, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2014, 2020, Red Hat 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 <sys/types.h>
27
28 #include "precompiled.hpp"
29 #include "jvm.h"
30 #include "asm/assembler.hpp"
31 #include "asm/assembler.inline.hpp"
32 #include "gc/shared/barrierSet.hpp"
33 #include "gc/shared/cardTable.hpp"
34 #include "gc/shared/barrierSetAssembler.hpp"
35 #include "gc/shared/cardTableBarrierSet.hpp"
36 #include "interpreter/interpreter.hpp"
37 #include "compiler/disassembler.hpp"
38 #include "memory/resourceArea.hpp"
39 #include "memory/universe.hpp"
40 #include "nativeInst_aarch64.hpp"
41 #include "oops/accessDecorators.hpp"
42 #include "oops/compressedOops.inline.hpp"
43 #include "oops/klass.inline.hpp"
44 #include "runtime/biasedLocking.hpp"
45 #include "runtime/icache.hpp"
46 #include "runtime/interfaceSupport.inline.hpp"
47 #include "runtime/jniHandles.inline.hpp"
48 #include "runtime/sharedRuntime.hpp"
49 #include "runtime/thread.hpp"
50 #include "utilities/powerOfTwo.hpp"
51 #ifdef COMPILER1
52 #include "c1/c1_LIRAssembler.hpp"
53 #endif
54 #ifdef COMPILER2
55 #include "oops/oop.hpp"
56 #include "opto/compile.hpp"
57 #include "opto/node.hpp"
58 #include "opto/output.hpp"
59 #endif
60
61 #ifdef PRODUCT
62 #define BLOCK_COMMENT(str) /* nothing */
63 #else
64 #define BLOCK_COMMENT(str) block_comment(str)
65 #endif
66 #define STOP(str) stop(str);
67 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
68
69 // Patch any kind of instruction; there may be several instructions.
70 // Return the total length (in bytes) of the instructions.
71 int MacroAssembler::pd_patch_instruction_size(address branch, address target) {
72 int instructions = 1;
73 assert((uint64_t)target < (1ull << 48), "48-bit overflow in address constant");
74 intptr_t offset = (target - branch) >> 2;
75 unsigned insn = *(unsigned*)branch;
76 if ((Instruction_aarch64::extract(insn, 29, 24) & 0b111011) == 0b011000) {
77 // Load register (literal)
78 Instruction_aarch64::spatch(branch, 23, 5, offset);
79 } else if (Instruction_aarch64::extract(insn, 30, 26) == 0b00101) {
80 // Unconditional branch (immediate)
81 Instruction_aarch64::spatch(branch, 25, 0, offset);
82 } else if (Instruction_aarch64::extract(insn, 31, 25) == 0b0101010) {
83 // Conditional branch (immediate)
84 Instruction_aarch64::spatch(branch, 23, 5, offset);
85 } else if (Instruction_aarch64::extract(insn, 30, 25) == 0b011010) {
86 // Compare & branch (immediate)
87 Instruction_aarch64::spatch(branch, 23, 5, offset);
88 } else if (Instruction_aarch64::extract(insn, 30, 25) == 0b011011) {
89 // Test & branch (immediate)
90 Instruction_aarch64::spatch(branch, 18, 5, offset);
91 } else if (Instruction_aarch64::extract(insn, 28, 24) == 0b10000) {
92 // PC-rel. addressing
93 offset = target-branch;
94 int shift = Instruction_aarch64::extract(insn, 31, 31);
95 if (shift) {
96 uint64_t dest = (uint64_t)target;
97 uint64_t pc_page = (uint64_t)branch >> 12;
98 uint64_t adr_page = (uint64_t)target >> 12;
99 unsigned offset_lo = dest & 0xfff;
100 offset = adr_page - pc_page;
101
102 // We handle 4 types of PC relative addressing
103 // 1 - adrp Rx, target_page
104 // ldr/str Ry, [Rx, #offset_in_page]
105 // 2 - adrp Rx, target_page
106 // add Ry, Rx, #offset_in_page
107 // 3 - adrp Rx, target_page (page aligned reloc, offset == 0)
108 // movk Rx, #imm16<<32
109 // 4 - adrp Rx, target_page (page aligned reloc, offset == 0)
110 // In the first 3 cases we must check that Rx is the same in the adrp and the
111 // subsequent ldr/str, add or movk instruction. Otherwise we could accidentally end
112 // up treating a type 4 relocation as a type 1, 2 or 3 just because it happened
113 // to be followed by a random unrelated ldr/str, add or movk instruction.
114 //
115 unsigned insn2 = ((unsigned*)branch)[1];
116 if (Instruction_aarch64::extract(insn2, 29, 24) == 0b111001 &&
117 Instruction_aarch64::extract(insn, 4, 0) ==
118 Instruction_aarch64::extract(insn2, 9, 5)) {
119 // Load/store register (unsigned immediate)
120 unsigned size = Instruction_aarch64::extract(insn2, 31, 30);
121 Instruction_aarch64::patch(branch + sizeof (unsigned),
122 21, 10, offset_lo >> size);
123 guarantee(((dest >> size) << size) == dest, "misaligned target");
124 instructions = 2;
125 } else if (Instruction_aarch64::extract(insn2, 31, 22) == 0b1001000100 &&
126 Instruction_aarch64::extract(insn, 4, 0) ==
127 Instruction_aarch64::extract(insn2, 4, 0)) {
128 // add (immediate)
129 Instruction_aarch64::patch(branch + sizeof (unsigned),
130 21, 10, offset_lo);
131 instructions = 2;
132 } else if (Instruction_aarch64::extract(insn2, 31, 21) == 0b11110010110 &&
133 Instruction_aarch64::extract(insn, 4, 0) ==
134 Instruction_aarch64::extract(insn2, 4, 0)) {
135 // movk #imm16<<32
136 Instruction_aarch64::patch(branch + 4, 20, 5, (uint64_t)target >> 32);
137 uintptr_t dest = ((uintptr_t)target & 0xffffffffULL) | ((uintptr_t)branch & 0xffff00000000ULL);
138 uintptr_t pc_page = (uintptr_t)branch >> 12;
139 uintptr_t adr_page = (uintptr_t)dest >> 12;
140 offset = adr_page - pc_page;
141 instructions = 2;
142 }
143 }
144 int offset_lo = offset & 3;
145 offset >>= 2;
146 Instruction_aarch64::spatch(branch, 23, 5, offset);
147 Instruction_aarch64::patch(branch, 30, 29, offset_lo);
148 } else if (Instruction_aarch64::extract(insn, 31, 21) == 0b11010010100) {
149 uint64_t dest = (uint64_t)target;
150 // Move wide constant
151 assert(nativeInstruction_at(branch+4)->is_movk(), "wrong insns in patch");
152 assert(nativeInstruction_at(branch+8)->is_movk(), "wrong insns in patch");
153 Instruction_aarch64::patch(branch, 20, 5, dest & 0xffff);
154 Instruction_aarch64::patch(branch+4, 20, 5, (dest >>= 16) & 0xffff);
155 Instruction_aarch64::patch(branch+8, 20, 5, (dest >>= 16) & 0xffff);
156 assert(target_addr_for_insn(branch) == target, "should be");
157 instructions = 3;
158 } else if (Instruction_aarch64::extract(insn, 31, 22) == 0b1011100101 &&
159 Instruction_aarch64::extract(insn, 4, 0) == 0b11111) {
160 // nothing to do
161 assert(target == 0, "did not expect to relocate target for polling page load");
162 } else {
163 ShouldNotReachHere();
164 }
165 return instructions * NativeInstruction::instruction_size;
166 }
167
168 int MacroAssembler::patch_oop(address insn_addr, address o) {
169 int instructions;
170 unsigned insn = *(unsigned*)insn_addr;
171 assert(nativeInstruction_at(insn_addr+4)->is_movk(), "wrong insns in patch");
172
173 // OOPs are either narrow (32 bits) or wide (48 bits). We encode
174 // narrow OOPs by setting the upper 16 bits in the first
175 // instruction.
176 if (Instruction_aarch64::extract(insn, 31, 21) == 0b11010010101) {
177 // Move narrow OOP
178 narrowOop n = CompressedOops::encode((oop)o);
179 Instruction_aarch64::patch(insn_addr, 20, 5, n >> 16);
180 Instruction_aarch64::patch(insn_addr+4, 20, 5, n & 0xffff);
181 instructions = 2;
182 } else {
183 // Move wide OOP
184 assert(nativeInstruction_at(insn_addr+8)->is_movk(), "wrong insns in patch");
185 uintptr_t dest = (uintptr_t)o;
186 Instruction_aarch64::patch(insn_addr, 20, 5, dest & 0xffff);
187 Instruction_aarch64::patch(insn_addr+4, 20, 5, (dest >>= 16) & 0xffff);
188 Instruction_aarch64::patch(insn_addr+8, 20, 5, (dest >>= 16) & 0xffff);
189 instructions = 3;
190 }
191 return instructions * NativeInstruction::instruction_size;
192 }
193
194 int MacroAssembler::patch_narrow_klass(address insn_addr, narrowKlass n) {
195 // Metatdata pointers are either narrow (32 bits) or wide (48 bits).
196 // We encode narrow ones by setting the upper 16 bits in the first
197 // instruction.
198 NativeInstruction *insn = nativeInstruction_at(insn_addr);
199 assert(Instruction_aarch64::extract(insn->encoding(), 31, 21) == 0b11010010101 &&
200 nativeInstruction_at(insn_addr+4)->is_movk(), "wrong insns in patch");
201
202 Instruction_aarch64::patch(insn_addr, 20, 5, n >> 16);
203 Instruction_aarch64::patch(insn_addr+4, 20, 5, n & 0xffff);
204 return 2 * NativeInstruction::instruction_size;
205 }
206
207 address MacroAssembler::target_addr_for_insn(address insn_addr, unsigned insn) {
208 intptr_t offset = 0;
209 if ((Instruction_aarch64::extract(insn, 29, 24) & 0b011011) == 0b00011000) {
210 // Load register (literal)
211 offset = Instruction_aarch64::sextract(insn, 23, 5);
212 return address(((uint64_t)insn_addr + (offset << 2)));
213 } else if (Instruction_aarch64::extract(insn, 30, 26) == 0b00101) {
214 // Unconditional branch (immediate)
215 offset = Instruction_aarch64::sextract(insn, 25, 0);
216 } else if (Instruction_aarch64::extract(insn, 31, 25) == 0b0101010) {
217 // Conditional branch (immediate)
218 offset = Instruction_aarch64::sextract(insn, 23, 5);
219 } else if (Instruction_aarch64::extract(insn, 30, 25) == 0b011010) {
220 // Compare & branch (immediate)
221 offset = Instruction_aarch64::sextract(insn, 23, 5);
222 } else if (Instruction_aarch64::extract(insn, 30, 25) == 0b011011) {
223 // Test & branch (immediate)
224 offset = Instruction_aarch64::sextract(insn, 18, 5);
225 } else if (Instruction_aarch64::extract(insn, 28, 24) == 0b10000) {
226 // PC-rel. addressing
227 offset = Instruction_aarch64::extract(insn, 30, 29);
228 offset |= Instruction_aarch64::sextract(insn, 23, 5) << 2;
229 int shift = Instruction_aarch64::extract(insn, 31, 31) ? 12 : 0;
230 if (shift) {
231 offset <<= shift;
232 uint64_t target_page = ((uint64_t)insn_addr) + offset;
233 target_page &= ((uint64_t)-1) << shift;
234 // Return the target address for the following sequences
235 // 1 - adrp Rx, target_page
236 // ldr/str Ry, [Rx, #offset_in_page]
237 // 2 - adrp Rx, target_page
238 // add Ry, Rx, #offset_in_page
239 // 3 - adrp Rx, target_page (page aligned reloc, offset == 0)
240 // movk Rx, #imm12<<32
241 // 4 - adrp Rx, target_page (page aligned reloc, offset == 0)
242 //
243 // In the first two cases we check that the register is the same and
244 // return the target_page + the offset within the page.
245 // Otherwise we assume it is a page aligned relocation and return
246 // the target page only.
247 //
248 unsigned insn2 = ((unsigned*)insn_addr)[1];
249 if (Instruction_aarch64::extract(insn2, 29, 24) == 0b111001 &&
250 Instruction_aarch64::extract(insn, 4, 0) ==
251 Instruction_aarch64::extract(insn2, 9, 5)) {
252 // Load/store register (unsigned immediate)
253 unsigned int byte_offset = Instruction_aarch64::extract(insn2, 21, 10);
254 unsigned int size = Instruction_aarch64::extract(insn2, 31, 30);
255 return address(target_page + (byte_offset << size));
256 } else if (Instruction_aarch64::extract(insn2, 31, 22) == 0b1001000100 &&
257 Instruction_aarch64::extract(insn, 4, 0) ==
258 Instruction_aarch64::extract(insn2, 4, 0)) {
259 // add (immediate)
260 unsigned int byte_offset = Instruction_aarch64::extract(insn2, 21, 10);
261 return address(target_page + byte_offset);
262 } else {
263 if (Instruction_aarch64::extract(insn2, 31, 21) == 0b11110010110 &&
264 Instruction_aarch64::extract(insn, 4, 0) ==
265 Instruction_aarch64::extract(insn2, 4, 0)) {
266 target_page = (target_page & 0xffffffff) |
267 ((uint64_t)Instruction_aarch64::extract(insn2, 20, 5) << 32);
268 }
269 return (address)target_page;
270 }
271 } else {
272 ShouldNotReachHere();
273 }
274 } else if (Instruction_aarch64::extract(insn, 31, 23) == 0b110100101) {
275 uint32_t *insns = (uint32_t *)insn_addr;
276 // Move wide constant: movz, movk, movk. See movptr().
277 assert(nativeInstruction_at(insns+1)->is_movk(), "wrong insns in patch");
278 assert(nativeInstruction_at(insns+2)->is_movk(), "wrong insns in patch");
279 return address(uint64_t(Instruction_aarch64::extract(insns[0], 20, 5))
280 + (uint64_t(Instruction_aarch64::extract(insns[1], 20, 5)) << 16)
281 + (uint64_t(Instruction_aarch64::extract(insns[2], 20, 5)) << 32));
282 } else if (Instruction_aarch64::extract(insn, 31, 22) == 0b1011100101 &&
283 Instruction_aarch64::extract(insn, 4, 0) == 0b11111) {
284 return 0;
285 } else {
286 ShouldNotReachHere();
287 }
288 return address(((uint64_t)insn_addr + (offset << 2)));
289 }
290
291 void MacroAssembler::safepoint_poll(Label& slow_path) {
292 ldr(rscratch1, Address(rthread, Thread::polling_page_offset()));
293 tbnz(rscratch1, exact_log2(SafepointMechanism::poll_bit()), slow_path);
294 }
295
296 // Just like safepoint_poll, but use an acquiring load for thread-
297 // local polling.
298 //
299 // We need an acquire here to ensure that any subsequent load of the
300 // global SafepointSynchronize::_state flag is ordered after this load
301 // of the local Thread::_polling page. We don't want this poll to
302 // return false (i.e. not safepointing) and a later poll of the global
303 // SafepointSynchronize::_state spuriously to return true.
304 //
305 // This is to avoid a race when we're in a native->Java transition
306 // racing the code which wakes up from a safepoint.
307 //
308 void MacroAssembler::safepoint_poll_acquire(Label& slow_path) {
309 lea(rscratch1, Address(rthread, Thread::polling_page_offset()));
310 ldar(rscratch1, rscratch1);
311 tbnz(rscratch1, exact_log2(SafepointMechanism::poll_bit()), slow_path);
312 }
313
314 void MacroAssembler::reset_last_Java_frame(bool clear_fp) {
315 // we must set sp to zero to clear frame
316 str(zr, Address(rthread, JavaThread::last_Java_sp_offset()));
317
318 // must clear fp, so that compiled frames are not confused; it is
319 // possible that we need it only for debugging
320 if (clear_fp) {
321 str(zr, Address(rthread, JavaThread::last_Java_fp_offset()));
322 }
323
324 // Always clear the pc because it could have been set by make_walkable()
325 str(zr, Address(rthread, JavaThread::last_Java_pc_offset()));
326 }
327
328 // Calls to C land
329 //
330 // When entering C land, the rfp, & resp of the last Java frame have to be recorded
331 // in the (thread-local) JavaThread object. When leaving C land, the last Java fp
332 // has to be reset to 0. This is required to allow proper stack traversal.
333 void MacroAssembler::set_last_Java_frame(Register last_java_sp,
334 Register last_java_fp,
335 Register last_java_pc,
336 Register scratch) {
337
338 if (last_java_pc->is_valid()) {
339 str(last_java_pc, Address(rthread,
340 JavaThread::frame_anchor_offset()
341 + JavaFrameAnchor::last_Java_pc_offset()));
342 }
343
344 // determine last_java_sp register
345 if (last_java_sp == sp) {
346 mov(scratch, sp);
347 last_java_sp = scratch;
348 } else if (!last_java_sp->is_valid()) {
349 last_java_sp = esp;
350 }
351
352 str(last_java_sp, Address(rthread, JavaThread::last_Java_sp_offset()));
353
354 // last_java_fp is optional
355 if (last_java_fp->is_valid()) {
356 str(last_java_fp, Address(rthread, JavaThread::last_Java_fp_offset()));
357 }
358 }
359
360 void MacroAssembler::set_last_Java_frame(Register last_java_sp,
361 Register last_java_fp,
362 address last_java_pc,
363 Register scratch) {
364 assert(last_java_pc != NULL, "must provide a valid PC");
365
366 adr(scratch, last_java_pc);
367 str(scratch, Address(rthread,
368 JavaThread::frame_anchor_offset()
369 + JavaFrameAnchor::last_Java_pc_offset()));
370
371 set_last_Java_frame(last_java_sp, last_java_fp, noreg, scratch);
372 }
373
374 void MacroAssembler::set_last_Java_frame(Register last_java_sp,
375 Register last_java_fp,
376 Label &L,
377 Register scratch) {
378 if (L.is_bound()) {
379 set_last_Java_frame(last_java_sp, last_java_fp, target(L), scratch);
380 } else {
381 InstructionMark im(this);
382 L.add_patch_at(code(), locator());
383 set_last_Java_frame(last_java_sp, last_java_fp, pc() /* Patched later */, scratch);
384 }
385 }
386
387 void MacroAssembler::far_call(Address entry, CodeBuffer *cbuf, Register tmp) {
388 assert(ReservedCodeCacheSize < 4*G, "branch out of range");
389 assert(CodeCache::find_blob(entry.target()) != NULL,
390 "destination of far call not found in code cache");
391 if (far_branches()) {
392 uintptr_t offset;
393 // We can use ADRP here because we know that the total size of
394 // the code cache cannot exceed 2Gb.
395 adrp(tmp, entry, offset);
396 add(tmp, tmp, offset);
397 if (cbuf) cbuf->set_insts_mark();
398 blr(tmp);
399 } else {
400 if (cbuf) cbuf->set_insts_mark();
401 bl(entry);
402 }
403 }
404
405 void MacroAssembler::far_jump(Address entry, CodeBuffer *cbuf, Register tmp) {
406 assert(ReservedCodeCacheSize < 4*G, "branch out of range");
407 assert(CodeCache::find_blob(entry.target()) != NULL,
408 "destination of far call not found in code cache");
409 if (far_branches()) {
410 uintptr_t offset;
411 // We can use ADRP here because we know that the total size of
412 // the code cache cannot exceed 2Gb.
413 adrp(tmp, entry, offset);
414 add(tmp, tmp, offset);
415 if (cbuf) cbuf->set_insts_mark();
416 br(tmp);
417 } else {
418 if (cbuf) cbuf->set_insts_mark();
419 b(entry);
420 }
421 }
422
423 void MacroAssembler::reserved_stack_check() {
424 // testing if reserved zone needs to be enabled
425 Label no_reserved_zone_enabling;
426
427 ldr(rscratch1, Address(rthread, JavaThread::reserved_stack_activation_offset()));
428 cmp(sp, rscratch1);
429 br(Assembler::LO, no_reserved_zone_enabling);
430
431 enter(); // LR and FP are live.
432 lea(rscratch1, CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone));
433 mov(c_rarg0, rthread);
434 blr(rscratch1);
435 leave();
436
437 // We have already removed our own frame.
438 // throw_delayed_StackOverflowError will think that it's been
439 // called by our caller.
440 lea(rscratch1, RuntimeAddress(StubRoutines::throw_delayed_StackOverflowError_entry()));
441 br(rscratch1);
442 should_not_reach_here();
443
444 bind(no_reserved_zone_enabling);
445 }
446
447 int MacroAssembler::biased_locking_enter(Register lock_reg,
448 Register obj_reg,
449 Register swap_reg,
450 Register tmp_reg,
451 bool swap_reg_contains_mark,
452 Label& done,
453 Label* slow_case,
454 BiasedLockingCounters* counters) {
455 assert(UseBiasedLocking, "why call this otherwise?");
456 assert_different_registers(lock_reg, obj_reg, swap_reg);
457
458 if (PrintBiasedLockingStatistics && counters == NULL)
459 counters = BiasedLocking::counters();
460
461 assert_different_registers(lock_reg, obj_reg, swap_reg, tmp_reg, rscratch1, rscratch2, noreg);
462 assert(markWord::age_shift == markWord::lock_bits + markWord::biased_lock_bits, "biased locking makes assumptions about bit layout");
463 Address mark_addr (obj_reg, oopDesc::mark_offset_in_bytes());
464 Address klass_addr (obj_reg, oopDesc::klass_offset_in_bytes());
465 Address saved_mark_addr(lock_reg, 0);
466
467 // Biased locking
468 // See whether the lock is currently biased toward our thread and
469 // whether the epoch is still valid
470 // Note that the runtime guarantees sufficient alignment of JavaThread
471 // pointers to allow age to be placed into low bits
472 // First check to see whether biasing is even enabled for this object
473 Label cas_label;
474 int null_check_offset = -1;
475 if (!swap_reg_contains_mark) {
476 null_check_offset = offset();
477 ldr(swap_reg, mark_addr);
478 }
479 andr(tmp_reg, swap_reg, markWord::biased_lock_mask_in_place);
480 cmp(tmp_reg, (u1)markWord::biased_lock_pattern);
481 br(Assembler::NE, cas_label);
482 // The bias pattern is present in the object's header. Need to check
483 // whether the bias owner and the epoch are both still current.
484 load_prototype_header(tmp_reg, obj_reg);
485 orr(tmp_reg, tmp_reg, rthread);
486 eor(tmp_reg, swap_reg, tmp_reg);
487 andr(tmp_reg, tmp_reg, ~((int) markWord::age_mask_in_place));
488 if (counters != NULL) {
489 Label around;
490 cbnz(tmp_reg, around);
491 atomic_incw(Address((address)counters->biased_lock_entry_count_addr()), tmp_reg, rscratch1, rscratch2);
492 b(done);
493 bind(around);
494 } else {
495 cbz(tmp_reg, done);
496 }
497
498 Label try_revoke_bias;
499 Label try_rebias;
500
501 // At this point we know that the header has the bias pattern and
502 // that we are not the bias owner in the current epoch. We need to
503 // figure out more details about the state of the header in order to
504 // know what operations can be legally performed on the object's
505 // header.
506
507 // If the low three bits in the xor result aren't clear, that means
508 // the prototype header is no longer biased and we have to revoke
509 // the bias on this object.
510 andr(rscratch1, tmp_reg, markWord::biased_lock_mask_in_place);
511 cbnz(rscratch1, try_revoke_bias);
512
513 // Biasing is still enabled for this data type. See whether the
514 // epoch of the current bias is still valid, meaning that the epoch
515 // bits of the mark word are equal to the epoch bits of the
516 // prototype header. (Note that the prototype header's epoch bits
517 // only change at a safepoint.) If not, attempt to rebias the object
518 // toward the current thread. Note that we must be absolutely sure
519 // that the current epoch is invalid in order to do this because
520 // otherwise the manipulations it performs on the mark word are
521 // illegal.
522 andr(rscratch1, tmp_reg, markWord::epoch_mask_in_place);
523 cbnz(rscratch1, try_rebias);
524
525 // The epoch of the current bias is still valid but we know nothing
526 // about the owner; it might be set or it might be clear. Try to
527 // acquire the bias of the object using an atomic operation. If this
528 // fails we will go in to the runtime to revoke the object's bias.
529 // Note that we first construct the presumed unbiased header so we
530 // don't accidentally blow away another thread's valid bias.
531 {
532 Label here;
533 mov(rscratch1, markWord::biased_lock_mask_in_place | markWord::age_mask_in_place | markWord::epoch_mask_in_place);
534 andr(swap_reg, swap_reg, rscratch1);
535 orr(tmp_reg, swap_reg, rthread);
536 cmpxchg_obj_header(swap_reg, tmp_reg, obj_reg, rscratch1, here, slow_case);
537 // If the biasing toward our thread failed, this means that
538 // another thread succeeded in biasing it toward itself and we
539 // need to revoke that bias. The revocation will occur in the
540 // interpreter runtime in the slow case.
541 bind(here);
542 if (counters != NULL) {
543 atomic_incw(Address((address)counters->anonymously_biased_lock_entry_count_addr()),
544 tmp_reg, rscratch1, rscratch2);
545 }
546 }
547 b(done);
548
549 bind(try_rebias);
550 // At this point we know the epoch has expired, meaning that the
551 // current "bias owner", if any, is actually invalid. Under these
552 // circumstances _only_, we are allowed to use the current header's
553 // value as the comparison value when doing the cas to acquire the
554 // bias in the current epoch. In other words, we allow transfer of
555 // the bias from one thread to another directly in this situation.
556 //
557 // FIXME: due to a lack of registers we currently blow away the age
558 // bits in this situation. Should attempt to preserve them.
559 {
560 Label here;
561 load_prototype_header(tmp_reg, obj_reg);
562 orr(tmp_reg, rthread, tmp_reg);
563 cmpxchg_obj_header(swap_reg, tmp_reg, obj_reg, rscratch1, here, slow_case);
564 // If the biasing toward our thread failed, then another thread
565 // succeeded in biasing it toward itself and we need to revoke that
566 // bias. The revocation will occur in the runtime in the slow case.
567 bind(here);
568 if (counters != NULL) {
569 atomic_incw(Address((address)counters->rebiased_lock_entry_count_addr()),
570 tmp_reg, rscratch1, rscratch2);
571 }
572 }
573 b(done);
574
575 bind(try_revoke_bias);
576 // The prototype mark in the klass doesn't have the bias bit set any
577 // more, indicating that objects of this data type are not supposed
578 // to be biased any more. We are going to try to reset the mark of
579 // this object to the prototype value and fall through to the
580 // CAS-based locking scheme. Note that if our CAS fails, it means
581 // that another thread raced us for the privilege of revoking the
582 // bias of this particular object, so it's okay to continue in the
583 // normal locking code.
584 //
585 // FIXME: due to a lack of registers we currently blow away the age
586 // bits in this situation. Should attempt to preserve them.
587 {
588 Label here, nope;
589 load_prototype_header(tmp_reg, obj_reg);
590 cmpxchg_obj_header(swap_reg, tmp_reg, obj_reg, rscratch1, here, &nope);
591 bind(here);
592
593 // Fall through to the normal CAS-based lock, because no matter what
594 // the result of the above CAS, some thread must have succeeded in
595 // removing the bias bit from the object's header.
596 if (counters != NULL) {
597 atomic_incw(Address((address)counters->revoked_lock_entry_count_addr()), tmp_reg,
598 rscratch1, rscratch2);
599 }
600 bind(nope);
601 }
602
603 bind(cas_label);
604
605 return null_check_offset;
606 }
607
608 void MacroAssembler::biased_locking_exit(Register obj_reg, Register temp_reg, Label& done) {
609 assert(UseBiasedLocking, "why call this otherwise?");
610
611 // Check for biased locking unlock case, which is a no-op
612 // Note: we do not have to check the thread ID for two reasons.
613 // First, the interpreter checks for IllegalMonitorStateException at
614 // a higher level. Second, if the bias was revoked while we held the
615 // lock, the object could not be rebiased toward another thread, so
616 // the bias bit would be clear.
617 ldr(temp_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
618 andr(temp_reg, temp_reg, markWord::biased_lock_mask_in_place);
619 cmp(temp_reg, (u1)markWord::biased_lock_pattern);
620 br(Assembler::EQ, done);
621 }
622
623 static void pass_arg0(MacroAssembler* masm, Register arg) {
624 if (c_rarg0 != arg ) {
625 masm->mov(c_rarg0, arg);
626 }
627 }
628
629 static void pass_arg1(MacroAssembler* masm, Register arg) {
630 if (c_rarg1 != arg ) {
631 masm->mov(c_rarg1, arg);
632 }
633 }
634
635 static void pass_arg2(MacroAssembler* masm, Register arg) {
636 if (c_rarg2 != arg ) {
637 masm->mov(c_rarg2, arg);
638 }
639 }
640
641 static void pass_arg3(MacroAssembler* masm, Register arg) {
642 if (c_rarg3 != arg ) {
643 masm->mov(c_rarg3, arg);
644 }
645 }
646
647 void MacroAssembler::call_VM_base(Register oop_result,
648 Register java_thread,
649 Register last_java_sp,
650 address entry_point,
651 int number_of_arguments,
652 bool check_exceptions) {
653 // determine java_thread register
654 if (!java_thread->is_valid()) {
655 java_thread = rthread;
656 }
657
658 // determine last_java_sp register
659 if (!last_java_sp->is_valid()) {
660 last_java_sp = esp;
661 }
662
663 // debugging support
664 assert(number_of_arguments >= 0 , "cannot have negative number of arguments");
665 assert(java_thread == rthread, "unexpected register");
666 #ifdef ASSERT
667 // TraceBytecodes does not use r12 but saves it over the call, so don't verify
668 // if ((UseCompressedOops || UseCompressedClassPointers) && !TraceBytecodes) verify_heapbase("call_VM_base: heap base corrupted?");
669 #endif // ASSERT
670
671 assert(java_thread != oop_result , "cannot use the same register for java_thread & oop_result");
672 assert(java_thread != last_java_sp, "cannot use the same register for java_thread & last_java_sp");
673
674 // push java thread (becomes first argument of C function)
675
676 mov(c_rarg0, java_thread);
677
678 // set last Java frame before call
679 assert(last_java_sp != rfp, "can't use rfp");
680
681 Label l;
682 set_last_Java_frame(last_java_sp, rfp, l, rscratch1);
683
684 // do the call, remove parameters
685 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments, &l);
686
687 // reset last Java frame
688 // Only interpreter should have to clear fp
689 reset_last_Java_frame(true);
690
691 // C++ interp handles this in the interpreter
692 check_and_handle_popframe(java_thread);
693 check_and_handle_earlyret(java_thread);
694
695 if (check_exceptions) {
696 // check for pending exceptions (java_thread is set upon return)
697 ldr(rscratch1, Address(java_thread, in_bytes(Thread::pending_exception_offset())));
698 Label ok;
699 cbz(rscratch1, ok);
700 lea(rscratch1, RuntimeAddress(StubRoutines::forward_exception_entry()));
701 br(rscratch1);
702 bind(ok);
703 }
704
705 // get oop result if there is one and reset the value in the thread
706 if (oop_result->is_valid()) {
707 get_vm_result(oop_result, java_thread);
708 }
709 }
710
711 void MacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) {
712 call_VM_base(oop_result, noreg, noreg, entry_point, number_of_arguments, check_exceptions);
713 }
714
715 // Maybe emit a call via a trampoline. If the code cache is small
716 // trampolines won't be emitted.
717
718 address MacroAssembler::trampoline_call(Address entry, CodeBuffer *cbuf) {
719 assert(JavaThread::current()->is_Compiler_thread(), "just checking");
720 assert(entry.rspec().type() == relocInfo::runtime_call_type
721 || entry.rspec().type() == relocInfo::opt_virtual_call_type
722 || entry.rspec().type() == relocInfo::static_call_type
723 || entry.rspec().type() == relocInfo::virtual_call_type, "wrong reloc type");
724
725 // We need a trampoline if branches are far.
726 if (far_branches()) {
727 bool in_scratch_emit_size = false;
728 #ifdef COMPILER2
729 // We don't want to emit a trampoline if C2 is generating dummy
730 // code during its branch shortening phase.
731 CompileTask* task = ciEnv::current()->task();
732 in_scratch_emit_size =
733 (task != NULL && is_c2_compile(task->comp_level()) &&
734 Compile::current()->output()->in_scratch_emit_size());
735 #endif
736 if (!in_scratch_emit_size) {
737 address stub = emit_trampoline_stub(offset(), entry.target());
738 if (stub == NULL) {
739 return NULL; // CodeCache is full
740 }
741 }
742 }
743
744 if (cbuf) cbuf->set_insts_mark();
745 relocate(entry.rspec());
746 if (!far_branches()) {
747 bl(entry.target());
748 } else {
749 bl(pc());
750 }
751 // just need to return a non-null address
752 return pc();
753 }
754
755
756 // Emit a trampoline stub for a call to a target which is too far away.
757 //
758 // code sequences:
759 //
760 // call-site:
761 // branch-and-link to <destination> or <trampoline stub>
762 //
763 // Related trampoline stub for this call site in the stub section:
764 // load the call target from the constant pool
765 // branch (LR still points to the call site above)
766
767 address MacroAssembler::emit_trampoline_stub(int insts_call_instruction_offset,
768 address dest) {
769 // Max stub size: alignment nop, TrampolineStub.
770 address stub = start_a_stub(NativeInstruction::instruction_size
771 + NativeCallTrampolineStub::instruction_size);
772 if (stub == NULL) {
773 return NULL; // CodeBuffer::expand failed
774 }
775
776 // Create a trampoline stub relocation which relates this trampoline stub
777 // with the call instruction at insts_call_instruction_offset in the
778 // instructions code-section.
779 align(wordSize);
780 relocate(trampoline_stub_Relocation::spec(code()->insts()->start()
781 + insts_call_instruction_offset));
782 const int stub_start_offset = offset();
783
784 // Now, create the trampoline stub's code:
785 // - load the call
786 // - call
787 Label target;
788 ldr(rscratch1, target);
789 br(rscratch1);
790 bind(target);
791 assert(offset() - stub_start_offset == NativeCallTrampolineStub::data_offset,
792 "should be");
793 emit_int64((int64_t)dest);
794
795 const address stub_start_addr = addr_at(stub_start_offset);
796
797 assert(is_NativeCallTrampolineStub_at(stub_start_addr), "doesn't look like a trampoline");
798
799 end_a_stub();
800 return stub_start_addr;
801 }
802
803 void MacroAssembler::emit_static_call_stub() {
804 // CompiledDirectStaticCall::set_to_interpreted knows the
805 // exact layout of this stub.
806
807 isb();
808 mov_metadata(rmethod, (Metadata*)NULL);
809
810 // Jump to the entry point of the i2c stub.
811 movptr(rscratch1, 0);
812 br(rscratch1);
813 }
814
815 void MacroAssembler::c2bool(Register x) {
816 // implements x == 0 ? 0 : 1
817 // note: must only look at least-significant byte of x
818 // since C-style booleans are stored in one byte
819 // only! (was bug)
820 tst(x, 0xff);
821 cset(x, Assembler::NE);
822 }
823
824 address MacroAssembler::ic_call(address entry, jint method_index) {
825 RelocationHolder rh = virtual_call_Relocation::spec(pc(), method_index);
826 // address const_ptr = long_constant((jlong)Universe::non_oop_word());
827 // uintptr_t offset;
828 // ldr_constant(rscratch2, const_ptr);
829 movptr(rscratch2, (uintptr_t)Universe::non_oop_word());
830 return trampoline_call(Address(entry, rh));
831 }
832
833 // Implementation of call_VM versions
834
835 void MacroAssembler::call_VM(Register oop_result,
836 address entry_point,
837 bool check_exceptions) {
838 call_VM_helper(oop_result, entry_point, 0, check_exceptions);
839 }
840
841 void MacroAssembler::call_VM(Register oop_result,
842 address entry_point,
843 Register arg_1,
844 bool check_exceptions) {
845 pass_arg1(this, arg_1);
846 call_VM_helper(oop_result, entry_point, 1, check_exceptions);
847 }
848
849 void MacroAssembler::call_VM(Register oop_result,
850 address entry_point,
851 Register arg_1,
852 Register arg_2,
853 bool check_exceptions) {
854 assert(arg_1 != c_rarg2, "smashed arg");
855 pass_arg2(this, arg_2);
856 pass_arg1(this, arg_1);
857 call_VM_helper(oop_result, entry_point, 2, check_exceptions);
858 }
859
860 void MacroAssembler::call_VM(Register oop_result,
861 address entry_point,
862 Register arg_1,
863 Register arg_2,
864 Register arg_3,
865 bool check_exceptions) {
866 assert(arg_1 != c_rarg3, "smashed arg");
867 assert(arg_2 != c_rarg3, "smashed arg");
868 pass_arg3(this, arg_3);
869
870 assert(arg_1 != c_rarg2, "smashed arg");
871 pass_arg2(this, arg_2);
872
873 pass_arg1(this, arg_1);
874 call_VM_helper(oop_result, entry_point, 3, check_exceptions);
875 }
876
877 void MacroAssembler::call_VM(Register oop_result,
878 Register last_java_sp,
879 address entry_point,
880 int number_of_arguments,
881 bool check_exceptions) {
882 call_VM_base(oop_result, rthread, last_java_sp, entry_point, number_of_arguments, check_exceptions);
883 }
884
885 void MacroAssembler::call_VM(Register oop_result,
886 Register last_java_sp,
887 address entry_point,
888 Register arg_1,
889 bool check_exceptions) {
890 pass_arg1(this, arg_1);
891 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
892 }
893
894 void MacroAssembler::call_VM(Register oop_result,
895 Register last_java_sp,
896 address entry_point,
897 Register arg_1,
898 Register arg_2,
899 bool check_exceptions) {
900
901 assert(arg_1 != c_rarg2, "smashed arg");
902 pass_arg2(this, arg_2);
903 pass_arg1(this, arg_1);
904 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
905 }
906
907 void MacroAssembler::call_VM(Register oop_result,
908 Register last_java_sp,
909 address entry_point,
910 Register arg_1,
911 Register arg_2,
912 Register arg_3,
913 bool check_exceptions) {
914 assert(arg_1 != c_rarg3, "smashed arg");
915 assert(arg_2 != c_rarg3, "smashed arg");
916 pass_arg3(this, arg_3);
917 assert(arg_1 != c_rarg2, "smashed arg");
918 pass_arg2(this, arg_2);
919 pass_arg1(this, arg_1);
920 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
921 }
922
923
924 void MacroAssembler::get_vm_result(Register oop_result, Register java_thread) {
925 ldr(oop_result, Address(java_thread, JavaThread::vm_result_offset()));
926 str(zr, Address(java_thread, JavaThread::vm_result_offset()));
927 verify_oop(oop_result, "broken oop in call_VM_base");
928 }
929
930 void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thread) {
931 ldr(metadata_result, Address(java_thread, JavaThread::vm_result_2_offset()));
932 str(zr, Address(java_thread, JavaThread::vm_result_2_offset()));
933 }
934
935 void MacroAssembler::align(int modulus) {
936 while (offset() % modulus != 0) nop();
937 }
938
939 // these are no-ops overridden by InterpreterMacroAssembler
940
941 void MacroAssembler::check_and_handle_earlyret(Register java_thread) { }
942
943 void MacroAssembler::check_and_handle_popframe(Register java_thread) { }
944
945
946 RegisterOrConstant MacroAssembler::delayed_value_impl(intptr_t* delayed_value_addr,
947 Register tmp,
948 int offset) {
949 intptr_t value = *delayed_value_addr;
950 if (value != 0)
951 return RegisterOrConstant(value + offset);
952
953 // load indirectly to solve generation ordering problem
954 ldr(tmp, ExternalAddress((address) delayed_value_addr));
955
956 if (offset != 0)
957 add(tmp, tmp, offset);
958
959 return RegisterOrConstant(tmp);
960 }
961
962 // Look up the method for a megamorphic invokeinterface call.
963 // The target method is determined by <intf_klass, itable_index>.
964 // The receiver klass is in recv_klass.
965 // On success, the result will be in method_result, and execution falls through.
966 // On failure, execution transfers to the given label.
967 void MacroAssembler::lookup_interface_method(Register recv_klass,
968 Register intf_klass,
969 RegisterOrConstant itable_index,
970 Register method_result,
971 Register scan_temp,
972 Label& L_no_such_interface,
973 bool return_method) {
974 assert_different_registers(recv_klass, intf_klass, scan_temp);
975 assert_different_registers(method_result, intf_klass, scan_temp);
976 assert(recv_klass != method_result || !return_method,
977 "recv_klass can be destroyed when method isn't needed");
978 assert(itable_index.is_constant() || itable_index.as_register() == method_result,
979 "caller must use same register for non-constant itable index as for method");
980
981 // Compute start of first itableOffsetEntry (which is at the end of the vtable)
982 int vtable_base = in_bytes(Klass::vtable_start_offset());
983 int itentry_off = itableMethodEntry::method_offset_in_bytes();
984 int scan_step = itableOffsetEntry::size() * wordSize;
985 int vte_size = vtableEntry::size_in_bytes();
986 assert(vte_size == wordSize, "else adjust times_vte_scale");
987
988 ldrw(scan_temp, Address(recv_klass, Klass::vtable_length_offset()));
989
990 // %%% Could store the aligned, prescaled offset in the klassoop.
991 // lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base));
992 lea(scan_temp, Address(recv_klass, scan_temp, Address::lsl(3)));
993 add(scan_temp, scan_temp, vtable_base);
994
995 if (return_method) {
996 // Adjust recv_klass by scaled itable_index, so we can free itable_index.
997 assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
998 // lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off));
999 lea(recv_klass, Address(recv_klass, itable_index, Address::lsl(3)));
1000 if (itentry_off)
1001 add(recv_klass, recv_klass, itentry_off);
1002 }
1003
1004 // for (scan = klass->itable(); scan->interface() != NULL; scan += scan_step) {
1005 // if (scan->interface() == intf) {
1006 // result = (klass + scan->offset() + itable_index);
1007 // }
1008 // }
1009 Label search, found_method;
1010
1011 for (int peel = 1; peel >= 0; peel--) {
1012 ldr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset_in_bytes()));
1013 cmp(intf_klass, method_result);
1014
1015 if (peel) {
1016 br(Assembler::EQ, found_method);
1017 } else {
1018 br(Assembler::NE, search);
1019 // (invert the test to fall through to found_method...)
1020 }
1021
1022 if (!peel) break;
1023
1024 bind(search);
1025
1026 // Check that the previous entry is non-null. A null entry means that
1027 // the receiver class doesn't implement the interface, and wasn't the
1028 // same as when the caller was compiled.
1029 cbz(method_result, L_no_such_interface);
1030 add(scan_temp, scan_temp, scan_step);
1031 }
1032
1033 bind(found_method);
1034
1035 // Got a hit.
1036 if (return_method) {
1037 ldrw(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset_in_bytes()));
1038 ldr(method_result, Address(recv_klass, scan_temp, Address::uxtw(0)));
1039 }
1040 }
1041
1042 // virtual method calling
1043 void MacroAssembler::lookup_virtual_method(Register recv_klass,
1044 RegisterOrConstant vtable_index,
1045 Register method_result) {
1046 const int base = in_bytes(Klass::vtable_start_offset());
1047 assert(vtableEntry::size() * wordSize == 8,
1048 "adjust the scaling in the code below");
1049 int vtable_offset_in_bytes = base + vtableEntry::method_offset_in_bytes();
1050
1051 if (vtable_index.is_register()) {
1052 lea(method_result, Address(recv_klass,
1053 vtable_index.as_register(),
1054 Address::lsl(LogBytesPerWord)));
1055 ldr(method_result, Address(method_result, vtable_offset_in_bytes));
1056 } else {
1057 vtable_offset_in_bytes += vtable_index.as_constant() * wordSize;
1058 ldr(method_result,
1059 form_address(rscratch1, recv_klass, vtable_offset_in_bytes, 0));
1060 }
1061 }
1062
1063 void MacroAssembler::check_klass_subtype(Register sub_klass,
1064 Register super_klass,
1065 Register temp_reg,
1066 Label& L_success) {
1067 Label L_failure;
1068 check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg, &L_success, &L_failure, NULL);
1069 check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, NULL);
1070 bind(L_failure);
1071 }
1072
1073
1074 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass,
1075 Register super_klass,
1076 Register temp_reg,
1077 Label* L_success,
1078 Label* L_failure,
1079 Label* L_slow_path,
1080 RegisterOrConstant super_check_offset) {
1081 assert_different_registers(sub_klass, super_klass, temp_reg);
1082 bool must_load_sco = (super_check_offset.constant_or_zero() == -1);
1083 if (super_check_offset.is_register()) {
1084 assert_different_registers(sub_klass, super_klass,
1085 super_check_offset.as_register());
1086 } else if (must_load_sco) {
1087 assert(temp_reg != noreg, "supply either a temp or a register offset");
1088 }
1089
1090 Label L_fallthrough;
1091 int label_nulls = 0;
1092 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
1093 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
1094 if (L_slow_path == NULL) { L_slow_path = &L_fallthrough; label_nulls++; }
1095 assert(label_nulls <= 1, "at most one NULL in the batch");
1096
1097 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
1098 int sco_offset = in_bytes(Klass::super_check_offset_offset());
1099 Address super_check_offset_addr(super_klass, sco_offset);
1100
1101 // Hacked jmp, which may only be used just before L_fallthrough.
1102 #define final_jmp(label) \
1103 if (&(label) == &L_fallthrough) { /*do nothing*/ } \
1104 else b(label) /*omit semi*/
1105
1106 // If the pointers are equal, we are done (e.g., String[] elements).
1107 // This self-check enables sharing of secondary supertype arrays among
1108 // non-primary types such as array-of-interface. Otherwise, each such
1109 // type would need its own customized SSA.
1110 // We move this check to the front of the fast path because many
1111 // type checks are in fact trivially successful in this manner,
1112 // so we get a nicely predicted branch right at the start of the check.
1113 cmp(sub_klass, super_klass);
1114 br(Assembler::EQ, *L_success);
1115
1116 // Check the supertype display:
1117 if (must_load_sco) {
1118 ldrw(temp_reg, super_check_offset_addr);
1119 super_check_offset = RegisterOrConstant(temp_reg);
1120 }
1121 Address super_check_addr(sub_klass, super_check_offset);
1122 ldr(rscratch1, super_check_addr);
1123 cmp(super_klass, rscratch1); // load displayed supertype
1124
1125 // This check has worked decisively for primary supers.
1126 // Secondary supers are sought in the super_cache ('super_cache_addr').
1127 // (Secondary supers are interfaces and very deeply nested subtypes.)
1128 // This works in the same check above because of a tricky aliasing
1129 // between the super_cache and the primary super display elements.
1130 // (The 'super_check_addr' can address either, as the case requires.)
1131 // Note that the cache is updated below if it does not help us find
1132 // what we need immediately.
1133 // So if it was a primary super, we can just fail immediately.
1134 // Otherwise, it's the slow path for us (no success at this point).
1135
1136 if (super_check_offset.is_register()) {
1137 br(Assembler::EQ, *L_success);
1138 subs(zr, super_check_offset.as_register(), sc_offset);
1139 if (L_failure == &L_fallthrough) {
1140 br(Assembler::EQ, *L_slow_path);
1141 } else {
1142 br(Assembler::NE, *L_failure);
1143 final_jmp(*L_slow_path);
1144 }
1145 } else if (super_check_offset.as_constant() == sc_offset) {
1146 // Need a slow path; fast failure is impossible.
1147 if (L_slow_path == &L_fallthrough) {
1148 br(Assembler::EQ, *L_success);
1149 } else {
1150 br(Assembler::NE, *L_slow_path);
1151 final_jmp(*L_success);
1152 }
1153 } else {
1154 // No slow path; it's a fast decision.
1155 if (L_failure == &L_fallthrough) {
1156 br(Assembler::EQ, *L_success);
1157 } else {
1158 br(Assembler::NE, *L_failure);
1159 final_jmp(*L_success);
1160 }
1161 }
1162
1163 bind(L_fallthrough);
1164
1165 #undef final_jmp
1166 }
1167
1168 // These two are taken from x86, but they look generally useful
1169
1170 // scans count pointer sized words at [addr] for occurence of value,
1171 // generic
1172 void MacroAssembler::repne_scan(Register addr, Register value, Register count,
1173 Register scratch) {
1174 Label Lloop, Lexit;
1175 cbz(count, Lexit);
1176 bind(Lloop);
1177 ldr(scratch, post(addr, wordSize));
1178 cmp(value, scratch);
1179 br(EQ, Lexit);
1180 sub(count, count, 1);
1181 cbnz(count, Lloop);
1182 bind(Lexit);
1183 }
1184
1185 // scans count 4 byte words at [addr] for occurence of value,
1186 // generic
1187 void MacroAssembler::repne_scanw(Register addr, Register value, Register count,
1188 Register scratch) {
1189 Label Lloop, Lexit;
1190 cbz(count, Lexit);
1191 bind(Lloop);
1192 ldrw(scratch, post(addr, wordSize));
1193 cmpw(value, scratch);
1194 br(EQ, Lexit);
1195 sub(count, count, 1);
1196 cbnz(count, Lloop);
1197 bind(Lexit);
1198 }
1199
1200 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass,
1201 Register super_klass,
1202 Register temp_reg,
1203 Register temp2_reg,
1204 Label* L_success,
1205 Label* L_failure,
1206 bool set_cond_codes) {
1207 assert_different_registers(sub_klass, super_klass, temp_reg);
1208 if (temp2_reg != noreg)
1209 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg, rscratch1);
1210 #define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg)
1211
1212 Label L_fallthrough;
1213 int label_nulls = 0;
1214 if (L_success == NULL) { L_success = &L_fallthrough; label_nulls++; }
1215 if (L_failure == NULL) { L_failure = &L_fallthrough; label_nulls++; }
1216 assert(label_nulls <= 1, "at most one NULL in the batch");
1217
1218 // a couple of useful fields in sub_klass:
1219 int ss_offset = in_bytes(Klass::secondary_supers_offset());
1220 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
1221 Address secondary_supers_addr(sub_klass, ss_offset);
1222 Address super_cache_addr( sub_klass, sc_offset);
1223
1224 BLOCK_COMMENT("check_klass_subtype_slow_path");
1225
1226 // Do a linear scan of the secondary super-klass chain.
1227 // This code is rarely used, so simplicity is a virtue here.
1228 // The repne_scan instruction uses fixed registers, which we must spill.
1229 // Don't worry too much about pre-existing connections with the input regs.
1230
1231 assert(sub_klass != r0, "killed reg"); // killed by mov(r0, super)
1232 assert(sub_klass != r2, "killed reg"); // killed by lea(r2, &pst_counter)
1233
1234 RegSet pushed_registers;
1235 if (!IS_A_TEMP(r2)) pushed_registers += r2;
1236 if (!IS_A_TEMP(r5)) pushed_registers += r5;
1237
1238 if (super_klass != r0 || UseCompressedOops) {
1239 if (!IS_A_TEMP(r0)) pushed_registers += r0;
1240 }
1241
1242 push(pushed_registers, sp);
1243
1244 // Get super_klass value into r0 (even if it was in r5 or r2).
1245 if (super_klass != r0) {
1246 mov(r0, super_klass);
1247 }
1248
1249 #ifndef PRODUCT
1250 mov(rscratch2, (address)&SharedRuntime::_partial_subtype_ctr);
1251 Address pst_counter_addr(rscratch2);
1252 ldr(rscratch1, pst_counter_addr);
1253 add(rscratch1, rscratch1, 1);
1254 str(rscratch1, pst_counter_addr);
1255 #endif //PRODUCT
1256
1257 // We will consult the secondary-super array.
1258 ldr(r5, secondary_supers_addr);
1259 // Load the array length.
1260 ldrw(r2, Address(r5, Array<Klass*>::length_offset_in_bytes()));
1261 // Skip to start of data.
1262 add(r5, r5, Array<Klass*>::base_offset_in_bytes());
1263
1264 cmp(sp, zr); // Clear Z flag; SP is never zero
1265 // Scan R2 words at [R5] for an occurrence of R0.
1266 // Set NZ/Z based on last compare.
1267 repne_scan(r5, r0, r2, rscratch1);
1268
1269 // Unspill the temp. registers:
1270 pop(pushed_registers, sp);
1271
1272 br(Assembler::NE, *L_failure);
1273
1274 // Success. Cache the super we found and proceed in triumph.
1275 str(super_klass, super_cache_addr);
1276
1277 if (L_success != &L_fallthrough) {
1278 b(*L_success);
1279 }
1280
1281 #undef IS_A_TEMP
1282
1283 bind(L_fallthrough);
1284 }
1285
1286 void MacroAssembler::clinit_barrier(Register klass, Register scratch, Label* L_fast_path, Label* L_slow_path) {
1287 assert(L_fast_path != NULL || L_slow_path != NULL, "at least one is required");
1288 assert_different_registers(klass, rthread, scratch);
1289
1290 Label L_fallthrough, L_tmp;
1291 if (L_fast_path == NULL) {
1292 L_fast_path = &L_fallthrough;
1293 } else if (L_slow_path == NULL) {
1294 L_slow_path = &L_fallthrough;
1295 }
1296 // Fast path check: class is fully initialized
1297 ldrb(scratch, Address(klass, InstanceKlass::init_state_offset()));
1298 subs(zr, scratch, InstanceKlass::fully_initialized);
1299 br(Assembler::EQ, *L_fast_path);
1300
1301 // Fast path check: current thread is initializer thread
1302 ldr(scratch, Address(klass, InstanceKlass::init_thread_offset()));
1303 cmp(rthread, scratch);
1304
1305 if (L_slow_path == &L_fallthrough) {
1306 br(Assembler::EQ, *L_fast_path);
1307 bind(*L_slow_path);
1308 } else if (L_fast_path == &L_fallthrough) {
1309 br(Assembler::NE, *L_slow_path);
1310 bind(*L_fast_path);
1311 } else {
1312 Unimplemented();
1313 }
1314 }
1315
1316 void MacroAssembler::verify_oop(Register reg, const char* s) {
1317 if (!VerifyOops) return;
1318
1319 // Pass register number to verify_oop_subroutine
1320 const char* b = NULL;
1321 {
1322 ResourceMark rm;
1323 stringStream ss;
1324 ss.print("verify_oop: %s: %s", reg->name(), s);
1325 b = code_string(ss.as_string());
1326 }
1327 BLOCK_COMMENT("verify_oop {");
1328
1329 stp(r0, rscratch1, Address(pre(sp, -2 * wordSize)));
1330 stp(rscratch2, lr, Address(pre(sp, -2 * wordSize)));
1331
1332 mov(r0, reg);
1333 movptr(rscratch1, (uintptr_t)(address)b);
1334
1335 // call indirectly to solve generation ordering problem
1336 lea(rscratch2, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
1337 ldr(rscratch2, Address(rscratch2));
1338 blr(rscratch2);
1339
1340 ldp(rscratch2, lr, Address(post(sp, 2 * wordSize)));
1341 ldp(r0, rscratch1, Address(post(sp, 2 * wordSize)));
1342
1343 BLOCK_COMMENT("} verify_oop");
1344 }
1345
1346 void MacroAssembler::verify_oop_addr(Address addr, const char* s) {
1347 if (!VerifyOops) return;
1348
1349 const char* b = NULL;
1350 {
1351 ResourceMark rm;
1352 stringStream ss;
1353 ss.print("verify_oop_addr: %s", s);
1354 b = code_string(ss.as_string());
1355 }
1356 BLOCK_COMMENT("verify_oop_addr {");
1357
1358 stp(r0, rscratch1, Address(pre(sp, -2 * wordSize)));
1359 stp(rscratch2, lr, Address(pre(sp, -2 * wordSize)));
1360
1361 // addr may contain sp so we will have to adjust it based on the
1362 // pushes that we just did.
1363 if (addr.uses(sp)) {
1364 lea(r0, addr);
1365 ldr(r0, Address(r0, 4 * wordSize));
1366 } else {
1367 ldr(r0, addr);
1368 }
1369 movptr(rscratch1, (uintptr_t)(address)b);
1370
1371 // call indirectly to solve generation ordering problem
1372 lea(rscratch2, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
1373 ldr(rscratch2, Address(rscratch2));
1374 blr(rscratch2);
1375
1376 ldp(rscratch2, lr, Address(post(sp, 2 * wordSize)));
1377 ldp(r0, rscratch1, Address(post(sp, 2 * wordSize)));
1378
1379 BLOCK_COMMENT("} verify_oop_addr");
1380 }
1381
1382 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot,
1383 int extra_slot_offset) {
1384 // cf. TemplateTable::prepare_invoke(), if (load_receiver).
1385 int stackElementSize = Interpreter::stackElementSize;
1386 int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0);
1387 #ifdef ASSERT
1388 int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1);
1389 assert(offset1 - offset == stackElementSize, "correct arithmetic");
1390 #endif
1391 if (arg_slot.is_constant()) {
1392 return Address(esp, arg_slot.as_constant() * stackElementSize
1393 + offset);
1394 } else {
1395 add(rscratch1, esp, arg_slot.as_register(),
1396 ext::uxtx, exact_log2(stackElementSize));
1397 return Address(rscratch1, offset);
1398 }
1399 }
1400
1401 void MacroAssembler::call_VM_leaf_base(address entry_point,
1402 int number_of_arguments,
1403 Label *retaddr) {
1404 Label E, L;
1405
1406 stp(rscratch1, rmethod, Address(pre(sp, -2 * wordSize)));
1407
1408 mov(rscratch1, entry_point);
1409 blr(rscratch1);
1410 if (retaddr)
1411 bind(*retaddr);
1412
1413 ldp(rscratch1, rmethod, Address(post(sp, 2 * wordSize)));
1414 maybe_isb();
1415 }
1416
1417 void MacroAssembler::call_VM_leaf(address entry_point, int number_of_arguments) {
1418 call_VM_leaf_base(entry_point, number_of_arguments);
1419 }
1420
1421 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0) {
1422 pass_arg0(this, arg_0);
1423 call_VM_leaf_base(entry_point, 1);
1424 }
1425
1426 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
1427 pass_arg0(this, arg_0);
1428 pass_arg1(this, arg_1);
1429 call_VM_leaf_base(entry_point, 2);
1430 }
1431
1432 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0,
1433 Register arg_1, Register arg_2) {
1434 pass_arg0(this, arg_0);
1435 pass_arg1(this, arg_1);
1436 pass_arg2(this, arg_2);
1437 call_VM_leaf_base(entry_point, 3);
1438 }
1439
1440 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) {
1441 pass_arg0(this, arg_0);
1442 MacroAssembler::call_VM_leaf_base(entry_point, 1);
1443 }
1444
1445 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
1446
1447 assert(arg_0 != c_rarg1, "smashed arg");
1448 pass_arg1(this, arg_1);
1449 pass_arg0(this, arg_0);
1450 MacroAssembler::call_VM_leaf_base(entry_point, 2);
1451 }
1452
1453 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) {
1454 assert(arg_0 != c_rarg2, "smashed arg");
1455 assert(arg_1 != c_rarg2, "smashed arg");
1456 pass_arg2(this, arg_2);
1457 assert(arg_0 != c_rarg1, "smashed arg");
1458 pass_arg1(this, arg_1);
1459 pass_arg0(this, arg_0);
1460 MacroAssembler::call_VM_leaf_base(entry_point, 3);
1461 }
1462
1463 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) {
1464 assert(arg_0 != c_rarg3, "smashed arg");
1465 assert(arg_1 != c_rarg3, "smashed arg");
1466 assert(arg_2 != c_rarg3, "smashed arg");
1467 pass_arg3(this, arg_3);
1468 assert(arg_0 != c_rarg2, "smashed arg");
1469 assert(arg_1 != c_rarg2, "smashed arg");
1470 pass_arg2(this, arg_2);
1471 assert(arg_0 != c_rarg1, "smashed arg");
1472 pass_arg1(this, arg_1);
1473 pass_arg0(this, arg_0);
1474 MacroAssembler::call_VM_leaf_base(entry_point, 4);
1475 }
1476
1477 void MacroAssembler::null_check(Register reg, int offset) {
1478 if (needs_explicit_null_check(offset)) {
1479 // provoke OS NULL exception if reg = NULL by
1480 // accessing M[reg] w/o changing any registers
1481 // NOTE: this is plenty to provoke a segv
1482 ldr(zr, Address(reg));
1483 } else {
1484 // nothing to do, (later) access of M[reg + offset]
1485 // will provoke OS NULL exception if reg = NULL
1486 }
1487 }
1488
1489 // MacroAssembler protected routines needed to implement
1490 // public methods
1491
1492 void MacroAssembler::mov(Register r, Address dest) {
1493 code_section()->relocate(pc(), dest.rspec());
1494 uint64_t imm64 = (uint64_t)dest.target();
1495 movptr(r, imm64);
1496 }
1497
1498 // Move a constant pointer into r. In AArch64 mode the virtual
1499 // address space is 48 bits in size, so we only need three
1500 // instructions to create a patchable instruction sequence that can
1501 // reach anywhere.
1502 void MacroAssembler::movptr(Register r, uintptr_t imm64) {
1503 #ifndef PRODUCT
1504 {
1505 char buffer[64];
1506 snprintf(buffer, sizeof(buffer), "0x%" PRIX64, imm64);
1507 block_comment(buffer);
1508 }
1509 #endif
1510 assert(imm64 < (1ull << 48), "48-bit overflow in address constant");
1511 movz(r, imm64 & 0xffff);
1512 imm64 >>= 16;
1513 movk(r, imm64 & 0xffff, 16);
1514 imm64 >>= 16;
1515 movk(r, imm64 & 0xffff, 32);
1516 }
1517
1518 // Macro to mov replicated immediate to vector register.
1519 // Vd will get the following values for different arrangements in T
1520 // imm32 == hex 000000gh T8B: Vd = ghghghghghghghgh
1521 // imm32 == hex 000000gh T16B: Vd = ghghghghghghghghghghghghghghghgh
1522 // imm32 == hex 0000efgh T4H: Vd = efghefghefghefgh
1523 // imm32 == hex 0000efgh T8H: Vd = efghefghefghefghefghefghefghefgh
1524 // imm32 == hex abcdefgh T2S: Vd = abcdefghabcdefgh
1525 // imm32 == hex abcdefgh T4S: Vd = abcdefghabcdefghabcdefghabcdefgh
1526 // T1D/T2D: invalid
1527 void MacroAssembler::mov(FloatRegister Vd, SIMD_Arrangement T, uint32_t imm32) {
1528 assert(T != T1D && T != T2D, "invalid arrangement");
1529 if (T == T8B || T == T16B) {
1530 assert((imm32 & ~0xff) == 0, "extraneous bits in unsigned imm32 (T8B/T16B)");
1531 movi(Vd, T, imm32 & 0xff, 0);
1532 return;
1533 }
1534 uint32_t nimm32 = ~imm32;
1535 if (T == T4H || T == T8H) {
1536 assert((imm32 & ~0xffff) == 0, "extraneous bits in unsigned imm32 (T4H/T8H)");
1537 imm32 &= 0xffff;
1538 nimm32 &= 0xffff;
1539 }
1540 uint32_t x = imm32;
1541 int movi_cnt = 0;
1542 int movn_cnt = 0;
1543 while (x) { if (x & 0xff) movi_cnt++; x >>= 8; }
1544 x = nimm32;
1545 while (x) { if (x & 0xff) movn_cnt++; x >>= 8; }
1546 if (movn_cnt < movi_cnt) imm32 = nimm32;
1547 unsigned lsl = 0;
1548 while (imm32 && (imm32 & 0xff) == 0) { lsl += 8; imm32 >>= 8; }
1549 if (movn_cnt < movi_cnt)
1550 mvni(Vd, T, imm32 & 0xff, lsl);
1551 else
1552 movi(Vd, T, imm32 & 0xff, lsl);
1553 imm32 >>= 8; lsl += 8;
1554 while (imm32) {
1555 while ((imm32 & 0xff) == 0) { lsl += 8; imm32 >>= 8; }
1556 if (movn_cnt < movi_cnt)
1557 bici(Vd, T, imm32 & 0xff, lsl);
1558 else
1559 orri(Vd, T, imm32 & 0xff, lsl);
1560 lsl += 8; imm32 >>= 8;
1561 }
1562 }
1563
1564 void MacroAssembler::mov_immediate64(Register dst, uint64_t imm64)
1565 {
1566 #ifndef PRODUCT
1567 {
1568 char buffer[64];
1569 snprintf(buffer, sizeof(buffer), "0x%" PRIX64, imm64);
1570 block_comment(buffer);
1571 }
1572 #endif
1573 if (operand_valid_for_logical_immediate(false, imm64)) {
1574 orr(dst, zr, imm64);
1575 } else {
1576 // we can use a combination of MOVZ or MOVN with
1577 // MOVK to build up the constant
1578 uint64_t imm_h[4];
1579 int zero_count = 0;
1580 int neg_count = 0;
1581 int i;
1582 for (i = 0; i < 4; i++) {
1583 imm_h[i] = ((imm64 >> (i * 16)) & 0xffffL);
1584 if (imm_h[i] == 0) {
1585 zero_count++;
1586 } else if (imm_h[i] == 0xffffL) {
1587 neg_count++;
1588 }
1589 }
1590 if (zero_count == 4) {
1591 // one MOVZ will do
1592 movz(dst, 0);
1593 } else if (neg_count == 4) {
1594 // one MOVN will do
1595 movn(dst, 0);
1596 } else if (zero_count == 3) {
1597 for (i = 0; i < 4; i++) {
1598 if (imm_h[i] != 0L) {
1599 movz(dst, (uint32_t)imm_h[i], (i << 4));
1600 break;
1601 }
1602 }
1603 } else if (neg_count == 3) {
1604 // one MOVN will do
1605 for (int i = 0; i < 4; i++) {
1606 if (imm_h[i] != 0xffffL) {
1607 movn(dst, (uint32_t)imm_h[i] ^ 0xffffL, (i << 4));
1608 break;
1609 }
1610 }
1611 } else if (zero_count == 2) {
1612 // one MOVZ and one MOVK will do
1613 for (i = 0; i < 3; i++) {
1614 if (imm_h[i] != 0L) {
1615 movz(dst, (uint32_t)imm_h[i], (i << 4));
1616 i++;
1617 break;
1618 }
1619 }
1620 for (;i < 4; i++) {
1621 if (imm_h[i] != 0L) {
1622 movk(dst, (uint32_t)imm_h[i], (i << 4));
1623 }
1624 }
1625 } else if (neg_count == 2) {
1626 // one MOVN and one MOVK will do
1627 for (i = 0; i < 4; i++) {
1628 if (imm_h[i] != 0xffffL) {
1629 movn(dst, (uint32_t)imm_h[i] ^ 0xffffL, (i << 4));
1630 i++;
1631 break;
1632 }
1633 }
1634 for (;i < 4; i++) {
1635 if (imm_h[i] != 0xffffL) {
1636 movk(dst, (uint32_t)imm_h[i], (i << 4));
1637 }
1638 }
1639 } else if (zero_count == 1) {
1640 // one MOVZ and two MOVKs will do
1641 for (i = 0; i < 4; i++) {
1642 if (imm_h[i] != 0L) {
1643 movz(dst, (uint32_t)imm_h[i], (i << 4));
1644 i++;
1645 break;
1646 }
1647 }
1648 for (;i < 4; i++) {
1649 if (imm_h[i] != 0x0L) {
1650 movk(dst, (uint32_t)imm_h[i], (i << 4));
1651 }
1652 }
1653 } else if (neg_count == 1) {
1654 // one MOVN and two MOVKs will do
1655 for (i = 0; i < 4; i++) {
1656 if (imm_h[i] != 0xffffL) {
1657 movn(dst, (uint32_t)imm_h[i] ^ 0xffffL, (i << 4));
1658 i++;
1659 break;
1660 }
1661 }
1662 for (;i < 4; i++) {
1663 if (imm_h[i] != 0xffffL) {
1664 movk(dst, (uint32_t)imm_h[i], (i << 4));
1665 }
1666 }
1667 } else {
1668 // use a MOVZ and 3 MOVKs (makes it easier to debug)
1669 movz(dst, (uint32_t)imm_h[0], 0);
1670 for (i = 1; i < 4; i++) {
1671 movk(dst, (uint32_t)imm_h[i], (i << 4));
1672 }
1673 }
1674 }
1675 }
1676
1677 void MacroAssembler::mov_immediate32(Register dst, uint32_t imm32)
1678 {
1679 #ifndef PRODUCT
1680 {
1681 char buffer[64];
1682 snprintf(buffer, sizeof(buffer), "0x%" PRIX32, imm32);
1683 block_comment(buffer);
1684 }
1685 #endif
1686 if (operand_valid_for_logical_immediate(true, imm32)) {
1687 orrw(dst, zr, imm32);
1688 } else {
1689 // we can use MOVZ, MOVN or two calls to MOVK to build up the
1690 // constant
1691 uint32_t imm_h[2];
1692 imm_h[0] = imm32 & 0xffff;
1693 imm_h[1] = ((imm32 >> 16) & 0xffff);
1694 if (imm_h[0] == 0) {
1695 movzw(dst, imm_h[1], 16);
1696 } else if (imm_h[0] == 0xffff) {
1697 movnw(dst, imm_h[1] ^ 0xffff, 16);
1698 } else if (imm_h[1] == 0) {
1699 movzw(dst, imm_h[0], 0);
1700 } else if (imm_h[1] == 0xffff) {
1701 movnw(dst, imm_h[0] ^ 0xffff, 0);
1702 } else {
1703 // use a MOVZ and MOVK (makes it easier to debug)
1704 movzw(dst, imm_h[0], 0);
1705 movkw(dst, imm_h[1], 16);
1706 }
1707 }
1708 }
1709
1710 // Form an address from base + offset in Rd. Rd may or may
1711 // not actually be used: you must use the Address that is returned.
1712 // It is up to you to ensure that the shift provided matches the size
1713 // of your data.
1714 Address MacroAssembler::form_address(Register Rd, Register base, int64_t byte_offset, int shift) {
1715 if (Address::offset_ok_for_immed(byte_offset, shift))
1716 // It fits; no need for any heroics
1717 return Address(base, byte_offset);
1718
1719 // Don't do anything clever with negative or misaligned offsets
1720 unsigned mask = (1 << shift) - 1;
1721 if (byte_offset < 0 || byte_offset & mask) {
1722 mov(Rd, byte_offset);
1723 add(Rd, base, Rd);
1724 return Address(Rd);
1725 }
1726
1727 // See if we can do this with two 12-bit offsets
1728 {
1729 uint64_t word_offset = byte_offset >> shift;
1730 uint64_t masked_offset = word_offset & 0xfff000;
1731 if (Address::offset_ok_for_immed(word_offset - masked_offset, 0)
1732 && Assembler::operand_valid_for_add_sub_immediate(masked_offset << shift)) {
1733 add(Rd, base, masked_offset << shift);
1734 word_offset -= masked_offset;
1735 return Address(Rd, word_offset << shift);
1736 }
1737 }
1738
1739 // Do it the hard way
1740 mov(Rd, byte_offset);
1741 add(Rd, base, Rd);
1742 return Address(Rd);
1743 }
1744
1745 void MacroAssembler::atomic_incw(Register counter_addr, Register tmp, Register tmp2) {
1746 if (UseLSE) {
1747 mov(tmp, 1);
1748 ldadd(Assembler::word, tmp, zr, counter_addr);
1749 return;
1750 }
1751 Label retry_load;
1752 if ((VM_Version::features() & VM_Version::CPU_STXR_PREFETCH))
1753 prfm(Address(counter_addr), PSTL1STRM);
1754 bind(retry_load);
1755 // flush and load exclusive from the memory location
1756 ldxrw(tmp, counter_addr);
1757 addw(tmp, tmp, 1);
1758 // if we store+flush with no intervening write tmp wil be zero
1759 stxrw(tmp2, tmp, counter_addr);
1760 cbnzw(tmp2, retry_load);
1761 }
1762
1763
1764 int MacroAssembler::corrected_idivl(Register result, Register ra, Register rb,
1765 bool want_remainder, Register scratch)
1766 {
1767 // Full implementation of Java idiv and irem. The function
1768 // returns the (pc) offset of the div instruction - may be needed
1769 // for implicit exceptions.
1770 //
1771 // constraint : ra/rb =/= scratch
1772 // normal case
1773 //
1774 // input : ra: dividend
1775 // rb: divisor
1776 //
1777 // result: either
1778 // quotient (= ra idiv rb)
1779 // remainder (= ra irem rb)
1780
1781 assert(ra != scratch && rb != scratch, "reg cannot be scratch");
1782
1783 int idivl_offset = offset();
1784 if (! want_remainder) {
1785 sdivw(result, ra, rb);
1786 } else {
1787 sdivw(scratch, ra, rb);
1788 Assembler::msubw(result, scratch, rb, ra);
1789 }
1790
1791 return idivl_offset;
1792 }
1793
1794 int MacroAssembler::corrected_idivq(Register result, Register ra, Register rb,
1795 bool want_remainder, Register scratch)
1796 {
1797 // Full implementation of Java ldiv and lrem. The function
1798 // returns the (pc) offset of the div instruction - may be needed
1799 // for implicit exceptions.
1800 //
1801 // constraint : ra/rb =/= scratch
1802 // normal case
1803 //
1804 // input : ra: dividend
1805 // rb: divisor
1806 //
1807 // result: either
1808 // quotient (= ra idiv rb)
1809 // remainder (= ra irem rb)
1810
1811 assert(ra != scratch && rb != scratch, "reg cannot be scratch");
1812
1813 int idivq_offset = offset();
1814 if (! want_remainder) {
1815 sdiv(result, ra, rb);
1816 } else {
1817 sdiv(scratch, ra, rb);
1818 Assembler::msub(result, scratch, rb, ra);
1819 }
1820
1821 return idivq_offset;
1822 }
1823
1824 void MacroAssembler::membar(Membar_mask_bits order_constraint) {
1825 address prev = pc() - NativeMembar::instruction_size;
1826 address last = code()->last_insn();
1827 if (last != NULL && nativeInstruction_at(last)->is_Membar() && prev == last) {
1828 NativeMembar *bar = NativeMembar_at(prev);
1829 // We are merging two memory barrier instructions. On AArch64 we
1830 // can do this simply by ORing them together.
1831 bar->set_kind(bar->get_kind() | order_constraint);
1832 BLOCK_COMMENT("merged membar");
1833 } else {
1834 code()->set_last_insn(pc());
1835 dmb(Assembler::barrier(order_constraint));
1836 }
1837 }
1838
1839 bool MacroAssembler::try_merge_ldst(Register rt, const Address &adr, size_t size_in_bytes, bool is_store) {
1840 if (ldst_can_merge(rt, adr, size_in_bytes, is_store)) {
1841 merge_ldst(rt, adr, size_in_bytes, is_store);
1842 code()->clear_last_insn();
1843 return true;
1844 } else {
1845 assert(size_in_bytes == 8 || size_in_bytes == 4, "only 8 bytes or 4 bytes load/store is supported.");
1846 const unsigned mask = size_in_bytes - 1;
1847 if (adr.getMode() == Address::base_plus_offset &&
1848 (adr.offset() & mask) == 0) { // only supports base_plus_offset.
1849 code()->set_last_insn(pc());
1850 }
1851 return false;
1852 }
1853 }
1854
1855 void MacroAssembler::ldr(Register Rx, const Address &adr) {
1856 // We always try to merge two adjacent loads into one ldp.
1857 if (!try_merge_ldst(Rx, adr, 8, false)) {
1858 Assembler::ldr(Rx, adr);
1859 }
1860 }
1861
1862 void MacroAssembler::ldrw(Register Rw, const Address &adr) {
1863 // We always try to merge two adjacent loads into one ldp.
1864 if (!try_merge_ldst(Rw, adr, 4, false)) {
1865 Assembler::ldrw(Rw, adr);
1866 }
1867 }
1868
1869 void MacroAssembler::str(Register Rx, const Address &adr) {
1870 // We always try to merge two adjacent stores into one stp.
1871 if (!try_merge_ldst(Rx, adr, 8, true)) {
1872 Assembler::str(Rx, adr);
1873 }
1874 }
1875
1876 void MacroAssembler::strw(Register Rw, const Address &adr) {
1877 // We always try to merge two adjacent stores into one stp.
1878 if (!try_merge_ldst(Rw, adr, 4, true)) {
1879 Assembler::strw(Rw, adr);
1880 }
1881 }
1882
1883 // MacroAssembler routines found actually to be needed
1884
1885 void MacroAssembler::push(Register src)
1886 {
1887 str(src, Address(pre(esp, -1 * wordSize)));
1888 }
1889
1890 void MacroAssembler::pop(Register dst)
1891 {
1892 ldr(dst, Address(post(esp, 1 * wordSize)));
1893 }
1894
1895 // Note: load_unsigned_short used to be called load_unsigned_word.
1896 int MacroAssembler::load_unsigned_short(Register dst, Address src) {
1897 int off = offset();
1898 ldrh(dst, src);
1899 return off;
1900 }
1901
1902 int MacroAssembler::load_unsigned_byte(Register dst, Address src) {
1903 int off = offset();
1904 ldrb(dst, src);
1905 return off;
1906 }
1907
1908 int MacroAssembler::load_signed_short(Register dst, Address src) {
1909 int off = offset();
1910 ldrsh(dst, src);
1911 return off;
1912 }
1913
1914 int MacroAssembler::load_signed_byte(Register dst, Address src) {
1915 int off = offset();
1916 ldrsb(dst, src);
1917 return off;
1918 }
1919
1920 int MacroAssembler::load_signed_short32(Register dst, Address src) {
1921 int off = offset();
1922 ldrshw(dst, src);
1923 return off;
1924 }
1925
1926 int MacroAssembler::load_signed_byte32(Register dst, Address src) {
1927 int off = offset();
1928 ldrsbw(dst, src);
1929 return off;
1930 }
1931
1932 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed, Register dst2) {
1933 switch (size_in_bytes) {
1934 case 8: ldr(dst, src); break;
1935 case 4: ldrw(dst, src); break;
1936 case 2: is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break;
1937 case 1: is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break;
1938 default: ShouldNotReachHere();
1939 }
1940 }
1941
1942 void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes, Register src2) {
1943 switch (size_in_bytes) {
1944 case 8: str(src, dst); break;
1945 case 4: strw(src, dst); break;
1946 case 2: strh(src, dst); break;
1947 case 1: strb(src, dst); break;
1948 default: ShouldNotReachHere();
1949 }
1950 }
1951
1952 void MacroAssembler::decrementw(Register reg, int value)
1953 {
1954 if (value < 0) { incrementw(reg, -value); return; }
1955 if (value == 0) { return; }
1956 if (value < (1 << 12)) { subw(reg, reg, value); return; }
1957 /* else */ {
1958 guarantee(reg != rscratch2, "invalid dst for register decrement");
1959 movw(rscratch2, (unsigned)value);
1960 subw(reg, reg, rscratch2);
1961 }
1962 }
1963
1964 void MacroAssembler::decrement(Register reg, int value)
1965 {
1966 if (value < 0) { increment(reg, -value); return; }
1967 if (value == 0) { return; }
1968 if (value < (1 << 12)) { sub(reg, reg, value); return; }
1969 /* else */ {
1970 assert(reg != rscratch2, "invalid dst for register decrement");
1971 mov(rscratch2, (uint64_t)value);
1972 sub(reg, reg, rscratch2);
1973 }
1974 }
1975
1976 void MacroAssembler::decrementw(Address dst, int value)
1977 {
1978 assert(!dst.uses(rscratch1), "invalid dst for address decrement");
1979 if (dst.getMode() == Address::literal) {
1980 assert(abs(value) < (1 << 12), "invalid value and address mode combination");
1981 lea(rscratch2, dst);
1982 dst = Address(rscratch2);
1983 }
1984 ldrw(rscratch1, dst);
1985 decrementw(rscratch1, value);
1986 strw(rscratch1, dst);
1987 }
1988
1989 void MacroAssembler::decrement(Address dst, int value)
1990 {
1991 assert(!dst.uses(rscratch1), "invalid address for decrement");
1992 if (dst.getMode() == Address::literal) {
1993 assert(abs(value) < (1 << 12), "invalid value and address mode combination");
1994 lea(rscratch2, dst);
1995 dst = Address(rscratch2);
1996 }
1997 ldr(rscratch1, dst);
1998 decrement(rscratch1, value);
1999 str(rscratch1, dst);
2000 }
2001
2002 void MacroAssembler::incrementw(Register reg, int value)
2003 {
2004 if (value < 0) { decrementw(reg, -value); return; }
2005 if (value == 0) { return; }
2006 if (value < (1 << 12)) { addw(reg, reg, value); return; }
2007 /* else */ {
2008 assert(reg != rscratch2, "invalid dst for register increment");
2009 movw(rscratch2, (unsigned)value);
2010 addw(reg, reg, rscratch2);
2011 }
2012 }
2013
2014 void MacroAssembler::increment(Register reg, int value)
2015 {
2016 if (value < 0) { decrement(reg, -value); return; }
2017 if (value == 0) { return; }
2018 if (value < (1 << 12)) { add(reg, reg, value); return; }
2019 /* else */ {
2020 assert(reg != rscratch2, "invalid dst for register increment");
2021 movw(rscratch2, (unsigned)value);
2022 add(reg, reg, rscratch2);
2023 }
2024 }
2025
2026 void MacroAssembler::incrementw(Address dst, int value)
2027 {
2028 assert(!dst.uses(rscratch1), "invalid dst for address increment");
2029 if (dst.getMode() == Address::literal) {
2030 assert(abs(value) < (1 << 12), "invalid value and address mode combination");
2031 lea(rscratch2, dst);
2032 dst = Address(rscratch2);
2033 }
2034 ldrw(rscratch1, dst);
2035 incrementw(rscratch1, value);
2036 strw(rscratch1, dst);
2037 }
2038
2039 void MacroAssembler::increment(Address dst, int value)
2040 {
2041 assert(!dst.uses(rscratch1), "invalid dst for address increment");
2042 if (dst.getMode() == Address::literal) {
2043 assert(abs(value) < (1 << 12), "invalid value and address mode combination");
2044 lea(rscratch2, dst);
2045 dst = Address(rscratch2);
2046 }
2047 ldr(rscratch1, dst);
2048 increment(rscratch1, value);
2049 str(rscratch1, dst);
2050 }
2051
2052
2053 void MacroAssembler::pusha() {
2054 push(0x7fffffff, sp);
2055 }
2056
2057 void MacroAssembler::popa() {
2058 pop(0x7fffffff, sp);
2059 }
2060
2061 // Push lots of registers in the bit set supplied. Don't push sp.
2062 // Return the number of words pushed
2063 int MacroAssembler::push(unsigned int bitset, Register stack) {
2064 int words_pushed = 0;
2065
2066 // Scan bitset to accumulate register pairs
2067 unsigned char regs[32];
2068 int count = 0;
2069 for (int reg = 0; reg <= 30; reg++) {
2070 if (1 & bitset)
2071 regs[count++] = reg;
2072 bitset >>= 1;
2073 }
2074 regs[count++] = zr->encoding_nocheck();
2075 count &= ~1; // Only push an even nuber of regs
2076
2077 if (count) {
2078 stp(as_Register(regs[0]), as_Register(regs[1]),
2079 Address(pre(stack, -count * wordSize)));
2080 words_pushed += 2;
2081 }
2082 for (int i = 2; i < count; i += 2) {
2083 stp(as_Register(regs[i]), as_Register(regs[i+1]),
2084 Address(stack, i * wordSize));
2085 words_pushed += 2;
2086 }
2087
2088 assert(words_pushed == count, "oops, pushed != count");
2089
2090 return count;
2091 }
2092
2093 int MacroAssembler::pop(unsigned int bitset, Register stack) {
2094 int words_pushed = 0;
2095
2096 // Scan bitset to accumulate register pairs
2097 unsigned char regs[32];
2098 int count = 0;
2099 for (int reg = 0; reg <= 30; reg++) {
2100 if (1 & bitset)
2101 regs[count++] = reg;
2102 bitset >>= 1;
2103 }
2104 regs[count++] = zr->encoding_nocheck();
2105 count &= ~1;
2106
2107 for (int i = 2; i < count; i += 2) {
2108 ldp(as_Register(regs[i]), as_Register(regs[i+1]),
2109 Address(stack, i * wordSize));
2110 words_pushed += 2;
2111 }
2112 if (count) {
2113 ldp(as_Register(regs[0]), as_Register(regs[1]),
2114 Address(post(stack, count * wordSize)));
2115 words_pushed += 2;
2116 }
2117
2118 assert(words_pushed == count, "oops, pushed != count");
2119
2120 return count;
2121 }
2122
2123 // Push lots of registers in the bit set supplied. Don't push sp.
2124 // Return the number of words pushed
2125 int MacroAssembler::push_fp(unsigned int bitset, Register stack) {
2126 int words_pushed = 0;
2127
2128 // Scan bitset to accumulate register pairs
2129 unsigned char regs[32];
2130 int count = 0;
2131 for (int reg = 0; reg <= 31; reg++) {
2132 if (1 & bitset)
2133 regs[count++] = reg;
2134 bitset >>= 1;
2135 }
2136
2137 if (count == 0) {
2138 return 0;
2139 }
2140
2141 if (count == 1) {
2142 strq(as_FloatRegister(regs[0]), Address(pre(stack, -wordSize * 2)));
2143 return 1;
2144 }
2145
2146 bool odd = (count & 1) == 1;
2147 int push_slots = count + (odd ? 1 : 0);
2148
2149 // Always pushing full 128 bit registers.
2150 stpq(as_FloatRegister(regs[0]), as_FloatRegister(regs[1]), Address(pre(stack, -push_slots * wordSize * 2)));
2151 words_pushed += 2;
2152
2153 for (int i = 2; i + 1 < count; i += 2) {
2154 stpq(as_FloatRegister(regs[i]), as_FloatRegister(regs[i+1]), Address(stack, i * wordSize * 2));
2155 words_pushed += 2;
2156 }
2157
2158 if (odd) {
2159 strq(as_FloatRegister(regs[count - 1]), Address(stack, (count - 1) * wordSize * 2));
2160 words_pushed++;
2161 }
2162
2163 assert(words_pushed == count, "oops, pushed(%d) != count(%d)", words_pushed, count);
2164 return count;
2165 }
2166
2167 int MacroAssembler::pop_fp(unsigned int bitset, Register stack) {
2168 int words_pushed = 0;
2169
2170 // Scan bitset to accumulate register pairs
2171 unsigned char regs[32];
2172 int count = 0;
2173 for (int reg = 0; reg <= 31; reg++) {
2174 if (1 & bitset)
2175 regs[count++] = reg;
2176 bitset >>= 1;
2177 }
2178
2179 if (count == 0) {
2180 return 0;
2181 }
2182
2183 if (count == 1) {
2184 ldrq(as_FloatRegister(regs[0]), Address(post(stack, wordSize * 2)));
2185 return 1;
2186 }
2187
2188 bool odd = (count & 1) == 1;
2189 int push_slots = count + (odd ? 1 : 0);
2190
2191 if (odd) {
2192 ldrq(as_FloatRegister(regs[count - 1]), Address(stack, (count - 1) * wordSize * 2));
2193 words_pushed++;
2194 }
2195
2196 for (int i = 2; i + 1 < count; i += 2) {
2197 ldpq(as_FloatRegister(regs[i]), as_FloatRegister(regs[i+1]), Address(stack, i * wordSize * 2));
2198 words_pushed += 2;
2199 }
2200
2201 ldpq(as_FloatRegister(regs[0]), as_FloatRegister(regs[1]), Address(post(stack, push_slots * wordSize * 2)));
2202 words_pushed += 2;
2203
2204 assert(words_pushed == count, "oops, pushed(%d) != count(%d)", words_pushed, count);
2205
2206 return count;
2207 }
2208
2209 #ifdef ASSERT
2210 void MacroAssembler::verify_heapbase(const char* msg) {
2211 #if 0
2212 assert (UseCompressedOops || UseCompressedClassPointers, "should be compressed");
2213 assert (Universe::heap() != NULL, "java heap should be initialized");
2214 if (!UseCompressedOops || Universe::ptr_base() == NULL) {
2215 // rheapbase is allocated as general register
2216 return;
2217 }
2218 if (CheckCompressedOops) {
2219 Label ok;
2220 push(1 << rscratch1->encoding(), sp); // cmpptr trashes rscratch1
2221 cmpptr(rheapbase, ExternalAddress((address)CompressedOops::ptrs_base_addr()));
2222 br(Assembler::EQ, ok);
2223 stop(msg);
2224 bind(ok);
2225 pop(1 << rscratch1->encoding(), sp);
2226 }
2227 #endif
2228 }
2229 #endif
2230
2231 void MacroAssembler::resolve_jobject(Register value, Register thread, Register tmp) {
2232 Label done, not_weak;
2233 cbz(value, done); // Use NULL as-is.
2234
2235 STATIC_ASSERT(JNIHandles::weak_tag_mask == 1u);
2236 tbz(r0, 0, not_weak); // Test for jweak tag.
2237
2238 // Resolve jweak.
2239 access_load_at(T_OBJECT, IN_NATIVE | ON_PHANTOM_OOP_REF, value,
2240 Address(value, -JNIHandles::weak_tag_value), tmp, thread);
2241 verify_oop(value);
2242 b(done);
2243
2244 bind(not_weak);
2245 // Resolve (untagged) jobject.
2246 access_load_at(T_OBJECT, IN_NATIVE, value, Address(value, 0), tmp, thread);
2247 verify_oop(value);
2248 bind(done);
2249 }
2250
2251 void MacroAssembler::stop(const char* msg) {
2252 BLOCK_COMMENT(msg);
2253 dcps1(0xdeae);
2254 emit_int64((uintptr_t)msg);
2255 }
2256
2257 void MacroAssembler::unimplemented(const char* what) {
2258 const char* buf = NULL;
2259 {
2260 ResourceMark rm;
2261 stringStream ss;
2262 ss.print("unimplemented: %s", what);
2263 buf = code_string(ss.as_string());
2264 }
2265 stop(buf);
2266 }
2267
2268 // If a constant does not fit in an immediate field, generate some
2269 // number of MOV instructions and then perform the operation.
2270 void MacroAssembler::wrap_add_sub_imm_insn(Register Rd, Register Rn, unsigned imm,
2271 add_sub_imm_insn insn1,
2272 add_sub_reg_insn insn2) {
2273 assert(Rd != zr, "Rd = zr and not setting flags?");
2274 if (operand_valid_for_add_sub_immediate((int)imm)) {
2275 (this->*insn1)(Rd, Rn, imm);
2276 } else {
2277 if (uabs(imm) < (1 << 24)) {
2278 (this->*insn1)(Rd, Rn, imm & -(1 << 12));
2279 (this->*insn1)(Rd, Rd, imm & ((1 << 12)-1));
2280 } else {
2281 assert_different_registers(Rd, Rn);
2282 mov(Rd, (uint64_t)imm);
2283 (this->*insn2)(Rd, Rn, Rd, LSL, 0);
2284 }
2285 }
2286 }
2287
2288 // Seperate vsn which sets the flags. Optimisations are more restricted
2289 // because we must set the flags correctly.
2290 void MacroAssembler::wrap_adds_subs_imm_insn(Register Rd, Register Rn, unsigned imm,
2291 add_sub_imm_insn insn1,
2292 add_sub_reg_insn insn2) {
2293 if (operand_valid_for_add_sub_immediate((int)imm)) {
2294 (this->*insn1)(Rd, Rn, imm);
2295 } else {
2296 assert_different_registers(Rd, Rn);
2297 assert(Rd != zr, "overflow in immediate operand");
2298 mov(Rd, (uint64_t)imm);
2299 (this->*insn2)(Rd, Rn, Rd, LSL, 0);
2300 }
2301 }
2302
2303
2304 void MacroAssembler::add(Register Rd, Register Rn, RegisterOrConstant increment) {
2305 if (increment.is_register()) {
2306 add(Rd, Rn, increment.as_register());
2307 } else {
2308 add(Rd, Rn, increment.as_constant());
2309 }
2310 }
2311
2312 void MacroAssembler::addw(Register Rd, Register Rn, RegisterOrConstant increment) {
2313 if (increment.is_register()) {
2314 addw(Rd, Rn, increment.as_register());
2315 } else {
2316 addw(Rd, Rn, increment.as_constant());
2317 }
2318 }
2319
2320 void MacroAssembler::sub(Register Rd, Register Rn, RegisterOrConstant decrement) {
2321 if (decrement.is_register()) {
2322 sub(Rd, Rn, decrement.as_register());
2323 } else {
2324 sub(Rd, Rn, decrement.as_constant());
2325 }
2326 }
2327
2328 void MacroAssembler::subw(Register Rd, Register Rn, RegisterOrConstant decrement) {
2329 if (decrement.is_register()) {
2330 subw(Rd, Rn, decrement.as_register());
2331 } else {
2332 subw(Rd, Rn, decrement.as_constant());
2333 }
2334 }
2335
2336 void MacroAssembler::reinit_heapbase()
2337 {
2338 if (UseCompressedOops) {
2339 if (Universe::is_fully_initialized()) {
2340 mov(rheapbase, CompressedOops::ptrs_base());
2341 } else {
2342 lea(rheapbase, ExternalAddress((address)CompressedOops::ptrs_base_addr()));
2343 ldr(rheapbase, Address(rheapbase));
2344 }
2345 }
2346 }
2347
2348 // this simulates the behaviour of the x86 cmpxchg instruction using a
2349 // load linked/store conditional pair. we use the acquire/release
2350 // versions of these instructions so that we flush pending writes as
2351 // per Java semantics.
2352
2353 // n.b the x86 version assumes the old value to be compared against is
2354 // in rax and updates rax with the value located in memory if the
2355 // cmpxchg fails. we supply a register for the old value explicitly
2356
2357 // the aarch64 load linked/store conditional instructions do not
2358 // accept an offset. so, unlike x86, we must provide a plain register
2359 // to identify the memory word to be compared/exchanged rather than a
2360 // register+offset Address.
2361
2362 void MacroAssembler::cmpxchgptr(Register oldv, Register newv, Register addr, Register tmp,
2363 Label &succeed, Label *fail) {
2364 // oldv holds comparison value
2365 // newv holds value to write in exchange
2366 // addr identifies memory word to compare against/update
2367 if (UseLSE) {
2368 mov(tmp, oldv);
2369 casal(Assembler::xword, oldv, newv, addr);
2370 cmp(tmp, oldv);
2371 br(Assembler::EQ, succeed);
2372 membar(AnyAny);
2373 } else {
2374 Label retry_load, nope;
2375 if ((VM_Version::features() & VM_Version::CPU_STXR_PREFETCH))
2376 prfm(Address(addr), PSTL1STRM);
2377 bind(retry_load);
2378 // flush and load exclusive from the memory location
2379 // and fail if it is not what we expect
2380 ldaxr(tmp, addr);
2381 cmp(tmp, oldv);
2382 br(Assembler::NE, nope);
2383 // if we store+flush with no intervening write tmp wil be zero
2384 stlxr(tmp, newv, addr);
2385 cbzw(tmp, succeed);
2386 // retry so we only ever return after a load fails to compare
2387 // ensures we don't return a stale value after a failed write.
2388 b(retry_load);
2389 // if the memory word differs we return it in oldv and signal a fail
2390 bind(nope);
2391 membar(AnyAny);
2392 mov(oldv, tmp);
2393 }
2394 if (fail)
2395 b(*fail);
2396 }
2397
2398 void MacroAssembler::cmpxchg_obj_header(Register oldv, Register newv, Register obj, Register tmp,
2399 Label &succeed, Label *fail) {
2400 assert(oopDesc::mark_offset_in_bytes() == 0, "assumption");
2401 cmpxchgptr(oldv, newv, obj, tmp, succeed, fail);
2402 }
2403
2404 void MacroAssembler::cmpxchgw(Register oldv, Register newv, Register addr, Register tmp,
2405 Label &succeed, Label *fail) {
2406 // oldv holds comparison value
2407 // newv holds value to write in exchange
2408 // addr identifies memory word to compare against/update
2409 // tmp returns 0/1 for success/failure
2410 if (UseLSE) {
2411 mov(tmp, oldv);
2412 casal(Assembler::word, oldv, newv, addr);
2413 cmp(tmp, oldv);
2414 br(Assembler::EQ, succeed);
2415 membar(AnyAny);
2416 } else {
2417 Label retry_load, nope;
2418 if ((VM_Version::features() & VM_Version::CPU_STXR_PREFETCH))
2419 prfm(Address(addr), PSTL1STRM);
2420 bind(retry_load);
2421 // flush and load exclusive from the memory location
2422 // and fail if it is not what we expect
2423 ldaxrw(tmp, addr);
2424 cmp(tmp, oldv);
2425 br(Assembler::NE, nope);
2426 // if we store+flush with no intervening write tmp wil be zero
2427 stlxrw(tmp, newv, addr);
2428 cbzw(tmp, succeed);
2429 // retry so we only ever return after a load fails to compare
2430 // ensures we don't return a stale value after a failed write.
2431 b(retry_load);
2432 // if the memory word differs we return it in oldv and signal a fail
2433 bind(nope);
2434 membar(AnyAny);
2435 mov(oldv, tmp);
2436 }
2437 if (fail)
2438 b(*fail);
2439 }
2440
2441 // A generic CAS; success or failure is in the EQ flag. A weak CAS
2442 // doesn't retry and may fail spuriously. If the oldval is wanted,
2443 // Pass a register for the result, otherwise pass noreg.
2444
2445 // Clobbers rscratch1
2446 void MacroAssembler::cmpxchg(Register addr, Register expected,
2447 Register new_val,
2448 enum operand_size size,
2449 bool acquire, bool release,
2450 bool weak,
2451 Register result) {
2452 if (result == noreg) result = rscratch1;
2453 BLOCK_COMMENT("cmpxchg {");
2454 if (UseLSE) {
2455 mov(result, expected);
2456 lse_cas(result, new_val, addr, size, acquire, release, /*not_pair*/ true);
2457 compare_eq(result, expected, size);
2458 } else {
2459 Label retry_load, done;
2460 if ((VM_Version::features() & VM_Version::CPU_STXR_PREFETCH))
2461 prfm(Address(addr), PSTL1STRM);
2462 bind(retry_load);
2463 load_exclusive(result, addr, size, acquire);
2464 compare_eq(result, expected, size);
2465 br(Assembler::NE, done);
2466 store_exclusive(rscratch1, new_val, addr, size, release);
2467 if (weak) {
2468 cmpw(rscratch1, 0u); // If the store fails, return NE to our caller.
2469 } else {
2470 cbnzw(rscratch1, retry_load);
2471 }
2472 bind(done);
2473 }
2474 BLOCK_COMMENT("} cmpxchg");
2475 }
2476
2477 // A generic comparison. Only compares for equality, clobbers rscratch1.
2478 void MacroAssembler::compare_eq(Register rm, Register rn, enum operand_size size) {
2479 if (size == xword) {
2480 cmp(rm, rn);
2481 } else if (size == word) {
2482 cmpw(rm, rn);
2483 } else if (size == halfword) {
2484 eorw(rscratch1, rm, rn);
2485 ands(zr, rscratch1, 0xffff);
2486 } else if (size == byte) {
2487 eorw(rscratch1, rm, rn);
2488 ands(zr, rscratch1, 0xff);
2489 } else {
2490 ShouldNotReachHere();
2491 }
2492 }
2493
2494
2495 static bool different(Register a, RegisterOrConstant b, Register c) {
2496 if (b.is_constant())
2497 return a != c;
2498 else
2499 return a != b.as_register() && a != c && b.as_register() != c;
2500 }
2501
2502 #define ATOMIC_OP(NAME, LDXR, OP, IOP, AOP, STXR, sz) \
2503 void MacroAssembler::atomic_##NAME(Register prev, RegisterOrConstant incr, Register addr) { \
2504 if (UseLSE) { \
2505 prev = prev->is_valid() ? prev : zr; \
2506 if (incr.is_register()) { \
2507 AOP(sz, incr.as_register(), prev, addr); \
2508 } else { \
2509 mov(rscratch2, incr.as_constant()); \
2510 AOP(sz, rscratch2, prev, addr); \
2511 } \
2512 return; \
2513 } \
2514 Register result = rscratch2; \
2515 if (prev->is_valid()) \
2516 result = different(prev, incr, addr) ? prev : rscratch2; \
2517 \
2518 Label retry_load; \
2519 if ((VM_Version::features() & VM_Version::CPU_STXR_PREFETCH)) \
2520 prfm(Address(addr), PSTL1STRM); \
2521 bind(retry_load); \
2522 LDXR(result, addr); \
2523 OP(rscratch1, result, incr); \
2524 STXR(rscratch2, rscratch1, addr); \
2525 cbnzw(rscratch2, retry_load); \
2526 if (prev->is_valid() && prev != result) { \
2527 IOP(prev, rscratch1, incr); \
2528 } \
2529 }
2530
2531 ATOMIC_OP(add, ldxr, add, sub, ldadd, stxr, Assembler::xword)
2532 ATOMIC_OP(addw, ldxrw, addw, subw, ldadd, stxrw, Assembler::word)
2533 ATOMIC_OP(addal, ldaxr, add, sub, ldaddal, stlxr, Assembler::xword)
2534 ATOMIC_OP(addalw, ldaxrw, addw, subw, ldaddal, stlxrw, Assembler::word)
2535
2536 #undef ATOMIC_OP
2537
2538 #define ATOMIC_XCHG(OP, AOP, LDXR, STXR, sz) \
2539 void MacroAssembler::atomic_##OP(Register prev, Register newv, Register addr) { \
2540 if (UseLSE) { \
2541 prev = prev->is_valid() ? prev : zr; \
2542 AOP(sz, newv, prev, addr); \
2543 return; \
2544 } \
2545 Register result = rscratch2; \
2546 if (prev->is_valid()) \
2547 result = different(prev, newv, addr) ? prev : rscratch2; \
2548 \
2549 Label retry_load; \
2550 if ((VM_Version::features() & VM_Version::CPU_STXR_PREFETCH)) \
2551 prfm(Address(addr), PSTL1STRM); \
2552 bind(retry_load); \
2553 LDXR(result, addr); \
2554 STXR(rscratch1, newv, addr); \
2555 cbnzw(rscratch1, retry_load); \
2556 if (prev->is_valid() && prev != result) \
2557 mov(prev, result); \
2558 }
2559
2560 ATOMIC_XCHG(xchg, swp, ldxr, stxr, Assembler::xword)
2561 ATOMIC_XCHG(xchgw, swp, ldxrw, stxrw, Assembler::word)
2562 ATOMIC_XCHG(xchgal, swpal, ldaxr, stlxr, Assembler::xword)
2563 ATOMIC_XCHG(xchgalw, swpal, ldaxrw, stlxrw, Assembler::word)
2564
2565 #undef ATOMIC_XCHG
2566
2567 #ifndef PRODUCT
2568 extern "C" void findpc(intptr_t x);
2569 #endif
2570
2571 void MacroAssembler::debug64(char* msg, int64_t pc, int64_t regs[])
2572 {
2573 // In order to get locks to work, we need to fake a in_VM state
2574 if (ShowMessageBoxOnError ) {
2575 JavaThread* thread = JavaThread::current();
2576 JavaThreadState saved_state = thread->thread_state();
2577 thread->set_thread_state(_thread_in_vm);
2578 #ifndef PRODUCT
2579 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) {
2580 ttyLocker ttyl;
2581 BytecodeCounter::print();
2582 }
2583 #endif
2584 if (os::message_box(msg, "Execution stopped, print registers?")) {
2585 ttyLocker ttyl;
2586 tty->print_cr(" pc = 0x%016lx", pc);
2587 #ifndef PRODUCT
2588 tty->cr();
2589 findpc(pc);
2590 tty->cr();
2591 #endif
2592 tty->print_cr(" r0 = 0x%016lx", regs[0]);
2593 tty->print_cr(" r1 = 0x%016lx", regs[1]);
2594 tty->print_cr(" r2 = 0x%016lx", regs[2]);
2595 tty->print_cr(" r3 = 0x%016lx", regs[3]);
2596 tty->print_cr(" r4 = 0x%016lx", regs[4]);
2597 tty->print_cr(" r5 = 0x%016lx", regs[5]);
2598 tty->print_cr(" r6 = 0x%016lx", regs[6]);
2599 tty->print_cr(" r7 = 0x%016lx", regs[7]);
2600 tty->print_cr(" r8 = 0x%016lx", regs[8]);
2601 tty->print_cr(" r9 = 0x%016lx", regs[9]);
2602 tty->print_cr("r10 = 0x%016lx", regs[10]);
2603 tty->print_cr("r11 = 0x%016lx", regs[11]);
2604 tty->print_cr("r12 = 0x%016lx", regs[12]);
2605 tty->print_cr("r13 = 0x%016lx", regs[13]);
2606 tty->print_cr("r14 = 0x%016lx", regs[14]);
2607 tty->print_cr("r15 = 0x%016lx", regs[15]);
2608 tty->print_cr("r16 = 0x%016lx", regs[16]);
2609 tty->print_cr("r17 = 0x%016lx", regs[17]);
2610 tty->print_cr("r18 = 0x%016lx", regs[18]);
2611 tty->print_cr("r19 = 0x%016lx", regs[19]);
2612 tty->print_cr("r20 = 0x%016lx", regs[20]);
2613 tty->print_cr("r21 = 0x%016lx", regs[21]);
2614 tty->print_cr("r22 = 0x%016lx", regs[22]);
2615 tty->print_cr("r23 = 0x%016lx", regs[23]);
2616 tty->print_cr("r24 = 0x%016lx", regs[24]);
2617 tty->print_cr("r25 = 0x%016lx", regs[25]);
2618 tty->print_cr("r26 = 0x%016lx", regs[26]);
2619 tty->print_cr("r27 = 0x%016lx", regs[27]);
2620 tty->print_cr("r28 = 0x%016lx", regs[28]);
2621 tty->print_cr("r30 = 0x%016lx", regs[30]);
2622 tty->print_cr("r31 = 0x%016lx", regs[31]);
2623 BREAKPOINT;
2624 }
2625 }
2626 fatal("DEBUG MESSAGE: %s", msg);
2627 }
2628
2629 void MacroAssembler::push_call_clobbered_registers_except(RegSet exclude) {
2630 int step = 4 * wordSize;
2631 push(RegSet::range(r0, r18) - RegSet::of(rscratch1, rscratch2) - exclude, sp);
2632 sub(sp, sp, step);
2633 mov(rscratch1, -step);
2634 // Push v0-v7, v16-v31.
2635 for (int i = 31; i>= 4; i -= 4) {
2636 if (i <= v7->encoding() || i >= v16->encoding())
2637 st1(as_FloatRegister(i-3), as_FloatRegister(i-2), as_FloatRegister(i-1),
2638 as_FloatRegister(i), T1D, Address(post(sp, rscratch1)));
2639 }
2640 st1(as_FloatRegister(0), as_FloatRegister(1), as_FloatRegister(2),
2641 as_FloatRegister(3), T1D, Address(sp));
2642 }
2643
2644 void MacroAssembler::pop_call_clobbered_registers_except(RegSet exclude) {
2645 for (int i = 0; i < 32; i += 4) {
2646 if (i <= v7->encoding() || i >= v16->encoding())
2647 ld1(as_FloatRegister(i), as_FloatRegister(i+1), as_FloatRegister(i+2),
2648 as_FloatRegister(i+3), T1D, Address(post(sp, 4 * wordSize)));
2649 }
2650
2651 pop(RegSet::range(r0, r18) - RegSet::of(rscratch1, rscratch2) - exclude, sp);
2652 }
2653
2654 void MacroAssembler::push_CPU_state(bool save_vectors) {
2655 int step = (save_vectors ? 8 : 4) * wordSize;
2656 push(0x3fffffff, sp); // integer registers except lr & sp
2657 mov(rscratch1, -step);
2658 sub(sp, sp, step);
2659 for (int i = 28; i >= 4; i -= 4) {
2660 st1(as_FloatRegister(i), as_FloatRegister(i+1), as_FloatRegister(i+2),
2661 as_FloatRegister(i+3), save_vectors ? T2D : T1D, Address(post(sp, rscratch1)));
2662 }
2663 st1(v0, v1, v2, v3, save_vectors ? T2D : T1D, sp);
2664 }
2665
2666 void MacroAssembler::pop_CPU_state(bool restore_vectors) {
2667 int step = (restore_vectors ? 8 : 4) * wordSize;
2668 for (int i = 0; i <= 28; i += 4)
2669 ld1(as_FloatRegister(i), as_FloatRegister(i+1), as_FloatRegister(i+2),
2670 as_FloatRegister(i+3), restore_vectors ? T2D : T1D, Address(post(sp, step)));
2671 pop(0x3fffffff, sp); // integer registers except lr & sp
2672 }
2673
2674 /**
2675 * Helpers for multiply_to_len().
2676 */
2677 void MacroAssembler::add2_with_carry(Register final_dest_hi, Register dest_hi, Register dest_lo,
2678 Register src1, Register src2) {
2679 adds(dest_lo, dest_lo, src1);
2680 adc(dest_hi, dest_hi, zr);
2681 adds(dest_lo, dest_lo, src2);
2682 adc(final_dest_hi, dest_hi, zr);
2683 }
2684
2685 // Generate an address from (r + r1 extend offset). "size" is the
2686 // size of the operand. The result may be in rscratch2.
2687 Address MacroAssembler::offsetted_address(Register r, Register r1,
2688 Address::extend ext, int offset, int size) {
2689 if (offset || (ext.shift() % size != 0)) {
2690 lea(rscratch2, Address(r, r1, ext));
2691 return Address(rscratch2, offset);
2692 } else {
2693 return Address(r, r1, ext);
2694 }
2695 }
2696
2697 Address MacroAssembler::spill_address(int size, int offset, Register tmp)
2698 {
2699 assert(offset >= 0, "spill to negative address?");
2700 // Offset reachable ?
2701 // Not aligned - 9 bits signed offset
2702 // Aligned - 12 bits unsigned offset shifted
2703 Register base = sp;
2704 if ((offset & (size-1)) && offset >= (1<<8)) {
2705 add(tmp, base, offset & ((1<<12)-1));
2706 base = tmp;
2707 offset &= -1u<<12;
2708 }
2709
2710 if (offset >= (1<<12) * size) {
2711 add(tmp, base, offset & (((1<<12)-1)<<12));
2712 base = tmp;
2713 offset &= ~(((1<<12)-1)<<12);
2714 }
2715
2716 return Address(base, offset);
2717 }
2718
2719 // Checks whether offset is aligned.
2720 // Returns true if it is, else false.
2721 bool MacroAssembler::merge_alignment_check(Register base,
2722 size_t size,
2723 int64_t cur_offset,
2724 int64_t prev_offset) const {
2725 if (AvoidUnalignedAccesses) {
2726 if (base == sp) {
2727 // Checks whether low offset if aligned to pair of registers.
2728 int64_t pair_mask = size * 2 - 1;
2729 int64_t offset = prev_offset > cur_offset ? cur_offset : prev_offset;
2730 return (offset & pair_mask) == 0;
2731 } else { // If base is not sp, we can't guarantee the access is aligned.
2732 return false;
2733 }
2734 } else {
2735 int64_t mask = size - 1;
2736 // Load/store pair instruction only supports element size aligned offset.
2737 return (cur_offset & mask) == 0 && (prev_offset & mask) == 0;
2738 }
2739 }
2740
2741 // Checks whether current and previous loads/stores can be merged.
2742 // Returns true if it can be merged, else false.
2743 bool MacroAssembler::ldst_can_merge(Register rt,
2744 const Address &adr,
2745 size_t cur_size_in_bytes,
2746 bool is_store) const {
2747 address prev = pc() - NativeInstruction::instruction_size;
2748 address last = code()->last_insn();
2749
2750 if (last == NULL || !nativeInstruction_at(last)->is_Imm_LdSt()) {
2751 return false;
2752 }
2753
2754 if (adr.getMode() != Address::base_plus_offset || prev != last) {
2755 return false;
2756 }
2757
2758 NativeLdSt* prev_ldst = NativeLdSt_at(prev);
2759 size_t prev_size_in_bytes = prev_ldst->size_in_bytes();
2760
2761 assert(prev_size_in_bytes == 4 || prev_size_in_bytes == 8, "only supports 64/32bit merging.");
2762 assert(cur_size_in_bytes == 4 || cur_size_in_bytes == 8, "only supports 64/32bit merging.");
2763
2764 if (cur_size_in_bytes != prev_size_in_bytes || is_store != prev_ldst->is_store()) {
2765 return false;
2766 }
2767
2768 int64_t max_offset = 63 * prev_size_in_bytes;
2769 int64_t min_offset = -64 * prev_size_in_bytes;
2770
2771 assert(prev_ldst->is_not_pre_post_index(), "pre-index or post-index is not supported to be merged.");
2772
2773 // Only same base can be merged.
2774 if (adr.base() != prev_ldst->base()) {
2775 return false;
2776 }
2777
2778 int64_t cur_offset = adr.offset();
2779 int64_t prev_offset = prev_ldst->offset();
2780 size_t diff = abs(cur_offset - prev_offset);
2781 if (diff != prev_size_in_bytes) {
2782 return false;
2783 }
2784
2785 // Following cases can not be merged:
2786 // ldr x2, [x2, #8]
2787 // ldr x3, [x2, #16]
2788 // or:
2789 // ldr x2, [x3, #8]
2790 // ldr x2, [x3, #16]
2791 // If t1 and t2 is the same in "ldp t1, t2, [xn, #imm]", we'll get SIGILL.
2792 if (!is_store && (adr.base() == prev_ldst->target() || rt == prev_ldst->target())) {
2793 return false;
2794 }
2795
2796 int64_t low_offset = prev_offset > cur_offset ? cur_offset : prev_offset;
2797 // Offset range must be in ldp/stp instruction's range.
2798 if (low_offset > max_offset || low_offset < min_offset) {
2799 return false;
2800 }
2801
2802 if (merge_alignment_check(adr.base(), prev_size_in_bytes, cur_offset, prev_offset)) {
2803 return true;
2804 }
2805
2806 return false;
2807 }
2808
2809 // Merge current load/store with previous load/store into ldp/stp.
2810 void MacroAssembler::merge_ldst(Register rt,
2811 const Address &adr,
2812 size_t cur_size_in_bytes,
2813 bool is_store) {
2814
2815 assert(ldst_can_merge(rt, adr, cur_size_in_bytes, is_store) == true, "cur and prev must be able to be merged.");
2816
2817 Register rt_low, rt_high;
2818 address prev = pc() - NativeInstruction::instruction_size;
2819 NativeLdSt* prev_ldst = NativeLdSt_at(prev);
2820
2821 int64_t offset;
2822
2823 if (adr.offset() < prev_ldst->offset()) {
2824 offset = adr.offset();
2825 rt_low = rt;
2826 rt_high = prev_ldst->target();
2827 } else {
2828 offset = prev_ldst->offset();
2829 rt_low = prev_ldst->target();
2830 rt_high = rt;
2831 }
2832
2833 Address adr_p = Address(prev_ldst->base(), offset);
2834 // Overwrite previous generated binary.
2835 code_section()->set_end(prev);
2836
2837 const int sz = prev_ldst->size_in_bytes();
2838 assert(sz == 8 || sz == 4, "only supports 64/32bit merging.");
2839 if (!is_store) {
2840 BLOCK_COMMENT("merged ldr pair");
2841 if (sz == 8) {
2842 ldp(rt_low, rt_high, adr_p);
2843 } else {
2844 ldpw(rt_low, rt_high, adr_p);
2845 }
2846 } else {
2847 BLOCK_COMMENT("merged str pair");
2848 if (sz == 8) {
2849 stp(rt_low, rt_high, adr_p);
2850 } else {
2851 stpw(rt_low, rt_high, adr_p);
2852 }
2853 }
2854 }
2855
2856 /**
2857 * Multiply 64 bit by 64 bit first loop.
2858 */
2859 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart, Register x_xstart,
2860 Register y, Register y_idx, Register z,
2861 Register carry, Register product,
2862 Register idx, Register kdx) {
2863 //
2864 // jlong carry, x[], y[], z[];
2865 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
2866 // huge_128 product = y[idx] * x[xstart] + carry;
2867 // z[kdx] = (jlong)product;
2868 // carry = (jlong)(product >>> 64);
2869 // }
2870 // z[xstart] = carry;
2871 //
2872
2873 Label L_first_loop, L_first_loop_exit;
2874 Label L_one_x, L_one_y, L_multiply;
2875
2876 subsw(xstart, xstart, 1);
2877 br(Assembler::MI, L_one_x);
2878
2879 lea(rscratch1, Address(x, xstart, Address::lsl(LogBytesPerInt)));
2880 ldr(x_xstart, Address(rscratch1));
2881 ror(x_xstart, x_xstart, 32); // convert big-endian to little-endian
2882
2883 bind(L_first_loop);
2884 subsw(idx, idx, 1);
2885 br(Assembler::MI, L_first_loop_exit);
2886 subsw(idx, idx, 1);
2887 br(Assembler::MI, L_one_y);
2888 lea(rscratch1, Address(y, idx, Address::uxtw(LogBytesPerInt)));
2889 ldr(y_idx, Address(rscratch1));
2890 ror(y_idx, y_idx, 32); // convert big-endian to little-endian
2891 bind(L_multiply);
2892
2893 // AArch64 has a multiply-accumulate instruction that we can't use
2894 // here because it has no way to process carries, so we have to use
2895 // separate add and adc instructions. Bah.
2896 umulh(rscratch1, x_xstart, y_idx); // x_xstart * y_idx -> rscratch1:product
2897 mul(product, x_xstart, y_idx);
2898 adds(product, product, carry);
2899 adc(carry, rscratch1, zr); // x_xstart * y_idx + carry -> carry:product
2900
2901 subw(kdx, kdx, 2);
2902 ror(product, product, 32); // back to big-endian
2903 str(product, offsetted_address(z, kdx, Address::uxtw(LogBytesPerInt), 0, BytesPerLong));
2904
2905 b(L_first_loop);
2906
2907 bind(L_one_y);
2908 ldrw(y_idx, Address(y, 0));
2909 b(L_multiply);
2910
2911 bind(L_one_x);
2912 ldrw(x_xstart, Address(x, 0));
2913 b(L_first_loop);
2914
2915 bind(L_first_loop_exit);
2916 }
2917
2918 /**
2919 * Multiply 128 bit by 128. Unrolled inner loop.
2920 *
2921 */
2922 void MacroAssembler::multiply_128_x_128_loop(Register y, Register z,
2923 Register carry, Register carry2,
2924 Register idx, Register jdx,
2925 Register yz_idx1, Register yz_idx2,
2926 Register tmp, Register tmp3, Register tmp4,
2927 Register tmp6, Register product_hi) {
2928
2929 // jlong carry, x[], y[], z[];
2930 // int kdx = ystart+1;
2931 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
2932 // huge_128 tmp3 = (y[idx+1] * product_hi) + z[kdx+idx+1] + carry;
2933 // jlong carry2 = (jlong)(tmp3 >>> 64);
2934 // huge_128 tmp4 = (y[idx] * product_hi) + z[kdx+idx] + carry2;
2935 // carry = (jlong)(tmp4 >>> 64);
2936 // z[kdx+idx+1] = (jlong)tmp3;
2937 // z[kdx+idx] = (jlong)tmp4;
2938 // }
2939 // idx += 2;
2940 // if (idx > 0) {
2941 // yz_idx1 = (y[idx] * product_hi) + z[kdx+idx] + carry;
2942 // z[kdx+idx] = (jlong)yz_idx1;
2943 // carry = (jlong)(yz_idx1 >>> 64);
2944 // }
2945 //
2946
2947 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
2948
2949 lsrw(jdx, idx, 2);
2950
2951 bind(L_third_loop);
2952
2953 subsw(jdx, jdx, 1);
2954 br(Assembler::MI, L_third_loop_exit);
2955 subw(idx, idx, 4);
2956
2957 lea(rscratch1, Address(y, idx, Address::uxtw(LogBytesPerInt)));
2958
2959 ldp(yz_idx2, yz_idx1, Address(rscratch1, 0));
2960
2961 lea(tmp6, Address(z, idx, Address::uxtw(LogBytesPerInt)));
2962
2963 ror(yz_idx1, yz_idx1, 32); // convert big-endian to little-endian
2964 ror(yz_idx2, yz_idx2, 32);
2965
2966 ldp(rscratch2, rscratch1, Address(tmp6, 0));
2967
2968 mul(tmp3, product_hi, yz_idx1); // yz_idx1 * product_hi -> tmp4:tmp3
2969 umulh(tmp4, product_hi, yz_idx1);
2970
2971 ror(rscratch1, rscratch1, 32); // convert big-endian to little-endian
2972 ror(rscratch2, rscratch2, 32);
2973
2974 mul(tmp, product_hi, yz_idx2); // yz_idx2 * product_hi -> carry2:tmp
2975 umulh(carry2, product_hi, yz_idx2);
2976
2977 // propagate sum of both multiplications into carry:tmp4:tmp3
2978 adds(tmp3, tmp3, carry);
2979 adc(tmp4, tmp4, zr);
2980 adds(tmp3, tmp3, rscratch1);
2981 adcs(tmp4, tmp4, tmp);
2982 adc(carry, carry2, zr);
2983 adds(tmp4, tmp4, rscratch2);
2984 adc(carry, carry, zr);
2985
2986 ror(tmp3, tmp3, 32); // convert little-endian to big-endian
2987 ror(tmp4, tmp4, 32);
2988 stp(tmp4, tmp3, Address(tmp6, 0));
2989
2990 b(L_third_loop);
2991 bind (L_third_loop_exit);
2992
2993 andw (idx, idx, 0x3);
2994 cbz(idx, L_post_third_loop_done);
2995
2996 Label L_check_1;
2997 subsw(idx, idx, 2);
2998 br(Assembler::MI, L_check_1);
2999
3000 lea(rscratch1, Address(y, idx, Address::uxtw(LogBytesPerInt)));
3001 ldr(yz_idx1, Address(rscratch1, 0));
3002 ror(yz_idx1, yz_idx1, 32);
3003 mul(tmp3, product_hi, yz_idx1); // yz_idx1 * product_hi -> tmp4:tmp3
3004 umulh(tmp4, product_hi, yz_idx1);
3005 lea(rscratch1, Address(z, idx, Address::uxtw(LogBytesPerInt)));
3006 ldr(yz_idx2, Address(rscratch1, 0));
3007 ror(yz_idx2, yz_idx2, 32);
3008
3009 add2_with_carry(carry, tmp4, tmp3, carry, yz_idx2);
3010
3011 ror(tmp3, tmp3, 32);
3012 str(tmp3, Address(rscratch1, 0));
3013
3014 bind (L_check_1);
3015
3016 andw (idx, idx, 0x1);
3017 subsw(idx, idx, 1);
3018 br(Assembler::MI, L_post_third_loop_done);
3019 ldrw(tmp4, Address(y, idx, Address::uxtw(LogBytesPerInt)));
3020 mul(tmp3, tmp4, product_hi); // tmp4 * product_hi -> carry2:tmp3
3021 umulh(carry2, tmp4, product_hi);
3022 ldrw(tmp4, Address(z, idx, Address::uxtw(LogBytesPerInt)));
3023
3024 add2_with_carry(carry2, tmp3, tmp4, carry);
3025
3026 strw(tmp3, Address(z, idx, Address::uxtw(LogBytesPerInt)));
3027 extr(carry, carry2, tmp3, 32);
3028
3029 bind(L_post_third_loop_done);
3030 }
3031
3032 /**
3033 * Code for BigInteger::multiplyToLen() instrinsic.
3034 *
3035 * r0: x
3036 * r1: xlen
3037 * r2: y
3038 * r3: ylen
3039 * r4: z
3040 * r5: zlen
3041 * r10: tmp1
3042 * r11: tmp2
3043 * r12: tmp3
3044 * r13: tmp4
3045 * r14: tmp5
3046 * r15: tmp6
3047 * r16: tmp7
3048 *
3049 */
3050 void MacroAssembler::multiply_to_len(Register x, Register xlen, Register y, Register ylen,
3051 Register z, Register zlen,
3052 Register tmp1, Register tmp2, Register tmp3, Register tmp4,
3053 Register tmp5, Register tmp6, Register product_hi) {
3054
3055 assert_different_registers(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6);
3056
3057 const Register idx = tmp1;
3058 const Register kdx = tmp2;
3059 const Register xstart = tmp3;
3060
3061 const Register y_idx = tmp4;
3062 const Register carry = tmp5;
3063 const Register product = xlen;
3064 const Register x_xstart = zlen; // reuse register
3065
3066 // First Loop.
3067 //
3068 // final static long LONG_MASK = 0xffffffffL;
3069 // int xstart = xlen - 1;
3070 // int ystart = ylen - 1;
3071 // long carry = 0;
3072 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
3073 // long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry;
3074 // z[kdx] = (int)product;
3075 // carry = product >>> 32;
3076 // }
3077 // z[xstart] = (int)carry;
3078 //
3079
3080 movw(idx, ylen); // idx = ylen;
3081 movw(kdx, zlen); // kdx = xlen+ylen;
3082 mov(carry, zr); // carry = 0;
3083
3084 Label L_done;
3085
3086 movw(xstart, xlen);
3087 subsw(xstart, xstart, 1);
3088 br(Assembler::MI, L_done);
3089
3090 multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx);
3091
3092 Label L_second_loop;
3093 cbzw(kdx, L_second_loop);
3094
3095 Label L_carry;
3096 subw(kdx, kdx, 1);
3097 cbzw(kdx, L_carry);
3098
3099 strw(carry, Address(z, kdx, Address::uxtw(LogBytesPerInt)));
3100 lsr(carry, carry, 32);
3101 subw(kdx, kdx, 1);
3102
3103 bind(L_carry);
3104 strw(carry, Address(z, kdx, Address::uxtw(LogBytesPerInt)));
3105
3106 // Second and third (nested) loops.
3107 //
3108 // for (int i = xstart-1; i >= 0; i--) { // Second loop
3109 // carry = 0;
3110 // for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop
3111 // long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) +
3112 // (z[k] & LONG_MASK) + carry;
3113 // z[k] = (int)product;
3114 // carry = product >>> 32;
3115 // }
3116 // z[i] = (int)carry;
3117 // }
3118 //
3119 // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = product_hi
3120
3121 const Register jdx = tmp1;
3122
3123 bind(L_second_loop);
3124 mov(carry, zr); // carry = 0;
3125 movw(jdx, ylen); // j = ystart+1
3126
3127 subsw(xstart, xstart, 1); // i = xstart-1;
3128 br(Assembler::MI, L_done);
3129
3130 str(z, Address(pre(sp, -4 * wordSize)));
3131
3132 Label L_last_x;
3133 lea(z, offsetted_address(z, xstart, Address::uxtw(LogBytesPerInt), 4, BytesPerInt)); // z = z + k - j
3134 subsw(xstart, xstart, 1); // i = xstart-1;
3135 br(Assembler::MI, L_last_x);
3136
3137 lea(rscratch1, Address(x, xstart, Address::uxtw(LogBytesPerInt)));
3138 ldr(product_hi, Address(rscratch1));
3139 ror(product_hi, product_hi, 32); // convert big-endian to little-endian
3140
3141 Label L_third_loop_prologue;
3142 bind(L_third_loop_prologue);
3143
3144 str(ylen, Address(sp, wordSize));
3145 stp(x, xstart, Address(sp, 2 * wordSize));
3146 multiply_128_x_128_loop(y, z, carry, x, jdx, ylen, product,
3147 tmp2, x_xstart, tmp3, tmp4, tmp6, product_hi);
3148 ldp(z, ylen, Address(post(sp, 2 * wordSize)));
3149 ldp(x, xlen, Address(post(sp, 2 * wordSize))); // copy old xstart -> xlen
3150
3151 addw(tmp3, xlen, 1);
3152 strw(carry, Address(z, tmp3, Address::uxtw(LogBytesPerInt)));
3153 subsw(tmp3, tmp3, 1);
3154 br(Assembler::MI, L_done);
3155
3156 lsr(carry, carry, 32);
3157 strw(carry, Address(z, tmp3, Address::uxtw(LogBytesPerInt)));
3158 b(L_second_loop);
3159
3160 // Next infrequent code is moved outside loops.
3161 bind(L_last_x);
3162 ldrw(product_hi, Address(x, 0));
3163 b(L_third_loop_prologue);
3164
3165 bind(L_done);
3166 }
3167
3168 // Code for BigInteger::mulAdd instrinsic
3169 // out = r0
3170 // in = r1
3171 // offset = r2 (already out.length-offset)
3172 // len = r3
3173 // k = r4
3174 //
3175 // pseudo code from java implementation:
3176 // carry = 0;
3177 // offset = out.length-offset - 1;
3178 // for (int j=len-1; j >= 0; j--) {
3179 // product = (in[j] & LONG_MASK) * kLong + (out[offset] & LONG_MASK) + carry;
3180 // out[offset--] = (int)product;
3181 // carry = product >>> 32;
3182 // }
3183 // return (int)carry;
3184 void MacroAssembler::mul_add(Register out, Register in, Register offset,
3185 Register len, Register k) {
3186 Label LOOP, END;
3187 // pre-loop
3188 cmp(len, zr); // cmp, not cbz/cbnz: to use condition twice => less branches
3189 csel(out, zr, out, Assembler::EQ);
3190 br(Assembler::EQ, END);
3191 add(in, in, len, LSL, 2); // in[j+1] address
3192 add(offset, out, offset, LSL, 2); // out[offset + 1] address
3193 mov(out, zr); // used to keep carry now
3194 BIND(LOOP);
3195 ldrw(rscratch1, Address(pre(in, -4)));
3196 madd(rscratch1, rscratch1, k, out);
3197 ldrw(rscratch2, Address(pre(offset, -4)));
3198 add(rscratch1, rscratch1, rscratch2);
3199 strw(rscratch1, Address(offset));
3200 lsr(out, rscratch1, 32);
3201 subs(len, len, 1);
3202 br(Assembler::NE, LOOP);
3203 BIND(END);
3204 }
3205
3206 /**
3207 * Emits code to update CRC-32 with a byte value according to constants in table
3208 *
3209 * @param [in,out]crc Register containing the crc.
3210 * @param [in]val Register containing the byte to fold into the CRC.
3211 * @param [in]table Register containing the table of crc constants.
3212 *
3213 * uint32_t crc;
3214 * val = crc_table[(val ^ crc) & 0xFF];
3215 * crc = val ^ (crc >> 8);
3216 *
3217 */
3218 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) {
3219 eor(val, val, crc);
3220 andr(val, val, 0xff);
3221 ldrw(val, Address(table, val, Address::lsl(2)));
3222 eor(crc, val, crc, Assembler::LSR, 8);
3223 }
3224
3225 /**
3226 * Emits code to update CRC-32 with a 32-bit value according to tables 0 to 3
3227 *
3228 * @param [in,out]crc Register containing the crc.
3229 * @param [in]v Register containing the 32-bit to fold into the CRC.
3230 * @param [in]table0 Register containing table 0 of crc constants.
3231 * @param [in]table1 Register containing table 1 of crc constants.
3232 * @param [in]table2 Register containing table 2 of crc constants.
3233 * @param [in]table3 Register containing table 3 of crc constants.
3234 *
3235 * uint32_t crc;
3236 * v = crc ^ v
3237 * crc = table3[v&0xff]^table2[(v>>8)&0xff]^table1[(v>>16)&0xff]^table0[v>>24]
3238 *
3239 */
3240 void MacroAssembler::update_word_crc32(Register crc, Register v, Register tmp,
3241 Register table0, Register table1, Register table2, Register table3,
3242 bool upper) {
3243 eor(v, crc, v, upper ? LSR:LSL, upper ? 32:0);
3244 uxtb(tmp, v);
3245 ldrw(crc, Address(table3, tmp, Address::lsl(2)));
3246 ubfx(tmp, v, 8, 8);
3247 ldrw(tmp, Address(table2, tmp, Address::lsl(2)));
3248 eor(crc, crc, tmp);
3249 ubfx(tmp, v, 16, 8);
3250 ldrw(tmp, Address(table1, tmp, Address::lsl(2)));
3251 eor(crc, crc, tmp);
3252 ubfx(tmp, v, 24, 8);
3253 ldrw(tmp, Address(table0, tmp, Address::lsl(2)));
3254 eor(crc, crc, tmp);
3255 }
3256
3257 void MacroAssembler::kernel_crc32_using_crc32(Register crc, Register buf,
3258 Register len, Register tmp0, Register tmp1, Register tmp2,
3259 Register tmp3) {
3260 Label CRC_by64_loop, CRC_by4_loop, CRC_by1_loop, CRC_less64, CRC_by64_pre, CRC_by32_loop, CRC_less32, L_exit;
3261 assert_different_registers(crc, buf, len, tmp0, tmp1, tmp2, tmp3);
3262
3263 mvnw(crc, crc);
3264
3265 subs(len, len, 128);
3266 br(Assembler::GE, CRC_by64_pre);
3267 BIND(CRC_less64);
3268 adds(len, len, 128-32);
3269 br(Assembler::GE, CRC_by32_loop);
3270 BIND(CRC_less32);
3271 adds(len, len, 32-4);
3272 br(Assembler::GE, CRC_by4_loop);
3273 adds(len, len, 4);
3274 br(Assembler::GT, CRC_by1_loop);
3275 b(L_exit);
3276
3277 BIND(CRC_by32_loop);
3278 ldp(tmp0, tmp1, Address(post(buf, 16)));
3279 subs(len, len, 32);
3280 crc32x(crc, crc, tmp0);
3281 ldr(tmp2, Address(post(buf, 8)));
3282 crc32x(crc, crc, tmp1);
3283 ldr(tmp3, Address(post(buf, 8)));
3284 crc32x(crc, crc, tmp2);
3285 crc32x(crc, crc, tmp3);
3286 br(Assembler::GE, CRC_by32_loop);
3287 cmn(len, 32);
3288 br(Assembler::NE, CRC_less32);
3289 b(L_exit);
3290
3291 BIND(CRC_by4_loop);
3292 ldrw(tmp0, Address(post(buf, 4)));
3293 subs(len, len, 4);
3294 crc32w(crc, crc, tmp0);
3295 br(Assembler::GE, CRC_by4_loop);
3296 adds(len, len, 4);
3297 br(Assembler::LE, L_exit);
3298 BIND(CRC_by1_loop);
3299 ldrb(tmp0, Address(post(buf, 1)));
3300 subs(len, len, 1);
3301 crc32b(crc, crc, tmp0);
3302 br(Assembler::GT, CRC_by1_loop);
3303 b(L_exit);
3304
3305 BIND(CRC_by64_pre);
3306 sub(buf, buf, 8);
3307 ldp(tmp0, tmp1, Address(buf, 8));
3308 crc32x(crc, crc, tmp0);
3309 ldr(tmp2, Address(buf, 24));
3310 crc32x(crc, crc, tmp1);
3311 ldr(tmp3, Address(buf, 32));
3312 crc32x(crc, crc, tmp2);
3313 ldr(tmp0, Address(buf, 40));
3314 crc32x(crc, crc, tmp3);
3315 ldr(tmp1, Address(buf, 48));
3316 crc32x(crc, crc, tmp0);
3317 ldr(tmp2, Address(buf, 56));
3318 crc32x(crc, crc, tmp1);
3319 ldr(tmp3, Address(pre(buf, 64)));
3320
3321 b(CRC_by64_loop);
3322
3323 align(CodeEntryAlignment);
3324 BIND(CRC_by64_loop);
3325 subs(len, len, 64);
3326 crc32x(crc, crc, tmp2);
3327 ldr(tmp0, Address(buf, 8));
3328 crc32x(crc, crc, tmp3);
3329 ldr(tmp1, Address(buf, 16));
3330 crc32x(crc, crc, tmp0);
3331 ldr(tmp2, Address(buf, 24));
3332 crc32x(crc, crc, tmp1);
3333 ldr(tmp3, Address(buf, 32));
3334 crc32x(crc, crc, tmp2);
3335 ldr(tmp0, Address(buf, 40));
3336 crc32x(crc, crc, tmp3);
3337 ldr(tmp1, Address(buf, 48));
3338 crc32x(crc, crc, tmp0);
3339 ldr(tmp2, Address(buf, 56));
3340 crc32x(crc, crc, tmp1);
3341 ldr(tmp3, Address(pre(buf, 64)));
3342 br(Assembler::GE, CRC_by64_loop);
3343
3344 // post-loop
3345 crc32x(crc, crc, tmp2);
3346 crc32x(crc, crc, tmp3);
3347
3348 sub(len, len, 64);
3349 add(buf, buf, 8);
3350 cmn(len, 128);
3351 br(Assembler::NE, CRC_less64);
3352 BIND(L_exit);
3353 mvnw(crc, crc);
3354 }
3355
3356 /**
3357 * @param crc register containing existing CRC (32-bit)
3358 * @param buf register pointing to input byte buffer (byte*)
3359 * @param len register containing number of bytes
3360 * @param table register that will contain address of CRC table
3361 * @param tmp scratch register
3362 */
3363 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len,
3364 Register table0, Register table1, Register table2, Register table3,
3365 Register tmp, Register tmp2, Register tmp3) {
3366 Label L_by16, L_by16_loop, L_by4, L_by4_loop, L_by1, L_by1_loop, L_exit;
3367 uint64_t offset;
3368
3369 if (UseCRC32) {
3370 kernel_crc32_using_crc32(crc, buf, len, table0, table1, table2, table3);
3371 return;
3372 }
3373
3374 mvnw(crc, crc);
3375
3376 adrp(table0, ExternalAddress(StubRoutines::crc_table_addr()), offset);
3377 if (offset) add(table0, table0, offset);
3378 add(table1, table0, 1*256*sizeof(juint));
3379 add(table2, table0, 2*256*sizeof(juint));
3380 add(table3, table0, 3*256*sizeof(juint));
3381
3382 if (UseNeon) {
3383 cmp(len, (u1)64);
3384 br(Assembler::LT, L_by16);
3385 eor(v16, T16B, v16, v16);
3386
3387 Label L_fold;
3388
3389 add(tmp, table0, 4*256*sizeof(juint)); // Point at the Neon constants
3390
3391 ld1(v0, v1, T2D, post(buf, 32));
3392 ld1r(v4, T2D, post(tmp, 8));
3393 ld1r(v5, T2D, post(tmp, 8));
3394 ld1r(v6, T2D, post(tmp, 8));
3395 ld1r(v7, T2D, post(tmp, 8));
3396 mov(v16, T4S, 0, crc);
3397
3398 eor(v0, T16B, v0, v16);
3399 sub(len, len, 64);
3400
3401 BIND(L_fold);
3402 pmull(v22, T8H, v0, v5, T8B);
3403 pmull(v20, T8H, v0, v7, T8B);
3404 pmull(v23, T8H, v0, v4, T8B);
3405 pmull(v21, T8H, v0, v6, T8B);
3406
3407 pmull2(v18, T8H, v0, v5, T16B);
3408 pmull2(v16, T8H, v0, v7, T16B);
3409 pmull2(v19, T8H, v0, v4, T16B);
3410 pmull2(v17, T8H, v0, v6, T16B);
3411
3412 uzp1(v24, T8H, v20, v22);
3413 uzp2(v25, T8H, v20, v22);
3414 eor(v20, T16B, v24, v25);
3415
3416 uzp1(v26, T8H, v16, v18);
3417 uzp2(v27, T8H, v16, v18);
3418 eor(v16, T16B, v26, v27);
3419
3420 ushll2(v22, T4S, v20, T8H, 8);
3421 ushll(v20, T4S, v20, T4H, 8);
3422
3423 ushll2(v18, T4S, v16, T8H, 8);
3424 ushll(v16, T4S, v16, T4H, 8);
3425
3426 eor(v22, T16B, v23, v22);
3427 eor(v18, T16B, v19, v18);
3428 eor(v20, T16B, v21, v20);
3429 eor(v16, T16B, v17, v16);
3430
3431 uzp1(v17, T2D, v16, v20);
3432 uzp2(v21, T2D, v16, v20);
3433 eor(v17, T16B, v17, v21);
3434
3435 ushll2(v20, T2D, v17, T4S, 16);
3436 ushll(v16, T2D, v17, T2S, 16);
3437
3438 eor(v20, T16B, v20, v22);
3439 eor(v16, T16B, v16, v18);
3440
3441 uzp1(v17, T2D, v20, v16);
3442 uzp2(v21, T2D, v20, v16);
3443 eor(v28, T16B, v17, v21);
3444
3445 pmull(v22, T8H, v1, v5, T8B);
3446 pmull(v20, T8H, v1, v7, T8B);
3447 pmull(v23, T8H, v1, v4, T8B);
3448 pmull(v21, T8H, v1, v6, T8B);
3449
3450 pmull2(v18, T8H, v1, v5, T16B);
3451 pmull2(v16, T8H, v1, v7, T16B);
3452 pmull2(v19, T8H, v1, v4, T16B);
3453 pmull2(v17, T8H, v1, v6, T16B);
3454
3455 ld1(v0, v1, T2D, post(buf, 32));
3456
3457 uzp1(v24, T8H, v20, v22);
3458 uzp2(v25, T8H, v20, v22);
3459 eor(v20, T16B, v24, v25);
3460
3461 uzp1(v26, T8H, v16, v18);
3462 uzp2(v27, T8H, v16, v18);
3463 eor(v16, T16B, v26, v27);
3464
3465 ushll2(v22, T4S, v20, T8H, 8);
3466 ushll(v20, T4S, v20, T4H, 8);
3467
3468 ushll2(v18, T4S, v16, T8H, 8);
3469 ushll(v16, T4S, v16, T4H, 8);
3470
3471 eor(v22, T16B, v23, v22);
3472 eor(v18, T16B, v19, v18);
3473 eor(v20, T16B, v21, v20);
3474 eor(v16, T16B, v17, v16);
3475
3476 uzp1(v17, T2D, v16, v20);
3477 uzp2(v21, T2D, v16, v20);
3478 eor(v16, T16B, v17, v21);
3479
3480 ushll2(v20, T2D, v16, T4S, 16);
3481 ushll(v16, T2D, v16, T2S, 16);
3482
3483 eor(v20, T16B, v22, v20);
3484 eor(v16, T16B, v16, v18);
3485
3486 uzp1(v17, T2D, v20, v16);
3487 uzp2(v21, T2D, v20, v16);
3488 eor(v20, T16B, v17, v21);
3489
3490 shl(v16, T2D, v28, 1);
3491 shl(v17, T2D, v20, 1);
3492
3493 eor(v0, T16B, v0, v16);
3494 eor(v1, T16B, v1, v17);
3495
3496 subs(len, len, 32);
3497 br(Assembler::GE, L_fold);
3498
3499 mov(crc, 0);
3500 mov(tmp, v0, T1D, 0);
3501 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false);
3502 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true);
3503 mov(tmp, v0, T1D, 1);
3504 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false);
3505 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true);
3506 mov(tmp, v1, T1D, 0);
3507 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false);
3508 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true);
3509 mov(tmp, v1, T1D, 1);
3510 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false);
3511 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true);
3512
3513 add(len, len, 32);
3514 }
3515
3516 BIND(L_by16);
3517 subs(len, len, 16);
3518 br(Assembler::GE, L_by16_loop);
3519 adds(len, len, 16-4);
3520 br(Assembler::GE, L_by4_loop);
3521 adds(len, len, 4);
3522 br(Assembler::GT, L_by1_loop);
3523 b(L_exit);
3524
3525 BIND(L_by4_loop);
3526 ldrw(tmp, Address(post(buf, 4)));
3527 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3);
3528 subs(len, len, 4);
3529 br(Assembler::GE, L_by4_loop);
3530 adds(len, len, 4);
3531 br(Assembler::LE, L_exit);
3532 BIND(L_by1_loop);
3533 subs(len, len, 1);
3534 ldrb(tmp, Address(post(buf, 1)));
3535 update_byte_crc32(crc, tmp, table0);
3536 br(Assembler::GT, L_by1_loop);
3537 b(L_exit);
3538
3539 align(CodeEntryAlignment);
3540 BIND(L_by16_loop);
3541 subs(len, len, 16);
3542 ldp(tmp, tmp3, Address(post(buf, 16)));
3543 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false);
3544 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true);
3545 update_word_crc32(crc, tmp3, tmp2, table0, table1, table2, table3, false);
3546 update_word_crc32(crc, tmp3, tmp2, table0, table1, table2, table3, true);
3547 br(Assembler::GE, L_by16_loop);
3548 adds(len, len, 16-4);
3549 br(Assembler::GE, L_by4_loop);
3550 adds(len, len, 4);
3551 br(Assembler::GT, L_by1_loop);
3552 BIND(L_exit);
3553 mvnw(crc, crc);
3554 }
3555
3556 void MacroAssembler::kernel_crc32c_using_crc32c(Register crc, Register buf,
3557 Register len, Register tmp0, Register tmp1, Register tmp2,
3558 Register tmp3) {
3559 Label CRC_by64_loop, CRC_by4_loop, CRC_by1_loop, CRC_less64, CRC_by64_pre, CRC_by32_loop, CRC_less32, L_exit;
3560 assert_different_registers(crc, buf, len, tmp0, tmp1, tmp2, tmp3);
3561
3562 subs(len, len, 128);
3563 br(Assembler::GE, CRC_by64_pre);
3564 BIND(CRC_less64);
3565 adds(len, len, 128-32);
3566 br(Assembler::GE, CRC_by32_loop);
3567 BIND(CRC_less32);
3568 adds(len, len, 32-4);
3569 br(Assembler::GE, CRC_by4_loop);
3570 adds(len, len, 4);
3571 br(Assembler::GT, CRC_by1_loop);
3572 b(L_exit);
3573
3574 BIND(CRC_by32_loop);
3575 ldp(tmp0, tmp1, Address(post(buf, 16)));
3576 subs(len, len, 32);
3577 crc32cx(crc, crc, tmp0);
3578 ldr(tmp2, Address(post(buf, 8)));
3579 crc32cx(crc, crc, tmp1);
3580 ldr(tmp3, Address(post(buf, 8)));
3581 crc32cx(crc, crc, tmp2);
3582 crc32cx(crc, crc, tmp3);
3583 br(Assembler::GE, CRC_by32_loop);
3584 cmn(len, 32);
3585 br(Assembler::NE, CRC_less32);
3586 b(L_exit);
3587
3588 BIND(CRC_by4_loop);
3589 ldrw(tmp0, Address(post(buf, 4)));
3590 subs(len, len, 4);
3591 crc32cw(crc, crc, tmp0);
3592 br(Assembler::GE, CRC_by4_loop);
3593 adds(len, len, 4);
3594 br(Assembler::LE, L_exit);
3595 BIND(CRC_by1_loop);
3596 ldrb(tmp0, Address(post(buf, 1)));
3597 subs(len, len, 1);
3598 crc32cb(crc, crc, tmp0);
3599 br(Assembler::GT, CRC_by1_loop);
3600 b(L_exit);
3601
3602 BIND(CRC_by64_pre);
3603 sub(buf, buf, 8);
3604 ldp(tmp0, tmp1, Address(buf, 8));
3605 crc32cx(crc, crc, tmp0);
3606 ldr(tmp2, Address(buf, 24));
3607 crc32cx(crc, crc, tmp1);
3608 ldr(tmp3, Address(buf, 32));
3609 crc32cx(crc, crc, tmp2);
3610 ldr(tmp0, Address(buf, 40));
3611 crc32cx(crc, crc, tmp3);
3612 ldr(tmp1, Address(buf, 48));
3613 crc32cx(crc, crc, tmp0);
3614 ldr(tmp2, Address(buf, 56));
3615 crc32cx(crc, crc, tmp1);
3616 ldr(tmp3, Address(pre(buf, 64)));
3617
3618 b(CRC_by64_loop);
3619
3620 align(CodeEntryAlignment);
3621 BIND(CRC_by64_loop);
3622 subs(len, len, 64);
3623 crc32cx(crc, crc, tmp2);
3624 ldr(tmp0, Address(buf, 8));
3625 crc32cx(crc, crc, tmp3);
3626 ldr(tmp1, Address(buf, 16));
3627 crc32cx(crc, crc, tmp0);
3628 ldr(tmp2, Address(buf, 24));
3629 crc32cx(crc, crc, tmp1);
3630 ldr(tmp3, Address(buf, 32));
3631 crc32cx(crc, crc, tmp2);
3632 ldr(tmp0, Address(buf, 40));
3633 crc32cx(crc, crc, tmp3);
3634 ldr(tmp1, Address(buf, 48));
3635 crc32cx(crc, crc, tmp0);
3636 ldr(tmp2, Address(buf, 56));
3637 crc32cx(crc, crc, tmp1);
3638 ldr(tmp3, Address(pre(buf, 64)));
3639 br(Assembler::GE, CRC_by64_loop);
3640
3641 // post-loop
3642 crc32cx(crc, crc, tmp2);
3643 crc32cx(crc, crc, tmp3);
3644
3645 sub(len, len, 64);
3646 add(buf, buf, 8);
3647 cmn(len, 128);
3648 br(Assembler::NE, CRC_less64);
3649 BIND(L_exit);
3650 }
3651
3652 /**
3653 * @param crc register containing existing CRC (32-bit)
3654 * @param buf register pointing to input byte buffer (byte*)
3655 * @param len register containing number of bytes
3656 * @param table register that will contain address of CRC table
3657 * @param tmp scratch register
3658 */
3659 void MacroAssembler::kernel_crc32c(Register crc, Register buf, Register len,
3660 Register table0, Register table1, Register table2, Register table3,
3661 Register tmp, Register tmp2, Register tmp3) {
3662 kernel_crc32c_using_crc32c(crc, buf, len, table0, table1, table2, table3);
3663 }
3664
3665
3666 SkipIfEqual::SkipIfEqual(
3667 MacroAssembler* masm, const bool* flag_addr, bool value) {
3668 _masm = masm;
3669 uint64_t offset;
3670 _masm->adrp(rscratch1, ExternalAddress((address)flag_addr), offset);
3671 _masm->ldrb(rscratch1, Address(rscratch1, offset));
3672 _masm->cbzw(rscratch1, _label);
3673 }
3674
3675 SkipIfEqual::~SkipIfEqual() {
3676 _masm->bind(_label);
3677 }
3678
3679 void MacroAssembler::addptr(const Address &dst, int32_t src) {
3680 Address adr;
3681 switch(dst.getMode()) {
3682 case Address::base_plus_offset:
3683 // This is the expected mode, although we allow all the other
3684 // forms below.
3685 adr = form_address(rscratch2, dst.base(), dst.offset(), LogBytesPerWord);
3686 break;
3687 default:
3688 lea(rscratch2, dst);
3689 adr = Address(rscratch2);
3690 break;
3691 }
3692 ldr(rscratch1, adr);
3693 add(rscratch1, rscratch1, src);
3694 str(rscratch1, adr);
3695 }
3696
3697 void MacroAssembler::cmpptr(Register src1, Address src2) {
3698 uint64_t offset;
3699 adrp(rscratch1, src2, offset);
3700 ldr(rscratch1, Address(rscratch1, offset));
3701 cmp(src1, rscratch1);
3702 }
3703
3704 void MacroAssembler::cmpoop(Register obj1, Register obj2) {
3705 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
3706 bs->obj_equals(this, obj1, obj2);
3707 }
3708
3709 void MacroAssembler::load_method_holder_cld(Register rresult, Register rmethod) {
3710 load_method_holder(rresult, rmethod);
3711 ldr(rresult, Address(rresult, InstanceKlass::class_loader_data_offset()));
3712 }
3713
3714 void MacroAssembler::load_method_holder(Register holder, Register method) {
3715 ldr(holder, Address(method, Method::const_offset())); // ConstMethod*
3716 ldr(holder, Address(holder, ConstMethod::constants_offset())); // ConstantPool*
3717 ldr(holder, Address(holder, ConstantPool::pool_holder_offset_in_bytes())); // InstanceKlass*
3718 }
3719
3720 void MacroAssembler::load_klass(Register dst, Register src) {
3721 if (UseCompressedClassPointers) {
3722 ldrw(dst, Address(src, oopDesc::klass_offset_in_bytes()));
3723 decode_klass_not_null(dst);
3724 } else {
3725 ldr(dst, Address(src, oopDesc::klass_offset_in_bytes()));
3726 }
3727 }
3728
3729 // ((OopHandle)result).resolve();
3730 void MacroAssembler::resolve_oop_handle(Register result, Register tmp) {
3731 // OopHandle::resolve is an indirection.
3732 access_load_at(T_OBJECT, IN_NATIVE, result, Address(result, 0), tmp, noreg);
3733 }
3734
3735 // ((WeakHandle)result).resolve();
3736 void MacroAssembler::resolve_weak_handle(Register rresult, Register rtmp) {
3737 assert_different_registers(rresult, rtmp);
3738 Label resolved;
3739
3740 // A null weak handle resolves to null.
3741 cbz(rresult, resolved);
3742
3743 // Only 64 bit platforms support GCs that require a tmp register
3744 // Only IN_HEAP loads require a thread_tmp register
3745 // WeakHandle::resolve is an indirection like jweak.
3746 access_load_at(T_OBJECT, IN_NATIVE | ON_PHANTOM_OOP_REF,
3747 rresult, Address(rresult), rtmp, /*tmp_thread*/noreg);
3748 bind(resolved);
3749 }
3750
3751 void MacroAssembler::load_mirror(Register dst, Register method, Register tmp) {
3752 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
3753 ldr(dst, Address(rmethod, Method::const_offset()));
3754 ldr(dst, Address(dst, ConstMethod::constants_offset()));
3755 ldr(dst, Address(dst, ConstantPool::pool_holder_offset_in_bytes()));
3756 ldr(dst, Address(dst, mirror_offset));
3757 resolve_oop_handle(dst, tmp);
3758 }
3759
3760 void MacroAssembler::cmp_klass(Register oop, Register trial_klass, Register tmp) {
3761 if (UseCompressedClassPointers) {
3762 ldrw(tmp, Address(oop, oopDesc::klass_offset_in_bytes()));
3763 if (CompressedKlassPointers::base() == NULL) {
3764 cmp(trial_klass, tmp, LSL, CompressedKlassPointers::shift());
3765 return;
3766 } else if (((uint64_t)CompressedKlassPointers::base() & 0xffffffff) == 0
3767 && CompressedKlassPointers::shift() == 0) {
3768 // Only the bottom 32 bits matter
3769 cmpw(trial_klass, tmp);
3770 return;
3771 }
3772 decode_klass_not_null(tmp);
3773 } else {
3774 ldr(tmp, Address(oop, oopDesc::klass_offset_in_bytes()));
3775 }
3776 cmp(trial_klass, tmp);
3777 }
3778
3779 void MacroAssembler::load_prototype_header(Register dst, Register src) {
3780 load_klass(dst, src);
3781 ldr(dst, Address(dst, Klass::prototype_header_offset()));
3782 }
3783
3784 void MacroAssembler::store_klass(Register dst, Register src) {
3785 // FIXME: Should this be a store release? concurrent gcs assumes
3786 // klass length is valid if klass field is not null.
3787 if (UseCompressedClassPointers) {
3788 encode_klass_not_null(src);
3789 strw(src, Address(dst, oopDesc::klass_offset_in_bytes()));
3790 } else {
3791 str(src, Address(dst, oopDesc::klass_offset_in_bytes()));
3792 }
3793 }
3794
3795 void MacroAssembler::store_klass_gap(Register dst, Register src) {
3796 if (UseCompressedClassPointers) {
3797 // Store to klass gap in destination
3798 strw(src, Address(dst, oopDesc::klass_gap_offset_in_bytes()));
3799 }
3800 }
3801
3802 // Algorithm must match CompressedOops::encode.
3803 void MacroAssembler::encode_heap_oop(Register d, Register s) {
3804 #ifdef ASSERT
3805 verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?");
3806 #endif
3807 verify_oop(s, "broken oop in encode_heap_oop");
3808 if (CompressedOops::base() == NULL) {
3809 if (CompressedOops::shift() != 0) {
3810 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
3811 lsr(d, s, LogMinObjAlignmentInBytes);
3812 } else {
3813 mov(d, s);
3814 }
3815 } else {
3816 subs(d, s, rheapbase);
3817 csel(d, d, zr, Assembler::HS);
3818 lsr(d, d, LogMinObjAlignmentInBytes);
3819
3820 /* Old algorithm: is this any worse?
3821 Label nonnull;
3822 cbnz(r, nonnull);
3823 sub(r, r, rheapbase);
3824 bind(nonnull);
3825 lsr(r, r, LogMinObjAlignmentInBytes);
3826 */
3827 }
3828 }
3829
3830 void MacroAssembler::encode_heap_oop_not_null(Register r) {
3831 #ifdef ASSERT
3832 verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?");
3833 if (CheckCompressedOops) {
3834 Label ok;
3835 cbnz(r, ok);
3836 stop("null oop passed to encode_heap_oop_not_null");
3837 bind(ok);
3838 }
3839 #endif
3840 verify_oop(r, "broken oop in encode_heap_oop_not_null");
3841 if (CompressedOops::base() != NULL) {
3842 sub(r, r, rheapbase);
3843 }
3844 if (CompressedOops::shift() != 0) {
3845 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
3846 lsr(r, r, LogMinObjAlignmentInBytes);
3847 }
3848 }
3849
3850 void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) {
3851 #ifdef ASSERT
3852 verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?");
3853 if (CheckCompressedOops) {
3854 Label ok;
3855 cbnz(src, ok);
3856 stop("null oop passed to encode_heap_oop_not_null2");
3857 bind(ok);
3858 }
3859 #endif
3860 verify_oop(src, "broken oop in encode_heap_oop_not_null2");
3861
3862 Register data = src;
3863 if (CompressedOops::base() != NULL) {
3864 sub(dst, src, rheapbase);
3865 data = dst;
3866 }
3867 if (CompressedOops::shift() != 0) {
3868 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
3869 lsr(dst, data, LogMinObjAlignmentInBytes);
3870 data = dst;
3871 }
3872 if (data == src)
3873 mov(dst, src);
3874 }
3875
3876 void MacroAssembler::decode_heap_oop(Register d, Register s) {
3877 #ifdef ASSERT
3878 verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?");
3879 #endif
3880 if (CompressedOops::base() == NULL) {
3881 if (CompressedOops::shift() != 0 || d != s) {
3882 lsl(d, s, CompressedOops::shift());
3883 }
3884 } else {
3885 Label done;
3886 if (d != s)
3887 mov(d, s);
3888 cbz(s, done);
3889 add(d, rheapbase, s, Assembler::LSL, LogMinObjAlignmentInBytes);
3890 bind(done);
3891 }
3892 verify_oop(d, "broken oop in decode_heap_oop");
3893 }
3894
3895 void MacroAssembler::decode_heap_oop_not_null(Register r) {
3896 assert (UseCompressedOops, "should only be used for compressed headers");
3897 assert (Universe::heap() != NULL, "java heap should be initialized");
3898 // Cannot assert, unverified entry point counts instructions (see .ad file)
3899 // vtableStubs also counts instructions in pd_code_size_limit.
3900 // Also do not verify_oop as this is called by verify_oop.
3901 if (CompressedOops::shift() != 0) {
3902 assert(LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
3903 if (CompressedOops::base() != NULL) {
3904 add(r, rheapbase, r, Assembler::LSL, LogMinObjAlignmentInBytes);
3905 } else {
3906 add(r, zr, r, Assembler::LSL, LogMinObjAlignmentInBytes);
3907 }
3908 } else {
3909 assert (CompressedOops::base() == NULL, "sanity");
3910 }
3911 }
3912
3913 void MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) {
3914 assert (UseCompressedOops, "should only be used for compressed headers");
3915 assert (Universe::heap() != NULL, "java heap should be initialized");
3916 // Cannot assert, unverified entry point counts instructions (see .ad file)
3917 // vtableStubs also counts instructions in pd_code_size_limit.
3918 // Also do not verify_oop as this is called by verify_oop.
3919 if (CompressedOops::shift() != 0) {
3920 assert(LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
3921 if (CompressedOops::base() != NULL) {
3922 add(dst, rheapbase, src, Assembler::LSL, LogMinObjAlignmentInBytes);
3923 } else {
3924 add(dst, zr, src, Assembler::LSL, LogMinObjAlignmentInBytes);
3925 }
3926 } else {
3927 assert (CompressedOops::base() == NULL, "sanity");
3928 if (dst != src) {
3929 mov(dst, src);
3930 }
3931 }
3932 }
3933
3934 MacroAssembler::KlassDecodeMode MacroAssembler::_klass_decode_mode(KlassDecodeNone);
3935
3936 MacroAssembler::KlassDecodeMode MacroAssembler::klass_decode_mode() {
3937 assert(UseCompressedClassPointers, "not using compressed class pointers");
3938 assert(Metaspace::initialized(), "metaspace not initialized yet");
3939
3940 if (_klass_decode_mode != KlassDecodeNone) {
3941 return _klass_decode_mode;
3942 }
3943
3944 assert(LogKlassAlignmentInBytes == CompressedKlassPointers::shift()
3945 || 0 == CompressedKlassPointers::shift(), "decode alg wrong");
3946
3947 if (CompressedKlassPointers::base() == NULL) {
3948 return (_klass_decode_mode = KlassDecodeZero);
3949 }
3950
3951 if (operand_valid_for_logical_immediate(
3952 /*is32*/false, (uint64_t)CompressedKlassPointers::base())) {
3953 const uint64_t range_mask =
3954 (1ULL << log2_intptr(CompressedKlassPointers::range())) - 1;
3955 if (((uint64_t)CompressedKlassPointers::base() & range_mask) == 0) {
3956 return (_klass_decode_mode = KlassDecodeXor);
3957 }
3958 }
3959
3960 const uint64_t shifted_base =
3961 (uint64_t)CompressedKlassPointers::base() >> CompressedKlassPointers::shift();
3962 guarantee((shifted_base & 0xffff0000ffffffff) == 0,
3963 "compressed class base bad alignment");
3964
3965 return (_klass_decode_mode = KlassDecodeMovk);
3966 }
3967
3968 void MacroAssembler::encode_klass_not_null(Register dst, Register src) {
3969 switch (klass_decode_mode()) {
3970 case KlassDecodeZero:
3971 if (CompressedKlassPointers::shift() != 0) {
3972 lsr(dst, src, LogKlassAlignmentInBytes);
3973 } else {
3974 if (dst != src) mov(dst, src);
3975 }
3976 break;
3977
3978 case KlassDecodeXor:
3979 if (CompressedKlassPointers::shift() != 0) {
3980 eor(dst, src, (uint64_t)CompressedKlassPointers::base());
3981 lsr(dst, dst, LogKlassAlignmentInBytes);
3982 } else {
3983 eor(dst, src, (uint64_t)CompressedKlassPointers::base());
3984 }
3985 break;
3986
3987 case KlassDecodeMovk:
3988 if (CompressedKlassPointers::shift() != 0) {
3989 ubfx(dst, src, LogKlassAlignmentInBytes, 32);
3990 } else {
3991 movw(dst, src);
3992 }
3993 break;
3994
3995 case KlassDecodeNone:
3996 ShouldNotReachHere();
3997 break;
3998 }
3999 }
4000
4001 void MacroAssembler::encode_klass_not_null(Register r) {
4002 encode_klass_not_null(r, r);
4003 }
4004
4005 void MacroAssembler::decode_klass_not_null(Register dst, Register src) {
4006 assert (UseCompressedClassPointers, "should only be used for compressed headers");
4007
4008 switch (klass_decode_mode()) {
4009 case KlassDecodeZero:
4010 if (CompressedKlassPointers::shift() != 0) {
4011 lsl(dst, src, LogKlassAlignmentInBytes);
4012 } else {
4013 if (dst != src) mov(dst, src);
4014 }
4015 break;
4016
4017 case KlassDecodeXor:
4018 if (CompressedKlassPointers::shift() != 0) {
4019 lsl(dst, src, LogKlassAlignmentInBytes);
4020 eor(dst, dst, (uint64_t)CompressedKlassPointers::base());
4021 } else {
4022 eor(dst, src, (uint64_t)CompressedKlassPointers::base());
4023 }
4024 break;
4025
4026 case KlassDecodeMovk: {
4027 const uint64_t shifted_base =
4028 (uint64_t)CompressedKlassPointers::base() >> CompressedKlassPointers::shift();
4029
4030 if (dst != src) movw(dst, src);
4031 movk(dst, shifted_base >> 32, 32);
4032
4033 if (CompressedKlassPointers::shift() != 0) {
4034 lsl(dst, dst, LogKlassAlignmentInBytes);
4035 }
4036
4037 break;
4038 }
4039
4040 case KlassDecodeNone:
4041 ShouldNotReachHere();
4042 break;
4043 }
4044 }
4045
4046 void MacroAssembler::decode_klass_not_null(Register r) {
4047 decode_klass_not_null(r, r);
4048 }
4049
4050 void MacroAssembler::set_narrow_oop(Register dst, jobject obj) {
4051 #ifdef ASSERT
4052 {
4053 ThreadInVMfromUnknown tiv;
4054 assert (UseCompressedOops, "should only be used for compressed oops");
4055 assert (Universe::heap() != NULL, "java heap should be initialized");
4056 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
4057 assert(Universe::heap()->is_in(JNIHandles::resolve(obj)), "should be real oop");
4058 }
4059 #endif
4060 int oop_index = oop_recorder()->find_index(obj);
4061 InstructionMark im(this);
4062 RelocationHolder rspec = oop_Relocation::spec(oop_index);
4063 code_section()->relocate(inst_mark(), rspec);
4064 movz(dst, 0xDEAD, 16);
4065 movk(dst, 0xBEEF);
4066 }
4067
4068 void MacroAssembler::set_narrow_klass(Register dst, Klass* k) {
4069 assert (UseCompressedClassPointers, "should only be used for compressed headers");
4070 assert (oop_recorder() != NULL, "this assembler needs an OopRecorder");
4071 int index = oop_recorder()->find_index(k);
4072 assert(! Universe::heap()->is_in(k), "should not be an oop");
4073
4074 InstructionMark im(this);
4075 RelocationHolder rspec = metadata_Relocation::spec(index);
4076 code_section()->relocate(inst_mark(), rspec);
4077 narrowKlass nk = CompressedKlassPointers::encode(k);
4078 movz(dst, (nk >> 16), 16);
4079 movk(dst, nk & 0xffff);
4080 }
4081
4082 void MacroAssembler::access_load_at(BasicType type, DecoratorSet decorators,
4083 Register dst, Address src,
4084 Register tmp1, Register thread_tmp) {
4085 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler();
4086 decorators = AccessInternal::decorator_fixup(decorators);
4087 bool as_raw = (decorators & AS_RAW) != 0;
4088 if (as_raw) {
4089 bs->BarrierSetAssembler::load_at(this, decorators, type, dst, src, tmp1, thread_tmp);
4090 } else {
4091 bs->load_at(this, decorators, type, dst, src, tmp1, thread_tmp);
4092 }
4093 }
4094
4095 void MacroAssembler::access_store_at(BasicType type, DecoratorSet decorators,
4096 Address dst, Register src,
4097 Register tmp1, Register thread_tmp) {
4098 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler();
4099 decorators = AccessInternal::decorator_fixup(decorators);
4100 bool as_raw = (decorators & AS_RAW) != 0;
4101 if (as_raw) {
4102 bs->BarrierSetAssembler::store_at(this, decorators, type, dst, src, tmp1, thread_tmp);
4103 } else {
4104 bs->store_at(this, decorators, type, dst, src, tmp1, thread_tmp);
4105 }
4106 }
4107
4108 void MacroAssembler::resolve(DecoratorSet decorators, Register obj) {
4109 // Use stronger ACCESS_WRITE|ACCESS_READ by default.
4110 if ((decorators & (ACCESS_READ | ACCESS_WRITE)) == 0) {
4111 decorators |= ACCESS_READ | ACCESS_WRITE;
4112 }
4113 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
4114 return bs->resolve(this, decorators, obj);
4115 }
4116
4117 void MacroAssembler::load_heap_oop(Register dst, Address src, Register tmp1,
4118 Register thread_tmp, DecoratorSet decorators) {
4119 access_load_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, thread_tmp);
4120 }
4121
4122 void MacroAssembler::load_heap_oop_not_null(Register dst, Address src, Register tmp1,
4123 Register thread_tmp, DecoratorSet decorators) {
4124 access_load_at(T_OBJECT, IN_HEAP | IS_NOT_NULL | decorators, dst, src, tmp1, thread_tmp);
4125 }
4126
4127 void MacroAssembler::store_heap_oop(Address dst, Register src, Register tmp1,
4128 Register thread_tmp, DecoratorSet decorators) {
4129 access_store_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, thread_tmp);
4130 }
4131
4132 // Used for storing NULLs.
4133 void MacroAssembler::store_heap_oop_null(Address dst) {
4134 access_store_at(T_OBJECT, IN_HEAP, dst, noreg, noreg, noreg);
4135 }
4136
4137 Address MacroAssembler::allocate_metadata_address(Metadata* obj) {
4138 assert(oop_recorder() != NULL, "this assembler needs a Recorder");
4139 int index = oop_recorder()->allocate_metadata_index(obj);
4140 RelocationHolder rspec = metadata_Relocation::spec(index);
4141 return Address((address)obj, rspec);
4142 }
4143
4144 // Move an oop into a register. immediate is true if we want
4145 // immediate instructions and nmethod entry barriers are not enabled.
4146 // i.e. we are not going to patch this instruction while the code is being
4147 // executed by another thread.
4148 void MacroAssembler::movoop(Register dst, jobject obj, bool immediate) {
4149 int oop_index;
4150 if (obj == NULL) {
4151 oop_index = oop_recorder()->allocate_oop_index(obj);
4152 } else {
4153 #ifdef ASSERT
4154 {
4155 ThreadInVMfromUnknown tiv;
4156 assert(Universe::heap()->is_in(JNIHandles::resolve(obj)), "should be real oop");
4157 }
4158 #endif
4159 oop_index = oop_recorder()->find_index(obj);
4160 }
4161 RelocationHolder rspec = oop_Relocation::spec(oop_index);
4162
4163 // nmethod entry barrier necessitate using the constant pool. They have to be
4164 // ordered with respected to oop accesses.
4165 // Using immediate literals would necessitate ISBs.
4166 if (BarrierSet::barrier_set()->barrier_set_nmethod() != NULL || !immediate) {
4167 address dummy = address(uintptr_t(pc()) & -wordSize); // A nearby aligned address
4168 ldr_constant(dst, Address(dummy, rspec));
4169 } else
4170 mov(dst, Address((address)obj, rspec));
4171
4172 }
4173
4174 // Move a metadata address into a register.
4175 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) {
4176 int oop_index;
4177 if (obj == NULL) {
4178 oop_index = oop_recorder()->allocate_metadata_index(obj);
4179 } else {
4180 oop_index = oop_recorder()->find_index(obj);
4181 }
4182 RelocationHolder rspec = metadata_Relocation::spec(oop_index);
4183 mov(dst, Address((address)obj, rspec));
4184 }
4185
4186 Address MacroAssembler::constant_oop_address(jobject obj) {
4187 #ifdef ASSERT
4188 {
4189 ThreadInVMfromUnknown tiv;
4190 assert(oop_recorder() != NULL, "this assembler needs an OopRecorder");
4191 assert(Universe::heap()->is_in(JNIHandles::resolve(obj)), "not an oop");
4192 }
4193 #endif
4194 int oop_index = oop_recorder()->find_index(obj);
4195 return Address((address)obj, oop_Relocation::spec(oop_index));
4196 }
4197
4198 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes.
4199 void MacroAssembler::tlab_allocate(Register obj,
4200 Register var_size_in_bytes,
4201 int con_size_in_bytes,
4202 Register t1,
4203 Register t2,
4204 Label& slow_case) {
4205 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler();
4206 bs->tlab_allocate(this, obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case);
4207 }
4208
4209 // Defines obj, preserves var_size_in_bytes
4210 void MacroAssembler::eden_allocate(Register obj,
4211 Register var_size_in_bytes,
4212 int con_size_in_bytes,
4213 Register t1,
4214 Label& slow_case) {
4215 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler();
4216 bs->eden_allocate(this, obj, var_size_in_bytes, con_size_in_bytes, t1, slow_case);
4217 }
4218
4219 // Zero words; len is in bytes
4220 // Destroys all registers except addr
4221 // len must be a nonzero multiple of wordSize
4222 void MacroAssembler::zero_memory(Register addr, Register len, Register t1) {
4223 assert_different_registers(addr, len, t1, rscratch1, rscratch2);
4224
4225 #ifdef ASSERT
4226 { Label L;
4227 tst(len, BytesPerWord - 1);
4228 br(Assembler::EQ, L);
4229 stop("len is not a multiple of BytesPerWord");
4230 bind(L);
4231 }
4232 #endif
4233
4234 #ifndef PRODUCT
4235 block_comment("zero memory");
4236 #endif
4237
4238 Label loop;
4239 Label entry;
4240
4241 // Algorithm:
4242 //
4243 // scratch1 = cnt & 7;
4244 // cnt -= scratch1;
4245 // p += scratch1;
4246 // switch (scratch1) {
4247 // do {
4248 // cnt -= 8;
4249 // p[-8] = 0;
4250 // case 7:
4251 // p[-7] = 0;
4252 // case 6:
4253 // p[-6] = 0;
4254 // // ...
4255 // case 1:
4256 // p[-1] = 0;
4257 // case 0:
4258 // p += 8;
4259 // } while (cnt);
4260 // }
4261
4262 const int unroll = 8; // Number of str(zr) instructions we'll unroll
4263
4264 lsr(len, len, LogBytesPerWord);
4265 andr(rscratch1, len, unroll - 1); // tmp1 = cnt % unroll
4266 sub(len, len, rscratch1); // cnt -= unroll
4267 // t1 always points to the end of the region we're about to zero
4268 add(t1, addr, rscratch1, Assembler::LSL, LogBytesPerWord);
4269 adr(rscratch2, entry);
4270 sub(rscratch2, rscratch2, rscratch1, Assembler::LSL, 2);
4271 br(rscratch2);
4272 bind(loop);
4273 sub(len, len, unroll);
4274 for (int i = -unroll; i < 0; i++)
4275 Assembler::str(zr, Address(t1, i * wordSize));
4276 bind(entry);
4277 add(t1, t1, unroll * wordSize);
4278 cbnz(len, loop);
4279 }
4280
4281 void MacroAssembler::verify_tlab() {
4282 #ifdef ASSERT
4283 if (UseTLAB && VerifyOops) {
4284 Label next, ok;
4285
4286 stp(rscratch2, rscratch1, Address(pre(sp, -16)));
4287
4288 ldr(rscratch2, Address(rthread, in_bytes(JavaThread::tlab_top_offset())));
4289 ldr(rscratch1, Address(rthread, in_bytes(JavaThread::tlab_start_offset())));
4290 cmp(rscratch2, rscratch1);
4291 br(Assembler::HS, next);
4292 STOP("assert(top >= start)");
4293 should_not_reach_here();
4294
4295 bind(next);
4296 ldr(rscratch2, Address(rthread, in_bytes(JavaThread::tlab_end_offset())));
4297 ldr(rscratch1, Address(rthread, in_bytes(JavaThread::tlab_top_offset())));
4298 cmp(rscratch2, rscratch1);
4299 br(Assembler::HS, ok);
4300 STOP("assert(top <= end)");
4301 should_not_reach_here();
4302
4303 bind(ok);
4304 ldp(rscratch2, rscratch1, Address(post(sp, 16)));
4305 }
4306 #endif
4307 }
4308
4309 // Writes to stack successive pages until offset reached to check for
4310 // stack overflow + shadow pages. This clobbers tmp.
4311 void MacroAssembler::bang_stack_size(Register size, Register tmp) {
4312 assert_different_registers(tmp, size, rscratch1);
4313 mov(tmp, sp);
4314 // Bang stack for total size given plus shadow page size.
4315 // Bang one page at a time because large size can bang beyond yellow and
4316 // red zones.
4317 Label loop;
4318 mov(rscratch1, os::vm_page_size());
4319 bind(loop);
4320 lea(tmp, Address(tmp, -os::vm_page_size()));
4321 subsw(size, size, rscratch1);
4322 str(size, Address(tmp));
4323 br(Assembler::GT, loop);
4324
4325 // Bang down shadow pages too.
4326 // At this point, (tmp-0) is the last address touched, so don't
4327 // touch it again. (It was touched as (tmp-pagesize) but then tmp
4328 // was post-decremented.) Skip this address by starting at i=1, and
4329 // touch a few more pages below. N.B. It is important to touch all
4330 // the way down to and including i=StackShadowPages.
4331 for (int i = 0; i < (int)(JavaThread::stack_shadow_zone_size() / os::vm_page_size()) - 1; i++) {
4332 // this could be any sized move but this is can be a debugging crumb
4333 // so the bigger the better.
4334 lea(tmp, Address(tmp, -os::vm_page_size()));
4335 str(size, Address(tmp));
4336 }
4337 }
4338
4339 // Move the address of the polling page into dest.
4340 void MacroAssembler::get_polling_page(Register dest, relocInfo::relocType rtype) {
4341 ldr(dest, Address(rthread, Thread::polling_page_offset()));
4342 }
4343
4344 // Move the address of the polling page into r, then read the polling
4345 // page.
4346 address MacroAssembler::fetch_and_read_polling_page(Register r, relocInfo::relocType rtype) {
4347 get_polling_page(r, rtype);
4348 return read_polling_page(r, rtype);
4349 }
4350
4351 // Read the polling page. The address of the polling page must
4352 // already be in r.
4353 address MacroAssembler::read_polling_page(Register r, relocInfo::relocType rtype) {
4354 InstructionMark im(this);
4355 code_section()->relocate(inst_mark(), rtype);
4356 ldrw(zr, Address(r, 0));
4357 return inst_mark();
4358 }
4359
4360 void MacroAssembler::adrp(Register reg1, const Address &dest, uint64_t &byte_offset) {
4361 relocInfo::relocType rtype = dest.rspec().reloc()->type();
4362 uint64_t low_page = (uint64_t)CodeCache::low_bound() >> 12;
4363 uint64_t high_page = (uint64_t)(CodeCache::high_bound()-1) >> 12;
4364 uint64_t dest_page = (uint64_t)dest.target() >> 12;
4365 int64_t offset_low = dest_page - low_page;
4366 int64_t offset_high = dest_page - high_page;
4367
4368 assert(is_valid_AArch64_address(dest.target()), "bad address");
4369 assert(dest.getMode() == Address::literal, "ADRP must be applied to a literal address");
4370
4371 InstructionMark im(this);
4372 code_section()->relocate(inst_mark(), dest.rspec());
4373 // 8143067: Ensure that the adrp can reach the dest from anywhere within
4374 // the code cache so that if it is relocated we know it will still reach
4375 if (offset_high >= -(1<<20) && offset_low < (1<<20)) {
4376 _adrp(reg1, dest.target());
4377 } else {
4378 uint64_t target = (uint64_t)dest.target();
4379 uint64_t adrp_target
4380 = (target & 0xffffffffULL) | ((uint64_t)pc() & 0xffff00000000ULL);
4381
4382 _adrp(reg1, (address)adrp_target);
4383 movk(reg1, target >> 32, 32);
4384 }
4385 byte_offset = (uint64_t)dest.target() & 0xfff;
4386 }
4387
4388 void MacroAssembler::load_byte_map_base(Register reg) {
4389 CardTable::CardValue* byte_map_base =
4390 ((CardTableBarrierSet*)(BarrierSet::barrier_set()))->card_table()->byte_map_base();
4391
4392 if (is_valid_AArch64_address((address)byte_map_base)) {
4393 // Strictly speaking the byte_map_base isn't an address at all,
4394 // and it might even be negative.
4395 uint64_t offset;
4396 adrp(reg, ExternalAddress((address)byte_map_base), offset);
4397 // We expect offset to be zero with most collectors.
4398 if (offset != 0) {
4399 add(reg, reg, offset);
4400 }
4401 } else {
4402 mov(reg, (uint64_t)byte_map_base);
4403 }
4404 }
4405
4406 void MacroAssembler::build_frame(int framesize) {
4407 assert(framesize > 0, "framesize must be > 0");
4408 if (framesize < ((1 << 9) + 2 * wordSize)) {
4409 sub(sp, sp, framesize);
4410 stp(rfp, lr, Address(sp, framesize - 2 * wordSize));
4411 if (PreserveFramePointer) add(rfp, sp, framesize - 2 * wordSize);
4412 } else {
4413 stp(rfp, lr, Address(pre(sp, -2 * wordSize)));
4414 if (PreserveFramePointer) mov(rfp, sp);
4415 if (framesize < ((1 << 12) + 2 * wordSize))
4416 sub(sp, sp, framesize - 2 * wordSize);
4417 else {
4418 mov(rscratch1, framesize - 2 * wordSize);
4419 sub(sp, sp, rscratch1);
4420 }
4421 }
4422 }
4423
4424 void MacroAssembler::remove_frame(int framesize) {
4425 assert(framesize > 0, "framesize must be > 0");
4426 if (framesize < ((1 << 9) + 2 * wordSize)) {
4427 ldp(rfp, lr, Address(sp, framesize - 2 * wordSize));
4428 add(sp, sp, framesize);
4429 } else {
4430 if (framesize < ((1 << 12) + 2 * wordSize))
4431 add(sp, sp, framesize - 2 * wordSize);
4432 else {
4433 mov(rscratch1, framesize - 2 * wordSize);
4434 add(sp, sp, rscratch1);
4435 }
4436 ldp(rfp, lr, Address(post(sp, 2 * wordSize)));
4437 }
4438 }
4439
4440
4441 // This method checks if provided byte array contains byte with highest bit set.
4442 void MacroAssembler::has_negatives(Register ary1, Register len, Register result) {
4443 // Simple and most common case of aligned small array which is not at the
4444 // end of memory page is placed here. All other cases are in stub.
4445 Label LOOP, END, STUB, STUB_LONG, SET_RESULT, DONE;
4446 const uint64_t UPPER_BIT_MASK=0x8080808080808080;
4447 assert_different_registers(ary1, len, result);
4448
4449 cmpw(len, 0);
4450 br(LE, SET_RESULT);
4451 cmpw(len, 4 * wordSize);
4452 br(GE, STUB_LONG); // size > 32 then go to stub
4453
4454 int shift = 64 - exact_log2(os::vm_page_size());
4455 lsl(rscratch1, ary1, shift);
4456 mov(rscratch2, (size_t)(4 * wordSize) << shift);
4457 adds(rscratch2, rscratch1, rscratch2); // At end of page?
4458 br(CS, STUB); // at the end of page then go to stub
4459 subs(len, len, wordSize);
4460 br(LT, END);
4461
4462 BIND(LOOP);
4463 ldr(rscratch1, Address(post(ary1, wordSize)));
4464 tst(rscratch1, UPPER_BIT_MASK);
4465 br(NE, SET_RESULT);
4466 subs(len, len, wordSize);
4467 br(GE, LOOP);
4468 cmpw(len, -wordSize);
4469 br(EQ, SET_RESULT);
4470
4471 BIND(END);
4472 ldr(result, Address(ary1));
4473 sub(len, zr, len, LSL, 3); // LSL 3 is to get bits from bytes
4474 lslv(result, result, len);
4475 tst(result, UPPER_BIT_MASK);
4476 b(SET_RESULT);
4477
4478 BIND(STUB);
4479 RuntimeAddress has_neg = RuntimeAddress(StubRoutines::aarch64::has_negatives());
4480 assert(has_neg.target() != NULL, "has_negatives stub has not been generated");
4481 trampoline_call(has_neg);
4482 b(DONE);
4483
4484 BIND(STUB_LONG);
4485 RuntimeAddress has_neg_long = RuntimeAddress(
4486 StubRoutines::aarch64::has_negatives_long());
4487 assert(has_neg_long.target() != NULL, "has_negatives stub has not been generated");
4488 trampoline_call(has_neg_long);
4489 b(DONE);
4490
4491 BIND(SET_RESULT);
4492 cset(result, NE); // set true or false
4493
4494 BIND(DONE);
4495 }
4496
4497 void MacroAssembler::arrays_equals(Register a1, Register a2, Register tmp3,
4498 Register tmp4, Register tmp5, Register result,
4499 Register cnt1, int elem_size) {
4500 Label DONE, SAME;
4501 Register tmp1 = rscratch1;
4502 Register tmp2 = rscratch2;
4503 Register cnt2 = tmp2; // cnt2 only used in array length compare
4504 int elem_per_word = wordSize/elem_size;
4505 int log_elem_size = exact_log2(elem_size);
4506 int length_offset = arrayOopDesc::length_offset_in_bytes();
4507 int base_offset
4508 = arrayOopDesc::base_offset_in_bytes(elem_size == 2 ? T_CHAR : T_BYTE);
4509 int stubBytesThreshold = 3 * 64 + (UseSIMDForArrayEquals ? 0 : 16);
4510
4511 assert(elem_size == 1 || elem_size == 2, "must be char or byte");
4512 assert_different_registers(a1, a2, result, cnt1, rscratch1, rscratch2);
4513
4514 #ifndef PRODUCT
4515 {
4516 const char kind = (elem_size == 2) ? 'U' : 'L';
4517 char comment[64];
4518 snprintf(comment, sizeof comment, "array_equals%c{", kind);
4519 BLOCK_COMMENT(comment);
4520 }
4521 #endif
4522
4523 // if (a1 == a2)
4524 // return true;
4525 cmpoop(a1, a2); // May have read barriers for a1 and a2.
4526 br(EQ, SAME);
4527
4528 if (UseSimpleArrayEquals) {
4529 Label NEXT_WORD, SHORT, TAIL03, TAIL01, A_MIGHT_BE_NULL, A_IS_NOT_NULL;
4530 // if (a1 == null || a2 == null)
4531 // return false;
4532 // a1 & a2 == 0 means (some-pointer is null) or
4533 // (very-rare-or-even-probably-impossible-pointer-values)
4534 // so, we can save one branch in most cases
4535 tst(a1, a2);
4536 mov(result, false);
4537 br(EQ, A_MIGHT_BE_NULL);
4538 // if (a1.length != a2.length)
4539 // return false;
4540 bind(A_IS_NOT_NULL);
4541 ldrw(cnt1, Address(a1, length_offset));
4542 ldrw(cnt2, Address(a2, length_offset));
4543 eorw(tmp5, cnt1, cnt2);
4544 cbnzw(tmp5, DONE);
4545 lea(a1, Address(a1, base_offset));
4546 lea(a2, Address(a2, base_offset));
4547 // Check for short strings, i.e. smaller than wordSize.
4548 subs(cnt1, cnt1, elem_per_word);
4549 br(Assembler::LT, SHORT);
4550 // Main 8 byte comparison loop.
4551 bind(NEXT_WORD); {
4552 ldr(tmp1, Address(post(a1, wordSize)));
4553 ldr(tmp2, Address(post(a2, wordSize)));
4554 subs(cnt1, cnt1, elem_per_word);
4555 eor(tmp5, tmp1, tmp2);
4556 cbnz(tmp5, DONE);
4557 } br(GT, NEXT_WORD);
4558 // Last longword. In the case where length == 4 we compare the
4559 // same longword twice, but that's still faster than another
4560 // conditional branch.
4561 // cnt1 could be 0, -1, -2, -3, -4 for chars; -4 only happens when
4562 // length == 4.
4563 if (log_elem_size > 0)
4564 lsl(cnt1, cnt1, log_elem_size);
4565 ldr(tmp3, Address(a1, cnt1));
4566 ldr(tmp4, Address(a2, cnt1));
4567 eor(tmp5, tmp3, tmp4);
4568 cbnz(tmp5, DONE);
4569 b(SAME);
4570 bind(A_MIGHT_BE_NULL);
4571 // in case both a1 and a2 are not-null, proceed with loads
4572 cbz(a1, DONE);
4573 cbz(a2, DONE);
4574 b(A_IS_NOT_NULL);
4575 bind(SHORT);
4576
4577 tbz(cnt1, 2 - log_elem_size, TAIL03); // 0-7 bytes left.
4578 {
4579 ldrw(tmp1, Address(post(a1, 4)));
4580 ldrw(tmp2, Address(post(a2, 4)));
4581 eorw(tmp5, tmp1, tmp2);
4582 cbnzw(tmp5, DONE);
4583 }
4584 bind(TAIL03);
4585 tbz(cnt1, 1 - log_elem_size, TAIL01); // 0-3 bytes left.
4586 {
4587 ldrh(tmp3, Address(post(a1, 2)));
4588 ldrh(tmp4, Address(post(a2, 2)));
4589 eorw(tmp5, tmp3, tmp4);
4590 cbnzw(tmp5, DONE);
4591 }
4592 bind(TAIL01);
4593 if (elem_size == 1) { // Only needed when comparing byte arrays.
4594 tbz(cnt1, 0, SAME); // 0-1 bytes left.
4595 {
4596 ldrb(tmp1, a1);
4597 ldrb(tmp2, a2);
4598 eorw(tmp5, tmp1, tmp2);
4599 cbnzw(tmp5, DONE);
4600 }
4601 }
4602 } else {
4603 Label NEXT_DWORD, SHORT, TAIL, TAIL2, STUB, EARLY_OUT,
4604 CSET_EQ, LAST_CHECK;
4605 mov(result, false);
4606 cbz(a1, DONE);
4607 ldrw(cnt1, Address(a1, length_offset));
4608 cbz(a2, DONE);
4609 ldrw(cnt2, Address(a2, length_offset));
4610 // on most CPUs a2 is still "locked"(surprisingly) in ldrw and it's
4611 // faster to perform another branch before comparing a1 and a2
4612 cmp(cnt1, (u1)elem_per_word);
4613 br(LE, SHORT); // short or same
4614 ldr(tmp3, Address(pre(a1, base_offset)));
4615 subs(zr, cnt1, stubBytesThreshold);
4616 br(GE, STUB);
4617 ldr(tmp4, Address(pre(a2, base_offset)));
4618 sub(tmp5, zr, cnt1, LSL, 3 + log_elem_size);
4619 cmp(cnt2, cnt1);
4620 br(NE, DONE);
4621
4622 // Main 16 byte comparison loop with 2 exits
4623 bind(NEXT_DWORD); {
4624 ldr(tmp1, Address(pre(a1, wordSize)));
4625 ldr(tmp2, Address(pre(a2, wordSize)));
4626 subs(cnt1, cnt1, 2 * elem_per_word);
4627 br(LE, TAIL);
4628 eor(tmp4, tmp3, tmp4);
4629 cbnz(tmp4, DONE);
4630 ldr(tmp3, Address(pre(a1, wordSize)));
4631 ldr(tmp4, Address(pre(a2, wordSize)));
4632 cmp(cnt1, (u1)elem_per_word);
4633 br(LE, TAIL2);
4634 cmp(tmp1, tmp2);
4635 } br(EQ, NEXT_DWORD);
4636 b(DONE);
4637
4638 bind(TAIL);
4639 eor(tmp4, tmp3, tmp4);
4640 eor(tmp2, tmp1, tmp2);
4641 lslv(tmp2, tmp2, tmp5);
4642 orr(tmp5, tmp4, tmp2);
4643 cmp(tmp5, zr);
4644 b(CSET_EQ);
4645
4646 bind(TAIL2);
4647 eor(tmp2, tmp1, tmp2);
4648 cbnz(tmp2, DONE);
4649 b(LAST_CHECK);
4650
4651 bind(STUB);
4652 ldr(tmp4, Address(pre(a2, base_offset)));
4653 cmp(cnt2, cnt1);
4654 br(NE, DONE);
4655 if (elem_size == 2) { // convert to byte counter
4656 lsl(cnt1, cnt1, 1);
4657 }
4658 eor(tmp5, tmp3, tmp4);
4659 cbnz(tmp5, DONE);
4660 RuntimeAddress stub = RuntimeAddress(StubRoutines::aarch64::large_array_equals());
4661 assert(stub.target() != NULL, "array_equals_long stub has not been generated");
4662 trampoline_call(stub);
4663 b(DONE);
4664
4665 bind(EARLY_OUT);
4666 // (a1 != null && a2 == null) || (a1 != null && a2 != null && a1 == a2)
4667 // so, if a2 == null => return false(0), else return true, so we can return a2
4668 mov(result, a2);
4669 b(DONE);
4670 bind(SHORT);
4671 cmp(cnt2, cnt1);
4672 br(NE, DONE);
4673 cbz(cnt1, SAME);
4674 sub(tmp5, zr, cnt1, LSL, 3 + log_elem_size);
4675 ldr(tmp3, Address(a1, base_offset));
4676 ldr(tmp4, Address(a2, base_offset));
4677 bind(LAST_CHECK);
4678 eor(tmp4, tmp3, tmp4);
4679 lslv(tmp5, tmp4, tmp5);
4680 cmp(tmp5, zr);
4681 bind(CSET_EQ);
4682 cset(result, EQ);
4683 b(DONE);
4684 }
4685
4686 bind(SAME);
4687 mov(result, true);
4688 // That's it.
4689 bind(DONE);
4690
4691 BLOCK_COMMENT("} array_equals");
4692 }
4693
4694 // Compare Strings
4695
4696 // For Strings we're passed the address of the first characters in a1
4697 // and a2 and the length in cnt1.
4698 // elem_size is the element size in bytes: either 1 or 2.
4699 // There are two implementations. For arrays >= 8 bytes, all
4700 // comparisons (including the final one, which may overlap) are
4701 // performed 8 bytes at a time. For strings < 8 bytes, we compare a
4702 // halfword, then a short, and then a byte.
4703
4704 void MacroAssembler::string_equals(Register a1, Register a2,
4705 Register result, Register cnt1, int elem_size)
4706 {
4707 Label SAME, DONE, SHORT, NEXT_WORD;
4708 Register tmp1 = rscratch1;
4709 Register tmp2 = rscratch2;
4710 Register cnt2 = tmp2; // cnt2 only used in array length compare
4711
4712 assert(elem_size == 1 || elem_size == 2, "must be 2 or 1 byte");
4713 assert_different_registers(a1, a2, result, cnt1, rscratch1, rscratch2);
4714
4715 #ifndef PRODUCT
4716 {
4717 const char kind = (elem_size == 2) ? 'U' : 'L';
4718 char comment[64];
4719 snprintf(comment, sizeof comment, "{string_equals%c", kind);
4720 BLOCK_COMMENT(comment);
4721 }
4722 #endif
4723
4724 mov(result, false);
4725
4726 // Check for short strings, i.e. smaller than wordSize.
4727 subs(cnt1, cnt1, wordSize);
4728 br(Assembler::LT, SHORT);
4729 // Main 8 byte comparison loop.
4730 bind(NEXT_WORD); {
4731 ldr(tmp1, Address(post(a1, wordSize)));
4732 ldr(tmp2, Address(post(a2, wordSize)));
4733 subs(cnt1, cnt1, wordSize);
4734 eor(tmp1, tmp1, tmp2);
4735 cbnz(tmp1, DONE);
4736 } br(GT, NEXT_WORD);
4737 // Last longword. In the case where length == 4 we compare the
4738 // same longword twice, but that's still faster than another
4739 // conditional branch.
4740 // cnt1 could be 0, -1, -2, -3, -4 for chars; -4 only happens when
4741 // length == 4.
4742 ldr(tmp1, Address(a1, cnt1));
4743 ldr(tmp2, Address(a2, cnt1));
4744 eor(tmp2, tmp1, tmp2);
4745 cbnz(tmp2, DONE);
4746 b(SAME);
4747
4748 bind(SHORT);
4749 Label TAIL03, TAIL01;
4750
4751 tbz(cnt1, 2, TAIL03); // 0-7 bytes left.
4752 {
4753 ldrw(tmp1, Address(post(a1, 4)));
4754 ldrw(tmp2, Address(post(a2, 4)));
4755 eorw(tmp1, tmp1, tmp2);
4756 cbnzw(tmp1, DONE);
4757 }
4758 bind(TAIL03);
4759 tbz(cnt1, 1, TAIL01); // 0-3 bytes left.
4760 {
4761 ldrh(tmp1, Address(post(a1, 2)));
4762 ldrh(tmp2, Address(post(a2, 2)));
4763 eorw(tmp1, tmp1, tmp2);
4764 cbnzw(tmp1, DONE);
4765 }
4766 bind(TAIL01);
4767 if (elem_size == 1) { // Only needed when comparing 1-byte elements
4768 tbz(cnt1, 0, SAME); // 0-1 bytes left.
4769 {
4770 ldrb(tmp1, a1);
4771 ldrb(tmp2, a2);
4772 eorw(tmp1, tmp1, tmp2);
4773 cbnzw(tmp1, DONE);
4774 }
4775 }
4776 // Arrays are equal.
4777 bind(SAME);
4778 mov(result, true);
4779
4780 // That's it.
4781 bind(DONE);
4782 BLOCK_COMMENT("} string_equals");
4783 }
4784
4785
4786 // The size of the blocks erased by the zero_blocks stub. We must
4787 // handle anything smaller than this ourselves in zero_words().
4788 const int MacroAssembler::zero_words_block_size = 8;
4789
4790 // zero_words() is used by C2 ClearArray patterns. It is as small as
4791 // possible, handling small word counts locally and delegating
4792 // anything larger to the zero_blocks stub. It is expanded many times
4793 // in compiled code, so it is important to keep it short.
4794
4795 // ptr: Address of a buffer to be zeroed.
4796 // cnt: Count in HeapWords.
4797 //
4798 // ptr, cnt, rscratch1, and rscratch2 are clobbered.
4799 void MacroAssembler::zero_words(Register ptr, Register cnt)
4800 {
4801 assert(is_power_of_2(zero_words_block_size), "adjust this");
4802 assert(ptr == r10 && cnt == r11, "mismatch in register usage");
4803
4804 BLOCK_COMMENT("zero_words {");
4805 cmp(cnt, (u1)zero_words_block_size);
4806 Label around;
4807 br(LO, around);
4808 {
4809 RuntimeAddress zero_blocks = RuntimeAddress(StubRoutines::aarch64::zero_blocks());
4810 assert(zero_blocks.target() != NULL, "zero_blocks stub has not been generated");
4811 if (StubRoutines::aarch64::complete()) {
4812 trampoline_call(zero_blocks);
4813 } else {
4814 bl(zero_blocks);
4815 }
4816 }
4817 bind(around);
4818 for (int i = zero_words_block_size >> 1; i > 1; i >>= 1) {
4819 Label l;
4820 tbz(cnt, exact_log2(i), l);
4821 for (int j = 0; j < i; j += 2) {
4822 stp(zr, zr, post(ptr, 16));
4823 }
4824 bind(l);
4825 }
4826 {
4827 Label l;
4828 tbz(cnt, 0, l);
4829 str(zr, Address(ptr));
4830 bind(l);
4831 }
4832 BLOCK_COMMENT("} zero_words");
4833 }
4834
4835 // base: Address of a buffer to be zeroed, 8 bytes aligned.
4836 // cnt: Immediate count in HeapWords.
4837 #define SmallArraySize (18 * BytesPerLong)
4838 void MacroAssembler::zero_words(Register base, uint64_t cnt)
4839 {
4840 BLOCK_COMMENT("zero_words {");
4841 int i = cnt & 1; // store any odd word to start
4842 if (i) str(zr, Address(base));
4843
4844 if (cnt <= SmallArraySize / BytesPerLong) {
4845 for (; i < (int)cnt; i += 2)
4846 stp(zr, zr, Address(base, i * wordSize));
4847 } else {
4848 const int unroll = 4; // Number of stp(zr, zr) instructions we'll unroll
4849 int remainder = cnt % (2 * unroll);
4850 for (; i < remainder; i += 2)
4851 stp(zr, zr, Address(base, i * wordSize));
4852
4853 Label loop;
4854 Register cnt_reg = rscratch1;
4855 Register loop_base = rscratch2;
4856 cnt = cnt - remainder;
4857 mov(cnt_reg, cnt);
4858 // adjust base and prebias by -2 * wordSize so we can pre-increment
4859 add(loop_base, base, (remainder - 2) * wordSize);
4860 bind(loop);
4861 sub(cnt_reg, cnt_reg, 2 * unroll);
4862 for (i = 1; i < unroll; i++)
4863 stp(zr, zr, Address(loop_base, 2 * i * wordSize));
4864 stp(zr, zr, Address(pre(loop_base, 2 * unroll * wordSize)));
4865 cbnz(cnt_reg, loop);
4866 }
4867 BLOCK_COMMENT("} zero_words");
4868 }
4869
4870 // Zero blocks of memory by using DC ZVA.
4871 //
4872 // Aligns the base address first sufficently for DC ZVA, then uses
4873 // DC ZVA repeatedly for every full block. cnt is the size to be
4874 // zeroed in HeapWords. Returns the count of words left to be zeroed
4875 // in cnt.
4876 //
4877 // NOTE: This is intended to be used in the zero_blocks() stub. If
4878 // you want to use it elsewhere, note that cnt must be >= 2*zva_length.
4879 void MacroAssembler::zero_dcache_blocks(Register base, Register cnt) {
4880 Register tmp = rscratch1;
4881 Register tmp2 = rscratch2;
4882 int zva_length = VM_Version::zva_length();
4883 Label initial_table_end, loop_zva;
4884 Label fini;
4885
4886 // Base must be 16 byte aligned. If not just return and let caller handle it
4887 tst(base, 0x0f);
4888 br(Assembler::NE, fini);
4889 // Align base with ZVA length.
4890 neg(tmp, base);
4891 andr(tmp, tmp, zva_length - 1);
4892
4893 // tmp: the number of bytes to be filled to align the base with ZVA length.
4894 add(base, base, tmp);
4895 sub(cnt, cnt, tmp, Assembler::ASR, 3);
4896 adr(tmp2, initial_table_end);
4897 sub(tmp2, tmp2, tmp, Assembler::LSR, 2);
4898 br(tmp2);
4899
4900 for (int i = -zva_length + 16; i < 0; i += 16)
4901 stp(zr, zr, Address(base, i));
4902 bind(initial_table_end);
4903
4904 sub(cnt, cnt, zva_length >> 3);
4905 bind(loop_zva);
4906 dc(Assembler::ZVA, base);
4907 subs(cnt, cnt, zva_length >> 3);
4908 add(base, base, zva_length);
4909 br(Assembler::GE, loop_zva);
4910 add(cnt, cnt, zva_length >> 3); // count not zeroed by DC ZVA
4911 bind(fini);
4912 }
4913
4914 // base: Address of a buffer to be filled, 8 bytes aligned.
4915 // cnt: Count in 8-byte unit.
4916 // value: Value to be filled with.
4917 // base will point to the end of the buffer after filling.
4918 void MacroAssembler::fill_words(Register base, Register cnt, Register value)
4919 {
4920 // Algorithm:
4921 //
4922 // scratch1 = cnt & 7;
4923 // cnt -= scratch1;
4924 // p += scratch1;
4925 // switch (scratch1) {
4926 // do {
4927 // cnt -= 8;
4928 // p[-8] = v;
4929 // case 7:
4930 // p[-7] = v;
4931 // case 6:
4932 // p[-6] = v;
4933 // // ...
4934 // case 1:
4935 // p[-1] = v;
4936 // case 0:
4937 // p += 8;
4938 // } while (cnt);
4939 // }
4940
4941 assert_different_registers(base, cnt, value, rscratch1, rscratch2);
4942
4943 Label fini, skip, entry, loop;
4944 const int unroll = 8; // Number of stp instructions we'll unroll
4945
4946 cbz(cnt, fini);
4947 tbz(base, 3, skip);
4948 str(value, Address(post(base, 8)));
4949 sub(cnt, cnt, 1);
4950 bind(skip);
4951
4952 andr(rscratch1, cnt, (unroll-1) * 2);
4953 sub(cnt, cnt, rscratch1);
4954 add(base, base, rscratch1, Assembler::LSL, 3);
4955 adr(rscratch2, entry);
4956 sub(rscratch2, rscratch2, rscratch1, Assembler::LSL, 1);
4957 br(rscratch2);
4958
4959 bind(loop);
4960 add(base, base, unroll * 16);
4961 for (int i = -unroll; i < 0; i++)
4962 stp(value, value, Address(base, i * 16));
4963 bind(entry);
4964 subs(cnt, cnt, unroll * 2);
4965 br(Assembler::GE, loop);
4966
4967 tbz(cnt, 0, fini);
4968 str(value, Address(post(base, 8)));
4969 bind(fini);
4970 }
4971
4972 // Intrinsic for sun/nio/cs/ISO_8859_1$Encoder.implEncodeISOArray and
4973 // java/lang/StringUTF16.compress.
4974 void MacroAssembler::encode_iso_array(Register src, Register dst,
4975 Register len, Register result,
4976 FloatRegister Vtmp1, FloatRegister Vtmp2,
4977 FloatRegister Vtmp3, FloatRegister Vtmp4)
4978 {
4979 Label DONE, SET_RESULT, NEXT_32, NEXT_32_PRFM, LOOP_8, NEXT_8, LOOP_1, NEXT_1,
4980 NEXT_32_START, NEXT_32_PRFM_START;
4981 Register tmp1 = rscratch1, tmp2 = rscratch2;
4982
4983 mov(result, len); // Save initial len
4984
4985 cmp(len, (u1)8); // handle shortest strings first
4986 br(LT, LOOP_1);
4987 cmp(len, (u1)32);
4988 br(LT, NEXT_8);
4989 // The following code uses the SIMD 'uzp1' and 'uzp2' instructions
4990 // to convert chars to bytes
4991 if (SoftwarePrefetchHintDistance >= 0) {
4992 ld1(Vtmp1, Vtmp2, Vtmp3, Vtmp4, T8H, src);
4993 subs(tmp2, len, SoftwarePrefetchHintDistance/2 + 16);
4994 br(LE, NEXT_32_START);
4995 b(NEXT_32_PRFM_START);
4996 BIND(NEXT_32_PRFM);
4997 ld1(Vtmp1, Vtmp2, Vtmp3, Vtmp4, T8H, src);
4998 BIND(NEXT_32_PRFM_START);
4999 prfm(Address(src, SoftwarePrefetchHintDistance));
5000 orr(v4, T16B, Vtmp1, Vtmp2);
5001 orr(v5, T16B, Vtmp3, Vtmp4);
5002 uzp1(Vtmp1, T16B, Vtmp1, Vtmp2);
5003 uzp1(Vtmp3, T16B, Vtmp3, Vtmp4);
5004 uzp2(v5, T16B, v4, v5); // high bytes
5005 umov(tmp2, v5, D, 1);
5006 fmovd(tmp1, v5);
5007 orr(tmp1, tmp1, tmp2);
5008 cbnz(tmp1, LOOP_8);
5009 stpq(Vtmp1, Vtmp3, dst);
5010 sub(len, len, 32);
5011 add(dst, dst, 32);
5012 add(src, src, 64);
5013 subs(tmp2, len, SoftwarePrefetchHintDistance/2 + 16);
5014 br(GE, NEXT_32_PRFM);
5015 cmp(len, (u1)32);
5016 br(LT, LOOP_8);
5017 BIND(NEXT_32);
5018 ld1(Vtmp1, Vtmp2, Vtmp3, Vtmp4, T8H, src);
5019 BIND(NEXT_32_START);
5020 } else {
5021 BIND(NEXT_32);
5022 ld1(Vtmp1, Vtmp2, Vtmp3, Vtmp4, T8H, src);
5023 }
5024 prfm(Address(src, SoftwarePrefetchHintDistance));
5025 uzp1(v4, T16B, Vtmp1, Vtmp2);
5026 uzp1(v5, T16B, Vtmp3, Vtmp4);
5027 orr(Vtmp1, T16B, Vtmp1, Vtmp2);
5028 orr(Vtmp3, T16B, Vtmp3, Vtmp4);
5029 uzp2(Vtmp1, T16B, Vtmp1, Vtmp3); // high bytes
5030 umov(tmp2, Vtmp1, D, 1);
5031 fmovd(tmp1, Vtmp1);
5032 orr(tmp1, tmp1, tmp2);
5033 cbnz(tmp1, LOOP_8);
5034 stpq(v4, v5, dst);
5035 sub(len, len, 32);
5036 add(dst, dst, 32);
5037 add(src, src, 64);
5038 cmp(len, (u1)32);
5039 br(GE, NEXT_32);
5040 cbz(len, DONE);
5041
5042 BIND(LOOP_8);
5043 cmp(len, (u1)8);
5044 br(LT, LOOP_1);
5045 BIND(NEXT_8);
5046 ld1(Vtmp1, T8H, src);
5047 uzp1(Vtmp2, T16B, Vtmp1, Vtmp1); // low bytes
5048 uzp2(Vtmp3, T16B, Vtmp1, Vtmp1); // high bytes
5049 fmovd(tmp1, Vtmp3);
5050 cbnz(tmp1, NEXT_1);
5051 strd(Vtmp2, dst);
5052
5053 sub(len, len, 8);
5054 add(dst, dst, 8);
5055 add(src, src, 16);
5056 cmp(len, (u1)8);
5057 br(GE, NEXT_8);
5058
5059 BIND(LOOP_1);
5060
5061 cbz(len, DONE);
5062 BIND(NEXT_1);
5063 ldrh(tmp1, Address(post(src, 2)));
5064 tst(tmp1, 0xff00);
5065 br(NE, SET_RESULT);
5066 strb(tmp1, Address(post(dst, 1)));
5067 subs(len, len, 1);
5068 br(GT, NEXT_1);
5069
5070 BIND(SET_RESULT);
5071 sub(result, result, len); // Return index where we stopped
5072 // Return len == 0 if we processed all
5073 // characters
5074 BIND(DONE);
5075 }
5076
5077
5078 // Inflate byte[] array to char[].
5079 void MacroAssembler::byte_array_inflate(Register src, Register dst, Register len,
5080 FloatRegister vtmp1, FloatRegister vtmp2, FloatRegister vtmp3,
5081 Register tmp4) {
5082 Label big, done, after_init, to_stub;
5083
5084 assert_different_registers(src, dst, len, tmp4, rscratch1);
5085
5086 fmovd(vtmp1, zr);
5087 lsrw(tmp4, len, 3);
5088 bind(after_init);
5089 cbnzw(tmp4, big);
5090 // Short string: less than 8 bytes.
5091 {
5092 Label loop, tiny;
5093
5094 cmpw(len, 4);
5095 br(LT, tiny);
5096 // Use SIMD to do 4 bytes.
5097 ldrs(vtmp2, post(src, 4));
5098 zip1(vtmp3, T8B, vtmp2, vtmp1);
5099 subw(len, len, 4);
5100 strd(vtmp3, post(dst, 8));
5101
5102 cbzw(len, done);
5103
5104 // Do the remaining bytes by steam.
5105 bind(loop);
5106 ldrb(tmp4, post(src, 1));
5107 strh(tmp4, post(dst, 2));
5108 subw(len, len, 1);
5109
5110 bind(tiny);
5111 cbnz(len, loop);
5112
5113 b(done);
5114 }
5115
5116 if (SoftwarePrefetchHintDistance >= 0) {
5117 bind(to_stub);
5118 RuntimeAddress stub = RuntimeAddress(StubRoutines::aarch64::large_byte_array_inflate());
5119 assert(stub.target() != NULL, "large_byte_array_inflate stub has not been generated");
5120 trampoline_call(stub);
5121 b(after_init);
5122 }
5123
5124 // Unpack the bytes 8 at a time.
5125 bind(big);
5126 {
5127 Label loop, around, loop_last, loop_start;
5128
5129 if (SoftwarePrefetchHintDistance >= 0) {
5130 const int large_loop_threshold = (64 + 16)/8;
5131 ldrd(vtmp2, post(src, 8));
5132 andw(len, len, 7);
5133 cmp(tmp4, (u1)large_loop_threshold);
5134 br(GE, to_stub);
5135 b(loop_start);
5136
5137 bind(loop);
5138 ldrd(vtmp2, post(src, 8));
5139 bind(loop_start);
5140 subs(tmp4, tmp4, 1);
5141 br(EQ, loop_last);
5142 zip1(vtmp2, T16B, vtmp2, vtmp1);
5143 ldrd(vtmp3, post(src, 8));
5144 st1(vtmp2, T8H, post(dst, 16));
5145 subs(tmp4, tmp4, 1);
5146 zip1(vtmp3, T16B, vtmp3, vtmp1);
5147 st1(vtmp3, T8H, post(dst, 16));
5148 br(NE, loop);
5149 b(around);
5150 bind(loop_last);
5151 zip1(vtmp2, T16B, vtmp2, vtmp1);
5152 st1(vtmp2, T8H, post(dst, 16));
5153 bind(around);
5154 cbz(len, done);
5155 } else {
5156 andw(len, len, 7);
5157 bind(loop);
5158 ldrd(vtmp2, post(src, 8));
5159 sub(tmp4, tmp4, 1);
5160 zip1(vtmp3, T16B, vtmp2, vtmp1);
5161 st1(vtmp3, T8H, post(dst, 16));
5162 cbnz(tmp4, loop);
5163 }
5164 }
5165
5166 // Do the tail of up to 8 bytes.
5167 add(src, src, len);
5168 ldrd(vtmp3, Address(src, -8));
5169 add(dst, dst, len, ext::uxtw, 1);
5170 zip1(vtmp3, T16B, vtmp3, vtmp1);
5171 strq(vtmp3, Address(dst, -16));
5172
5173 bind(done);
5174 }
5175
5176 // Compress char[] array to byte[].
5177 void MacroAssembler::char_array_compress(Register src, Register dst, Register len,
5178 FloatRegister tmp1Reg, FloatRegister tmp2Reg,
5179 FloatRegister tmp3Reg, FloatRegister tmp4Reg,
5180 Register result) {
5181 encode_iso_array(src, dst, len, result,
5182 tmp1Reg, tmp2Reg, tmp3Reg, tmp4Reg);
5183 cmp(len, zr);
5184 csel(result, result, zr, EQ);
5185 }
5186
5187 // get_thread() can be called anywhere inside generated code so we
5188 // need to save whatever non-callee save context might get clobbered
5189 // by the call to JavaThread::aarch64_get_thread_helper() or, indeed,
5190 // the call setup code.
5191 //
5192 // aarch64_get_thread_helper() clobbers only r0, r1, and flags.
5193 //
5194 void MacroAssembler::get_thread(Register dst) {
5195 RegSet saved_regs = RegSet::range(r0, r1) + lr - dst;
5196 push(saved_regs, sp);
5197
5198 mov(lr, CAST_FROM_FN_PTR(address, JavaThread::aarch64_get_thread_helper));
5199 blr(lr);
5200 if (dst != c_rarg0) {
5201 mov(dst, c_rarg0);
5202 }
5203
5204 pop(saved_regs, sp);
5205 }
5206
5207 void MacroAssembler::cache_wb(Address line) {
5208 assert(line.getMode() == Address::base_plus_offset, "mode should be base_plus_offset");
5209 assert(line.index() == noreg, "index should be noreg");
5210 assert(line.offset() == 0, "offset should be 0");
5211 // would like to assert this
5212 // assert(line._ext.shift == 0, "shift should be zero");
5213 if (VM_Version::supports_dcpop()) {
5214 // writeback using clear virtual address to point of persistence
5215 dc(Assembler::CVAP, line.base());
5216 } else {
5217 // no need to generate anything as Unsafe.writebackMemory should
5218 // never invoke this stub
5219 }
5220 }
5221
5222 void MacroAssembler::cache_wbsync(bool is_pre) {
5223 // we only need a barrier post sync
5224 if (!is_pre) {
5225 membar(Assembler::AnyAny);
5226 }
5227 }
5228
5229 void MacroAssembler::verify_sve_vector_length() {
5230 Label verify_ok;
5231 assert(UseSVE > 0, "should only be used for SVE");
5232 movw(rscratch1, zr);
5233 sve_inc(rscratch1, B);
5234 subsw(zr, rscratch1, VM_Version::get_initial_sve_vector_length());
5235 br(EQ, verify_ok);
5236 stop("Error: SVE vector length has changed since jvm startup");
5237 bind(verify_ok);
5238 }