1 /*
2 * Copyright (c) 1997, 2023, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2014, 2021, 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 "asm/assembler.hpp"
30 #include "asm/assembler.inline.hpp"
31 #include "ci/ciEnv.hpp"
32 #include "compiler/compileTask.hpp"
33 #include "compiler/disassembler.hpp"
34 #include "compiler/oopMap.hpp"
35 #include "gc/shared/barrierSet.hpp"
36 #include "gc/shared/barrierSetAssembler.hpp"
37 #include "gc/shared/cardTableBarrierSet.hpp"
38 #include "gc/shared/cardTable.hpp"
39 #include "gc/shared/collectedHeap.hpp"
40 #include "gc/shared/tlab_globals.hpp"
41 #include "interpreter/bytecodeHistogram.hpp"
42 #include "interpreter/interpreter.hpp"
43 #include "jvm.h"
44 #include "memory/resourceArea.hpp"
45 #include "memory/universe.hpp"
46 #include "nativeInst_aarch64.hpp"
47 #include "oops/accessDecorators.hpp"
48 #include "oops/compressedKlass.inline.hpp"
49 #include "oops/compressedOops.inline.hpp"
50 #include "oops/klass.inline.hpp"
51 #include "runtime/continuation.hpp"
52 #include "runtime/icache.hpp"
53 #include "runtime/interfaceSupport.inline.hpp"
54 #include "runtime/javaThread.hpp"
55 #include "runtime/jniHandles.inline.hpp"
56 #include "runtime/sharedRuntime.hpp"
57 #include "runtime/stubRoutines.hpp"
58 #include "utilities/powerOfTwo.hpp"
59 #ifdef COMPILER1
60 #include "c1/c1_LIRAssembler.hpp"
61 #endif
62 #ifdef COMPILER2
63 #include "oops/oop.hpp"
64 #include "opto/compile.hpp"
65 #include "opto/node.hpp"
66 #include "opto/output.hpp"
67 #endif
68
69 #ifdef PRODUCT
70 #define BLOCK_COMMENT(str) /* nothing */
71 #else
72 #define BLOCK_COMMENT(str) block_comment(str)
73 #endif
74 #define STOP(str) stop(str);
75 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
76
77 #ifdef ASSERT
78 extern "C" void disnm(intptr_t p);
79 #endif
80 // Target-dependent relocation processing
81 //
82 // Instruction sequences whose target may need to be retrieved or
83 // patched are distinguished by their leading instruction, sorting
84 // them into three main instruction groups and related subgroups.
85 //
86 // 1) Branch, Exception and System (insn count = 1)
87 // 1a) Unconditional branch (immediate):
88 // b/bl imm19
89 // 1b) Compare & branch (immediate):
90 // cbz/cbnz Rt imm19
91 // 1c) Test & branch (immediate):
92 // tbz/tbnz Rt imm14
93 // 1d) Conditional branch (immediate):
94 // b.cond imm19
95 //
96 // 2) Loads and Stores (insn count = 1)
97 // 2a) Load register literal:
98 // ldr Rt imm19
99 //
100 // 3) Data Processing Immediate (insn count = 2 or 3)
101 // 3a) PC-rel. addressing
102 // adr/adrp Rx imm21; ldr/str Ry Rx #imm12
103 // adr/adrp Rx imm21; add Ry Rx #imm12
104 // adr/adrp Rx imm21; movk Rx #imm16<<32; ldr/str Ry, [Rx, #offset_in_page]
105 // adr/adrp Rx imm21
106 // adr/adrp Rx imm21; movk Rx #imm16<<32
107 // adr/adrp Rx imm21; movk Rx #imm16<<32; add Ry, Rx, #offset_in_page
108 // The latter form can only happen when the target is an
109 // ExternalAddress, and (by definition) ExternalAddresses don't
110 // move. Because of that property, there is never any need to
111 // patch the last of the three instructions. However,
112 // MacroAssembler::target_addr_for_insn takes all three
113 // instructions into account and returns the correct address.
114 // 3b) Move wide (immediate)
115 // movz Rx #imm16; movk Rx #imm16 << 16; movk Rx #imm16 << 32;
116 //
117 // A switch on a subset of the instruction's bits provides an
118 // efficient dispatch to these subcases.
119 //
120 // insn[28:26] -> main group ('x' == don't care)
121 // 00x -> UNALLOCATED
122 // 100 -> Data Processing Immediate
123 // 101 -> Branch, Exception and System
124 // x1x -> Loads and Stores
125 //
126 // insn[30:25] -> subgroup ('_' == group, 'x' == don't care).
127 // n.b. in some cases extra bits need to be checked to verify the
128 // instruction is as expected
129 //
130 // 1) ... xx101x Branch, Exception and System
131 // 1a) 00___x Unconditional branch (immediate)
132 // 1b) 01___0 Compare & branch (immediate)
133 // 1c) 01___1 Test & branch (immediate)
134 // 1d) 10___0 Conditional branch (immediate)
135 // other Should not happen
136 //
137 // 2) ... xxx1x0 Loads and Stores
138 // 2a) xx1__00 Load/Store register (insn[28] == 1 && insn[24] == 0)
139 // 2aa) x01__00 Load register literal (i.e. requires insn[29] == 0)
140 // strictly should be 64 bit non-FP/SIMD i.e.
141 // 0101_000 (i.e. requires insn[31:24] == 01011000)
142 //
143 // 3) ... xx100x Data Processing Immediate
144 // 3a) xx___00 PC-rel. addressing (n.b. requires insn[24] == 0)
145 // 3b) xx___101 Move wide (immediate) (n.b. requires insn[24:23] == 01)
146 // strictly should be 64 bit movz #imm16<<0
147 // 110___10100 (i.e. requires insn[31:21] == 11010010100)
148 //
149 class RelocActions {
150 protected:
151 typedef int (*reloc_insn)(address insn_addr, address &target);
152
153 virtual reloc_insn adrpMem() = 0;
154 virtual reloc_insn adrpAdd() = 0;
155 virtual reloc_insn adrpMovk() = 0;
156
157 const address _insn_addr;
158 const uint32_t _insn;
159
160 static uint32_t insn_at(address insn_addr, int n) {
161 return ((uint32_t*)insn_addr)[n];
162 }
163 uint32_t insn_at(int n) const {
164 return insn_at(_insn_addr, n);
165 }
166
167 public:
168
169 RelocActions(address insn_addr) : _insn_addr(insn_addr), _insn(insn_at(insn_addr, 0)) {}
170 RelocActions(address insn_addr, uint32_t insn)
171 : _insn_addr(insn_addr), _insn(insn) {}
172
173 virtual int unconditionalBranch(address insn_addr, address &target) = 0;
174 virtual int conditionalBranch(address insn_addr, address &target) = 0;
175 virtual int testAndBranch(address insn_addr, address &target) = 0;
176 virtual int loadStore(address insn_addr, address &target) = 0;
177 virtual int adr(address insn_addr, address &target) = 0;
178 virtual int adrp(address insn_addr, address &target, reloc_insn inner) = 0;
179 virtual int immediate(address insn_addr, address &target) = 0;
180 virtual void verify(address insn_addr, address &target) = 0;
181
182 int ALWAYSINLINE run(address insn_addr, address &target) {
183 int instructions = 1;
184
185 uint32_t dispatch = Instruction_aarch64::extract(_insn, 30, 25);
186 switch(dispatch) {
187 case 0b001010:
188 case 0b001011: {
189 instructions = unconditionalBranch(insn_addr, target);
190 break;
191 }
192 case 0b101010: // Conditional branch (immediate)
193 case 0b011010: { // Compare & branch (immediate)
194 instructions = conditionalBranch(insn_addr, target);
195 break;
196 }
197 case 0b011011: {
198 instructions = testAndBranch(insn_addr, target);
199 break;
200 }
201 case 0b001100:
202 case 0b001110:
203 case 0b011100:
204 case 0b011110:
205 case 0b101100:
206 case 0b101110:
207 case 0b111100:
208 case 0b111110: {
209 // load/store
210 if ((Instruction_aarch64::extract(_insn, 29, 24) & 0b111011) == 0b011000) {
211 // Load register (literal)
212 instructions = loadStore(insn_addr, target);
213 break;
214 } else {
215 // nothing to do
216 assert(target == 0, "did not expect to relocate target for polling page load");
217 }
218 break;
219 }
220 case 0b001000:
221 case 0b011000:
222 case 0b101000:
223 case 0b111000: {
224 // adr/adrp
225 assert(Instruction_aarch64::extract(_insn, 28, 24) == 0b10000, "must be");
226 int shift = Instruction_aarch64::extract(_insn, 31, 31);
227 if (shift) {
228 uint32_t insn2 = insn_at(1);
229 if (Instruction_aarch64::extract(insn2, 29, 24) == 0b111001 &&
230 Instruction_aarch64::extract(_insn, 4, 0) ==
231 Instruction_aarch64::extract(insn2, 9, 5)) {
232 instructions = adrp(insn_addr, target, adrpMem());
233 } else if (Instruction_aarch64::extract(insn2, 31, 22) == 0b1001000100 &&
234 Instruction_aarch64::extract(_insn, 4, 0) ==
235 Instruction_aarch64::extract(insn2, 4, 0)) {
236 instructions = adrp(insn_addr, target, adrpAdd());
237 } else if (Instruction_aarch64::extract(insn2, 31, 21) == 0b11110010110 &&
238 Instruction_aarch64::extract(_insn, 4, 0) ==
239 Instruction_aarch64::extract(insn2, 4, 0)) {
240 instructions = adrp(insn_addr, target, adrpMovk());
241 } else {
242 ShouldNotReachHere();
243 }
244 } else {
245 instructions = adr(insn_addr, target);
246 }
247 break;
248 }
249 case 0b001001:
250 case 0b011001:
251 case 0b101001:
252 case 0b111001: {
253 instructions = immediate(insn_addr, target);
254 break;
255 }
256 default: {
257 ShouldNotReachHere();
258 }
259 }
260
261 verify(insn_addr, target);
262 return instructions * NativeInstruction::instruction_size;
263 }
264 };
265
266 class Patcher : public RelocActions {
267 virtual reloc_insn adrpMem() { return &Patcher::adrpMem_impl; }
268 virtual reloc_insn adrpAdd() { return &Patcher::adrpAdd_impl; }
269 virtual reloc_insn adrpMovk() { return &Patcher::adrpMovk_impl; }
270
271 public:
272 Patcher(address insn_addr) : RelocActions(insn_addr) {}
273
274 virtual int unconditionalBranch(address insn_addr, address &target) {
275 intptr_t offset = (target - insn_addr) >> 2;
276 Instruction_aarch64::spatch(insn_addr, 25, 0, offset);
277 return 1;
278 }
279 virtual int conditionalBranch(address insn_addr, address &target) {
280 intptr_t offset = (target - insn_addr) >> 2;
281 Instruction_aarch64::spatch(insn_addr, 23, 5, offset);
282 return 1;
283 }
284 virtual int testAndBranch(address insn_addr, address &target) {
285 intptr_t offset = (target - insn_addr) >> 2;
286 Instruction_aarch64::spatch(insn_addr, 18, 5, offset);
287 return 1;
288 }
289 virtual int loadStore(address insn_addr, address &target) {
290 intptr_t offset = (target - insn_addr) >> 2;
291 Instruction_aarch64::spatch(insn_addr, 23, 5, offset);
292 return 1;
293 }
294 virtual int adr(address insn_addr, address &target) {
295 #ifdef ASSERT
296 assert(Instruction_aarch64::extract(_insn, 28, 24) == 0b10000, "must be");
297 #endif
298 // PC-rel. addressing
299 ptrdiff_t offset = target - insn_addr;
300 int offset_lo = offset & 3;
301 offset >>= 2;
302 Instruction_aarch64::spatch(insn_addr, 23, 5, offset);
303 Instruction_aarch64::patch(insn_addr, 30, 29, offset_lo);
304 return 1;
305 }
306 virtual int adrp(address insn_addr, address &target, reloc_insn inner) {
307 int instructions = 1;
308 #ifdef ASSERT
309 assert(Instruction_aarch64::extract(_insn, 28, 24) == 0b10000, "must be");
310 #endif
311 ptrdiff_t offset = target - insn_addr;
312 instructions = 2;
313 precond(inner != nullptr);
314 // Give the inner reloc a chance to modify the target.
315 address adjusted_target = target;
316 instructions = (*inner)(insn_addr, adjusted_target);
317 uintptr_t pc_page = (uintptr_t)insn_addr >> 12;
318 uintptr_t adr_page = (uintptr_t)adjusted_target >> 12;
319 offset = adr_page - pc_page;
320 int offset_lo = offset & 3;
321 offset >>= 2;
322 Instruction_aarch64::spatch(insn_addr, 23, 5, offset);
323 Instruction_aarch64::patch(insn_addr, 30, 29, offset_lo);
324 return instructions;
325 }
326 static int adrpMem_impl(address insn_addr, address &target) {
327 uintptr_t dest = (uintptr_t)target;
328 int offset_lo = dest & 0xfff;
329 uint32_t insn2 = insn_at(insn_addr, 1);
330 uint32_t size = Instruction_aarch64::extract(insn2, 31, 30);
331 Instruction_aarch64::patch(insn_addr + sizeof (uint32_t), 21, 10, offset_lo >> size);
332 guarantee(((dest >> size) << size) == dest, "misaligned target");
333 return 2;
334 }
335 static int adrpAdd_impl(address insn_addr, address &target) {
336 uintptr_t dest = (uintptr_t)target;
337 int offset_lo = dest & 0xfff;
338 Instruction_aarch64::patch(insn_addr + sizeof (uint32_t), 21, 10, offset_lo);
339 return 2;
340 }
341 static int adrpMovk_impl(address insn_addr, address &target) {
342 uintptr_t dest = uintptr_t(target);
343 Instruction_aarch64::patch(insn_addr + sizeof (uint32_t), 20, 5, (uintptr_t)target >> 32);
344 dest = (dest & 0xffffffffULL) | (uintptr_t(insn_addr) & 0xffff00000000ULL);
345 target = address(dest);
346 return 2;
347 }
348 virtual int immediate(address insn_addr, address &target) {
349 assert(Instruction_aarch64::extract(_insn, 31, 21) == 0b11010010100, "must be");
350 uint64_t dest = (uint64_t)target;
351 // Move wide constant
352 assert(nativeInstruction_at(insn_addr+4)->is_movk(), "wrong insns in patch");
353 assert(nativeInstruction_at(insn_addr+8)->is_movk(), "wrong insns in patch");
354 Instruction_aarch64::patch(insn_addr, 20, 5, dest & 0xffff);
355 Instruction_aarch64::patch(insn_addr+4, 20, 5, (dest >>= 16) & 0xffff);
356 Instruction_aarch64::patch(insn_addr+8, 20, 5, (dest >>= 16) & 0xffff);
357 return 3;
358 }
359 virtual void verify(address insn_addr, address &target) {
360 #ifdef ASSERT
361 address address_is = MacroAssembler::target_addr_for_insn(insn_addr);
362 if (!(address_is == target)) {
363 tty->print_cr("%p at %p should be %p", address_is, insn_addr, target);
364 disnm((intptr_t)insn_addr);
365 assert(address_is == target, "should be");
366 }
367 #endif
368 }
369 };
370
371 // If insn1 and insn2 use the same register to form an address, either
372 // by an offsetted LDR or a simple ADD, return the offset. If the
373 // second instruction is an LDR, the offset may be scaled.
374 static bool offset_for(uint32_t insn1, uint32_t insn2, ptrdiff_t &byte_offset) {
375 if (Instruction_aarch64::extract(insn2, 29, 24) == 0b111001 &&
376 Instruction_aarch64::extract(insn1, 4, 0) ==
377 Instruction_aarch64::extract(insn2, 9, 5)) {
378 // Load/store register (unsigned immediate)
379 byte_offset = Instruction_aarch64::extract(insn2, 21, 10);
380 uint32_t size = Instruction_aarch64::extract(insn2, 31, 30);
381 byte_offset <<= size;
382 return true;
383 } else if (Instruction_aarch64::extract(insn2, 31, 22) == 0b1001000100 &&
384 Instruction_aarch64::extract(insn1, 4, 0) ==
385 Instruction_aarch64::extract(insn2, 4, 0)) {
386 // add (immediate)
387 byte_offset = Instruction_aarch64::extract(insn2, 21, 10);
388 return true;
389 }
390 return false;
391 }
392
393 class Decoder : public RelocActions {
394 virtual reloc_insn adrpMem() { return &Decoder::adrpMem_impl; }
395 virtual reloc_insn adrpAdd() { return &Decoder::adrpAdd_impl; }
396 virtual reloc_insn adrpMovk() { return &Decoder::adrpMovk_impl; }
397
398 public:
399 Decoder(address insn_addr, uint32_t insn) : RelocActions(insn_addr, insn) {}
400
401 virtual int loadStore(address insn_addr, address &target) {
402 intptr_t offset = Instruction_aarch64::sextract(_insn, 23, 5);
403 target = insn_addr + (offset << 2);
404 return 1;
405 }
406 virtual int unconditionalBranch(address insn_addr, address &target) {
407 intptr_t offset = Instruction_aarch64::sextract(_insn, 25, 0);
408 target = insn_addr + (offset << 2);
409 return 1;
410 }
411 virtual int conditionalBranch(address insn_addr, address &target) {
412 intptr_t offset = Instruction_aarch64::sextract(_insn, 23, 5);
413 target = address(((uint64_t)insn_addr + (offset << 2)));
414 return 1;
415 }
416 virtual int testAndBranch(address insn_addr, address &target) {
417 intptr_t offset = Instruction_aarch64::sextract(_insn, 18, 5);
418 target = address(((uint64_t)insn_addr + (offset << 2)));
419 return 1;
420 }
421 virtual int adr(address insn_addr, address &target) {
422 // PC-rel. addressing
423 intptr_t offset = Instruction_aarch64::extract(_insn, 30, 29);
424 offset |= Instruction_aarch64::sextract(_insn, 23, 5) << 2;
425 target = address((uint64_t)insn_addr + offset);
426 return 1;
427 }
428 virtual int adrp(address insn_addr, address &target, reloc_insn inner) {
429 assert(Instruction_aarch64::extract(_insn, 28, 24) == 0b10000, "must be");
430 intptr_t offset = Instruction_aarch64::extract(_insn, 30, 29);
431 offset |= Instruction_aarch64::sextract(_insn, 23, 5) << 2;
432 int shift = 12;
433 offset <<= shift;
434 uint64_t target_page = ((uint64_t)insn_addr) + offset;
435 target_page &= ((uint64_t)-1) << shift;
436 uint32_t insn2 = insn_at(1);
437 target = address(target_page);
438 precond(inner != nullptr);
439 (*inner)(insn_addr, target);
440 return 2;
441 }
442 static int adrpMem_impl(address insn_addr, address &target) {
443 uint32_t insn2 = insn_at(insn_addr, 1);
444 // Load/store register (unsigned immediate)
445 ptrdiff_t byte_offset = Instruction_aarch64::extract(insn2, 21, 10);
446 uint32_t size = Instruction_aarch64::extract(insn2, 31, 30);
447 byte_offset <<= size;
448 target += byte_offset;
449 return 2;
450 }
451 static int adrpAdd_impl(address insn_addr, address &target) {
452 uint32_t insn2 = insn_at(insn_addr, 1);
453 // add (immediate)
454 ptrdiff_t byte_offset = Instruction_aarch64::extract(insn2, 21, 10);
455 target += byte_offset;
456 return 2;
457 }
458 static int adrpMovk_impl(address insn_addr, address &target) {
459 uint32_t insn2 = insn_at(insn_addr, 1);
460 uint64_t dest = uint64_t(target);
461 dest = (dest & 0xffff0000ffffffff) |
462 ((uint64_t)Instruction_aarch64::extract(insn2, 20, 5) << 32);
463 target = address(dest);
464
465 // We know the destination 4k page. Maybe we have a third
466 // instruction.
467 uint32_t insn = insn_at(insn_addr, 0);
468 uint32_t insn3 = insn_at(insn_addr, 2);
469 ptrdiff_t byte_offset;
470 if (offset_for(insn, insn3, byte_offset)) {
471 target += byte_offset;
472 return 3;
473 } else {
474 return 2;
475 }
476 }
477 virtual int immediate(address insn_addr, address &target) {
478 uint32_t *insns = (uint32_t *)insn_addr;
479 assert(Instruction_aarch64::extract(_insn, 31, 21) == 0b11010010100, "must be");
480 // Move wide constant: movz, movk, movk. See movptr().
481 assert(nativeInstruction_at(insns+1)->is_movk(), "wrong insns in patch");
482 assert(nativeInstruction_at(insns+2)->is_movk(), "wrong insns in patch");
483 target = address(uint64_t(Instruction_aarch64::extract(_insn, 20, 5))
484 + (uint64_t(Instruction_aarch64::extract(insns[1], 20, 5)) << 16)
485 + (uint64_t(Instruction_aarch64::extract(insns[2], 20, 5)) << 32));
486 assert(nativeInstruction_at(insn_addr+4)->is_movk(), "wrong insns in patch");
487 assert(nativeInstruction_at(insn_addr+8)->is_movk(), "wrong insns in patch");
488 return 3;
489 }
490 virtual void verify(address insn_addr, address &target) {
491 }
492 };
493
494 address MacroAssembler::target_addr_for_insn(address insn_addr, uint32_t insn) {
495 Decoder decoder(insn_addr, insn);
496 address target;
497 decoder.run(insn_addr, target);
498 return target;
499 }
500
501 // Patch any kind of instruction; there may be several instructions.
502 // Return the total length (in bytes) of the instructions.
503 int MacroAssembler::pd_patch_instruction_size(address insn_addr, address target) {
504 Patcher patcher(insn_addr);
505 return patcher.run(insn_addr, target);
506 }
507
508 int MacroAssembler::patch_oop(address insn_addr, address o) {
509 int instructions;
510 unsigned insn = *(unsigned*)insn_addr;
511 assert(nativeInstruction_at(insn_addr+4)->is_movk(), "wrong insns in patch");
512
513 // OOPs are either narrow (32 bits) or wide (48 bits). We encode
514 // narrow OOPs by setting the upper 16 bits in the first
515 // instruction.
516 if (Instruction_aarch64::extract(insn, 31, 21) == 0b11010010101) {
517 // Move narrow OOP
518 uint32_t n = CompressedOops::narrow_oop_value(cast_to_oop(o));
519 Instruction_aarch64::patch(insn_addr, 20, 5, n >> 16);
520 Instruction_aarch64::patch(insn_addr+4, 20, 5, n & 0xffff);
521 instructions = 2;
522 } else {
523 // Move wide OOP
524 assert(nativeInstruction_at(insn_addr+8)->is_movk(), "wrong insns in patch");
525 uintptr_t dest = (uintptr_t)o;
526 Instruction_aarch64::patch(insn_addr, 20, 5, dest & 0xffff);
527 Instruction_aarch64::patch(insn_addr+4, 20, 5, (dest >>= 16) & 0xffff);
528 Instruction_aarch64::patch(insn_addr+8, 20, 5, (dest >>= 16) & 0xffff);
529 instructions = 3;
530 }
531 return instructions * NativeInstruction::instruction_size;
532 }
533
534 int MacroAssembler::patch_narrow_klass(address insn_addr, narrowKlass n) {
535 // Metadata pointers are either narrow (32 bits) or wide (48 bits).
536 // We encode narrow ones by setting the upper 16 bits in the first
537 // instruction.
538 NativeInstruction *insn = nativeInstruction_at(insn_addr);
539 assert(Instruction_aarch64::extract(insn->encoding(), 31, 21) == 0b11010010101 &&
540 nativeInstruction_at(insn_addr+4)->is_movk(), "wrong insns in patch");
541
542 Instruction_aarch64::patch(insn_addr, 20, 5, n >> 16);
543 Instruction_aarch64::patch(insn_addr+4, 20, 5, n & 0xffff);
544 return 2 * NativeInstruction::instruction_size;
545 }
546
547 address MacroAssembler::target_addr_for_insn_or_null(address insn_addr, unsigned insn) {
548 if (NativeInstruction::is_ldrw_to_zr(address(&insn))) {
549 return nullptr;
550 }
551 return MacroAssembler::target_addr_for_insn(insn_addr, insn);
552 }
553
554 void MacroAssembler::safepoint_poll(Label& slow_path, bool at_return, bool acquire, bool in_nmethod, Register tmp) {
555 if (acquire) {
556 lea(tmp, Address(rthread, JavaThread::polling_word_offset()));
557 ldar(tmp, tmp);
558 } else {
559 ldr(tmp, Address(rthread, JavaThread::polling_word_offset()));
560 }
561 if (at_return) {
562 // Note that when in_nmethod is set, the stack pointer is incremented before the poll. Therefore,
563 // we may safely use the sp instead to perform the stack watermark check.
564 cmp(in_nmethod ? sp : rfp, tmp);
565 br(Assembler::HI, slow_path);
566 } else {
567 tbnz(tmp, log2i_exact(SafepointMechanism::poll_bit()), slow_path);
568 }
569 }
570
571 void MacroAssembler::rt_call(address dest, Register tmp) {
572 CodeBlob *cb = CodeCache::find_blob(dest);
573 if (cb) {
574 far_call(RuntimeAddress(dest));
575 } else {
576 lea(tmp, RuntimeAddress(dest));
577 blr(tmp);
578 }
579 }
580
581 void MacroAssembler::push_cont_fastpath(Register java_thread) {
582 if (!Continuations::enabled()) return;
583 Label done;
584 ldr(rscratch1, Address(java_thread, JavaThread::cont_fastpath_offset()));
585 cmp(sp, rscratch1);
586 br(Assembler::LS, done);
587 mov(rscratch1, sp); // we can't use sp as the source in str
588 str(rscratch1, Address(java_thread, JavaThread::cont_fastpath_offset()));
589 bind(done);
590 }
591
592 void MacroAssembler::pop_cont_fastpath(Register java_thread) {
593 if (!Continuations::enabled()) return;
594 Label done;
595 ldr(rscratch1, Address(java_thread, JavaThread::cont_fastpath_offset()));
596 cmp(sp, rscratch1);
597 br(Assembler::LO, done);
598 str(zr, Address(java_thread, JavaThread::cont_fastpath_offset()));
599 bind(done);
600 }
601
602 void MacroAssembler::reset_last_Java_frame(bool clear_fp) {
603 // we must set sp to zero to clear frame
604 str(zr, Address(rthread, JavaThread::last_Java_sp_offset()));
605
606 // must clear fp, so that compiled frames are not confused; it is
607 // possible that we need it only for debugging
608 if (clear_fp) {
609 str(zr, Address(rthread, JavaThread::last_Java_fp_offset()));
610 }
611
612 // Always clear the pc because it could have been set by make_walkable()
613 str(zr, Address(rthread, JavaThread::last_Java_pc_offset()));
614 }
615
616 // Calls to C land
617 //
618 // When entering C land, the rfp, & resp of the last Java frame have to be recorded
619 // in the (thread-local) JavaThread object. When leaving C land, the last Java fp
620 // has to be reset to 0. This is required to allow proper stack traversal.
621 void MacroAssembler::set_last_Java_frame(Register last_java_sp,
622 Register last_java_fp,
623 Register last_java_pc,
624 Register scratch) {
625
626 if (last_java_pc->is_valid()) {
627 str(last_java_pc, Address(rthread,
628 JavaThread::frame_anchor_offset()
629 + JavaFrameAnchor::last_Java_pc_offset()));
630 }
631
632 // determine last_java_sp register
633 if (last_java_sp == sp) {
634 mov(scratch, sp);
635 last_java_sp = scratch;
636 } else if (!last_java_sp->is_valid()) {
637 last_java_sp = esp;
638 }
639
640 str(last_java_sp, Address(rthread, JavaThread::last_Java_sp_offset()));
641
642 // last_java_fp is optional
643 if (last_java_fp->is_valid()) {
644 str(last_java_fp, Address(rthread, JavaThread::last_Java_fp_offset()));
645 }
646 }
647
648 void MacroAssembler::set_last_Java_frame(Register last_java_sp,
649 Register last_java_fp,
650 address last_java_pc,
651 Register scratch) {
652 assert(last_java_pc != nullptr, "must provide a valid PC");
653
654 adr(scratch, last_java_pc);
655 str(scratch, Address(rthread,
656 JavaThread::frame_anchor_offset()
657 + JavaFrameAnchor::last_Java_pc_offset()));
658
659 set_last_Java_frame(last_java_sp, last_java_fp, noreg, scratch);
660 }
661
662 void MacroAssembler::set_last_Java_frame(Register last_java_sp,
663 Register last_java_fp,
664 Label &L,
665 Register scratch) {
666 if (L.is_bound()) {
667 set_last_Java_frame(last_java_sp, last_java_fp, target(L), scratch);
668 } else {
669 InstructionMark im(this);
670 L.add_patch_at(code(), locator());
671 set_last_Java_frame(last_java_sp, last_java_fp, pc() /* Patched later */, scratch);
672 }
673 }
674
675 static inline bool target_needs_far_branch(address addr) {
676 // codecache size <= 128M
677 if (!MacroAssembler::far_branches()) {
678 return false;
679 }
680 // codecache size > 240M
681 if (MacroAssembler::codestub_branch_needs_far_jump()) {
682 return true;
683 }
684 // codecache size: 128M..240M
685 return !CodeCache::is_non_nmethod(addr);
686 }
687
688 void MacroAssembler::far_call(Address entry, Register tmp) {
689 assert(ReservedCodeCacheSize < 4*G, "branch out of range");
690 assert(CodeCache::find_blob(entry.target()) != nullptr,
691 "destination of far call not found in code cache");
692 assert(entry.rspec().type() == relocInfo::external_word_type
693 || entry.rspec().type() == relocInfo::runtime_call_type
694 || entry.rspec().type() == relocInfo::none, "wrong entry relocInfo type");
695 if (target_needs_far_branch(entry.target())) {
696 uint64_t offset;
697 // We can use ADRP here because we know that the total size of
698 // the code cache cannot exceed 2Gb (ADRP limit is 4GB).
699 adrp(tmp, entry, offset);
700 add(tmp, tmp, offset);
701 blr(tmp);
702 } else {
703 bl(entry);
704 }
705 }
706
707 int MacroAssembler::far_jump(Address entry, Register tmp) {
708 assert(ReservedCodeCacheSize < 4*G, "branch out of range");
709 assert(CodeCache::find_blob(entry.target()) != nullptr,
710 "destination of far call not found in code cache");
711 assert(entry.rspec().type() == relocInfo::external_word_type
712 || entry.rspec().type() == relocInfo::runtime_call_type
713 || entry.rspec().type() == relocInfo::none, "wrong entry relocInfo type");
714 address start = pc();
715 if (target_needs_far_branch(entry.target())) {
716 uint64_t offset;
717 // We can use ADRP here because we know that the total size of
718 // the code cache cannot exceed 2Gb (ADRP limit is 4GB).
719 adrp(tmp, entry, offset);
720 add(tmp, tmp, offset);
721 br(tmp);
722 } else {
723 b(entry);
724 }
725 return pc() - start;
726 }
727
728 void MacroAssembler::reserved_stack_check() {
729 // testing if reserved zone needs to be enabled
730 Label no_reserved_zone_enabling;
731
732 ldr(rscratch1, Address(rthread, JavaThread::reserved_stack_activation_offset()));
733 cmp(sp, rscratch1);
734 br(Assembler::LO, no_reserved_zone_enabling);
735
736 enter(); // LR and FP are live.
737 lea(rscratch1, CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone));
738 mov(c_rarg0, rthread);
739 blr(rscratch1);
740 leave();
741
742 // We have already removed our own frame.
743 // throw_delayed_StackOverflowError will think that it's been
744 // called by our caller.
745 lea(rscratch1, RuntimeAddress(StubRoutines::throw_delayed_StackOverflowError_entry()));
746 br(rscratch1);
747 should_not_reach_here();
748
749 bind(no_reserved_zone_enabling);
750 }
751
752 static void pass_arg0(MacroAssembler* masm, Register arg) {
753 if (c_rarg0 != arg ) {
754 masm->mov(c_rarg0, arg);
755 }
756 }
757
758 static void pass_arg1(MacroAssembler* masm, Register arg) {
759 if (c_rarg1 != arg ) {
760 masm->mov(c_rarg1, arg);
761 }
762 }
763
764 static void pass_arg2(MacroAssembler* masm, Register arg) {
765 if (c_rarg2 != arg ) {
766 masm->mov(c_rarg2, arg);
767 }
768 }
769
770 static void pass_arg3(MacroAssembler* masm, Register arg) {
771 if (c_rarg3 != arg ) {
772 masm->mov(c_rarg3, arg);
773 }
774 }
775
776 void MacroAssembler::call_VM_base(Register oop_result,
777 Register java_thread,
778 Register last_java_sp,
779 address entry_point,
780 int number_of_arguments,
781 bool check_exceptions) {
782 // determine java_thread register
783 if (!java_thread->is_valid()) {
784 java_thread = rthread;
785 }
786
787 // determine last_java_sp register
788 if (!last_java_sp->is_valid()) {
789 last_java_sp = esp;
790 }
791
792 // debugging support
793 assert(number_of_arguments >= 0 , "cannot have negative number of arguments");
794 assert(java_thread == rthread, "unexpected register");
795 #ifdef ASSERT
796 // TraceBytecodes does not use r12 but saves it over the call, so don't verify
797 // if ((UseCompressedOops || UseCompressedClassPointers) && !TraceBytecodes) verify_heapbase("call_VM_base: heap base corrupted?");
798 #endif // ASSERT
799
800 assert(java_thread != oop_result , "cannot use the same register for java_thread & oop_result");
801 assert(java_thread != last_java_sp, "cannot use the same register for java_thread & last_java_sp");
802
803 // push java thread (becomes first argument of C function)
804
805 mov(c_rarg0, java_thread);
806
807 // set last Java frame before call
808 assert(last_java_sp != rfp, "can't use rfp");
809
810 Label l;
811 set_last_Java_frame(last_java_sp, rfp, l, rscratch1);
812
813 // do the call, remove parameters
814 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments, &l);
815
816 // lr could be poisoned with PAC signature during throw_pending_exception
817 // if it was tail-call optimized by compiler, since lr is not callee-saved
818 // reload it with proper value
819 adr(lr, l);
820
821 // reset last Java frame
822 // Only interpreter should have to clear fp
823 reset_last_Java_frame(true);
824
825 // C++ interp handles this in the interpreter
826 check_and_handle_popframe(java_thread);
827 check_and_handle_earlyret(java_thread);
828
829 if (check_exceptions) {
830 // check for pending exceptions (java_thread is set upon return)
831 ldr(rscratch1, Address(java_thread, in_bytes(Thread::pending_exception_offset())));
832 Label ok;
833 cbz(rscratch1, ok);
834 lea(rscratch1, RuntimeAddress(StubRoutines::forward_exception_entry()));
835 br(rscratch1);
836 bind(ok);
837 }
838
839 // get oop result if there is one and reset the value in the thread
840 if (oop_result->is_valid()) {
841 get_vm_result(oop_result, java_thread);
842 }
843 }
844
845 void MacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) {
846 call_VM_base(oop_result, noreg, noreg, entry_point, number_of_arguments, check_exceptions);
847 }
848
849 // Check the entry target is always reachable from any branch.
850 static bool is_always_within_branch_range(Address entry) {
851 const address target = entry.target();
852
853 if (!CodeCache::contains(target)) {
854 // We always use trampolines for callees outside CodeCache.
855 assert(entry.rspec().type() == relocInfo::runtime_call_type, "non-runtime call of an external target");
856 return false;
857 }
858
859 if (!MacroAssembler::far_branches()) {
860 return true;
861 }
862
863 if (entry.rspec().type() == relocInfo::runtime_call_type) {
864 // Runtime calls are calls of a non-compiled method (stubs, adapters).
865 // Non-compiled methods stay forever in CodeCache.
866 // We check whether the longest possible branch is within the branch range.
867 assert(CodeCache::find_blob(target) != nullptr &&
868 !CodeCache::find_blob(target)->is_compiled(),
869 "runtime call of compiled method");
870 const address right_longest_branch_start = CodeCache::high_bound() - NativeInstruction::instruction_size;
871 const address left_longest_branch_start = CodeCache::low_bound();
872 const bool is_reachable = Assembler::reachable_from_branch_at(left_longest_branch_start, target) &&
873 Assembler::reachable_from_branch_at(right_longest_branch_start, target);
874 return is_reachable;
875 }
876
877 return false;
878 }
879
880 // Maybe emit a call via a trampoline. If the code cache is small
881 // trampolines won't be emitted.
882 address MacroAssembler::trampoline_call(Address entry) {
883 assert(entry.rspec().type() == relocInfo::runtime_call_type
884 || entry.rspec().type() == relocInfo::opt_virtual_call_type
885 || entry.rspec().type() == relocInfo::static_call_type
886 || entry.rspec().type() == relocInfo::virtual_call_type, "wrong reloc type");
887
888 address target = entry.target();
889
890 if (!is_always_within_branch_range(entry)) {
891 if (!in_scratch_emit_size()) {
892 // We don't want to emit a trampoline if C2 is generating dummy
893 // code during its branch shortening phase.
894 if (entry.rspec().type() == relocInfo::runtime_call_type) {
895 assert(CodeBuffer::supports_shared_stubs(), "must support shared stubs");
896 code()->share_trampoline_for(entry.target(), offset());
897 } else {
898 address stub = emit_trampoline_stub(offset(), target);
899 if (stub == nullptr) {
900 postcond(pc() == badAddress);
901 return nullptr; // CodeCache is full
902 }
903 }
904 }
905 target = pc();
906 }
907
908 address call_pc = pc();
909 relocate(entry.rspec());
910 bl(target);
911
912 postcond(pc() != badAddress);
913 return call_pc;
914 }
915
916 // Emit a trampoline stub for a call to a target which is too far away.
917 //
918 // code sequences:
919 //
920 // call-site:
921 // branch-and-link to <destination> or <trampoline stub>
922 //
923 // Related trampoline stub for this call site in the stub section:
924 // load the call target from the constant pool
925 // branch (LR still points to the call site above)
926
927 address MacroAssembler::emit_trampoline_stub(int insts_call_instruction_offset,
928 address dest) {
929 // Max stub size: alignment nop, TrampolineStub.
930 address stub = start_a_stub(max_trampoline_stub_size());
931 if (stub == nullptr) {
932 return nullptr; // CodeBuffer::expand failed
933 }
934
935 // Create a trampoline stub relocation which relates this trampoline stub
936 // with the call instruction at insts_call_instruction_offset in the
937 // instructions code-section.
938 align(wordSize);
939 relocate(trampoline_stub_Relocation::spec(code()->insts()->start()
940 + insts_call_instruction_offset));
941 const int stub_start_offset = offset();
942
943 // Now, create the trampoline stub's code:
944 // - load the call
945 // - call
946 Label target;
947 ldr(rscratch1, target);
948 br(rscratch1);
949 bind(target);
950 assert(offset() - stub_start_offset == NativeCallTrampolineStub::data_offset,
951 "should be");
952 emit_int64((int64_t)dest);
953
954 const address stub_start_addr = addr_at(stub_start_offset);
955
956 assert(is_NativeCallTrampolineStub_at(stub_start_addr), "doesn't look like a trampoline");
957
958 end_a_stub();
959 return stub_start_addr;
960 }
961
962 int MacroAssembler::max_trampoline_stub_size() {
963 // Max stub size: alignment nop, TrampolineStub.
964 return NativeInstruction::instruction_size + NativeCallTrampolineStub::instruction_size;
965 }
966
967 void MacroAssembler::emit_static_call_stub() {
968 // CompiledDirectStaticCall::set_to_interpreted knows the
969 // exact layout of this stub.
970
971 isb();
972 mov_metadata(rmethod, nullptr);
973
974 // Jump to the entry point of the c2i stub.
975 movptr(rscratch1, 0);
976 br(rscratch1);
977 }
978
979 int MacroAssembler::static_call_stub_size() {
980 // isb; movk; movz; movz; movk; movz; movz; br
981 return 8 * NativeInstruction::instruction_size;
982 }
983
984 void MacroAssembler::c2bool(Register x) {
985 // implements x == 0 ? 0 : 1
986 // note: must only look at least-significant byte of x
987 // since C-style booleans are stored in one byte
988 // only! (was bug)
989 tst(x, 0xff);
990 cset(x, Assembler::NE);
991 }
992
993 address MacroAssembler::ic_call(address entry, jint method_index) {
994 RelocationHolder rh = virtual_call_Relocation::spec(pc(), method_index);
995 // address const_ptr = long_constant((jlong)Universe::non_oop_word());
996 // uintptr_t offset;
997 // ldr_constant(rscratch2, const_ptr);
998 movptr(rscratch2, (uintptr_t)Universe::non_oop_word());
999 return trampoline_call(Address(entry, rh));
1000 }
1001
1002 // Implementation of call_VM versions
1003
1004 void MacroAssembler::call_VM(Register oop_result,
1005 address entry_point,
1006 bool check_exceptions) {
1007 call_VM_helper(oop_result, entry_point, 0, check_exceptions);
1008 }
1009
1010 void MacroAssembler::call_VM(Register oop_result,
1011 address entry_point,
1012 Register arg_1,
1013 bool check_exceptions) {
1014 pass_arg1(this, arg_1);
1015 call_VM_helper(oop_result, entry_point, 1, check_exceptions);
1016 }
1017
1018 void MacroAssembler::call_VM(Register oop_result,
1019 address entry_point,
1020 Register arg_1,
1021 Register arg_2,
1022 bool check_exceptions) {
1023 assert_different_registers(arg_1, c_rarg2);
1024 pass_arg2(this, arg_2);
1025 pass_arg1(this, arg_1);
1026 call_VM_helper(oop_result, entry_point, 2, check_exceptions);
1027 }
1028
1029 void MacroAssembler::call_VM(Register oop_result,
1030 address entry_point,
1031 Register arg_1,
1032 Register arg_2,
1033 Register arg_3,
1034 bool check_exceptions) {
1035 assert_different_registers(arg_1, c_rarg2, c_rarg3);
1036 assert_different_registers(arg_2, c_rarg3);
1037 pass_arg3(this, arg_3);
1038
1039 pass_arg2(this, arg_2);
1040
1041 pass_arg1(this, arg_1);
1042 call_VM_helper(oop_result, entry_point, 3, check_exceptions);
1043 }
1044
1045 void MacroAssembler::call_VM(Register oop_result,
1046 Register last_java_sp,
1047 address entry_point,
1048 int number_of_arguments,
1049 bool check_exceptions) {
1050 call_VM_base(oop_result, rthread, last_java_sp, entry_point, number_of_arguments, check_exceptions);
1051 }
1052
1053 void MacroAssembler::call_VM(Register oop_result,
1054 Register last_java_sp,
1055 address entry_point,
1056 Register arg_1,
1057 bool check_exceptions) {
1058 pass_arg1(this, arg_1);
1059 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
1060 }
1061
1062 void MacroAssembler::call_VM(Register oop_result,
1063 Register last_java_sp,
1064 address entry_point,
1065 Register arg_1,
1066 Register arg_2,
1067 bool check_exceptions) {
1068
1069 assert_different_registers(arg_1, c_rarg2);
1070 pass_arg2(this, arg_2);
1071 pass_arg1(this, arg_1);
1072 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
1073 }
1074
1075 void MacroAssembler::call_VM(Register oop_result,
1076 Register last_java_sp,
1077 address entry_point,
1078 Register arg_1,
1079 Register arg_2,
1080 Register arg_3,
1081 bool check_exceptions) {
1082 assert_different_registers(arg_1, c_rarg2, c_rarg3);
1083 assert_different_registers(arg_2, c_rarg3);
1084 pass_arg3(this, arg_3);
1085 pass_arg2(this, arg_2);
1086 pass_arg1(this, arg_1);
1087 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
1088 }
1089
1090
1091 void MacroAssembler::get_vm_result(Register oop_result, Register java_thread) {
1092 ldr(oop_result, Address(java_thread, JavaThread::vm_result_offset()));
1093 str(zr, Address(java_thread, JavaThread::vm_result_offset()));
1094 verify_oop_msg(oop_result, "broken oop in call_VM_base");
1095 }
1096
1097 void MacroAssembler::get_vm_result_2(Register metadata_result, Register java_thread) {
1098 ldr(metadata_result, Address(java_thread, JavaThread::vm_result_2_offset()));
1099 str(zr, Address(java_thread, JavaThread::vm_result_2_offset()));
1100 }
1101
1102 void MacroAssembler::align(int modulus) {
1103 while (offset() % modulus != 0) nop();
1104 }
1105
1106 void MacroAssembler::post_call_nop() {
1107 if (!Continuations::enabled()) {
1108 return;
1109 }
1110 InstructionMark im(this);
1111 relocate(post_call_nop_Relocation::spec());
1112 InlineSkippedInstructionsCounter skipCounter(this);
1113 nop();
1114 movk(zr, 0);
1115 movk(zr, 0);
1116 }
1117
1118 // these are no-ops overridden by InterpreterMacroAssembler
1119
1120 void MacroAssembler::check_and_handle_earlyret(Register java_thread) { }
1121
1122 void MacroAssembler::check_and_handle_popframe(Register java_thread) { }
1123
1124 // Look up the method for a megamorphic invokeinterface call.
1125 // The target method is determined by <intf_klass, itable_index>.
1126 // The receiver klass is in recv_klass.
1127 // On success, the result will be in method_result, and execution falls through.
1128 // On failure, execution transfers to the given label.
1129 void MacroAssembler::lookup_interface_method(Register recv_klass,
1130 Register intf_klass,
1131 RegisterOrConstant itable_index,
1132 Register method_result,
1133 Register scan_temp,
1134 Label& L_no_such_interface,
1135 bool return_method) {
1136 assert_different_registers(recv_klass, intf_klass, scan_temp);
1137 assert_different_registers(method_result, intf_klass, scan_temp);
1138 assert(recv_klass != method_result || !return_method,
1139 "recv_klass can be destroyed when method isn't needed");
1140 assert(itable_index.is_constant() || itable_index.as_register() == method_result,
1141 "caller must use same register for non-constant itable index as for method");
1142
1143 // Compute start of first itableOffsetEntry (which is at the end of the vtable)
1144 int vtable_base = in_bytes(Klass::vtable_start_offset());
1145 int itentry_off = in_bytes(itableMethodEntry::method_offset());
1146 int scan_step = itableOffsetEntry::size() * wordSize;
1147 int vte_size = vtableEntry::size_in_bytes();
1148 assert(vte_size == wordSize, "else adjust times_vte_scale");
1149
1150 ldrw(scan_temp, Address(recv_klass, Klass::vtable_length_offset()));
1151
1152 // %%% Could store the aligned, prescaled offset in the klassoop.
1153 // lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base));
1154 lea(scan_temp, Address(recv_klass, scan_temp, Address::lsl(3)));
1155 add(scan_temp, scan_temp, vtable_base);
1156
1157 if (return_method) {
1158 // Adjust recv_klass by scaled itable_index, so we can free itable_index.
1159 assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
1160 // lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off));
1161 lea(recv_klass, Address(recv_klass, itable_index, Address::lsl(3)));
1162 if (itentry_off)
1163 add(recv_klass, recv_klass, itentry_off);
1164 }
1165
1166 // for (scan = klass->itable(); scan->interface() != nullptr; scan += scan_step) {
1167 // if (scan->interface() == intf) {
1168 // result = (klass + scan->offset() + itable_index);
1169 // }
1170 // }
1171 Label search, found_method;
1172
1173 ldr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset()));
1174 cmp(intf_klass, method_result);
1175 br(Assembler::EQ, found_method);
1176 bind(search);
1177 // Check that the previous entry is non-null. A null entry means that
1178 // the receiver class doesn't implement the interface, and wasn't the
1179 // same as when the caller was compiled.
1180 cbz(method_result, L_no_such_interface);
1181 if (itableOffsetEntry::interface_offset() != 0) {
1182 add(scan_temp, scan_temp, scan_step);
1183 ldr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset()));
1184 } else {
1185 ldr(method_result, Address(pre(scan_temp, scan_step)));
1186 }
1187 cmp(intf_klass, method_result);
1188 br(Assembler::NE, search);
1189
1190 bind(found_method);
1191
1192 // Got a hit.
1193 if (return_method) {
1194 ldrw(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset()));
1195 ldr(method_result, Address(recv_klass, scan_temp, Address::uxtw(0)));
1196 }
1197 }
1198
1199 // Look up the method for a megamorphic invokeinterface call in a single pass over itable:
1200 // - check recv_klass (actual object class) is a subtype of resolved_klass from CompiledICHolder
1201 // - find a holder_klass (class that implements the method) vtable offset and get the method from vtable by index
1202 // The target method is determined by <holder_klass, itable_index>.
1203 // The receiver klass is in recv_klass.
1204 // On success, the result will be in method_result, and execution falls through.
1205 // On failure, execution transfers to the given label.
1206 void MacroAssembler::lookup_interface_method_stub(Register recv_klass,
1207 Register holder_klass,
1208 Register resolved_klass,
1209 Register method_result,
1210 Register temp_itbl_klass,
1211 Register scan_temp,
1212 int itable_index,
1213 Label& L_no_such_interface) {
1214 // 'method_result' is only used as output register at the very end of this method.
1215 // Until then we can reuse it as 'holder_offset'.
1216 Register holder_offset = method_result;
1217 assert_different_registers(resolved_klass, recv_klass, holder_klass, temp_itbl_klass, scan_temp, holder_offset);
1218
1219 int vtable_start_offset = in_bytes(Klass::vtable_start_offset());
1220 int itable_offset_entry_size = itableOffsetEntry::size() * wordSize;
1221 int ioffset = in_bytes(itableOffsetEntry::interface_offset());
1222 int ooffset = in_bytes(itableOffsetEntry::offset_offset());
1223
1224 Label L_loop_search_resolved_entry, L_resolved_found, L_holder_found;
1225
1226 ldrw(scan_temp, Address(recv_klass, Klass::vtable_length_offset()));
1227 add(recv_klass, recv_klass, vtable_start_offset + ioffset);
1228 // itableOffsetEntry[] itable = recv_klass + Klass::vtable_start_offset() + sizeof(vtableEntry) * recv_klass->_vtable_len;
1229 // temp_itbl_klass = itable[0]._interface;
1230 int vtblEntrySize = vtableEntry::size_in_bytes();
1231 assert(vtblEntrySize == wordSize, "ldr lsl shift amount must be 3");
1232 ldr(temp_itbl_klass, Address(recv_klass, scan_temp, Address::lsl(exact_log2(vtblEntrySize))));
1233 mov(holder_offset, zr);
1234 // scan_temp = &(itable[0]._interface)
1235 lea(scan_temp, Address(recv_klass, scan_temp, Address::lsl(exact_log2(vtblEntrySize))));
1236
1237 // Initial checks:
1238 // - if (holder_klass != resolved_klass), go to "scan for resolved"
1239 // - if (itable[0] == holder_klass), shortcut to "holder found"
1240 // - if (itable[0] == 0), no such interface
1241 cmp(resolved_klass, holder_klass);
1242 br(Assembler::NE, L_loop_search_resolved_entry);
1243 cmp(holder_klass, temp_itbl_klass);
1244 br(Assembler::EQ, L_holder_found);
1245 cbz(temp_itbl_klass, L_no_such_interface);
1246
1247 // Loop: Look for holder_klass record in itable
1248 // do {
1249 // temp_itbl_klass = *(scan_temp += itable_offset_entry_size);
1250 // if (temp_itbl_klass == holder_klass) {
1251 // goto L_holder_found; // Found!
1252 // }
1253 // } while (temp_itbl_klass != 0);
1254 // goto L_no_such_interface // Not found.
1255 Label L_search_holder;
1256 bind(L_search_holder);
1257 ldr(temp_itbl_klass, Address(pre(scan_temp, itable_offset_entry_size)));
1258 cmp(holder_klass, temp_itbl_klass);
1259 br(Assembler::EQ, L_holder_found);
1260 cbnz(temp_itbl_klass, L_search_holder);
1261
1262 b(L_no_such_interface);
1263
1264 // Loop: Look for resolved_class record in itable
1265 // while (true) {
1266 // temp_itbl_klass = *(scan_temp += itable_offset_entry_size);
1267 // if (temp_itbl_klass == 0) {
1268 // goto L_no_such_interface;
1269 // }
1270 // if (temp_itbl_klass == resolved_klass) {
1271 // goto L_resolved_found; // Found!
1272 // }
1273 // if (temp_itbl_klass == holder_klass) {
1274 // holder_offset = scan_temp;
1275 // }
1276 // }
1277 //
1278 Label L_loop_search_resolved;
1279 bind(L_loop_search_resolved);
1280 ldr(temp_itbl_klass, Address(pre(scan_temp, itable_offset_entry_size)));
1281 bind(L_loop_search_resolved_entry);
1282 cbz(temp_itbl_klass, L_no_such_interface);
1283 cmp(resolved_klass, temp_itbl_klass);
1284 br(Assembler::EQ, L_resolved_found);
1285 cmp(holder_klass, temp_itbl_klass);
1286 br(Assembler::NE, L_loop_search_resolved);
1287 mov(holder_offset, scan_temp);
1288 b(L_loop_search_resolved);
1289
1290 // See if we already have a holder klass. If not, go and scan for it.
1291 bind(L_resolved_found);
1292 cbz(holder_offset, L_search_holder);
1293 mov(scan_temp, holder_offset);
1294
1295 // Finally, scan_temp contains holder_klass vtable offset
1296 bind(L_holder_found);
1297 ldrw(method_result, Address(scan_temp, ooffset - ioffset));
1298 add(recv_klass, recv_klass, itable_index * wordSize + in_bytes(itableMethodEntry::method_offset())
1299 - vtable_start_offset - ioffset); // substract offsets to restore the original value of recv_klass
1300 ldr(method_result, Address(recv_klass, method_result, Address::uxtw(0)));
1301 }
1302
1303 // virtual method calling
1304 void MacroAssembler::lookup_virtual_method(Register recv_klass,
1305 RegisterOrConstant vtable_index,
1306 Register method_result) {
1307 assert(vtableEntry::size() * wordSize == 8,
1308 "adjust the scaling in the code below");
1309 int64_t vtable_offset_in_bytes = in_bytes(Klass::vtable_start_offset() + vtableEntry::method_offset());
1310
1311 if (vtable_index.is_register()) {
1312 lea(method_result, Address(recv_klass,
1313 vtable_index.as_register(),
1314 Address::lsl(LogBytesPerWord)));
1315 ldr(method_result, Address(method_result, vtable_offset_in_bytes));
1316 } else {
1317 vtable_offset_in_bytes += vtable_index.as_constant() * wordSize;
1318 ldr(method_result,
1319 form_address(rscratch1, recv_klass, vtable_offset_in_bytes, 0));
1320 }
1321 }
1322
1323 void MacroAssembler::check_klass_subtype(Register sub_klass,
1324 Register super_klass,
1325 Register temp_reg,
1326 Label& L_success) {
1327 Label L_failure;
1328 check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg, &L_success, &L_failure, nullptr);
1329 check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, nullptr);
1330 bind(L_failure);
1331 }
1332
1333
1334 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass,
1335 Register super_klass,
1336 Register temp_reg,
1337 Label* L_success,
1338 Label* L_failure,
1339 Label* L_slow_path,
1340 RegisterOrConstant super_check_offset) {
1341 assert_different_registers(sub_klass, super_klass, temp_reg);
1342 bool must_load_sco = (super_check_offset.constant_or_zero() == -1);
1343 if (super_check_offset.is_register()) {
1344 assert_different_registers(sub_klass, super_klass,
1345 super_check_offset.as_register());
1346 } else if (must_load_sco) {
1347 assert(temp_reg != noreg, "supply either a temp or a register offset");
1348 }
1349
1350 Label L_fallthrough;
1351 int label_nulls = 0;
1352 if (L_success == nullptr) { L_success = &L_fallthrough; label_nulls++; }
1353 if (L_failure == nullptr) { L_failure = &L_fallthrough; label_nulls++; }
1354 if (L_slow_path == nullptr) { L_slow_path = &L_fallthrough; label_nulls++; }
1355 assert(label_nulls <= 1, "at most one null in the batch");
1356
1357 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
1358 int sco_offset = in_bytes(Klass::super_check_offset_offset());
1359 Address super_check_offset_addr(super_klass, sco_offset);
1360
1361 // Hacked jmp, which may only be used just before L_fallthrough.
1362 #define final_jmp(label) \
1363 if (&(label) == &L_fallthrough) { /*do nothing*/ } \
1364 else b(label) /*omit semi*/
1365
1366 // If the pointers are equal, we are done (e.g., String[] elements).
1367 // This self-check enables sharing of secondary supertype arrays among
1368 // non-primary types such as array-of-interface. Otherwise, each such
1369 // type would need its own customized SSA.
1370 // We move this check to the front of the fast path because many
1371 // type checks are in fact trivially successful in this manner,
1372 // so we get a nicely predicted branch right at the start of the check.
1373 cmp(sub_klass, super_klass);
1374 br(Assembler::EQ, *L_success);
1375
1376 // Check the supertype display:
1377 if (must_load_sco) {
1378 ldrw(temp_reg, super_check_offset_addr);
1379 super_check_offset = RegisterOrConstant(temp_reg);
1380 }
1381 Address super_check_addr(sub_klass, super_check_offset);
1382 ldr(rscratch1, super_check_addr);
1383 cmp(super_klass, rscratch1); // load displayed supertype
1384
1385 // This check has worked decisively for primary supers.
1386 // Secondary supers are sought in the super_cache ('super_cache_addr').
1387 // (Secondary supers are interfaces and very deeply nested subtypes.)
1388 // This works in the same check above because of a tricky aliasing
1389 // between the super_cache and the primary super display elements.
1390 // (The 'super_check_addr' can address either, as the case requires.)
1391 // Note that the cache is updated below if it does not help us find
1392 // what we need immediately.
1393 // So if it was a primary super, we can just fail immediately.
1394 // Otherwise, it's the slow path for us (no success at this point).
1395
1396 if (super_check_offset.is_register()) {
1397 br(Assembler::EQ, *L_success);
1398 subs(zr, super_check_offset.as_register(), sc_offset);
1399 if (L_failure == &L_fallthrough) {
1400 br(Assembler::EQ, *L_slow_path);
1401 } else {
1402 br(Assembler::NE, *L_failure);
1403 final_jmp(*L_slow_path);
1404 }
1405 } else if (super_check_offset.as_constant() == sc_offset) {
1406 // Need a slow path; fast failure is impossible.
1407 if (L_slow_path == &L_fallthrough) {
1408 br(Assembler::EQ, *L_success);
1409 } else {
1410 br(Assembler::NE, *L_slow_path);
1411 final_jmp(*L_success);
1412 }
1413 } else {
1414 // No slow path; it's a fast decision.
1415 if (L_failure == &L_fallthrough) {
1416 br(Assembler::EQ, *L_success);
1417 } else {
1418 br(Assembler::NE, *L_failure);
1419 final_jmp(*L_success);
1420 }
1421 }
1422
1423 bind(L_fallthrough);
1424
1425 #undef final_jmp
1426 }
1427
1428 // These two are taken from x86, but they look generally useful
1429
1430 // scans count pointer sized words at [addr] for occurrence of value,
1431 // generic
1432 void MacroAssembler::repne_scan(Register addr, Register value, Register count,
1433 Register scratch) {
1434 Label Lloop, Lexit;
1435 cbz(count, Lexit);
1436 bind(Lloop);
1437 ldr(scratch, post(addr, wordSize));
1438 cmp(value, scratch);
1439 br(EQ, Lexit);
1440 sub(count, count, 1);
1441 cbnz(count, Lloop);
1442 bind(Lexit);
1443 }
1444
1445 // scans count 4 byte words at [addr] for occurrence of value,
1446 // generic
1447 void MacroAssembler::repne_scanw(Register addr, Register value, Register count,
1448 Register scratch) {
1449 Label Lloop, Lexit;
1450 cbz(count, Lexit);
1451 bind(Lloop);
1452 ldrw(scratch, post(addr, wordSize));
1453 cmpw(value, scratch);
1454 br(EQ, Lexit);
1455 sub(count, count, 1);
1456 cbnz(count, Lloop);
1457 bind(Lexit);
1458 }
1459
1460 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass,
1461 Register super_klass,
1462 Register temp_reg,
1463 Register temp2_reg,
1464 Label* L_success,
1465 Label* L_failure,
1466 bool set_cond_codes) {
1467 assert_different_registers(sub_klass, super_klass, temp_reg);
1468 if (temp2_reg != noreg)
1469 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg, rscratch1);
1470 #define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg)
1471
1472 Label L_fallthrough;
1473 int label_nulls = 0;
1474 if (L_success == nullptr) { L_success = &L_fallthrough; label_nulls++; }
1475 if (L_failure == nullptr) { L_failure = &L_fallthrough; label_nulls++; }
1476 assert(label_nulls <= 1, "at most one null in the batch");
1477
1478 // a couple of useful fields in sub_klass:
1479 int ss_offset = in_bytes(Klass::secondary_supers_offset());
1480 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
1481 Address secondary_supers_addr(sub_klass, ss_offset);
1482 Address super_cache_addr( sub_klass, sc_offset);
1483
1484 BLOCK_COMMENT("check_klass_subtype_slow_path");
1485
1486 // Do a linear scan of the secondary super-klass chain.
1487 // This code is rarely used, so simplicity is a virtue here.
1488 // The repne_scan instruction uses fixed registers, which we must spill.
1489 // Don't worry too much about pre-existing connections with the input regs.
1490
1491 assert(sub_klass != r0, "killed reg"); // killed by mov(r0, super)
1492 assert(sub_klass != r2, "killed reg"); // killed by lea(r2, &pst_counter)
1493
1494 RegSet pushed_registers;
1495 if (!IS_A_TEMP(r2)) pushed_registers += r2;
1496 if (!IS_A_TEMP(r5)) pushed_registers += r5;
1497
1498 if (super_klass != r0) {
1499 if (!IS_A_TEMP(r0)) pushed_registers += r0;
1500 }
1501
1502 push(pushed_registers, sp);
1503
1504 // Get super_klass value into r0 (even if it was in r5 or r2).
1505 if (super_klass != r0) {
1506 mov(r0, super_klass);
1507 }
1508
1509 #ifndef PRODUCT
1510 mov(rscratch2, (address)&SharedRuntime::_partial_subtype_ctr);
1511 Address pst_counter_addr(rscratch2);
1512 ldr(rscratch1, pst_counter_addr);
1513 add(rscratch1, rscratch1, 1);
1514 str(rscratch1, pst_counter_addr);
1515 #endif //PRODUCT
1516
1517 // We will consult the secondary-super array.
1518 ldr(r5, secondary_supers_addr);
1519 // Load the array length.
1520 ldrw(r2, Address(r5, Array<Klass*>::length_offset_in_bytes()));
1521 // Skip to start of data.
1522 add(r5, r5, Array<Klass*>::base_offset_in_bytes());
1523
1524 cmp(sp, zr); // Clear Z flag; SP is never zero
1525 // Scan R2 words at [R5] for an occurrence of R0.
1526 // Set NZ/Z based on last compare.
1527 repne_scan(r5, r0, r2, rscratch1);
1528
1529 // Unspill the temp. registers:
1530 pop(pushed_registers, sp);
1531
1532 br(Assembler::NE, *L_failure);
1533
1534 // Success. Cache the super we found and proceed in triumph.
1535 str(super_klass, super_cache_addr);
1536
1537 if (L_success != &L_fallthrough) {
1538 b(*L_success);
1539 }
1540
1541 #undef IS_A_TEMP
1542
1543 bind(L_fallthrough);
1544 }
1545
1546 void MacroAssembler::clinit_barrier(Register klass, Register scratch, Label* L_fast_path, Label* L_slow_path) {
1547 assert(L_fast_path != nullptr || L_slow_path != nullptr, "at least one is required");
1548 assert_different_registers(klass, rthread, scratch);
1549
1550 Label L_fallthrough, L_tmp;
1551 if (L_fast_path == nullptr) {
1552 L_fast_path = &L_fallthrough;
1553 } else if (L_slow_path == nullptr) {
1554 L_slow_path = &L_fallthrough;
1555 }
1556 // Fast path check: class is fully initialized
1557 ldrb(scratch, Address(klass, InstanceKlass::init_state_offset()));
1558 subs(zr, scratch, InstanceKlass::fully_initialized);
1559 br(Assembler::EQ, *L_fast_path);
1560
1561 // Fast path check: current thread is initializer thread
1562 ldr(scratch, Address(klass, InstanceKlass::init_thread_offset()));
1563 cmp(rthread, scratch);
1564
1565 if (L_slow_path == &L_fallthrough) {
1566 br(Assembler::EQ, *L_fast_path);
1567 bind(*L_slow_path);
1568 } else if (L_fast_path == &L_fallthrough) {
1569 br(Assembler::NE, *L_slow_path);
1570 bind(*L_fast_path);
1571 } else {
1572 Unimplemented();
1573 }
1574 }
1575
1576 void MacroAssembler::_verify_oop(Register reg, const char* s, const char* file, int line) {
1577 if (!VerifyOops) return;
1578
1579 // Pass register number to verify_oop_subroutine
1580 const char* b = nullptr;
1581 {
1582 ResourceMark rm;
1583 stringStream ss;
1584 ss.print("verify_oop: %s: %s (%s:%d)", reg->name(), s, file, line);
1585 b = code_string(ss.as_string());
1586 }
1587 BLOCK_COMMENT("verify_oop {");
1588
1589 strip_return_address(); // This might happen within a stack frame.
1590 protect_return_address();
1591 stp(r0, rscratch1, Address(pre(sp, -2 * wordSize)));
1592 stp(rscratch2, lr, Address(pre(sp, -2 * wordSize)));
1593
1594 mov(r0, reg);
1595 movptr(rscratch1, (uintptr_t)(address)b);
1596
1597 // call indirectly to solve generation ordering problem
1598 lea(rscratch2, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
1599 ldr(rscratch2, Address(rscratch2));
1600 blr(rscratch2);
1601
1602 ldp(rscratch2, lr, Address(post(sp, 2 * wordSize)));
1603 ldp(r0, rscratch1, Address(post(sp, 2 * wordSize)));
1604 authenticate_return_address();
1605
1606 BLOCK_COMMENT("} verify_oop");
1607 }
1608
1609 void MacroAssembler::_verify_oop_addr(Address addr, const char* s, const char* file, int line) {
1610 if (!VerifyOops) return;
1611
1612 const char* b = nullptr;
1613 {
1614 ResourceMark rm;
1615 stringStream ss;
1616 ss.print("verify_oop_addr: %s (%s:%d)", s, file, line);
1617 b = code_string(ss.as_string());
1618 }
1619 BLOCK_COMMENT("verify_oop_addr {");
1620
1621 strip_return_address(); // This might happen within a stack frame.
1622 protect_return_address();
1623 stp(r0, rscratch1, Address(pre(sp, -2 * wordSize)));
1624 stp(rscratch2, lr, Address(pre(sp, -2 * wordSize)));
1625
1626 // addr may contain sp so we will have to adjust it based on the
1627 // pushes that we just did.
1628 if (addr.uses(sp)) {
1629 lea(r0, addr);
1630 ldr(r0, Address(r0, 4 * wordSize));
1631 } else {
1632 ldr(r0, addr);
1633 }
1634 movptr(rscratch1, (uintptr_t)(address)b);
1635
1636 // call indirectly to solve generation ordering problem
1637 lea(rscratch2, ExternalAddress(StubRoutines::verify_oop_subroutine_entry_address()));
1638 ldr(rscratch2, Address(rscratch2));
1639 blr(rscratch2);
1640
1641 ldp(rscratch2, lr, Address(post(sp, 2 * wordSize)));
1642 ldp(r0, rscratch1, Address(post(sp, 2 * wordSize)));
1643 authenticate_return_address();
1644
1645 BLOCK_COMMENT("} verify_oop_addr");
1646 }
1647
1648 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot,
1649 int extra_slot_offset) {
1650 // cf. TemplateTable::prepare_invoke(), if (load_receiver).
1651 int stackElementSize = Interpreter::stackElementSize;
1652 int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0);
1653 #ifdef ASSERT
1654 int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1);
1655 assert(offset1 - offset == stackElementSize, "correct arithmetic");
1656 #endif
1657 if (arg_slot.is_constant()) {
1658 return Address(esp, arg_slot.as_constant() * stackElementSize
1659 + offset);
1660 } else {
1661 add(rscratch1, esp, arg_slot.as_register(),
1662 ext::uxtx, exact_log2(stackElementSize));
1663 return Address(rscratch1, offset);
1664 }
1665 }
1666
1667 void MacroAssembler::call_VM_leaf_base(address entry_point,
1668 int number_of_arguments,
1669 Label *retaddr) {
1670 Label E, L;
1671
1672 stp(rscratch1, rmethod, Address(pre(sp, -2 * wordSize)));
1673
1674 mov(rscratch1, entry_point);
1675 blr(rscratch1);
1676 if (retaddr)
1677 bind(*retaddr);
1678
1679 ldp(rscratch1, rmethod, Address(post(sp, 2 * wordSize)));
1680 }
1681
1682 void MacroAssembler::call_VM_leaf(address entry_point, int number_of_arguments) {
1683 call_VM_leaf_base(entry_point, number_of_arguments);
1684 }
1685
1686 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0) {
1687 pass_arg0(this, arg_0);
1688 call_VM_leaf_base(entry_point, 1);
1689 }
1690
1691 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
1692 assert_different_registers(arg_1, c_rarg0);
1693 pass_arg0(this, arg_0);
1694 pass_arg1(this, arg_1);
1695 call_VM_leaf_base(entry_point, 2);
1696 }
1697
1698 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0,
1699 Register arg_1, Register arg_2) {
1700 assert_different_registers(arg_1, c_rarg0);
1701 assert_different_registers(arg_2, c_rarg0, c_rarg1);
1702 pass_arg0(this, arg_0);
1703 pass_arg1(this, arg_1);
1704 pass_arg2(this, arg_2);
1705 call_VM_leaf_base(entry_point, 3);
1706 }
1707
1708 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) {
1709 pass_arg0(this, arg_0);
1710 MacroAssembler::call_VM_leaf_base(entry_point, 1);
1711 }
1712
1713 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
1714
1715 assert_different_registers(arg_0, c_rarg1);
1716 pass_arg1(this, arg_1);
1717 pass_arg0(this, arg_0);
1718 MacroAssembler::call_VM_leaf_base(entry_point, 2);
1719 }
1720
1721 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) {
1722 assert_different_registers(arg_0, c_rarg1, c_rarg2);
1723 assert_different_registers(arg_1, c_rarg2);
1724 pass_arg2(this, arg_2);
1725 pass_arg1(this, arg_1);
1726 pass_arg0(this, arg_0);
1727 MacroAssembler::call_VM_leaf_base(entry_point, 3);
1728 }
1729
1730 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) {
1731 assert_different_registers(arg_0, c_rarg1, c_rarg2, c_rarg3);
1732 assert_different_registers(arg_1, c_rarg2, c_rarg3);
1733 assert_different_registers(arg_2, c_rarg3);
1734 pass_arg3(this, arg_3);
1735 pass_arg2(this, arg_2);
1736 pass_arg1(this, arg_1);
1737 pass_arg0(this, arg_0);
1738 MacroAssembler::call_VM_leaf_base(entry_point, 4);
1739 }
1740
1741 void MacroAssembler::null_check(Register reg, int offset) {
1742 if (needs_explicit_null_check(offset)) {
1743 // provoke OS null exception if reg is null by
1744 // accessing M[reg] w/o changing any registers
1745 // NOTE: this is plenty to provoke a segv
1746 ldr(zr, Address(reg));
1747 } else {
1748 // nothing to do, (later) access of M[reg + offset]
1749 // will provoke OS null exception if reg is null
1750 }
1751 }
1752
1753 // MacroAssembler protected routines needed to implement
1754 // public methods
1755
1756 void MacroAssembler::mov(Register r, Address dest) {
1757 code_section()->relocate(pc(), dest.rspec());
1758 uint64_t imm64 = (uint64_t)dest.target();
1759 movptr(r, imm64);
1760 }
1761
1762 // Move a constant pointer into r. In AArch64 mode the virtual
1763 // address space is 48 bits in size, so we only need three
1764 // instructions to create a patchable instruction sequence that can
1765 // reach anywhere.
1766 void MacroAssembler::movptr(Register r, uintptr_t imm64) {
1767 #ifndef PRODUCT
1768 {
1769 char buffer[64];
1770 snprintf(buffer, sizeof(buffer), "0x%" PRIX64, (uint64_t)imm64);
1771 block_comment(buffer);
1772 }
1773 #endif
1774 assert(imm64 < (1ull << 48), "48-bit overflow in address constant");
1775 movz(r, imm64 & 0xffff);
1776 imm64 >>= 16;
1777 movk(r, imm64 & 0xffff, 16);
1778 imm64 >>= 16;
1779 movk(r, imm64 & 0xffff, 32);
1780 }
1781
1782 // Macro to mov replicated immediate to vector register.
1783 // imm64: only the lower 8/16/32 bits are considered for B/H/S type. That is,
1784 // the upper 56/48/32 bits must be zeros for B/H/S type.
1785 // Vd will get the following values for different arrangements in T
1786 // imm64 == hex 000000gh T8B: Vd = ghghghghghghghgh
1787 // imm64 == hex 000000gh T16B: Vd = ghghghghghghghghghghghghghghghgh
1788 // imm64 == hex 0000efgh T4H: Vd = efghefghefghefgh
1789 // imm64 == hex 0000efgh T8H: Vd = efghefghefghefghefghefghefghefgh
1790 // imm64 == hex abcdefgh T2S: Vd = abcdefghabcdefgh
1791 // imm64 == hex abcdefgh T4S: Vd = abcdefghabcdefghabcdefghabcdefgh
1792 // imm64 == hex abcdefgh T1D: Vd = 00000000abcdefgh
1793 // imm64 == hex abcdefgh T2D: Vd = 00000000abcdefgh00000000abcdefgh
1794 // Clobbers rscratch1
1795 void MacroAssembler::mov(FloatRegister Vd, SIMD_Arrangement T, uint64_t imm64) {
1796 assert(T != T1Q, "unsupported");
1797 if (T == T1D || T == T2D) {
1798 int imm = operand_valid_for_movi_immediate(imm64, T);
1799 if (-1 != imm) {
1800 movi(Vd, T, imm);
1801 } else {
1802 mov(rscratch1, imm64);
1803 dup(Vd, T, rscratch1);
1804 }
1805 return;
1806 }
1807
1808 #ifdef ASSERT
1809 if (T == T8B || T == T16B) assert((imm64 & ~0xff) == 0, "extraneous bits (T8B/T16B)");
1810 if (T == T4H || T == T8H) assert((imm64 & ~0xffff) == 0, "extraneous bits (T4H/T8H)");
1811 if (T == T2S || T == T4S) assert((imm64 & ~0xffffffff) == 0, "extraneous bits (T2S/T4S)");
1812 #endif
1813 int shift = operand_valid_for_movi_immediate(imm64, T);
1814 uint32_t imm32 = imm64 & 0xffffffffULL;
1815 if (shift >= 0) {
1816 movi(Vd, T, (imm32 >> shift) & 0xff, shift);
1817 } else {
1818 movw(rscratch1, imm32);
1819 dup(Vd, T, rscratch1);
1820 }
1821 }
1822
1823 void MacroAssembler::mov_immediate64(Register dst, uint64_t imm64)
1824 {
1825 #ifndef PRODUCT
1826 {
1827 char buffer[64];
1828 snprintf(buffer, sizeof(buffer), "0x%" PRIX64, imm64);
1829 block_comment(buffer);
1830 }
1831 #endif
1832 if (operand_valid_for_logical_immediate(false, imm64)) {
1833 orr(dst, zr, imm64);
1834 } else {
1835 // we can use a combination of MOVZ or MOVN with
1836 // MOVK to build up the constant
1837 uint64_t imm_h[4];
1838 int zero_count = 0;
1839 int neg_count = 0;
1840 int i;
1841 for (i = 0; i < 4; i++) {
1842 imm_h[i] = ((imm64 >> (i * 16)) & 0xffffL);
1843 if (imm_h[i] == 0) {
1844 zero_count++;
1845 } else if (imm_h[i] == 0xffffL) {
1846 neg_count++;
1847 }
1848 }
1849 if (zero_count == 4) {
1850 // one MOVZ will do
1851 movz(dst, 0);
1852 } else if (neg_count == 4) {
1853 // one MOVN will do
1854 movn(dst, 0);
1855 } else if (zero_count == 3) {
1856 for (i = 0; i < 4; i++) {
1857 if (imm_h[i] != 0L) {
1858 movz(dst, (uint32_t)imm_h[i], (i << 4));
1859 break;
1860 }
1861 }
1862 } else if (neg_count == 3) {
1863 // one MOVN will do
1864 for (int i = 0; i < 4; i++) {
1865 if (imm_h[i] != 0xffffL) {
1866 movn(dst, (uint32_t)imm_h[i] ^ 0xffffL, (i << 4));
1867 break;
1868 }
1869 }
1870 } else if (zero_count == 2) {
1871 // one MOVZ and one MOVK will do
1872 for (i = 0; i < 3; i++) {
1873 if (imm_h[i] != 0L) {
1874 movz(dst, (uint32_t)imm_h[i], (i << 4));
1875 i++;
1876 break;
1877 }
1878 }
1879 for (;i < 4; i++) {
1880 if (imm_h[i] != 0L) {
1881 movk(dst, (uint32_t)imm_h[i], (i << 4));
1882 }
1883 }
1884 } else if (neg_count == 2) {
1885 // one MOVN and one MOVK will do
1886 for (i = 0; i < 4; i++) {
1887 if (imm_h[i] != 0xffffL) {
1888 movn(dst, (uint32_t)imm_h[i] ^ 0xffffL, (i << 4));
1889 i++;
1890 break;
1891 }
1892 }
1893 for (;i < 4; i++) {
1894 if (imm_h[i] != 0xffffL) {
1895 movk(dst, (uint32_t)imm_h[i], (i << 4));
1896 }
1897 }
1898 } else if (zero_count == 1) {
1899 // one MOVZ and two MOVKs will do
1900 for (i = 0; i < 4; i++) {
1901 if (imm_h[i] != 0L) {
1902 movz(dst, (uint32_t)imm_h[i], (i << 4));
1903 i++;
1904 break;
1905 }
1906 }
1907 for (;i < 4; i++) {
1908 if (imm_h[i] != 0x0L) {
1909 movk(dst, (uint32_t)imm_h[i], (i << 4));
1910 }
1911 }
1912 } else if (neg_count == 1) {
1913 // one MOVN and two MOVKs will do
1914 for (i = 0; i < 4; i++) {
1915 if (imm_h[i] != 0xffffL) {
1916 movn(dst, (uint32_t)imm_h[i] ^ 0xffffL, (i << 4));
1917 i++;
1918 break;
1919 }
1920 }
1921 for (;i < 4; i++) {
1922 if (imm_h[i] != 0xffffL) {
1923 movk(dst, (uint32_t)imm_h[i], (i << 4));
1924 }
1925 }
1926 } else {
1927 // use a MOVZ and 3 MOVKs (makes it easier to debug)
1928 movz(dst, (uint32_t)imm_h[0], 0);
1929 for (i = 1; i < 4; i++) {
1930 movk(dst, (uint32_t)imm_h[i], (i << 4));
1931 }
1932 }
1933 }
1934 }
1935
1936 void MacroAssembler::mov_immediate32(Register dst, uint32_t imm32)
1937 {
1938 #ifndef PRODUCT
1939 {
1940 char buffer[64];
1941 snprintf(buffer, sizeof(buffer), "0x%" PRIX32, imm32);
1942 block_comment(buffer);
1943 }
1944 #endif
1945 if (operand_valid_for_logical_immediate(true, imm32)) {
1946 orrw(dst, zr, imm32);
1947 } else {
1948 // we can use MOVZ, MOVN or two calls to MOVK to build up the
1949 // constant
1950 uint32_t imm_h[2];
1951 imm_h[0] = imm32 & 0xffff;
1952 imm_h[1] = ((imm32 >> 16) & 0xffff);
1953 if (imm_h[0] == 0) {
1954 movzw(dst, imm_h[1], 16);
1955 } else if (imm_h[0] == 0xffff) {
1956 movnw(dst, imm_h[1] ^ 0xffff, 16);
1957 } else if (imm_h[1] == 0) {
1958 movzw(dst, imm_h[0], 0);
1959 } else if (imm_h[1] == 0xffff) {
1960 movnw(dst, imm_h[0] ^ 0xffff, 0);
1961 } else {
1962 // use a MOVZ and MOVK (makes it easier to debug)
1963 movzw(dst, imm_h[0], 0);
1964 movkw(dst, imm_h[1], 16);
1965 }
1966 }
1967 }
1968
1969 // Form an address from base + offset in Rd. Rd may or may
1970 // not actually be used: you must use the Address that is returned.
1971 // It is up to you to ensure that the shift provided matches the size
1972 // of your data.
1973 Address MacroAssembler::form_address(Register Rd, Register base, int64_t byte_offset, int shift) {
1974 if (Address::offset_ok_for_immed(byte_offset, shift))
1975 // It fits; no need for any heroics
1976 return Address(base, byte_offset);
1977
1978 // Don't do anything clever with negative or misaligned offsets
1979 unsigned mask = (1 << shift) - 1;
1980 if (byte_offset < 0 || byte_offset & mask) {
1981 mov(Rd, byte_offset);
1982 add(Rd, base, Rd);
1983 return Address(Rd);
1984 }
1985
1986 // See if we can do this with two 12-bit offsets
1987 {
1988 uint64_t word_offset = byte_offset >> shift;
1989 uint64_t masked_offset = word_offset & 0xfff000;
1990 if (Address::offset_ok_for_immed(word_offset - masked_offset, 0)
1991 && Assembler::operand_valid_for_add_sub_immediate(masked_offset << shift)) {
1992 add(Rd, base, masked_offset << shift);
1993 word_offset -= masked_offset;
1994 return Address(Rd, word_offset << shift);
1995 }
1996 }
1997
1998 // Do it the hard way
1999 mov(Rd, byte_offset);
2000 add(Rd, base, Rd);
2001 return Address(Rd);
2002 }
2003
2004 int MacroAssembler::corrected_idivl(Register result, Register ra, Register rb,
2005 bool want_remainder, Register scratch)
2006 {
2007 // Full implementation of Java idiv and irem. The function
2008 // returns the (pc) offset of the div instruction - may be needed
2009 // for implicit exceptions.
2010 //
2011 // constraint : ra/rb =/= scratch
2012 // normal case
2013 //
2014 // input : ra: dividend
2015 // rb: divisor
2016 //
2017 // result: either
2018 // quotient (= ra idiv rb)
2019 // remainder (= ra irem rb)
2020
2021 assert(ra != scratch && rb != scratch, "reg cannot be scratch");
2022
2023 int idivl_offset = offset();
2024 if (! want_remainder) {
2025 sdivw(result, ra, rb);
2026 } else {
2027 sdivw(scratch, ra, rb);
2028 Assembler::msubw(result, scratch, rb, ra);
2029 }
2030
2031 return idivl_offset;
2032 }
2033
2034 int MacroAssembler::corrected_idivq(Register result, Register ra, Register rb,
2035 bool want_remainder, Register scratch)
2036 {
2037 // Full implementation of Java ldiv and lrem. The function
2038 // returns the (pc) offset of the div instruction - may be needed
2039 // for implicit exceptions.
2040 //
2041 // constraint : ra/rb =/= scratch
2042 // normal case
2043 //
2044 // input : ra: dividend
2045 // rb: divisor
2046 //
2047 // result: either
2048 // quotient (= ra idiv rb)
2049 // remainder (= ra irem rb)
2050
2051 assert(ra != scratch && rb != scratch, "reg cannot be scratch");
2052
2053 int idivq_offset = offset();
2054 if (! want_remainder) {
2055 sdiv(result, ra, rb);
2056 } else {
2057 sdiv(scratch, ra, rb);
2058 Assembler::msub(result, scratch, rb, ra);
2059 }
2060
2061 return idivq_offset;
2062 }
2063
2064 void MacroAssembler::membar(Membar_mask_bits order_constraint) {
2065 address prev = pc() - NativeMembar::instruction_size;
2066 address last = code()->last_insn();
2067 if (last != nullptr && nativeInstruction_at(last)->is_Membar() && prev == last) {
2068 NativeMembar *bar = NativeMembar_at(prev);
2069 // We are merging two memory barrier instructions. On AArch64 we
2070 // can do this simply by ORing them together.
2071 bar->set_kind(bar->get_kind() | order_constraint);
2072 BLOCK_COMMENT("merged membar");
2073 } else {
2074 code()->set_last_insn(pc());
2075 dmb(Assembler::barrier(order_constraint));
2076 }
2077 }
2078
2079 bool MacroAssembler::try_merge_ldst(Register rt, const Address &adr, size_t size_in_bytes, bool is_store) {
2080 if (ldst_can_merge(rt, adr, size_in_bytes, is_store)) {
2081 merge_ldst(rt, adr, size_in_bytes, is_store);
2082 code()->clear_last_insn();
2083 return true;
2084 } else {
2085 assert(size_in_bytes == 8 || size_in_bytes == 4, "only 8 bytes or 4 bytes load/store is supported.");
2086 const uint64_t mask = size_in_bytes - 1;
2087 if (adr.getMode() == Address::base_plus_offset &&
2088 (adr.offset() & mask) == 0) { // only supports base_plus_offset.
2089 code()->set_last_insn(pc());
2090 }
2091 return false;
2092 }
2093 }
2094
2095 void MacroAssembler::ldr(Register Rx, const Address &adr) {
2096 // We always try to merge two adjacent loads into one ldp.
2097 if (!try_merge_ldst(Rx, adr, 8, false)) {
2098 Assembler::ldr(Rx, adr);
2099 }
2100 }
2101
2102 void MacroAssembler::ldrw(Register Rw, const Address &adr) {
2103 // We always try to merge two adjacent loads into one ldp.
2104 if (!try_merge_ldst(Rw, adr, 4, false)) {
2105 Assembler::ldrw(Rw, adr);
2106 }
2107 }
2108
2109 void MacroAssembler::str(Register Rx, const Address &adr) {
2110 // We always try to merge two adjacent stores into one stp.
2111 if (!try_merge_ldst(Rx, adr, 8, true)) {
2112 Assembler::str(Rx, adr);
2113 }
2114 }
2115
2116 void MacroAssembler::strw(Register Rw, const Address &adr) {
2117 // We always try to merge two adjacent stores into one stp.
2118 if (!try_merge_ldst(Rw, adr, 4, true)) {
2119 Assembler::strw(Rw, adr);
2120 }
2121 }
2122
2123 // MacroAssembler routines found actually to be needed
2124
2125 void MacroAssembler::push(Register src)
2126 {
2127 str(src, Address(pre(esp, -1 * wordSize)));
2128 }
2129
2130 void MacroAssembler::pop(Register dst)
2131 {
2132 ldr(dst, Address(post(esp, 1 * wordSize)));
2133 }
2134
2135 // Note: load_unsigned_short used to be called load_unsigned_word.
2136 int MacroAssembler::load_unsigned_short(Register dst, Address src) {
2137 int off = offset();
2138 ldrh(dst, src);
2139 return off;
2140 }
2141
2142 int MacroAssembler::load_unsigned_byte(Register dst, Address src) {
2143 int off = offset();
2144 ldrb(dst, src);
2145 return off;
2146 }
2147
2148 int MacroAssembler::load_signed_short(Register dst, Address src) {
2149 int off = offset();
2150 ldrsh(dst, src);
2151 return off;
2152 }
2153
2154 int MacroAssembler::load_signed_byte(Register dst, Address src) {
2155 int off = offset();
2156 ldrsb(dst, src);
2157 return off;
2158 }
2159
2160 int MacroAssembler::load_signed_short32(Register dst, Address src) {
2161 int off = offset();
2162 ldrshw(dst, src);
2163 return off;
2164 }
2165
2166 int MacroAssembler::load_signed_byte32(Register dst, Address src) {
2167 int off = offset();
2168 ldrsbw(dst, src);
2169 return off;
2170 }
2171
2172 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed) {
2173 switch (size_in_bytes) {
2174 case 8: ldr(dst, src); break;
2175 case 4: ldrw(dst, src); break;
2176 case 2: is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break;
2177 case 1: is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break;
2178 default: ShouldNotReachHere();
2179 }
2180 }
2181
2182 void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes) {
2183 switch (size_in_bytes) {
2184 case 8: str(src, dst); break;
2185 case 4: strw(src, dst); break;
2186 case 2: strh(src, dst); break;
2187 case 1: strb(src, dst); break;
2188 default: ShouldNotReachHere();
2189 }
2190 }
2191
2192 void MacroAssembler::decrementw(Register reg, int value)
2193 {
2194 if (value < 0) { incrementw(reg, -value); return; }
2195 if (value == 0) { return; }
2196 if (value < (1 << 12)) { subw(reg, reg, value); return; }
2197 /* else */ {
2198 guarantee(reg != rscratch2, "invalid dst for register decrement");
2199 movw(rscratch2, (unsigned)value);
2200 subw(reg, reg, rscratch2);
2201 }
2202 }
2203
2204 void MacroAssembler::decrement(Register reg, int value)
2205 {
2206 if (value < 0) { increment(reg, -value); return; }
2207 if (value == 0) { return; }
2208 if (value < (1 << 12)) { sub(reg, reg, value); return; }
2209 /* else */ {
2210 assert(reg != rscratch2, "invalid dst for register decrement");
2211 mov(rscratch2, (uint64_t)value);
2212 sub(reg, reg, rscratch2);
2213 }
2214 }
2215
2216 void MacroAssembler::decrementw(Address dst, int value)
2217 {
2218 assert(!dst.uses(rscratch1), "invalid dst for address decrement");
2219 if (dst.getMode() == Address::literal) {
2220 assert(abs(value) < (1 << 12), "invalid value and address mode combination");
2221 lea(rscratch2, dst);
2222 dst = Address(rscratch2);
2223 }
2224 ldrw(rscratch1, dst);
2225 decrementw(rscratch1, value);
2226 strw(rscratch1, dst);
2227 }
2228
2229 void MacroAssembler::decrement(Address dst, int value)
2230 {
2231 assert(!dst.uses(rscratch1), "invalid address for decrement");
2232 if (dst.getMode() == Address::literal) {
2233 assert(abs(value) < (1 << 12), "invalid value and address mode combination");
2234 lea(rscratch2, dst);
2235 dst = Address(rscratch2);
2236 }
2237 ldr(rscratch1, dst);
2238 decrement(rscratch1, value);
2239 str(rscratch1, dst);
2240 }
2241
2242 void MacroAssembler::incrementw(Register reg, int value)
2243 {
2244 if (value < 0) { decrementw(reg, -value); return; }
2245 if (value == 0) { return; }
2246 if (value < (1 << 12)) { addw(reg, reg, value); return; }
2247 /* else */ {
2248 assert(reg != rscratch2, "invalid dst for register increment");
2249 movw(rscratch2, (unsigned)value);
2250 addw(reg, reg, rscratch2);
2251 }
2252 }
2253
2254 void MacroAssembler::increment(Register reg, int value)
2255 {
2256 if (value < 0) { decrement(reg, -value); return; }
2257 if (value == 0) { return; }
2258 if (value < (1 << 12)) { add(reg, reg, value); return; }
2259 /* else */ {
2260 assert(reg != rscratch2, "invalid dst for register increment");
2261 movw(rscratch2, (unsigned)value);
2262 add(reg, reg, rscratch2);
2263 }
2264 }
2265
2266 void MacroAssembler::incrementw(Address dst, int value)
2267 {
2268 assert(!dst.uses(rscratch1), "invalid dst for address increment");
2269 if (dst.getMode() == Address::literal) {
2270 assert(abs(value) < (1 << 12), "invalid value and address mode combination");
2271 lea(rscratch2, dst);
2272 dst = Address(rscratch2);
2273 }
2274 ldrw(rscratch1, dst);
2275 incrementw(rscratch1, value);
2276 strw(rscratch1, dst);
2277 }
2278
2279 void MacroAssembler::increment(Address dst, int value)
2280 {
2281 assert(!dst.uses(rscratch1), "invalid dst for address increment");
2282 if (dst.getMode() == Address::literal) {
2283 assert(abs(value) < (1 << 12), "invalid value and address mode combination");
2284 lea(rscratch2, dst);
2285 dst = Address(rscratch2);
2286 }
2287 ldr(rscratch1, dst);
2288 increment(rscratch1, value);
2289 str(rscratch1, dst);
2290 }
2291
2292 // Push lots of registers in the bit set supplied. Don't push sp.
2293 // Return the number of words pushed
2294 int MacroAssembler::push(unsigned int bitset, Register stack) {
2295 int words_pushed = 0;
2296
2297 // Scan bitset to accumulate register pairs
2298 unsigned char regs[32];
2299 int count = 0;
2300 for (int reg = 0; reg <= 30; reg++) {
2301 if (1 & bitset)
2302 regs[count++] = reg;
2303 bitset >>= 1;
2304 }
2305 regs[count++] = zr->raw_encoding();
2306 count &= ~1; // Only push an even number of regs
2307
2308 if (count) {
2309 stp(as_Register(regs[0]), as_Register(regs[1]),
2310 Address(pre(stack, -count * wordSize)));
2311 words_pushed += 2;
2312 }
2313 for (int i = 2; i < count; i += 2) {
2314 stp(as_Register(regs[i]), as_Register(regs[i+1]),
2315 Address(stack, i * wordSize));
2316 words_pushed += 2;
2317 }
2318
2319 assert(words_pushed == count, "oops, pushed != count");
2320
2321 return count;
2322 }
2323
2324 int MacroAssembler::pop(unsigned int bitset, Register stack) {
2325 int words_pushed = 0;
2326
2327 // Scan bitset to accumulate register pairs
2328 unsigned char regs[32];
2329 int count = 0;
2330 for (int reg = 0; reg <= 30; reg++) {
2331 if (1 & bitset)
2332 regs[count++] = reg;
2333 bitset >>= 1;
2334 }
2335 regs[count++] = zr->raw_encoding();
2336 count &= ~1;
2337
2338 for (int i = 2; i < count; i += 2) {
2339 ldp(as_Register(regs[i]), as_Register(regs[i+1]),
2340 Address(stack, i * wordSize));
2341 words_pushed += 2;
2342 }
2343 if (count) {
2344 ldp(as_Register(regs[0]), as_Register(regs[1]),
2345 Address(post(stack, count * wordSize)));
2346 words_pushed += 2;
2347 }
2348
2349 assert(words_pushed == count, "oops, pushed != count");
2350
2351 return count;
2352 }
2353
2354 // Push lots of registers in the bit set supplied. Don't push sp.
2355 // Return the number of dwords pushed
2356 int MacroAssembler::push_fp(unsigned int bitset, Register stack) {
2357 int words_pushed = 0;
2358 bool use_sve = false;
2359 int sve_vector_size_in_bytes = 0;
2360
2361 #ifdef COMPILER2
2362 use_sve = Matcher::supports_scalable_vector();
2363 sve_vector_size_in_bytes = Matcher::scalable_vector_reg_size(T_BYTE);
2364 #endif
2365
2366 // Scan bitset to accumulate register pairs
2367 unsigned char regs[32];
2368 int count = 0;
2369 for (int reg = 0; reg <= 31; reg++) {
2370 if (1 & bitset)
2371 regs[count++] = reg;
2372 bitset >>= 1;
2373 }
2374
2375 if (count == 0) {
2376 return 0;
2377 }
2378
2379 // SVE
2380 if (use_sve && sve_vector_size_in_bytes > 16) {
2381 sub(stack, stack, sve_vector_size_in_bytes * count);
2382 for (int i = 0; i < count; i++) {
2383 sve_str(as_FloatRegister(regs[i]), Address(stack, i));
2384 }
2385 return count * sve_vector_size_in_bytes / 8;
2386 }
2387
2388 // NEON
2389 if (count == 1) {
2390 strq(as_FloatRegister(regs[0]), Address(pre(stack, -wordSize * 2)));
2391 return 2;
2392 }
2393
2394 bool odd = (count & 1) == 1;
2395 int push_slots = count + (odd ? 1 : 0);
2396
2397 // Always pushing full 128 bit registers.
2398 stpq(as_FloatRegister(regs[0]), as_FloatRegister(regs[1]), Address(pre(stack, -push_slots * wordSize * 2)));
2399 words_pushed += 2;
2400
2401 for (int i = 2; i + 1 < count; i += 2) {
2402 stpq(as_FloatRegister(regs[i]), as_FloatRegister(regs[i+1]), Address(stack, i * wordSize * 2));
2403 words_pushed += 2;
2404 }
2405
2406 if (odd) {
2407 strq(as_FloatRegister(regs[count - 1]), Address(stack, (count - 1) * wordSize * 2));
2408 words_pushed++;
2409 }
2410
2411 assert(words_pushed == count, "oops, pushed(%d) != count(%d)", words_pushed, count);
2412 return count * 2;
2413 }
2414
2415 // Return the number of dwords popped
2416 int MacroAssembler::pop_fp(unsigned int bitset, Register stack) {
2417 int words_pushed = 0;
2418 bool use_sve = false;
2419 int sve_vector_size_in_bytes = 0;
2420
2421 #ifdef COMPILER2
2422 use_sve = Matcher::supports_scalable_vector();
2423 sve_vector_size_in_bytes = Matcher::scalable_vector_reg_size(T_BYTE);
2424 #endif
2425 // Scan bitset to accumulate register pairs
2426 unsigned char regs[32];
2427 int count = 0;
2428 for (int reg = 0; reg <= 31; reg++) {
2429 if (1 & bitset)
2430 regs[count++] = reg;
2431 bitset >>= 1;
2432 }
2433
2434 if (count == 0) {
2435 return 0;
2436 }
2437
2438 // SVE
2439 if (use_sve && sve_vector_size_in_bytes > 16) {
2440 for (int i = count - 1; i >= 0; i--) {
2441 sve_ldr(as_FloatRegister(regs[i]), Address(stack, i));
2442 }
2443 add(stack, stack, sve_vector_size_in_bytes * count);
2444 return count * sve_vector_size_in_bytes / 8;
2445 }
2446
2447 // NEON
2448 if (count == 1) {
2449 ldrq(as_FloatRegister(regs[0]), Address(post(stack, wordSize * 2)));
2450 return 2;
2451 }
2452
2453 bool odd = (count & 1) == 1;
2454 int push_slots = count + (odd ? 1 : 0);
2455
2456 if (odd) {
2457 ldrq(as_FloatRegister(regs[count - 1]), Address(stack, (count - 1) * wordSize * 2));
2458 words_pushed++;
2459 }
2460
2461 for (int i = 2; i + 1 < count; i += 2) {
2462 ldpq(as_FloatRegister(regs[i]), as_FloatRegister(regs[i+1]), Address(stack, i * wordSize * 2));
2463 words_pushed += 2;
2464 }
2465
2466 ldpq(as_FloatRegister(regs[0]), as_FloatRegister(regs[1]), Address(post(stack, push_slots * wordSize * 2)));
2467 words_pushed += 2;
2468
2469 assert(words_pushed == count, "oops, pushed(%d) != count(%d)", words_pushed, count);
2470
2471 return count * 2;
2472 }
2473
2474 // Return the number of dwords pushed
2475 int MacroAssembler::push_p(unsigned int bitset, Register stack) {
2476 bool use_sve = false;
2477 int sve_predicate_size_in_slots = 0;
2478
2479 #ifdef COMPILER2
2480 use_sve = Matcher::supports_scalable_vector();
2481 if (use_sve) {
2482 sve_predicate_size_in_slots = Matcher::scalable_predicate_reg_slots();
2483 }
2484 #endif
2485
2486 if (!use_sve) {
2487 return 0;
2488 }
2489
2490 unsigned char regs[PRegister::number_of_registers];
2491 int count = 0;
2492 for (int reg = 0; reg < PRegister::number_of_registers; reg++) {
2493 if (1 & bitset)
2494 regs[count++] = reg;
2495 bitset >>= 1;
2496 }
2497
2498 if (count == 0) {
2499 return 0;
2500 }
2501
2502 int total_push_bytes = align_up(sve_predicate_size_in_slots *
2503 VMRegImpl::stack_slot_size * count, 16);
2504 sub(stack, stack, total_push_bytes);
2505 for (int i = 0; i < count; i++) {
2506 sve_str(as_PRegister(regs[i]), Address(stack, i));
2507 }
2508 return total_push_bytes / 8;
2509 }
2510
2511 // Return the number of dwords popped
2512 int MacroAssembler::pop_p(unsigned int bitset, Register stack) {
2513 bool use_sve = false;
2514 int sve_predicate_size_in_slots = 0;
2515
2516 #ifdef COMPILER2
2517 use_sve = Matcher::supports_scalable_vector();
2518 if (use_sve) {
2519 sve_predicate_size_in_slots = Matcher::scalable_predicate_reg_slots();
2520 }
2521 #endif
2522
2523 if (!use_sve) {
2524 return 0;
2525 }
2526
2527 unsigned char regs[PRegister::number_of_registers];
2528 int count = 0;
2529 for (int reg = 0; reg < PRegister::number_of_registers; reg++) {
2530 if (1 & bitset)
2531 regs[count++] = reg;
2532 bitset >>= 1;
2533 }
2534
2535 if (count == 0) {
2536 return 0;
2537 }
2538
2539 int total_pop_bytes = align_up(sve_predicate_size_in_slots *
2540 VMRegImpl::stack_slot_size * count, 16);
2541 for (int i = count - 1; i >= 0; i--) {
2542 sve_ldr(as_PRegister(regs[i]), Address(stack, i));
2543 }
2544 add(stack, stack, total_pop_bytes);
2545 return total_pop_bytes / 8;
2546 }
2547
2548 #ifdef ASSERT
2549 void MacroAssembler::verify_heapbase(const char* msg) {
2550 #if 0
2551 assert (UseCompressedOops || UseCompressedClassPointers, "should be compressed");
2552 assert (Universe::heap() != nullptr, "java heap should be initialized");
2553 if (!UseCompressedOops || Universe::ptr_base() == nullptr) {
2554 // rheapbase is allocated as general register
2555 return;
2556 }
2557 if (CheckCompressedOops) {
2558 Label ok;
2559 push(1 << rscratch1->encoding(), sp); // cmpptr trashes rscratch1
2560 cmpptr(rheapbase, ExternalAddress(CompressedOops::ptrs_base_addr()));
2561 br(Assembler::EQ, ok);
2562 stop(msg);
2563 bind(ok);
2564 pop(1 << rscratch1->encoding(), sp);
2565 }
2566 #endif
2567 }
2568 #endif
2569
2570 void MacroAssembler::resolve_jobject(Register value, Register tmp1, Register tmp2) {
2571 assert_different_registers(value, tmp1, tmp2);
2572 Label done, tagged, weak_tagged;
2573
2574 cbz(value, done); // Use null as-is.
2575 tst(value, JNIHandles::tag_mask); // Test for tag.
2576 br(Assembler::NE, tagged);
2577
2578 // Resolve local handle
2579 access_load_at(T_OBJECT, IN_NATIVE | AS_RAW, value, Address(value, 0), tmp1, tmp2);
2580 verify_oop(value);
2581 b(done);
2582
2583 bind(tagged);
2584 STATIC_ASSERT(JNIHandles::TypeTag::weak_global == 0b1);
2585 tbnz(value, 0, weak_tagged); // Test for weak tag.
2586
2587 // Resolve global handle
2588 access_load_at(T_OBJECT, IN_NATIVE, value, Address(value, -JNIHandles::TypeTag::global), tmp1, tmp2);
2589 verify_oop(value);
2590 b(done);
2591
2592 bind(weak_tagged);
2593 // Resolve jweak.
2594 access_load_at(T_OBJECT, IN_NATIVE | ON_PHANTOM_OOP_REF,
2595 value, Address(value, -JNIHandles::TypeTag::weak_global), tmp1, tmp2);
2596 verify_oop(value);
2597
2598 bind(done);
2599 }
2600
2601 void MacroAssembler::resolve_global_jobject(Register value, Register tmp1, Register tmp2) {
2602 assert_different_registers(value, tmp1, tmp2);
2603 Label done;
2604
2605 cbz(value, done); // Use null as-is.
2606
2607 #ifdef ASSERT
2608 {
2609 STATIC_ASSERT(JNIHandles::TypeTag::global == 0b10);
2610 Label valid_global_tag;
2611 tbnz(value, 1, valid_global_tag); // Test for global tag
2612 stop("non global jobject using resolve_global_jobject");
2613 bind(valid_global_tag);
2614 }
2615 #endif
2616
2617 // Resolve global handle
2618 access_load_at(T_OBJECT, IN_NATIVE, value, Address(value, -JNIHandles::TypeTag::global), tmp1, tmp2);
2619 verify_oop(value);
2620
2621 bind(done);
2622 }
2623
2624 void MacroAssembler::stop(const char* msg) {
2625 BLOCK_COMMENT(msg);
2626 dcps1(0xdeae);
2627 emit_int64((uintptr_t)msg);
2628 }
2629
2630 void MacroAssembler::unimplemented(const char* what) {
2631 const char* buf = nullptr;
2632 {
2633 ResourceMark rm;
2634 stringStream ss;
2635 ss.print("unimplemented: %s", what);
2636 buf = code_string(ss.as_string());
2637 }
2638 stop(buf);
2639 }
2640
2641 void MacroAssembler::_assert_asm(Assembler::Condition cc, const char* msg) {
2642 #ifdef ASSERT
2643 Label OK;
2644 br(cc, OK);
2645 stop(msg);
2646 bind(OK);
2647 #endif
2648 }
2649
2650 // If a constant does not fit in an immediate field, generate some
2651 // number of MOV instructions and then perform the operation.
2652 void MacroAssembler::wrap_add_sub_imm_insn(Register Rd, Register Rn, uint64_t imm,
2653 add_sub_imm_insn insn1,
2654 add_sub_reg_insn insn2,
2655 bool is32) {
2656 assert(Rd != zr, "Rd = zr and not setting flags?");
2657 bool fits = operand_valid_for_add_sub_immediate(is32 ? (int32_t)imm : imm);
2658 if (fits) {
2659 (this->*insn1)(Rd, Rn, imm);
2660 } else {
2661 if (uabs(imm) < (1 << 24)) {
2662 (this->*insn1)(Rd, Rn, imm & -(1 << 12));
2663 (this->*insn1)(Rd, Rd, imm & ((1 << 12)-1));
2664 } else {
2665 assert_different_registers(Rd, Rn);
2666 mov(Rd, imm);
2667 (this->*insn2)(Rd, Rn, Rd, LSL, 0);
2668 }
2669 }
2670 }
2671
2672 // Separate vsn which sets the flags. Optimisations are more restricted
2673 // because we must set the flags correctly.
2674 void MacroAssembler::wrap_adds_subs_imm_insn(Register Rd, Register Rn, uint64_t imm,
2675 add_sub_imm_insn insn1,
2676 add_sub_reg_insn insn2,
2677 bool is32) {
2678 bool fits = operand_valid_for_add_sub_immediate(is32 ? (int32_t)imm : imm);
2679 if (fits) {
2680 (this->*insn1)(Rd, Rn, imm);
2681 } else {
2682 assert_different_registers(Rd, Rn);
2683 assert(Rd != zr, "overflow in immediate operand");
2684 mov(Rd, imm);
2685 (this->*insn2)(Rd, Rn, Rd, LSL, 0);
2686 }
2687 }
2688
2689
2690 void MacroAssembler::add(Register Rd, Register Rn, RegisterOrConstant increment) {
2691 if (increment.is_register()) {
2692 add(Rd, Rn, increment.as_register());
2693 } else {
2694 add(Rd, Rn, increment.as_constant());
2695 }
2696 }
2697
2698 void MacroAssembler::addw(Register Rd, Register Rn, RegisterOrConstant increment) {
2699 if (increment.is_register()) {
2700 addw(Rd, Rn, increment.as_register());
2701 } else {
2702 addw(Rd, Rn, increment.as_constant());
2703 }
2704 }
2705
2706 void MacroAssembler::sub(Register Rd, Register Rn, RegisterOrConstant decrement) {
2707 if (decrement.is_register()) {
2708 sub(Rd, Rn, decrement.as_register());
2709 } else {
2710 sub(Rd, Rn, decrement.as_constant());
2711 }
2712 }
2713
2714 void MacroAssembler::subw(Register Rd, Register Rn, RegisterOrConstant decrement) {
2715 if (decrement.is_register()) {
2716 subw(Rd, Rn, decrement.as_register());
2717 } else {
2718 subw(Rd, Rn, decrement.as_constant());
2719 }
2720 }
2721
2722 void MacroAssembler::reinit_heapbase()
2723 {
2724 if (UseCompressedOops) {
2725 if (Universe::is_fully_initialized()) {
2726 mov(rheapbase, CompressedOops::ptrs_base());
2727 } else {
2728 lea(rheapbase, ExternalAddress(CompressedOops::ptrs_base_addr()));
2729 ldr(rheapbase, Address(rheapbase));
2730 }
2731 }
2732 }
2733
2734 // this simulates the behaviour of the x86 cmpxchg instruction using a
2735 // load linked/store conditional pair. we use the acquire/release
2736 // versions of these instructions so that we flush pending writes as
2737 // per Java semantics.
2738
2739 // n.b the x86 version assumes the old value to be compared against is
2740 // in rax and updates rax with the value located in memory if the
2741 // cmpxchg fails. we supply a register for the old value explicitly
2742
2743 // the aarch64 load linked/store conditional instructions do not
2744 // accept an offset. so, unlike x86, we must provide a plain register
2745 // to identify the memory word to be compared/exchanged rather than a
2746 // register+offset Address.
2747
2748 void MacroAssembler::cmpxchgptr(Register oldv, Register newv, Register addr, Register tmp,
2749 Label &succeed, Label *fail) {
2750 // oldv holds comparison value
2751 // newv holds value to write in exchange
2752 // addr identifies memory word to compare against/update
2753 if (UseLSE) {
2754 mov(tmp, oldv);
2755 casal(Assembler::xword, oldv, newv, addr);
2756 cmp(tmp, oldv);
2757 br(Assembler::EQ, succeed);
2758 membar(AnyAny);
2759 } else {
2760 Label retry_load, nope;
2761 prfm(Address(addr), PSTL1STRM);
2762 bind(retry_load);
2763 // flush and load exclusive from the memory location
2764 // and fail if it is not what we expect
2765 ldaxr(tmp, addr);
2766 cmp(tmp, oldv);
2767 br(Assembler::NE, nope);
2768 // if we store+flush with no intervening write tmp will be zero
2769 stlxr(tmp, newv, addr);
2770 cbzw(tmp, succeed);
2771 // retry so we only ever return after a load fails to compare
2772 // ensures we don't return a stale value after a failed write.
2773 b(retry_load);
2774 // if the memory word differs we return it in oldv and signal a fail
2775 bind(nope);
2776 membar(AnyAny);
2777 mov(oldv, tmp);
2778 }
2779 if (fail)
2780 b(*fail);
2781 }
2782
2783 void MacroAssembler::cmpxchg_obj_header(Register oldv, Register newv, Register obj, Register tmp,
2784 Label &succeed, Label *fail) {
2785 assert(oopDesc::mark_offset_in_bytes() == 0, "assumption");
2786 cmpxchgptr(oldv, newv, obj, tmp, succeed, fail);
2787 }
2788
2789 void MacroAssembler::cmpxchgw(Register oldv, Register newv, Register addr, Register tmp,
2790 Label &succeed, Label *fail) {
2791 // oldv holds comparison value
2792 // newv holds value to write in exchange
2793 // addr identifies memory word to compare against/update
2794 // tmp returns 0/1 for success/failure
2795 if (UseLSE) {
2796 mov(tmp, oldv);
2797 casal(Assembler::word, oldv, newv, addr);
2798 cmp(tmp, oldv);
2799 br(Assembler::EQ, succeed);
2800 membar(AnyAny);
2801 } else {
2802 Label retry_load, nope;
2803 prfm(Address(addr), PSTL1STRM);
2804 bind(retry_load);
2805 // flush and load exclusive from the memory location
2806 // and fail if it is not what we expect
2807 ldaxrw(tmp, addr);
2808 cmp(tmp, oldv);
2809 br(Assembler::NE, nope);
2810 // if we store+flush with no intervening write tmp will be zero
2811 stlxrw(tmp, newv, addr);
2812 cbzw(tmp, succeed);
2813 // retry so we only ever return after a load fails to compare
2814 // ensures we don't return a stale value after a failed write.
2815 b(retry_load);
2816 // if the memory word differs we return it in oldv and signal a fail
2817 bind(nope);
2818 membar(AnyAny);
2819 mov(oldv, tmp);
2820 }
2821 if (fail)
2822 b(*fail);
2823 }
2824
2825 // A generic CAS; success or failure is in the EQ flag. A weak CAS
2826 // doesn't retry and may fail spuriously. If the oldval is wanted,
2827 // Pass a register for the result, otherwise pass noreg.
2828
2829 // Clobbers rscratch1
2830 void MacroAssembler::cmpxchg(Register addr, Register expected,
2831 Register new_val,
2832 enum operand_size size,
2833 bool acquire, bool release,
2834 bool weak,
2835 Register result) {
2836 if (result == noreg) result = rscratch1;
2837 BLOCK_COMMENT("cmpxchg {");
2838 if (UseLSE) {
2839 mov(result, expected);
2840 lse_cas(result, new_val, addr, size, acquire, release, /*not_pair*/ true);
2841 compare_eq(result, expected, size);
2842 #ifdef ASSERT
2843 // Poison rscratch1 which is written on !UseLSE branch
2844 mov(rscratch1, 0x1f1f1f1f1f1f1f1f);
2845 #endif
2846 } else {
2847 Label retry_load, done;
2848 prfm(Address(addr), PSTL1STRM);
2849 bind(retry_load);
2850 load_exclusive(result, addr, size, acquire);
2851 compare_eq(result, expected, size);
2852 br(Assembler::NE, done);
2853 store_exclusive(rscratch1, new_val, addr, size, release);
2854 if (weak) {
2855 cmpw(rscratch1, 0u); // If the store fails, return NE to our caller.
2856 } else {
2857 cbnzw(rscratch1, retry_load);
2858 }
2859 bind(done);
2860 }
2861 BLOCK_COMMENT("} cmpxchg");
2862 }
2863
2864 // A generic comparison. Only compares for equality, clobbers rscratch1.
2865 void MacroAssembler::compare_eq(Register rm, Register rn, enum operand_size size) {
2866 if (size == xword) {
2867 cmp(rm, rn);
2868 } else if (size == word) {
2869 cmpw(rm, rn);
2870 } else if (size == halfword) {
2871 eorw(rscratch1, rm, rn);
2872 ands(zr, rscratch1, 0xffff);
2873 } else if (size == byte) {
2874 eorw(rscratch1, rm, rn);
2875 ands(zr, rscratch1, 0xff);
2876 } else {
2877 ShouldNotReachHere();
2878 }
2879 }
2880
2881
2882 static bool different(Register a, RegisterOrConstant b, Register c) {
2883 if (b.is_constant())
2884 return a != c;
2885 else
2886 return a != b.as_register() && a != c && b.as_register() != c;
2887 }
2888
2889 #define ATOMIC_OP(NAME, LDXR, OP, IOP, AOP, STXR, sz) \
2890 void MacroAssembler::atomic_##NAME(Register prev, RegisterOrConstant incr, Register addr) { \
2891 if (UseLSE) { \
2892 prev = prev->is_valid() ? prev : zr; \
2893 if (incr.is_register()) { \
2894 AOP(sz, incr.as_register(), prev, addr); \
2895 } else { \
2896 mov(rscratch2, incr.as_constant()); \
2897 AOP(sz, rscratch2, prev, addr); \
2898 } \
2899 return; \
2900 } \
2901 Register result = rscratch2; \
2902 if (prev->is_valid()) \
2903 result = different(prev, incr, addr) ? prev : rscratch2; \
2904 \
2905 Label retry_load; \
2906 prfm(Address(addr), PSTL1STRM); \
2907 bind(retry_load); \
2908 LDXR(result, addr); \
2909 OP(rscratch1, result, incr); \
2910 STXR(rscratch2, rscratch1, addr); \
2911 cbnzw(rscratch2, retry_load); \
2912 if (prev->is_valid() && prev != result) { \
2913 IOP(prev, rscratch1, incr); \
2914 } \
2915 }
2916
2917 ATOMIC_OP(add, ldxr, add, sub, ldadd, stxr, Assembler::xword)
2918 ATOMIC_OP(addw, ldxrw, addw, subw, ldadd, stxrw, Assembler::word)
2919 ATOMIC_OP(addal, ldaxr, add, sub, ldaddal, stlxr, Assembler::xword)
2920 ATOMIC_OP(addalw, ldaxrw, addw, subw, ldaddal, stlxrw, Assembler::word)
2921
2922 #undef ATOMIC_OP
2923
2924 #define ATOMIC_XCHG(OP, AOP, LDXR, STXR, sz) \
2925 void MacroAssembler::atomic_##OP(Register prev, Register newv, Register addr) { \
2926 if (UseLSE) { \
2927 prev = prev->is_valid() ? prev : zr; \
2928 AOP(sz, newv, prev, addr); \
2929 return; \
2930 } \
2931 Register result = rscratch2; \
2932 if (prev->is_valid()) \
2933 result = different(prev, newv, addr) ? prev : rscratch2; \
2934 \
2935 Label retry_load; \
2936 prfm(Address(addr), PSTL1STRM); \
2937 bind(retry_load); \
2938 LDXR(result, addr); \
2939 STXR(rscratch1, newv, addr); \
2940 cbnzw(rscratch1, retry_load); \
2941 if (prev->is_valid() && prev != result) \
2942 mov(prev, result); \
2943 }
2944
2945 ATOMIC_XCHG(xchg, swp, ldxr, stxr, Assembler::xword)
2946 ATOMIC_XCHG(xchgw, swp, ldxrw, stxrw, Assembler::word)
2947 ATOMIC_XCHG(xchgl, swpl, ldxr, stlxr, Assembler::xword)
2948 ATOMIC_XCHG(xchglw, swpl, ldxrw, stlxrw, Assembler::word)
2949 ATOMIC_XCHG(xchgal, swpal, ldaxr, stlxr, Assembler::xword)
2950 ATOMIC_XCHG(xchgalw, swpal, ldaxrw, stlxrw, Assembler::word)
2951
2952 #undef ATOMIC_XCHG
2953
2954 #ifndef PRODUCT
2955 extern "C" void findpc(intptr_t x);
2956 #endif
2957
2958 void MacroAssembler::debug64(char* msg, int64_t pc, int64_t regs[])
2959 {
2960 // In order to get locks to work, we need to fake a in_VM state
2961 if (ShowMessageBoxOnError ) {
2962 JavaThread* thread = JavaThread::current();
2963 JavaThreadState saved_state = thread->thread_state();
2964 thread->set_thread_state(_thread_in_vm);
2965 #ifndef PRODUCT
2966 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) {
2967 ttyLocker ttyl;
2968 BytecodeCounter::print();
2969 }
2970 #endif
2971 if (os::message_box(msg, "Execution stopped, print registers?")) {
2972 ttyLocker ttyl;
2973 tty->print_cr(" pc = 0x%016" PRIx64, pc);
2974 #ifndef PRODUCT
2975 tty->cr();
2976 findpc(pc);
2977 tty->cr();
2978 #endif
2979 tty->print_cr(" r0 = 0x%016" PRIx64, regs[0]);
2980 tty->print_cr(" r1 = 0x%016" PRIx64, regs[1]);
2981 tty->print_cr(" r2 = 0x%016" PRIx64, regs[2]);
2982 tty->print_cr(" r3 = 0x%016" PRIx64, regs[3]);
2983 tty->print_cr(" r4 = 0x%016" PRIx64, regs[4]);
2984 tty->print_cr(" r5 = 0x%016" PRIx64, regs[5]);
2985 tty->print_cr(" r6 = 0x%016" PRIx64, regs[6]);
2986 tty->print_cr(" r7 = 0x%016" PRIx64, regs[7]);
2987 tty->print_cr(" r8 = 0x%016" PRIx64, regs[8]);
2988 tty->print_cr(" r9 = 0x%016" PRIx64, regs[9]);
2989 tty->print_cr("r10 = 0x%016" PRIx64, regs[10]);
2990 tty->print_cr("r11 = 0x%016" PRIx64, regs[11]);
2991 tty->print_cr("r12 = 0x%016" PRIx64, regs[12]);
2992 tty->print_cr("r13 = 0x%016" PRIx64, regs[13]);
2993 tty->print_cr("r14 = 0x%016" PRIx64, regs[14]);
2994 tty->print_cr("r15 = 0x%016" PRIx64, regs[15]);
2995 tty->print_cr("r16 = 0x%016" PRIx64, regs[16]);
2996 tty->print_cr("r17 = 0x%016" PRIx64, regs[17]);
2997 tty->print_cr("r18 = 0x%016" PRIx64, regs[18]);
2998 tty->print_cr("r19 = 0x%016" PRIx64, regs[19]);
2999 tty->print_cr("r20 = 0x%016" PRIx64, regs[20]);
3000 tty->print_cr("r21 = 0x%016" PRIx64, regs[21]);
3001 tty->print_cr("r22 = 0x%016" PRIx64, regs[22]);
3002 tty->print_cr("r23 = 0x%016" PRIx64, regs[23]);
3003 tty->print_cr("r24 = 0x%016" PRIx64, regs[24]);
3004 tty->print_cr("r25 = 0x%016" PRIx64, regs[25]);
3005 tty->print_cr("r26 = 0x%016" PRIx64, regs[26]);
3006 tty->print_cr("r27 = 0x%016" PRIx64, regs[27]);
3007 tty->print_cr("r28 = 0x%016" PRIx64, regs[28]);
3008 tty->print_cr("r30 = 0x%016" PRIx64, regs[30]);
3009 tty->print_cr("r31 = 0x%016" PRIx64, regs[31]);
3010 BREAKPOINT;
3011 }
3012 }
3013 fatal("DEBUG MESSAGE: %s", msg);
3014 }
3015
3016 RegSet MacroAssembler::call_clobbered_gp_registers() {
3017 RegSet regs = RegSet::range(r0, r17) - RegSet::of(rscratch1, rscratch2);
3018 #ifndef R18_RESERVED
3019 regs += r18_tls;
3020 #endif
3021 return regs;
3022 }
3023
3024 void MacroAssembler::push_call_clobbered_registers_except(RegSet exclude) {
3025 int step = 4 * wordSize;
3026 push(call_clobbered_gp_registers() - exclude, sp);
3027 sub(sp, sp, step);
3028 mov(rscratch1, -step);
3029 // Push v0-v7, v16-v31.
3030 for (int i = 31; i>= 4; i -= 4) {
3031 if (i <= v7->encoding() || i >= v16->encoding())
3032 st1(as_FloatRegister(i-3), as_FloatRegister(i-2), as_FloatRegister(i-1),
3033 as_FloatRegister(i), T1D, Address(post(sp, rscratch1)));
3034 }
3035 st1(as_FloatRegister(0), as_FloatRegister(1), as_FloatRegister(2),
3036 as_FloatRegister(3), T1D, Address(sp));
3037 }
3038
3039 void MacroAssembler::pop_call_clobbered_registers_except(RegSet exclude) {
3040 for (int i = 0; i < 32; i += 4) {
3041 if (i <= v7->encoding() || i >= v16->encoding())
3042 ld1(as_FloatRegister(i), as_FloatRegister(i+1), as_FloatRegister(i+2),
3043 as_FloatRegister(i+3), T1D, Address(post(sp, 4 * wordSize)));
3044 }
3045
3046 reinitialize_ptrue();
3047
3048 pop(call_clobbered_gp_registers() - exclude, sp);
3049 }
3050
3051 void MacroAssembler::push_CPU_state(bool save_vectors, bool use_sve,
3052 int sve_vector_size_in_bytes, int total_predicate_in_bytes) {
3053 push(RegSet::range(r0, r29), sp); // integer registers except lr & sp
3054 if (save_vectors && use_sve && sve_vector_size_in_bytes > 16) {
3055 sub(sp, sp, sve_vector_size_in_bytes * FloatRegister::number_of_registers);
3056 for (int i = 0; i < FloatRegister::number_of_registers; i++) {
3057 sve_str(as_FloatRegister(i), Address(sp, i));
3058 }
3059 } else {
3060 int step = (save_vectors ? 8 : 4) * wordSize;
3061 mov(rscratch1, -step);
3062 sub(sp, sp, step);
3063 for (int i = 28; i >= 4; i -= 4) {
3064 st1(as_FloatRegister(i), as_FloatRegister(i+1), as_FloatRegister(i+2),
3065 as_FloatRegister(i+3), save_vectors ? T2D : T1D, Address(post(sp, rscratch1)));
3066 }
3067 st1(v0, v1, v2, v3, save_vectors ? T2D : T1D, sp);
3068 }
3069 if (save_vectors && use_sve && total_predicate_in_bytes > 0) {
3070 sub(sp, sp, total_predicate_in_bytes);
3071 for (int i = 0; i < PRegister::number_of_registers; i++) {
3072 sve_str(as_PRegister(i), Address(sp, i));
3073 }
3074 }
3075 }
3076
3077 void MacroAssembler::pop_CPU_state(bool restore_vectors, bool use_sve,
3078 int sve_vector_size_in_bytes, int total_predicate_in_bytes) {
3079 if (restore_vectors && use_sve && total_predicate_in_bytes > 0) {
3080 for (int i = PRegister::number_of_registers - 1; i >= 0; i--) {
3081 sve_ldr(as_PRegister(i), Address(sp, i));
3082 }
3083 add(sp, sp, total_predicate_in_bytes);
3084 }
3085 if (restore_vectors && use_sve && sve_vector_size_in_bytes > 16) {
3086 for (int i = FloatRegister::number_of_registers - 1; i >= 0; i--) {
3087 sve_ldr(as_FloatRegister(i), Address(sp, i));
3088 }
3089 add(sp, sp, sve_vector_size_in_bytes * FloatRegister::number_of_registers);
3090 } else {
3091 int step = (restore_vectors ? 8 : 4) * wordSize;
3092 for (int i = 0; i <= 28; i += 4)
3093 ld1(as_FloatRegister(i), as_FloatRegister(i+1), as_FloatRegister(i+2),
3094 as_FloatRegister(i+3), restore_vectors ? T2D : T1D, Address(post(sp, step)));
3095 }
3096
3097 // We may use predicate registers and rely on ptrue with SVE,
3098 // regardless of wide vector (> 8 bytes) used or not.
3099 if (use_sve) {
3100 reinitialize_ptrue();
3101 }
3102
3103 // integer registers except lr & sp
3104 pop(RegSet::range(r0, r17), sp);
3105 #ifdef R18_RESERVED
3106 ldp(zr, r19, Address(post(sp, 2 * wordSize)));
3107 pop(RegSet::range(r20, r29), sp);
3108 #else
3109 pop(RegSet::range(r18_tls, r29), sp);
3110 #endif
3111 }
3112
3113 /**
3114 * Helpers for multiply_to_len().
3115 */
3116 void MacroAssembler::add2_with_carry(Register final_dest_hi, Register dest_hi, Register dest_lo,
3117 Register src1, Register src2) {
3118 adds(dest_lo, dest_lo, src1);
3119 adc(dest_hi, dest_hi, zr);
3120 adds(dest_lo, dest_lo, src2);
3121 adc(final_dest_hi, dest_hi, zr);
3122 }
3123
3124 // Generate an address from (r + r1 extend offset). "size" is the
3125 // size of the operand. The result may be in rscratch2.
3126 Address MacroAssembler::offsetted_address(Register r, Register r1,
3127 Address::extend ext, int offset, int size) {
3128 if (offset || (ext.shift() % size != 0)) {
3129 lea(rscratch2, Address(r, r1, ext));
3130 return Address(rscratch2, offset);
3131 } else {
3132 return Address(r, r1, ext);
3133 }
3134 }
3135
3136 Address MacroAssembler::spill_address(int size, int offset, Register tmp)
3137 {
3138 assert(offset >= 0, "spill to negative address?");
3139 // Offset reachable ?
3140 // Not aligned - 9 bits signed offset
3141 // Aligned - 12 bits unsigned offset shifted
3142 Register base = sp;
3143 if ((offset & (size-1)) && offset >= (1<<8)) {
3144 add(tmp, base, offset & ((1<<12)-1));
3145 base = tmp;
3146 offset &= -1u<<12;
3147 }
3148
3149 if (offset >= (1<<12) * size) {
3150 add(tmp, base, offset & (((1<<12)-1)<<12));
3151 base = tmp;
3152 offset &= ~(((1<<12)-1)<<12);
3153 }
3154
3155 return Address(base, offset);
3156 }
3157
3158 Address MacroAssembler::sve_spill_address(int sve_reg_size_in_bytes, int offset, Register tmp) {
3159 assert(offset >= 0, "spill to negative address?");
3160
3161 Register base = sp;
3162
3163 // An immediate offset in the range 0 to 255 which is multiplied
3164 // by the current vector or predicate register size in bytes.
3165 if (offset % sve_reg_size_in_bytes == 0 && offset < ((1<<8)*sve_reg_size_in_bytes)) {
3166 return Address(base, offset / sve_reg_size_in_bytes);
3167 }
3168
3169 add(tmp, base, offset);
3170 return Address(tmp);
3171 }
3172
3173 // Checks whether offset is aligned.
3174 // Returns true if it is, else false.
3175 bool MacroAssembler::merge_alignment_check(Register base,
3176 size_t size,
3177 int64_t cur_offset,
3178 int64_t prev_offset) const {
3179 if (AvoidUnalignedAccesses) {
3180 if (base == sp) {
3181 // Checks whether low offset if aligned to pair of registers.
3182 int64_t pair_mask = size * 2 - 1;
3183 int64_t offset = prev_offset > cur_offset ? cur_offset : prev_offset;
3184 return (offset & pair_mask) == 0;
3185 } else { // If base is not sp, we can't guarantee the access is aligned.
3186 return false;
3187 }
3188 } else {
3189 int64_t mask = size - 1;
3190 // Load/store pair instruction only supports element size aligned offset.
3191 return (cur_offset & mask) == 0 && (prev_offset & mask) == 0;
3192 }
3193 }
3194
3195 // Checks whether current and previous loads/stores can be merged.
3196 // Returns true if it can be merged, else false.
3197 bool MacroAssembler::ldst_can_merge(Register rt,
3198 const Address &adr,
3199 size_t cur_size_in_bytes,
3200 bool is_store) const {
3201 address prev = pc() - NativeInstruction::instruction_size;
3202 address last = code()->last_insn();
3203
3204 if (last == nullptr || !nativeInstruction_at(last)->is_Imm_LdSt()) {
3205 return false;
3206 }
3207
3208 if (adr.getMode() != Address::base_plus_offset || prev != last) {
3209 return false;
3210 }
3211
3212 NativeLdSt* prev_ldst = NativeLdSt_at(prev);
3213 size_t prev_size_in_bytes = prev_ldst->size_in_bytes();
3214
3215 assert(prev_size_in_bytes == 4 || prev_size_in_bytes == 8, "only supports 64/32bit merging.");
3216 assert(cur_size_in_bytes == 4 || cur_size_in_bytes == 8, "only supports 64/32bit merging.");
3217
3218 if (cur_size_in_bytes != prev_size_in_bytes || is_store != prev_ldst->is_store()) {
3219 return false;
3220 }
3221
3222 int64_t max_offset = 63 * prev_size_in_bytes;
3223 int64_t min_offset = -64 * prev_size_in_bytes;
3224
3225 assert(prev_ldst->is_not_pre_post_index(), "pre-index or post-index is not supported to be merged.");
3226
3227 // Only same base can be merged.
3228 if (adr.base() != prev_ldst->base()) {
3229 return false;
3230 }
3231
3232 int64_t cur_offset = adr.offset();
3233 int64_t prev_offset = prev_ldst->offset();
3234 size_t diff = abs(cur_offset - prev_offset);
3235 if (diff != prev_size_in_bytes) {
3236 return false;
3237 }
3238
3239 // Following cases can not be merged:
3240 // ldr x2, [x2, #8]
3241 // ldr x3, [x2, #16]
3242 // or:
3243 // ldr x2, [x3, #8]
3244 // ldr x2, [x3, #16]
3245 // If t1 and t2 is the same in "ldp t1, t2, [xn, #imm]", we'll get SIGILL.
3246 if (!is_store && (adr.base() == prev_ldst->target() || rt == prev_ldst->target())) {
3247 return false;
3248 }
3249
3250 int64_t low_offset = prev_offset > cur_offset ? cur_offset : prev_offset;
3251 // Offset range must be in ldp/stp instruction's range.
3252 if (low_offset > max_offset || low_offset < min_offset) {
3253 return false;
3254 }
3255
3256 if (merge_alignment_check(adr.base(), prev_size_in_bytes, cur_offset, prev_offset)) {
3257 return true;
3258 }
3259
3260 return false;
3261 }
3262
3263 // Merge current load/store with previous load/store into ldp/stp.
3264 void MacroAssembler::merge_ldst(Register rt,
3265 const Address &adr,
3266 size_t cur_size_in_bytes,
3267 bool is_store) {
3268
3269 assert(ldst_can_merge(rt, adr, cur_size_in_bytes, is_store) == true, "cur and prev must be able to be merged.");
3270
3271 Register rt_low, rt_high;
3272 address prev = pc() - NativeInstruction::instruction_size;
3273 NativeLdSt* prev_ldst = NativeLdSt_at(prev);
3274
3275 int64_t offset;
3276
3277 if (adr.offset() < prev_ldst->offset()) {
3278 offset = adr.offset();
3279 rt_low = rt;
3280 rt_high = prev_ldst->target();
3281 } else {
3282 offset = prev_ldst->offset();
3283 rt_low = prev_ldst->target();
3284 rt_high = rt;
3285 }
3286
3287 Address adr_p = Address(prev_ldst->base(), offset);
3288 // Overwrite previous generated binary.
3289 code_section()->set_end(prev);
3290
3291 const size_t sz = prev_ldst->size_in_bytes();
3292 assert(sz == 8 || sz == 4, "only supports 64/32bit merging.");
3293 if (!is_store) {
3294 BLOCK_COMMENT("merged ldr pair");
3295 if (sz == 8) {
3296 ldp(rt_low, rt_high, adr_p);
3297 } else {
3298 ldpw(rt_low, rt_high, adr_p);
3299 }
3300 } else {
3301 BLOCK_COMMENT("merged str pair");
3302 if (sz == 8) {
3303 stp(rt_low, rt_high, adr_p);
3304 } else {
3305 stpw(rt_low, rt_high, adr_p);
3306 }
3307 }
3308 }
3309
3310 /**
3311 * Multiply 64 bit by 64 bit first loop.
3312 */
3313 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart, Register x_xstart,
3314 Register y, Register y_idx, Register z,
3315 Register carry, Register product,
3316 Register idx, Register kdx) {
3317 //
3318 // jlong carry, x[], y[], z[];
3319 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
3320 // huge_128 product = y[idx] * x[xstart] + carry;
3321 // z[kdx] = (jlong)product;
3322 // carry = (jlong)(product >>> 64);
3323 // }
3324 // z[xstart] = carry;
3325 //
3326
3327 Label L_first_loop, L_first_loop_exit;
3328 Label L_one_x, L_one_y, L_multiply;
3329
3330 subsw(xstart, xstart, 1);
3331 br(Assembler::MI, L_one_x);
3332
3333 lea(rscratch1, Address(x, xstart, Address::lsl(LogBytesPerInt)));
3334 ldr(x_xstart, Address(rscratch1));
3335 ror(x_xstart, x_xstart, 32); // convert big-endian to little-endian
3336
3337 bind(L_first_loop);
3338 subsw(idx, idx, 1);
3339 br(Assembler::MI, L_first_loop_exit);
3340 subsw(idx, idx, 1);
3341 br(Assembler::MI, L_one_y);
3342 lea(rscratch1, Address(y, idx, Address::uxtw(LogBytesPerInt)));
3343 ldr(y_idx, Address(rscratch1));
3344 ror(y_idx, y_idx, 32); // convert big-endian to little-endian
3345 bind(L_multiply);
3346
3347 // AArch64 has a multiply-accumulate instruction that we can't use
3348 // here because it has no way to process carries, so we have to use
3349 // separate add and adc instructions. Bah.
3350 umulh(rscratch1, x_xstart, y_idx); // x_xstart * y_idx -> rscratch1:product
3351 mul(product, x_xstart, y_idx);
3352 adds(product, product, carry);
3353 adc(carry, rscratch1, zr); // x_xstart * y_idx + carry -> carry:product
3354
3355 subw(kdx, kdx, 2);
3356 ror(product, product, 32); // back to big-endian
3357 str(product, offsetted_address(z, kdx, Address::uxtw(LogBytesPerInt), 0, BytesPerLong));
3358
3359 b(L_first_loop);
3360
3361 bind(L_one_y);
3362 ldrw(y_idx, Address(y, 0));
3363 b(L_multiply);
3364
3365 bind(L_one_x);
3366 ldrw(x_xstart, Address(x, 0));
3367 b(L_first_loop);
3368
3369 bind(L_first_loop_exit);
3370 }
3371
3372 /**
3373 * Multiply 128 bit by 128. Unrolled inner loop.
3374 *
3375 */
3376 void MacroAssembler::multiply_128_x_128_loop(Register y, Register z,
3377 Register carry, Register carry2,
3378 Register idx, Register jdx,
3379 Register yz_idx1, Register yz_idx2,
3380 Register tmp, Register tmp3, Register tmp4,
3381 Register tmp6, Register product_hi) {
3382
3383 // jlong carry, x[], y[], z[];
3384 // int kdx = ystart+1;
3385 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
3386 // huge_128 tmp3 = (y[idx+1] * product_hi) + z[kdx+idx+1] + carry;
3387 // jlong carry2 = (jlong)(tmp3 >>> 64);
3388 // huge_128 tmp4 = (y[idx] * product_hi) + z[kdx+idx] + carry2;
3389 // carry = (jlong)(tmp4 >>> 64);
3390 // z[kdx+idx+1] = (jlong)tmp3;
3391 // z[kdx+idx] = (jlong)tmp4;
3392 // }
3393 // idx += 2;
3394 // if (idx > 0) {
3395 // yz_idx1 = (y[idx] * product_hi) + z[kdx+idx] + carry;
3396 // z[kdx+idx] = (jlong)yz_idx1;
3397 // carry = (jlong)(yz_idx1 >>> 64);
3398 // }
3399 //
3400
3401 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
3402
3403 lsrw(jdx, idx, 2);
3404
3405 bind(L_third_loop);
3406
3407 subsw(jdx, jdx, 1);
3408 br(Assembler::MI, L_third_loop_exit);
3409 subw(idx, idx, 4);
3410
3411 lea(rscratch1, Address(y, idx, Address::uxtw(LogBytesPerInt)));
3412
3413 ldp(yz_idx2, yz_idx1, Address(rscratch1, 0));
3414
3415 lea(tmp6, Address(z, idx, Address::uxtw(LogBytesPerInt)));
3416
3417 ror(yz_idx1, yz_idx1, 32); // convert big-endian to little-endian
3418 ror(yz_idx2, yz_idx2, 32);
3419
3420 ldp(rscratch2, rscratch1, Address(tmp6, 0));
3421
3422 mul(tmp3, product_hi, yz_idx1); // yz_idx1 * product_hi -> tmp4:tmp3
3423 umulh(tmp4, product_hi, yz_idx1);
3424
3425 ror(rscratch1, rscratch1, 32); // convert big-endian to little-endian
3426 ror(rscratch2, rscratch2, 32);
3427
3428 mul(tmp, product_hi, yz_idx2); // yz_idx2 * product_hi -> carry2:tmp
3429 umulh(carry2, product_hi, yz_idx2);
3430
3431 // propagate sum of both multiplications into carry:tmp4:tmp3
3432 adds(tmp3, tmp3, carry);
3433 adc(tmp4, tmp4, zr);
3434 adds(tmp3, tmp3, rscratch1);
3435 adcs(tmp4, tmp4, tmp);
3436 adc(carry, carry2, zr);
3437 adds(tmp4, tmp4, rscratch2);
3438 adc(carry, carry, zr);
3439
3440 ror(tmp3, tmp3, 32); // convert little-endian to big-endian
3441 ror(tmp4, tmp4, 32);
3442 stp(tmp4, tmp3, Address(tmp6, 0));
3443
3444 b(L_third_loop);
3445 bind (L_third_loop_exit);
3446
3447 andw (idx, idx, 0x3);
3448 cbz(idx, L_post_third_loop_done);
3449
3450 Label L_check_1;
3451 subsw(idx, idx, 2);
3452 br(Assembler::MI, L_check_1);
3453
3454 lea(rscratch1, Address(y, idx, Address::uxtw(LogBytesPerInt)));
3455 ldr(yz_idx1, Address(rscratch1, 0));
3456 ror(yz_idx1, yz_idx1, 32);
3457 mul(tmp3, product_hi, yz_idx1); // yz_idx1 * product_hi -> tmp4:tmp3
3458 umulh(tmp4, product_hi, yz_idx1);
3459 lea(rscratch1, Address(z, idx, Address::uxtw(LogBytesPerInt)));
3460 ldr(yz_idx2, Address(rscratch1, 0));
3461 ror(yz_idx2, yz_idx2, 32);
3462
3463 add2_with_carry(carry, tmp4, tmp3, carry, yz_idx2);
3464
3465 ror(tmp3, tmp3, 32);
3466 str(tmp3, Address(rscratch1, 0));
3467
3468 bind (L_check_1);
3469
3470 andw (idx, idx, 0x1);
3471 subsw(idx, idx, 1);
3472 br(Assembler::MI, L_post_third_loop_done);
3473 ldrw(tmp4, Address(y, idx, Address::uxtw(LogBytesPerInt)));
3474 mul(tmp3, tmp4, product_hi); // tmp4 * product_hi -> carry2:tmp3
3475 umulh(carry2, tmp4, product_hi);
3476 ldrw(tmp4, Address(z, idx, Address::uxtw(LogBytesPerInt)));
3477
3478 add2_with_carry(carry2, tmp3, tmp4, carry);
3479
3480 strw(tmp3, Address(z, idx, Address::uxtw(LogBytesPerInt)));
3481 extr(carry, carry2, tmp3, 32);
3482
3483 bind(L_post_third_loop_done);
3484 }
3485
3486 /**
3487 * Code for BigInteger::multiplyToLen() intrinsic.
3488 *
3489 * r0: x
3490 * r1: xlen
3491 * r2: y
3492 * r3: ylen
3493 * r4: z
3494 * r5: zlen
3495 * r10: tmp1
3496 * r11: tmp2
3497 * r12: tmp3
3498 * r13: tmp4
3499 * r14: tmp5
3500 * r15: tmp6
3501 * r16: tmp7
3502 *
3503 */
3504 void MacroAssembler::multiply_to_len(Register x, Register xlen, Register y, Register ylen,
3505 Register z, Register zlen,
3506 Register tmp1, Register tmp2, Register tmp3, Register tmp4,
3507 Register tmp5, Register tmp6, Register product_hi) {
3508
3509 assert_different_registers(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6);
3510
3511 const Register idx = tmp1;
3512 const Register kdx = tmp2;
3513 const Register xstart = tmp3;
3514
3515 const Register y_idx = tmp4;
3516 const Register carry = tmp5;
3517 const Register product = xlen;
3518 const Register x_xstart = zlen; // reuse register
3519
3520 // First Loop.
3521 //
3522 // final static long LONG_MASK = 0xffffffffL;
3523 // int xstart = xlen - 1;
3524 // int ystart = ylen - 1;
3525 // long carry = 0;
3526 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
3527 // long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry;
3528 // z[kdx] = (int)product;
3529 // carry = product >>> 32;
3530 // }
3531 // z[xstart] = (int)carry;
3532 //
3533
3534 movw(idx, ylen); // idx = ylen;
3535 movw(kdx, zlen); // kdx = xlen+ylen;
3536 mov(carry, zr); // carry = 0;
3537
3538 Label L_done;
3539
3540 movw(xstart, xlen);
3541 subsw(xstart, xstart, 1);
3542 br(Assembler::MI, L_done);
3543
3544 multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx);
3545
3546 Label L_second_loop;
3547 cbzw(kdx, L_second_loop);
3548
3549 Label L_carry;
3550 subw(kdx, kdx, 1);
3551 cbzw(kdx, L_carry);
3552
3553 strw(carry, Address(z, kdx, Address::uxtw(LogBytesPerInt)));
3554 lsr(carry, carry, 32);
3555 subw(kdx, kdx, 1);
3556
3557 bind(L_carry);
3558 strw(carry, Address(z, kdx, Address::uxtw(LogBytesPerInt)));
3559
3560 // Second and third (nested) loops.
3561 //
3562 // for (int i = xstart-1; i >= 0; i--) { // Second loop
3563 // carry = 0;
3564 // for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop
3565 // long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) +
3566 // (z[k] & LONG_MASK) + carry;
3567 // z[k] = (int)product;
3568 // carry = product >>> 32;
3569 // }
3570 // z[i] = (int)carry;
3571 // }
3572 //
3573 // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = product_hi
3574
3575 const Register jdx = tmp1;
3576
3577 bind(L_second_loop);
3578 mov(carry, zr); // carry = 0;
3579 movw(jdx, ylen); // j = ystart+1
3580
3581 subsw(xstart, xstart, 1); // i = xstart-1;
3582 br(Assembler::MI, L_done);
3583
3584 str(z, Address(pre(sp, -4 * wordSize)));
3585
3586 Label L_last_x;
3587 lea(z, offsetted_address(z, xstart, Address::uxtw(LogBytesPerInt), 4, BytesPerInt)); // z = z + k - j
3588 subsw(xstart, xstart, 1); // i = xstart-1;
3589 br(Assembler::MI, L_last_x);
3590
3591 lea(rscratch1, Address(x, xstart, Address::uxtw(LogBytesPerInt)));
3592 ldr(product_hi, Address(rscratch1));
3593 ror(product_hi, product_hi, 32); // convert big-endian to little-endian
3594
3595 Label L_third_loop_prologue;
3596 bind(L_third_loop_prologue);
3597
3598 str(ylen, Address(sp, wordSize));
3599 stp(x, xstart, Address(sp, 2 * wordSize));
3600 multiply_128_x_128_loop(y, z, carry, x, jdx, ylen, product,
3601 tmp2, x_xstart, tmp3, tmp4, tmp6, product_hi);
3602 ldp(z, ylen, Address(post(sp, 2 * wordSize)));
3603 ldp(x, xlen, Address(post(sp, 2 * wordSize))); // copy old xstart -> xlen
3604
3605 addw(tmp3, xlen, 1);
3606 strw(carry, Address(z, tmp3, Address::uxtw(LogBytesPerInt)));
3607 subsw(tmp3, tmp3, 1);
3608 br(Assembler::MI, L_done);
3609
3610 lsr(carry, carry, 32);
3611 strw(carry, Address(z, tmp3, Address::uxtw(LogBytesPerInt)));
3612 b(L_second_loop);
3613
3614 // Next infrequent code is moved outside loops.
3615 bind(L_last_x);
3616 ldrw(product_hi, Address(x, 0));
3617 b(L_third_loop_prologue);
3618
3619 bind(L_done);
3620 }
3621
3622 // Code for BigInteger::mulAdd intrinsic
3623 // out = r0
3624 // in = r1
3625 // offset = r2 (already out.length-offset)
3626 // len = r3
3627 // k = r4
3628 //
3629 // pseudo code from java implementation:
3630 // carry = 0;
3631 // offset = out.length-offset - 1;
3632 // for (int j=len-1; j >= 0; j--) {
3633 // product = (in[j] & LONG_MASK) * kLong + (out[offset] & LONG_MASK) + carry;
3634 // out[offset--] = (int)product;
3635 // carry = product >>> 32;
3636 // }
3637 // return (int)carry;
3638 void MacroAssembler::mul_add(Register out, Register in, Register offset,
3639 Register len, Register k) {
3640 Label LOOP, END;
3641 // pre-loop
3642 cmp(len, zr); // cmp, not cbz/cbnz: to use condition twice => less branches
3643 csel(out, zr, out, Assembler::EQ);
3644 br(Assembler::EQ, END);
3645 add(in, in, len, LSL, 2); // in[j+1] address
3646 add(offset, out, offset, LSL, 2); // out[offset + 1] address
3647 mov(out, zr); // used to keep carry now
3648 BIND(LOOP);
3649 ldrw(rscratch1, Address(pre(in, -4)));
3650 madd(rscratch1, rscratch1, k, out);
3651 ldrw(rscratch2, Address(pre(offset, -4)));
3652 add(rscratch1, rscratch1, rscratch2);
3653 strw(rscratch1, Address(offset));
3654 lsr(out, rscratch1, 32);
3655 subs(len, len, 1);
3656 br(Assembler::NE, LOOP);
3657 BIND(END);
3658 }
3659
3660 /**
3661 * Emits code to update CRC-32 with a byte value according to constants in table
3662 *
3663 * @param [in,out]crc Register containing the crc.
3664 * @param [in]val Register containing the byte to fold into the CRC.
3665 * @param [in]table Register containing the table of crc constants.
3666 *
3667 * uint32_t crc;
3668 * val = crc_table[(val ^ crc) & 0xFF];
3669 * crc = val ^ (crc >> 8);
3670 *
3671 */
3672 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) {
3673 eor(val, val, crc);
3674 andr(val, val, 0xff);
3675 ldrw(val, Address(table, val, Address::lsl(2)));
3676 eor(crc, val, crc, Assembler::LSR, 8);
3677 }
3678
3679 /**
3680 * Emits code to update CRC-32 with a 32-bit value according to tables 0 to 3
3681 *
3682 * @param [in,out]crc Register containing the crc.
3683 * @param [in]v Register containing the 32-bit to fold into the CRC.
3684 * @param [in]table0 Register containing table 0 of crc constants.
3685 * @param [in]table1 Register containing table 1 of crc constants.
3686 * @param [in]table2 Register containing table 2 of crc constants.
3687 * @param [in]table3 Register containing table 3 of crc constants.
3688 *
3689 * uint32_t crc;
3690 * v = crc ^ v
3691 * crc = table3[v&0xff]^table2[(v>>8)&0xff]^table1[(v>>16)&0xff]^table0[v>>24]
3692 *
3693 */
3694 void MacroAssembler::update_word_crc32(Register crc, Register v, Register tmp,
3695 Register table0, Register table1, Register table2, Register table3,
3696 bool upper) {
3697 eor(v, crc, v, upper ? LSR:LSL, upper ? 32:0);
3698 uxtb(tmp, v);
3699 ldrw(crc, Address(table3, tmp, Address::lsl(2)));
3700 ubfx(tmp, v, 8, 8);
3701 ldrw(tmp, Address(table2, tmp, Address::lsl(2)));
3702 eor(crc, crc, tmp);
3703 ubfx(tmp, v, 16, 8);
3704 ldrw(tmp, Address(table1, tmp, Address::lsl(2)));
3705 eor(crc, crc, tmp);
3706 ubfx(tmp, v, 24, 8);
3707 ldrw(tmp, Address(table0, tmp, Address::lsl(2)));
3708 eor(crc, crc, tmp);
3709 }
3710
3711 void MacroAssembler::kernel_crc32_using_crypto_pmull(Register crc, Register buf,
3712 Register len, Register tmp0, Register tmp1, Register tmp2, Register tmp3) {
3713 Label CRC_by4_loop, CRC_by1_loop, CRC_less128, CRC_by128_pre, CRC_by32_loop, CRC_less32, L_exit;
3714 assert_different_registers(crc, buf, len, tmp0, tmp1, tmp2);
3715
3716 subs(tmp0, len, 384);
3717 mvnw(crc, crc);
3718 br(Assembler::GE, CRC_by128_pre);
3719 BIND(CRC_less128);
3720 subs(len, len, 32);
3721 br(Assembler::GE, CRC_by32_loop);
3722 BIND(CRC_less32);
3723 adds(len, len, 32 - 4);
3724 br(Assembler::GE, CRC_by4_loop);
3725 adds(len, len, 4);
3726 br(Assembler::GT, CRC_by1_loop);
3727 b(L_exit);
3728
3729 BIND(CRC_by32_loop);
3730 ldp(tmp0, tmp1, Address(buf));
3731 crc32x(crc, crc, tmp0);
3732 ldp(tmp2, tmp3, Address(buf, 16));
3733 crc32x(crc, crc, tmp1);
3734 add(buf, buf, 32);
3735 crc32x(crc, crc, tmp2);
3736 subs(len, len, 32);
3737 crc32x(crc, crc, tmp3);
3738 br(Assembler::GE, CRC_by32_loop);
3739 cmn(len, (u1)32);
3740 br(Assembler::NE, CRC_less32);
3741 b(L_exit);
3742
3743 BIND(CRC_by4_loop);
3744 ldrw(tmp0, Address(post(buf, 4)));
3745 subs(len, len, 4);
3746 crc32w(crc, crc, tmp0);
3747 br(Assembler::GE, CRC_by4_loop);
3748 adds(len, len, 4);
3749 br(Assembler::LE, L_exit);
3750 BIND(CRC_by1_loop);
3751 ldrb(tmp0, Address(post(buf, 1)));
3752 subs(len, len, 1);
3753 crc32b(crc, crc, tmp0);
3754 br(Assembler::GT, CRC_by1_loop);
3755 b(L_exit);
3756
3757 BIND(CRC_by128_pre);
3758 kernel_crc32_common_fold_using_crypto_pmull(crc, buf, len, tmp0, tmp1, tmp2,
3759 4*256*sizeof(juint) + 8*sizeof(juint));
3760 mov(crc, 0);
3761 crc32x(crc, crc, tmp0);
3762 crc32x(crc, crc, tmp1);
3763
3764 cbnz(len, CRC_less128);
3765
3766 BIND(L_exit);
3767 mvnw(crc, crc);
3768 }
3769
3770 void MacroAssembler::kernel_crc32_using_crc32(Register crc, Register buf,
3771 Register len, Register tmp0, Register tmp1, Register tmp2,
3772 Register tmp3) {
3773 Label CRC_by64_loop, CRC_by4_loop, CRC_by1_loop, CRC_less64, CRC_by64_pre, CRC_by32_loop, CRC_less32, L_exit;
3774 assert_different_registers(crc, buf, len, tmp0, tmp1, tmp2, tmp3);
3775
3776 mvnw(crc, crc);
3777
3778 subs(len, len, 128);
3779 br(Assembler::GE, CRC_by64_pre);
3780 BIND(CRC_less64);
3781 adds(len, len, 128-32);
3782 br(Assembler::GE, CRC_by32_loop);
3783 BIND(CRC_less32);
3784 adds(len, len, 32-4);
3785 br(Assembler::GE, CRC_by4_loop);
3786 adds(len, len, 4);
3787 br(Assembler::GT, CRC_by1_loop);
3788 b(L_exit);
3789
3790 BIND(CRC_by32_loop);
3791 ldp(tmp0, tmp1, Address(post(buf, 16)));
3792 subs(len, len, 32);
3793 crc32x(crc, crc, tmp0);
3794 ldr(tmp2, Address(post(buf, 8)));
3795 crc32x(crc, crc, tmp1);
3796 ldr(tmp3, Address(post(buf, 8)));
3797 crc32x(crc, crc, tmp2);
3798 crc32x(crc, crc, tmp3);
3799 br(Assembler::GE, CRC_by32_loop);
3800 cmn(len, (u1)32);
3801 br(Assembler::NE, CRC_less32);
3802 b(L_exit);
3803
3804 BIND(CRC_by4_loop);
3805 ldrw(tmp0, Address(post(buf, 4)));
3806 subs(len, len, 4);
3807 crc32w(crc, crc, tmp0);
3808 br(Assembler::GE, CRC_by4_loop);
3809 adds(len, len, 4);
3810 br(Assembler::LE, L_exit);
3811 BIND(CRC_by1_loop);
3812 ldrb(tmp0, Address(post(buf, 1)));
3813 subs(len, len, 1);
3814 crc32b(crc, crc, tmp0);
3815 br(Assembler::GT, CRC_by1_loop);
3816 b(L_exit);
3817
3818 BIND(CRC_by64_pre);
3819 sub(buf, buf, 8);
3820 ldp(tmp0, tmp1, Address(buf, 8));
3821 crc32x(crc, crc, tmp0);
3822 ldr(tmp2, Address(buf, 24));
3823 crc32x(crc, crc, tmp1);
3824 ldr(tmp3, Address(buf, 32));
3825 crc32x(crc, crc, tmp2);
3826 ldr(tmp0, Address(buf, 40));
3827 crc32x(crc, crc, tmp3);
3828 ldr(tmp1, Address(buf, 48));
3829 crc32x(crc, crc, tmp0);
3830 ldr(tmp2, Address(buf, 56));
3831 crc32x(crc, crc, tmp1);
3832 ldr(tmp3, Address(pre(buf, 64)));
3833
3834 b(CRC_by64_loop);
3835
3836 align(CodeEntryAlignment);
3837 BIND(CRC_by64_loop);
3838 subs(len, len, 64);
3839 crc32x(crc, crc, tmp2);
3840 ldr(tmp0, Address(buf, 8));
3841 crc32x(crc, crc, tmp3);
3842 ldr(tmp1, Address(buf, 16));
3843 crc32x(crc, crc, tmp0);
3844 ldr(tmp2, Address(buf, 24));
3845 crc32x(crc, crc, tmp1);
3846 ldr(tmp3, Address(buf, 32));
3847 crc32x(crc, crc, tmp2);
3848 ldr(tmp0, Address(buf, 40));
3849 crc32x(crc, crc, tmp3);
3850 ldr(tmp1, Address(buf, 48));
3851 crc32x(crc, crc, tmp0);
3852 ldr(tmp2, Address(buf, 56));
3853 crc32x(crc, crc, tmp1);
3854 ldr(tmp3, Address(pre(buf, 64)));
3855 br(Assembler::GE, CRC_by64_loop);
3856
3857 // post-loop
3858 crc32x(crc, crc, tmp2);
3859 crc32x(crc, crc, tmp3);
3860
3861 sub(len, len, 64);
3862 add(buf, buf, 8);
3863 cmn(len, (u1)128);
3864 br(Assembler::NE, CRC_less64);
3865 BIND(L_exit);
3866 mvnw(crc, crc);
3867 }
3868
3869 /**
3870 * @param crc register containing existing CRC (32-bit)
3871 * @param buf register pointing to input byte buffer (byte*)
3872 * @param len register containing number of bytes
3873 * @param table register that will contain address of CRC table
3874 * @param tmp scratch register
3875 */
3876 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len,
3877 Register table0, Register table1, Register table2, Register table3,
3878 Register tmp, Register tmp2, Register tmp3) {
3879 Label L_by16, L_by16_loop, L_by4, L_by4_loop, L_by1, L_by1_loop, L_exit;
3880
3881 if (UseCryptoPmullForCRC32) {
3882 kernel_crc32_using_crypto_pmull(crc, buf, len, table0, table1, table2, table3);
3883 return;
3884 }
3885
3886 if (UseCRC32) {
3887 kernel_crc32_using_crc32(crc, buf, len, table0, table1, table2, table3);
3888 return;
3889 }
3890
3891 mvnw(crc, crc);
3892
3893 {
3894 uint64_t offset;
3895 adrp(table0, ExternalAddress(StubRoutines::crc_table_addr()), offset);
3896 add(table0, table0, offset);
3897 }
3898 add(table1, table0, 1*256*sizeof(juint));
3899 add(table2, table0, 2*256*sizeof(juint));
3900 add(table3, table0, 3*256*sizeof(juint));
3901
3902 if (UseNeon) {
3903 cmp(len, (u1)64);
3904 br(Assembler::LT, L_by16);
3905 eor(v16, T16B, v16, v16);
3906
3907 Label L_fold;
3908
3909 add(tmp, table0, 4*256*sizeof(juint)); // Point at the Neon constants
3910
3911 ld1(v0, v1, T2D, post(buf, 32));
3912 ld1r(v4, T2D, post(tmp, 8));
3913 ld1r(v5, T2D, post(tmp, 8));
3914 ld1r(v6, T2D, post(tmp, 8));
3915 ld1r(v7, T2D, post(tmp, 8));
3916 mov(v16, S, 0, crc);
3917
3918 eor(v0, T16B, v0, v16);
3919 sub(len, len, 64);
3920
3921 BIND(L_fold);
3922 pmull(v22, T8H, v0, v5, T8B);
3923 pmull(v20, T8H, v0, v7, T8B);
3924 pmull(v23, T8H, v0, v4, T8B);
3925 pmull(v21, T8H, v0, v6, T8B);
3926
3927 pmull2(v18, T8H, v0, v5, T16B);
3928 pmull2(v16, T8H, v0, v7, T16B);
3929 pmull2(v19, T8H, v0, v4, T16B);
3930 pmull2(v17, T8H, v0, v6, T16B);
3931
3932 uzp1(v24, T8H, v20, v22);
3933 uzp2(v25, T8H, v20, v22);
3934 eor(v20, T16B, v24, v25);
3935
3936 uzp1(v26, T8H, v16, v18);
3937 uzp2(v27, T8H, v16, v18);
3938 eor(v16, T16B, v26, v27);
3939
3940 ushll2(v22, T4S, v20, T8H, 8);
3941 ushll(v20, T4S, v20, T4H, 8);
3942
3943 ushll2(v18, T4S, v16, T8H, 8);
3944 ushll(v16, T4S, v16, T4H, 8);
3945
3946 eor(v22, T16B, v23, v22);
3947 eor(v18, T16B, v19, v18);
3948 eor(v20, T16B, v21, v20);
3949 eor(v16, T16B, v17, v16);
3950
3951 uzp1(v17, T2D, v16, v20);
3952 uzp2(v21, T2D, v16, v20);
3953 eor(v17, T16B, v17, v21);
3954
3955 ushll2(v20, T2D, v17, T4S, 16);
3956 ushll(v16, T2D, v17, T2S, 16);
3957
3958 eor(v20, T16B, v20, v22);
3959 eor(v16, T16B, v16, v18);
3960
3961 uzp1(v17, T2D, v20, v16);
3962 uzp2(v21, T2D, v20, v16);
3963 eor(v28, T16B, v17, v21);
3964
3965 pmull(v22, T8H, v1, v5, T8B);
3966 pmull(v20, T8H, v1, v7, T8B);
3967 pmull(v23, T8H, v1, v4, T8B);
3968 pmull(v21, T8H, v1, v6, T8B);
3969
3970 pmull2(v18, T8H, v1, v5, T16B);
3971 pmull2(v16, T8H, v1, v7, T16B);
3972 pmull2(v19, T8H, v1, v4, T16B);
3973 pmull2(v17, T8H, v1, v6, T16B);
3974
3975 ld1(v0, v1, T2D, post(buf, 32));
3976
3977 uzp1(v24, T8H, v20, v22);
3978 uzp2(v25, T8H, v20, v22);
3979 eor(v20, T16B, v24, v25);
3980
3981 uzp1(v26, T8H, v16, v18);
3982 uzp2(v27, T8H, v16, v18);
3983 eor(v16, T16B, v26, v27);
3984
3985 ushll2(v22, T4S, v20, T8H, 8);
3986 ushll(v20, T4S, v20, T4H, 8);
3987
3988 ushll2(v18, T4S, v16, T8H, 8);
3989 ushll(v16, T4S, v16, T4H, 8);
3990
3991 eor(v22, T16B, v23, v22);
3992 eor(v18, T16B, v19, v18);
3993 eor(v20, T16B, v21, v20);
3994 eor(v16, T16B, v17, v16);
3995
3996 uzp1(v17, T2D, v16, v20);
3997 uzp2(v21, T2D, v16, v20);
3998 eor(v16, T16B, v17, v21);
3999
4000 ushll2(v20, T2D, v16, T4S, 16);
4001 ushll(v16, T2D, v16, T2S, 16);
4002
4003 eor(v20, T16B, v22, v20);
4004 eor(v16, T16B, v16, v18);
4005
4006 uzp1(v17, T2D, v20, v16);
4007 uzp2(v21, T2D, v20, v16);
4008 eor(v20, T16B, v17, v21);
4009
4010 shl(v16, T2D, v28, 1);
4011 shl(v17, T2D, v20, 1);
4012
4013 eor(v0, T16B, v0, v16);
4014 eor(v1, T16B, v1, v17);
4015
4016 subs(len, len, 32);
4017 br(Assembler::GE, L_fold);
4018
4019 mov(crc, 0);
4020 mov(tmp, v0, D, 0);
4021 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false);
4022 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true);
4023 mov(tmp, v0, D, 1);
4024 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false);
4025 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true);
4026 mov(tmp, v1, D, 0);
4027 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false);
4028 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true);
4029 mov(tmp, v1, D, 1);
4030 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false);
4031 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true);
4032
4033 add(len, len, 32);
4034 }
4035
4036 BIND(L_by16);
4037 subs(len, len, 16);
4038 br(Assembler::GE, L_by16_loop);
4039 adds(len, len, 16-4);
4040 br(Assembler::GE, L_by4_loop);
4041 adds(len, len, 4);
4042 br(Assembler::GT, L_by1_loop);
4043 b(L_exit);
4044
4045 BIND(L_by4_loop);
4046 ldrw(tmp, Address(post(buf, 4)));
4047 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3);
4048 subs(len, len, 4);
4049 br(Assembler::GE, L_by4_loop);
4050 adds(len, len, 4);
4051 br(Assembler::LE, L_exit);
4052 BIND(L_by1_loop);
4053 subs(len, len, 1);
4054 ldrb(tmp, Address(post(buf, 1)));
4055 update_byte_crc32(crc, tmp, table0);
4056 br(Assembler::GT, L_by1_loop);
4057 b(L_exit);
4058
4059 align(CodeEntryAlignment);
4060 BIND(L_by16_loop);
4061 subs(len, len, 16);
4062 ldp(tmp, tmp3, Address(post(buf, 16)));
4063 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false);
4064 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true);
4065 update_word_crc32(crc, tmp3, tmp2, table0, table1, table2, table3, false);
4066 update_word_crc32(crc, tmp3, tmp2, table0, table1, table2, table3, true);
4067 br(Assembler::GE, L_by16_loop);
4068 adds(len, len, 16-4);
4069 br(Assembler::GE, L_by4_loop);
4070 adds(len, len, 4);
4071 br(Assembler::GT, L_by1_loop);
4072 BIND(L_exit);
4073 mvnw(crc, crc);
4074 }
4075
4076 void MacroAssembler::kernel_crc32c_using_crypto_pmull(Register crc, Register buf,
4077 Register len, Register tmp0, Register tmp1, Register tmp2, Register tmp3) {
4078 Label CRC_by4_loop, CRC_by1_loop, CRC_less128, CRC_by128_pre, CRC_by32_loop, CRC_less32, L_exit;
4079 assert_different_registers(crc, buf, len, tmp0, tmp1, tmp2);
4080
4081 subs(tmp0, len, 384);
4082 br(Assembler::GE, CRC_by128_pre);
4083 BIND(CRC_less128);
4084 subs(len, len, 32);
4085 br(Assembler::GE, CRC_by32_loop);
4086 BIND(CRC_less32);
4087 adds(len, len, 32 - 4);
4088 br(Assembler::GE, CRC_by4_loop);
4089 adds(len, len, 4);
4090 br(Assembler::GT, CRC_by1_loop);
4091 b(L_exit);
4092
4093 BIND(CRC_by32_loop);
4094 ldp(tmp0, tmp1, Address(buf));
4095 crc32cx(crc, crc, tmp0);
4096 ldr(tmp2, Address(buf, 16));
4097 crc32cx(crc, crc, tmp1);
4098 ldr(tmp3, Address(buf, 24));
4099 crc32cx(crc, crc, tmp2);
4100 add(buf, buf, 32);
4101 subs(len, len, 32);
4102 crc32cx(crc, crc, tmp3);
4103 br(Assembler::GE, CRC_by32_loop);
4104 cmn(len, (u1)32);
4105 br(Assembler::NE, CRC_less32);
4106 b(L_exit);
4107
4108 BIND(CRC_by4_loop);
4109 ldrw(tmp0, Address(post(buf, 4)));
4110 subs(len, len, 4);
4111 crc32cw(crc, crc, tmp0);
4112 br(Assembler::GE, CRC_by4_loop);
4113 adds(len, len, 4);
4114 br(Assembler::LE, L_exit);
4115 BIND(CRC_by1_loop);
4116 ldrb(tmp0, Address(post(buf, 1)));
4117 subs(len, len, 1);
4118 crc32cb(crc, crc, tmp0);
4119 br(Assembler::GT, CRC_by1_loop);
4120 b(L_exit);
4121
4122 BIND(CRC_by128_pre);
4123 kernel_crc32_common_fold_using_crypto_pmull(crc, buf, len, tmp0, tmp1, tmp2,
4124 4*256*sizeof(juint) + 8*sizeof(juint) + 0x50);
4125 mov(crc, 0);
4126 crc32cx(crc, crc, tmp0);
4127 crc32cx(crc, crc, tmp1);
4128
4129 cbnz(len, CRC_less128);
4130
4131 BIND(L_exit);
4132 }
4133
4134 void MacroAssembler::kernel_crc32c_using_crc32c(Register crc, Register buf,
4135 Register len, Register tmp0, Register tmp1, Register tmp2,
4136 Register tmp3) {
4137 Label CRC_by64_loop, CRC_by4_loop, CRC_by1_loop, CRC_less64, CRC_by64_pre, CRC_by32_loop, CRC_less32, L_exit;
4138 assert_different_registers(crc, buf, len, tmp0, tmp1, tmp2, tmp3);
4139
4140 subs(len, len, 128);
4141 br(Assembler::GE, CRC_by64_pre);
4142 BIND(CRC_less64);
4143 adds(len, len, 128-32);
4144 br(Assembler::GE, CRC_by32_loop);
4145 BIND(CRC_less32);
4146 adds(len, len, 32-4);
4147 br(Assembler::GE, CRC_by4_loop);
4148 adds(len, len, 4);
4149 br(Assembler::GT, CRC_by1_loop);
4150 b(L_exit);
4151
4152 BIND(CRC_by32_loop);
4153 ldp(tmp0, tmp1, Address(post(buf, 16)));
4154 subs(len, len, 32);
4155 crc32cx(crc, crc, tmp0);
4156 ldr(tmp2, Address(post(buf, 8)));
4157 crc32cx(crc, crc, tmp1);
4158 ldr(tmp3, Address(post(buf, 8)));
4159 crc32cx(crc, crc, tmp2);
4160 crc32cx(crc, crc, tmp3);
4161 br(Assembler::GE, CRC_by32_loop);
4162 cmn(len, (u1)32);
4163 br(Assembler::NE, CRC_less32);
4164 b(L_exit);
4165
4166 BIND(CRC_by4_loop);
4167 ldrw(tmp0, Address(post(buf, 4)));
4168 subs(len, len, 4);
4169 crc32cw(crc, crc, tmp0);
4170 br(Assembler::GE, CRC_by4_loop);
4171 adds(len, len, 4);
4172 br(Assembler::LE, L_exit);
4173 BIND(CRC_by1_loop);
4174 ldrb(tmp0, Address(post(buf, 1)));
4175 subs(len, len, 1);
4176 crc32cb(crc, crc, tmp0);
4177 br(Assembler::GT, CRC_by1_loop);
4178 b(L_exit);
4179
4180 BIND(CRC_by64_pre);
4181 sub(buf, buf, 8);
4182 ldp(tmp0, tmp1, Address(buf, 8));
4183 crc32cx(crc, crc, tmp0);
4184 ldr(tmp2, Address(buf, 24));
4185 crc32cx(crc, crc, tmp1);
4186 ldr(tmp3, Address(buf, 32));
4187 crc32cx(crc, crc, tmp2);
4188 ldr(tmp0, Address(buf, 40));
4189 crc32cx(crc, crc, tmp3);
4190 ldr(tmp1, Address(buf, 48));
4191 crc32cx(crc, crc, tmp0);
4192 ldr(tmp2, Address(buf, 56));
4193 crc32cx(crc, crc, tmp1);
4194 ldr(tmp3, Address(pre(buf, 64)));
4195
4196 b(CRC_by64_loop);
4197
4198 align(CodeEntryAlignment);
4199 BIND(CRC_by64_loop);
4200 subs(len, len, 64);
4201 crc32cx(crc, crc, tmp2);
4202 ldr(tmp0, Address(buf, 8));
4203 crc32cx(crc, crc, tmp3);
4204 ldr(tmp1, Address(buf, 16));
4205 crc32cx(crc, crc, tmp0);
4206 ldr(tmp2, Address(buf, 24));
4207 crc32cx(crc, crc, tmp1);
4208 ldr(tmp3, Address(buf, 32));
4209 crc32cx(crc, crc, tmp2);
4210 ldr(tmp0, Address(buf, 40));
4211 crc32cx(crc, crc, tmp3);
4212 ldr(tmp1, Address(buf, 48));
4213 crc32cx(crc, crc, tmp0);
4214 ldr(tmp2, Address(buf, 56));
4215 crc32cx(crc, crc, tmp1);
4216 ldr(tmp3, Address(pre(buf, 64)));
4217 br(Assembler::GE, CRC_by64_loop);
4218
4219 // post-loop
4220 crc32cx(crc, crc, tmp2);
4221 crc32cx(crc, crc, tmp3);
4222
4223 sub(len, len, 64);
4224 add(buf, buf, 8);
4225 cmn(len, (u1)128);
4226 br(Assembler::NE, CRC_less64);
4227 BIND(L_exit);
4228 }
4229
4230 /**
4231 * @param crc register containing existing CRC (32-bit)
4232 * @param buf register pointing to input byte buffer (byte*)
4233 * @param len register containing number of bytes
4234 * @param table register that will contain address of CRC table
4235 * @param tmp scratch register
4236 */
4237 void MacroAssembler::kernel_crc32c(Register crc, Register buf, Register len,
4238 Register table0, Register table1, Register table2, Register table3,
4239 Register tmp, Register tmp2, Register tmp3) {
4240 if (UseCryptoPmullForCRC32) {
4241 kernel_crc32c_using_crypto_pmull(crc, buf, len, table0, table1, table2, table3);
4242 } else {
4243 kernel_crc32c_using_crc32c(crc, buf, len, table0, table1, table2, table3);
4244 }
4245 }
4246
4247 void MacroAssembler::kernel_crc32_common_fold_using_crypto_pmull(Register crc, Register buf,
4248 Register len, Register tmp0, Register tmp1, Register tmp2, size_t table_offset) {
4249 Label CRC_by128_loop;
4250 assert_different_registers(crc, buf, len, tmp0, tmp1, tmp2);
4251
4252 sub(len, len, 256);
4253 Register table = tmp0;
4254 {
4255 uint64_t offset;
4256 adrp(table, ExternalAddress(StubRoutines::crc_table_addr()), offset);
4257 add(table, table, offset);
4258 }
4259 add(table, table, table_offset);
4260
4261 sub(buf, buf, 0x10);
4262 ldrq(v1, Address(buf, 0x10));
4263 ldrq(v2, Address(buf, 0x20));
4264 ldrq(v3, Address(buf, 0x30));
4265 ldrq(v4, Address(buf, 0x40));
4266 ldrq(v5, Address(buf, 0x50));
4267 ldrq(v6, Address(buf, 0x60));
4268 ldrq(v7, Address(buf, 0x70));
4269 ldrq(v8, Address(pre(buf, 0x80)));
4270
4271 movi(v25, T4S, 0);
4272 mov(v25, S, 0, crc);
4273 eor(v1, T16B, v1, v25);
4274
4275 ldrq(v0, Address(table));
4276 b(CRC_by128_loop);
4277
4278 align(OptoLoopAlignment);
4279 BIND(CRC_by128_loop);
4280 pmull (v9, T1Q, v1, v0, T1D);
4281 pmull2(v10, T1Q, v1, v0, T2D);
4282 ldrq(v1, Address(buf, 0x10));
4283 eor3(v1, T16B, v9, v10, v1);
4284
4285 pmull (v11, T1Q, v2, v0, T1D);
4286 pmull2(v12, T1Q, v2, v0, T2D);
4287 ldrq(v2, Address(buf, 0x20));
4288 eor3(v2, T16B, v11, v12, v2);
4289
4290 pmull (v13, T1Q, v3, v0, T1D);
4291 pmull2(v14, T1Q, v3, v0, T2D);
4292 ldrq(v3, Address(buf, 0x30));
4293 eor3(v3, T16B, v13, v14, v3);
4294
4295 pmull (v15, T1Q, v4, v0, T1D);
4296 pmull2(v16, T1Q, v4, v0, T2D);
4297 ldrq(v4, Address(buf, 0x40));
4298 eor3(v4, T16B, v15, v16, v4);
4299
4300 pmull (v17, T1Q, v5, v0, T1D);
4301 pmull2(v18, T1Q, v5, v0, T2D);
4302 ldrq(v5, Address(buf, 0x50));
4303 eor3(v5, T16B, v17, v18, v5);
4304
4305 pmull (v19, T1Q, v6, v0, T1D);
4306 pmull2(v20, T1Q, v6, v0, T2D);
4307 ldrq(v6, Address(buf, 0x60));
4308 eor3(v6, T16B, v19, v20, v6);
4309
4310 pmull (v21, T1Q, v7, v0, T1D);
4311 pmull2(v22, T1Q, v7, v0, T2D);
4312 ldrq(v7, Address(buf, 0x70));
4313 eor3(v7, T16B, v21, v22, v7);
4314
4315 pmull (v23, T1Q, v8, v0, T1D);
4316 pmull2(v24, T1Q, v8, v0, T2D);
4317 ldrq(v8, Address(pre(buf, 0x80)));
4318 eor3(v8, T16B, v23, v24, v8);
4319
4320 subs(len, len, 0x80);
4321 br(Assembler::GE, CRC_by128_loop);
4322
4323 // fold into 512 bits
4324 ldrq(v0, Address(table, 0x10));
4325
4326 pmull (v10, T1Q, v1, v0, T1D);
4327 pmull2(v11, T1Q, v1, v0, T2D);
4328 eor3(v1, T16B, v10, v11, v5);
4329
4330 pmull (v12, T1Q, v2, v0, T1D);
4331 pmull2(v13, T1Q, v2, v0, T2D);
4332 eor3(v2, T16B, v12, v13, v6);
4333
4334 pmull (v14, T1Q, v3, v0, T1D);
4335 pmull2(v15, T1Q, v3, v0, T2D);
4336 eor3(v3, T16B, v14, v15, v7);
4337
4338 pmull (v16, T1Q, v4, v0, T1D);
4339 pmull2(v17, T1Q, v4, v0, T2D);
4340 eor3(v4, T16B, v16, v17, v8);
4341
4342 // fold into 128 bits
4343 ldrq(v5, Address(table, 0x20));
4344 pmull (v10, T1Q, v1, v5, T1D);
4345 pmull2(v11, T1Q, v1, v5, T2D);
4346 eor3(v4, T16B, v4, v10, v11);
4347
4348 ldrq(v6, Address(table, 0x30));
4349 pmull (v12, T1Q, v2, v6, T1D);
4350 pmull2(v13, T1Q, v2, v6, T2D);
4351 eor3(v4, T16B, v4, v12, v13);
4352
4353 ldrq(v7, Address(table, 0x40));
4354 pmull (v14, T1Q, v3, v7, T1D);
4355 pmull2(v15, T1Q, v3, v7, T2D);
4356 eor3(v1, T16B, v4, v14, v15);
4357
4358 add(len, len, 0x80);
4359 add(buf, buf, 0x10);
4360
4361 mov(tmp0, v1, D, 0);
4362 mov(tmp1, v1, D, 1);
4363 }
4364
4365 SkipIfEqual::SkipIfEqual(
4366 MacroAssembler* masm, const bool* flag_addr, bool value) {
4367 _masm = masm;
4368 uint64_t offset;
4369 _masm->adrp(rscratch1, ExternalAddress((address)flag_addr), offset);
4370 _masm->ldrb(rscratch1, Address(rscratch1, offset));
4371 if (value) {
4372 _masm->cbnzw(rscratch1, _label);
4373 } else {
4374 _masm->cbzw(rscratch1, _label);
4375 }
4376 }
4377
4378 SkipIfEqual::~SkipIfEqual() {
4379 _masm->bind(_label);
4380 }
4381
4382 void MacroAssembler::addptr(const Address &dst, int32_t src) {
4383 Address adr;
4384 switch(dst.getMode()) {
4385 case Address::base_plus_offset:
4386 // This is the expected mode, although we allow all the other
4387 // forms below.
4388 adr = form_address(rscratch2, dst.base(), dst.offset(), LogBytesPerWord);
4389 break;
4390 default:
4391 lea(rscratch2, dst);
4392 adr = Address(rscratch2);
4393 break;
4394 }
4395 ldr(rscratch1, adr);
4396 add(rscratch1, rscratch1, src);
4397 str(rscratch1, adr);
4398 }
4399
4400 void MacroAssembler::cmpptr(Register src1, Address src2) {
4401 uint64_t offset;
4402 adrp(rscratch1, src2, offset);
4403 ldr(rscratch1, Address(rscratch1, offset));
4404 cmp(src1, rscratch1);
4405 }
4406
4407 void MacroAssembler::cmpoop(Register obj1, Register obj2) {
4408 cmp(obj1, obj2);
4409 }
4410
4411 void MacroAssembler::load_method_holder_cld(Register rresult, Register rmethod) {
4412 load_method_holder(rresult, rmethod);
4413 ldr(rresult, Address(rresult, InstanceKlass::class_loader_data_offset()));
4414 }
4415
4416 void MacroAssembler::load_method_holder(Register holder, Register method) {
4417 ldr(holder, Address(method, Method::const_offset())); // ConstMethod*
4418 ldr(holder, Address(holder, ConstMethod::constants_offset())); // ConstantPool*
4419 ldr(holder, Address(holder, ConstantPool::pool_holder_offset())); // InstanceKlass*
4420 }
4421
4422 void MacroAssembler::load_klass(Register dst, Register src) {
4423 if (UseCompressedClassPointers) {
4424 ldrw(dst, Address(src, oopDesc::klass_offset_in_bytes()));
4425 decode_klass_not_null(dst);
4426 } else {
4427 ldr(dst, Address(src, oopDesc::klass_offset_in_bytes()));
4428 }
4429 }
4430
4431 // ((OopHandle)result).resolve();
4432 void MacroAssembler::resolve_oop_handle(Register result, Register tmp1, Register tmp2) {
4433 // OopHandle::resolve is an indirection.
4434 access_load_at(T_OBJECT, IN_NATIVE, result, Address(result, 0), tmp1, tmp2);
4435 }
4436
4437 // ((WeakHandle)result).resolve();
4438 void MacroAssembler::resolve_weak_handle(Register result, Register tmp1, Register tmp2) {
4439 assert_different_registers(result, tmp1, tmp2);
4440 Label resolved;
4441
4442 // A null weak handle resolves to null.
4443 cbz(result, resolved);
4444
4445 // Only 64 bit platforms support GCs that require a tmp register
4446 // WeakHandle::resolve is an indirection like jweak.
4447 access_load_at(T_OBJECT, IN_NATIVE | ON_PHANTOM_OOP_REF,
4448 result, Address(result), tmp1, tmp2);
4449 bind(resolved);
4450 }
4451
4452 void MacroAssembler::load_mirror(Register dst, Register method, Register tmp1, Register tmp2) {
4453 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
4454 ldr(dst, Address(rmethod, Method::const_offset()));
4455 ldr(dst, Address(dst, ConstMethod::constants_offset()));
4456 ldr(dst, Address(dst, ConstantPool::pool_holder_offset()));
4457 ldr(dst, Address(dst, mirror_offset));
4458 resolve_oop_handle(dst, tmp1, tmp2);
4459 }
4460
4461 void MacroAssembler::cmp_klass(Register oop, Register trial_klass, Register tmp) {
4462 if (UseCompressedClassPointers) {
4463 ldrw(tmp, Address(oop, oopDesc::klass_offset_in_bytes()));
4464 if (CompressedKlassPointers::base() == nullptr) {
4465 cmp(trial_klass, tmp, LSL, CompressedKlassPointers::shift());
4466 return;
4467 } else if (((uint64_t)CompressedKlassPointers::base() & 0xffffffff) == 0
4468 && CompressedKlassPointers::shift() == 0) {
4469 // Only the bottom 32 bits matter
4470 cmpw(trial_klass, tmp);
4471 return;
4472 }
4473 decode_klass_not_null(tmp);
4474 } else {
4475 ldr(tmp, Address(oop, oopDesc::klass_offset_in_bytes()));
4476 }
4477 cmp(trial_klass, tmp);
4478 }
4479
4480 void MacroAssembler::store_klass(Register dst, Register src) {
4481 // FIXME: Should this be a store release? concurrent gcs assumes
4482 // klass length is valid if klass field is not null.
4483 if (UseCompressedClassPointers) {
4484 encode_klass_not_null(src);
4485 strw(src, Address(dst, oopDesc::klass_offset_in_bytes()));
4486 } else {
4487 str(src, Address(dst, oopDesc::klass_offset_in_bytes()));
4488 }
4489 }
4490
4491 void MacroAssembler::store_klass_gap(Register dst, Register src) {
4492 if (UseCompressedClassPointers) {
4493 // Store to klass gap in destination
4494 strw(src, Address(dst, oopDesc::klass_gap_offset_in_bytes()));
4495 }
4496 }
4497
4498 // Algorithm must match CompressedOops::encode.
4499 void MacroAssembler::encode_heap_oop(Register d, Register s) {
4500 #ifdef ASSERT
4501 verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?");
4502 #endif
4503 verify_oop_msg(s, "broken oop in encode_heap_oop");
4504 if (CompressedOops::base() == nullptr) {
4505 if (CompressedOops::shift() != 0) {
4506 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
4507 lsr(d, s, LogMinObjAlignmentInBytes);
4508 } else {
4509 mov(d, s);
4510 }
4511 } else {
4512 subs(d, s, rheapbase);
4513 csel(d, d, zr, Assembler::HS);
4514 lsr(d, d, LogMinObjAlignmentInBytes);
4515
4516 /* Old algorithm: is this any worse?
4517 Label nonnull;
4518 cbnz(r, nonnull);
4519 sub(r, r, rheapbase);
4520 bind(nonnull);
4521 lsr(r, r, LogMinObjAlignmentInBytes);
4522 */
4523 }
4524 }
4525
4526 void MacroAssembler::encode_heap_oop_not_null(Register r) {
4527 #ifdef ASSERT
4528 verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?");
4529 if (CheckCompressedOops) {
4530 Label ok;
4531 cbnz(r, ok);
4532 stop("null oop passed to encode_heap_oop_not_null");
4533 bind(ok);
4534 }
4535 #endif
4536 verify_oop_msg(r, "broken oop in encode_heap_oop_not_null");
4537 if (CompressedOops::base() != nullptr) {
4538 sub(r, r, rheapbase);
4539 }
4540 if (CompressedOops::shift() != 0) {
4541 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
4542 lsr(r, r, LogMinObjAlignmentInBytes);
4543 }
4544 }
4545
4546 void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) {
4547 #ifdef ASSERT
4548 verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?");
4549 if (CheckCompressedOops) {
4550 Label ok;
4551 cbnz(src, ok);
4552 stop("null oop passed to encode_heap_oop_not_null2");
4553 bind(ok);
4554 }
4555 #endif
4556 verify_oop_msg(src, "broken oop in encode_heap_oop_not_null2");
4557
4558 Register data = src;
4559 if (CompressedOops::base() != nullptr) {
4560 sub(dst, src, rheapbase);
4561 data = dst;
4562 }
4563 if (CompressedOops::shift() != 0) {
4564 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
4565 lsr(dst, data, LogMinObjAlignmentInBytes);
4566 data = dst;
4567 }
4568 if (data == src)
4569 mov(dst, src);
4570 }
4571
4572 void MacroAssembler::decode_heap_oop(Register d, Register s) {
4573 #ifdef ASSERT
4574 verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?");
4575 #endif
4576 if (CompressedOops::base() == nullptr) {
4577 if (CompressedOops::shift() != 0 || d != s) {
4578 lsl(d, s, CompressedOops::shift());
4579 }
4580 } else {
4581 Label done;
4582 if (d != s)
4583 mov(d, s);
4584 cbz(s, done);
4585 add(d, rheapbase, s, Assembler::LSL, LogMinObjAlignmentInBytes);
4586 bind(done);
4587 }
4588 verify_oop_msg(d, "broken oop in decode_heap_oop");
4589 }
4590
4591 void MacroAssembler::decode_heap_oop_not_null(Register r) {
4592 assert (UseCompressedOops, "should only be used for compressed headers");
4593 assert (Universe::heap() != nullptr, "java heap should be initialized");
4594 // Cannot assert, unverified entry point counts instructions (see .ad file)
4595 // vtableStubs also counts instructions in pd_code_size_limit.
4596 // Also do not verify_oop as this is called by verify_oop.
4597 if (CompressedOops::shift() != 0) {
4598 assert(LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
4599 if (CompressedOops::base() != nullptr) {
4600 add(r, rheapbase, r, Assembler::LSL, LogMinObjAlignmentInBytes);
4601 } else {
4602 add(r, zr, r, Assembler::LSL, LogMinObjAlignmentInBytes);
4603 }
4604 } else {
4605 assert (CompressedOops::base() == nullptr, "sanity");
4606 }
4607 }
4608
4609 void MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) {
4610 assert (UseCompressedOops, "should only be used for compressed headers");
4611 assert (Universe::heap() != nullptr, "java heap should be initialized");
4612 // Cannot assert, unverified entry point counts instructions (see .ad file)
4613 // vtableStubs also counts instructions in pd_code_size_limit.
4614 // Also do not verify_oop as this is called by verify_oop.
4615 if (CompressedOops::shift() != 0) {
4616 assert(LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
4617 if (CompressedOops::base() != nullptr) {
4618 add(dst, rheapbase, src, Assembler::LSL, LogMinObjAlignmentInBytes);
4619 } else {
4620 add(dst, zr, src, Assembler::LSL, LogMinObjAlignmentInBytes);
4621 }
4622 } else {
4623 assert (CompressedOops::base() == nullptr, "sanity");
4624 if (dst != src) {
4625 mov(dst, src);
4626 }
4627 }
4628 }
4629
4630 MacroAssembler::KlassDecodeMode MacroAssembler::_klass_decode_mode(KlassDecodeNone);
4631
4632 MacroAssembler::KlassDecodeMode MacroAssembler::klass_decode_mode() {
4633 assert(UseCompressedClassPointers, "not using compressed class pointers");
4634 assert(Metaspace::initialized(), "metaspace not initialized yet");
4635
4636 if (_klass_decode_mode != KlassDecodeNone) {
4637 return _klass_decode_mode;
4638 }
4639
4640 assert(LogKlassAlignmentInBytes == CompressedKlassPointers::shift()
4641 || 0 == CompressedKlassPointers::shift(), "decode alg wrong");
4642
4643 if (CompressedKlassPointers::base() == nullptr) {
4644 return (_klass_decode_mode = KlassDecodeZero);
4645 }
4646
4647 if (operand_valid_for_logical_immediate(
4648 /*is32*/false, (uint64_t)CompressedKlassPointers::base())) {
4649 const uint64_t range_mask =
4650 (1ULL << log2i(CompressedKlassPointers::range())) - 1;
4651 if (((uint64_t)CompressedKlassPointers::base() & range_mask) == 0) {
4652 return (_klass_decode_mode = KlassDecodeXor);
4653 }
4654 }
4655
4656 const uint64_t shifted_base =
4657 (uint64_t)CompressedKlassPointers::base() >> CompressedKlassPointers::shift();
4658 guarantee((shifted_base & 0xffff0000ffffffff) == 0,
4659 "compressed class base bad alignment");
4660
4661 return (_klass_decode_mode = KlassDecodeMovk);
4662 }
4663
4664 void MacroAssembler::encode_klass_not_null(Register dst, Register src) {
4665 switch (klass_decode_mode()) {
4666 case KlassDecodeZero:
4667 if (CompressedKlassPointers::shift() != 0) {
4668 lsr(dst, src, LogKlassAlignmentInBytes);
4669 } else {
4670 if (dst != src) mov(dst, src);
4671 }
4672 break;
4673
4674 case KlassDecodeXor:
4675 if (CompressedKlassPointers::shift() != 0) {
4676 eor(dst, src, (uint64_t)CompressedKlassPointers::base());
4677 lsr(dst, dst, LogKlassAlignmentInBytes);
4678 } else {
4679 eor(dst, src, (uint64_t)CompressedKlassPointers::base());
4680 }
4681 break;
4682
4683 case KlassDecodeMovk:
4684 if (CompressedKlassPointers::shift() != 0) {
4685 ubfx(dst, src, LogKlassAlignmentInBytes, 32);
4686 } else {
4687 movw(dst, src);
4688 }
4689 break;
4690
4691 case KlassDecodeNone:
4692 ShouldNotReachHere();
4693 break;
4694 }
4695 }
4696
4697 void MacroAssembler::encode_klass_not_null(Register r) {
4698 encode_klass_not_null(r, r);
4699 }
4700
4701 void MacroAssembler::decode_klass_not_null(Register dst, Register src) {
4702 assert (UseCompressedClassPointers, "should only be used for compressed headers");
4703
4704 switch (klass_decode_mode()) {
4705 case KlassDecodeZero:
4706 if (CompressedKlassPointers::shift() != 0) {
4707 lsl(dst, src, LogKlassAlignmentInBytes);
4708 } else {
4709 if (dst != src) mov(dst, src);
4710 }
4711 break;
4712
4713 case KlassDecodeXor:
4714 if (CompressedKlassPointers::shift() != 0) {
4715 lsl(dst, src, LogKlassAlignmentInBytes);
4716 eor(dst, dst, (uint64_t)CompressedKlassPointers::base());
4717 } else {
4718 eor(dst, src, (uint64_t)CompressedKlassPointers::base());
4719 }
4720 break;
4721
4722 case KlassDecodeMovk: {
4723 const uint64_t shifted_base =
4724 (uint64_t)CompressedKlassPointers::base() >> CompressedKlassPointers::shift();
4725
4726 if (dst != src) movw(dst, src);
4727 movk(dst, shifted_base >> 32, 32);
4728
4729 if (CompressedKlassPointers::shift() != 0) {
4730 lsl(dst, dst, LogKlassAlignmentInBytes);
4731 }
4732
4733 break;
4734 }
4735
4736 case KlassDecodeNone:
4737 ShouldNotReachHere();
4738 break;
4739 }
4740 }
4741
4742 void MacroAssembler::decode_klass_not_null(Register r) {
4743 decode_klass_not_null(r, r);
4744 }
4745
4746 void MacroAssembler::set_narrow_oop(Register dst, jobject obj) {
4747 #ifdef ASSERT
4748 {
4749 ThreadInVMfromUnknown tiv;
4750 assert (UseCompressedOops, "should only be used for compressed oops");
4751 assert (Universe::heap() != nullptr, "java heap should be initialized");
4752 assert (oop_recorder() != nullptr, "this assembler needs an OopRecorder");
4753 assert(Universe::heap()->is_in(JNIHandles::resolve(obj)), "should be real oop");
4754 }
4755 #endif
4756 int oop_index = oop_recorder()->find_index(obj);
4757 InstructionMark im(this);
4758 RelocationHolder rspec = oop_Relocation::spec(oop_index);
4759 code_section()->relocate(inst_mark(), rspec);
4760 movz(dst, 0xDEAD, 16);
4761 movk(dst, 0xBEEF);
4762 }
4763
4764 void MacroAssembler::set_narrow_klass(Register dst, Klass* k) {
4765 assert (UseCompressedClassPointers, "should only be used for compressed headers");
4766 assert (oop_recorder() != nullptr, "this assembler needs an OopRecorder");
4767 int index = oop_recorder()->find_index(k);
4768 assert(! Universe::heap()->is_in(k), "should not be an oop");
4769
4770 InstructionMark im(this);
4771 RelocationHolder rspec = metadata_Relocation::spec(index);
4772 code_section()->relocate(inst_mark(), rspec);
4773 narrowKlass nk = CompressedKlassPointers::encode(k);
4774 movz(dst, (nk >> 16), 16);
4775 movk(dst, nk & 0xffff);
4776 }
4777
4778 void MacroAssembler::access_load_at(BasicType type, DecoratorSet decorators,
4779 Register dst, Address src,
4780 Register tmp1, Register tmp2) {
4781 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler();
4782 decorators = AccessInternal::decorator_fixup(decorators, type);
4783 bool as_raw = (decorators & AS_RAW) != 0;
4784 if (as_raw) {
4785 bs->BarrierSetAssembler::load_at(this, decorators, type, dst, src, tmp1, tmp2);
4786 } else {
4787 bs->load_at(this, decorators, type, dst, src, tmp1, tmp2);
4788 }
4789 }
4790
4791 void MacroAssembler::access_store_at(BasicType type, DecoratorSet decorators,
4792 Address dst, Register val,
4793 Register tmp1, Register tmp2, Register tmp3) {
4794 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler();
4795 decorators = AccessInternal::decorator_fixup(decorators, type);
4796 bool as_raw = (decorators & AS_RAW) != 0;
4797 if (as_raw) {
4798 bs->BarrierSetAssembler::store_at(this, decorators, type, dst, val, tmp1, tmp2, tmp3);
4799 } else {
4800 bs->store_at(this, decorators, type, dst, val, tmp1, tmp2, tmp3);
4801 }
4802 }
4803
4804 void MacroAssembler::load_heap_oop(Register dst, Address src, Register tmp1,
4805 Register tmp2, DecoratorSet decorators) {
4806 access_load_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, tmp2);
4807 }
4808
4809 void MacroAssembler::load_heap_oop_not_null(Register dst, Address src, Register tmp1,
4810 Register tmp2, DecoratorSet decorators) {
4811 access_load_at(T_OBJECT, IN_HEAP | IS_NOT_NULL | decorators, dst, src, tmp1, tmp2);
4812 }
4813
4814 void MacroAssembler::store_heap_oop(Address dst, Register val, Register tmp1,
4815 Register tmp2, Register tmp3, DecoratorSet decorators) {
4816 access_store_at(T_OBJECT, IN_HEAP | decorators, dst, val, tmp1, tmp2, tmp3);
4817 }
4818
4819 // Used for storing nulls.
4820 void MacroAssembler::store_heap_oop_null(Address dst) {
4821 access_store_at(T_OBJECT, IN_HEAP, dst, noreg, noreg, noreg, noreg);
4822 }
4823
4824 Address MacroAssembler::allocate_metadata_address(Metadata* obj) {
4825 assert(oop_recorder() != nullptr, "this assembler needs a Recorder");
4826 int index = oop_recorder()->allocate_metadata_index(obj);
4827 RelocationHolder rspec = metadata_Relocation::spec(index);
4828 return Address((address)obj, rspec);
4829 }
4830
4831 // Move an oop into a register.
4832 void MacroAssembler::movoop(Register dst, jobject obj) {
4833 int oop_index;
4834 if (obj == nullptr) {
4835 oop_index = oop_recorder()->allocate_oop_index(obj);
4836 } else {
4837 #ifdef ASSERT
4838 {
4839 ThreadInVMfromUnknown tiv;
4840 assert(Universe::heap()->is_in(JNIHandles::resolve(obj)), "should be real oop");
4841 }
4842 #endif
4843 oop_index = oop_recorder()->find_index(obj);
4844 }
4845 RelocationHolder rspec = oop_Relocation::spec(oop_index);
4846
4847 if (BarrierSet::barrier_set()->barrier_set_assembler()->supports_instruction_patching()) {
4848 mov(dst, Address((address)obj, rspec));
4849 } else {
4850 address dummy = address(uintptr_t(pc()) & -wordSize); // A nearby aligned address
4851 ldr_constant(dst, Address(dummy, rspec));
4852 }
4853
4854 }
4855
4856 // Move a metadata address into a register.
4857 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) {
4858 int oop_index;
4859 if (obj == nullptr) {
4860 oop_index = oop_recorder()->allocate_metadata_index(obj);
4861 } else {
4862 oop_index = oop_recorder()->find_index(obj);
4863 }
4864 RelocationHolder rspec = metadata_Relocation::spec(oop_index);
4865 mov(dst, Address((address)obj, rspec));
4866 }
4867
4868 Address MacroAssembler::constant_oop_address(jobject obj) {
4869 #ifdef ASSERT
4870 {
4871 ThreadInVMfromUnknown tiv;
4872 assert(oop_recorder() != nullptr, "this assembler needs an OopRecorder");
4873 assert(Universe::heap()->is_in(JNIHandles::resolve(obj)), "not an oop");
4874 }
4875 #endif
4876 int oop_index = oop_recorder()->find_index(obj);
4877 return Address((address)obj, oop_Relocation::spec(oop_index));
4878 }
4879
4880 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes.
4881 void MacroAssembler::tlab_allocate(Register obj,
4882 Register var_size_in_bytes,
4883 int con_size_in_bytes,
4884 Register t1,
4885 Register t2,
4886 Label& slow_case) {
4887 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler();
4888 bs->tlab_allocate(this, obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case);
4889 }
4890
4891 void MacroAssembler::verify_tlab() {
4892 #ifdef ASSERT
4893 if (UseTLAB && VerifyOops) {
4894 Label next, ok;
4895
4896 stp(rscratch2, rscratch1, Address(pre(sp, -16)));
4897
4898 ldr(rscratch2, Address(rthread, in_bytes(JavaThread::tlab_top_offset())));
4899 ldr(rscratch1, Address(rthread, in_bytes(JavaThread::tlab_start_offset())));
4900 cmp(rscratch2, rscratch1);
4901 br(Assembler::HS, next);
4902 STOP("assert(top >= start)");
4903 should_not_reach_here();
4904
4905 bind(next);
4906 ldr(rscratch2, Address(rthread, in_bytes(JavaThread::tlab_end_offset())));
4907 ldr(rscratch1, Address(rthread, in_bytes(JavaThread::tlab_top_offset())));
4908 cmp(rscratch2, rscratch1);
4909 br(Assembler::HS, ok);
4910 STOP("assert(top <= end)");
4911 should_not_reach_here();
4912
4913 bind(ok);
4914 ldp(rscratch2, rscratch1, Address(post(sp, 16)));
4915 }
4916 #endif
4917 }
4918
4919 // Writes to stack successive pages until offset reached to check for
4920 // stack overflow + shadow pages. This clobbers tmp.
4921 void MacroAssembler::bang_stack_size(Register size, Register tmp) {
4922 assert_different_registers(tmp, size, rscratch1);
4923 mov(tmp, sp);
4924 // Bang stack for total size given plus shadow page size.
4925 // Bang one page at a time because large size can bang beyond yellow and
4926 // red zones.
4927 Label loop;
4928 mov(rscratch1, (int)os::vm_page_size());
4929 bind(loop);
4930 lea(tmp, Address(tmp, -(int)os::vm_page_size()));
4931 subsw(size, size, rscratch1);
4932 str(size, Address(tmp));
4933 br(Assembler::GT, loop);
4934
4935 // Bang down shadow pages too.
4936 // At this point, (tmp-0) is the last address touched, so don't
4937 // touch it again. (It was touched as (tmp-pagesize) but then tmp
4938 // was post-decremented.) Skip this address by starting at i=1, and
4939 // touch a few more pages below. N.B. It is important to touch all
4940 // the way down to and including i=StackShadowPages.
4941 for (int i = 0; i < (int)(StackOverflow::stack_shadow_zone_size() / (int)os::vm_page_size()) - 1; i++) {
4942 // this could be any sized move but this is can be a debugging crumb
4943 // so the bigger the better.
4944 lea(tmp, Address(tmp, -(int)os::vm_page_size()));
4945 str(size, Address(tmp));
4946 }
4947 }
4948
4949 // Move the address of the polling page into dest.
4950 void MacroAssembler::get_polling_page(Register dest, relocInfo::relocType rtype) {
4951 ldr(dest, Address(rthread, JavaThread::polling_page_offset()));
4952 }
4953
4954 // Read the polling page. The address of the polling page must
4955 // already be in r.
4956 address MacroAssembler::read_polling_page(Register r, relocInfo::relocType rtype) {
4957 address mark;
4958 {
4959 InstructionMark im(this);
4960 code_section()->relocate(inst_mark(), rtype);
4961 ldrw(zr, Address(r, 0));
4962 mark = inst_mark();
4963 }
4964 verify_cross_modify_fence_not_required();
4965 return mark;
4966 }
4967
4968 void MacroAssembler::adrp(Register reg1, const Address &dest, uint64_t &byte_offset) {
4969 relocInfo::relocType rtype = dest.rspec().reloc()->type();
4970 uint64_t low_page = (uint64_t)CodeCache::low_bound() >> 12;
4971 uint64_t high_page = (uint64_t)(CodeCache::high_bound()-1) >> 12;
4972 uint64_t dest_page = (uint64_t)dest.target() >> 12;
4973 int64_t offset_low = dest_page - low_page;
4974 int64_t offset_high = dest_page - high_page;
4975
4976 assert(is_valid_AArch64_address(dest.target()), "bad address");
4977 assert(dest.getMode() == Address::literal, "ADRP must be applied to a literal address");
4978
4979 InstructionMark im(this);
4980 code_section()->relocate(inst_mark(), dest.rspec());
4981 // 8143067: Ensure that the adrp can reach the dest from anywhere within
4982 // the code cache so that if it is relocated we know it will still reach
4983 if (offset_high >= -(1<<20) && offset_low < (1<<20)) {
4984 _adrp(reg1, dest.target());
4985 } else {
4986 uint64_t target = (uint64_t)dest.target();
4987 uint64_t adrp_target
4988 = (target & 0xffffffffULL) | ((uint64_t)pc() & 0xffff00000000ULL);
4989
4990 _adrp(reg1, (address)adrp_target);
4991 movk(reg1, target >> 32, 32);
4992 }
4993 byte_offset = (uint64_t)dest.target() & 0xfff;
4994 }
4995
4996 void MacroAssembler::load_byte_map_base(Register reg) {
4997 CardTable::CardValue* byte_map_base =
4998 ((CardTableBarrierSet*)(BarrierSet::barrier_set()))->card_table()->byte_map_base();
4999
5000 // Strictly speaking the byte_map_base isn't an address at all, and it might
5001 // even be negative. It is thus materialised as a constant.
5002 mov(reg, (uint64_t)byte_map_base);
5003 }
5004
5005 void MacroAssembler::build_frame(int framesize) {
5006 assert(framesize >= 2 * wordSize, "framesize must include space for FP/LR");
5007 assert(framesize % (2*wordSize) == 0, "must preserve 2*wordSize alignment");
5008 protect_return_address();
5009 if (framesize < ((1 << 9) + 2 * wordSize)) {
5010 sub(sp, sp, framesize);
5011 stp(rfp, lr, Address(sp, framesize - 2 * wordSize));
5012 if (PreserveFramePointer) add(rfp, sp, framesize - 2 * wordSize);
5013 } else {
5014 stp(rfp, lr, Address(pre(sp, -2 * wordSize)));
5015 if (PreserveFramePointer) mov(rfp, sp);
5016 if (framesize < ((1 << 12) + 2 * wordSize))
5017 sub(sp, sp, framesize - 2 * wordSize);
5018 else {
5019 mov(rscratch1, framesize - 2 * wordSize);
5020 sub(sp, sp, rscratch1);
5021 }
5022 }
5023 verify_cross_modify_fence_not_required();
5024 }
5025
5026 void MacroAssembler::remove_frame(int framesize) {
5027 assert(framesize >= 2 * wordSize, "framesize must include space for FP/LR");
5028 assert(framesize % (2*wordSize) == 0, "must preserve 2*wordSize alignment");
5029 if (framesize < ((1 << 9) + 2 * wordSize)) {
5030 ldp(rfp, lr, Address(sp, framesize - 2 * wordSize));
5031 add(sp, sp, framesize);
5032 } else {
5033 if (framesize < ((1 << 12) + 2 * wordSize))
5034 add(sp, sp, framesize - 2 * wordSize);
5035 else {
5036 mov(rscratch1, framesize - 2 * wordSize);
5037 add(sp, sp, rscratch1);
5038 }
5039 ldp(rfp, lr, Address(post(sp, 2 * wordSize)));
5040 }
5041 authenticate_return_address();
5042 }
5043
5044
5045 // This method counts leading positive bytes (highest bit not set) in provided byte array
5046 address MacroAssembler::count_positives(Register ary1, Register len, Register result) {
5047 // Simple and most common case of aligned small array which is not at the
5048 // end of memory page is placed here. All other cases are in stub.
5049 Label LOOP, END, STUB, STUB_LONG, SET_RESULT, DONE;
5050 const uint64_t UPPER_BIT_MASK=0x8080808080808080;
5051 assert_different_registers(ary1, len, result);
5052
5053 mov(result, len);
5054 cmpw(len, 0);
5055 br(LE, DONE);
5056 cmpw(len, 4 * wordSize);
5057 br(GE, STUB_LONG); // size > 32 then go to stub
5058
5059 int shift = 64 - exact_log2(os::vm_page_size());
5060 lsl(rscratch1, ary1, shift);
5061 mov(rscratch2, (size_t)(4 * wordSize) << shift);
5062 adds(rscratch2, rscratch1, rscratch2); // At end of page?
5063 br(CS, STUB); // at the end of page then go to stub
5064 subs(len, len, wordSize);
5065 br(LT, END);
5066
5067 BIND(LOOP);
5068 ldr(rscratch1, Address(post(ary1, wordSize)));
5069 tst(rscratch1, UPPER_BIT_MASK);
5070 br(NE, SET_RESULT);
5071 subs(len, len, wordSize);
5072 br(GE, LOOP);
5073 cmpw(len, -wordSize);
5074 br(EQ, DONE);
5075
5076 BIND(END);
5077 ldr(rscratch1, Address(ary1));
5078 sub(rscratch2, zr, len, LSL, 3); // LSL 3 is to get bits from bytes
5079 lslv(rscratch1, rscratch1, rscratch2);
5080 tst(rscratch1, UPPER_BIT_MASK);
5081 br(NE, SET_RESULT);
5082 b(DONE);
5083
5084 BIND(STUB);
5085 RuntimeAddress count_pos = RuntimeAddress(StubRoutines::aarch64::count_positives());
5086 assert(count_pos.target() != nullptr, "count_positives stub has not been generated");
5087 address tpc1 = trampoline_call(count_pos);
5088 if (tpc1 == nullptr) {
5089 DEBUG_ONLY(reset_labels(STUB_LONG, SET_RESULT, DONE));
5090 postcond(pc() == badAddress);
5091 return nullptr;
5092 }
5093 b(DONE);
5094
5095 BIND(STUB_LONG);
5096 RuntimeAddress count_pos_long = RuntimeAddress(StubRoutines::aarch64::count_positives_long());
5097 assert(count_pos_long.target() != nullptr, "count_positives_long stub has not been generated");
5098 address tpc2 = trampoline_call(count_pos_long);
5099 if (tpc2 == nullptr) {
5100 DEBUG_ONLY(reset_labels(SET_RESULT, DONE));
5101 postcond(pc() == badAddress);
5102 return nullptr;
5103 }
5104 b(DONE);
5105
5106 BIND(SET_RESULT);
5107
5108 add(len, len, wordSize);
5109 sub(result, result, len);
5110
5111 BIND(DONE);
5112 postcond(pc() != badAddress);
5113 return pc();
5114 }
5115
5116 // Clobbers: rscratch1, rscratch2, rflags
5117 // May also clobber v0-v7 when (!UseSimpleArrayEquals && UseSIMDForArrayEquals)
5118 address MacroAssembler::arrays_equals(Register a1, Register a2, Register tmp3,
5119 Register tmp4, Register tmp5, Register result,
5120 Register cnt1, int elem_size) {
5121 Label DONE, SAME;
5122 Register tmp1 = rscratch1;
5123 Register tmp2 = rscratch2;
5124 Register cnt2 = tmp2; // cnt2 only used in array length compare
5125 int elem_per_word = wordSize/elem_size;
5126 int log_elem_size = exact_log2(elem_size);
5127 int length_offset = arrayOopDesc::length_offset_in_bytes();
5128 int base_offset
5129 = arrayOopDesc::base_offset_in_bytes(elem_size == 2 ? T_CHAR : T_BYTE);
5130 int stubBytesThreshold = 3 * 64 + (UseSIMDForArrayEquals ? 0 : 16);
5131
5132 assert(elem_size == 1 || elem_size == 2, "must be char or byte");
5133 assert_different_registers(a1, a2, result, cnt1, rscratch1, rscratch2);
5134
5135 #ifndef PRODUCT
5136 {
5137 const char kind = (elem_size == 2) ? 'U' : 'L';
5138 char comment[64];
5139 snprintf(comment, sizeof comment, "array_equals%c{", kind);
5140 BLOCK_COMMENT(comment);
5141 }
5142 #endif
5143
5144 // if (a1 == a2)
5145 // return true;
5146 cmpoop(a1, a2); // May have read barriers for a1 and a2.
5147 br(EQ, SAME);
5148
5149 if (UseSimpleArrayEquals) {
5150 Label NEXT_WORD, SHORT, TAIL03, TAIL01, A_MIGHT_BE_NULL, A_IS_NOT_NULL;
5151 // if (a1 == nullptr || a2 == nullptr)
5152 // return false;
5153 // a1 & a2 == 0 means (some-pointer is null) or
5154 // (very-rare-or-even-probably-impossible-pointer-values)
5155 // so, we can save one branch in most cases
5156 tst(a1, a2);
5157 mov(result, false);
5158 br(EQ, A_MIGHT_BE_NULL);
5159 // if (a1.length != a2.length)
5160 // return false;
5161 bind(A_IS_NOT_NULL);
5162 ldrw(cnt1, Address(a1, length_offset));
5163 ldrw(cnt2, Address(a2, length_offset));
5164 eorw(tmp5, cnt1, cnt2);
5165 cbnzw(tmp5, DONE);
5166 lea(a1, Address(a1, base_offset));
5167 lea(a2, Address(a2, base_offset));
5168 // Check for short strings, i.e. smaller than wordSize.
5169 subs(cnt1, cnt1, elem_per_word);
5170 br(Assembler::LT, SHORT);
5171 // Main 8 byte comparison loop.
5172 bind(NEXT_WORD); {
5173 ldr(tmp1, Address(post(a1, wordSize)));
5174 ldr(tmp2, Address(post(a2, wordSize)));
5175 subs(cnt1, cnt1, elem_per_word);
5176 eor(tmp5, tmp1, tmp2);
5177 cbnz(tmp5, DONE);
5178 } br(GT, NEXT_WORD);
5179 // Last longword. In the case where length == 4 we compare the
5180 // same longword twice, but that's still faster than another
5181 // conditional branch.
5182 // cnt1 could be 0, -1, -2, -3, -4 for chars; -4 only happens when
5183 // length == 4.
5184 if (log_elem_size > 0)
5185 lsl(cnt1, cnt1, log_elem_size);
5186 ldr(tmp3, Address(a1, cnt1));
5187 ldr(tmp4, Address(a2, cnt1));
5188 eor(tmp5, tmp3, tmp4);
5189 cbnz(tmp5, DONE);
5190 b(SAME);
5191 bind(A_MIGHT_BE_NULL);
5192 // in case both a1 and a2 are not-null, proceed with loads
5193 cbz(a1, DONE);
5194 cbz(a2, DONE);
5195 b(A_IS_NOT_NULL);
5196 bind(SHORT);
5197
5198 tbz(cnt1, 2 - log_elem_size, TAIL03); // 0-7 bytes left.
5199 {
5200 ldrw(tmp1, Address(post(a1, 4)));
5201 ldrw(tmp2, Address(post(a2, 4)));
5202 eorw(tmp5, tmp1, tmp2);
5203 cbnzw(tmp5, DONE);
5204 }
5205 bind(TAIL03);
5206 tbz(cnt1, 1 - log_elem_size, TAIL01); // 0-3 bytes left.
5207 {
5208 ldrh(tmp3, Address(post(a1, 2)));
5209 ldrh(tmp4, Address(post(a2, 2)));
5210 eorw(tmp5, tmp3, tmp4);
5211 cbnzw(tmp5, DONE);
5212 }
5213 bind(TAIL01);
5214 if (elem_size == 1) { // Only needed when comparing byte arrays.
5215 tbz(cnt1, 0, SAME); // 0-1 bytes left.
5216 {
5217 ldrb(tmp1, a1);
5218 ldrb(tmp2, a2);
5219 eorw(tmp5, tmp1, tmp2);
5220 cbnzw(tmp5, DONE);
5221 }
5222 }
5223 } else {
5224 Label NEXT_DWORD, SHORT, TAIL, TAIL2, STUB,
5225 CSET_EQ, LAST_CHECK;
5226 mov(result, false);
5227 cbz(a1, DONE);
5228 ldrw(cnt1, Address(a1, length_offset));
5229 cbz(a2, DONE);
5230 ldrw(cnt2, Address(a2, length_offset));
5231 // on most CPUs a2 is still "locked"(surprisingly) in ldrw and it's
5232 // faster to perform another branch before comparing a1 and a2
5233 cmp(cnt1, (u1)elem_per_word);
5234 br(LE, SHORT); // short or same
5235 ldr(tmp3, Address(pre(a1, base_offset)));
5236 subs(zr, cnt1, stubBytesThreshold);
5237 br(GE, STUB);
5238 ldr(tmp4, Address(pre(a2, base_offset)));
5239 sub(tmp5, zr, cnt1, LSL, 3 + log_elem_size);
5240 cmp(cnt2, cnt1);
5241 br(NE, DONE);
5242
5243 // Main 16 byte comparison loop with 2 exits
5244 bind(NEXT_DWORD); {
5245 ldr(tmp1, Address(pre(a1, wordSize)));
5246 ldr(tmp2, Address(pre(a2, wordSize)));
5247 subs(cnt1, cnt1, 2 * elem_per_word);
5248 br(LE, TAIL);
5249 eor(tmp4, tmp3, tmp4);
5250 cbnz(tmp4, DONE);
5251 ldr(tmp3, Address(pre(a1, wordSize)));
5252 ldr(tmp4, Address(pre(a2, wordSize)));
5253 cmp(cnt1, (u1)elem_per_word);
5254 br(LE, TAIL2);
5255 cmp(tmp1, tmp2);
5256 } br(EQ, NEXT_DWORD);
5257 b(DONE);
5258
5259 bind(TAIL);
5260 eor(tmp4, tmp3, tmp4);
5261 eor(tmp2, tmp1, tmp2);
5262 lslv(tmp2, tmp2, tmp5);
5263 orr(tmp5, tmp4, tmp2);
5264 cmp(tmp5, zr);
5265 b(CSET_EQ);
5266
5267 bind(TAIL2);
5268 eor(tmp2, tmp1, tmp2);
5269 cbnz(tmp2, DONE);
5270 b(LAST_CHECK);
5271
5272 bind(STUB);
5273 ldr(tmp4, Address(pre(a2, base_offset)));
5274 cmp(cnt2, cnt1);
5275 br(NE, DONE);
5276 if (elem_size == 2) { // convert to byte counter
5277 lsl(cnt1, cnt1, 1);
5278 }
5279 eor(tmp5, tmp3, tmp4);
5280 cbnz(tmp5, DONE);
5281 RuntimeAddress stub = RuntimeAddress(StubRoutines::aarch64::large_array_equals());
5282 assert(stub.target() != nullptr, "array_equals_long stub has not been generated");
5283 address tpc = trampoline_call(stub);
5284 if (tpc == nullptr) {
5285 DEBUG_ONLY(reset_labels(SHORT, LAST_CHECK, CSET_EQ, SAME, DONE));
5286 postcond(pc() == badAddress);
5287 return nullptr;
5288 }
5289 b(DONE);
5290
5291 // (a1 != null && a2 == null) || (a1 != null && a2 != null && a1 == a2)
5292 // so, if a2 == null => return false(0), else return true, so we can return a2
5293 mov(result, a2);
5294 b(DONE);
5295 bind(SHORT);
5296 cmp(cnt2, cnt1);
5297 br(NE, DONE);
5298 cbz(cnt1, SAME);
5299 sub(tmp5, zr, cnt1, LSL, 3 + log_elem_size);
5300 ldr(tmp3, Address(a1, base_offset));
5301 ldr(tmp4, Address(a2, base_offset));
5302 bind(LAST_CHECK);
5303 eor(tmp4, tmp3, tmp4);
5304 lslv(tmp5, tmp4, tmp5);
5305 cmp(tmp5, zr);
5306 bind(CSET_EQ);
5307 cset(result, EQ);
5308 b(DONE);
5309 }
5310
5311 bind(SAME);
5312 mov(result, true);
5313 // That's it.
5314 bind(DONE);
5315
5316 BLOCK_COMMENT("} array_equals");
5317 postcond(pc() != badAddress);
5318 return pc();
5319 }
5320
5321 // Compare Strings
5322
5323 // For Strings we're passed the address of the first characters in a1
5324 // and a2 and the length in cnt1.
5325 // elem_size is the element size in bytes: either 1 or 2.
5326 // There are two implementations. For arrays >= 8 bytes, all
5327 // comparisons (including the final one, which may overlap) are
5328 // performed 8 bytes at a time. For strings < 8 bytes, we compare a
5329 // halfword, then a short, and then a byte.
5330
5331 void MacroAssembler::string_equals(Register a1, Register a2,
5332 Register result, Register cnt1, int elem_size)
5333 {
5334 Label SAME, DONE, SHORT, NEXT_WORD;
5335 Register tmp1 = rscratch1;
5336 Register tmp2 = rscratch2;
5337 Register cnt2 = tmp2; // cnt2 only used in array length compare
5338
5339 assert(elem_size == 1 || elem_size == 2, "must be 2 or 1 byte");
5340 assert_different_registers(a1, a2, result, cnt1, rscratch1, rscratch2);
5341
5342 #ifndef PRODUCT
5343 {
5344 const char kind = (elem_size == 2) ? 'U' : 'L';
5345 char comment[64];
5346 snprintf(comment, sizeof comment, "{string_equals%c", kind);
5347 BLOCK_COMMENT(comment);
5348 }
5349 #endif
5350
5351 mov(result, false);
5352
5353 // Check for short strings, i.e. smaller than wordSize.
5354 subs(cnt1, cnt1, wordSize);
5355 br(Assembler::LT, SHORT);
5356 // Main 8 byte comparison loop.
5357 bind(NEXT_WORD); {
5358 ldr(tmp1, Address(post(a1, wordSize)));
5359 ldr(tmp2, Address(post(a2, wordSize)));
5360 subs(cnt1, cnt1, wordSize);
5361 eor(tmp1, tmp1, tmp2);
5362 cbnz(tmp1, DONE);
5363 } br(GT, NEXT_WORD);
5364 // Last longword. In the case where length == 4 we compare the
5365 // same longword twice, but that's still faster than another
5366 // conditional branch.
5367 // cnt1 could be 0, -1, -2, -3, -4 for chars; -4 only happens when
5368 // length == 4.
5369 ldr(tmp1, Address(a1, cnt1));
5370 ldr(tmp2, Address(a2, cnt1));
5371 eor(tmp2, tmp1, tmp2);
5372 cbnz(tmp2, DONE);
5373 b(SAME);
5374
5375 bind(SHORT);
5376 Label TAIL03, TAIL01;
5377
5378 tbz(cnt1, 2, TAIL03); // 0-7 bytes left.
5379 {
5380 ldrw(tmp1, Address(post(a1, 4)));
5381 ldrw(tmp2, Address(post(a2, 4)));
5382 eorw(tmp1, tmp1, tmp2);
5383 cbnzw(tmp1, DONE);
5384 }
5385 bind(TAIL03);
5386 tbz(cnt1, 1, TAIL01); // 0-3 bytes left.
5387 {
5388 ldrh(tmp1, Address(post(a1, 2)));
5389 ldrh(tmp2, Address(post(a2, 2)));
5390 eorw(tmp1, tmp1, tmp2);
5391 cbnzw(tmp1, DONE);
5392 }
5393 bind(TAIL01);
5394 if (elem_size == 1) { // Only needed when comparing 1-byte elements
5395 tbz(cnt1, 0, SAME); // 0-1 bytes left.
5396 {
5397 ldrb(tmp1, a1);
5398 ldrb(tmp2, a2);
5399 eorw(tmp1, tmp1, tmp2);
5400 cbnzw(tmp1, DONE);
5401 }
5402 }
5403 // Arrays are equal.
5404 bind(SAME);
5405 mov(result, true);
5406
5407 // That's it.
5408 bind(DONE);
5409 BLOCK_COMMENT("} string_equals");
5410 }
5411
5412
5413 // The size of the blocks erased by the zero_blocks stub. We must
5414 // handle anything smaller than this ourselves in zero_words().
5415 const int MacroAssembler::zero_words_block_size = 8;
5416
5417 // zero_words() is used by C2 ClearArray patterns and by
5418 // C1_MacroAssembler. It is as small as possible, handling small word
5419 // counts locally and delegating anything larger to the zero_blocks
5420 // stub. It is expanded many times in compiled code, so it is
5421 // important to keep it short.
5422
5423 // ptr: Address of a buffer to be zeroed.
5424 // cnt: Count in HeapWords.
5425 //
5426 // ptr, cnt, rscratch1, and rscratch2 are clobbered.
5427 address MacroAssembler::zero_words(Register ptr, Register cnt)
5428 {
5429 assert(is_power_of_2(zero_words_block_size), "adjust this");
5430
5431 BLOCK_COMMENT("zero_words {");
5432 assert(ptr == r10 && cnt == r11, "mismatch in register usage");
5433 RuntimeAddress zero_blocks = RuntimeAddress(StubRoutines::aarch64::zero_blocks());
5434 assert(zero_blocks.target() != nullptr, "zero_blocks stub has not been generated");
5435
5436 subs(rscratch1, cnt, zero_words_block_size);
5437 Label around;
5438 br(LO, around);
5439 {
5440 RuntimeAddress zero_blocks = RuntimeAddress(StubRoutines::aarch64::zero_blocks());
5441 assert(zero_blocks.target() != nullptr, "zero_blocks stub has not been generated");
5442 // Make sure this is a C2 compilation. C1 allocates space only for
5443 // trampoline stubs generated by Call LIR ops, and in any case it
5444 // makes sense for a C1 compilation task to proceed as quickly as
5445 // possible.
5446 CompileTask* task;
5447 if (StubRoutines::aarch64::complete()
5448 && Thread::current()->is_Compiler_thread()
5449 && (task = ciEnv::current()->task())
5450 && is_c2_compile(task->comp_level())) {
5451 address tpc = trampoline_call(zero_blocks);
5452 if (tpc == nullptr) {
5453 DEBUG_ONLY(reset_labels(around));
5454 return nullptr;
5455 }
5456 } else {
5457 far_call(zero_blocks);
5458 }
5459 }
5460 bind(around);
5461
5462 // We have a few words left to do. zero_blocks has adjusted r10 and r11
5463 // for us.
5464 for (int i = zero_words_block_size >> 1; i > 1; i >>= 1) {
5465 Label l;
5466 tbz(cnt, exact_log2(i), l);
5467 for (int j = 0; j < i; j += 2) {
5468 stp(zr, zr, post(ptr, 2 * BytesPerWord));
5469 }
5470 bind(l);
5471 }
5472 {
5473 Label l;
5474 tbz(cnt, 0, l);
5475 str(zr, Address(ptr));
5476 bind(l);
5477 }
5478
5479 BLOCK_COMMENT("} zero_words");
5480 return pc();
5481 }
5482
5483 // base: Address of a buffer to be zeroed, 8 bytes aligned.
5484 // cnt: Immediate count in HeapWords.
5485 //
5486 // r10, r11, rscratch1, and rscratch2 are clobbered.
5487 address MacroAssembler::zero_words(Register base, uint64_t cnt)
5488 {
5489 assert(wordSize <= BlockZeroingLowLimit,
5490 "increase BlockZeroingLowLimit");
5491 address result = nullptr;
5492 if (cnt <= (uint64_t)BlockZeroingLowLimit / BytesPerWord) {
5493 #ifndef PRODUCT
5494 {
5495 char buf[64];
5496 snprintf(buf, sizeof buf, "zero_words (count = %" PRIu64 ") {", cnt);
5497 BLOCK_COMMENT(buf);
5498 }
5499 #endif
5500 if (cnt >= 16) {
5501 uint64_t loops = cnt/16;
5502 if (loops > 1) {
5503 mov(rscratch2, loops - 1);
5504 }
5505 {
5506 Label loop;
5507 bind(loop);
5508 for (int i = 0; i < 16; i += 2) {
5509 stp(zr, zr, Address(base, i * BytesPerWord));
5510 }
5511 add(base, base, 16 * BytesPerWord);
5512 if (loops > 1) {
5513 subs(rscratch2, rscratch2, 1);
5514 br(GE, loop);
5515 }
5516 }
5517 }
5518 cnt %= 16;
5519 int i = cnt & 1; // store any odd word to start
5520 if (i) str(zr, Address(base));
5521 for (; i < (int)cnt; i += 2) {
5522 stp(zr, zr, Address(base, i * wordSize));
5523 }
5524 BLOCK_COMMENT("} zero_words");
5525 result = pc();
5526 } else {
5527 mov(r10, base); mov(r11, cnt);
5528 result = zero_words(r10, r11);
5529 }
5530 return result;
5531 }
5532
5533 // Zero blocks of memory by using DC ZVA.
5534 //
5535 // Aligns the base address first sufficiently for DC ZVA, then uses
5536 // DC ZVA repeatedly for every full block. cnt is the size to be
5537 // zeroed in HeapWords. Returns the count of words left to be zeroed
5538 // in cnt.
5539 //
5540 // NOTE: This is intended to be used in the zero_blocks() stub. If
5541 // you want to use it elsewhere, note that cnt must be >= 2*zva_length.
5542 void MacroAssembler::zero_dcache_blocks(Register base, Register cnt) {
5543 Register tmp = rscratch1;
5544 Register tmp2 = rscratch2;
5545 int zva_length = VM_Version::zva_length();
5546 Label initial_table_end, loop_zva;
5547 Label fini;
5548
5549 // Base must be 16 byte aligned. If not just return and let caller handle it
5550 tst(base, 0x0f);
5551 br(Assembler::NE, fini);
5552 // Align base with ZVA length.
5553 neg(tmp, base);
5554 andr(tmp, tmp, zva_length - 1);
5555
5556 // tmp: the number of bytes to be filled to align the base with ZVA length.
5557 add(base, base, tmp);
5558 sub(cnt, cnt, tmp, Assembler::ASR, 3);
5559 adr(tmp2, initial_table_end);
5560 sub(tmp2, tmp2, tmp, Assembler::LSR, 2);
5561 br(tmp2);
5562
5563 for (int i = -zva_length + 16; i < 0; i += 16)
5564 stp(zr, zr, Address(base, i));
5565 bind(initial_table_end);
5566
5567 sub(cnt, cnt, zva_length >> 3);
5568 bind(loop_zva);
5569 dc(Assembler::ZVA, base);
5570 subs(cnt, cnt, zva_length >> 3);
5571 add(base, base, zva_length);
5572 br(Assembler::GE, loop_zva);
5573 add(cnt, cnt, zva_length >> 3); // count not zeroed by DC ZVA
5574 bind(fini);
5575 }
5576
5577 // base: Address of a buffer to be filled, 8 bytes aligned.
5578 // cnt: Count in 8-byte unit.
5579 // value: Value to be filled with.
5580 // base will point to the end of the buffer after filling.
5581 void MacroAssembler::fill_words(Register base, Register cnt, Register value)
5582 {
5583 // Algorithm:
5584 //
5585 // if (cnt == 0) {
5586 // return;
5587 // }
5588 // if ((p & 8) != 0) {
5589 // *p++ = v;
5590 // }
5591 //
5592 // scratch1 = cnt & 14;
5593 // cnt -= scratch1;
5594 // p += scratch1;
5595 // switch (scratch1 / 2) {
5596 // do {
5597 // cnt -= 16;
5598 // p[-16] = v;
5599 // p[-15] = v;
5600 // case 7:
5601 // p[-14] = v;
5602 // p[-13] = v;
5603 // case 6:
5604 // p[-12] = v;
5605 // p[-11] = v;
5606 // // ...
5607 // case 1:
5608 // p[-2] = v;
5609 // p[-1] = v;
5610 // case 0:
5611 // p += 16;
5612 // } while (cnt);
5613 // }
5614 // if ((cnt & 1) == 1) {
5615 // *p++ = v;
5616 // }
5617
5618 assert_different_registers(base, cnt, value, rscratch1, rscratch2);
5619
5620 Label fini, skip, entry, loop;
5621 const int unroll = 8; // Number of stp instructions we'll unroll
5622
5623 cbz(cnt, fini);
5624 tbz(base, 3, skip);
5625 str(value, Address(post(base, 8)));
5626 sub(cnt, cnt, 1);
5627 bind(skip);
5628
5629 andr(rscratch1, cnt, (unroll-1) * 2);
5630 sub(cnt, cnt, rscratch1);
5631 add(base, base, rscratch1, Assembler::LSL, 3);
5632 adr(rscratch2, entry);
5633 sub(rscratch2, rscratch2, rscratch1, Assembler::LSL, 1);
5634 br(rscratch2);
5635
5636 bind(loop);
5637 add(base, base, unroll * 16);
5638 for (int i = -unroll; i < 0; i++)
5639 stp(value, value, Address(base, i * 16));
5640 bind(entry);
5641 subs(cnt, cnt, unroll * 2);
5642 br(Assembler::GE, loop);
5643
5644 tbz(cnt, 0, fini);
5645 str(value, Address(post(base, 8)));
5646 bind(fini);
5647 }
5648
5649 // Intrinsic for
5650 //
5651 // - sun/nio/cs/ISO_8859_1$Encoder.implEncodeISOArray
5652 // return the number of characters copied.
5653 // - java/lang/StringUTF16.compress
5654 // return zero (0) if copy fails, otherwise 'len'.
5655 //
5656 // This version always returns the number of characters copied, and does not
5657 // clobber the 'len' register. A successful copy will complete with the post-
5658 // condition: 'res' == 'len', while an unsuccessful copy will exit with the
5659 // post-condition: 0 <= 'res' < 'len'.
5660 //
5661 // NOTE: Attempts to use 'ld2' (and 'umaxv' in the ISO part) has proven to
5662 // degrade performance (on Ampere Altra - Neoverse N1), to an extent
5663 // beyond the acceptable, even though the footprint would be smaller.
5664 // Using 'umaxv' in the ASCII-case comes with a small penalty but does
5665 // avoid additional bloat.
5666 //
5667 // Clobbers: src, dst, res, rscratch1, rscratch2, rflags
5668 void MacroAssembler::encode_iso_array(Register src, Register dst,
5669 Register len, Register res, bool ascii,
5670 FloatRegister vtmp0, FloatRegister vtmp1,
5671 FloatRegister vtmp2, FloatRegister vtmp3,
5672 FloatRegister vtmp4, FloatRegister vtmp5)
5673 {
5674 Register cnt = res;
5675 Register max = rscratch1;
5676 Register chk = rscratch2;
5677
5678 prfm(Address(src), PLDL1STRM);
5679 movw(cnt, len);
5680
5681 #define ASCII(insn) do { if (ascii) { insn; } } while (0)
5682
5683 Label LOOP_32, DONE_32, FAIL_32;
5684
5685 BIND(LOOP_32);
5686 {
5687 cmpw(cnt, 32);
5688 br(LT, DONE_32);
5689 ld1(vtmp0, vtmp1, vtmp2, vtmp3, T8H, Address(post(src, 64)));
5690 // Extract lower bytes.
5691 FloatRegister vlo0 = vtmp4;
5692 FloatRegister vlo1 = vtmp5;
5693 uzp1(vlo0, T16B, vtmp0, vtmp1);
5694 uzp1(vlo1, T16B, vtmp2, vtmp3);
5695 // Merge bits...
5696 orr(vtmp0, T16B, vtmp0, vtmp1);
5697 orr(vtmp2, T16B, vtmp2, vtmp3);
5698 // Extract merged upper bytes.
5699 FloatRegister vhix = vtmp0;
5700 uzp2(vhix, T16B, vtmp0, vtmp2);
5701 // ISO-check on hi-parts (all zero).
5702 // ASCII-check on lo-parts (no sign).
5703 FloatRegister vlox = vtmp1; // Merge lower bytes.
5704 ASCII(orr(vlox, T16B, vlo0, vlo1));
5705 umov(chk, vhix, D, 1); ASCII(cm(LT, vlox, T16B, vlox));
5706 fmovd(max, vhix); ASCII(umaxv(vlox, T16B, vlox));
5707 orr(chk, chk, max); ASCII(umov(max, vlox, B, 0));
5708 ASCII(orr(chk, chk, max));
5709 cbnz(chk, FAIL_32);
5710 subw(cnt, cnt, 32);
5711 st1(vlo0, vlo1, T16B, Address(post(dst, 32)));
5712 b(LOOP_32);
5713 }
5714 BIND(FAIL_32);
5715 sub(src, src, 64);
5716 BIND(DONE_32);
5717
5718 Label LOOP_8, SKIP_8;
5719
5720 BIND(LOOP_8);
5721 {
5722 cmpw(cnt, 8);
5723 br(LT, SKIP_8);
5724 FloatRegister vhi = vtmp0;
5725 FloatRegister vlo = vtmp1;
5726 ld1(vtmp3, T8H, src);
5727 uzp1(vlo, T16B, vtmp3, vtmp3);
5728 uzp2(vhi, T16B, vtmp3, vtmp3);
5729 // ISO-check on hi-parts (all zero).
5730 // ASCII-check on lo-parts (no sign).
5731 ASCII(cm(LT, vtmp2, T16B, vlo));
5732 fmovd(chk, vhi); ASCII(umaxv(vtmp2, T16B, vtmp2));
5733 ASCII(umov(max, vtmp2, B, 0));
5734 ASCII(orr(chk, chk, max));
5735 cbnz(chk, SKIP_8);
5736
5737 strd(vlo, Address(post(dst, 8)));
5738 subw(cnt, cnt, 8);
5739 add(src, src, 16);
5740 b(LOOP_8);
5741 }
5742 BIND(SKIP_8);
5743
5744 #undef ASCII
5745
5746 Label LOOP, DONE;
5747
5748 cbz(cnt, DONE);
5749 BIND(LOOP);
5750 {
5751 Register chr = rscratch1;
5752 ldrh(chr, Address(post(src, 2)));
5753 tst(chr, ascii ? 0xff80 : 0xff00);
5754 br(NE, DONE);
5755 strb(chr, Address(post(dst, 1)));
5756 subs(cnt, cnt, 1);
5757 br(GT, LOOP);
5758 }
5759 BIND(DONE);
5760 // Return index where we stopped.
5761 subw(res, len, cnt);
5762 }
5763
5764 // Inflate byte[] array to char[].
5765 // Clobbers: src, dst, len, rflags, rscratch1, v0-v6
5766 address MacroAssembler::byte_array_inflate(Register src, Register dst, Register len,
5767 FloatRegister vtmp1, FloatRegister vtmp2,
5768 FloatRegister vtmp3, Register tmp4) {
5769 Label big, done, after_init, to_stub;
5770
5771 assert_different_registers(src, dst, len, tmp4, rscratch1);
5772
5773 fmovd(vtmp1, 0.0);
5774 lsrw(tmp4, len, 3);
5775 bind(after_init);
5776 cbnzw(tmp4, big);
5777 // Short string: less than 8 bytes.
5778 {
5779 Label loop, tiny;
5780
5781 cmpw(len, 4);
5782 br(LT, tiny);
5783 // Use SIMD to do 4 bytes.
5784 ldrs(vtmp2, post(src, 4));
5785 zip1(vtmp3, T8B, vtmp2, vtmp1);
5786 subw(len, len, 4);
5787 strd(vtmp3, post(dst, 8));
5788
5789 cbzw(len, done);
5790
5791 // Do the remaining bytes by steam.
5792 bind(loop);
5793 ldrb(tmp4, post(src, 1));
5794 strh(tmp4, post(dst, 2));
5795 subw(len, len, 1);
5796
5797 bind(tiny);
5798 cbnz(len, loop);
5799
5800 b(done);
5801 }
5802
5803 if (SoftwarePrefetchHintDistance >= 0) {
5804 bind(to_stub);
5805 RuntimeAddress stub = RuntimeAddress(StubRoutines::aarch64::large_byte_array_inflate());
5806 assert(stub.target() != nullptr, "large_byte_array_inflate stub has not been generated");
5807 address tpc = trampoline_call(stub);
5808 if (tpc == nullptr) {
5809 DEBUG_ONLY(reset_labels(big, done));
5810 postcond(pc() == badAddress);
5811 return nullptr;
5812 }
5813 b(after_init);
5814 }
5815
5816 // Unpack the bytes 8 at a time.
5817 bind(big);
5818 {
5819 Label loop, around, loop_last, loop_start;
5820
5821 if (SoftwarePrefetchHintDistance >= 0) {
5822 const int large_loop_threshold = (64 + 16)/8;
5823 ldrd(vtmp2, post(src, 8));
5824 andw(len, len, 7);
5825 cmp(tmp4, (u1)large_loop_threshold);
5826 br(GE, to_stub);
5827 b(loop_start);
5828
5829 bind(loop);
5830 ldrd(vtmp2, post(src, 8));
5831 bind(loop_start);
5832 subs(tmp4, tmp4, 1);
5833 br(EQ, loop_last);
5834 zip1(vtmp2, T16B, vtmp2, vtmp1);
5835 ldrd(vtmp3, post(src, 8));
5836 st1(vtmp2, T8H, post(dst, 16));
5837 subs(tmp4, tmp4, 1);
5838 zip1(vtmp3, T16B, vtmp3, vtmp1);
5839 st1(vtmp3, T8H, post(dst, 16));
5840 br(NE, loop);
5841 b(around);
5842 bind(loop_last);
5843 zip1(vtmp2, T16B, vtmp2, vtmp1);
5844 st1(vtmp2, T8H, post(dst, 16));
5845 bind(around);
5846 cbz(len, done);
5847 } else {
5848 andw(len, len, 7);
5849 bind(loop);
5850 ldrd(vtmp2, post(src, 8));
5851 sub(tmp4, tmp4, 1);
5852 zip1(vtmp3, T16B, vtmp2, vtmp1);
5853 st1(vtmp3, T8H, post(dst, 16));
5854 cbnz(tmp4, loop);
5855 }
5856 }
5857
5858 // Do the tail of up to 8 bytes.
5859 add(src, src, len);
5860 ldrd(vtmp3, Address(src, -8));
5861 add(dst, dst, len, ext::uxtw, 1);
5862 zip1(vtmp3, T16B, vtmp3, vtmp1);
5863 strq(vtmp3, Address(dst, -16));
5864
5865 bind(done);
5866 postcond(pc() != badAddress);
5867 return pc();
5868 }
5869
5870 // Compress char[] array to byte[].
5871 void MacroAssembler::char_array_compress(Register src, Register dst, Register len,
5872 Register res,
5873 FloatRegister tmp0, FloatRegister tmp1,
5874 FloatRegister tmp2, FloatRegister tmp3,
5875 FloatRegister tmp4, FloatRegister tmp5) {
5876 encode_iso_array(src, dst, len, res, false, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5);
5877 // Adjust result: res == len ? len : 0
5878 cmp(len, res);
5879 csel(res, res, zr, EQ);
5880 }
5881
5882 // java.math.round(double a)
5883 // Returns the closest long to the argument, with ties rounding to
5884 // positive infinity. This requires some fiddling for corner
5885 // cases. We take care to avoid double rounding in e.g. (jlong)(a + 0.5).
5886 void MacroAssembler::java_round_double(Register dst, FloatRegister src,
5887 FloatRegister ftmp) {
5888 Label DONE;
5889 BLOCK_COMMENT("java_round_double: { ");
5890 fmovd(rscratch1, src);
5891 // Use RoundToNearestTiesAway unless src small and -ve.
5892 fcvtasd(dst, src);
5893 // Test if src >= 0 || abs(src) >= 0x1.0p52
5894 eor(rscratch1, rscratch1, UCONST64(1) << 63); // flip sign bit
5895 mov(rscratch2, julong_cast(0x1.0p52));
5896 cmp(rscratch1, rscratch2);
5897 br(HS, DONE); {
5898 // src < 0 && abs(src) < 0x1.0p52
5899 // src may have a fractional part, so add 0.5
5900 fmovd(ftmp, 0.5);
5901 faddd(ftmp, src, ftmp);
5902 // Convert double to jlong, use RoundTowardsNegative
5903 fcvtmsd(dst, ftmp);
5904 }
5905 bind(DONE);
5906 BLOCK_COMMENT("} java_round_double");
5907 }
5908
5909 void MacroAssembler::java_round_float(Register dst, FloatRegister src,
5910 FloatRegister ftmp) {
5911 Label DONE;
5912 BLOCK_COMMENT("java_round_float: { ");
5913 fmovs(rscratch1, src);
5914 // Use RoundToNearestTiesAway unless src small and -ve.
5915 fcvtassw(dst, src);
5916 // Test if src >= 0 || abs(src) >= 0x1.0p23
5917 eor(rscratch1, rscratch1, 0x80000000); // flip sign bit
5918 mov(rscratch2, jint_cast(0x1.0p23f));
5919 cmp(rscratch1, rscratch2);
5920 br(HS, DONE); {
5921 // src < 0 && |src| < 0x1.0p23
5922 // src may have a fractional part, so add 0.5
5923 fmovs(ftmp, 0.5f);
5924 fadds(ftmp, src, ftmp);
5925 // Convert float to jint, use RoundTowardsNegative
5926 fcvtmssw(dst, ftmp);
5927 }
5928 bind(DONE);
5929 BLOCK_COMMENT("} java_round_float");
5930 }
5931
5932 // get_thread() can be called anywhere inside generated code so we
5933 // need to save whatever non-callee save context might get clobbered
5934 // by the call to JavaThread::aarch64_get_thread_helper() or, indeed,
5935 // the call setup code.
5936 //
5937 // On Linux, aarch64_get_thread_helper() clobbers only r0, r1, and flags.
5938 // On other systems, the helper is a usual C function.
5939 //
5940 void MacroAssembler::get_thread(Register dst) {
5941 RegSet saved_regs =
5942 LINUX_ONLY(RegSet::range(r0, r1) + lr - dst)
5943 NOT_LINUX (RegSet::range(r0, r17) + lr - dst);
5944
5945 protect_return_address();
5946 push(saved_regs, sp);
5947
5948 mov(lr, CAST_FROM_FN_PTR(address, JavaThread::aarch64_get_thread_helper));
5949 blr(lr);
5950 if (dst != c_rarg0) {
5951 mov(dst, c_rarg0);
5952 }
5953
5954 pop(saved_regs, sp);
5955 authenticate_return_address();
5956 }
5957
5958 void MacroAssembler::cache_wb(Address line) {
5959 assert(line.getMode() == Address::base_plus_offset, "mode should be base_plus_offset");
5960 assert(line.index() == noreg, "index should be noreg");
5961 assert(line.offset() == 0, "offset should be 0");
5962 // would like to assert this
5963 // assert(line._ext.shift == 0, "shift should be zero");
5964 if (VM_Version::supports_dcpop()) {
5965 // writeback using clear virtual address to point of persistence
5966 dc(Assembler::CVAP, line.base());
5967 } else {
5968 // no need to generate anything as Unsafe.writebackMemory should
5969 // never invoke this stub
5970 }
5971 }
5972
5973 void MacroAssembler::cache_wbsync(bool is_pre) {
5974 // we only need a barrier post sync
5975 if (!is_pre) {
5976 membar(Assembler::AnyAny);
5977 }
5978 }
5979
5980 void MacroAssembler::verify_sve_vector_length(Register tmp) {
5981 // Make sure that native code does not change SVE vector length.
5982 if (!UseSVE) return;
5983 Label verify_ok;
5984 movw(tmp, zr);
5985 sve_inc(tmp, B);
5986 subsw(zr, tmp, VM_Version::get_initial_sve_vector_length());
5987 br(EQ, verify_ok);
5988 stop("Error: SVE vector length has changed since jvm startup");
5989 bind(verify_ok);
5990 }
5991
5992 void MacroAssembler::verify_ptrue() {
5993 Label verify_ok;
5994 if (!UseSVE) {
5995 return;
5996 }
5997 sve_cntp(rscratch1, B, ptrue, ptrue); // get true elements count.
5998 sve_dec(rscratch1, B);
5999 cbz(rscratch1, verify_ok);
6000 stop("Error: the preserved predicate register (p7) elements are not all true");
6001 bind(verify_ok);
6002 }
6003
6004 void MacroAssembler::safepoint_isb() {
6005 isb();
6006 #ifndef PRODUCT
6007 if (VerifyCrossModifyFence) {
6008 // Clear the thread state.
6009 strb(zr, Address(rthread, in_bytes(JavaThread::requires_cross_modify_fence_offset())));
6010 }
6011 #endif
6012 }
6013
6014 #ifndef PRODUCT
6015 void MacroAssembler::verify_cross_modify_fence_not_required() {
6016 if (VerifyCrossModifyFence) {
6017 // Check if thread needs a cross modify fence.
6018 ldrb(rscratch1, Address(rthread, in_bytes(JavaThread::requires_cross_modify_fence_offset())));
6019 Label fence_not_required;
6020 cbz(rscratch1, fence_not_required);
6021 // If it does then fail.
6022 lea(rscratch1, CAST_FROM_FN_PTR(address, JavaThread::verify_cross_modify_fence_failure));
6023 mov(c_rarg0, rthread);
6024 blr(rscratch1);
6025 bind(fence_not_required);
6026 }
6027 }
6028 #endif
6029
6030 void MacroAssembler::spin_wait() {
6031 for (int i = 0; i < VM_Version::spin_wait_desc().inst_count(); ++i) {
6032 switch (VM_Version::spin_wait_desc().inst()) {
6033 case SpinWait::NOP:
6034 nop();
6035 break;
6036 case SpinWait::ISB:
6037 isb();
6038 break;
6039 case SpinWait::YIELD:
6040 yield();
6041 break;
6042 default:
6043 ShouldNotReachHere();
6044 }
6045 }
6046 }
6047
6048 // Stack frame creation/removal
6049
6050 void MacroAssembler::enter(bool strip_ret_addr) {
6051 if (strip_ret_addr) {
6052 // Addresses can only be signed once. If there are multiple nested frames being created
6053 // in the same function, then the return address needs stripping first.
6054 strip_return_address();
6055 }
6056 protect_return_address();
6057 stp(rfp, lr, Address(pre(sp, -2 * wordSize)));
6058 mov(rfp, sp);
6059 }
6060
6061 void MacroAssembler::leave() {
6062 mov(sp, rfp);
6063 ldp(rfp, lr, Address(post(sp, 2 * wordSize)));
6064 authenticate_return_address();
6065 }
6066
6067 // ROP Protection
6068 // Use the AArch64 PAC feature to add ROP protection for generated code. Use whenever creating/
6069 // destroying stack frames or whenever directly loading/storing the LR to memory.
6070 // If ROP protection is not set then these functions are no-ops.
6071 // For more details on PAC see pauth_aarch64.hpp.
6072
6073 // Sign the LR. Use during construction of a stack frame, before storing the LR to memory.
6074 // Uses value zero as the modifier.
6075 //
6076 void MacroAssembler::protect_return_address() {
6077 if (VM_Version::use_rop_protection()) {
6078 check_return_address();
6079 paciaz();
6080 }
6081 }
6082
6083 // Sign the return value in the given register. Use before updating the LR in the existing stack
6084 // frame for the current function.
6085 // Uses value zero as the modifier.
6086 //
6087 void MacroAssembler::protect_return_address(Register return_reg) {
6088 if (VM_Version::use_rop_protection()) {
6089 check_return_address(return_reg);
6090 paciza(return_reg);
6091 }
6092 }
6093
6094 // Authenticate the LR. Use before function return, after restoring FP and loading LR from memory.
6095 // Uses value zero as the modifier.
6096 //
6097 void MacroAssembler::authenticate_return_address() {
6098 if (VM_Version::use_rop_protection()) {
6099 autiaz();
6100 check_return_address();
6101 }
6102 }
6103
6104 // Authenticate the return value in the given register. Use before updating the LR in the existing
6105 // stack frame for the current function.
6106 // Uses value zero as the modifier.
6107 //
6108 void MacroAssembler::authenticate_return_address(Register return_reg) {
6109 if (VM_Version::use_rop_protection()) {
6110 autiza(return_reg);
6111 check_return_address(return_reg);
6112 }
6113 }
6114
6115 // Strip any PAC data from LR without performing any authentication. Use with caution - only if
6116 // there is no guaranteed way of authenticating the LR.
6117 //
6118 void MacroAssembler::strip_return_address() {
6119 if (VM_Version::use_rop_protection()) {
6120 xpaclri();
6121 }
6122 }
6123
6124 #ifndef PRODUCT
6125 // PAC failures can be difficult to debug. After an authentication failure, a segfault will only
6126 // occur when the pointer is used - ie when the program returns to the invalid LR. At this point
6127 // it is difficult to debug back to the callee function.
6128 // This function simply loads from the address in the given register.
6129 // Use directly after authentication to catch authentication failures.
6130 // Also use before signing to check that the pointer is valid and hasn't already been signed.
6131 //
6132 void MacroAssembler::check_return_address(Register return_reg) {
6133 if (VM_Version::use_rop_protection()) {
6134 ldr(zr, Address(return_reg));
6135 }
6136 }
6137 #endif
6138
6139 // The java_calling_convention describes stack locations as ideal slots on
6140 // a frame with no abi restrictions. Since we must observe abi restrictions
6141 // (like the placement of the register window) the slots must be biased by
6142 // the following value.
6143 static int reg2offset_in(VMReg r) {
6144 // Account for saved rfp and lr
6145 // This should really be in_preserve_stack_slots
6146 return (r->reg2stack() + 4) * VMRegImpl::stack_slot_size;
6147 }
6148
6149 static int reg2offset_out(VMReg r) {
6150 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
6151 }
6152
6153 // On 64bit we will store integer like items to the stack as
6154 // 64bits items (AArch64 ABI) even though java would only store
6155 // 32bits for a parameter. On 32bit it will simply be 32bits
6156 // So this routine will do 32->32 on 32bit and 32->64 on 64bit
6157 void MacroAssembler::move32_64(VMRegPair src, VMRegPair dst, Register tmp) {
6158 if (src.first()->is_stack()) {
6159 if (dst.first()->is_stack()) {
6160 // stack to stack
6161 ldr(tmp, Address(rfp, reg2offset_in(src.first())));
6162 str(tmp, Address(sp, reg2offset_out(dst.first())));
6163 } else {
6164 // stack to reg
6165 ldrsw(dst.first()->as_Register(), Address(rfp, reg2offset_in(src.first())));
6166 }
6167 } else if (dst.first()->is_stack()) {
6168 // reg to stack
6169 str(src.first()->as_Register(), Address(sp, reg2offset_out(dst.first())));
6170 } else {
6171 if (dst.first() != src.first()) {
6172 sxtw(dst.first()->as_Register(), src.first()->as_Register());
6173 }
6174 }
6175 }
6176
6177 // An oop arg. Must pass a handle not the oop itself
6178 void MacroAssembler::object_move(
6179 OopMap* map,
6180 int oop_handle_offset,
6181 int framesize_in_slots,
6182 VMRegPair src,
6183 VMRegPair dst,
6184 bool is_receiver,
6185 int* receiver_offset) {
6186
6187 // must pass a handle. First figure out the location we use as a handle
6188
6189 Register rHandle = dst.first()->is_stack() ? rscratch2 : dst.first()->as_Register();
6190
6191 // See if oop is null if it is we need no handle
6192
6193 if (src.first()->is_stack()) {
6194
6195 // Oop is already on the stack as an argument
6196 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots();
6197 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots));
6198 if (is_receiver) {
6199 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size;
6200 }
6201
6202 ldr(rscratch1, Address(rfp, reg2offset_in(src.first())));
6203 lea(rHandle, Address(rfp, reg2offset_in(src.first())));
6204 // conditionally move a null
6205 cmp(rscratch1, zr);
6206 csel(rHandle, zr, rHandle, Assembler::EQ);
6207 } else {
6208
6209 // Oop is in an a register we must store it to the space we reserve
6210 // on the stack for oop_handles and pass a handle if oop is non-null
6211
6212 const Register rOop = src.first()->as_Register();
6213 int oop_slot;
6214 if (rOop == j_rarg0)
6215 oop_slot = 0;
6216 else if (rOop == j_rarg1)
6217 oop_slot = 1;
6218 else if (rOop == j_rarg2)
6219 oop_slot = 2;
6220 else if (rOop == j_rarg3)
6221 oop_slot = 3;
6222 else if (rOop == j_rarg4)
6223 oop_slot = 4;
6224 else if (rOop == j_rarg5)
6225 oop_slot = 5;
6226 else if (rOop == j_rarg6)
6227 oop_slot = 6;
6228 else {
6229 assert(rOop == j_rarg7, "wrong register");
6230 oop_slot = 7;
6231 }
6232
6233 oop_slot = oop_slot * VMRegImpl::slots_per_word + oop_handle_offset;
6234 int offset = oop_slot*VMRegImpl::stack_slot_size;
6235
6236 map->set_oop(VMRegImpl::stack2reg(oop_slot));
6237 // Store oop in handle area, may be null
6238 str(rOop, Address(sp, offset));
6239 if (is_receiver) {
6240 *receiver_offset = offset;
6241 }
6242
6243 cmp(rOop, zr);
6244 lea(rHandle, Address(sp, offset));
6245 // conditionally move a null
6246 csel(rHandle, zr, rHandle, Assembler::EQ);
6247 }
6248
6249 // If arg is on the stack then place it otherwise it is already in correct reg.
6250 if (dst.first()->is_stack()) {
6251 str(rHandle, Address(sp, reg2offset_out(dst.first())));
6252 }
6253 }
6254
6255 // A float arg may have to do float reg int reg conversion
6256 void MacroAssembler::float_move(VMRegPair src, VMRegPair dst, Register tmp) {
6257 if (src.first()->is_stack()) {
6258 if (dst.first()->is_stack()) {
6259 ldrw(tmp, Address(rfp, reg2offset_in(src.first())));
6260 strw(tmp, Address(sp, reg2offset_out(dst.first())));
6261 } else {
6262 ldrs(dst.first()->as_FloatRegister(), Address(rfp, reg2offset_in(src.first())));
6263 }
6264 } else if (src.first() != dst.first()) {
6265 if (src.is_single_phys_reg() && dst.is_single_phys_reg())
6266 fmovs(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister());
6267 else
6268 strs(src.first()->as_FloatRegister(), Address(sp, reg2offset_out(dst.first())));
6269 }
6270 }
6271
6272 // A long move
6273 void MacroAssembler::long_move(VMRegPair src, VMRegPair dst, Register tmp) {
6274 if (src.first()->is_stack()) {
6275 if (dst.first()->is_stack()) {
6276 // stack to stack
6277 ldr(tmp, Address(rfp, reg2offset_in(src.first())));
6278 str(tmp, Address(sp, reg2offset_out(dst.first())));
6279 } else {
6280 // stack to reg
6281 ldr(dst.first()->as_Register(), Address(rfp, reg2offset_in(src.first())));
6282 }
6283 } else if (dst.first()->is_stack()) {
6284 // reg to stack
6285 // Do we really have to sign extend???
6286 // __ movslq(src.first()->as_Register(), src.first()->as_Register());
6287 str(src.first()->as_Register(), Address(sp, reg2offset_out(dst.first())));
6288 } else {
6289 if (dst.first() != src.first()) {
6290 mov(dst.first()->as_Register(), src.first()->as_Register());
6291 }
6292 }
6293 }
6294
6295
6296 // A double move
6297 void MacroAssembler::double_move(VMRegPair src, VMRegPair dst, Register tmp) {
6298 if (src.first()->is_stack()) {
6299 if (dst.first()->is_stack()) {
6300 ldr(tmp, Address(rfp, reg2offset_in(src.first())));
6301 str(tmp, Address(sp, reg2offset_out(dst.first())));
6302 } else {
6303 ldrd(dst.first()->as_FloatRegister(), Address(rfp, reg2offset_in(src.first())));
6304 }
6305 } else if (src.first() != dst.first()) {
6306 if (src.is_single_phys_reg() && dst.is_single_phys_reg())
6307 fmovd(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister());
6308 else
6309 strd(src.first()->as_FloatRegister(), Address(sp, reg2offset_out(dst.first())));
6310 }
6311 }
6312
6313 // Implements lightweight-locking.
6314 // Branches to slow upon failure to lock the object, with ZF cleared.
6315 // Falls through upon success with ZF set.
6316 //
6317 // - obj: the object to be locked
6318 // - hdr: the header, already loaded from obj, will be destroyed
6319 // - t1, t2: temporary registers, will be destroyed
6320 void MacroAssembler::lightweight_lock(Register obj, Register hdr, Register t1, Register t2, Label& slow) {
6321 assert(LockingMode == LM_LIGHTWEIGHT, "only used with new lightweight locking");
6322 assert_different_registers(obj, hdr, t1, t2, rscratch1);
6323
6324 // Check if we would have space on lock-stack for the object.
6325 ldrw(t1, Address(rthread, JavaThread::lock_stack_top_offset()));
6326 cmpw(t1, (unsigned)LockStack::end_offset() - 1);
6327 br(Assembler::GT, slow);
6328
6329 // Load (object->mark() | 1) into hdr
6330 orr(hdr, hdr, markWord::unlocked_value);
6331 // Clear lock-bits, into t2
6332 eor(t2, hdr, markWord::unlocked_value);
6333 // Try to swing header from unlocked to locked
6334 // Clobbers rscratch1 when UseLSE is false
6335 cmpxchg(/*addr*/ obj, /*expected*/ hdr, /*new*/ t2, Assembler::xword,
6336 /*acquire*/ true, /*release*/ true, /*weak*/ false, t1);
6337 br(Assembler::NE, slow);
6338
6339 // After successful lock, push object on lock-stack
6340 ldrw(t1, Address(rthread, JavaThread::lock_stack_top_offset()));
6341 str(obj, Address(rthread, t1));
6342 addw(t1, t1, oopSize);
6343 strw(t1, Address(rthread, JavaThread::lock_stack_top_offset()));
6344 }
6345
6346 // Implements lightweight-unlocking.
6347 // Branches to slow upon failure, with ZF cleared.
6348 // Falls through upon success, with ZF set.
6349 //
6350 // - obj: the object to be unlocked
6351 // - hdr: the (pre-loaded) header of the object
6352 // - t1, t2: temporary registers
6353 void MacroAssembler::lightweight_unlock(Register obj, Register hdr, Register t1, Register t2, Label& slow) {
6354 assert(LockingMode == LM_LIGHTWEIGHT, "only used with new lightweight locking");
6355 assert_different_registers(obj, hdr, t1, t2, rscratch1);
6356
6357 #ifdef ASSERT
6358 {
6359 // The following checks rely on the fact that LockStack is only ever modified by
6360 // its owning thread, even if the lock got inflated concurrently; removal of LockStack
6361 // entries after inflation will happen delayed in that case.
6362
6363 // Check for lock-stack underflow.
6364 Label stack_ok;
6365 ldrw(t1, Address(rthread, JavaThread::lock_stack_top_offset()));
6366 cmpw(t1, (unsigned)LockStack::start_offset());
6367 br(Assembler::GT, stack_ok);
6368 STOP("Lock-stack underflow");
6369 bind(stack_ok);
6370 }
6371 {
6372 // Check if the top of the lock-stack matches the unlocked object.
6373 Label tos_ok;
6374 subw(t1, t1, oopSize);
6375 ldr(t1, Address(rthread, t1));
6376 cmpoop(t1, obj);
6377 br(Assembler::EQ, tos_ok);
6378 STOP("Top of lock-stack does not match the unlocked object");
6379 bind(tos_ok);
6380 }
6381 {
6382 // Check that hdr is fast-locked.
6383 Label hdr_ok;
6384 tst(hdr, markWord::lock_mask_in_place);
6385 br(Assembler::EQ, hdr_ok);
6386 STOP("Header is not fast-locked");
6387 bind(hdr_ok);
6388 }
6389 #endif
6390
6391 // Load the new header (unlocked) into t1
6392 orr(t1, hdr, markWord::unlocked_value);
6393
6394 // Try to swing header from locked to unlocked
6395 // Clobbers rscratch1 when UseLSE is false
6396 cmpxchg(obj, hdr, t1, Assembler::xword,
6397 /*acquire*/ true, /*release*/ true, /*weak*/ false, t2);
6398 br(Assembler::NE, slow);
6399
6400 // After successful unlock, pop object from lock-stack
6401 ldrw(t1, Address(rthread, JavaThread::lock_stack_top_offset()));
6402 subw(t1, t1, oopSize);
6403 #ifdef ASSERT
6404 str(zr, Address(rthread, t1));
6405 #endif
6406 strw(t1, Address(rthread, JavaThread::lock_stack_top_offset()));
6407 }