1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2014, 2024, 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 "asm/assembler.hpp"
27 #include "asm/assembler.inline.hpp"
28 #include "ci/ciEnv.hpp"
29 #include "code/compiledIC.hpp"
30 #include "compiler/compileTask.hpp"
31 #include "compiler/disassembler.hpp"
32 #include "compiler/oopMap.hpp"
33 #include "gc/shared/barrierSet.hpp"
34 #include "gc/shared/barrierSetAssembler.hpp"
35 #include "gc/shared/cardTableBarrierSet.hpp"
36 #include "gc/shared/cardTable.hpp"
37 #include "gc/shared/collectedHeap.hpp"
38 #include "gc/shared/tlab_globals.hpp"
39 #include "interpreter/bytecodeHistogram.hpp"
40 #include "interpreter/interpreter.hpp"
41 #include "interpreter/interpreterRuntime.hpp"
42 #include "jvm.h"
43 #include "memory/resourceArea.hpp"
44 #include "memory/universe.hpp"
45 #include "nativeInst_aarch64.hpp"
46 #include "oops/accessDecorators.hpp"
47 #include "oops/compressedKlass.inline.hpp"
48 #include "oops/compressedOops.inline.hpp"
49 #include "oops/klass.inline.hpp"
50 #include "runtime/continuation.hpp"
51 #include "runtime/icache.hpp"
52 #include "runtime/interfaceSupport.inline.hpp"
53 #include "runtime/javaThread.hpp"
54 #include "runtime/jniHandles.inline.hpp"
55 #include "runtime/sharedRuntime.hpp"
56 #include "runtime/stubRoutines.hpp"
57 #include "utilities/globalDefinitions.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 #include <sys/types.h>
70
71 #ifdef PRODUCT
72 #define BLOCK_COMMENT(str) /* nothing */
73 #else
74 #define BLOCK_COMMENT(str) block_comment(str)
75 #endif
76 #define STOP(str) stop(str);
77 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
78
79 #ifdef ASSERT
80 extern "C" void disnm(intptr_t p);
81 #endif
82 // Target-dependent relocation processing
83 //
84 // Instruction sequences whose target may need to be retrieved or
85 // patched are distinguished by their leading instruction, sorting
86 // them into three main instruction groups and related subgroups.
87 //
88 // 1) Branch, Exception and System (insn count = 1)
89 // 1a) Unconditional branch (immediate):
90 // b/bl imm19
91 // 1b) Compare & branch (immediate):
92 // cbz/cbnz Rt imm19
93 // 1c) Test & branch (immediate):
94 // tbz/tbnz Rt imm14
95 // 1d) Conditional branch (immediate):
96 // b.cond imm19
97 //
98 // 2) Loads and Stores (insn count = 1)
99 // 2a) Load register literal:
100 // ldr Rt imm19
101 //
102 // 3) Data Processing Immediate (insn count = 2 or 3)
103 // 3a) PC-rel. addressing
104 // adr/adrp Rx imm21; ldr/str Ry Rx #imm12
105 // adr/adrp Rx imm21; add Ry Rx #imm12
106 // adr/adrp Rx imm21; movk Rx #imm16<<32; ldr/str Ry, [Rx, #offset_in_page]
107 // adr/adrp Rx imm21
108 // adr/adrp Rx imm21; movk Rx #imm16<<32
109 // adr/adrp Rx imm21; movk Rx #imm16<<32; add Ry, Rx, #offset_in_page
110 // The latter form can only happen when the target is an
111 // ExternalAddress, and (by definition) ExternalAddresses don't
112 // move. Because of that property, there is never any need to
113 // patch the last of the three instructions. However,
114 // MacroAssembler::target_addr_for_insn takes all three
115 // instructions into account and returns the correct address.
116 // 3b) Move wide (immediate)
117 // movz Rx #imm16; movk Rx #imm16 << 16; movk Rx #imm16 << 32;
118 //
119 // A switch on a subset of the instruction's bits provides an
120 // efficient dispatch to these subcases.
121 //
122 // insn[28:26] -> main group ('x' == don't care)
123 // 00x -> UNALLOCATED
124 // 100 -> Data Processing Immediate
125 // 101 -> Branch, Exception and System
126 // x1x -> Loads and Stores
127 //
128 // insn[30:25] -> subgroup ('_' == group, 'x' == don't care).
129 // n.b. in some cases extra bits need to be checked to verify the
130 // instruction is as expected
131 //
132 // 1) ... xx101x Branch, Exception and System
133 // 1a) 00___x Unconditional branch (immediate)
134 // 1b) 01___0 Compare & branch (immediate)
135 // 1c) 01___1 Test & branch (immediate)
136 // 1d) 10___0 Conditional branch (immediate)
137 // other Should not happen
138 //
139 // 2) ... xxx1x0 Loads and Stores
140 // 2a) xx1__00 Load/Store register (insn[28] == 1 && insn[24] == 0)
141 // 2aa) x01__00 Load register literal (i.e. requires insn[29] == 0)
142 // strictly should be 64 bit non-FP/SIMD i.e.
143 // 0101_000 (i.e. requires insn[31:24] == 01011000)
144 //
145 // 3) ... xx100x Data Processing Immediate
146 // 3a) xx___00 PC-rel. addressing (n.b. requires insn[24] == 0)
147 // 3b) xx___101 Move wide (immediate) (n.b. requires insn[24:23] == 01)
148 // strictly should be 64 bit movz #imm16<<0
149 // 110___10100 (i.e. requires insn[31:21] == 11010010100)
150 //
151 class RelocActions {
152 protected:
153 typedef int (*reloc_insn)(address insn_addr, address &target);
154
155 virtual reloc_insn adrpMem() = 0;
156 virtual reloc_insn adrpAdd() = 0;
157 virtual reloc_insn adrpMovk() = 0;
158
159 const address _insn_addr;
160 const uint32_t _insn;
161
162 static uint32_t insn_at(address insn_addr, int n) {
163 return ((uint32_t*)insn_addr)[n];
164 }
165 uint32_t insn_at(int n) const {
166 return insn_at(_insn_addr, n);
167 }
168
169 public:
170
171 RelocActions(address insn_addr) : _insn_addr(insn_addr), _insn(insn_at(insn_addr, 0)) {}
172 RelocActions(address insn_addr, uint32_t insn)
173 : _insn_addr(insn_addr), _insn(insn) {}
174
175 virtual int unconditionalBranch(address insn_addr, address &target) = 0;
176 virtual int conditionalBranch(address insn_addr, address &target) = 0;
177 virtual int testAndBranch(address insn_addr, address &target) = 0;
178 virtual int loadStore(address insn_addr, address &target) = 0;
179 virtual int adr(address insn_addr, address &target) = 0;
180 virtual int adrp(address insn_addr, address &target, reloc_insn inner) = 0;
181 virtual int immediate(address insn_addr, address &target) = 0;
182 virtual void verify(address insn_addr, address &target) = 0;
183
184 int ALWAYSINLINE run(address insn_addr, address &target) {
185 int instructions = 1;
186
187 uint32_t dispatch = Instruction_aarch64::extract(_insn, 30, 25);
188 switch(dispatch) {
189 case 0b001010:
190 case 0b001011: {
191 instructions = unconditionalBranch(insn_addr, target);
192 break;
193 }
194 case 0b101010: // Conditional branch (immediate)
195 case 0b011010: { // Compare & branch (immediate)
196 instructions = conditionalBranch(insn_addr, target);
197 break;
198 }
199 case 0b011011: {
200 instructions = testAndBranch(insn_addr, target);
201 break;
202 }
203 case 0b001100:
204 case 0b001110:
205 case 0b011100:
206 case 0b011110:
207 case 0b101100:
208 case 0b101110:
209 case 0b111100:
210 case 0b111110: {
211 // load/store
212 if ((Instruction_aarch64::extract(_insn, 29, 24) & 0b111011) == 0b011000) {
213 // Load register (literal)
214 instructions = loadStore(insn_addr, target);
215 break;
216 } else {
217 // nothing to do
218 assert(target == nullptr, "did not expect to relocate target for polling page load");
219 }
220 break;
221 }
222 case 0b001000:
223 case 0b011000:
224 case 0b101000:
225 case 0b111000: {
226 // adr/adrp
227 assert(Instruction_aarch64::extract(_insn, 28, 24) == 0b10000, "must be");
228 int shift = Instruction_aarch64::extract(_insn, 31, 31);
229 if (shift) {
230 uint32_t insn2 = insn_at(1);
231 if (Instruction_aarch64::extract(insn2, 29, 24) == 0b111001 &&
232 Instruction_aarch64::extract(_insn, 4, 0) ==
233 Instruction_aarch64::extract(insn2, 9, 5)) {
234 instructions = adrp(insn_addr, target, adrpMem());
235 } else if (Instruction_aarch64::extract(insn2, 31, 22) == 0b1001000100 &&
236 Instruction_aarch64::extract(_insn, 4, 0) ==
237 Instruction_aarch64::extract(insn2, 4, 0)) {
238 instructions = adrp(insn_addr, target, adrpAdd());
239 } else if (Instruction_aarch64::extract(insn2, 31, 21) == 0b11110010110 &&
240 Instruction_aarch64::extract(_insn, 4, 0) ==
241 Instruction_aarch64::extract(insn2, 4, 0)) {
242 instructions = adrp(insn_addr, target, adrpMovk());
243 } else {
244 ShouldNotReachHere();
245 }
246 } else {
247 instructions = adr(insn_addr, target);
248 }
249 break;
250 }
251 case 0b001001:
252 case 0b011001:
253 case 0b101001:
254 case 0b111001: {
255 instructions = immediate(insn_addr, target);
256 break;
257 }
258 default: {
259 ShouldNotReachHere();
260 }
261 }
262
263 verify(insn_addr, target);
264 return instructions * NativeInstruction::instruction_size;
265 }
266 };
267
268 class Patcher : public RelocActions {
269 virtual reloc_insn adrpMem() { return &Patcher::adrpMem_impl; }
270 virtual reloc_insn adrpAdd() { return &Patcher::adrpAdd_impl; }
271 virtual reloc_insn adrpMovk() { return &Patcher::adrpMovk_impl; }
272
273 public:
274 Patcher(address insn_addr) : RelocActions(insn_addr) {}
275
276 virtual int unconditionalBranch(address insn_addr, address &target) {
277 intptr_t offset = (target - insn_addr) >> 2;
278 Instruction_aarch64::spatch(insn_addr, 25, 0, offset);
279 return 1;
280 }
281 virtual int conditionalBranch(address insn_addr, address &target) {
282 intptr_t offset = (target - insn_addr) >> 2;
283 Instruction_aarch64::spatch(insn_addr, 23, 5, offset);
284 return 1;
285 }
286 virtual int testAndBranch(address insn_addr, address &target) {
287 intptr_t offset = (target - insn_addr) >> 2;
288 Instruction_aarch64::spatch(insn_addr, 18, 5, offset);
289 return 1;
290 }
291 virtual int loadStore(address insn_addr, address &target) {
292 intptr_t offset = (target - insn_addr) >> 2;
293 Instruction_aarch64::spatch(insn_addr, 23, 5, offset);
294 return 1;
295 }
296 virtual int adr(address insn_addr, address &target) {
297 #ifdef ASSERT
298 assert(Instruction_aarch64::extract(_insn, 28, 24) == 0b10000, "must be");
299 #endif
300 // PC-rel. addressing
301 ptrdiff_t offset = target - insn_addr;
302 int offset_lo = offset & 3;
303 offset >>= 2;
304 Instruction_aarch64::spatch(insn_addr, 23, 5, offset);
305 Instruction_aarch64::patch(insn_addr, 30, 29, offset_lo);
306 return 1;
307 }
308 virtual int adrp(address insn_addr, address &target, reloc_insn inner) {
309 int instructions = 1;
310 #ifdef ASSERT
311 assert(Instruction_aarch64::extract(_insn, 28, 24) == 0b10000, "must be");
312 #endif
313 ptrdiff_t offset = target - insn_addr;
314 instructions = 2;
315 precond(inner != nullptr);
316 // Give the inner reloc a chance to modify the target.
317 address adjusted_target = target;
318 instructions = (*inner)(insn_addr, adjusted_target);
319 uintptr_t pc_page = (uintptr_t)insn_addr >> 12;
320 uintptr_t adr_page = (uintptr_t)adjusted_target >> 12;
321 offset = adr_page - pc_page;
322 int offset_lo = offset & 3;
323 offset >>= 2;
324 Instruction_aarch64::spatch(insn_addr, 23, 5, offset);
325 Instruction_aarch64::patch(insn_addr, 30, 29, offset_lo);
326 return instructions;
327 }
328 static int adrpMem_impl(address insn_addr, address &target) {
329 uintptr_t dest = (uintptr_t)target;
330 int offset_lo = dest & 0xfff;
331 uint32_t insn2 = insn_at(insn_addr, 1);
332 uint32_t size = Instruction_aarch64::extract(insn2, 31, 30);
333 Instruction_aarch64::patch(insn_addr + sizeof (uint32_t), 21, 10, offset_lo >> size);
334 guarantee(((dest >> size) << size) == dest, "misaligned target");
335 return 2;
336 }
337 static int adrpAdd_impl(address insn_addr, address &target) {
338 uintptr_t dest = (uintptr_t)target;
339 int offset_lo = dest & 0xfff;
340 Instruction_aarch64::patch(insn_addr + sizeof (uint32_t), 21, 10, offset_lo);
341 return 2;
342 }
343 static int adrpMovk_impl(address insn_addr, address &target) {
344 uintptr_t dest = uintptr_t(target);
345 Instruction_aarch64::patch(insn_addr + sizeof (uint32_t), 20, 5, (uintptr_t)target >> 32);
346 dest = (dest & 0xffffffffULL) | (uintptr_t(insn_addr) & 0xffff00000000ULL);
347 target = address(dest);
348 return 2;
349 }
350 virtual int immediate(address insn_addr, address &target) {
351 assert(Instruction_aarch64::extract(_insn, 31, 21) == 0b11010010100, "must be");
352 uint64_t dest = (uint64_t)target;
353 // Move wide constant
354 assert(nativeInstruction_at(insn_addr+4)->is_movk(), "wrong insns in patch");
355 assert(nativeInstruction_at(insn_addr+8)->is_movk(), "wrong insns in patch");
356 Instruction_aarch64::patch(insn_addr, 20, 5, dest & 0xffff);
357 Instruction_aarch64::patch(insn_addr+4, 20, 5, (dest >>= 16) & 0xffff);
358 Instruction_aarch64::patch(insn_addr+8, 20, 5, (dest >>= 16) & 0xffff);
359 return 3;
360 }
361 virtual void verify(address insn_addr, address &target) {
362 #ifdef ASSERT
363 address address_is = MacroAssembler::target_addr_for_insn(insn_addr);
364 if (!(address_is == target)) {
365 tty->print_cr("%p at %p should be %p", address_is, insn_addr, target);
366 disnm((intptr_t)insn_addr);
367 assert(address_is == target, "should be");
368 }
369 #endif
370 }
371 };
372
373 // If insn1 and insn2 use the same register to form an address, either
374 // by an offsetted LDR or a simple ADD, return the offset. If the
375 // second instruction is an LDR, the offset may be scaled.
376 static bool offset_for(uint32_t insn1, uint32_t insn2, ptrdiff_t &byte_offset) {
377 if (Instruction_aarch64::extract(insn2, 29, 24) == 0b111001 &&
378 Instruction_aarch64::extract(insn1, 4, 0) ==
379 Instruction_aarch64::extract(insn2, 9, 5)) {
380 // Load/store register (unsigned immediate)
381 byte_offset = Instruction_aarch64::extract(insn2, 21, 10);
382 uint32_t size = Instruction_aarch64::extract(insn2, 31, 30);
383 byte_offset <<= size;
384 return true;
385 } else if (Instruction_aarch64::extract(insn2, 31, 22) == 0b1001000100 &&
386 Instruction_aarch64::extract(insn1, 4, 0) ==
387 Instruction_aarch64::extract(insn2, 4, 0)) {
388 // add (immediate)
389 byte_offset = Instruction_aarch64::extract(insn2, 21, 10);
390 return true;
391 }
392 return false;
393 }
394
395 class AArch64Decoder : public RelocActions {
396 virtual reloc_insn adrpMem() { return &AArch64Decoder::adrpMem_impl; }
397 virtual reloc_insn adrpAdd() { return &AArch64Decoder::adrpAdd_impl; }
398 virtual reloc_insn adrpMovk() { return &AArch64Decoder::adrpMovk_impl; }
399
400 public:
401 AArch64Decoder(address insn_addr, uint32_t insn) : RelocActions(insn_addr, insn) {}
402
403 virtual int loadStore(address insn_addr, address &target) {
404 intptr_t offset = Instruction_aarch64::sextract(_insn, 23, 5);
405 target = insn_addr + (offset << 2);
406 return 1;
407 }
408 virtual int unconditionalBranch(address insn_addr, address &target) {
409 intptr_t offset = Instruction_aarch64::sextract(_insn, 25, 0);
410 target = insn_addr + (offset << 2);
411 return 1;
412 }
413 virtual int conditionalBranch(address insn_addr, address &target) {
414 intptr_t offset = Instruction_aarch64::sextract(_insn, 23, 5);
415 target = address(((uint64_t)insn_addr + (offset << 2)));
416 return 1;
417 }
418 virtual int testAndBranch(address insn_addr, address &target) {
419 intptr_t offset = Instruction_aarch64::sextract(_insn, 18, 5);
420 target = address(((uint64_t)insn_addr + (offset << 2)));
421 return 1;
422 }
423 virtual int adr(address insn_addr, address &target) {
424 // PC-rel. addressing
425 intptr_t offset = Instruction_aarch64::extract(_insn, 30, 29);
426 offset |= Instruction_aarch64::sextract(_insn, 23, 5) << 2;
427 target = address((uint64_t)insn_addr + offset);
428 return 1;
429 }
430 virtual int adrp(address insn_addr, address &target, reloc_insn inner) {
431 assert(Instruction_aarch64::extract(_insn, 28, 24) == 0b10000, "must be");
432 intptr_t offset = Instruction_aarch64::extract(_insn, 30, 29);
433 offset |= Instruction_aarch64::sextract(_insn, 23, 5) << 2;
434 int shift = 12;
435 offset <<= shift;
436 uint64_t target_page = ((uint64_t)insn_addr) + offset;
437 target_page &= ((uint64_t)-1) << shift;
438 uint32_t insn2 = insn_at(1);
439 target = address(target_page);
440 precond(inner != nullptr);
441 (*inner)(insn_addr, target);
442 return 2;
443 }
444 static int adrpMem_impl(address insn_addr, address &target) {
445 uint32_t insn2 = insn_at(insn_addr, 1);
446 // Load/store register (unsigned immediate)
447 ptrdiff_t byte_offset = Instruction_aarch64::extract(insn2, 21, 10);
448 uint32_t size = Instruction_aarch64::extract(insn2, 31, 30);
449 byte_offset <<= size;
450 target += byte_offset;
451 return 2;
452 }
453 static int adrpAdd_impl(address insn_addr, address &target) {
454 uint32_t insn2 = insn_at(insn_addr, 1);
455 // add (immediate)
456 ptrdiff_t byte_offset = Instruction_aarch64::extract(insn2, 21, 10);
457 target += byte_offset;
458 return 2;
459 }
460 static int adrpMovk_impl(address insn_addr, address &target) {
461 uint32_t insn2 = insn_at(insn_addr, 1);
462 uint64_t dest = uint64_t(target);
463 dest = (dest & 0xffff0000ffffffff) |
464 ((uint64_t)Instruction_aarch64::extract(insn2, 20, 5) << 32);
465 target = address(dest);
466
467 // We know the destination 4k page. Maybe we have a third
468 // instruction.
469 uint32_t insn = insn_at(insn_addr, 0);
470 uint32_t insn3 = insn_at(insn_addr, 2);
471 ptrdiff_t byte_offset;
472 if (offset_for(insn, insn3, byte_offset)) {
473 target += byte_offset;
474 return 3;
475 } else {
476 return 2;
477 }
478 }
479 virtual int immediate(address insn_addr, address &target) {
480 uint32_t *insns = (uint32_t *)insn_addr;
481 assert(Instruction_aarch64::extract(_insn, 31, 21) == 0b11010010100, "must be");
482 // Move wide constant: movz, movk, movk. See movptr().
483 assert(nativeInstruction_at(insns+1)->is_movk(), "wrong insns in patch");
484 assert(nativeInstruction_at(insns+2)->is_movk(), "wrong insns in patch");
485 target = address(uint64_t(Instruction_aarch64::extract(_insn, 20, 5))
486 + (uint64_t(Instruction_aarch64::extract(insns[1], 20, 5)) << 16)
487 + (uint64_t(Instruction_aarch64::extract(insns[2], 20, 5)) << 32));
488 assert(nativeInstruction_at(insn_addr+4)->is_movk(), "wrong insns in patch");
489 assert(nativeInstruction_at(insn_addr+8)->is_movk(), "wrong insns in patch");
490 return 3;
491 }
492 virtual void verify(address insn_addr, address &target) {
493 }
494 };
495
496 address MacroAssembler::target_addr_for_insn(address insn_addr, uint32_t insn) {
497 AArch64Decoder decoder(insn_addr, insn);
498 address target;
499 decoder.run(insn_addr, target);
500 return target;
501 }
502
503 // Patch any kind of instruction; there may be several instructions.
504 // Return the total length (in bytes) of the instructions.
505 int MacroAssembler::pd_patch_instruction_size(address insn_addr, address target) {
506 Patcher patcher(insn_addr);
507 return patcher.run(insn_addr, target);
508 }
509
510 int MacroAssembler::patch_oop(address insn_addr, address o) {
511 int instructions;
512 unsigned insn = *(unsigned*)insn_addr;
513 assert(nativeInstruction_at(insn_addr+4)->is_movk(), "wrong insns in patch");
514
515 // OOPs are either narrow (32 bits) or wide (48 bits). We encode
516 // narrow OOPs by setting the upper 16 bits in the first
517 // instruction.
518 if (Instruction_aarch64::extract(insn, 31, 21) == 0b11010010101) {
519 // Move narrow OOP
520 uint32_t n = CompressedOops::narrow_oop_value(cast_to_oop(o));
521 Instruction_aarch64::patch(insn_addr, 20, 5, n >> 16);
522 Instruction_aarch64::patch(insn_addr+4, 20, 5, n & 0xffff);
523 instructions = 2;
524 } else {
525 // Move wide OOP
526 assert(nativeInstruction_at(insn_addr+8)->is_movk(), "wrong insns in patch");
527 uintptr_t dest = (uintptr_t)o;
528 Instruction_aarch64::patch(insn_addr, 20, 5, dest & 0xffff);
529 Instruction_aarch64::patch(insn_addr+4, 20, 5, (dest >>= 16) & 0xffff);
530 Instruction_aarch64::patch(insn_addr+8, 20, 5, (dest >>= 16) & 0xffff);
531 instructions = 3;
532 }
533 return instructions * NativeInstruction::instruction_size;
534 }
535
536 int MacroAssembler::patch_narrow_klass(address insn_addr, narrowKlass n) {
537 // Metadata pointers are either narrow (32 bits) or wide (48 bits).
538 // We encode narrow ones by setting the upper 16 bits in the first
539 // instruction.
540 NativeInstruction *insn = nativeInstruction_at(insn_addr);
541 assert(Instruction_aarch64::extract(insn->encoding(), 31, 21) == 0b11010010101 &&
542 nativeInstruction_at(insn_addr+4)->is_movk(), "wrong insns in patch");
543
544 Instruction_aarch64::patch(insn_addr, 20, 5, n >> 16);
545 Instruction_aarch64::patch(insn_addr+4, 20, 5, n & 0xffff);
546 return 2 * NativeInstruction::instruction_size;
547 }
548
549 address MacroAssembler::target_addr_for_insn_or_null(address insn_addr, unsigned insn) {
550 if (NativeInstruction::is_ldrw_to_zr(address(&insn))) {
551 return nullptr;
552 }
553 return MacroAssembler::target_addr_for_insn(insn_addr, insn);
554 }
555
556 void MacroAssembler::safepoint_poll(Label& slow_path, bool at_return, bool acquire, bool in_nmethod, Register tmp) {
557 if (acquire) {
558 lea(tmp, Address(rthread, JavaThread::polling_word_offset()));
559 ldar(tmp, tmp);
560 } else {
561 ldr(tmp, Address(rthread, JavaThread::polling_word_offset()));
562 }
563 if (at_return) {
564 // Note that when in_nmethod is set, the stack pointer is incremented before the poll. Therefore,
565 // we may safely use the sp instead to perform the stack watermark check.
566 cmp(in_nmethod ? sp : rfp, tmp);
567 br(Assembler::HI, slow_path);
568 } else {
569 tbnz(tmp, log2i_exact(SafepointMechanism::poll_bit()), slow_path);
570 }
571 }
572
573 void MacroAssembler::rt_call(address dest, Register tmp) {
574 CodeBlob *cb = CodeCache::find_blob(dest);
575 if (cb) {
576 far_call(RuntimeAddress(dest));
577 } else {
578 lea(tmp, RuntimeAddress(dest));
579 blr(tmp);
580 }
581 }
582
583 void MacroAssembler::push_cont_fastpath(Register java_thread) {
584 if (!Continuations::enabled()) return;
585 Label done;
586 ldr(rscratch1, Address(java_thread, JavaThread::cont_fastpath_offset()));
587 cmp(sp, rscratch1);
588 br(Assembler::LS, done);
589 mov(rscratch1, sp); // we can't use sp as the source in str
590 str(rscratch1, Address(java_thread, JavaThread::cont_fastpath_offset()));
591 bind(done);
592 }
593
594 void MacroAssembler::pop_cont_fastpath(Register java_thread) {
595 if (!Continuations::enabled()) return;
596 Label done;
597 ldr(rscratch1, Address(java_thread, JavaThread::cont_fastpath_offset()));
598 cmp(sp, rscratch1);
599 br(Assembler::LO, done);
600 str(zr, Address(java_thread, JavaThread::cont_fastpath_offset()));
601 bind(done);
602 }
603
604 void MacroAssembler::reset_last_Java_frame(bool clear_fp) {
605 // we must set sp to zero to clear frame
606 str(zr, Address(rthread, JavaThread::last_Java_sp_offset()));
607
608 // must clear fp, so that compiled frames are not confused; it is
609 // possible that we need it only for debugging
610 if (clear_fp) {
611 str(zr, Address(rthread, JavaThread::last_Java_fp_offset()));
612 }
613
614 // Always clear the pc because it could have been set by make_walkable()
615 str(zr, Address(rthread, JavaThread::last_Java_pc_offset()));
616 }
617
618 // Calls to C land
619 //
620 // When entering C land, the rfp, & resp of the last Java frame have to be recorded
621 // in the (thread-local) JavaThread object. When leaving C land, the last Java fp
622 // has to be reset to 0. This is required to allow proper stack traversal.
623 void MacroAssembler::set_last_Java_frame(Register last_java_sp,
624 Register last_java_fp,
625 Register last_java_pc,
626 Register scratch) {
627
628 if (last_java_pc->is_valid()) {
629 str(last_java_pc, Address(rthread,
630 JavaThread::frame_anchor_offset()
631 + JavaFrameAnchor::last_Java_pc_offset()));
632 }
633
634 // determine last_java_sp register
635 if (last_java_sp == sp) {
636 mov(scratch, sp);
637 last_java_sp = scratch;
638 } else if (!last_java_sp->is_valid()) {
639 last_java_sp = esp;
640 }
641
642 str(last_java_sp, Address(rthread, JavaThread::last_Java_sp_offset()));
643
644 // last_java_fp is optional
645 if (last_java_fp->is_valid()) {
646 str(last_java_fp, Address(rthread, JavaThread::last_Java_fp_offset()));
647 }
648 }
649
650 void MacroAssembler::set_last_Java_frame(Register last_java_sp,
651 Register last_java_fp,
652 address last_java_pc,
653 Register scratch) {
654 assert(last_java_pc != nullptr, "must provide a valid PC");
655
656 adr(scratch, last_java_pc);
657 str(scratch, Address(rthread,
658 JavaThread::frame_anchor_offset()
659 + JavaFrameAnchor::last_Java_pc_offset()));
660
661 set_last_Java_frame(last_java_sp, last_java_fp, noreg, scratch);
662 }
663
664 void MacroAssembler::set_last_Java_frame(Register last_java_sp,
665 Register last_java_fp,
666 Label &L,
667 Register scratch) {
668 if (L.is_bound()) {
669 set_last_Java_frame(last_java_sp, last_java_fp, target(L), scratch);
670 } else {
671 InstructionMark im(this);
672 L.add_patch_at(code(), locator());
673 set_last_Java_frame(last_java_sp, last_java_fp, pc() /* Patched later */, scratch);
674 }
675 }
676
677 static inline bool target_needs_far_branch(address addr) {
678 if (AOTCodeCache::is_on_for_dump()) {
679 return true;
680 }
681 // codecache size <= 128M
682 if (!MacroAssembler::far_branches()) {
683 return false;
684 }
685 // codecache size > 240M
686 if (MacroAssembler::codestub_branch_needs_far_jump()) {
687 return true;
688 }
689 // codecache size: 128M..240M
690 return !CodeCache::is_non_nmethod(addr);
691 }
692
693 void MacroAssembler::far_call(Address entry, Register tmp) {
694 assert(ReservedCodeCacheSize < 4*G, "branch out of range");
695 assert(CodeCache::find_blob(entry.target()) != nullptr,
696 "destination of far call not found in code cache");
697 assert(entry.rspec().type() == relocInfo::external_word_type
698 || entry.rspec().type() == relocInfo::runtime_call_type
699 || entry.rspec().type() == relocInfo::none, "wrong entry relocInfo type");
700 if (target_needs_far_branch(entry.target())) {
701 uint64_t offset;
702 // We can use ADRP here because we know that the total size of
703 // the code cache cannot exceed 2Gb (ADRP limit is 4GB).
704 adrp(tmp, entry, offset);
705 add(tmp, tmp, offset);
706 blr(tmp);
707 } else {
708 bl(entry);
709 }
710 }
711
712 int MacroAssembler::far_jump(Address entry, Register tmp) {
713 assert(ReservedCodeCacheSize < 4*G, "branch out of range");
714 assert(CodeCache::find_blob(entry.target()) != nullptr,
715 "destination of far call not found in code cache");
716 assert(entry.rspec().type() == relocInfo::external_word_type
717 || entry.rspec().type() == relocInfo::runtime_call_type
718 || entry.rspec().type() == relocInfo::none, "wrong entry relocInfo type");
719 address start = pc();
720 if (target_needs_far_branch(entry.target())) {
721 uint64_t offset;
722 // We can use ADRP here because we know that the total size of
723 // the code cache cannot exceed 2Gb (ADRP limit is 4GB).
724 adrp(tmp, entry, offset);
725 add(tmp, tmp, offset);
726 br(tmp);
727 } else {
728 b(entry);
729 }
730 return pc() - start;
731 }
732
733 void MacroAssembler::reserved_stack_check() {
734 // testing if reserved zone needs to be enabled
735 Label no_reserved_zone_enabling;
736
737 ldr(rscratch1, Address(rthread, JavaThread::reserved_stack_activation_offset()));
738 cmp(sp, rscratch1);
739 br(Assembler::LO, no_reserved_zone_enabling);
740
741 enter(); // LR and FP are live.
742 lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::enable_stack_reserved_zone)));
743 mov(c_rarg0, rthread);
744 blr(rscratch1);
745 leave();
746
747 // We have already removed our own frame.
748 // throw_delayed_StackOverflowError will think that it's been
749 // called by our caller.
750 lea(rscratch1, RuntimeAddress(SharedRuntime::throw_delayed_StackOverflowError_entry()));
751 br(rscratch1);
752 should_not_reach_here();
753
754 bind(no_reserved_zone_enabling);
755 }
756
757 static void pass_arg0(MacroAssembler* masm, Register arg) {
758 if (c_rarg0 != arg ) {
759 masm->mov(c_rarg0, arg);
760 }
761 }
762
763 static void pass_arg1(MacroAssembler* masm, Register arg) {
764 if (c_rarg1 != arg ) {
765 masm->mov(c_rarg1, arg);
766 }
767 }
768
769 static void pass_arg2(MacroAssembler* masm, Register arg) {
770 if (c_rarg2 != arg ) {
771 masm->mov(c_rarg2, arg);
772 }
773 }
774
775 static void pass_arg3(MacroAssembler* masm, Register arg) {
776 if (c_rarg3 != arg ) {
777 masm->mov(c_rarg3, arg);
778 }
779 }
780
781 static bool is_preemptable(address entry_point) {
782 return entry_point == CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter);
783 }
784
785 void MacroAssembler::call_VM_base(Register oop_result,
786 Register java_thread,
787 Register last_java_sp,
788 address entry_point,
789 int number_of_arguments,
790 bool check_exceptions) {
791 // determine java_thread register
792 if (!java_thread->is_valid()) {
793 java_thread = rthread;
794 }
795
796 // determine last_java_sp register
797 if (!last_java_sp->is_valid()) {
798 last_java_sp = esp;
799 }
800
801 // debugging support
802 assert(number_of_arguments >= 0 , "cannot have negative number of arguments");
803 assert(java_thread == rthread, "unexpected register");
804 #ifdef ASSERT
805 // TraceBytecodes does not use r12 but saves it over the call, so don't verify
806 // if ((UseCompressedOops || UseCompressedClassPointers) && !TraceBytecodes) verify_heapbase("call_VM_base: heap base corrupted?");
807 #endif // ASSERT
808
809 assert(java_thread != oop_result , "cannot use the same register for java_thread & oop_result");
810 assert(java_thread != last_java_sp, "cannot use the same register for java_thread & last_java_sp");
811
812 // push java thread (becomes first argument of C function)
813
814 mov(c_rarg0, java_thread);
815
816 // set last Java frame before call
817 assert(last_java_sp != rfp, "can't use rfp");
818
819 Label l;
820 if (is_preemptable(entry_point)) {
821 // skip setting last_pc since we already set it to desired value.
822 set_last_Java_frame(last_java_sp, rfp, noreg, rscratch1);
823 } else {
824 set_last_Java_frame(last_java_sp, rfp, l, rscratch1);
825 }
826
827 // do the call, remove parameters
828 MacroAssembler::call_VM_leaf_base(entry_point, number_of_arguments, &l);
829
830 // lr could be poisoned with PAC signature during throw_pending_exception
831 // if it was tail-call optimized by compiler, since lr is not callee-saved
832 // reload it with proper value
833 adr(lr, l);
834
835 // reset last Java frame
836 // Only interpreter should have to clear fp
837 reset_last_Java_frame(true);
838
839 // C++ interp handles this in the interpreter
840 check_and_handle_popframe(java_thread);
841 check_and_handle_earlyret(java_thread);
842
843 if (check_exceptions) {
844 // check for pending exceptions (java_thread is set upon return)
845 ldr(rscratch1, Address(java_thread, in_bytes(Thread::pending_exception_offset())));
846 Label ok;
847 cbz(rscratch1, ok);
848 lea(rscratch1, RuntimeAddress(StubRoutines::forward_exception_entry()));
849 br(rscratch1);
850 bind(ok);
851 }
852
853 // get oop result if there is one and reset the value in the thread
854 if (oop_result->is_valid()) {
855 get_vm_result_oop(oop_result, java_thread);
856 }
857 }
858
859 void MacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) {
860 call_VM_base(oop_result, noreg, noreg, entry_point, number_of_arguments, check_exceptions);
861 }
862
863 // Check the entry target is always reachable from any branch.
864 static bool is_always_within_branch_range(Address entry) {
865 if (AOTCodeCache::is_on_for_dump()) {
866 return false;
867 }
868 const address target = entry.target();
869
870 if (!CodeCache::contains(target)) {
871 // We always use trampolines for callees outside CodeCache.
872 assert(entry.rspec().type() == relocInfo::runtime_call_type, "non-runtime call of an external target");
873 return false;
874 }
875
876 if (!MacroAssembler::far_branches()) {
877 return true;
878 }
879
880 if (entry.rspec().type() == relocInfo::runtime_call_type) {
881 // Runtime calls are calls of a non-compiled method (stubs, adapters).
882 // Non-compiled methods stay forever in CodeCache.
883 // We check whether the longest possible branch is within the branch range.
884 assert(CodeCache::find_blob(target) != nullptr &&
885 !CodeCache::find_blob(target)->is_nmethod(),
886 "runtime call of compiled method");
887 const address right_longest_branch_start = CodeCache::high_bound() - NativeInstruction::instruction_size;
888 const address left_longest_branch_start = CodeCache::low_bound();
889 const bool is_reachable = Assembler::reachable_from_branch_at(left_longest_branch_start, target) &&
890 Assembler::reachable_from_branch_at(right_longest_branch_start, target);
891 return is_reachable;
892 }
893
894 return false;
895 }
896
897 // Maybe emit a call via a trampoline. If the code cache is small
898 // trampolines won't be emitted.
899 address MacroAssembler::trampoline_call(Address entry) {
900 assert(entry.rspec().type() == relocInfo::runtime_call_type
901 || entry.rspec().type() == relocInfo::opt_virtual_call_type
902 || entry.rspec().type() == relocInfo::static_call_type
903 || entry.rspec().type() == relocInfo::virtual_call_type, "wrong reloc type");
904
905 address target = entry.target();
906
907 if (!is_always_within_branch_range(entry)) {
908 if (!in_scratch_emit_size()) {
909 // We don't want to emit a trampoline if C2 is generating dummy
910 // code during its branch shortening phase.
911 if (entry.rspec().type() == relocInfo::runtime_call_type) {
912 assert(CodeBuffer::supports_shared_stubs(), "must support shared stubs");
913 code()->share_trampoline_for(entry.target(), offset());
914 } else {
915 address stub = emit_trampoline_stub(offset(), target);
916 if (stub == nullptr) {
917 postcond(pc() == badAddress);
918 return nullptr; // CodeCache is full
919 }
920 }
921 }
922 target = pc();
923 }
924
925 address call_pc = pc();
926 relocate(entry.rspec());
927 bl(target);
928
929 postcond(pc() != badAddress);
930 return call_pc;
931 }
932
933 // Emit a trampoline stub for a call to a target which is too far away.
934 //
935 // code sequences:
936 //
937 // call-site:
938 // branch-and-link to <destination> or <trampoline stub>
939 //
940 // Related trampoline stub for this call site in the stub section:
941 // load the call target from the constant pool
942 // branch (LR still points to the call site above)
943
944 address MacroAssembler::emit_trampoline_stub(int insts_call_instruction_offset,
945 address dest) {
946 // Max stub size: alignment nop, TrampolineStub.
947 address stub = start_a_stub(max_trampoline_stub_size());
948 if (stub == nullptr) {
949 return nullptr; // CodeBuffer::expand failed
950 }
951
952 // Create a trampoline stub relocation which relates this trampoline stub
953 // with the call instruction at insts_call_instruction_offset in the
954 // instructions code-section.
955 align(wordSize);
956 relocate(trampoline_stub_Relocation::spec(code()->insts()->start()
957 + insts_call_instruction_offset));
958 const int stub_start_offset = offset();
959
960 // Now, create the trampoline stub's code:
961 // - load the call
962 // - call
963 Label target;
964 ldr(rscratch1, target);
965 br(rscratch1);
966 bind(target);
967 assert(offset() - stub_start_offset == NativeCallTrampolineStub::data_offset,
968 "should be");
969 emit_int64((int64_t)dest);
970
971 const address stub_start_addr = addr_at(stub_start_offset);
972
973 assert(is_NativeCallTrampolineStub_at(stub_start_addr), "doesn't look like a trampoline");
974
975 end_a_stub();
976 return stub_start_addr;
977 }
978
979 int MacroAssembler::max_trampoline_stub_size() {
980 // Max stub size: alignment nop, TrampolineStub.
981 return NativeInstruction::instruction_size + NativeCallTrampolineStub::instruction_size;
982 }
983
984 void MacroAssembler::emit_static_call_stub() {
985 // CompiledDirectCall::set_to_interpreted knows the
986 // exact layout of this stub.
987
988 isb();
989 mov_metadata(rmethod, nullptr);
990
991 // Jump to the entry point of the c2i stub.
992 movptr(rscratch1, 0);
993 br(rscratch1);
994 }
995
996 int MacroAssembler::static_call_stub_size() {
997 // isb; movk; movz; movz; movk; movz; movz; br
998 return 8 * NativeInstruction::instruction_size;
999 }
1000
1001 void MacroAssembler::c2bool(Register x) {
1002 // implements x == 0 ? 0 : 1
1003 // note: must only look at least-significant byte of x
1004 // since C-style booleans are stored in one byte
1005 // only! (was bug)
1006 tst(x, 0xff);
1007 cset(x, Assembler::NE);
1008 }
1009
1010 address MacroAssembler::ic_call(address entry, jint method_index) {
1011 RelocationHolder rh = virtual_call_Relocation::spec(pc(), method_index);
1012 movptr(rscratch2, (intptr_t)Universe::non_oop_word());
1013 return trampoline_call(Address(entry, rh));
1014 }
1015
1016 int MacroAssembler::ic_check_size() {
1017 int extra_instructions = UseCompactObjectHeaders ? 1 : 0;
1018 if (target_needs_far_branch(CAST_FROM_FN_PTR(address, SharedRuntime::get_ic_miss_stub()))) {
1019 return NativeInstruction::instruction_size * (7 + extra_instructions);
1020 } else {
1021 return NativeInstruction::instruction_size * (5 + extra_instructions);
1022 }
1023 }
1024
1025 int MacroAssembler::ic_check(int end_alignment) {
1026 Register receiver = j_rarg0;
1027 Register data = rscratch2;
1028 Register tmp1 = rscratch1;
1029 Register tmp2 = r10;
1030
1031 // The UEP of a code blob ensures that the VEP is padded. However, the padding of the UEP is placed
1032 // before the inline cache check, so we don't have to execute any nop instructions when dispatching
1033 // through the UEP, yet we can ensure that the VEP is aligned appropriately. That's why we align
1034 // before the inline cache check here, and not after
1035 align(end_alignment, offset() + ic_check_size());
1036
1037 int uep_offset = offset();
1038
1039 if (UseCompactObjectHeaders) {
1040 load_narrow_klass_compact(tmp1, receiver);
1041 ldrw(tmp2, Address(data, CompiledICData::speculated_klass_offset()));
1042 cmpw(tmp1, tmp2);
1043 } else if (UseCompressedClassPointers) {
1044 ldrw(tmp1, Address(receiver, oopDesc::klass_offset_in_bytes()));
1045 ldrw(tmp2, Address(data, CompiledICData::speculated_klass_offset()));
1046 cmpw(tmp1, tmp2);
1047 } else {
1048 ldr(tmp1, Address(receiver, oopDesc::klass_offset_in_bytes()));
1049 ldr(tmp2, Address(data, CompiledICData::speculated_klass_offset()));
1050 cmp(tmp1, tmp2);
1051 }
1052
1053 Label dont;
1054 br(Assembler::EQ, dont);
1055 far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
1056 bind(dont);
1057 assert((offset() % end_alignment) == 0, "Misaligned verified entry point");
1058
1059 return uep_offset;
1060 }
1061
1062 // Implementation of call_VM versions
1063
1064 void MacroAssembler::call_VM(Register oop_result,
1065 address entry_point,
1066 bool check_exceptions) {
1067 call_VM_helper(oop_result, entry_point, 0, check_exceptions);
1068 }
1069
1070 void MacroAssembler::call_VM(Register oop_result,
1071 address entry_point,
1072 Register arg_1,
1073 bool check_exceptions) {
1074 pass_arg1(this, arg_1);
1075 call_VM_helper(oop_result, entry_point, 1, check_exceptions);
1076 }
1077
1078 void MacroAssembler::call_VM(Register oop_result,
1079 address entry_point,
1080 Register arg_1,
1081 Register arg_2,
1082 bool check_exceptions) {
1083 assert_different_registers(arg_1, c_rarg2);
1084 pass_arg2(this, arg_2);
1085 pass_arg1(this, arg_1);
1086 call_VM_helper(oop_result, entry_point, 2, check_exceptions);
1087 }
1088
1089 void MacroAssembler::call_VM(Register oop_result,
1090 address entry_point,
1091 Register arg_1,
1092 Register arg_2,
1093 Register arg_3,
1094 bool check_exceptions) {
1095 assert_different_registers(arg_1, c_rarg2, c_rarg3);
1096 assert_different_registers(arg_2, c_rarg3);
1097 pass_arg3(this, arg_3);
1098
1099 pass_arg2(this, arg_2);
1100
1101 pass_arg1(this, arg_1);
1102 call_VM_helper(oop_result, entry_point, 3, check_exceptions);
1103 }
1104
1105 void MacroAssembler::call_VM(Register oop_result,
1106 Register last_java_sp,
1107 address entry_point,
1108 int number_of_arguments,
1109 bool check_exceptions) {
1110 call_VM_base(oop_result, rthread, last_java_sp, entry_point, number_of_arguments, check_exceptions);
1111 }
1112
1113 void MacroAssembler::call_VM(Register oop_result,
1114 Register last_java_sp,
1115 address entry_point,
1116 Register arg_1,
1117 bool check_exceptions) {
1118 pass_arg1(this, arg_1);
1119 call_VM(oop_result, last_java_sp, entry_point, 1, check_exceptions);
1120 }
1121
1122 void MacroAssembler::call_VM(Register oop_result,
1123 Register last_java_sp,
1124 address entry_point,
1125 Register arg_1,
1126 Register arg_2,
1127 bool check_exceptions) {
1128
1129 assert_different_registers(arg_1, c_rarg2);
1130 pass_arg2(this, arg_2);
1131 pass_arg1(this, arg_1);
1132 call_VM(oop_result, last_java_sp, entry_point, 2, check_exceptions);
1133 }
1134
1135 void MacroAssembler::call_VM(Register oop_result,
1136 Register last_java_sp,
1137 address entry_point,
1138 Register arg_1,
1139 Register arg_2,
1140 Register arg_3,
1141 bool check_exceptions) {
1142 assert_different_registers(arg_1, c_rarg2, c_rarg3);
1143 assert_different_registers(arg_2, c_rarg3);
1144 pass_arg3(this, arg_3);
1145 pass_arg2(this, arg_2);
1146 pass_arg1(this, arg_1);
1147 call_VM(oop_result, last_java_sp, entry_point, 3, check_exceptions);
1148 }
1149
1150
1151 void MacroAssembler::get_vm_result_oop(Register oop_result, Register java_thread) {
1152 ldr(oop_result, Address(java_thread, JavaThread::vm_result_oop_offset()));
1153 str(zr, Address(java_thread, JavaThread::vm_result_oop_offset()));
1154 verify_oop_msg(oop_result, "broken oop in call_VM_base");
1155 }
1156
1157 void MacroAssembler::get_vm_result_metadata(Register metadata_result, Register java_thread) {
1158 ldr(metadata_result, Address(java_thread, JavaThread::vm_result_metadata_offset()));
1159 str(zr, Address(java_thread, JavaThread::vm_result_metadata_offset()));
1160 }
1161
1162 void MacroAssembler::align(int modulus) {
1163 align(modulus, offset());
1164 }
1165
1166 // Ensure that the code at target bytes offset from the current offset() is aligned
1167 // according to modulus.
1168 void MacroAssembler::align(int modulus, int target) {
1169 int delta = target - offset();
1170 while ((offset() + delta) % modulus != 0) nop();
1171 }
1172
1173 void MacroAssembler::post_call_nop() {
1174 if (!Continuations::enabled()) {
1175 return;
1176 }
1177 InstructionMark im(this);
1178 relocate(post_call_nop_Relocation::spec());
1179 InlineSkippedInstructionsCounter skipCounter(this);
1180 nop();
1181 movk(zr, 0);
1182 movk(zr, 0);
1183 }
1184
1185 // these are no-ops overridden by InterpreterMacroAssembler
1186
1187 void MacroAssembler::check_and_handle_earlyret(Register java_thread) { }
1188
1189 void MacroAssembler::check_and_handle_popframe(Register java_thread) { }
1190
1191 // Look up the method for a megamorphic invokeinterface call.
1192 // The target method is determined by <intf_klass, itable_index>.
1193 // The receiver klass is in recv_klass.
1194 // On success, the result will be in method_result, and execution falls through.
1195 // On failure, execution transfers to the given label.
1196 void MacroAssembler::lookup_interface_method(Register recv_klass,
1197 Register intf_klass,
1198 RegisterOrConstant itable_index,
1199 Register method_result,
1200 Register scan_temp,
1201 Label& L_no_such_interface,
1202 bool return_method) {
1203 assert_different_registers(recv_klass, intf_klass, scan_temp);
1204 assert_different_registers(method_result, intf_klass, scan_temp);
1205 assert(recv_klass != method_result || !return_method,
1206 "recv_klass can be destroyed when method isn't needed");
1207 assert(itable_index.is_constant() || itable_index.as_register() == method_result,
1208 "caller must use same register for non-constant itable index as for method");
1209
1210 // Compute start of first itableOffsetEntry (which is at the end of the vtable)
1211 int vtable_base = in_bytes(Klass::vtable_start_offset());
1212 int itentry_off = in_bytes(itableMethodEntry::method_offset());
1213 int scan_step = itableOffsetEntry::size() * wordSize;
1214 int vte_size = vtableEntry::size_in_bytes();
1215 assert(vte_size == wordSize, "else adjust times_vte_scale");
1216
1217 ldrw(scan_temp, Address(recv_klass, Klass::vtable_length_offset()));
1218
1219 // Could store the aligned, prescaled offset in the klass.
1220 // lea(scan_temp, Address(recv_klass, scan_temp, times_vte_scale, vtable_base));
1221 lea(scan_temp, Address(recv_klass, scan_temp, Address::lsl(3)));
1222 add(scan_temp, scan_temp, vtable_base);
1223
1224 if (return_method) {
1225 // Adjust recv_klass by scaled itable_index, so we can free itable_index.
1226 assert(itableMethodEntry::size() * wordSize == wordSize, "adjust the scaling in the code below");
1227 // lea(recv_klass, Address(recv_klass, itable_index, Address::times_ptr, itentry_off));
1228 lea(recv_klass, Address(recv_klass, itable_index, Address::lsl(3)));
1229 if (itentry_off)
1230 add(recv_klass, recv_klass, itentry_off);
1231 }
1232
1233 // for (scan = klass->itable(); scan->interface() != nullptr; scan += scan_step) {
1234 // if (scan->interface() == intf) {
1235 // result = (klass + scan->offset() + itable_index);
1236 // }
1237 // }
1238 Label search, found_method;
1239
1240 ldr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset()));
1241 cmp(intf_klass, method_result);
1242 br(Assembler::EQ, found_method);
1243 bind(search);
1244 // Check that the previous entry is non-null. A null entry means that
1245 // the receiver class doesn't implement the interface, and wasn't the
1246 // same as when the caller was compiled.
1247 cbz(method_result, L_no_such_interface);
1248 if (itableOffsetEntry::interface_offset() != 0) {
1249 add(scan_temp, scan_temp, scan_step);
1250 ldr(method_result, Address(scan_temp, itableOffsetEntry::interface_offset()));
1251 } else {
1252 ldr(method_result, Address(pre(scan_temp, scan_step)));
1253 }
1254 cmp(intf_klass, method_result);
1255 br(Assembler::NE, search);
1256
1257 bind(found_method);
1258
1259 // Got a hit.
1260 if (return_method) {
1261 ldrw(scan_temp, Address(scan_temp, itableOffsetEntry::offset_offset()));
1262 ldr(method_result, Address(recv_klass, scan_temp, Address::uxtw(0)));
1263 }
1264 }
1265
1266 // Look up the method for a megamorphic invokeinterface call in a single pass over itable:
1267 // - check recv_klass (actual object class) is a subtype of resolved_klass from CompiledICData
1268 // - find a holder_klass (class that implements the method) vtable offset and get the method from vtable by index
1269 // The target method is determined by <holder_klass, itable_index>.
1270 // The receiver klass is in recv_klass.
1271 // On success, the result will be in method_result, and execution falls through.
1272 // On failure, execution transfers to the given label.
1273 void MacroAssembler::lookup_interface_method_stub(Register recv_klass,
1274 Register holder_klass,
1275 Register resolved_klass,
1276 Register method_result,
1277 Register temp_itbl_klass,
1278 Register scan_temp,
1279 int itable_index,
1280 Label& L_no_such_interface) {
1281 // 'method_result' is only used as output register at the very end of this method.
1282 // Until then we can reuse it as 'holder_offset'.
1283 Register holder_offset = method_result;
1284 assert_different_registers(resolved_klass, recv_klass, holder_klass, temp_itbl_klass, scan_temp, holder_offset);
1285
1286 int vtable_start_offset = in_bytes(Klass::vtable_start_offset());
1287 int itable_offset_entry_size = itableOffsetEntry::size() * wordSize;
1288 int ioffset = in_bytes(itableOffsetEntry::interface_offset());
1289 int ooffset = in_bytes(itableOffsetEntry::offset_offset());
1290
1291 Label L_loop_search_resolved_entry, L_resolved_found, L_holder_found;
1292
1293 ldrw(scan_temp, Address(recv_klass, Klass::vtable_length_offset()));
1294 add(recv_klass, recv_klass, vtable_start_offset + ioffset);
1295 // itableOffsetEntry[] itable = recv_klass + Klass::vtable_start_offset() + sizeof(vtableEntry) * recv_klass->_vtable_len;
1296 // temp_itbl_klass = itable[0]._interface;
1297 int vtblEntrySize = vtableEntry::size_in_bytes();
1298 assert(vtblEntrySize == wordSize, "ldr lsl shift amount must be 3");
1299 ldr(temp_itbl_klass, Address(recv_klass, scan_temp, Address::lsl(exact_log2(vtblEntrySize))));
1300 mov(holder_offset, zr);
1301 // scan_temp = &(itable[0]._interface)
1302 lea(scan_temp, Address(recv_klass, scan_temp, Address::lsl(exact_log2(vtblEntrySize))));
1303
1304 // Initial checks:
1305 // - if (holder_klass != resolved_klass), go to "scan for resolved"
1306 // - if (itable[0] == holder_klass), shortcut to "holder found"
1307 // - if (itable[0] == 0), no such interface
1308 cmp(resolved_klass, holder_klass);
1309 br(Assembler::NE, L_loop_search_resolved_entry);
1310 cmp(holder_klass, temp_itbl_klass);
1311 br(Assembler::EQ, L_holder_found);
1312 cbz(temp_itbl_klass, L_no_such_interface);
1313
1314 // Loop: Look for holder_klass record in itable
1315 // do {
1316 // temp_itbl_klass = *(scan_temp += itable_offset_entry_size);
1317 // if (temp_itbl_klass == holder_klass) {
1318 // goto L_holder_found; // Found!
1319 // }
1320 // } while (temp_itbl_klass != 0);
1321 // goto L_no_such_interface // Not found.
1322 Label L_search_holder;
1323 bind(L_search_holder);
1324 ldr(temp_itbl_klass, Address(pre(scan_temp, itable_offset_entry_size)));
1325 cmp(holder_klass, temp_itbl_klass);
1326 br(Assembler::EQ, L_holder_found);
1327 cbnz(temp_itbl_klass, L_search_holder);
1328
1329 b(L_no_such_interface);
1330
1331 // Loop: Look for resolved_class record in itable
1332 // while (true) {
1333 // temp_itbl_klass = *(scan_temp += itable_offset_entry_size);
1334 // if (temp_itbl_klass == 0) {
1335 // goto L_no_such_interface;
1336 // }
1337 // if (temp_itbl_klass == resolved_klass) {
1338 // goto L_resolved_found; // Found!
1339 // }
1340 // if (temp_itbl_klass == holder_klass) {
1341 // holder_offset = scan_temp;
1342 // }
1343 // }
1344 //
1345 Label L_loop_search_resolved;
1346 bind(L_loop_search_resolved);
1347 ldr(temp_itbl_klass, Address(pre(scan_temp, itable_offset_entry_size)));
1348 bind(L_loop_search_resolved_entry);
1349 cbz(temp_itbl_klass, L_no_such_interface);
1350 cmp(resolved_klass, temp_itbl_klass);
1351 br(Assembler::EQ, L_resolved_found);
1352 cmp(holder_klass, temp_itbl_klass);
1353 br(Assembler::NE, L_loop_search_resolved);
1354 mov(holder_offset, scan_temp);
1355 b(L_loop_search_resolved);
1356
1357 // See if we already have a holder klass. If not, go and scan for it.
1358 bind(L_resolved_found);
1359 cbz(holder_offset, L_search_holder);
1360 mov(scan_temp, holder_offset);
1361
1362 // Finally, scan_temp contains holder_klass vtable offset
1363 bind(L_holder_found);
1364 ldrw(method_result, Address(scan_temp, ooffset - ioffset));
1365 add(recv_klass, recv_klass, itable_index * wordSize + in_bytes(itableMethodEntry::method_offset())
1366 - vtable_start_offset - ioffset); // substract offsets to restore the original value of recv_klass
1367 ldr(method_result, Address(recv_klass, method_result, Address::uxtw(0)));
1368 }
1369
1370 // virtual method calling
1371 void MacroAssembler::lookup_virtual_method(Register recv_klass,
1372 RegisterOrConstant vtable_index,
1373 Register method_result) {
1374 assert(vtableEntry::size() * wordSize == 8,
1375 "adjust the scaling in the code below");
1376 int64_t vtable_offset_in_bytes = in_bytes(Klass::vtable_start_offset() + vtableEntry::method_offset());
1377
1378 if (vtable_index.is_register()) {
1379 lea(method_result, Address(recv_klass,
1380 vtable_index.as_register(),
1381 Address::lsl(LogBytesPerWord)));
1382 ldr(method_result, Address(method_result, vtable_offset_in_bytes));
1383 } else {
1384 vtable_offset_in_bytes += vtable_index.as_constant() * wordSize;
1385 ldr(method_result,
1386 form_address(rscratch1, recv_klass, vtable_offset_in_bytes, 0));
1387 }
1388 }
1389
1390 void MacroAssembler::check_klass_subtype(Register sub_klass,
1391 Register super_klass,
1392 Register temp_reg,
1393 Label& L_success) {
1394 Label L_failure;
1395 check_klass_subtype_fast_path(sub_klass, super_klass, temp_reg, &L_success, &L_failure, nullptr);
1396 check_klass_subtype_slow_path(sub_klass, super_klass, temp_reg, noreg, &L_success, nullptr);
1397 bind(L_failure);
1398 }
1399
1400
1401 void MacroAssembler::check_klass_subtype_fast_path(Register sub_klass,
1402 Register super_klass,
1403 Register temp_reg,
1404 Label* L_success,
1405 Label* L_failure,
1406 Label* L_slow_path,
1407 Register super_check_offset) {
1408 assert_different_registers(sub_klass, super_klass, temp_reg, super_check_offset);
1409 bool must_load_sco = ! super_check_offset->is_valid();
1410 if (must_load_sco) {
1411 assert(temp_reg != noreg, "supply either a temp or a register offset");
1412 }
1413
1414 Label L_fallthrough;
1415 int label_nulls = 0;
1416 if (L_success == nullptr) { L_success = &L_fallthrough; label_nulls++; }
1417 if (L_failure == nullptr) { L_failure = &L_fallthrough; label_nulls++; }
1418 if (L_slow_path == nullptr) { L_slow_path = &L_fallthrough; label_nulls++; }
1419 assert(label_nulls <= 1, "at most one null in the batch");
1420
1421 int sco_offset = in_bytes(Klass::super_check_offset_offset());
1422 Address super_check_offset_addr(super_klass, sco_offset);
1423
1424 // Hacked jmp, which may only be used just before L_fallthrough.
1425 #define final_jmp(label) \
1426 if (&(label) == &L_fallthrough) { /*do nothing*/ } \
1427 else b(label) /*omit semi*/
1428
1429 // If the pointers are equal, we are done (e.g., String[] elements).
1430 // This self-check enables sharing of secondary supertype arrays among
1431 // non-primary types such as array-of-interface. Otherwise, each such
1432 // type would need its own customized SSA.
1433 // We move this check to the front of the fast path because many
1434 // type checks are in fact trivially successful in this manner,
1435 // so we get a nicely predicted branch right at the start of the check.
1436 cmp(sub_klass, super_klass);
1437 br(Assembler::EQ, *L_success);
1438
1439 // Check the supertype display:
1440 if (must_load_sco) {
1441 ldrw(temp_reg, super_check_offset_addr);
1442 super_check_offset = temp_reg;
1443 }
1444
1445 Address super_check_addr(sub_klass, super_check_offset);
1446 ldr(rscratch1, super_check_addr);
1447 cmp(super_klass, rscratch1); // load displayed supertype
1448 br(Assembler::EQ, *L_success);
1449
1450 // This check has worked decisively for primary supers.
1451 // Secondary supers are sought in the super_cache ('super_cache_addr').
1452 // (Secondary supers are interfaces and very deeply nested subtypes.)
1453 // This works in the same check above because of a tricky aliasing
1454 // between the super_cache and the primary super display elements.
1455 // (The 'super_check_addr' can address either, as the case requires.)
1456 // Note that the cache is updated below if it does not help us find
1457 // what we need immediately.
1458 // So if it was a primary super, we can just fail immediately.
1459 // Otherwise, it's the slow path for us (no success at this point).
1460
1461 sub(rscratch1, super_check_offset, in_bytes(Klass::secondary_super_cache_offset()));
1462 if (L_failure == &L_fallthrough) {
1463 cbz(rscratch1, *L_slow_path);
1464 } else {
1465 cbnz(rscratch1, *L_failure);
1466 final_jmp(*L_slow_path);
1467 }
1468
1469 bind(L_fallthrough);
1470
1471 #undef final_jmp
1472 }
1473
1474 // These two are taken from x86, but they look generally useful
1475
1476 // scans count pointer sized words at [addr] for occurrence of value,
1477 // generic
1478 void MacroAssembler::repne_scan(Register addr, Register value, Register count,
1479 Register scratch) {
1480 Label Lloop, Lexit;
1481 cbz(count, Lexit);
1482 bind(Lloop);
1483 ldr(scratch, post(addr, wordSize));
1484 cmp(value, scratch);
1485 br(EQ, Lexit);
1486 sub(count, count, 1);
1487 cbnz(count, Lloop);
1488 bind(Lexit);
1489 }
1490
1491 // scans count 4 byte words at [addr] for occurrence of value,
1492 // generic
1493 void MacroAssembler::repne_scanw(Register addr, Register value, Register count,
1494 Register scratch) {
1495 Label Lloop, Lexit;
1496 cbz(count, Lexit);
1497 bind(Lloop);
1498 ldrw(scratch, post(addr, wordSize));
1499 cmpw(value, scratch);
1500 br(EQ, Lexit);
1501 sub(count, count, 1);
1502 cbnz(count, Lloop);
1503 bind(Lexit);
1504 }
1505
1506 void MacroAssembler::check_klass_subtype_slow_path_linear(Register sub_klass,
1507 Register super_klass,
1508 Register temp_reg,
1509 Register temp2_reg,
1510 Label* L_success,
1511 Label* L_failure,
1512 bool set_cond_codes) {
1513 // NB! Callers may assume that, when temp2_reg is a valid register,
1514 // this code sets it to a nonzero value.
1515
1516 assert_different_registers(sub_klass, super_klass, temp_reg);
1517 if (temp2_reg != noreg)
1518 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg, rscratch1);
1519 #define IS_A_TEMP(reg) ((reg) == temp_reg || (reg) == temp2_reg)
1520
1521 Label L_fallthrough;
1522 int label_nulls = 0;
1523 if (L_success == nullptr) { L_success = &L_fallthrough; label_nulls++; }
1524 if (L_failure == nullptr) { L_failure = &L_fallthrough; label_nulls++; }
1525 assert(label_nulls <= 1, "at most one null in the batch");
1526
1527 // a couple of useful fields in sub_klass:
1528 int ss_offset = in_bytes(Klass::secondary_supers_offset());
1529 int sc_offset = in_bytes(Klass::secondary_super_cache_offset());
1530 Address secondary_supers_addr(sub_klass, ss_offset);
1531 Address super_cache_addr( sub_klass, sc_offset);
1532
1533 BLOCK_COMMENT("check_klass_subtype_slow_path");
1534
1535 // Do a linear scan of the secondary super-klass chain.
1536 // This code is rarely used, so simplicity is a virtue here.
1537 // The repne_scan instruction uses fixed registers, which we must spill.
1538 // Don't worry too much about pre-existing connections with the input regs.
1539
1540 assert(sub_klass != r0, "killed reg"); // killed by mov(r0, super)
1541 assert(sub_klass != r2, "killed reg"); // killed by lea(r2, &pst_counter)
1542
1543 RegSet pushed_registers;
1544 if (!IS_A_TEMP(r2)) pushed_registers += r2;
1545 if (!IS_A_TEMP(r5)) pushed_registers += r5;
1546
1547 if (super_klass != r0) {
1548 if (!IS_A_TEMP(r0)) pushed_registers += r0;
1549 }
1550
1551 push(pushed_registers, sp);
1552
1553 // Get super_klass value into r0 (even if it was in r5 or r2).
1554 if (super_klass != r0) {
1555 mov(r0, super_klass);
1556 }
1557
1558 #ifndef PRODUCT
1559 incrementw(ExternalAddress((address)&SharedRuntime::_partial_subtype_ctr));
1560 #endif //PRODUCT
1561
1562 // We will consult the secondary-super array.
1563 ldr(r5, secondary_supers_addr);
1564 // Load the array length.
1565 ldrw(r2, Address(r5, Array<Klass*>::length_offset_in_bytes()));
1566 // Skip to start of data.
1567 add(r5, r5, Array<Klass*>::base_offset_in_bytes());
1568
1569 cmp(sp, zr); // Clear Z flag; SP is never zero
1570 // Scan R2 words at [R5] for an occurrence of R0.
1571 // Set NZ/Z based on last compare.
1572 repne_scan(r5, r0, r2, rscratch1);
1573
1574 // Unspill the temp. registers:
1575 pop(pushed_registers, sp);
1576
1577 br(Assembler::NE, *L_failure);
1578
1579 // Success. Cache the super we found and proceed in triumph.
1580
1581 if (UseSecondarySupersCache) {
1582 str(super_klass, super_cache_addr);
1583 }
1584
1585 if (L_success != &L_fallthrough) {
1586 b(*L_success);
1587 }
1588
1589 #undef IS_A_TEMP
1590
1591 bind(L_fallthrough);
1592 }
1593
1594 // If Register r is invalid, remove a new register from
1595 // available_regs, and add new register to regs_to_push.
1596 Register MacroAssembler::allocate_if_noreg(Register r,
1597 RegSetIterator<Register> &available_regs,
1598 RegSet ®s_to_push) {
1599 if (!r->is_valid()) {
1600 r = *available_regs++;
1601 regs_to_push += r;
1602 }
1603 return r;
1604 }
1605
1606 // check_klass_subtype_slow_path_table() looks for super_klass in the
1607 // hash table belonging to super_klass, branching to L_success or
1608 // L_failure as appropriate. This is essentially a shim which
1609 // allocates registers as necessary then calls
1610 // lookup_secondary_supers_table() to do the work. Any of the temp
1611 // regs may be noreg, in which case this logic will chooses some
1612 // registers push and pop them from the stack.
1613 void MacroAssembler::check_klass_subtype_slow_path_table(Register sub_klass,
1614 Register super_klass,
1615 Register temp_reg,
1616 Register temp2_reg,
1617 Register temp3_reg,
1618 Register result_reg,
1619 FloatRegister vtemp,
1620 Label* L_success,
1621 Label* L_failure,
1622 bool set_cond_codes) {
1623 RegSet temps = RegSet::of(temp_reg, temp2_reg, temp3_reg);
1624
1625 assert_different_registers(sub_klass, super_klass, temp_reg, temp2_reg, rscratch1);
1626
1627 Label L_fallthrough;
1628 int label_nulls = 0;
1629 if (L_success == nullptr) { L_success = &L_fallthrough; label_nulls++; }
1630 if (L_failure == nullptr) { L_failure = &L_fallthrough; label_nulls++; }
1631 assert(label_nulls <= 1, "at most one null in the batch");
1632
1633 BLOCK_COMMENT("check_klass_subtype_slow_path");
1634
1635 RegSetIterator<Register> available_regs
1636 = (RegSet::range(r0, r15) - temps - sub_klass - super_klass).begin();
1637
1638 RegSet pushed_regs;
1639
1640 temp_reg = allocate_if_noreg(temp_reg, available_regs, pushed_regs);
1641 temp2_reg = allocate_if_noreg(temp2_reg, available_regs, pushed_regs);
1642 temp3_reg = allocate_if_noreg(temp3_reg, available_regs, pushed_regs);
1643 result_reg = allocate_if_noreg(result_reg, available_regs, pushed_regs);
1644
1645 push(pushed_regs, sp);
1646
1647 lookup_secondary_supers_table_var(sub_klass,
1648 super_klass,
1649 temp_reg, temp2_reg, temp3_reg, vtemp, result_reg,
1650 nullptr);
1651 cmp(result_reg, zr);
1652
1653 // Unspill the temp. registers:
1654 pop(pushed_regs, sp);
1655
1656 // NB! Callers may assume that, when set_cond_codes is true, this
1657 // code sets temp2_reg to a nonzero value.
1658 if (set_cond_codes) {
1659 mov(temp2_reg, 1);
1660 }
1661
1662 br(Assembler::NE, *L_failure);
1663
1664 if (L_success != &L_fallthrough) {
1665 b(*L_success);
1666 }
1667
1668 bind(L_fallthrough);
1669 }
1670
1671 void MacroAssembler::check_klass_subtype_slow_path(Register sub_klass,
1672 Register super_klass,
1673 Register temp_reg,
1674 Register temp2_reg,
1675 Label* L_success,
1676 Label* L_failure,
1677 bool set_cond_codes) {
1678 if (UseSecondarySupersTable) {
1679 check_klass_subtype_slow_path_table
1680 (sub_klass, super_klass, temp_reg, temp2_reg, /*temp3*/noreg, /*result*/noreg,
1681 /*vtemp*/fnoreg,
1682 L_success, L_failure, set_cond_codes);
1683 } else {
1684 check_klass_subtype_slow_path_linear
1685 (sub_klass, super_klass, temp_reg, temp2_reg, L_success, L_failure, set_cond_codes);
1686 }
1687 }
1688
1689
1690 // Ensure that the inline code and the stub are using the same registers.
1691 #define LOOKUP_SECONDARY_SUPERS_TABLE_REGISTERS \
1692 do { \
1693 assert(r_super_klass == r0 && \
1694 r_array_base == r1 && \
1695 r_array_length == r2 && \
1696 (r_array_index == r3 || r_array_index == noreg) && \
1697 (r_sub_klass == r4 || r_sub_klass == noreg) && \
1698 (r_bitmap == rscratch2 || r_bitmap == noreg) && \
1699 (result == r5 || result == noreg), "registers must match aarch64.ad"); \
1700 } while(0)
1701
1702 bool MacroAssembler::lookup_secondary_supers_table_const(Register r_sub_klass,
1703 Register r_super_klass,
1704 Register temp1,
1705 Register temp2,
1706 Register temp3,
1707 FloatRegister vtemp,
1708 Register result,
1709 u1 super_klass_slot,
1710 bool stub_is_near) {
1711 assert_different_registers(r_sub_klass, temp1, temp2, temp3, result, rscratch1, rscratch2);
1712
1713 Label L_fallthrough;
1714
1715 BLOCK_COMMENT("lookup_secondary_supers_table {");
1716
1717 const Register
1718 r_array_base = temp1, // r1
1719 r_array_length = temp2, // r2
1720 r_array_index = temp3, // r3
1721 r_bitmap = rscratch2;
1722
1723 LOOKUP_SECONDARY_SUPERS_TABLE_REGISTERS;
1724
1725 u1 bit = super_klass_slot;
1726
1727 // Make sure that result is nonzero if the TBZ below misses.
1728 mov(result, 1);
1729
1730 // We're going to need the bitmap in a vector reg and in a core reg,
1731 // so load both now.
1732 ldr(r_bitmap, Address(r_sub_klass, Klass::secondary_supers_bitmap_offset()));
1733 if (bit != 0) {
1734 ldrd(vtemp, Address(r_sub_klass, Klass::secondary_supers_bitmap_offset()));
1735 }
1736 // First check the bitmap to see if super_klass might be present. If
1737 // the bit is zero, we are certain that super_klass is not one of
1738 // the secondary supers.
1739 tbz(r_bitmap, bit, L_fallthrough);
1740
1741 // Get the first array index that can contain super_klass into r_array_index.
1742 if (bit != 0) {
1743 shld(vtemp, vtemp, Klass::SECONDARY_SUPERS_TABLE_MASK - bit);
1744 cnt(vtemp, T8B, vtemp);
1745 addv(vtemp, T8B, vtemp);
1746 fmovd(r_array_index, vtemp);
1747 } else {
1748 mov(r_array_index, (u1)1);
1749 }
1750 // NB! r_array_index is off by 1. It is compensated by keeping r_array_base off by 1 word.
1751
1752 // We will consult the secondary-super array.
1753 ldr(r_array_base, Address(r_sub_klass, in_bytes(Klass::secondary_supers_offset())));
1754
1755 // The value i in r_array_index is >= 1, so even though r_array_base
1756 // points to the length, we don't need to adjust it to point to the
1757 // data.
1758 assert(Array<Klass*>::base_offset_in_bytes() == wordSize, "Adjust this code");
1759 assert(Array<Klass*>::length_offset_in_bytes() == 0, "Adjust this code");
1760
1761 ldr(result, Address(r_array_base, r_array_index, Address::lsl(LogBytesPerWord)));
1762 eor(result, result, r_super_klass);
1763 cbz(result, L_fallthrough); // Found a match
1764
1765 // Is there another entry to check? Consult the bitmap.
1766 tbz(r_bitmap, (bit + 1) & Klass::SECONDARY_SUPERS_TABLE_MASK, L_fallthrough);
1767
1768 // Linear probe.
1769 if (bit != 0) {
1770 ror(r_bitmap, r_bitmap, bit);
1771 }
1772
1773 // The slot we just inspected is at secondary_supers[r_array_index - 1].
1774 // The next slot to be inspected, by the stub we're about to call,
1775 // is secondary_supers[r_array_index]. Bits 0 and 1 in the bitmap
1776 // have been checked.
1777 Address stub = RuntimeAddress(StubRoutines::lookup_secondary_supers_table_slow_path_stub());
1778 if (stub_is_near) {
1779 bl(stub);
1780 } else {
1781 address call = trampoline_call(stub);
1782 if (call == nullptr) {
1783 return false; // trampoline allocation failed
1784 }
1785 }
1786
1787 BLOCK_COMMENT("} lookup_secondary_supers_table");
1788
1789 bind(L_fallthrough);
1790
1791 if (VerifySecondarySupers) {
1792 verify_secondary_supers_table(r_sub_klass, r_super_klass, // r4, r0
1793 temp1, temp2, result); // r1, r2, r5
1794 }
1795 return true;
1796 }
1797
1798 // At runtime, return 0 in result if r_super_klass is a superclass of
1799 // r_sub_klass, otherwise return nonzero. Use this version of
1800 // lookup_secondary_supers_table() if you don't know ahead of time
1801 // which superclass will be searched for. Used by interpreter and
1802 // runtime stubs. It is larger and has somewhat greater latency than
1803 // the version above, which takes a constant super_klass_slot.
1804 void MacroAssembler::lookup_secondary_supers_table_var(Register r_sub_klass,
1805 Register r_super_klass,
1806 Register temp1,
1807 Register temp2,
1808 Register temp3,
1809 FloatRegister vtemp,
1810 Register result,
1811 Label *L_success) {
1812 assert_different_registers(r_sub_klass, temp1, temp2, temp3, result, rscratch1, rscratch2);
1813
1814 Label L_fallthrough;
1815
1816 BLOCK_COMMENT("lookup_secondary_supers_table {");
1817
1818 const Register
1819 r_array_index = temp3,
1820 slot = rscratch1,
1821 r_bitmap = rscratch2;
1822
1823 ldrb(slot, Address(r_super_klass, Klass::hash_slot_offset()));
1824
1825 // Make sure that result is nonzero if the test below misses.
1826 mov(result, 1);
1827
1828 ldr(r_bitmap, Address(r_sub_klass, Klass::secondary_supers_bitmap_offset()));
1829
1830 // First check the bitmap to see if super_klass might be present. If
1831 // the bit is zero, we are certain that super_klass is not one of
1832 // the secondary supers.
1833
1834 // This next instruction is equivalent to:
1835 // mov(tmp_reg, (u1)(Klass::SECONDARY_SUPERS_TABLE_SIZE - 1));
1836 // sub(temp2, tmp_reg, slot);
1837 eor(temp2, slot, (u1)(Klass::SECONDARY_SUPERS_TABLE_SIZE - 1));
1838 lslv(temp2, r_bitmap, temp2);
1839 tbz(temp2, Klass::SECONDARY_SUPERS_TABLE_SIZE - 1, L_fallthrough);
1840
1841 bool must_save_v0 = (vtemp == fnoreg);
1842 if (must_save_v0) {
1843 // temp1 and result are free, so use them to preserve vtemp
1844 vtemp = v0;
1845 mov(temp1, vtemp, D, 0);
1846 mov(result, vtemp, D, 1);
1847 }
1848
1849 // Get the first array index that can contain super_klass into r_array_index.
1850 mov(vtemp, D, 0, temp2);
1851 cnt(vtemp, T8B, vtemp);
1852 addv(vtemp, T8B, vtemp);
1853 mov(r_array_index, vtemp, D, 0);
1854
1855 if (must_save_v0) {
1856 mov(vtemp, D, 0, temp1 );
1857 mov(vtemp, D, 1, result);
1858 }
1859
1860 // NB! r_array_index is off by 1. It is compensated by keeping r_array_base off by 1 word.
1861
1862 const Register
1863 r_array_base = temp1,
1864 r_array_length = temp2;
1865
1866 // The value i in r_array_index is >= 1, so even though r_array_base
1867 // points to the length, we don't need to adjust it to point to the
1868 // data.
1869 assert(Array<Klass*>::base_offset_in_bytes() == wordSize, "Adjust this code");
1870 assert(Array<Klass*>::length_offset_in_bytes() == 0, "Adjust this code");
1871
1872 // We will consult the secondary-super array.
1873 ldr(r_array_base, Address(r_sub_klass, in_bytes(Klass::secondary_supers_offset())));
1874
1875 ldr(result, Address(r_array_base, r_array_index, Address::lsl(LogBytesPerWord)));
1876 eor(result, result, r_super_klass);
1877 cbz(result, L_success ? *L_success : L_fallthrough); // Found a match
1878
1879 // Is there another entry to check? Consult the bitmap.
1880 rorv(r_bitmap, r_bitmap, slot);
1881 // rol(r_bitmap, r_bitmap, 1);
1882 tbz(r_bitmap, 1, L_fallthrough);
1883
1884 // The slot we just inspected is at secondary_supers[r_array_index - 1].
1885 // The next slot to be inspected, by the logic we're about to call,
1886 // is secondary_supers[r_array_index]. Bits 0 and 1 in the bitmap
1887 // have been checked.
1888 lookup_secondary_supers_table_slow_path(r_super_klass, r_array_base, r_array_index,
1889 r_bitmap, r_array_length, result, /*is_stub*/false);
1890
1891 BLOCK_COMMENT("} lookup_secondary_supers_table");
1892
1893 bind(L_fallthrough);
1894
1895 if (VerifySecondarySupers) {
1896 verify_secondary_supers_table(r_sub_klass, r_super_klass, // r4, r0
1897 temp1, temp2, result); // r1, r2, r5
1898 }
1899
1900 if (L_success) {
1901 cbz(result, *L_success);
1902 }
1903 }
1904
1905 // Called by code generated by check_klass_subtype_slow_path
1906 // above. This is called when there is a collision in the hashed
1907 // lookup in the secondary supers array.
1908 void MacroAssembler::lookup_secondary_supers_table_slow_path(Register r_super_klass,
1909 Register r_array_base,
1910 Register r_array_index,
1911 Register r_bitmap,
1912 Register temp1,
1913 Register result,
1914 bool is_stub) {
1915 assert_different_registers(r_super_klass, r_array_base, r_array_index, r_bitmap, temp1, result, rscratch1);
1916
1917 const Register
1918 r_array_length = temp1,
1919 r_sub_klass = noreg; // unused
1920
1921 if (is_stub) {
1922 LOOKUP_SECONDARY_SUPERS_TABLE_REGISTERS;
1923 }
1924
1925 Label L_fallthrough, L_huge;
1926
1927 // Load the array length.
1928 ldrw(r_array_length, Address(r_array_base, Array<Klass*>::length_offset_in_bytes()));
1929 // And adjust the array base to point to the data.
1930 // NB! Effectively increments current slot index by 1.
1931 assert(Array<Klass*>::base_offset_in_bytes() == wordSize, "");
1932 add(r_array_base, r_array_base, Array<Klass*>::base_offset_in_bytes());
1933
1934 // The bitmap is full to bursting.
1935 // Implicit invariant: BITMAP_FULL implies (length > 0)
1936 assert(Klass::SECONDARY_SUPERS_BITMAP_FULL == ~uintx(0), "");
1937 cmpw(r_array_length, (u1)(Klass::SECONDARY_SUPERS_TABLE_SIZE - 2));
1938 br(GT, L_huge);
1939
1940 // NB! Our caller has checked bits 0 and 1 in the bitmap. The
1941 // current slot (at secondary_supers[r_array_index]) has not yet
1942 // been inspected, and r_array_index may be out of bounds if we
1943 // wrapped around the end of the array.
1944
1945 { // This is conventional linear probing, but instead of terminating
1946 // when a null entry is found in the table, we maintain a bitmap
1947 // in which a 0 indicates missing entries.
1948 // As long as the bitmap is not completely full,
1949 // array_length == popcount(bitmap). The array_length check above
1950 // guarantees there are 0s in the bitmap, so the loop eventually
1951 // terminates.
1952 Label L_loop;
1953 bind(L_loop);
1954
1955 // Check for wraparound.
1956 cmp(r_array_index, r_array_length);
1957 csel(r_array_index, zr, r_array_index, GE);
1958
1959 ldr(rscratch1, Address(r_array_base, r_array_index, Address::lsl(LogBytesPerWord)));
1960 eor(result, rscratch1, r_super_klass);
1961 cbz(result, L_fallthrough);
1962
1963 tbz(r_bitmap, 2, L_fallthrough); // look-ahead check (Bit 2); result is non-zero
1964
1965 ror(r_bitmap, r_bitmap, 1);
1966 add(r_array_index, r_array_index, 1);
1967 b(L_loop);
1968 }
1969
1970 { // Degenerate case: more than 64 secondary supers.
1971 // FIXME: We could do something smarter here, maybe a vectorized
1972 // comparison or a binary search, but is that worth any added
1973 // complexity?
1974 bind(L_huge);
1975 cmp(sp, zr); // Clear Z flag; SP is never zero
1976 repne_scan(r_array_base, r_super_klass, r_array_length, rscratch1);
1977 cset(result, NE); // result == 0 iff we got a match.
1978 }
1979
1980 bind(L_fallthrough);
1981 }
1982
1983 // Make sure that the hashed lookup and a linear scan agree.
1984 void MacroAssembler::verify_secondary_supers_table(Register r_sub_klass,
1985 Register r_super_klass,
1986 Register temp1,
1987 Register temp2,
1988 Register result) {
1989 assert_different_registers(r_sub_klass, r_super_klass, temp1, temp2, result, rscratch1);
1990
1991 const Register
1992 r_array_base = temp1,
1993 r_array_length = temp2,
1994 r_array_index = noreg, // unused
1995 r_bitmap = noreg; // unused
1996
1997 BLOCK_COMMENT("verify_secondary_supers_table {");
1998
1999 // We will consult the secondary-super array.
2000 ldr(r_array_base, Address(r_sub_klass, in_bytes(Klass::secondary_supers_offset())));
2001
2002 // Load the array length.
2003 ldrw(r_array_length, Address(r_array_base, Array<Klass*>::length_offset_in_bytes()));
2004 // And adjust the array base to point to the data.
2005 add(r_array_base, r_array_base, Array<Klass*>::base_offset_in_bytes());
2006
2007 cmp(sp, zr); // Clear Z flag; SP is never zero
2008 // Scan R2 words at [R5] for an occurrence of R0.
2009 // Set NZ/Z based on last compare.
2010 repne_scan(/*addr*/r_array_base, /*value*/r_super_klass, /*count*/r_array_length, rscratch2);
2011 // rscratch1 == 0 iff we got a match.
2012 cset(rscratch1, NE);
2013
2014 Label passed;
2015 cmp(result, zr);
2016 cset(result, NE); // normalize result to 0/1 for comparison
2017
2018 cmp(rscratch1, result);
2019 br(EQ, passed);
2020 {
2021 mov(r0, r_super_klass); // r0 <- r0
2022 mov(r1, r_sub_klass); // r1 <- r4
2023 mov(r2, /*expected*/rscratch1); // r2 <- r8
2024 mov(r3, result); // r3 <- r5
2025 mov(r4, (address)("mismatch")); // r4 <- const
2026 rt_call(CAST_FROM_FN_PTR(address, Klass::on_secondary_supers_verification_failure), rscratch2);
2027 should_not_reach_here();
2028 }
2029 bind(passed);
2030
2031 BLOCK_COMMENT("} verify_secondary_supers_table");
2032 }
2033
2034 void MacroAssembler::clinit_barrier(Register klass, Register scratch, Label* L_fast_path, Label* L_slow_path) {
2035 assert(L_fast_path != nullptr || L_slow_path != nullptr, "at least one is required");
2036 assert_different_registers(klass, rthread, scratch);
2037
2038 Label L_fallthrough, L_tmp;
2039 if (L_fast_path == nullptr) {
2040 L_fast_path = &L_fallthrough;
2041 } else if (L_slow_path == nullptr) {
2042 L_slow_path = &L_fallthrough;
2043 }
2044 // Fast path check: class is fully initialized
2045 lea(scratch, Address(klass, InstanceKlass::init_state_offset()));
2046 ldarb(scratch, scratch);
2047 cmp(scratch, InstanceKlass::fully_initialized);
2048 br(Assembler::EQ, *L_fast_path);
2049
2050 // Fast path check: current thread is initializer thread
2051 ldr(scratch, Address(klass, InstanceKlass::init_thread_offset()));
2052 cmp(rthread, scratch);
2053
2054 if (L_slow_path == &L_fallthrough) {
2055 br(Assembler::EQ, *L_fast_path);
2056 bind(*L_slow_path);
2057 } else if (L_fast_path == &L_fallthrough) {
2058 br(Assembler::NE, *L_slow_path);
2059 bind(*L_fast_path);
2060 } else {
2061 Unimplemented();
2062 }
2063 }
2064
2065 void MacroAssembler::_verify_oop(Register reg, const char* s, const char* file, int line) {
2066 if (!VerifyOops) return;
2067
2068 // Pass register number to verify_oop_subroutine
2069 const char* b = nullptr;
2070 {
2071 ResourceMark rm;
2072 stringStream ss;
2073 ss.print("verify_oop: %s: %s (%s:%d)", reg->name(), s, file, line);
2074 b = code_string(ss.as_string());
2075 }
2076 BLOCK_COMMENT("verify_oop {");
2077
2078 strip_return_address(); // This might happen within a stack frame.
2079 protect_return_address();
2080 stp(r0, rscratch1, Address(pre(sp, -2 * wordSize)));
2081 stp(rscratch2, lr, Address(pre(sp, -2 * wordSize)));
2082
2083 mov(r0, reg);
2084 movptr(rscratch1, (uintptr_t)(address)b);
2085
2086 // call indirectly to solve generation ordering problem
2087 lea(rscratch2, RuntimeAddress(StubRoutines::verify_oop_subroutine_entry_address()));
2088 ldr(rscratch2, Address(rscratch2));
2089 blr(rscratch2);
2090
2091 ldp(rscratch2, lr, Address(post(sp, 2 * wordSize)));
2092 ldp(r0, rscratch1, Address(post(sp, 2 * wordSize)));
2093 authenticate_return_address();
2094
2095 BLOCK_COMMENT("} verify_oop");
2096 }
2097
2098 void MacroAssembler::_verify_oop_addr(Address addr, const char* s, const char* file, int line) {
2099 if (!VerifyOops) return;
2100
2101 const char* b = nullptr;
2102 {
2103 ResourceMark rm;
2104 stringStream ss;
2105 ss.print("verify_oop_addr: %s (%s:%d)", s, file, line);
2106 b = code_string(ss.as_string());
2107 }
2108 BLOCK_COMMENT("verify_oop_addr {");
2109
2110 strip_return_address(); // This might happen within a stack frame.
2111 protect_return_address();
2112 stp(r0, rscratch1, Address(pre(sp, -2 * wordSize)));
2113 stp(rscratch2, lr, Address(pre(sp, -2 * wordSize)));
2114
2115 // addr may contain sp so we will have to adjust it based on the
2116 // pushes that we just did.
2117 if (addr.uses(sp)) {
2118 lea(r0, addr);
2119 ldr(r0, Address(r0, 4 * wordSize));
2120 } else {
2121 ldr(r0, addr);
2122 }
2123 movptr(rscratch1, (uintptr_t)(address)b);
2124
2125 // call indirectly to solve generation ordering problem
2126 lea(rscratch2, RuntimeAddress(StubRoutines::verify_oop_subroutine_entry_address()));
2127 ldr(rscratch2, Address(rscratch2));
2128 blr(rscratch2);
2129
2130 ldp(rscratch2, lr, Address(post(sp, 2 * wordSize)));
2131 ldp(r0, rscratch1, Address(post(sp, 2 * wordSize)));
2132 authenticate_return_address();
2133
2134 BLOCK_COMMENT("} verify_oop_addr");
2135 }
2136
2137 Address MacroAssembler::argument_address(RegisterOrConstant arg_slot,
2138 int extra_slot_offset) {
2139 // cf. TemplateTable::prepare_invoke(), if (load_receiver).
2140 int stackElementSize = Interpreter::stackElementSize;
2141 int offset = Interpreter::expr_offset_in_bytes(extra_slot_offset+0);
2142 #ifdef ASSERT
2143 int offset1 = Interpreter::expr_offset_in_bytes(extra_slot_offset+1);
2144 assert(offset1 - offset == stackElementSize, "correct arithmetic");
2145 #endif
2146 if (arg_slot.is_constant()) {
2147 return Address(esp, arg_slot.as_constant() * stackElementSize
2148 + offset);
2149 } else {
2150 add(rscratch1, esp, arg_slot.as_register(),
2151 ext::uxtx, exact_log2(stackElementSize));
2152 return Address(rscratch1, offset);
2153 }
2154 }
2155
2156 void MacroAssembler::call_VM_leaf_base(address entry_point,
2157 int number_of_arguments,
2158 Label *retaddr) {
2159 Label E, L;
2160
2161 stp(rscratch1, rmethod, Address(pre(sp, -2 * wordSize)));
2162
2163 mov(rscratch1, RuntimeAddress(entry_point));
2164 blr(rscratch1);
2165 if (retaddr)
2166 bind(*retaddr);
2167
2168 ldp(rscratch1, rmethod, Address(post(sp, 2 * wordSize)));
2169 }
2170
2171 void MacroAssembler::call_VM_leaf(address entry_point, int number_of_arguments) {
2172 call_VM_leaf_base(entry_point, number_of_arguments);
2173 }
2174
2175 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0) {
2176 pass_arg0(this, arg_0);
2177 call_VM_leaf_base(entry_point, 1);
2178 }
2179
2180 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
2181 assert_different_registers(arg_1, c_rarg0);
2182 pass_arg0(this, arg_0);
2183 pass_arg1(this, arg_1);
2184 call_VM_leaf_base(entry_point, 2);
2185 }
2186
2187 void MacroAssembler::call_VM_leaf(address entry_point, Register arg_0,
2188 Register arg_1, Register arg_2) {
2189 assert_different_registers(arg_1, c_rarg0);
2190 assert_different_registers(arg_2, c_rarg0, c_rarg1);
2191 pass_arg0(this, arg_0);
2192 pass_arg1(this, arg_1);
2193 pass_arg2(this, arg_2);
2194 call_VM_leaf_base(entry_point, 3);
2195 }
2196
2197 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0) {
2198 pass_arg0(this, arg_0);
2199 MacroAssembler::call_VM_leaf_base(entry_point, 1);
2200 }
2201
2202 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1) {
2203
2204 assert_different_registers(arg_0, c_rarg1);
2205 pass_arg1(this, arg_1);
2206 pass_arg0(this, arg_0);
2207 MacroAssembler::call_VM_leaf_base(entry_point, 2);
2208 }
2209
2210 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2) {
2211 assert_different_registers(arg_0, c_rarg1, c_rarg2);
2212 assert_different_registers(arg_1, c_rarg2);
2213 pass_arg2(this, arg_2);
2214 pass_arg1(this, arg_1);
2215 pass_arg0(this, arg_0);
2216 MacroAssembler::call_VM_leaf_base(entry_point, 3);
2217 }
2218
2219 void MacroAssembler::super_call_VM_leaf(address entry_point, Register arg_0, Register arg_1, Register arg_2, Register arg_3) {
2220 assert_different_registers(arg_0, c_rarg1, c_rarg2, c_rarg3);
2221 assert_different_registers(arg_1, c_rarg2, c_rarg3);
2222 assert_different_registers(arg_2, c_rarg3);
2223 pass_arg3(this, arg_3);
2224 pass_arg2(this, arg_2);
2225 pass_arg1(this, arg_1);
2226 pass_arg0(this, arg_0);
2227 MacroAssembler::call_VM_leaf_base(entry_point, 4);
2228 }
2229
2230 void MacroAssembler::null_check(Register reg, int offset) {
2231 if (needs_explicit_null_check(offset)) {
2232 // provoke OS null exception if reg is null by
2233 // accessing M[reg] w/o changing any registers
2234 // NOTE: this is plenty to provoke a segv
2235 ldr(zr, Address(reg));
2236 } else {
2237 // nothing to do, (later) access of M[reg + offset]
2238 // will provoke OS null exception if reg is null
2239 }
2240 }
2241
2242 // MacroAssembler protected routines needed to implement
2243 // public methods
2244
2245 void MacroAssembler::mov(Register r, Address dest) {
2246 code_section()->relocate(pc(), dest.rspec());
2247 uint64_t imm64 = (uint64_t)dest.target();
2248 movptr(r, imm64);
2249 }
2250
2251 // Move a constant pointer into r. In AArch64 mode the virtual
2252 // address space is 48 bits in size, so we only need three
2253 // instructions to create a patchable instruction sequence that can
2254 // reach anywhere.
2255 void MacroAssembler::movptr(Register r, uintptr_t imm64) {
2256 #ifndef PRODUCT
2257 {
2258 char buffer[64];
2259 snprintf(buffer, sizeof(buffer), "0x%" PRIX64, (uint64_t)imm64);
2260 block_comment(buffer);
2261 }
2262 #endif
2263 assert(imm64 < (1ull << 48), "48-bit overflow in address constant");
2264 movz(r, imm64 & 0xffff);
2265 imm64 >>= 16;
2266 movk(r, imm64 & 0xffff, 16);
2267 imm64 >>= 16;
2268 movk(r, imm64 & 0xffff, 32);
2269 }
2270
2271 // Macro to mov replicated immediate to vector register.
2272 // imm64: only the lower 8/16/32 bits are considered for B/H/S type. That is,
2273 // the upper 56/48/32 bits must be zeros for B/H/S type.
2274 // Vd will get the following values for different arrangements in T
2275 // imm64 == hex 000000gh T8B: Vd = ghghghghghghghgh
2276 // imm64 == hex 000000gh T16B: Vd = ghghghghghghghghghghghghghghghgh
2277 // imm64 == hex 0000efgh T4H: Vd = efghefghefghefgh
2278 // imm64 == hex 0000efgh T8H: Vd = efghefghefghefghefghefghefghefgh
2279 // imm64 == hex abcdefgh T2S: Vd = abcdefghabcdefgh
2280 // imm64 == hex abcdefgh T4S: Vd = abcdefghabcdefghabcdefghabcdefgh
2281 // imm64 == hex abcdefgh T1D: Vd = 00000000abcdefgh
2282 // imm64 == hex abcdefgh T2D: Vd = 00000000abcdefgh00000000abcdefgh
2283 // Clobbers rscratch1
2284 void MacroAssembler::mov(FloatRegister Vd, SIMD_Arrangement T, uint64_t imm64) {
2285 assert(T != T1Q, "unsupported");
2286 if (T == T1D || T == T2D) {
2287 int imm = operand_valid_for_movi_immediate(imm64, T);
2288 if (-1 != imm) {
2289 movi(Vd, T, imm);
2290 } else {
2291 mov(rscratch1, imm64);
2292 dup(Vd, T, rscratch1);
2293 }
2294 return;
2295 }
2296
2297 #ifdef ASSERT
2298 if (T == T8B || T == T16B) assert((imm64 & ~0xff) == 0, "extraneous bits (T8B/T16B)");
2299 if (T == T4H || T == T8H) assert((imm64 & ~0xffff) == 0, "extraneous bits (T4H/T8H)");
2300 if (T == T2S || T == T4S) assert((imm64 & ~0xffffffff) == 0, "extraneous bits (T2S/T4S)");
2301 #endif
2302 int shift = operand_valid_for_movi_immediate(imm64, T);
2303 uint32_t imm32 = imm64 & 0xffffffffULL;
2304 if (shift >= 0) {
2305 movi(Vd, T, (imm32 >> shift) & 0xff, shift);
2306 } else {
2307 movw(rscratch1, imm32);
2308 dup(Vd, T, rscratch1);
2309 }
2310 }
2311
2312 void MacroAssembler::mov_immediate64(Register dst, uint64_t imm64)
2313 {
2314 #ifndef PRODUCT
2315 {
2316 char buffer[64];
2317 snprintf(buffer, sizeof(buffer), "0x%" PRIX64, imm64);
2318 block_comment(buffer);
2319 }
2320 #endif
2321 if (operand_valid_for_logical_immediate(false, imm64)) {
2322 orr(dst, zr, imm64);
2323 } else {
2324 // we can use a combination of MOVZ or MOVN with
2325 // MOVK to build up the constant
2326 uint64_t imm_h[4];
2327 int zero_count = 0;
2328 int neg_count = 0;
2329 int i;
2330 for (i = 0; i < 4; i++) {
2331 imm_h[i] = ((imm64 >> (i * 16)) & 0xffffL);
2332 if (imm_h[i] == 0) {
2333 zero_count++;
2334 } else if (imm_h[i] == 0xffffL) {
2335 neg_count++;
2336 }
2337 }
2338 if (zero_count == 4) {
2339 // one MOVZ will do
2340 movz(dst, 0);
2341 } else if (neg_count == 4) {
2342 // one MOVN will do
2343 movn(dst, 0);
2344 } else if (zero_count == 3) {
2345 for (i = 0; i < 4; i++) {
2346 if (imm_h[i] != 0L) {
2347 movz(dst, (uint32_t)imm_h[i], (i << 4));
2348 break;
2349 }
2350 }
2351 } else if (neg_count == 3) {
2352 // one MOVN will do
2353 for (int i = 0; i < 4; i++) {
2354 if (imm_h[i] != 0xffffL) {
2355 movn(dst, (uint32_t)imm_h[i] ^ 0xffffL, (i << 4));
2356 break;
2357 }
2358 }
2359 } else if (zero_count == 2) {
2360 // one MOVZ and one MOVK will do
2361 for (i = 0; i < 3; i++) {
2362 if (imm_h[i] != 0L) {
2363 movz(dst, (uint32_t)imm_h[i], (i << 4));
2364 i++;
2365 break;
2366 }
2367 }
2368 for (;i < 4; i++) {
2369 if (imm_h[i] != 0L) {
2370 movk(dst, (uint32_t)imm_h[i], (i << 4));
2371 }
2372 }
2373 } else if (neg_count == 2) {
2374 // one MOVN and one MOVK will do
2375 for (i = 0; i < 4; i++) {
2376 if (imm_h[i] != 0xffffL) {
2377 movn(dst, (uint32_t)imm_h[i] ^ 0xffffL, (i << 4));
2378 i++;
2379 break;
2380 }
2381 }
2382 for (;i < 4; i++) {
2383 if (imm_h[i] != 0xffffL) {
2384 movk(dst, (uint32_t)imm_h[i], (i << 4));
2385 }
2386 }
2387 } else if (zero_count == 1) {
2388 // one MOVZ and two MOVKs will do
2389 for (i = 0; i < 4; i++) {
2390 if (imm_h[i] != 0L) {
2391 movz(dst, (uint32_t)imm_h[i], (i << 4));
2392 i++;
2393 break;
2394 }
2395 }
2396 for (;i < 4; i++) {
2397 if (imm_h[i] != 0x0L) {
2398 movk(dst, (uint32_t)imm_h[i], (i << 4));
2399 }
2400 }
2401 } else if (neg_count == 1) {
2402 // one MOVN and two MOVKs will do
2403 for (i = 0; i < 4; i++) {
2404 if (imm_h[i] != 0xffffL) {
2405 movn(dst, (uint32_t)imm_h[i] ^ 0xffffL, (i << 4));
2406 i++;
2407 break;
2408 }
2409 }
2410 for (;i < 4; i++) {
2411 if (imm_h[i] != 0xffffL) {
2412 movk(dst, (uint32_t)imm_h[i], (i << 4));
2413 }
2414 }
2415 } else {
2416 // use a MOVZ and 3 MOVKs (makes it easier to debug)
2417 movz(dst, (uint32_t)imm_h[0], 0);
2418 for (i = 1; i < 4; i++) {
2419 movk(dst, (uint32_t)imm_h[i], (i << 4));
2420 }
2421 }
2422 }
2423 }
2424
2425 void MacroAssembler::mov_immediate32(Register dst, uint32_t imm32)
2426 {
2427 #ifndef PRODUCT
2428 {
2429 char buffer[64];
2430 snprintf(buffer, sizeof(buffer), "0x%" PRIX32, imm32);
2431 block_comment(buffer);
2432 }
2433 #endif
2434 if (operand_valid_for_logical_immediate(true, imm32)) {
2435 orrw(dst, zr, imm32);
2436 } else {
2437 // we can use MOVZ, MOVN or two calls to MOVK to build up the
2438 // constant
2439 uint32_t imm_h[2];
2440 imm_h[0] = imm32 & 0xffff;
2441 imm_h[1] = ((imm32 >> 16) & 0xffff);
2442 if (imm_h[0] == 0) {
2443 movzw(dst, imm_h[1], 16);
2444 } else if (imm_h[0] == 0xffff) {
2445 movnw(dst, imm_h[1] ^ 0xffff, 16);
2446 } else if (imm_h[1] == 0) {
2447 movzw(dst, imm_h[0], 0);
2448 } else if (imm_h[1] == 0xffff) {
2449 movnw(dst, imm_h[0] ^ 0xffff, 0);
2450 } else {
2451 // use a MOVZ and MOVK (makes it easier to debug)
2452 movzw(dst, imm_h[0], 0);
2453 movkw(dst, imm_h[1], 16);
2454 }
2455 }
2456 }
2457
2458 // Form an address from base + offset in Rd. Rd may or may
2459 // not actually be used: you must use the Address that is returned.
2460 // It is up to you to ensure that the shift provided matches the size
2461 // of your data.
2462 Address MacroAssembler::form_address(Register Rd, Register base, int64_t byte_offset, int shift) {
2463 if (Address::offset_ok_for_immed(byte_offset, shift))
2464 // It fits; no need for any heroics
2465 return Address(base, byte_offset);
2466
2467 // Don't do anything clever with negative or misaligned offsets
2468 unsigned mask = (1 << shift) - 1;
2469 if (byte_offset < 0 || byte_offset & mask) {
2470 mov(Rd, byte_offset);
2471 add(Rd, base, Rd);
2472 return Address(Rd);
2473 }
2474
2475 // See if we can do this with two 12-bit offsets
2476 {
2477 uint64_t word_offset = byte_offset >> shift;
2478 uint64_t masked_offset = word_offset & 0xfff000;
2479 if (Address::offset_ok_for_immed(word_offset - masked_offset, 0)
2480 && Assembler::operand_valid_for_add_sub_immediate(masked_offset << shift)) {
2481 add(Rd, base, masked_offset << shift);
2482 word_offset -= masked_offset;
2483 return Address(Rd, word_offset << shift);
2484 }
2485 }
2486
2487 // Do it the hard way
2488 mov(Rd, byte_offset);
2489 add(Rd, base, Rd);
2490 return Address(Rd);
2491 }
2492
2493 int MacroAssembler::corrected_idivl(Register result, Register ra, Register rb,
2494 bool want_remainder, Register scratch)
2495 {
2496 // Full implementation of Java idiv and irem. The function
2497 // returns the (pc) offset of the div instruction - may be needed
2498 // for implicit exceptions.
2499 //
2500 // constraint : ra/rb =/= scratch
2501 // normal case
2502 //
2503 // input : ra: dividend
2504 // rb: divisor
2505 //
2506 // result: either
2507 // quotient (= ra idiv rb)
2508 // remainder (= ra irem rb)
2509
2510 assert(ra != scratch && rb != scratch, "reg cannot be scratch");
2511
2512 int idivl_offset = offset();
2513 if (! want_remainder) {
2514 sdivw(result, ra, rb);
2515 } else {
2516 sdivw(scratch, ra, rb);
2517 Assembler::msubw(result, scratch, rb, ra);
2518 }
2519
2520 return idivl_offset;
2521 }
2522
2523 int MacroAssembler::corrected_idivq(Register result, Register ra, Register rb,
2524 bool want_remainder, Register scratch)
2525 {
2526 // Full implementation of Java ldiv and lrem. The function
2527 // returns the (pc) offset of the div instruction - may be needed
2528 // for implicit exceptions.
2529 //
2530 // constraint : ra/rb =/= scratch
2531 // normal case
2532 //
2533 // input : ra: dividend
2534 // rb: divisor
2535 //
2536 // result: either
2537 // quotient (= ra idiv rb)
2538 // remainder (= ra irem rb)
2539
2540 assert(ra != scratch && rb != scratch, "reg cannot be scratch");
2541
2542 int idivq_offset = offset();
2543 if (! want_remainder) {
2544 sdiv(result, ra, rb);
2545 } else {
2546 sdiv(scratch, ra, rb);
2547 Assembler::msub(result, scratch, rb, ra);
2548 }
2549
2550 return idivq_offset;
2551 }
2552
2553 void MacroAssembler::membar(Membar_mask_bits order_constraint) {
2554 address prev = pc() - NativeMembar::instruction_size;
2555 address last = code()->last_insn();
2556 if (last != nullptr && nativeInstruction_at(last)->is_Membar() && prev == last) {
2557 NativeMembar *bar = NativeMembar_at(prev);
2558 if (AlwaysMergeDMB) {
2559 bar->set_kind(bar->get_kind() | order_constraint);
2560 BLOCK_COMMENT("merged membar(always)");
2561 return;
2562 }
2563 // Don't promote DMB ST|DMB LD to DMB (a full barrier) because
2564 // doing so would introduce a StoreLoad which the caller did not
2565 // intend
2566 if (bar->get_kind() == order_constraint
2567 || bar->get_kind() == AnyAny
2568 || order_constraint == AnyAny) {
2569 // We are merging two memory barrier instructions. On AArch64 we
2570 // can do this simply by ORing them together.
2571 bar->set_kind(bar->get_kind() | order_constraint);
2572 BLOCK_COMMENT("merged membar");
2573 return;
2574 } else {
2575 // A special case like "DMB ST;DMB LD;DMB ST", the last DMB can be skipped
2576 // We need check the last 2 instructions
2577 address prev2 = prev - NativeMembar::instruction_size;
2578 if (last != code()->last_label() && nativeInstruction_at(prev2)->is_Membar()) {
2579 NativeMembar *bar2 = NativeMembar_at(prev2);
2580 assert(bar2->get_kind() == order_constraint, "it should be merged before");
2581 BLOCK_COMMENT("merged membar(elided)");
2582 return;
2583 }
2584 }
2585 }
2586 code()->set_last_insn(pc());
2587 dmb(Assembler::barrier(order_constraint));
2588 }
2589
2590 bool MacroAssembler::try_merge_ldst(Register rt, const Address &adr, size_t size_in_bytes, bool is_store) {
2591 if (ldst_can_merge(rt, adr, size_in_bytes, is_store)) {
2592 merge_ldst(rt, adr, size_in_bytes, is_store);
2593 code()->clear_last_insn();
2594 return true;
2595 } else {
2596 assert(size_in_bytes == 8 || size_in_bytes == 4, "only 8 bytes or 4 bytes load/store is supported.");
2597 const uint64_t mask = size_in_bytes - 1;
2598 if (adr.getMode() == Address::base_plus_offset &&
2599 (adr.offset() & mask) == 0) { // only supports base_plus_offset.
2600 code()->set_last_insn(pc());
2601 }
2602 return false;
2603 }
2604 }
2605
2606 void MacroAssembler::ldr(Register Rx, const Address &adr) {
2607 // We always try to merge two adjacent loads into one ldp.
2608 if (!try_merge_ldst(Rx, adr, 8, false)) {
2609 Assembler::ldr(Rx, adr);
2610 }
2611 }
2612
2613 void MacroAssembler::ldrw(Register Rw, const Address &adr) {
2614 // We always try to merge two adjacent loads into one ldp.
2615 if (!try_merge_ldst(Rw, adr, 4, false)) {
2616 Assembler::ldrw(Rw, adr);
2617 }
2618 }
2619
2620 void MacroAssembler::str(Register Rx, const Address &adr) {
2621 // We always try to merge two adjacent stores into one stp.
2622 if (!try_merge_ldst(Rx, adr, 8, true)) {
2623 Assembler::str(Rx, adr);
2624 }
2625 }
2626
2627 void MacroAssembler::strw(Register Rw, const Address &adr) {
2628 // We always try to merge two adjacent stores into one stp.
2629 if (!try_merge_ldst(Rw, adr, 4, true)) {
2630 Assembler::strw(Rw, adr);
2631 }
2632 }
2633
2634 // MacroAssembler routines found actually to be needed
2635
2636 void MacroAssembler::push(Register src)
2637 {
2638 str(src, Address(pre(esp, -1 * wordSize)));
2639 }
2640
2641 void MacroAssembler::pop(Register dst)
2642 {
2643 ldr(dst, Address(post(esp, 1 * wordSize)));
2644 }
2645
2646 // Note: load_unsigned_short used to be called load_unsigned_word.
2647 int MacroAssembler::load_unsigned_short(Register dst, Address src) {
2648 int off = offset();
2649 ldrh(dst, src);
2650 return off;
2651 }
2652
2653 int MacroAssembler::load_unsigned_byte(Register dst, Address src) {
2654 int off = offset();
2655 ldrb(dst, src);
2656 return off;
2657 }
2658
2659 int MacroAssembler::load_signed_short(Register dst, Address src) {
2660 int off = offset();
2661 ldrsh(dst, src);
2662 return off;
2663 }
2664
2665 int MacroAssembler::load_signed_byte(Register dst, Address src) {
2666 int off = offset();
2667 ldrsb(dst, src);
2668 return off;
2669 }
2670
2671 int MacroAssembler::load_signed_short32(Register dst, Address src) {
2672 int off = offset();
2673 ldrshw(dst, src);
2674 return off;
2675 }
2676
2677 int MacroAssembler::load_signed_byte32(Register dst, Address src) {
2678 int off = offset();
2679 ldrsbw(dst, src);
2680 return off;
2681 }
2682
2683 void MacroAssembler::load_sized_value(Register dst, Address src, size_t size_in_bytes, bool is_signed) {
2684 switch (size_in_bytes) {
2685 case 8: ldr(dst, src); break;
2686 case 4: ldrw(dst, src); break;
2687 case 2: is_signed ? load_signed_short(dst, src) : load_unsigned_short(dst, src); break;
2688 case 1: is_signed ? load_signed_byte( dst, src) : load_unsigned_byte( dst, src); break;
2689 default: ShouldNotReachHere();
2690 }
2691 }
2692
2693 void MacroAssembler::store_sized_value(Address dst, Register src, size_t size_in_bytes) {
2694 switch (size_in_bytes) {
2695 case 8: str(src, dst); break;
2696 case 4: strw(src, dst); break;
2697 case 2: strh(src, dst); break;
2698 case 1: strb(src, dst); break;
2699 default: ShouldNotReachHere();
2700 }
2701 }
2702
2703 void MacroAssembler::decrementw(Register reg, int value)
2704 {
2705 if (value < 0) { incrementw(reg, -value); return; }
2706 if (value == 0) { return; }
2707 if (value < (1 << 12)) { subw(reg, reg, value); return; }
2708 /* else */ {
2709 guarantee(reg != rscratch2, "invalid dst for register decrement");
2710 movw(rscratch2, (unsigned)value);
2711 subw(reg, reg, rscratch2);
2712 }
2713 }
2714
2715 void MacroAssembler::decrement(Register reg, int value)
2716 {
2717 if (value < 0) { increment(reg, -value); return; }
2718 if (value == 0) { return; }
2719 if (value < (1 << 12)) { sub(reg, reg, value); return; }
2720 /* else */ {
2721 assert(reg != rscratch2, "invalid dst for register decrement");
2722 mov(rscratch2, (uint64_t)value);
2723 sub(reg, reg, rscratch2);
2724 }
2725 }
2726
2727 void MacroAssembler::decrementw(Address dst, int value)
2728 {
2729 assert(!dst.uses(rscratch1), "invalid dst for address decrement");
2730 if (dst.getMode() == Address::literal) {
2731 assert(abs(value) < (1 << 12), "invalid value and address mode combination");
2732 lea(rscratch2, dst);
2733 dst = Address(rscratch2);
2734 }
2735 ldrw(rscratch1, dst);
2736 decrementw(rscratch1, value);
2737 strw(rscratch1, dst);
2738 }
2739
2740 void MacroAssembler::decrement(Address dst, int value)
2741 {
2742 assert(!dst.uses(rscratch1), "invalid address for decrement");
2743 if (dst.getMode() == Address::literal) {
2744 assert(abs(value) < (1 << 12), "invalid value and address mode combination");
2745 lea(rscratch2, dst);
2746 dst = Address(rscratch2);
2747 }
2748 ldr(rscratch1, dst);
2749 decrement(rscratch1, value);
2750 str(rscratch1, dst);
2751 }
2752
2753 void MacroAssembler::incrementw(Register reg, int value)
2754 {
2755 if (value < 0) { decrementw(reg, -value); return; }
2756 if (value == 0) { return; }
2757 if (value < (1 << 12)) { addw(reg, reg, value); return; }
2758 /* else */ {
2759 assert(reg != rscratch2, "invalid dst for register increment");
2760 movw(rscratch2, (unsigned)value);
2761 addw(reg, reg, rscratch2);
2762 }
2763 }
2764
2765 void MacroAssembler::increment(Register reg, int value)
2766 {
2767 if (value < 0) { decrement(reg, -value); return; }
2768 if (value == 0) { return; }
2769 if (value < (1 << 12)) { add(reg, reg, value); return; }
2770 /* else */ {
2771 assert(reg != rscratch2, "invalid dst for register increment");
2772 movw(rscratch2, (unsigned)value);
2773 add(reg, reg, rscratch2);
2774 }
2775 }
2776
2777 void MacroAssembler::incrementw(Address dst, int value)
2778 {
2779 assert(!dst.uses(rscratch1), "invalid dst for address increment");
2780 if (dst.getMode() == Address::literal) {
2781 assert(abs(value) < (1 << 12), "invalid value and address mode combination");
2782 lea(rscratch2, dst);
2783 dst = Address(rscratch2);
2784 }
2785 ldrw(rscratch1, dst);
2786 incrementw(rscratch1, value);
2787 strw(rscratch1, dst);
2788 }
2789
2790 void MacroAssembler::increment(Address dst, int value)
2791 {
2792 assert(!dst.uses(rscratch1), "invalid dst for address increment");
2793 if (dst.getMode() == Address::literal) {
2794 assert(abs(value) < (1 << 12), "invalid value and address mode combination");
2795 lea(rscratch2, dst);
2796 dst = Address(rscratch2);
2797 }
2798 ldr(rscratch1, dst);
2799 increment(rscratch1, value);
2800 str(rscratch1, dst);
2801 }
2802
2803 // Push lots of registers in the bit set supplied. Don't push sp.
2804 // Return the number of words pushed
2805 int MacroAssembler::push(unsigned int bitset, Register stack) {
2806 int words_pushed = 0;
2807
2808 // Scan bitset to accumulate register pairs
2809 unsigned char regs[32];
2810 int count = 0;
2811 for (int reg = 0; reg <= 30; reg++) {
2812 if (1 & bitset)
2813 regs[count++] = reg;
2814 bitset >>= 1;
2815 }
2816 regs[count++] = zr->raw_encoding();
2817 count &= ~1; // Only push an even number of regs
2818
2819 if (count) {
2820 stp(as_Register(regs[0]), as_Register(regs[1]),
2821 Address(pre(stack, -count * wordSize)));
2822 words_pushed += 2;
2823 }
2824 for (int i = 2; i < count; i += 2) {
2825 stp(as_Register(regs[i]), as_Register(regs[i+1]),
2826 Address(stack, i * wordSize));
2827 words_pushed += 2;
2828 }
2829
2830 assert(words_pushed == count, "oops, pushed != count");
2831
2832 return count;
2833 }
2834
2835 int MacroAssembler::pop(unsigned int bitset, Register stack) {
2836 int words_pushed = 0;
2837
2838 // Scan bitset to accumulate register pairs
2839 unsigned char regs[32];
2840 int count = 0;
2841 for (int reg = 0; reg <= 30; reg++) {
2842 if (1 & bitset)
2843 regs[count++] = reg;
2844 bitset >>= 1;
2845 }
2846 regs[count++] = zr->raw_encoding();
2847 count &= ~1;
2848
2849 for (int i = 2; i < count; i += 2) {
2850 ldp(as_Register(regs[i]), as_Register(regs[i+1]),
2851 Address(stack, i * wordSize));
2852 words_pushed += 2;
2853 }
2854 if (count) {
2855 ldp(as_Register(regs[0]), as_Register(regs[1]),
2856 Address(post(stack, count * wordSize)));
2857 words_pushed += 2;
2858 }
2859
2860 assert(words_pushed == count, "oops, pushed != count");
2861
2862 return count;
2863 }
2864
2865 // Push lots of registers in the bit set supplied. Don't push sp.
2866 // Return the number of dwords pushed
2867 int MacroAssembler::push_fp(unsigned int bitset, Register stack, FpPushPopMode mode) {
2868 int words_pushed = 0;
2869 bool use_sve = false;
2870 int sve_vector_size_in_bytes = 0;
2871
2872 #ifdef COMPILER2
2873 use_sve = Matcher::supports_scalable_vector();
2874 sve_vector_size_in_bytes = Matcher::scalable_vector_reg_size(T_BYTE);
2875 #endif
2876
2877 // Scan bitset to accumulate register pairs
2878 unsigned char regs[32];
2879 int count = 0;
2880 for (int reg = 0; reg <= 31; reg++) {
2881 if (1 & bitset)
2882 regs[count++] = reg;
2883 bitset >>= 1;
2884 }
2885
2886 if (count == 0) {
2887 return 0;
2888 }
2889
2890 if (mode == PushPopFull) {
2891 if (use_sve && sve_vector_size_in_bytes > 16) {
2892 mode = PushPopSVE;
2893 } else {
2894 mode = PushPopNeon;
2895 }
2896 }
2897
2898 #ifndef PRODUCT
2899 {
2900 char buffer[48];
2901 if (mode == PushPopSVE) {
2902 snprintf(buffer, sizeof(buffer), "push_fp: %d SVE registers", count);
2903 } else if (mode == PushPopNeon) {
2904 snprintf(buffer, sizeof(buffer), "push_fp: %d Neon registers", count);
2905 } else {
2906 snprintf(buffer, sizeof(buffer), "push_fp: %d fp registers", count);
2907 }
2908 block_comment(buffer);
2909 }
2910 #endif
2911
2912 if (mode == PushPopSVE) {
2913 sub(stack, stack, sve_vector_size_in_bytes * count);
2914 for (int i = 0; i < count; i++) {
2915 sve_str(as_FloatRegister(regs[i]), Address(stack, i));
2916 }
2917 return count * sve_vector_size_in_bytes / 8;
2918 }
2919
2920 if (mode == PushPopNeon) {
2921 if (count == 1) {
2922 strq(as_FloatRegister(regs[0]), Address(pre(stack, -wordSize * 2)));
2923 return 2;
2924 }
2925
2926 bool odd = (count & 1) == 1;
2927 int push_slots = count + (odd ? 1 : 0);
2928
2929 // Always pushing full 128 bit registers.
2930 stpq(as_FloatRegister(regs[0]), as_FloatRegister(regs[1]), Address(pre(stack, -push_slots * wordSize * 2)));
2931 words_pushed += 2;
2932
2933 for (int i = 2; i + 1 < count; i += 2) {
2934 stpq(as_FloatRegister(regs[i]), as_FloatRegister(regs[i+1]), Address(stack, i * wordSize * 2));
2935 words_pushed += 2;
2936 }
2937
2938 if (odd) {
2939 strq(as_FloatRegister(regs[count - 1]), Address(stack, (count - 1) * wordSize * 2));
2940 words_pushed++;
2941 }
2942
2943 assert(words_pushed == count, "oops, pushed(%d) != count(%d)", words_pushed, count);
2944 return count * 2;
2945 }
2946
2947 if (mode == PushPopFp) {
2948 bool odd = (count & 1) == 1;
2949 int push_slots = count + (odd ? 1 : 0);
2950
2951 if (count == 1) {
2952 // Stack pointer must be 16 bytes aligned
2953 strd(as_FloatRegister(regs[0]), Address(pre(stack, -push_slots * wordSize)));
2954 return 1;
2955 }
2956
2957 stpd(as_FloatRegister(regs[0]), as_FloatRegister(regs[1]), Address(pre(stack, -push_slots * wordSize)));
2958 words_pushed += 2;
2959
2960 for (int i = 2; i + 1 < count; i += 2) {
2961 stpd(as_FloatRegister(regs[i]), as_FloatRegister(regs[i+1]), Address(stack, i * wordSize));
2962 words_pushed += 2;
2963 }
2964
2965 if (odd) {
2966 // Stack pointer must be 16 bytes aligned
2967 strd(as_FloatRegister(regs[count - 1]), Address(stack, (count - 1) * wordSize));
2968 words_pushed++;
2969 }
2970
2971 assert(words_pushed == count, "oops, pushed != count");
2972
2973 return count;
2974 }
2975
2976 return 0;
2977 }
2978
2979 // Return the number of dwords popped
2980 int MacroAssembler::pop_fp(unsigned int bitset, Register stack, FpPushPopMode mode) {
2981 int words_pushed = 0;
2982 bool use_sve = false;
2983 int sve_vector_size_in_bytes = 0;
2984
2985 #ifdef COMPILER2
2986 use_sve = Matcher::supports_scalable_vector();
2987 sve_vector_size_in_bytes = Matcher::scalable_vector_reg_size(T_BYTE);
2988 #endif
2989 // Scan bitset to accumulate register pairs
2990 unsigned char regs[32];
2991 int count = 0;
2992 for (int reg = 0; reg <= 31; reg++) {
2993 if (1 & bitset)
2994 regs[count++] = reg;
2995 bitset >>= 1;
2996 }
2997
2998 if (count == 0) {
2999 return 0;
3000 }
3001
3002 if (mode == PushPopFull) {
3003 if (use_sve && sve_vector_size_in_bytes > 16) {
3004 mode = PushPopSVE;
3005 } else {
3006 mode = PushPopNeon;
3007 }
3008 }
3009
3010 #ifndef PRODUCT
3011 {
3012 char buffer[48];
3013 if (mode == PushPopSVE) {
3014 snprintf(buffer, sizeof(buffer), "pop_fp: %d SVE registers", count);
3015 } else if (mode == PushPopNeon) {
3016 snprintf(buffer, sizeof(buffer), "pop_fp: %d Neon registers", count);
3017 } else {
3018 snprintf(buffer, sizeof(buffer), "pop_fp: %d fp registers", count);
3019 }
3020 block_comment(buffer);
3021 }
3022 #endif
3023
3024 if (mode == PushPopSVE) {
3025 for (int i = count - 1; i >= 0; i--) {
3026 sve_ldr(as_FloatRegister(regs[i]), Address(stack, i));
3027 }
3028 add(stack, stack, sve_vector_size_in_bytes * count);
3029 return count * sve_vector_size_in_bytes / 8;
3030 }
3031
3032 if (mode == PushPopNeon) {
3033 if (count == 1) {
3034 ldrq(as_FloatRegister(regs[0]), Address(post(stack, wordSize * 2)));
3035 return 2;
3036 }
3037
3038 bool odd = (count & 1) == 1;
3039 int push_slots = count + (odd ? 1 : 0);
3040
3041 if (odd) {
3042 ldrq(as_FloatRegister(regs[count - 1]), Address(stack, (count - 1) * wordSize * 2));
3043 words_pushed++;
3044 }
3045
3046 for (int i = 2; i + 1 < count; i += 2) {
3047 ldpq(as_FloatRegister(regs[i]), as_FloatRegister(regs[i+1]), Address(stack, i * wordSize * 2));
3048 words_pushed += 2;
3049 }
3050
3051 ldpq(as_FloatRegister(regs[0]), as_FloatRegister(regs[1]), Address(post(stack, push_slots * wordSize * 2)));
3052 words_pushed += 2;
3053
3054 assert(words_pushed == count, "oops, pushed(%d) != count(%d)", words_pushed, count);
3055
3056 return count * 2;
3057 }
3058
3059 if (mode == PushPopFp) {
3060 bool odd = (count & 1) == 1;
3061 int push_slots = count + (odd ? 1 : 0);
3062
3063 if (count == 1) {
3064 ldrd(as_FloatRegister(regs[0]), Address(post(stack, push_slots * wordSize)));
3065 return 1;
3066 }
3067
3068 if (odd) {
3069 ldrd(as_FloatRegister(regs[count - 1]), Address(stack, (count - 1) * wordSize));
3070 words_pushed++;
3071 }
3072
3073 for (int i = 2; i + 1 < count; i += 2) {
3074 ldpd(as_FloatRegister(regs[i]), as_FloatRegister(regs[i+1]), Address(stack, i * wordSize));
3075 words_pushed += 2;
3076 }
3077
3078 ldpd(as_FloatRegister(regs[0]), as_FloatRegister(regs[1]), Address(post(stack, push_slots * wordSize)));
3079 words_pushed += 2;
3080
3081 assert(words_pushed == count, "oops, pushed != count");
3082
3083 return count;
3084 }
3085
3086 return 0;
3087 }
3088
3089 // Return the number of dwords pushed
3090 int MacroAssembler::push_p(unsigned int bitset, Register stack) {
3091 bool use_sve = false;
3092 int sve_predicate_size_in_slots = 0;
3093
3094 #ifdef COMPILER2
3095 use_sve = Matcher::supports_scalable_vector();
3096 if (use_sve) {
3097 sve_predicate_size_in_slots = Matcher::scalable_predicate_reg_slots();
3098 }
3099 #endif
3100
3101 if (!use_sve) {
3102 return 0;
3103 }
3104
3105 unsigned char regs[PRegister::number_of_registers];
3106 int count = 0;
3107 for (int reg = 0; reg < PRegister::number_of_registers; reg++) {
3108 if (1 & bitset)
3109 regs[count++] = reg;
3110 bitset >>= 1;
3111 }
3112
3113 if (count == 0) {
3114 return 0;
3115 }
3116
3117 int total_push_bytes = align_up(sve_predicate_size_in_slots *
3118 VMRegImpl::stack_slot_size * count, 16);
3119 sub(stack, stack, total_push_bytes);
3120 for (int i = 0; i < count; i++) {
3121 sve_str(as_PRegister(regs[i]), Address(stack, i));
3122 }
3123 return total_push_bytes / 8;
3124 }
3125
3126 // Return the number of dwords popped
3127 int MacroAssembler::pop_p(unsigned int bitset, Register stack) {
3128 bool use_sve = false;
3129 int sve_predicate_size_in_slots = 0;
3130
3131 #ifdef COMPILER2
3132 use_sve = Matcher::supports_scalable_vector();
3133 if (use_sve) {
3134 sve_predicate_size_in_slots = Matcher::scalable_predicate_reg_slots();
3135 }
3136 #endif
3137
3138 if (!use_sve) {
3139 return 0;
3140 }
3141
3142 unsigned char regs[PRegister::number_of_registers];
3143 int count = 0;
3144 for (int reg = 0; reg < PRegister::number_of_registers; reg++) {
3145 if (1 & bitset)
3146 regs[count++] = reg;
3147 bitset >>= 1;
3148 }
3149
3150 if (count == 0) {
3151 return 0;
3152 }
3153
3154 int total_pop_bytes = align_up(sve_predicate_size_in_slots *
3155 VMRegImpl::stack_slot_size * count, 16);
3156 for (int i = count - 1; i >= 0; i--) {
3157 sve_ldr(as_PRegister(regs[i]), Address(stack, i));
3158 }
3159 add(stack, stack, total_pop_bytes);
3160 return total_pop_bytes / 8;
3161 }
3162
3163 #ifdef ASSERT
3164 void MacroAssembler::verify_heapbase(const char* msg) {
3165 #if 0
3166 assert (UseCompressedOops || UseCompressedClassPointers, "should be compressed");
3167 assert (Universe::heap() != nullptr, "java heap should be initialized");
3168 if (!UseCompressedOops || Universe::ptr_base() == nullptr) {
3169 // rheapbase is allocated as general register
3170 return;
3171 }
3172 if (CheckCompressedOops) {
3173 Label ok;
3174 push(1 << rscratch1->encoding(), sp); // cmpptr trashes rscratch1
3175 cmpptr(rheapbase, ExternalAddress(CompressedOops::base_addr()));
3176 br(Assembler::EQ, ok);
3177 stop(msg);
3178 bind(ok);
3179 pop(1 << rscratch1->encoding(), sp);
3180 }
3181 #endif
3182 }
3183 #endif
3184
3185 void MacroAssembler::resolve_jobject(Register value, Register tmp1, Register tmp2) {
3186 assert_different_registers(value, tmp1, tmp2);
3187 Label done, tagged, weak_tagged;
3188
3189 cbz(value, done); // Use null as-is.
3190 tst(value, JNIHandles::tag_mask); // Test for tag.
3191 br(Assembler::NE, tagged);
3192
3193 // Resolve local handle
3194 access_load_at(T_OBJECT, IN_NATIVE | AS_RAW, value, Address(value, 0), tmp1, tmp2);
3195 verify_oop(value);
3196 b(done);
3197
3198 bind(tagged);
3199 STATIC_ASSERT(JNIHandles::TypeTag::weak_global == 0b1);
3200 tbnz(value, 0, weak_tagged); // Test for weak tag.
3201
3202 // Resolve global handle
3203 access_load_at(T_OBJECT, IN_NATIVE, value, Address(value, -JNIHandles::TypeTag::global), tmp1, tmp2);
3204 verify_oop(value);
3205 b(done);
3206
3207 bind(weak_tagged);
3208 // Resolve jweak.
3209 access_load_at(T_OBJECT, IN_NATIVE | ON_PHANTOM_OOP_REF,
3210 value, Address(value, -JNIHandles::TypeTag::weak_global), tmp1, tmp2);
3211 verify_oop(value);
3212
3213 bind(done);
3214 }
3215
3216 void MacroAssembler::resolve_global_jobject(Register value, Register tmp1, Register tmp2) {
3217 assert_different_registers(value, tmp1, tmp2);
3218 Label done;
3219
3220 cbz(value, done); // Use null as-is.
3221
3222 #ifdef ASSERT
3223 {
3224 STATIC_ASSERT(JNIHandles::TypeTag::global == 0b10);
3225 Label valid_global_tag;
3226 tbnz(value, 1, valid_global_tag); // Test for global tag
3227 stop("non global jobject using resolve_global_jobject");
3228 bind(valid_global_tag);
3229 }
3230 #endif
3231
3232 // Resolve global handle
3233 access_load_at(T_OBJECT, IN_NATIVE, value, Address(value, -JNIHandles::TypeTag::global), tmp1, tmp2);
3234 verify_oop(value);
3235
3236 bind(done);
3237 }
3238
3239 void MacroAssembler::stop(const char* msg) {
3240 // Skip AOT caching C strings in scratch buffer.
3241 const char* str = (code_section()->scratch_emit()) ? msg : AOTCodeCache::add_C_string(msg);
3242 BLOCK_COMMENT(str);
3243 // load msg into r0 so we can access it from the signal handler
3244 // ExternalAddress enables saving and restoring via the code cache
3245 lea(c_rarg0, ExternalAddress((address) str));
3246 dcps1(0xdeae);
3247 }
3248
3249 void MacroAssembler::unimplemented(const char* what) {
3250 const char* buf = nullptr;
3251 {
3252 ResourceMark rm;
3253 stringStream ss;
3254 ss.print("unimplemented: %s", what);
3255 buf = code_string(ss.as_string());
3256 }
3257 stop(buf);
3258 }
3259
3260 void MacroAssembler::_assert_asm(Assembler::Condition cc, const char* msg) {
3261 #ifdef ASSERT
3262 Label OK;
3263 br(cc, OK);
3264 stop(msg);
3265 bind(OK);
3266 #endif
3267 }
3268
3269 // If a constant does not fit in an immediate field, generate some
3270 // number of MOV instructions and then perform the operation.
3271 void MacroAssembler::wrap_add_sub_imm_insn(Register Rd, Register Rn, uint64_t imm,
3272 add_sub_imm_insn insn1,
3273 add_sub_reg_insn insn2,
3274 bool is32) {
3275 assert(Rd != zr, "Rd = zr and not setting flags?");
3276 bool fits = operand_valid_for_add_sub_immediate(is32 ? (int32_t)imm : imm);
3277 if (fits) {
3278 (this->*insn1)(Rd, Rn, imm);
3279 } else {
3280 if (g_uabs(imm) < (1 << 24)) {
3281 (this->*insn1)(Rd, Rn, imm & -(1 << 12));
3282 (this->*insn1)(Rd, Rd, imm & ((1 << 12)-1));
3283 } else {
3284 assert_different_registers(Rd, Rn);
3285 mov(Rd, imm);
3286 (this->*insn2)(Rd, Rn, Rd, LSL, 0);
3287 }
3288 }
3289 }
3290
3291 // Separate vsn which sets the flags. Optimisations are more restricted
3292 // because we must set the flags correctly.
3293 void MacroAssembler::wrap_adds_subs_imm_insn(Register Rd, Register Rn, uint64_t imm,
3294 add_sub_imm_insn insn1,
3295 add_sub_reg_insn insn2,
3296 bool is32) {
3297 bool fits = operand_valid_for_add_sub_immediate(is32 ? (int32_t)imm : imm);
3298 if (fits) {
3299 (this->*insn1)(Rd, Rn, imm);
3300 } else {
3301 assert_different_registers(Rd, Rn);
3302 assert(Rd != zr, "overflow in immediate operand");
3303 mov(Rd, imm);
3304 (this->*insn2)(Rd, Rn, Rd, LSL, 0);
3305 }
3306 }
3307
3308
3309 void MacroAssembler::add(Register Rd, Register Rn, RegisterOrConstant increment) {
3310 if (increment.is_register()) {
3311 add(Rd, Rn, increment.as_register());
3312 } else {
3313 add(Rd, Rn, increment.as_constant());
3314 }
3315 }
3316
3317 void MacroAssembler::addw(Register Rd, Register Rn, RegisterOrConstant increment) {
3318 if (increment.is_register()) {
3319 addw(Rd, Rn, increment.as_register());
3320 } else {
3321 addw(Rd, Rn, increment.as_constant());
3322 }
3323 }
3324
3325 void MacroAssembler::sub(Register Rd, Register Rn, RegisterOrConstant decrement) {
3326 if (decrement.is_register()) {
3327 sub(Rd, Rn, decrement.as_register());
3328 } else {
3329 sub(Rd, Rn, decrement.as_constant());
3330 }
3331 }
3332
3333 void MacroAssembler::subw(Register Rd, Register Rn, RegisterOrConstant decrement) {
3334 if (decrement.is_register()) {
3335 subw(Rd, Rn, decrement.as_register());
3336 } else {
3337 subw(Rd, Rn, decrement.as_constant());
3338 }
3339 }
3340
3341 void MacroAssembler::reinit_heapbase()
3342 {
3343 if (UseCompressedOops) {
3344 if (Universe::is_fully_initialized()) {
3345 mov(rheapbase, CompressedOops::base());
3346 } else {
3347 lea(rheapbase, ExternalAddress(CompressedOops::base_addr()));
3348 ldr(rheapbase, Address(rheapbase));
3349 }
3350 }
3351 }
3352
3353 // this simulates the behaviour of the x86 cmpxchg instruction using a
3354 // load linked/store conditional pair. we use the acquire/release
3355 // versions of these instructions so that we flush pending writes as
3356 // per Java semantics.
3357
3358 // n.b the x86 version assumes the old value to be compared against is
3359 // in rax and updates rax with the value located in memory if the
3360 // cmpxchg fails. we supply a register for the old value explicitly
3361
3362 // the aarch64 load linked/store conditional instructions do not
3363 // accept an offset. so, unlike x86, we must provide a plain register
3364 // to identify the memory word to be compared/exchanged rather than a
3365 // register+offset Address.
3366
3367 void MacroAssembler::cmpxchgptr(Register oldv, Register newv, Register addr, Register tmp,
3368 Label &succeed, Label *fail) {
3369 // oldv holds comparison value
3370 // newv holds value to write in exchange
3371 // addr identifies memory word to compare against/update
3372 if (UseLSE) {
3373 mov(tmp, oldv);
3374 casal(Assembler::xword, oldv, newv, addr);
3375 cmp(tmp, oldv);
3376 br(Assembler::EQ, succeed);
3377 membar(AnyAny);
3378 } else {
3379 Label retry_load, nope;
3380 prfm(Address(addr), PSTL1STRM);
3381 bind(retry_load);
3382 // flush and load exclusive from the memory location
3383 // and fail if it is not what we expect
3384 ldaxr(tmp, addr);
3385 cmp(tmp, oldv);
3386 br(Assembler::NE, nope);
3387 // if we store+flush with no intervening write tmp will be zero
3388 stlxr(tmp, newv, addr);
3389 cbzw(tmp, succeed);
3390 // retry so we only ever return after a load fails to compare
3391 // ensures we don't return a stale value after a failed write.
3392 b(retry_load);
3393 // if the memory word differs we return it in oldv and signal a fail
3394 bind(nope);
3395 membar(AnyAny);
3396 mov(oldv, tmp);
3397 }
3398 if (fail)
3399 b(*fail);
3400 }
3401
3402 void MacroAssembler::cmpxchg_obj_header(Register oldv, Register newv, Register obj, Register tmp,
3403 Label &succeed, Label *fail) {
3404 assert(oopDesc::mark_offset_in_bytes() == 0, "assumption");
3405 cmpxchgptr(oldv, newv, obj, tmp, succeed, fail);
3406 }
3407
3408 void MacroAssembler::cmpxchgw(Register oldv, Register newv, Register addr, Register tmp,
3409 Label &succeed, Label *fail) {
3410 // oldv holds comparison value
3411 // newv holds value to write in exchange
3412 // addr identifies memory word to compare against/update
3413 // tmp returns 0/1 for success/failure
3414 if (UseLSE) {
3415 mov(tmp, oldv);
3416 casal(Assembler::word, oldv, newv, addr);
3417 cmp(tmp, oldv);
3418 br(Assembler::EQ, succeed);
3419 membar(AnyAny);
3420 } else {
3421 Label retry_load, nope;
3422 prfm(Address(addr), PSTL1STRM);
3423 bind(retry_load);
3424 // flush and load exclusive from the memory location
3425 // and fail if it is not what we expect
3426 ldaxrw(tmp, addr);
3427 cmp(tmp, oldv);
3428 br(Assembler::NE, nope);
3429 // if we store+flush with no intervening write tmp will be zero
3430 stlxrw(tmp, newv, addr);
3431 cbzw(tmp, succeed);
3432 // retry so we only ever return after a load fails to compare
3433 // ensures we don't return a stale value after a failed write.
3434 b(retry_load);
3435 // if the memory word differs we return it in oldv and signal a fail
3436 bind(nope);
3437 membar(AnyAny);
3438 mov(oldv, tmp);
3439 }
3440 if (fail)
3441 b(*fail);
3442 }
3443
3444 // A generic CAS; success or failure is in the EQ flag. A weak CAS
3445 // doesn't retry and may fail spuriously. If the oldval is wanted,
3446 // Pass a register for the result, otherwise pass noreg.
3447
3448 // Clobbers rscratch1
3449 void MacroAssembler::cmpxchg(Register addr, Register expected,
3450 Register new_val,
3451 enum operand_size size,
3452 bool acquire, bool release,
3453 bool weak,
3454 Register result) {
3455 if (result == noreg) result = rscratch1;
3456 BLOCK_COMMENT("cmpxchg {");
3457 if (UseLSE) {
3458 mov(result, expected);
3459 lse_cas(result, new_val, addr, size, acquire, release, /*not_pair*/ true);
3460 compare_eq(result, expected, size);
3461 #ifdef ASSERT
3462 // Poison rscratch1 which is written on !UseLSE branch
3463 mov(rscratch1, 0x1f1f1f1f1f1f1f1f);
3464 #endif
3465 } else {
3466 Label retry_load, done;
3467 prfm(Address(addr), PSTL1STRM);
3468 bind(retry_load);
3469 load_exclusive(result, addr, size, acquire);
3470 compare_eq(result, expected, size);
3471 br(Assembler::NE, done);
3472 store_exclusive(rscratch1, new_val, addr, size, release);
3473 if (weak) {
3474 cmpw(rscratch1, 0u); // If the store fails, return NE to our caller.
3475 } else {
3476 cbnzw(rscratch1, retry_load);
3477 }
3478 bind(done);
3479 }
3480 BLOCK_COMMENT("} cmpxchg");
3481 }
3482
3483 // A generic comparison. Only compares for equality, clobbers rscratch1.
3484 void MacroAssembler::compare_eq(Register rm, Register rn, enum operand_size size) {
3485 if (size == xword) {
3486 cmp(rm, rn);
3487 } else if (size == word) {
3488 cmpw(rm, rn);
3489 } else if (size == halfword) {
3490 eorw(rscratch1, rm, rn);
3491 ands(zr, rscratch1, 0xffff);
3492 } else if (size == byte) {
3493 eorw(rscratch1, rm, rn);
3494 ands(zr, rscratch1, 0xff);
3495 } else {
3496 ShouldNotReachHere();
3497 }
3498 }
3499
3500
3501 static bool different(Register a, RegisterOrConstant b, Register c) {
3502 if (b.is_constant())
3503 return a != c;
3504 else
3505 return a != b.as_register() && a != c && b.as_register() != c;
3506 }
3507
3508 #define ATOMIC_OP(NAME, LDXR, OP, IOP, AOP, STXR, sz) \
3509 void MacroAssembler::atomic_##NAME(Register prev, RegisterOrConstant incr, Register addr) { \
3510 if (UseLSE) { \
3511 prev = prev->is_valid() ? prev : zr; \
3512 if (incr.is_register()) { \
3513 AOP(sz, incr.as_register(), prev, addr); \
3514 } else { \
3515 mov(rscratch2, incr.as_constant()); \
3516 AOP(sz, rscratch2, prev, addr); \
3517 } \
3518 return; \
3519 } \
3520 Register result = rscratch2; \
3521 if (prev->is_valid()) \
3522 result = different(prev, incr, addr) ? prev : rscratch2; \
3523 \
3524 Label retry_load; \
3525 prfm(Address(addr), PSTL1STRM); \
3526 bind(retry_load); \
3527 LDXR(result, addr); \
3528 OP(rscratch1, result, incr); \
3529 STXR(rscratch2, rscratch1, addr); \
3530 cbnzw(rscratch2, retry_load); \
3531 if (prev->is_valid() && prev != result) { \
3532 IOP(prev, rscratch1, incr); \
3533 } \
3534 }
3535
3536 ATOMIC_OP(add, ldxr, add, sub, ldadd, stxr, Assembler::xword)
3537 ATOMIC_OP(addw, ldxrw, addw, subw, ldadd, stxrw, Assembler::word)
3538 ATOMIC_OP(addal, ldaxr, add, sub, ldaddal, stlxr, Assembler::xword)
3539 ATOMIC_OP(addalw, ldaxrw, addw, subw, ldaddal, stlxrw, Assembler::word)
3540
3541 #undef ATOMIC_OP
3542
3543 #define ATOMIC_XCHG(OP, AOP, LDXR, STXR, sz) \
3544 void MacroAssembler::atomic_##OP(Register prev, Register newv, Register addr) { \
3545 if (UseLSE) { \
3546 prev = prev->is_valid() ? prev : zr; \
3547 AOP(sz, newv, prev, addr); \
3548 return; \
3549 } \
3550 Register result = rscratch2; \
3551 if (prev->is_valid()) \
3552 result = different(prev, newv, addr) ? prev : rscratch2; \
3553 \
3554 Label retry_load; \
3555 prfm(Address(addr), PSTL1STRM); \
3556 bind(retry_load); \
3557 LDXR(result, addr); \
3558 STXR(rscratch1, newv, addr); \
3559 cbnzw(rscratch1, retry_load); \
3560 if (prev->is_valid() && prev != result) \
3561 mov(prev, result); \
3562 }
3563
3564 ATOMIC_XCHG(xchg, swp, ldxr, stxr, Assembler::xword)
3565 ATOMIC_XCHG(xchgw, swp, ldxrw, stxrw, Assembler::word)
3566 ATOMIC_XCHG(xchgl, swpl, ldxr, stlxr, Assembler::xword)
3567 ATOMIC_XCHG(xchglw, swpl, ldxrw, stlxrw, Assembler::word)
3568 ATOMIC_XCHG(xchgal, swpal, ldaxr, stlxr, Assembler::xword)
3569 ATOMIC_XCHG(xchgalw, swpal, ldaxrw, stlxrw, Assembler::word)
3570
3571 #undef ATOMIC_XCHG
3572
3573 #ifndef PRODUCT
3574 extern "C" void findpc(intptr_t x);
3575 #endif
3576
3577 void MacroAssembler::debug64(char* msg, int64_t pc, int64_t regs[])
3578 {
3579 // In order to get locks to work, we need to fake a in_VM state
3580 if (ShowMessageBoxOnError ) {
3581 JavaThread* thread = JavaThread::current();
3582 JavaThreadState saved_state = thread->thread_state();
3583 thread->set_thread_state(_thread_in_vm);
3584 #ifndef PRODUCT
3585 if (CountBytecodes || TraceBytecodes || StopInterpreterAt) {
3586 ttyLocker ttyl;
3587 BytecodeCounter::print();
3588 }
3589 #endif
3590 if (os::message_box(msg, "Execution stopped, print registers?")) {
3591 ttyLocker ttyl;
3592 tty->print_cr(" pc = 0x%016" PRIx64, pc);
3593 #ifndef PRODUCT
3594 tty->cr();
3595 findpc(pc);
3596 tty->cr();
3597 #endif
3598 tty->print_cr(" r0 = 0x%016" PRIx64, regs[0]);
3599 tty->print_cr(" r1 = 0x%016" PRIx64, regs[1]);
3600 tty->print_cr(" r2 = 0x%016" PRIx64, regs[2]);
3601 tty->print_cr(" r3 = 0x%016" PRIx64, regs[3]);
3602 tty->print_cr(" r4 = 0x%016" PRIx64, regs[4]);
3603 tty->print_cr(" r5 = 0x%016" PRIx64, regs[5]);
3604 tty->print_cr(" r6 = 0x%016" PRIx64, regs[6]);
3605 tty->print_cr(" r7 = 0x%016" PRIx64, regs[7]);
3606 tty->print_cr(" r8 = 0x%016" PRIx64, regs[8]);
3607 tty->print_cr(" r9 = 0x%016" PRIx64, regs[9]);
3608 tty->print_cr("r10 = 0x%016" PRIx64, regs[10]);
3609 tty->print_cr("r11 = 0x%016" PRIx64, regs[11]);
3610 tty->print_cr("r12 = 0x%016" PRIx64, regs[12]);
3611 tty->print_cr("r13 = 0x%016" PRIx64, regs[13]);
3612 tty->print_cr("r14 = 0x%016" PRIx64, regs[14]);
3613 tty->print_cr("r15 = 0x%016" PRIx64, regs[15]);
3614 tty->print_cr("r16 = 0x%016" PRIx64, regs[16]);
3615 tty->print_cr("r17 = 0x%016" PRIx64, regs[17]);
3616 tty->print_cr("r18 = 0x%016" PRIx64, regs[18]);
3617 tty->print_cr("r19 = 0x%016" PRIx64, regs[19]);
3618 tty->print_cr("r20 = 0x%016" PRIx64, regs[20]);
3619 tty->print_cr("r21 = 0x%016" PRIx64, regs[21]);
3620 tty->print_cr("r22 = 0x%016" PRIx64, regs[22]);
3621 tty->print_cr("r23 = 0x%016" PRIx64, regs[23]);
3622 tty->print_cr("r24 = 0x%016" PRIx64, regs[24]);
3623 tty->print_cr("r25 = 0x%016" PRIx64, regs[25]);
3624 tty->print_cr("r26 = 0x%016" PRIx64, regs[26]);
3625 tty->print_cr("r27 = 0x%016" PRIx64, regs[27]);
3626 tty->print_cr("r28 = 0x%016" PRIx64, regs[28]);
3627 tty->print_cr("r30 = 0x%016" PRIx64, regs[30]);
3628 tty->print_cr("r31 = 0x%016" PRIx64, regs[31]);
3629 BREAKPOINT;
3630 }
3631 }
3632 fatal("DEBUG MESSAGE: %s", msg);
3633 }
3634
3635 RegSet MacroAssembler::call_clobbered_gp_registers() {
3636 RegSet regs = RegSet::range(r0, r17) - RegSet::of(rscratch1, rscratch2);
3637 #ifndef R18_RESERVED
3638 regs += r18_tls;
3639 #endif
3640 return regs;
3641 }
3642
3643 void MacroAssembler::push_call_clobbered_registers_except(RegSet exclude) {
3644 int step = 4 * wordSize;
3645 push(call_clobbered_gp_registers() - exclude, sp);
3646 sub(sp, sp, step);
3647 mov(rscratch1, -step);
3648 // Push v0-v7, v16-v31.
3649 for (int i = 31; i>= 4; i -= 4) {
3650 if (i <= v7->encoding() || i >= v16->encoding())
3651 st1(as_FloatRegister(i-3), as_FloatRegister(i-2), as_FloatRegister(i-1),
3652 as_FloatRegister(i), T1D, Address(post(sp, rscratch1)));
3653 }
3654 st1(as_FloatRegister(0), as_FloatRegister(1), as_FloatRegister(2),
3655 as_FloatRegister(3), T1D, Address(sp));
3656 }
3657
3658 void MacroAssembler::pop_call_clobbered_registers_except(RegSet exclude) {
3659 for (int i = 0; i < 32; i += 4) {
3660 if (i <= v7->encoding() || i >= v16->encoding())
3661 ld1(as_FloatRegister(i), as_FloatRegister(i+1), as_FloatRegister(i+2),
3662 as_FloatRegister(i+3), T1D, Address(post(sp, 4 * wordSize)));
3663 }
3664
3665 reinitialize_ptrue();
3666
3667 pop(call_clobbered_gp_registers() - exclude, sp);
3668 }
3669
3670 void MacroAssembler::push_CPU_state(bool save_vectors, bool use_sve,
3671 int sve_vector_size_in_bytes, int total_predicate_in_bytes) {
3672 push(RegSet::range(r0, r29), sp); // integer registers except lr & sp
3673 if (save_vectors && use_sve && sve_vector_size_in_bytes > 16) {
3674 sub(sp, sp, sve_vector_size_in_bytes * FloatRegister::number_of_registers);
3675 for (int i = 0; i < FloatRegister::number_of_registers; i++) {
3676 sve_str(as_FloatRegister(i), Address(sp, i));
3677 }
3678 } else {
3679 int step = (save_vectors ? 8 : 4) * wordSize;
3680 mov(rscratch1, -step);
3681 sub(sp, sp, step);
3682 for (int i = 28; i >= 4; i -= 4) {
3683 st1(as_FloatRegister(i), as_FloatRegister(i+1), as_FloatRegister(i+2),
3684 as_FloatRegister(i+3), save_vectors ? T2D : T1D, Address(post(sp, rscratch1)));
3685 }
3686 st1(v0, v1, v2, v3, save_vectors ? T2D : T1D, sp);
3687 }
3688 if (save_vectors && use_sve && total_predicate_in_bytes > 0) {
3689 sub(sp, sp, total_predicate_in_bytes);
3690 for (int i = 0; i < PRegister::number_of_registers; i++) {
3691 sve_str(as_PRegister(i), Address(sp, i));
3692 }
3693 }
3694 }
3695
3696 void MacroAssembler::pop_CPU_state(bool restore_vectors, bool use_sve,
3697 int sve_vector_size_in_bytes, int total_predicate_in_bytes) {
3698 if (restore_vectors && use_sve && total_predicate_in_bytes > 0) {
3699 for (int i = PRegister::number_of_registers - 1; i >= 0; i--) {
3700 sve_ldr(as_PRegister(i), Address(sp, i));
3701 }
3702 add(sp, sp, total_predicate_in_bytes);
3703 }
3704 if (restore_vectors && use_sve && sve_vector_size_in_bytes > 16) {
3705 for (int i = FloatRegister::number_of_registers - 1; i >= 0; i--) {
3706 sve_ldr(as_FloatRegister(i), Address(sp, i));
3707 }
3708 add(sp, sp, sve_vector_size_in_bytes * FloatRegister::number_of_registers);
3709 } else {
3710 int step = (restore_vectors ? 8 : 4) * wordSize;
3711 for (int i = 0; i <= 28; i += 4)
3712 ld1(as_FloatRegister(i), as_FloatRegister(i+1), as_FloatRegister(i+2),
3713 as_FloatRegister(i+3), restore_vectors ? T2D : T1D, Address(post(sp, step)));
3714 }
3715
3716 // We may use predicate registers and rely on ptrue with SVE,
3717 // regardless of wide vector (> 8 bytes) used or not.
3718 if (use_sve) {
3719 reinitialize_ptrue();
3720 }
3721
3722 // integer registers except lr & sp
3723 pop(RegSet::range(r0, r17), sp);
3724 #ifdef R18_RESERVED
3725 ldp(zr, r19, Address(post(sp, 2 * wordSize)));
3726 pop(RegSet::range(r20, r29), sp);
3727 #else
3728 pop(RegSet::range(r18_tls, r29), sp);
3729 #endif
3730 }
3731
3732 /**
3733 * Helpers for multiply_to_len().
3734 */
3735 void MacroAssembler::add2_with_carry(Register final_dest_hi, Register dest_hi, Register dest_lo,
3736 Register src1, Register src2) {
3737 adds(dest_lo, dest_lo, src1);
3738 adc(dest_hi, dest_hi, zr);
3739 adds(dest_lo, dest_lo, src2);
3740 adc(final_dest_hi, dest_hi, zr);
3741 }
3742
3743 // Generate an address from (r + r1 extend offset). "size" is the
3744 // size of the operand. The result may be in rscratch2.
3745 Address MacroAssembler::offsetted_address(Register r, Register r1,
3746 Address::extend ext, int offset, int size) {
3747 if (offset || (ext.shift() % size != 0)) {
3748 lea(rscratch2, Address(r, r1, ext));
3749 return Address(rscratch2, offset);
3750 } else {
3751 return Address(r, r1, ext);
3752 }
3753 }
3754
3755 Address MacroAssembler::spill_address(int size, int offset, Register tmp)
3756 {
3757 assert(offset >= 0, "spill to negative address?");
3758 // Offset reachable ?
3759 // Not aligned - 9 bits signed offset
3760 // Aligned - 12 bits unsigned offset shifted
3761 Register base = sp;
3762 if ((offset & (size-1)) && offset >= (1<<8)) {
3763 add(tmp, base, offset & ((1<<12)-1));
3764 base = tmp;
3765 offset &= -1u<<12;
3766 }
3767
3768 if (offset >= (1<<12) * size) {
3769 add(tmp, base, offset & (((1<<12)-1)<<12));
3770 base = tmp;
3771 offset &= ~(((1<<12)-1)<<12);
3772 }
3773
3774 return Address(base, offset);
3775 }
3776
3777 Address MacroAssembler::sve_spill_address(int sve_reg_size_in_bytes, int offset, Register tmp) {
3778 assert(offset >= 0, "spill to negative address?");
3779
3780 Register base = sp;
3781
3782 // An immediate offset in the range 0 to 255 which is multiplied
3783 // by the current vector or predicate register size in bytes.
3784 if (offset % sve_reg_size_in_bytes == 0 && offset < ((1<<8)*sve_reg_size_in_bytes)) {
3785 return Address(base, offset / sve_reg_size_in_bytes);
3786 }
3787
3788 add(tmp, base, offset);
3789 return Address(tmp);
3790 }
3791
3792 // Checks whether offset is aligned.
3793 // Returns true if it is, else false.
3794 bool MacroAssembler::merge_alignment_check(Register base,
3795 size_t size,
3796 int64_t cur_offset,
3797 int64_t prev_offset) const {
3798 if (AvoidUnalignedAccesses) {
3799 if (base == sp) {
3800 // Checks whether low offset if aligned to pair of registers.
3801 int64_t pair_mask = size * 2 - 1;
3802 int64_t offset = prev_offset > cur_offset ? cur_offset : prev_offset;
3803 return (offset & pair_mask) == 0;
3804 } else { // If base is not sp, we can't guarantee the access is aligned.
3805 return false;
3806 }
3807 } else {
3808 int64_t mask = size - 1;
3809 // Load/store pair instruction only supports element size aligned offset.
3810 return (cur_offset & mask) == 0 && (prev_offset & mask) == 0;
3811 }
3812 }
3813
3814 // Checks whether current and previous loads/stores can be merged.
3815 // Returns true if it can be merged, else false.
3816 bool MacroAssembler::ldst_can_merge(Register rt,
3817 const Address &adr,
3818 size_t cur_size_in_bytes,
3819 bool is_store) const {
3820 address prev = pc() - NativeInstruction::instruction_size;
3821 address last = code()->last_insn();
3822
3823 if (last == nullptr || !nativeInstruction_at(last)->is_Imm_LdSt()) {
3824 return false;
3825 }
3826
3827 if (adr.getMode() != Address::base_plus_offset || prev != last) {
3828 return false;
3829 }
3830
3831 NativeLdSt* prev_ldst = NativeLdSt_at(prev);
3832 size_t prev_size_in_bytes = prev_ldst->size_in_bytes();
3833
3834 assert(prev_size_in_bytes == 4 || prev_size_in_bytes == 8, "only supports 64/32bit merging.");
3835 assert(cur_size_in_bytes == 4 || cur_size_in_bytes == 8, "only supports 64/32bit merging.");
3836
3837 if (cur_size_in_bytes != prev_size_in_bytes || is_store != prev_ldst->is_store()) {
3838 return false;
3839 }
3840
3841 int64_t max_offset = 63 * prev_size_in_bytes;
3842 int64_t min_offset = -64 * prev_size_in_bytes;
3843
3844 assert(prev_ldst->is_not_pre_post_index(), "pre-index or post-index is not supported to be merged.");
3845
3846 // Only same base can be merged.
3847 if (adr.base() != prev_ldst->base()) {
3848 return false;
3849 }
3850
3851 int64_t cur_offset = adr.offset();
3852 int64_t prev_offset = prev_ldst->offset();
3853 size_t diff = abs(cur_offset - prev_offset);
3854 if (diff != prev_size_in_bytes) {
3855 return false;
3856 }
3857
3858 // Following cases can not be merged:
3859 // ldr x2, [x2, #8]
3860 // ldr x3, [x2, #16]
3861 // or:
3862 // ldr x2, [x3, #8]
3863 // ldr x2, [x3, #16]
3864 // If t1 and t2 is the same in "ldp t1, t2, [xn, #imm]", we'll get SIGILL.
3865 if (!is_store && (adr.base() == prev_ldst->target() || rt == prev_ldst->target())) {
3866 return false;
3867 }
3868
3869 int64_t low_offset = prev_offset > cur_offset ? cur_offset : prev_offset;
3870 // Offset range must be in ldp/stp instruction's range.
3871 if (low_offset > max_offset || low_offset < min_offset) {
3872 return false;
3873 }
3874
3875 if (merge_alignment_check(adr.base(), prev_size_in_bytes, cur_offset, prev_offset)) {
3876 return true;
3877 }
3878
3879 return false;
3880 }
3881
3882 // Merge current load/store with previous load/store into ldp/stp.
3883 void MacroAssembler::merge_ldst(Register rt,
3884 const Address &adr,
3885 size_t cur_size_in_bytes,
3886 bool is_store) {
3887
3888 assert(ldst_can_merge(rt, adr, cur_size_in_bytes, is_store) == true, "cur and prev must be able to be merged.");
3889
3890 Register rt_low, rt_high;
3891 address prev = pc() - NativeInstruction::instruction_size;
3892 NativeLdSt* prev_ldst = NativeLdSt_at(prev);
3893
3894 int64_t offset;
3895
3896 if (adr.offset() < prev_ldst->offset()) {
3897 offset = adr.offset();
3898 rt_low = rt;
3899 rt_high = prev_ldst->target();
3900 } else {
3901 offset = prev_ldst->offset();
3902 rt_low = prev_ldst->target();
3903 rt_high = rt;
3904 }
3905
3906 Address adr_p = Address(prev_ldst->base(), offset);
3907 // Overwrite previous generated binary.
3908 code_section()->set_end(prev);
3909
3910 const size_t sz = prev_ldst->size_in_bytes();
3911 assert(sz == 8 || sz == 4, "only supports 64/32bit merging.");
3912 if (!is_store) {
3913 BLOCK_COMMENT("merged ldr pair");
3914 if (sz == 8) {
3915 ldp(rt_low, rt_high, adr_p);
3916 } else {
3917 ldpw(rt_low, rt_high, adr_p);
3918 }
3919 } else {
3920 BLOCK_COMMENT("merged str pair");
3921 if (sz == 8) {
3922 stp(rt_low, rt_high, adr_p);
3923 } else {
3924 stpw(rt_low, rt_high, adr_p);
3925 }
3926 }
3927 }
3928
3929 /**
3930 * Multiply 64 bit by 64 bit first loop.
3931 */
3932 void MacroAssembler::multiply_64_x_64_loop(Register x, Register xstart, Register x_xstart,
3933 Register y, Register y_idx, Register z,
3934 Register carry, Register product,
3935 Register idx, Register kdx) {
3936 //
3937 // jlong carry, x[], y[], z[];
3938 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
3939 // huge_128 product = y[idx] * x[xstart] + carry;
3940 // z[kdx] = (jlong)product;
3941 // carry = (jlong)(product >>> 64);
3942 // }
3943 // z[xstart] = carry;
3944 //
3945
3946 Label L_first_loop, L_first_loop_exit;
3947 Label L_one_x, L_one_y, L_multiply;
3948
3949 subsw(xstart, xstart, 1);
3950 br(Assembler::MI, L_one_x);
3951
3952 lea(rscratch1, Address(x, xstart, Address::lsl(LogBytesPerInt)));
3953 ldr(x_xstart, Address(rscratch1));
3954 ror(x_xstart, x_xstart, 32); // convert big-endian to little-endian
3955
3956 bind(L_first_loop);
3957 subsw(idx, idx, 1);
3958 br(Assembler::MI, L_first_loop_exit);
3959 subsw(idx, idx, 1);
3960 br(Assembler::MI, L_one_y);
3961 lea(rscratch1, Address(y, idx, Address::uxtw(LogBytesPerInt)));
3962 ldr(y_idx, Address(rscratch1));
3963 ror(y_idx, y_idx, 32); // convert big-endian to little-endian
3964 bind(L_multiply);
3965
3966 // AArch64 has a multiply-accumulate instruction that we can't use
3967 // here because it has no way to process carries, so we have to use
3968 // separate add and adc instructions. Bah.
3969 umulh(rscratch1, x_xstart, y_idx); // x_xstart * y_idx -> rscratch1:product
3970 mul(product, x_xstart, y_idx);
3971 adds(product, product, carry);
3972 adc(carry, rscratch1, zr); // x_xstart * y_idx + carry -> carry:product
3973
3974 subw(kdx, kdx, 2);
3975 ror(product, product, 32); // back to big-endian
3976 str(product, offsetted_address(z, kdx, Address::uxtw(LogBytesPerInt), 0, BytesPerLong));
3977
3978 b(L_first_loop);
3979
3980 bind(L_one_y);
3981 ldrw(y_idx, Address(y, 0));
3982 b(L_multiply);
3983
3984 bind(L_one_x);
3985 ldrw(x_xstart, Address(x, 0));
3986 b(L_first_loop);
3987
3988 bind(L_first_loop_exit);
3989 }
3990
3991 /**
3992 * Multiply 128 bit by 128. Unrolled inner loop.
3993 *
3994 */
3995 void MacroAssembler::multiply_128_x_128_loop(Register y, Register z,
3996 Register carry, Register carry2,
3997 Register idx, Register jdx,
3998 Register yz_idx1, Register yz_idx2,
3999 Register tmp, Register tmp3, Register tmp4,
4000 Register tmp6, Register product_hi) {
4001
4002 // jlong carry, x[], y[], z[];
4003 // int kdx = ystart+1;
4004 // for (int idx=ystart-2; idx >= 0; idx -= 2) { // Third loop
4005 // huge_128 tmp3 = (y[idx+1] * product_hi) + z[kdx+idx+1] + carry;
4006 // jlong carry2 = (jlong)(tmp3 >>> 64);
4007 // huge_128 tmp4 = (y[idx] * product_hi) + z[kdx+idx] + carry2;
4008 // carry = (jlong)(tmp4 >>> 64);
4009 // z[kdx+idx+1] = (jlong)tmp3;
4010 // z[kdx+idx] = (jlong)tmp4;
4011 // }
4012 // idx += 2;
4013 // if (idx > 0) {
4014 // yz_idx1 = (y[idx] * product_hi) + z[kdx+idx] + carry;
4015 // z[kdx+idx] = (jlong)yz_idx1;
4016 // carry = (jlong)(yz_idx1 >>> 64);
4017 // }
4018 //
4019
4020 Label L_third_loop, L_third_loop_exit, L_post_third_loop_done;
4021
4022 lsrw(jdx, idx, 2);
4023
4024 bind(L_third_loop);
4025
4026 subsw(jdx, jdx, 1);
4027 br(Assembler::MI, L_third_loop_exit);
4028 subw(idx, idx, 4);
4029
4030 lea(rscratch1, Address(y, idx, Address::uxtw(LogBytesPerInt)));
4031
4032 ldp(yz_idx2, yz_idx1, Address(rscratch1, 0));
4033
4034 lea(tmp6, Address(z, idx, Address::uxtw(LogBytesPerInt)));
4035
4036 ror(yz_idx1, yz_idx1, 32); // convert big-endian to little-endian
4037 ror(yz_idx2, yz_idx2, 32);
4038
4039 ldp(rscratch2, rscratch1, Address(tmp6, 0));
4040
4041 mul(tmp3, product_hi, yz_idx1); // yz_idx1 * product_hi -> tmp4:tmp3
4042 umulh(tmp4, product_hi, yz_idx1);
4043
4044 ror(rscratch1, rscratch1, 32); // convert big-endian to little-endian
4045 ror(rscratch2, rscratch2, 32);
4046
4047 mul(tmp, product_hi, yz_idx2); // yz_idx2 * product_hi -> carry2:tmp
4048 umulh(carry2, product_hi, yz_idx2);
4049
4050 // propagate sum of both multiplications into carry:tmp4:tmp3
4051 adds(tmp3, tmp3, carry);
4052 adc(tmp4, tmp4, zr);
4053 adds(tmp3, tmp3, rscratch1);
4054 adcs(tmp4, tmp4, tmp);
4055 adc(carry, carry2, zr);
4056 adds(tmp4, tmp4, rscratch2);
4057 adc(carry, carry, zr);
4058
4059 ror(tmp3, tmp3, 32); // convert little-endian to big-endian
4060 ror(tmp4, tmp4, 32);
4061 stp(tmp4, tmp3, Address(tmp6, 0));
4062
4063 b(L_third_loop);
4064 bind (L_third_loop_exit);
4065
4066 andw (idx, idx, 0x3);
4067 cbz(idx, L_post_third_loop_done);
4068
4069 Label L_check_1;
4070 subsw(idx, idx, 2);
4071 br(Assembler::MI, L_check_1);
4072
4073 lea(rscratch1, Address(y, idx, Address::uxtw(LogBytesPerInt)));
4074 ldr(yz_idx1, Address(rscratch1, 0));
4075 ror(yz_idx1, yz_idx1, 32);
4076 mul(tmp3, product_hi, yz_idx1); // yz_idx1 * product_hi -> tmp4:tmp3
4077 umulh(tmp4, product_hi, yz_idx1);
4078 lea(rscratch1, Address(z, idx, Address::uxtw(LogBytesPerInt)));
4079 ldr(yz_idx2, Address(rscratch1, 0));
4080 ror(yz_idx2, yz_idx2, 32);
4081
4082 add2_with_carry(carry, tmp4, tmp3, carry, yz_idx2);
4083
4084 ror(tmp3, tmp3, 32);
4085 str(tmp3, Address(rscratch1, 0));
4086
4087 bind (L_check_1);
4088
4089 andw (idx, idx, 0x1);
4090 subsw(idx, idx, 1);
4091 br(Assembler::MI, L_post_third_loop_done);
4092 ldrw(tmp4, Address(y, idx, Address::uxtw(LogBytesPerInt)));
4093 mul(tmp3, tmp4, product_hi); // tmp4 * product_hi -> carry2:tmp3
4094 umulh(carry2, tmp4, product_hi);
4095 ldrw(tmp4, Address(z, idx, Address::uxtw(LogBytesPerInt)));
4096
4097 add2_with_carry(carry2, tmp3, tmp4, carry);
4098
4099 strw(tmp3, Address(z, idx, Address::uxtw(LogBytesPerInt)));
4100 extr(carry, carry2, tmp3, 32);
4101
4102 bind(L_post_third_loop_done);
4103 }
4104
4105 /**
4106 * Code for BigInteger::multiplyToLen() intrinsic.
4107 *
4108 * r0: x
4109 * r1: xlen
4110 * r2: y
4111 * r3: ylen
4112 * r4: z
4113 * r5: tmp0
4114 * r10: tmp1
4115 * r11: tmp2
4116 * r12: tmp3
4117 * r13: tmp4
4118 * r14: tmp5
4119 * r15: tmp6
4120 * r16: tmp7
4121 *
4122 */
4123 void MacroAssembler::multiply_to_len(Register x, Register xlen, Register y, Register ylen,
4124 Register z, Register tmp0,
4125 Register tmp1, Register tmp2, Register tmp3, Register tmp4,
4126 Register tmp5, Register tmp6, Register product_hi) {
4127
4128 assert_different_registers(x, xlen, y, ylen, z, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, product_hi);
4129
4130 const Register idx = tmp1;
4131 const Register kdx = tmp2;
4132 const Register xstart = tmp3;
4133
4134 const Register y_idx = tmp4;
4135 const Register carry = tmp5;
4136 const Register product = xlen;
4137 const Register x_xstart = tmp0;
4138
4139 // First Loop.
4140 //
4141 // final static long LONG_MASK = 0xffffffffL;
4142 // int xstart = xlen - 1;
4143 // int ystart = ylen - 1;
4144 // long carry = 0;
4145 // for (int idx=ystart, kdx=ystart+1+xstart; idx >= 0; idx-, kdx--) {
4146 // long product = (y[idx] & LONG_MASK) * (x[xstart] & LONG_MASK) + carry;
4147 // z[kdx] = (int)product;
4148 // carry = product >>> 32;
4149 // }
4150 // z[xstart] = (int)carry;
4151 //
4152
4153 movw(idx, ylen); // idx = ylen;
4154 addw(kdx, xlen, ylen); // kdx = xlen+ylen;
4155 mov(carry, zr); // carry = 0;
4156
4157 Label L_done;
4158
4159 movw(xstart, xlen);
4160 subsw(xstart, xstart, 1);
4161 br(Assembler::MI, L_done);
4162
4163 multiply_64_x_64_loop(x, xstart, x_xstart, y, y_idx, z, carry, product, idx, kdx);
4164
4165 Label L_second_loop;
4166 cbzw(kdx, L_second_loop);
4167
4168 Label L_carry;
4169 subw(kdx, kdx, 1);
4170 cbzw(kdx, L_carry);
4171
4172 strw(carry, Address(z, kdx, Address::uxtw(LogBytesPerInt)));
4173 lsr(carry, carry, 32);
4174 subw(kdx, kdx, 1);
4175
4176 bind(L_carry);
4177 strw(carry, Address(z, kdx, Address::uxtw(LogBytesPerInt)));
4178
4179 // Second and third (nested) loops.
4180 //
4181 // for (int i = xstart-1; i >= 0; i--) { // Second loop
4182 // carry = 0;
4183 // for (int jdx=ystart, k=ystart+1+i; jdx >= 0; jdx--, k--) { // Third loop
4184 // long product = (y[jdx] & LONG_MASK) * (x[i] & LONG_MASK) +
4185 // (z[k] & LONG_MASK) + carry;
4186 // z[k] = (int)product;
4187 // carry = product >>> 32;
4188 // }
4189 // z[i] = (int)carry;
4190 // }
4191 //
4192 // i = xlen, j = tmp1, k = tmp2, carry = tmp5, x[i] = product_hi
4193
4194 const Register jdx = tmp1;
4195
4196 bind(L_second_loop);
4197 mov(carry, zr); // carry = 0;
4198 movw(jdx, ylen); // j = ystart+1
4199
4200 subsw(xstart, xstart, 1); // i = xstart-1;
4201 br(Assembler::MI, L_done);
4202
4203 str(z, Address(pre(sp, -4 * wordSize)));
4204
4205 Label L_last_x;
4206 lea(z, offsetted_address(z, xstart, Address::uxtw(LogBytesPerInt), 4, BytesPerInt)); // z = z + k - j
4207 subsw(xstart, xstart, 1); // i = xstart-1;
4208 br(Assembler::MI, L_last_x);
4209
4210 lea(rscratch1, Address(x, xstart, Address::uxtw(LogBytesPerInt)));
4211 ldr(product_hi, Address(rscratch1));
4212 ror(product_hi, product_hi, 32); // convert big-endian to little-endian
4213
4214 Label L_third_loop_prologue;
4215 bind(L_third_loop_prologue);
4216
4217 str(ylen, Address(sp, wordSize));
4218 stp(x, xstart, Address(sp, 2 * wordSize));
4219 multiply_128_x_128_loop(y, z, carry, x, jdx, ylen, product,
4220 tmp2, x_xstart, tmp3, tmp4, tmp6, product_hi);
4221 ldp(z, ylen, Address(post(sp, 2 * wordSize)));
4222 ldp(x, xlen, Address(post(sp, 2 * wordSize))); // copy old xstart -> xlen
4223
4224 addw(tmp3, xlen, 1);
4225 strw(carry, Address(z, tmp3, Address::uxtw(LogBytesPerInt)));
4226 subsw(tmp3, tmp3, 1);
4227 br(Assembler::MI, L_done);
4228
4229 lsr(carry, carry, 32);
4230 strw(carry, Address(z, tmp3, Address::uxtw(LogBytesPerInt)));
4231 b(L_second_loop);
4232
4233 // Next infrequent code is moved outside loops.
4234 bind(L_last_x);
4235 ldrw(product_hi, Address(x, 0));
4236 b(L_third_loop_prologue);
4237
4238 bind(L_done);
4239 }
4240
4241 // Code for BigInteger::mulAdd intrinsic
4242 // out = r0
4243 // in = r1
4244 // offset = r2 (already out.length-offset)
4245 // len = r3
4246 // k = r4
4247 //
4248 // pseudo code from java implementation:
4249 // carry = 0;
4250 // offset = out.length-offset - 1;
4251 // for (int j=len-1; j >= 0; j--) {
4252 // product = (in[j] & LONG_MASK) * kLong + (out[offset] & LONG_MASK) + carry;
4253 // out[offset--] = (int)product;
4254 // carry = product >>> 32;
4255 // }
4256 // return (int)carry;
4257 void MacroAssembler::mul_add(Register out, Register in, Register offset,
4258 Register len, Register k) {
4259 Label LOOP, END;
4260 // pre-loop
4261 cmp(len, zr); // cmp, not cbz/cbnz: to use condition twice => less branches
4262 csel(out, zr, out, Assembler::EQ);
4263 br(Assembler::EQ, END);
4264 add(in, in, len, LSL, 2); // in[j+1] address
4265 add(offset, out, offset, LSL, 2); // out[offset + 1] address
4266 mov(out, zr); // used to keep carry now
4267 BIND(LOOP);
4268 ldrw(rscratch1, Address(pre(in, -4)));
4269 madd(rscratch1, rscratch1, k, out);
4270 ldrw(rscratch2, Address(pre(offset, -4)));
4271 add(rscratch1, rscratch1, rscratch2);
4272 strw(rscratch1, Address(offset));
4273 lsr(out, rscratch1, 32);
4274 subs(len, len, 1);
4275 br(Assembler::NE, LOOP);
4276 BIND(END);
4277 }
4278
4279 /**
4280 * Emits code to update CRC-32 with a byte value according to constants in table
4281 *
4282 * @param [in,out]crc Register containing the crc.
4283 * @param [in]val Register containing the byte to fold into the CRC.
4284 * @param [in]table Register containing the table of crc constants.
4285 *
4286 * uint32_t crc;
4287 * val = crc_table[(val ^ crc) & 0xFF];
4288 * crc = val ^ (crc >> 8);
4289 *
4290 */
4291 void MacroAssembler::update_byte_crc32(Register crc, Register val, Register table) {
4292 eor(val, val, crc);
4293 andr(val, val, 0xff);
4294 ldrw(val, Address(table, val, Address::lsl(2)));
4295 eor(crc, val, crc, Assembler::LSR, 8);
4296 }
4297
4298 /**
4299 * Emits code to update CRC-32 with a 32-bit value according to tables 0 to 3
4300 *
4301 * @param [in,out]crc Register containing the crc.
4302 * @param [in]v Register containing the 32-bit to fold into the CRC.
4303 * @param [in]table0 Register containing table 0 of crc constants.
4304 * @param [in]table1 Register containing table 1 of crc constants.
4305 * @param [in]table2 Register containing table 2 of crc constants.
4306 * @param [in]table3 Register containing table 3 of crc constants.
4307 *
4308 * uint32_t crc;
4309 * v = crc ^ v
4310 * crc = table3[v&0xff]^table2[(v>>8)&0xff]^table1[(v>>16)&0xff]^table0[v>>24]
4311 *
4312 */
4313 void MacroAssembler::update_word_crc32(Register crc, Register v, Register tmp,
4314 Register table0, Register table1, Register table2, Register table3,
4315 bool upper) {
4316 eor(v, crc, v, upper ? LSR:LSL, upper ? 32:0);
4317 uxtb(tmp, v);
4318 ldrw(crc, Address(table3, tmp, Address::lsl(2)));
4319 ubfx(tmp, v, 8, 8);
4320 ldrw(tmp, Address(table2, tmp, Address::lsl(2)));
4321 eor(crc, crc, tmp);
4322 ubfx(tmp, v, 16, 8);
4323 ldrw(tmp, Address(table1, tmp, Address::lsl(2)));
4324 eor(crc, crc, tmp);
4325 ubfx(tmp, v, 24, 8);
4326 ldrw(tmp, Address(table0, tmp, Address::lsl(2)));
4327 eor(crc, crc, tmp);
4328 }
4329
4330 void MacroAssembler::kernel_crc32_using_crypto_pmull(Register crc, Register buf,
4331 Register len, Register tmp0, Register tmp1, Register tmp2, Register tmp3) {
4332 Label CRC_by4_loop, CRC_by1_loop, CRC_less128, CRC_by128_pre, CRC_by32_loop, CRC_less32, L_exit;
4333 assert_different_registers(crc, buf, len, tmp0, tmp1, tmp2);
4334
4335 subs(tmp0, len, 384);
4336 mvnw(crc, crc);
4337 br(Assembler::GE, CRC_by128_pre);
4338 BIND(CRC_less128);
4339 subs(len, len, 32);
4340 br(Assembler::GE, CRC_by32_loop);
4341 BIND(CRC_less32);
4342 adds(len, len, 32 - 4);
4343 br(Assembler::GE, CRC_by4_loop);
4344 adds(len, len, 4);
4345 br(Assembler::GT, CRC_by1_loop);
4346 b(L_exit);
4347
4348 BIND(CRC_by32_loop);
4349 ldp(tmp0, tmp1, Address(buf));
4350 crc32x(crc, crc, tmp0);
4351 ldp(tmp2, tmp3, Address(buf, 16));
4352 crc32x(crc, crc, tmp1);
4353 add(buf, buf, 32);
4354 crc32x(crc, crc, tmp2);
4355 subs(len, len, 32);
4356 crc32x(crc, crc, tmp3);
4357 br(Assembler::GE, CRC_by32_loop);
4358 cmn(len, (u1)32);
4359 br(Assembler::NE, CRC_less32);
4360 b(L_exit);
4361
4362 BIND(CRC_by4_loop);
4363 ldrw(tmp0, Address(post(buf, 4)));
4364 subs(len, len, 4);
4365 crc32w(crc, crc, tmp0);
4366 br(Assembler::GE, CRC_by4_loop);
4367 adds(len, len, 4);
4368 br(Assembler::LE, L_exit);
4369 BIND(CRC_by1_loop);
4370 ldrb(tmp0, Address(post(buf, 1)));
4371 subs(len, len, 1);
4372 crc32b(crc, crc, tmp0);
4373 br(Assembler::GT, CRC_by1_loop);
4374 b(L_exit);
4375
4376 BIND(CRC_by128_pre);
4377 kernel_crc32_common_fold_using_crypto_pmull(crc, buf, len, tmp0, tmp1, tmp2,
4378 4*256*sizeof(juint) + 8*sizeof(juint));
4379 mov(crc, 0);
4380 crc32x(crc, crc, tmp0);
4381 crc32x(crc, crc, tmp1);
4382
4383 cbnz(len, CRC_less128);
4384
4385 BIND(L_exit);
4386 mvnw(crc, crc);
4387 }
4388
4389 void MacroAssembler::kernel_crc32_using_crc32(Register crc, Register buf,
4390 Register len, Register tmp0, Register tmp1, Register tmp2,
4391 Register tmp3) {
4392 Label CRC_by64_loop, CRC_by4_loop, CRC_by1_loop, CRC_less64, CRC_by64_pre, CRC_by32_loop, CRC_less32, L_exit;
4393 assert_different_registers(crc, buf, len, tmp0, tmp1, tmp2, tmp3);
4394
4395 mvnw(crc, crc);
4396
4397 subs(len, len, 128);
4398 br(Assembler::GE, CRC_by64_pre);
4399 BIND(CRC_less64);
4400 adds(len, len, 128-32);
4401 br(Assembler::GE, CRC_by32_loop);
4402 BIND(CRC_less32);
4403 adds(len, len, 32-4);
4404 br(Assembler::GE, CRC_by4_loop);
4405 adds(len, len, 4);
4406 br(Assembler::GT, CRC_by1_loop);
4407 b(L_exit);
4408
4409 BIND(CRC_by32_loop);
4410 ldp(tmp0, tmp1, Address(post(buf, 16)));
4411 subs(len, len, 32);
4412 crc32x(crc, crc, tmp0);
4413 ldr(tmp2, Address(post(buf, 8)));
4414 crc32x(crc, crc, tmp1);
4415 ldr(tmp3, Address(post(buf, 8)));
4416 crc32x(crc, crc, tmp2);
4417 crc32x(crc, crc, tmp3);
4418 br(Assembler::GE, CRC_by32_loop);
4419 cmn(len, (u1)32);
4420 br(Assembler::NE, CRC_less32);
4421 b(L_exit);
4422
4423 BIND(CRC_by4_loop);
4424 ldrw(tmp0, Address(post(buf, 4)));
4425 subs(len, len, 4);
4426 crc32w(crc, crc, tmp0);
4427 br(Assembler::GE, CRC_by4_loop);
4428 adds(len, len, 4);
4429 br(Assembler::LE, L_exit);
4430 BIND(CRC_by1_loop);
4431 ldrb(tmp0, Address(post(buf, 1)));
4432 subs(len, len, 1);
4433 crc32b(crc, crc, tmp0);
4434 br(Assembler::GT, CRC_by1_loop);
4435 b(L_exit);
4436
4437 BIND(CRC_by64_pre);
4438 sub(buf, buf, 8);
4439 ldp(tmp0, tmp1, Address(buf, 8));
4440 crc32x(crc, crc, tmp0);
4441 ldr(tmp2, Address(buf, 24));
4442 crc32x(crc, crc, tmp1);
4443 ldr(tmp3, Address(buf, 32));
4444 crc32x(crc, crc, tmp2);
4445 ldr(tmp0, Address(buf, 40));
4446 crc32x(crc, crc, tmp3);
4447 ldr(tmp1, Address(buf, 48));
4448 crc32x(crc, crc, tmp0);
4449 ldr(tmp2, Address(buf, 56));
4450 crc32x(crc, crc, tmp1);
4451 ldr(tmp3, Address(pre(buf, 64)));
4452
4453 b(CRC_by64_loop);
4454
4455 align(CodeEntryAlignment);
4456 BIND(CRC_by64_loop);
4457 subs(len, len, 64);
4458 crc32x(crc, crc, tmp2);
4459 ldr(tmp0, Address(buf, 8));
4460 crc32x(crc, crc, tmp3);
4461 ldr(tmp1, Address(buf, 16));
4462 crc32x(crc, crc, tmp0);
4463 ldr(tmp2, Address(buf, 24));
4464 crc32x(crc, crc, tmp1);
4465 ldr(tmp3, Address(buf, 32));
4466 crc32x(crc, crc, tmp2);
4467 ldr(tmp0, Address(buf, 40));
4468 crc32x(crc, crc, tmp3);
4469 ldr(tmp1, Address(buf, 48));
4470 crc32x(crc, crc, tmp0);
4471 ldr(tmp2, Address(buf, 56));
4472 crc32x(crc, crc, tmp1);
4473 ldr(tmp3, Address(pre(buf, 64)));
4474 br(Assembler::GE, CRC_by64_loop);
4475
4476 // post-loop
4477 crc32x(crc, crc, tmp2);
4478 crc32x(crc, crc, tmp3);
4479
4480 sub(len, len, 64);
4481 add(buf, buf, 8);
4482 cmn(len, (u1)128);
4483 br(Assembler::NE, CRC_less64);
4484 BIND(L_exit);
4485 mvnw(crc, crc);
4486 }
4487
4488 /**
4489 * @param crc register containing existing CRC (32-bit)
4490 * @param buf register pointing to input byte buffer (byte*)
4491 * @param len register containing number of bytes
4492 * @param table register that will contain address of CRC table
4493 * @param tmp scratch register
4494 */
4495 void MacroAssembler::kernel_crc32(Register crc, Register buf, Register len,
4496 Register table0, Register table1, Register table2, Register table3,
4497 Register tmp, Register tmp2, Register tmp3) {
4498 Label L_by16, L_by16_loop, L_by4, L_by4_loop, L_by1, L_by1_loop, L_exit;
4499
4500 if (UseCryptoPmullForCRC32) {
4501 kernel_crc32_using_crypto_pmull(crc, buf, len, table0, table1, table2, table3);
4502 return;
4503 }
4504
4505 if (UseCRC32) {
4506 kernel_crc32_using_crc32(crc, buf, len, table0, table1, table2, table3);
4507 return;
4508 }
4509
4510 mvnw(crc, crc);
4511
4512 {
4513 uint64_t offset;
4514 adrp(table0, ExternalAddress(StubRoutines::crc_table_addr()), offset);
4515 add(table0, table0, offset);
4516 }
4517 add(table1, table0, 1*256*sizeof(juint));
4518 add(table2, table0, 2*256*sizeof(juint));
4519 add(table3, table0, 3*256*sizeof(juint));
4520
4521 { // Neon code start
4522 cmp(len, (u1)64);
4523 br(Assembler::LT, L_by16);
4524 eor(v16, T16B, v16, v16);
4525
4526 Label L_fold;
4527
4528 add(tmp, table0, 4*256*sizeof(juint)); // Point at the Neon constants
4529
4530 ld1(v0, v1, T2D, post(buf, 32));
4531 ld1r(v4, T2D, post(tmp, 8));
4532 ld1r(v5, T2D, post(tmp, 8));
4533 ld1r(v6, T2D, post(tmp, 8));
4534 ld1r(v7, T2D, post(tmp, 8));
4535 mov(v16, S, 0, crc);
4536
4537 eor(v0, T16B, v0, v16);
4538 sub(len, len, 64);
4539
4540 BIND(L_fold);
4541 pmull(v22, T8H, v0, v5, T8B);
4542 pmull(v20, T8H, v0, v7, T8B);
4543 pmull(v23, T8H, v0, v4, T8B);
4544 pmull(v21, T8H, v0, v6, T8B);
4545
4546 pmull2(v18, T8H, v0, v5, T16B);
4547 pmull2(v16, T8H, v0, v7, T16B);
4548 pmull2(v19, T8H, v0, v4, T16B);
4549 pmull2(v17, T8H, v0, v6, T16B);
4550
4551 uzp1(v24, T8H, v20, v22);
4552 uzp2(v25, T8H, v20, v22);
4553 eor(v20, T16B, v24, v25);
4554
4555 uzp1(v26, T8H, v16, v18);
4556 uzp2(v27, T8H, v16, v18);
4557 eor(v16, T16B, v26, v27);
4558
4559 ushll2(v22, T4S, v20, T8H, 8);
4560 ushll(v20, T4S, v20, T4H, 8);
4561
4562 ushll2(v18, T4S, v16, T8H, 8);
4563 ushll(v16, T4S, v16, T4H, 8);
4564
4565 eor(v22, T16B, v23, v22);
4566 eor(v18, T16B, v19, v18);
4567 eor(v20, T16B, v21, v20);
4568 eor(v16, T16B, v17, v16);
4569
4570 uzp1(v17, T2D, v16, v20);
4571 uzp2(v21, T2D, v16, v20);
4572 eor(v17, T16B, v17, v21);
4573
4574 ushll2(v20, T2D, v17, T4S, 16);
4575 ushll(v16, T2D, v17, T2S, 16);
4576
4577 eor(v20, T16B, v20, v22);
4578 eor(v16, T16B, v16, v18);
4579
4580 uzp1(v17, T2D, v20, v16);
4581 uzp2(v21, T2D, v20, v16);
4582 eor(v28, T16B, v17, v21);
4583
4584 pmull(v22, T8H, v1, v5, T8B);
4585 pmull(v20, T8H, v1, v7, T8B);
4586 pmull(v23, T8H, v1, v4, T8B);
4587 pmull(v21, T8H, v1, v6, T8B);
4588
4589 pmull2(v18, T8H, v1, v5, T16B);
4590 pmull2(v16, T8H, v1, v7, T16B);
4591 pmull2(v19, T8H, v1, v4, T16B);
4592 pmull2(v17, T8H, v1, v6, T16B);
4593
4594 ld1(v0, v1, T2D, post(buf, 32));
4595
4596 uzp1(v24, T8H, v20, v22);
4597 uzp2(v25, T8H, v20, v22);
4598 eor(v20, T16B, v24, v25);
4599
4600 uzp1(v26, T8H, v16, v18);
4601 uzp2(v27, T8H, v16, v18);
4602 eor(v16, T16B, v26, v27);
4603
4604 ushll2(v22, T4S, v20, T8H, 8);
4605 ushll(v20, T4S, v20, T4H, 8);
4606
4607 ushll2(v18, T4S, v16, T8H, 8);
4608 ushll(v16, T4S, v16, T4H, 8);
4609
4610 eor(v22, T16B, v23, v22);
4611 eor(v18, T16B, v19, v18);
4612 eor(v20, T16B, v21, v20);
4613 eor(v16, T16B, v17, v16);
4614
4615 uzp1(v17, T2D, v16, v20);
4616 uzp2(v21, T2D, v16, v20);
4617 eor(v16, T16B, v17, v21);
4618
4619 ushll2(v20, T2D, v16, T4S, 16);
4620 ushll(v16, T2D, v16, T2S, 16);
4621
4622 eor(v20, T16B, v22, v20);
4623 eor(v16, T16B, v16, v18);
4624
4625 uzp1(v17, T2D, v20, v16);
4626 uzp2(v21, T2D, v20, v16);
4627 eor(v20, T16B, v17, v21);
4628
4629 shl(v16, T2D, v28, 1);
4630 shl(v17, T2D, v20, 1);
4631
4632 eor(v0, T16B, v0, v16);
4633 eor(v1, T16B, v1, v17);
4634
4635 subs(len, len, 32);
4636 br(Assembler::GE, L_fold);
4637
4638 mov(crc, 0);
4639 mov(tmp, v0, D, 0);
4640 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false);
4641 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true);
4642 mov(tmp, v0, D, 1);
4643 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false);
4644 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true);
4645 mov(tmp, v1, D, 0);
4646 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false);
4647 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true);
4648 mov(tmp, v1, D, 1);
4649 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false);
4650 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true);
4651
4652 add(len, len, 32);
4653 } // Neon code end
4654
4655 BIND(L_by16);
4656 subs(len, len, 16);
4657 br(Assembler::GE, L_by16_loop);
4658 adds(len, len, 16-4);
4659 br(Assembler::GE, L_by4_loop);
4660 adds(len, len, 4);
4661 br(Assembler::GT, L_by1_loop);
4662 b(L_exit);
4663
4664 BIND(L_by4_loop);
4665 ldrw(tmp, Address(post(buf, 4)));
4666 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3);
4667 subs(len, len, 4);
4668 br(Assembler::GE, L_by4_loop);
4669 adds(len, len, 4);
4670 br(Assembler::LE, L_exit);
4671 BIND(L_by1_loop);
4672 subs(len, len, 1);
4673 ldrb(tmp, Address(post(buf, 1)));
4674 update_byte_crc32(crc, tmp, table0);
4675 br(Assembler::GT, L_by1_loop);
4676 b(L_exit);
4677
4678 align(CodeEntryAlignment);
4679 BIND(L_by16_loop);
4680 subs(len, len, 16);
4681 ldp(tmp, tmp3, Address(post(buf, 16)));
4682 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, false);
4683 update_word_crc32(crc, tmp, tmp2, table0, table1, table2, table3, true);
4684 update_word_crc32(crc, tmp3, tmp2, table0, table1, table2, table3, false);
4685 update_word_crc32(crc, tmp3, tmp2, table0, table1, table2, table3, true);
4686 br(Assembler::GE, L_by16_loop);
4687 adds(len, len, 16-4);
4688 br(Assembler::GE, L_by4_loop);
4689 adds(len, len, 4);
4690 br(Assembler::GT, L_by1_loop);
4691 BIND(L_exit);
4692 mvnw(crc, crc);
4693 }
4694
4695 void MacroAssembler::kernel_crc32c_using_crypto_pmull(Register crc, Register buf,
4696 Register len, Register tmp0, Register tmp1, Register tmp2, Register tmp3) {
4697 Label CRC_by4_loop, CRC_by1_loop, CRC_less128, CRC_by128_pre, CRC_by32_loop, CRC_less32, L_exit;
4698 assert_different_registers(crc, buf, len, tmp0, tmp1, tmp2);
4699
4700 subs(tmp0, len, 384);
4701 br(Assembler::GE, CRC_by128_pre);
4702 BIND(CRC_less128);
4703 subs(len, len, 32);
4704 br(Assembler::GE, CRC_by32_loop);
4705 BIND(CRC_less32);
4706 adds(len, len, 32 - 4);
4707 br(Assembler::GE, CRC_by4_loop);
4708 adds(len, len, 4);
4709 br(Assembler::GT, CRC_by1_loop);
4710 b(L_exit);
4711
4712 BIND(CRC_by32_loop);
4713 ldp(tmp0, tmp1, Address(buf));
4714 crc32cx(crc, crc, tmp0);
4715 ldr(tmp2, Address(buf, 16));
4716 crc32cx(crc, crc, tmp1);
4717 ldr(tmp3, Address(buf, 24));
4718 crc32cx(crc, crc, tmp2);
4719 add(buf, buf, 32);
4720 subs(len, len, 32);
4721 crc32cx(crc, crc, tmp3);
4722 br(Assembler::GE, CRC_by32_loop);
4723 cmn(len, (u1)32);
4724 br(Assembler::NE, CRC_less32);
4725 b(L_exit);
4726
4727 BIND(CRC_by4_loop);
4728 ldrw(tmp0, Address(post(buf, 4)));
4729 subs(len, len, 4);
4730 crc32cw(crc, crc, tmp0);
4731 br(Assembler::GE, CRC_by4_loop);
4732 adds(len, len, 4);
4733 br(Assembler::LE, L_exit);
4734 BIND(CRC_by1_loop);
4735 ldrb(tmp0, Address(post(buf, 1)));
4736 subs(len, len, 1);
4737 crc32cb(crc, crc, tmp0);
4738 br(Assembler::GT, CRC_by1_loop);
4739 b(L_exit);
4740
4741 BIND(CRC_by128_pre);
4742 kernel_crc32_common_fold_using_crypto_pmull(crc, buf, len, tmp0, tmp1, tmp2,
4743 4*256*sizeof(juint) + 8*sizeof(juint) + 0x50);
4744 mov(crc, 0);
4745 crc32cx(crc, crc, tmp0);
4746 crc32cx(crc, crc, tmp1);
4747
4748 cbnz(len, CRC_less128);
4749
4750 BIND(L_exit);
4751 }
4752
4753 void MacroAssembler::kernel_crc32c_using_crc32c(Register crc, Register buf,
4754 Register len, Register tmp0, Register tmp1, Register tmp2,
4755 Register tmp3) {
4756 Label CRC_by64_loop, CRC_by4_loop, CRC_by1_loop, CRC_less64, CRC_by64_pre, CRC_by32_loop, CRC_less32, L_exit;
4757 assert_different_registers(crc, buf, len, tmp0, tmp1, tmp2, tmp3);
4758
4759 subs(len, len, 128);
4760 br(Assembler::GE, CRC_by64_pre);
4761 BIND(CRC_less64);
4762 adds(len, len, 128-32);
4763 br(Assembler::GE, CRC_by32_loop);
4764 BIND(CRC_less32);
4765 adds(len, len, 32-4);
4766 br(Assembler::GE, CRC_by4_loop);
4767 adds(len, len, 4);
4768 br(Assembler::GT, CRC_by1_loop);
4769 b(L_exit);
4770
4771 BIND(CRC_by32_loop);
4772 ldp(tmp0, tmp1, Address(post(buf, 16)));
4773 subs(len, len, 32);
4774 crc32cx(crc, crc, tmp0);
4775 ldr(tmp2, Address(post(buf, 8)));
4776 crc32cx(crc, crc, tmp1);
4777 ldr(tmp3, Address(post(buf, 8)));
4778 crc32cx(crc, crc, tmp2);
4779 crc32cx(crc, crc, tmp3);
4780 br(Assembler::GE, CRC_by32_loop);
4781 cmn(len, (u1)32);
4782 br(Assembler::NE, CRC_less32);
4783 b(L_exit);
4784
4785 BIND(CRC_by4_loop);
4786 ldrw(tmp0, Address(post(buf, 4)));
4787 subs(len, len, 4);
4788 crc32cw(crc, crc, tmp0);
4789 br(Assembler::GE, CRC_by4_loop);
4790 adds(len, len, 4);
4791 br(Assembler::LE, L_exit);
4792 BIND(CRC_by1_loop);
4793 ldrb(tmp0, Address(post(buf, 1)));
4794 subs(len, len, 1);
4795 crc32cb(crc, crc, tmp0);
4796 br(Assembler::GT, CRC_by1_loop);
4797 b(L_exit);
4798
4799 BIND(CRC_by64_pre);
4800 sub(buf, buf, 8);
4801 ldp(tmp0, tmp1, Address(buf, 8));
4802 crc32cx(crc, crc, tmp0);
4803 ldr(tmp2, Address(buf, 24));
4804 crc32cx(crc, crc, tmp1);
4805 ldr(tmp3, Address(buf, 32));
4806 crc32cx(crc, crc, tmp2);
4807 ldr(tmp0, Address(buf, 40));
4808 crc32cx(crc, crc, tmp3);
4809 ldr(tmp1, Address(buf, 48));
4810 crc32cx(crc, crc, tmp0);
4811 ldr(tmp2, Address(buf, 56));
4812 crc32cx(crc, crc, tmp1);
4813 ldr(tmp3, Address(pre(buf, 64)));
4814
4815 b(CRC_by64_loop);
4816
4817 align(CodeEntryAlignment);
4818 BIND(CRC_by64_loop);
4819 subs(len, len, 64);
4820 crc32cx(crc, crc, tmp2);
4821 ldr(tmp0, Address(buf, 8));
4822 crc32cx(crc, crc, tmp3);
4823 ldr(tmp1, Address(buf, 16));
4824 crc32cx(crc, crc, tmp0);
4825 ldr(tmp2, Address(buf, 24));
4826 crc32cx(crc, crc, tmp1);
4827 ldr(tmp3, Address(buf, 32));
4828 crc32cx(crc, crc, tmp2);
4829 ldr(tmp0, Address(buf, 40));
4830 crc32cx(crc, crc, tmp3);
4831 ldr(tmp1, Address(buf, 48));
4832 crc32cx(crc, crc, tmp0);
4833 ldr(tmp2, Address(buf, 56));
4834 crc32cx(crc, crc, tmp1);
4835 ldr(tmp3, Address(pre(buf, 64)));
4836 br(Assembler::GE, CRC_by64_loop);
4837
4838 // post-loop
4839 crc32cx(crc, crc, tmp2);
4840 crc32cx(crc, crc, tmp3);
4841
4842 sub(len, len, 64);
4843 add(buf, buf, 8);
4844 cmn(len, (u1)128);
4845 br(Assembler::NE, CRC_less64);
4846 BIND(L_exit);
4847 }
4848
4849 /**
4850 * @param crc register containing existing CRC (32-bit)
4851 * @param buf register pointing to input byte buffer (byte*)
4852 * @param len register containing number of bytes
4853 * @param table register that will contain address of CRC table
4854 * @param tmp scratch register
4855 */
4856 void MacroAssembler::kernel_crc32c(Register crc, Register buf, Register len,
4857 Register table0, Register table1, Register table2, Register table3,
4858 Register tmp, Register tmp2, Register tmp3) {
4859 if (UseCryptoPmullForCRC32) {
4860 kernel_crc32c_using_crypto_pmull(crc, buf, len, table0, table1, table2, table3);
4861 } else {
4862 kernel_crc32c_using_crc32c(crc, buf, len, table0, table1, table2, table3);
4863 }
4864 }
4865
4866 void MacroAssembler::kernel_crc32_common_fold_using_crypto_pmull(Register crc, Register buf,
4867 Register len, Register tmp0, Register tmp1, Register tmp2, size_t table_offset) {
4868 Label CRC_by128_loop;
4869 assert_different_registers(crc, buf, len, tmp0, tmp1, tmp2);
4870
4871 sub(len, len, 256);
4872 Register table = tmp0;
4873 {
4874 uint64_t offset;
4875 adrp(table, ExternalAddress(StubRoutines::crc_table_addr()), offset);
4876 add(table, table, offset);
4877 }
4878 add(table, table, table_offset);
4879
4880 // Registers v0..v7 are used as data registers.
4881 // Registers v16..v31 are used as tmp registers.
4882 sub(buf, buf, 0x10);
4883 ldrq(v0, Address(buf, 0x10));
4884 ldrq(v1, Address(buf, 0x20));
4885 ldrq(v2, Address(buf, 0x30));
4886 ldrq(v3, Address(buf, 0x40));
4887 ldrq(v4, Address(buf, 0x50));
4888 ldrq(v5, Address(buf, 0x60));
4889 ldrq(v6, Address(buf, 0x70));
4890 ldrq(v7, Address(pre(buf, 0x80)));
4891
4892 movi(v31, T4S, 0);
4893 mov(v31, S, 0, crc);
4894 eor(v0, T16B, v0, v31);
4895
4896 // Register v16 contains constants from the crc table.
4897 ldrq(v16, Address(table));
4898 b(CRC_by128_loop);
4899
4900 align(OptoLoopAlignment);
4901 BIND(CRC_by128_loop);
4902 pmull (v17, T1Q, v0, v16, T1D);
4903 pmull2(v18, T1Q, v0, v16, T2D);
4904 ldrq(v0, Address(buf, 0x10));
4905 eor3(v0, T16B, v17, v18, v0);
4906
4907 pmull (v19, T1Q, v1, v16, T1D);
4908 pmull2(v20, T1Q, v1, v16, T2D);
4909 ldrq(v1, Address(buf, 0x20));
4910 eor3(v1, T16B, v19, v20, v1);
4911
4912 pmull (v21, T1Q, v2, v16, T1D);
4913 pmull2(v22, T1Q, v2, v16, T2D);
4914 ldrq(v2, Address(buf, 0x30));
4915 eor3(v2, T16B, v21, v22, v2);
4916
4917 pmull (v23, T1Q, v3, v16, T1D);
4918 pmull2(v24, T1Q, v3, v16, T2D);
4919 ldrq(v3, Address(buf, 0x40));
4920 eor3(v3, T16B, v23, v24, v3);
4921
4922 pmull (v25, T1Q, v4, v16, T1D);
4923 pmull2(v26, T1Q, v4, v16, T2D);
4924 ldrq(v4, Address(buf, 0x50));
4925 eor3(v4, T16B, v25, v26, v4);
4926
4927 pmull (v27, T1Q, v5, v16, T1D);
4928 pmull2(v28, T1Q, v5, v16, T2D);
4929 ldrq(v5, Address(buf, 0x60));
4930 eor3(v5, T16B, v27, v28, v5);
4931
4932 pmull (v29, T1Q, v6, v16, T1D);
4933 pmull2(v30, T1Q, v6, v16, T2D);
4934 ldrq(v6, Address(buf, 0x70));
4935 eor3(v6, T16B, v29, v30, v6);
4936
4937 // Reuse registers v23, v24.
4938 // Using them won't block the first instruction of the next iteration.
4939 pmull (v23, T1Q, v7, v16, T1D);
4940 pmull2(v24, T1Q, v7, v16, T2D);
4941 ldrq(v7, Address(pre(buf, 0x80)));
4942 eor3(v7, T16B, v23, v24, v7);
4943
4944 subs(len, len, 0x80);
4945 br(Assembler::GE, CRC_by128_loop);
4946
4947 // fold into 512 bits
4948 // Use v31 for constants because v16 can be still in use.
4949 ldrq(v31, Address(table, 0x10));
4950
4951 pmull (v17, T1Q, v0, v31, T1D);
4952 pmull2(v18, T1Q, v0, v31, T2D);
4953 eor3(v0, T16B, v17, v18, v4);
4954
4955 pmull (v19, T1Q, v1, v31, T1D);
4956 pmull2(v20, T1Q, v1, v31, T2D);
4957 eor3(v1, T16B, v19, v20, v5);
4958
4959 pmull (v21, T1Q, v2, v31, T1D);
4960 pmull2(v22, T1Q, v2, v31, T2D);
4961 eor3(v2, T16B, v21, v22, v6);
4962
4963 pmull (v23, T1Q, v3, v31, T1D);
4964 pmull2(v24, T1Q, v3, v31, T2D);
4965 eor3(v3, T16B, v23, v24, v7);
4966
4967 // fold into 128 bits
4968 // Use v17 for constants because v31 can be still in use.
4969 ldrq(v17, Address(table, 0x20));
4970 pmull (v25, T1Q, v0, v17, T1D);
4971 pmull2(v26, T1Q, v0, v17, T2D);
4972 eor3(v3, T16B, v3, v25, v26);
4973
4974 // Use v18 for constants because v17 can be still in use.
4975 ldrq(v18, Address(table, 0x30));
4976 pmull (v27, T1Q, v1, v18, T1D);
4977 pmull2(v28, T1Q, v1, v18, T2D);
4978 eor3(v3, T16B, v3, v27, v28);
4979
4980 // Use v19 for constants because v18 can be still in use.
4981 ldrq(v19, Address(table, 0x40));
4982 pmull (v29, T1Q, v2, v19, T1D);
4983 pmull2(v30, T1Q, v2, v19, T2D);
4984 eor3(v0, T16B, v3, v29, v30);
4985
4986 add(len, len, 0x80);
4987 add(buf, buf, 0x10);
4988
4989 mov(tmp0, v0, D, 0);
4990 mov(tmp1, v0, D, 1);
4991 }
4992
4993 void MacroAssembler::addptr(const Address &dst, int32_t src) {
4994 Address adr;
4995 switch(dst.getMode()) {
4996 case Address::base_plus_offset:
4997 // This is the expected mode, although we allow all the other
4998 // forms below.
4999 adr = form_address(rscratch2, dst.base(), dst.offset(), LogBytesPerWord);
5000 break;
5001 default:
5002 lea(rscratch2, dst);
5003 adr = Address(rscratch2);
5004 break;
5005 }
5006 ldr(rscratch1, adr);
5007 add(rscratch1, rscratch1, src);
5008 str(rscratch1, adr);
5009 }
5010
5011 void MacroAssembler::cmpptr(Register src1, Address src2) {
5012 uint64_t offset;
5013 adrp(rscratch1, src2, offset);
5014 ldr(rscratch1, Address(rscratch1, offset));
5015 cmp(src1, rscratch1);
5016 }
5017
5018 void MacroAssembler::cmpoop(Register obj1, Register obj2) {
5019 cmp(obj1, obj2);
5020 }
5021
5022 void MacroAssembler::load_method_holder_cld(Register rresult, Register rmethod) {
5023 load_method_holder(rresult, rmethod);
5024 ldr(rresult, Address(rresult, InstanceKlass::class_loader_data_offset()));
5025 }
5026
5027 void MacroAssembler::load_method_holder(Register holder, Register method) {
5028 ldr(holder, Address(method, Method::const_offset())); // ConstMethod*
5029 ldr(holder, Address(holder, ConstMethod::constants_offset())); // ConstantPool*
5030 ldr(holder, Address(holder, ConstantPool::pool_holder_offset())); // InstanceKlass*
5031 }
5032
5033 // Loads the obj's Klass* into dst.
5034 // Preserves all registers (incl src, rscratch1 and rscratch2).
5035 // Input:
5036 // src - the oop we want to load the klass from.
5037 // dst - output narrow klass.
5038 void MacroAssembler::load_narrow_klass_compact(Register dst, Register src) {
5039 assert(UseCompactObjectHeaders, "expects UseCompactObjectHeaders");
5040 ldr(dst, Address(src, oopDesc::mark_offset_in_bytes()));
5041 lsr(dst, dst, markWord::klass_shift);
5042 }
5043
5044 void MacroAssembler::load_klass(Register dst, Register src) {
5045 if (UseCompactObjectHeaders) {
5046 load_narrow_klass_compact(dst, src);
5047 decode_klass_not_null(dst);
5048 } else if (UseCompressedClassPointers) {
5049 ldrw(dst, Address(src, oopDesc::klass_offset_in_bytes()));
5050 decode_klass_not_null(dst);
5051 } else {
5052 ldr(dst, Address(src, oopDesc::klass_offset_in_bytes()));
5053 }
5054 }
5055
5056 void MacroAssembler::restore_cpu_control_state_after_jni(Register tmp1, Register tmp2) {
5057 if (RestoreMXCSROnJNICalls) {
5058 Label OK;
5059 get_fpcr(tmp1);
5060 mov(tmp2, tmp1);
5061 // Set FPCR to the state we need. We do want Round to Nearest. We
5062 // don't want non-IEEE rounding modes or floating-point traps.
5063 bfi(tmp1, zr, 22, 4); // Clear DN, FZ, and Rmode
5064 bfi(tmp1, zr, 8, 5); // Clear exception-control bits (8-12)
5065 bfi(tmp1, zr, 0, 2); // Clear AH:FIZ
5066 eor(tmp2, tmp1, tmp2);
5067 cbz(tmp2, OK); // Only reset FPCR if it's wrong
5068 set_fpcr(tmp1);
5069 bind(OK);
5070 }
5071 }
5072
5073 // ((OopHandle)result).resolve();
5074 void MacroAssembler::resolve_oop_handle(Register result, Register tmp1, Register tmp2) {
5075 // OopHandle::resolve is an indirection.
5076 access_load_at(T_OBJECT, IN_NATIVE, result, Address(result, 0), tmp1, tmp2);
5077 }
5078
5079 // ((WeakHandle)result).resolve();
5080 void MacroAssembler::resolve_weak_handle(Register result, Register tmp1, Register tmp2) {
5081 assert_different_registers(result, tmp1, tmp2);
5082 Label resolved;
5083
5084 // A null weak handle resolves to null.
5085 cbz(result, resolved);
5086
5087 // Only 64 bit platforms support GCs that require a tmp register
5088 // WeakHandle::resolve is an indirection like jweak.
5089 access_load_at(T_OBJECT, IN_NATIVE | ON_PHANTOM_OOP_REF,
5090 result, Address(result), tmp1, tmp2);
5091 bind(resolved);
5092 }
5093
5094 void MacroAssembler::load_mirror(Register dst, Register method, Register tmp1, Register tmp2) {
5095 const int mirror_offset = in_bytes(Klass::java_mirror_offset());
5096 ldr(dst, Address(rmethod, Method::const_offset()));
5097 ldr(dst, Address(dst, ConstMethod::constants_offset()));
5098 ldr(dst, Address(dst, ConstantPool::pool_holder_offset()));
5099 ldr(dst, Address(dst, mirror_offset));
5100 resolve_oop_handle(dst, tmp1, tmp2);
5101 }
5102
5103 void MacroAssembler::cmp_klass(Register obj, Register klass, Register tmp) {
5104 assert_different_registers(obj, klass, tmp);
5105 if (UseCompressedClassPointers) {
5106 if (UseCompactObjectHeaders) {
5107 load_narrow_klass_compact(tmp, obj);
5108 } else {
5109 ldrw(tmp, Address(obj, oopDesc::klass_offset_in_bytes()));
5110 }
5111 if (CompressedKlassPointers::base() == nullptr) {
5112 cmp(klass, tmp, LSL, CompressedKlassPointers::shift());
5113 return;
5114 } else if (((uint64_t)CompressedKlassPointers::base() & 0xffffffff) == 0
5115 && CompressedKlassPointers::shift() == 0) {
5116 // Only the bottom 32 bits matter
5117 cmpw(klass, tmp);
5118 return;
5119 }
5120 decode_klass_not_null(tmp);
5121 } else {
5122 ldr(tmp, Address(obj, oopDesc::klass_offset_in_bytes()));
5123 }
5124 cmp(klass, tmp);
5125 }
5126
5127 void MacroAssembler::cmp_klasses_from_objects(Register obj1, Register obj2, Register tmp1, Register tmp2) {
5128 if (UseCompactObjectHeaders) {
5129 load_narrow_klass_compact(tmp1, obj1);
5130 load_narrow_klass_compact(tmp2, obj2);
5131 cmpw(tmp1, tmp2);
5132 } else if (UseCompressedClassPointers) {
5133 ldrw(tmp1, Address(obj1, oopDesc::klass_offset_in_bytes()));
5134 ldrw(tmp2, Address(obj2, oopDesc::klass_offset_in_bytes()));
5135 cmpw(tmp1, tmp2);
5136 } else {
5137 ldr(tmp1, Address(obj1, oopDesc::klass_offset_in_bytes()));
5138 ldr(tmp2, Address(obj2, oopDesc::klass_offset_in_bytes()));
5139 cmp(tmp1, tmp2);
5140 }
5141 }
5142
5143 void MacroAssembler::store_klass(Register dst, Register src) {
5144 // FIXME: Should this be a store release? concurrent gcs assumes
5145 // klass length is valid if klass field is not null.
5146 assert(!UseCompactObjectHeaders, "not with compact headers");
5147 if (UseCompressedClassPointers) {
5148 encode_klass_not_null(src);
5149 strw(src, Address(dst, oopDesc::klass_offset_in_bytes()));
5150 } else {
5151 str(src, Address(dst, oopDesc::klass_offset_in_bytes()));
5152 }
5153 }
5154
5155 void MacroAssembler::store_klass_gap(Register dst, Register src) {
5156 assert(!UseCompactObjectHeaders, "not with compact headers");
5157 if (UseCompressedClassPointers) {
5158 // Store to klass gap in destination
5159 strw(src, Address(dst, oopDesc::klass_gap_offset_in_bytes()));
5160 }
5161 }
5162
5163 // Algorithm must match CompressedOops::encode.
5164 void MacroAssembler::encode_heap_oop(Register d, Register s) {
5165 #ifdef ASSERT
5166 verify_heapbase("MacroAssembler::encode_heap_oop: heap base corrupted?");
5167 #endif
5168 verify_oop_msg(s, "broken oop in encode_heap_oop");
5169 if (CompressedOops::base() == nullptr) {
5170 if (CompressedOops::shift() != 0) {
5171 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
5172 lsr(d, s, LogMinObjAlignmentInBytes);
5173 } else {
5174 mov(d, s);
5175 }
5176 } else {
5177 subs(d, s, rheapbase);
5178 csel(d, d, zr, Assembler::HS);
5179 lsr(d, d, LogMinObjAlignmentInBytes);
5180
5181 /* Old algorithm: is this any worse?
5182 Label nonnull;
5183 cbnz(r, nonnull);
5184 sub(r, r, rheapbase);
5185 bind(nonnull);
5186 lsr(r, r, LogMinObjAlignmentInBytes);
5187 */
5188 }
5189 }
5190
5191 void MacroAssembler::encode_heap_oop_not_null(Register r) {
5192 #ifdef ASSERT
5193 verify_heapbase("MacroAssembler::encode_heap_oop_not_null: heap base corrupted?");
5194 if (CheckCompressedOops) {
5195 Label ok;
5196 cbnz(r, ok);
5197 stop("null oop passed to encode_heap_oop_not_null");
5198 bind(ok);
5199 }
5200 #endif
5201 verify_oop_msg(r, "broken oop in encode_heap_oop_not_null");
5202 if (CompressedOops::base() != nullptr) {
5203 sub(r, r, rheapbase);
5204 }
5205 if (CompressedOops::shift() != 0) {
5206 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
5207 lsr(r, r, LogMinObjAlignmentInBytes);
5208 }
5209 }
5210
5211 void MacroAssembler::encode_heap_oop_not_null(Register dst, Register src) {
5212 #ifdef ASSERT
5213 verify_heapbase("MacroAssembler::encode_heap_oop_not_null2: heap base corrupted?");
5214 if (CheckCompressedOops) {
5215 Label ok;
5216 cbnz(src, ok);
5217 stop("null oop passed to encode_heap_oop_not_null2");
5218 bind(ok);
5219 }
5220 #endif
5221 verify_oop_msg(src, "broken oop in encode_heap_oop_not_null2");
5222
5223 Register data = src;
5224 if (CompressedOops::base() != nullptr) {
5225 sub(dst, src, rheapbase);
5226 data = dst;
5227 }
5228 if (CompressedOops::shift() != 0) {
5229 assert (LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
5230 lsr(dst, data, LogMinObjAlignmentInBytes);
5231 data = dst;
5232 }
5233 if (data == src)
5234 mov(dst, src);
5235 }
5236
5237 void MacroAssembler::decode_heap_oop(Register d, Register s) {
5238 #ifdef ASSERT
5239 verify_heapbase("MacroAssembler::decode_heap_oop: heap base corrupted?");
5240 #endif
5241 if (CompressedOops::base() == nullptr) {
5242 if (CompressedOops::shift() != 0) {
5243 lsl(d, s, CompressedOops::shift());
5244 } else if (d != s) {
5245 mov(d, s);
5246 }
5247 } else {
5248 Label done;
5249 if (d != s)
5250 mov(d, s);
5251 cbz(s, done);
5252 add(d, rheapbase, s, Assembler::LSL, LogMinObjAlignmentInBytes);
5253 bind(done);
5254 }
5255 verify_oop_msg(d, "broken oop in decode_heap_oop");
5256 }
5257
5258 void MacroAssembler::decode_heap_oop_not_null(Register r) {
5259 assert (UseCompressedOops, "should only be used for compressed headers");
5260 assert (Universe::heap() != nullptr, "java heap should be initialized");
5261 // Cannot assert, unverified entry point counts instructions (see .ad file)
5262 // vtableStubs also counts instructions in pd_code_size_limit.
5263 // Also do not verify_oop as this is called by verify_oop.
5264 if (CompressedOops::shift() != 0) {
5265 assert(LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
5266 if (CompressedOops::base() != nullptr) {
5267 add(r, rheapbase, r, Assembler::LSL, LogMinObjAlignmentInBytes);
5268 } else {
5269 add(r, zr, r, Assembler::LSL, LogMinObjAlignmentInBytes);
5270 }
5271 } else {
5272 assert (CompressedOops::base() == nullptr, "sanity");
5273 }
5274 }
5275
5276 void MacroAssembler::decode_heap_oop_not_null(Register dst, Register src) {
5277 assert (UseCompressedOops, "should only be used for compressed headers");
5278 assert (Universe::heap() != nullptr, "java heap should be initialized");
5279 // Cannot assert, unverified entry point counts instructions (see .ad file)
5280 // vtableStubs also counts instructions in pd_code_size_limit.
5281 // Also do not verify_oop as this is called by verify_oop.
5282 if (CompressedOops::shift() != 0) {
5283 assert(LogMinObjAlignmentInBytes == CompressedOops::shift(), "decode alg wrong");
5284 if (CompressedOops::base() != nullptr) {
5285 add(dst, rheapbase, src, Assembler::LSL, LogMinObjAlignmentInBytes);
5286 } else {
5287 add(dst, zr, src, Assembler::LSL, LogMinObjAlignmentInBytes);
5288 }
5289 } else {
5290 assert (CompressedOops::base() == nullptr, "sanity");
5291 if (dst != src) {
5292 mov(dst, src);
5293 }
5294 }
5295 }
5296
5297 MacroAssembler::KlassDecodeMode MacroAssembler::_klass_decode_mode(KlassDecodeNone);
5298
5299 MacroAssembler::KlassDecodeMode MacroAssembler::klass_decode_mode() {
5300 assert(Metaspace::initialized(), "metaspace not initialized yet");
5301 assert(_klass_decode_mode != KlassDecodeNone, "should be initialized");
5302 return _klass_decode_mode;
5303 }
5304
5305 MacroAssembler::KlassDecodeMode MacroAssembler::klass_decode_mode(address base, int shift, const size_t range) {
5306 assert(UseCompressedClassPointers, "not using compressed class pointers");
5307
5308 // KlassDecodeMode shouldn't be set already.
5309 assert(_klass_decode_mode == KlassDecodeNone, "set once");
5310
5311 if (base == nullptr) {
5312 return KlassDecodeZero;
5313 }
5314
5315 if (operand_valid_for_logical_immediate(
5316 /*is32*/false, (uint64_t)base)) {
5317 const uint64_t range_mask = right_n_bits(log2i_ceil(range));
5318 if (((uint64_t)base & range_mask) == 0) {
5319 return KlassDecodeXor;
5320 }
5321 }
5322
5323 const uint64_t shifted_base =
5324 (uint64_t)base >> shift;
5325 if ((shifted_base & 0xffff0000ffffffff) == 0) {
5326 return KlassDecodeMovk;
5327 }
5328
5329 // No valid encoding.
5330 return KlassDecodeNone;
5331 }
5332
5333 // Check if one of the above decoding modes will work for given base, shift and range.
5334 bool MacroAssembler::check_klass_decode_mode(address base, int shift, const size_t range) {
5335 return klass_decode_mode(base, shift, range) != KlassDecodeNone;
5336 }
5337
5338 bool MacroAssembler::set_klass_decode_mode(address base, int shift, const size_t range) {
5339 _klass_decode_mode = klass_decode_mode(base, shift, range);
5340 return _klass_decode_mode != KlassDecodeNone;
5341 }
5342
5343 void MacroAssembler::encode_klass_not_null(Register dst, Register src) {
5344 switch (klass_decode_mode()) {
5345 case KlassDecodeZero:
5346 if (CompressedKlassPointers::shift() != 0) {
5347 lsr(dst, src, CompressedKlassPointers::shift());
5348 } else {
5349 if (dst != src) mov(dst, src);
5350 }
5351 break;
5352
5353 case KlassDecodeXor:
5354 if (CompressedKlassPointers::shift() != 0) {
5355 eor(dst, src, (uint64_t)CompressedKlassPointers::base());
5356 lsr(dst, dst, CompressedKlassPointers::shift());
5357 } else {
5358 eor(dst, src, (uint64_t)CompressedKlassPointers::base());
5359 }
5360 break;
5361
5362 case KlassDecodeMovk:
5363 if (CompressedKlassPointers::shift() != 0) {
5364 ubfx(dst, src, CompressedKlassPointers::shift(), 32);
5365 } else {
5366 movw(dst, src);
5367 }
5368 break;
5369
5370 case KlassDecodeNone:
5371 ShouldNotReachHere();
5372 break;
5373 }
5374 }
5375
5376 void MacroAssembler::encode_klass_not_null(Register r) {
5377 encode_klass_not_null(r, r);
5378 }
5379
5380 void MacroAssembler::decode_klass_not_null(Register dst, Register src) {
5381 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5382
5383 switch (klass_decode_mode()) {
5384 case KlassDecodeZero:
5385 if (CompressedKlassPointers::shift() != 0) {
5386 lsl(dst, src, CompressedKlassPointers::shift());
5387 } else {
5388 if (dst != src) mov(dst, src);
5389 }
5390 break;
5391
5392 case KlassDecodeXor:
5393 if (CompressedKlassPointers::shift() != 0) {
5394 lsl(dst, src, CompressedKlassPointers::shift());
5395 eor(dst, dst, (uint64_t)CompressedKlassPointers::base());
5396 } else {
5397 eor(dst, src, (uint64_t)CompressedKlassPointers::base());
5398 }
5399 break;
5400
5401 case KlassDecodeMovk: {
5402 const uint64_t shifted_base =
5403 (uint64_t)CompressedKlassPointers::base() >> CompressedKlassPointers::shift();
5404
5405 if (dst != src) movw(dst, src);
5406 movk(dst, shifted_base >> 32, 32);
5407
5408 if (CompressedKlassPointers::shift() != 0) {
5409 lsl(dst, dst, CompressedKlassPointers::shift());
5410 }
5411
5412 break;
5413 }
5414
5415 case KlassDecodeNone:
5416 ShouldNotReachHere();
5417 break;
5418 }
5419 }
5420
5421 void MacroAssembler::decode_klass_not_null(Register r) {
5422 decode_klass_not_null(r, r);
5423 }
5424
5425 void MacroAssembler::set_narrow_oop(Register dst, jobject obj) {
5426 #ifdef ASSERT
5427 {
5428 ThreadInVMfromUnknown tiv;
5429 assert (UseCompressedOops, "should only be used for compressed oops");
5430 assert (Universe::heap() != nullptr, "java heap should be initialized");
5431 assert (oop_recorder() != nullptr, "this assembler needs an OopRecorder");
5432 assert(Universe::heap()->is_in(JNIHandles::resolve(obj)), "should be real oop");
5433 }
5434 #endif
5435 int oop_index = oop_recorder()->find_index(obj);
5436 InstructionMark im(this);
5437 RelocationHolder rspec = oop_Relocation::spec(oop_index);
5438 code_section()->relocate(inst_mark(), rspec);
5439 movz(dst, 0xDEAD, 16);
5440 movk(dst, 0xBEEF);
5441 }
5442
5443 void MacroAssembler::set_narrow_klass(Register dst, Klass* k) {
5444 assert (UseCompressedClassPointers, "should only be used for compressed headers");
5445 assert (oop_recorder() != nullptr, "this assembler needs an OopRecorder");
5446 int index = oop_recorder()->find_index(k);
5447 assert(! Universe::heap()->is_in(k), "should not be an oop");
5448
5449 InstructionMark im(this);
5450 RelocationHolder rspec = metadata_Relocation::spec(index);
5451 code_section()->relocate(inst_mark(), rspec);
5452 narrowKlass nk = CompressedKlassPointers::encode(k);
5453 movz(dst, (nk >> 16), 16);
5454 movk(dst, nk & 0xffff);
5455 }
5456
5457 void MacroAssembler::access_load_at(BasicType type, DecoratorSet decorators,
5458 Register dst, Address src,
5459 Register tmp1, Register tmp2) {
5460 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler();
5461 decorators = AccessInternal::decorator_fixup(decorators, type);
5462 bool as_raw = (decorators & AS_RAW) != 0;
5463 if (as_raw) {
5464 bs->BarrierSetAssembler::load_at(this, decorators, type, dst, src, tmp1, tmp2);
5465 } else {
5466 bs->load_at(this, decorators, type, dst, src, tmp1, tmp2);
5467 }
5468 }
5469
5470 void MacroAssembler::access_store_at(BasicType type, DecoratorSet decorators,
5471 Address dst, Register val,
5472 Register tmp1, Register tmp2, Register tmp3) {
5473 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler();
5474 decorators = AccessInternal::decorator_fixup(decorators, type);
5475 bool as_raw = (decorators & AS_RAW) != 0;
5476 if (as_raw) {
5477 bs->BarrierSetAssembler::store_at(this, decorators, type, dst, val, tmp1, tmp2, tmp3);
5478 } else {
5479 bs->store_at(this, decorators, type, dst, val, tmp1, tmp2, tmp3);
5480 }
5481 }
5482
5483 void MacroAssembler::load_heap_oop(Register dst, Address src, Register tmp1,
5484 Register tmp2, DecoratorSet decorators) {
5485 access_load_at(T_OBJECT, IN_HEAP | decorators, dst, src, tmp1, tmp2);
5486 }
5487
5488 void MacroAssembler::load_heap_oop_not_null(Register dst, Address src, Register tmp1,
5489 Register tmp2, DecoratorSet decorators) {
5490 access_load_at(T_OBJECT, IN_HEAP | IS_NOT_NULL | decorators, dst, src, tmp1, tmp2);
5491 }
5492
5493 void MacroAssembler::store_heap_oop(Address dst, Register val, Register tmp1,
5494 Register tmp2, Register tmp3, DecoratorSet decorators) {
5495 access_store_at(T_OBJECT, IN_HEAP | decorators, dst, val, tmp1, tmp2, tmp3);
5496 }
5497
5498 // Used for storing nulls.
5499 void MacroAssembler::store_heap_oop_null(Address dst) {
5500 access_store_at(T_OBJECT, IN_HEAP, dst, noreg, noreg, noreg, noreg);
5501 }
5502
5503 Address MacroAssembler::allocate_metadata_address(Metadata* obj) {
5504 assert(oop_recorder() != nullptr, "this assembler needs a Recorder");
5505 int index = oop_recorder()->allocate_metadata_index(obj);
5506 RelocationHolder rspec = metadata_Relocation::spec(index);
5507 return Address((address)obj, rspec);
5508 }
5509
5510 // Move an oop into a register.
5511 void MacroAssembler::movoop(Register dst, jobject obj) {
5512 int oop_index;
5513 if (obj == nullptr) {
5514 oop_index = oop_recorder()->allocate_oop_index(obj);
5515 } else {
5516 #ifdef ASSERT
5517 {
5518 ThreadInVMfromUnknown tiv;
5519 assert(Universe::heap()->is_in(JNIHandles::resolve(obj)), "should be real oop");
5520 }
5521 #endif
5522 oop_index = oop_recorder()->find_index(obj);
5523 }
5524 RelocationHolder rspec = oop_Relocation::spec(oop_index);
5525
5526 if (BarrierSet::barrier_set()->barrier_set_assembler()->supports_instruction_patching()) {
5527 mov(dst, Address((address)obj, rspec));
5528 } else {
5529 address dummy = address(uintptr_t(pc()) & -wordSize); // A nearby aligned address
5530 ldr(dst, Address(dummy, rspec));
5531 }
5532 }
5533
5534 // Move a metadata address into a register.
5535 void MacroAssembler::mov_metadata(Register dst, Metadata* obj) {
5536 int oop_index;
5537 if (obj == nullptr) {
5538 oop_index = oop_recorder()->allocate_metadata_index(obj);
5539 } else {
5540 oop_index = oop_recorder()->find_index(obj);
5541 }
5542 RelocationHolder rspec = metadata_Relocation::spec(oop_index);
5543 mov(dst, Address((address)obj, rspec));
5544 }
5545
5546 Address MacroAssembler::constant_oop_address(jobject obj) {
5547 #ifdef ASSERT
5548 {
5549 ThreadInVMfromUnknown tiv;
5550 assert(oop_recorder() != nullptr, "this assembler needs an OopRecorder");
5551 assert(Universe::heap()->is_in(JNIHandles::resolve(obj)), "not an oop");
5552 }
5553 #endif
5554 int oop_index = oop_recorder()->find_index(obj);
5555 return Address((address)obj, oop_Relocation::spec(oop_index));
5556 }
5557
5558 // Defines obj, preserves var_size_in_bytes, okay for t2 == var_size_in_bytes.
5559 void MacroAssembler::tlab_allocate(Register obj,
5560 Register var_size_in_bytes,
5561 int con_size_in_bytes,
5562 Register t1,
5563 Register t2,
5564 Label& slow_case) {
5565 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler();
5566 bs->tlab_allocate(this, obj, var_size_in_bytes, con_size_in_bytes, t1, t2, slow_case);
5567 }
5568
5569 void MacroAssembler::inc_held_monitor_count(Register tmp) {
5570 Address dst(rthread, JavaThread::held_monitor_count_offset());
5571 #ifdef ASSERT
5572 ldr(tmp, dst);
5573 increment(tmp);
5574 str(tmp, dst);
5575 Label ok;
5576 tbz(tmp, 63, ok);
5577 STOP("assert(held monitor count underflow)");
5578 should_not_reach_here();
5579 bind(ok);
5580 #else
5581 increment(dst);
5582 #endif
5583 }
5584
5585 void MacroAssembler::dec_held_monitor_count(Register tmp) {
5586 Address dst(rthread, JavaThread::held_monitor_count_offset());
5587 #ifdef ASSERT
5588 ldr(tmp, dst);
5589 decrement(tmp);
5590 str(tmp, dst);
5591 Label ok;
5592 tbz(tmp, 63, ok);
5593 STOP("assert(held monitor count underflow)");
5594 should_not_reach_here();
5595 bind(ok);
5596 #else
5597 decrement(dst);
5598 #endif
5599 }
5600
5601 void MacroAssembler::verify_tlab() {
5602 #ifdef ASSERT
5603 if (UseTLAB && VerifyOops) {
5604 Label next, ok;
5605
5606 stp(rscratch2, rscratch1, Address(pre(sp, -16)));
5607
5608 ldr(rscratch2, Address(rthread, in_bytes(JavaThread::tlab_top_offset())));
5609 ldr(rscratch1, Address(rthread, in_bytes(JavaThread::tlab_start_offset())));
5610 cmp(rscratch2, rscratch1);
5611 br(Assembler::HS, next);
5612 STOP("assert(top >= start)");
5613 should_not_reach_here();
5614
5615 bind(next);
5616 ldr(rscratch2, Address(rthread, in_bytes(JavaThread::tlab_end_offset())));
5617 ldr(rscratch1, Address(rthread, in_bytes(JavaThread::tlab_top_offset())));
5618 cmp(rscratch2, rscratch1);
5619 br(Assembler::HS, ok);
5620 STOP("assert(top <= end)");
5621 should_not_reach_here();
5622
5623 bind(ok);
5624 ldp(rscratch2, rscratch1, Address(post(sp, 16)));
5625 }
5626 #endif
5627 }
5628
5629 // Writes to stack successive pages until offset reached to check for
5630 // stack overflow + shadow pages. This clobbers tmp.
5631 void MacroAssembler::bang_stack_size(Register size, Register tmp) {
5632 assert_different_registers(tmp, size, rscratch1);
5633 mov(tmp, sp);
5634 // Bang stack for total size given plus shadow page size.
5635 // Bang one page at a time because large size can bang beyond yellow and
5636 // red zones.
5637 Label loop;
5638 mov(rscratch1, (int)os::vm_page_size());
5639 bind(loop);
5640 lea(tmp, Address(tmp, -(int)os::vm_page_size()));
5641 subsw(size, size, rscratch1);
5642 str(size, Address(tmp));
5643 br(Assembler::GT, loop);
5644
5645 // Bang down shadow pages too.
5646 // At this point, (tmp-0) is the last address touched, so don't
5647 // touch it again. (It was touched as (tmp-pagesize) but then tmp
5648 // was post-decremented.) Skip this address by starting at i=1, and
5649 // touch a few more pages below. N.B. It is important to touch all
5650 // the way down to and including i=StackShadowPages.
5651 for (int i = 0; i < (int)(StackOverflow::stack_shadow_zone_size() / (int)os::vm_page_size()) - 1; i++) {
5652 // this could be any sized move but this is can be a debugging crumb
5653 // so the bigger the better.
5654 lea(tmp, Address(tmp, -(int)os::vm_page_size()));
5655 str(size, Address(tmp));
5656 }
5657 }
5658
5659 // Move the address of the polling page into dest.
5660 void MacroAssembler::get_polling_page(Register dest, relocInfo::relocType rtype) {
5661 ldr(dest, Address(rthread, JavaThread::polling_page_offset()));
5662 }
5663
5664 // Read the polling page. The address of the polling page must
5665 // already be in r.
5666 address MacroAssembler::read_polling_page(Register r, relocInfo::relocType rtype) {
5667 address mark;
5668 {
5669 InstructionMark im(this);
5670 code_section()->relocate(inst_mark(), rtype);
5671 ldrw(zr, Address(r, 0));
5672 mark = inst_mark();
5673 }
5674 verify_cross_modify_fence_not_required();
5675 return mark;
5676 }
5677
5678 void MacroAssembler::adrp(Register reg1, const Address &dest, uint64_t &byte_offset) {
5679 relocInfo::relocType rtype = dest.rspec().reloc()->type();
5680 uint64_t low_page = (uint64_t)CodeCache::low_bound() >> 12;
5681 uint64_t high_page = (uint64_t)(CodeCache::high_bound()-1) >> 12;
5682 uint64_t dest_page = (uint64_t)dest.target() >> 12;
5683 int64_t offset_low = dest_page - low_page;
5684 int64_t offset_high = dest_page - high_page;
5685
5686 assert(is_valid_AArch64_address(dest.target()), "bad address");
5687 assert(dest.getMode() == Address::literal, "ADRP must be applied to a literal address");
5688
5689 InstructionMark im(this);
5690 code_section()->relocate(inst_mark(), dest.rspec());
5691 // 8143067: Ensure that the adrp can reach the dest from anywhere within
5692 // the code cache so that if it is relocated we know it will still reach
5693 if (offset_high >= -(1<<20) && offset_low < (1<<20)) {
5694 _adrp(reg1, dest.target());
5695 } else {
5696 uint64_t target = (uint64_t)dest.target();
5697 uint64_t adrp_target
5698 = (target & 0xffffffffULL) | ((uint64_t)pc() & 0xffff00000000ULL);
5699
5700 _adrp(reg1, (address)adrp_target);
5701 movk(reg1, target >> 32, 32);
5702 }
5703 byte_offset = (uint64_t)dest.target() & 0xfff;
5704 }
5705
5706 void MacroAssembler::load_byte_map_base(Register reg) {
5707 CardTable::CardValue* byte_map_base =
5708 ((CardTableBarrierSet*)(BarrierSet::barrier_set()))->card_table()->byte_map_base();
5709
5710 // Strictly speaking the byte_map_base isn't an address at all, and it might
5711 // even be negative. It is thus materialised as a constant.
5712 mov(reg, (uint64_t)byte_map_base);
5713 }
5714
5715 void MacroAssembler::build_frame(int framesize) {
5716 assert(framesize >= 2 * wordSize, "framesize must include space for FP/LR");
5717 assert(framesize % (2*wordSize) == 0, "must preserve 2*wordSize alignment");
5718 protect_return_address();
5719 if (framesize < ((1 << 9) + 2 * wordSize)) {
5720 sub(sp, sp, framesize);
5721 stp(rfp, lr, Address(sp, framesize - 2 * wordSize));
5722 if (PreserveFramePointer) add(rfp, sp, framesize - 2 * wordSize);
5723 } else {
5724 stp(rfp, lr, Address(pre(sp, -2 * wordSize)));
5725 if (PreserveFramePointer) mov(rfp, sp);
5726 if (framesize < ((1 << 12) + 2 * wordSize))
5727 sub(sp, sp, framesize - 2 * wordSize);
5728 else {
5729 mov(rscratch1, framesize - 2 * wordSize);
5730 sub(sp, sp, rscratch1);
5731 }
5732 }
5733 verify_cross_modify_fence_not_required();
5734 }
5735
5736 void MacroAssembler::remove_frame(int framesize) {
5737 assert(framesize >= 2 * wordSize, "framesize must include space for FP/LR");
5738 assert(framesize % (2*wordSize) == 0, "must preserve 2*wordSize alignment");
5739 if (framesize < ((1 << 9) + 2 * wordSize)) {
5740 ldp(rfp, lr, Address(sp, framesize - 2 * wordSize));
5741 add(sp, sp, framesize);
5742 } else {
5743 if (framesize < ((1 << 12) + 2 * wordSize))
5744 add(sp, sp, framesize - 2 * wordSize);
5745 else {
5746 mov(rscratch1, framesize - 2 * wordSize);
5747 add(sp, sp, rscratch1);
5748 }
5749 ldp(rfp, lr, Address(post(sp, 2 * wordSize)));
5750 }
5751 authenticate_return_address();
5752 }
5753
5754
5755 // This method counts leading positive bytes (highest bit not set) in provided byte array
5756 address MacroAssembler::count_positives(Register ary1, Register len, Register result) {
5757 // Simple and most common case of aligned small array which is not at the
5758 // end of memory page is placed here. All other cases are in stub.
5759 Label LOOP, END, STUB, STUB_LONG, SET_RESULT, DONE;
5760 const uint64_t UPPER_BIT_MASK=0x8080808080808080;
5761 assert_different_registers(ary1, len, result);
5762
5763 mov(result, len);
5764 cmpw(len, 0);
5765 br(LE, DONE);
5766 cmpw(len, 4 * wordSize);
5767 br(GE, STUB_LONG); // size > 32 then go to stub
5768
5769 int shift = 64 - exact_log2(os::vm_page_size());
5770 lsl(rscratch1, ary1, shift);
5771 mov(rscratch2, (size_t)(4 * wordSize) << shift);
5772 adds(rscratch2, rscratch1, rscratch2); // At end of page?
5773 br(CS, STUB); // at the end of page then go to stub
5774 subs(len, len, wordSize);
5775 br(LT, END);
5776
5777 BIND(LOOP);
5778 ldr(rscratch1, Address(post(ary1, wordSize)));
5779 tst(rscratch1, UPPER_BIT_MASK);
5780 br(NE, SET_RESULT);
5781 subs(len, len, wordSize);
5782 br(GE, LOOP);
5783 cmpw(len, -wordSize);
5784 br(EQ, DONE);
5785
5786 BIND(END);
5787 ldr(rscratch1, Address(ary1));
5788 sub(rscratch2, zr, len, LSL, 3); // LSL 3 is to get bits from bytes
5789 lslv(rscratch1, rscratch1, rscratch2);
5790 tst(rscratch1, UPPER_BIT_MASK);
5791 br(NE, SET_RESULT);
5792 b(DONE);
5793
5794 BIND(STUB);
5795 RuntimeAddress count_pos = RuntimeAddress(StubRoutines::aarch64::count_positives());
5796 assert(count_pos.target() != nullptr, "count_positives stub has not been generated");
5797 address tpc1 = trampoline_call(count_pos);
5798 if (tpc1 == nullptr) {
5799 DEBUG_ONLY(reset_labels(STUB_LONG, SET_RESULT, DONE));
5800 postcond(pc() == badAddress);
5801 return nullptr;
5802 }
5803 b(DONE);
5804
5805 BIND(STUB_LONG);
5806 RuntimeAddress count_pos_long = RuntimeAddress(StubRoutines::aarch64::count_positives_long());
5807 assert(count_pos_long.target() != nullptr, "count_positives_long stub has not been generated");
5808 address tpc2 = trampoline_call(count_pos_long);
5809 if (tpc2 == nullptr) {
5810 DEBUG_ONLY(reset_labels(SET_RESULT, DONE));
5811 postcond(pc() == badAddress);
5812 return nullptr;
5813 }
5814 b(DONE);
5815
5816 BIND(SET_RESULT);
5817
5818 add(len, len, wordSize);
5819 sub(result, result, len);
5820
5821 BIND(DONE);
5822 postcond(pc() != badAddress);
5823 return pc();
5824 }
5825
5826 // Clobbers: rscratch1, rscratch2, rflags
5827 // May also clobber v0-v7 when (!UseSimpleArrayEquals && UseSIMDForArrayEquals)
5828 address MacroAssembler::arrays_equals(Register a1, Register a2, Register tmp3,
5829 Register tmp4, Register tmp5, Register result,
5830 Register cnt1, int elem_size) {
5831 Label DONE, SAME;
5832 Register tmp1 = rscratch1;
5833 Register tmp2 = rscratch2;
5834 int elem_per_word = wordSize/elem_size;
5835 int log_elem_size = exact_log2(elem_size);
5836 int klass_offset = arrayOopDesc::klass_offset_in_bytes();
5837 int length_offset = arrayOopDesc::length_offset_in_bytes();
5838 int base_offset
5839 = arrayOopDesc::base_offset_in_bytes(elem_size == 2 ? T_CHAR : T_BYTE);
5840 // When the length offset is not aligned to 8 bytes,
5841 // then we align it down. This is valid because the new
5842 // offset will always be the klass which is the same
5843 // for type arrays.
5844 int start_offset = align_down(length_offset, BytesPerWord);
5845 int extra_length = base_offset - start_offset;
5846 assert(start_offset == length_offset || start_offset == klass_offset,
5847 "start offset must be 8-byte-aligned or be the klass offset");
5848 assert(base_offset != start_offset, "must include the length field");
5849 extra_length = extra_length / elem_size; // We count in elements, not bytes.
5850 int stubBytesThreshold = 3 * 64 + (UseSIMDForArrayEquals ? 0 : 16);
5851
5852 assert(elem_size == 1 || elem_size == 2, "must be char or byte");
5853 assert_different_registers(a1, a2, result, cnt1, rscratch1, rscratch2);
5854
5855 #ifndef PRODUCT
5856 {
5857 const char kind = (elem_size == 2) ? 'U' : 'L';
5858 char comment[64];
5859 snprintf(comment, sizeof comment, "array_equals%c{", kind);
5860 BLOCK_COMMENT(comment);
5861 }
5862 #endif
5863
5864 // if (a1 == a2)
5865 // return true;
5866 cmpoop(a1, a2); // May have read barriers for a1 and a2.
5867 br(EQ, SAME);
5868
5869 if (UseSimpleArrayEquals) {
5870 Label NEXT_WORD, SHORT, TAIL03, TAIL01, A_MIGHT_BE_NULL, A_IS_NOT_NULL;
5871 // if (a1 == nullptr || a2 == nullptr)
5872 // return false;
5873 // a1 & a2 == 0 means (some-pointer is null) or
5874 // (very-rare-or-even-probably-impossible-pointer-values)
5875 // so, we can save one branch in most cases
5876 tst(a1, a2);
5877 mov(result, false);
5878 br(EQ, A_MIGHT_BE_NULL);
5879 // if (a1.length != a2.length)
5880 // return false;
5881 bind(A_IS_NOT_NULL);
5882 ldrw(cnt1, Address(a1, length_offset));
5883 // Increase loop counter by diff between base- and actual start-offset.
5884 addw(cnt1, cnt1, extra_length);
5885 lea(a1, Address(a1, start_offset));
5886 lea(a2, Address(a2, start_offset));
5887 // Check for short strings, i.e. smaller than wordSize.
5888 subs(cnt1, cnt1, elem_per_word);
5889 br(Assembler::LT, SHORT);
5890 // Main 8 byte comparison loop.
5891 bind(NEXT_WORD); {
5892 ldr(tmp1, Address(post(a1, wordSize)));
5893 ldr(tmp2, Address(post(a2, wordSize)));
5894 subs(cnt1, cnt1, elem_per_word);
5895 eor(tmp5, tmp1, tmp2);
5896 cbnz(tmp5, DONE);
5897 } br(GT, NEXT_WORD);
5898 // Last longword. In the case where length == 4 we compare the
5899 // same longword twice, but that's still faster than another
5900 // conditional branch.
5901 // cnt1 could be 0, -1, -2, -3, -4 for chars; -4 only happens when
5902 // length == 4.
5903 if (log_elem_size > 0)
5904 lsl(cnt1, cnt1, log_elem_size);
5905 ldr(tmp3, Address(a1, cnt1));
5906 ldr(tmp4, Address(a2, cnt1));
5907 eor(tmp5, tmp3, tmp4);
5908 cbnz(tmp5, DONE);
5909 b(SAME);
5910 bind(A_MIGHT_BE_NULL);
5911 // in case both a1 and a2 are not-null, proceed with loads
5912 cbz(a1, DONE);
5913 cbz(a2, DONE);
5914 b(A_IS_NOT_NULL);
5915 bind(SHORT);
5916
5917 tbz(cnt1, 2 - log_elem_size, TAIL03); // 0-7 bytes left.
5918 {
5919 ldrw(tmp1, Address(post(a1, 4)));
5920 ldrw(tmp2, Address(post(a2, 4)));
5921 eorw(tmp5, tmp1, tmp2);
5922 cbnzw(tmp5, DONE);
5923 }
5924 bind(TAIL03);
5925 tbz(cnt1, 1 - log_elem_size, TAIL01); // 0-3 bytes left.
5926 {
5927 ldrh(tmp3, Address(post(a1, 2)));
5928 ldrh(tmp4, Address(post(a2, 2)));
5929 eorw(tmp5, tmp3, tmp4);
5930 cbnzw(tmp5, DONE);
5931 }
5932 bind(TAIL01);
5933 if (elem_size == 1) { // Only needed when comparing byte arrays.
5934 tbz(cnt1, 0, SAME); // 0-1 bytes left.
5935 {
5936 ldrb(tmp1, a1);
5937 ldrb(tmp2, a2);
5938 eorw(tmp5, tmp1, tmp2);
5939 cbnzw(tmp5, DONE);
5940 }
5941 }
5942 } else {
5943 Label NEXT_DWORD, SHORT, TAIL, TAIL2, STUB,
5944 CSET_EQ, LAST_CHECK;
5945 mov(result, false);
5946 cbz(a1, DONE);
5947 ldrw(cnt1, Address(a1, length_offset));
5948 cbz(a2, DONE);
5949 // Increase loop counter by diff between base- and actual start-offset.
5950 addw(cnt1, cnt1, extra_length);
5951
5952 // on most CPUs a2 is still "locked"(surprisingly) in ldrw and it's
5953 // faster to perform another branch before comparing a1 and a2
5954 cmp(cnt1, (u1)elem_per_word);
5955 br(LE, SHORT); // short or same
5956 ldr(tmp3, Address(pre(a1, start_offset)));
5957 subs(zr, cnt1, stubBytesThreshold);
5958 br(GE, STUB);
5959 ldr(tmp4, Address(pre(a2, start_offset)));
5960 sub(tmp5, zr, cnt1, LSL, 3 + log_elem_size);
5961
5962 // Main 16 byte comparison loop with 2 exits
5963 bind(NEXT_DWORD); {
5964 ldr(tmp1, Address(pre(a1, wordSize)));
5965 ldr(tmp2, Address(pre(a2, wordSize)));
5966 subs(cnt1, cnt1, 2 * elem_per_word);
5967 br(LE, TAIL);
5968 eor(tmp4, tmp3, tmp4);
5969 cbnz(tmp4, DONE);
5970 ldr(tmp3, Address(pre(a1, wordSize)));
5971 ldr(tmp4, Address(pre(a2, wordSize)));
5972 cmp(cnt1, (u1)elem_per_word);
5973 br(LE, TAIL2);
5974 cmp(tmp1, tmp2);
5975 } br(EQ, NEXT_DWORD);
5976 b(DONE);
5977
5978 bind(TAIL);
5979 eor(tmp4, tmp3, tmp4);
5980 eor(tmp2, tmp1, tmp2);
5981 lslv(tmp2, tmp2, tmp5);
5982 orr(tmp5, tmp4, tmp2);
5983 cmp(tmp5, zr);
5984 b(CSET_EQ);
5985
5986 bind(TAIL2);
5987 eor(tmp2, tmp1, tmp2);
5988 cbnz(tmp2, DONE);
5989 b(LAST_CHECK);
5990
5991 bind(STUB);
5992 ldr(tmp4, Address(pre(a2, start_offset)));
5993 if (elem_size == 2) { // convert to byte counter
5994 lsl(cnt1, cnt1, 1);
5995 }
5996 eor(tmp5, tmp3, tmp4);
5997 cbnz(tmp5, DONE);
5998 RuntimeAddress stub = RuntimeAddress(StubRoutines::aarch64::large_array_equals());
5999 assert(stub.target() != nullptr, "array_equals_long stub has not been generated");
6000 address tpc = trampoline_call(stub);
6001 if (tpc == nullptr) {
6002 DEBUG_ONLY(reset_labels(SHORT, LAST_CHECK, CSET_EQ, SAME, DONE));
6003 postcond(pc() == badAddress);
6004 return nullptr;
6005 }
6006 b(DONE);
6007
6008 // (a1 != null && a2 == null) || (a1 != null && a2 != null && a1 == a2)
6009 // so, if a2 == null => return false(0), else return true, so we can return a2
6010 mov(result, a2);
6011 b(DONE);
6012 bind(SHORT);
6013 sub(tmp5, zr, cnt1, LSL, 3 + log_elem_size);
6014 ldr(tmp3, Address(a1, start_offset));
6015 ldr(tmp4, Address(a2, start_offset));
6016 bind(LAST_CHECK);
6017 eor(tmp4, tmp3, tmp4);
6018 lslv(tmp5, tmp4, tmp5);
6019 cmp(tmp5, zr);
6020 bind(CSET_EQ);
6021 cset(result, EQ);
6022 b(DONE);
6023 }
6024
6025 bind(SAME);
6026 mov(result, true);
6027 // That's it.
6028 bind(DONE);
6029
6030 BLOCK_COMMENT("} array_equals");
6031 postcond(pc() != badAddress);
6032 return pc();
6033 }
6034
6035 // Compare Strings
6036
6037 // For Strings we're passed the address of the first characters in a1
6038 // and a2 and the length in cnt1.
6039 // There are two implementations. For arrays >= 8 bytes, all
6040 // comparisons (including the final one, which may overlap) are
6041 // performed 8 bytes at a time. For strings < 8 bytes, we compare a
6042 // halfword, then a short, and then a byte.
6043
6044 void MacroAssembler::string_equals(Register a1, Register a2,
6045 Register result, Register cnt1)
6046 {
6047 Label SAME, DONE, SHORT, NEXT_WORD;
6048 Register tmp1 = rscratch1;
6049 Register tmp2 = rscratch2;
6050 Register cnt2 = tmp2; // cnt2 only used in array length compare
6051
6052 assert_different_registers(a1, a2, result, cnt1, rscratch1, rscratch2);
6053
6054 #ifndef PRODUCT
6055 {
6056 char comment[64];
6057 snprintf(comment, sizeof comment, "{string_equalsL");
6058 BLOCK_COMMENT(comment);
6059 }
6060 #endif
6061
6062 mov(result, false);
6063
6064 // Check for short strings, i.e. smaller than wordSize.
6065 subs(cnt1, cnt1, wordSize);
6066 br(Assembler::LT, SHORT);
6067 // Main 8 byte comparison loop.
6068 bind(NEXT_WORD); {
6069 ldr(tmp1, Address(post(a1, wordSize)));
6070 ldr(tmp2, Address(post(a2, wordSize)));
6071 subs(cnt1, cnt1, wordSize);
6072 eor(tmp1, tmp1, tmp2);
6073 cbnz(tmp1, DONE);
6074 } br(GT, NEXT_WORD);
6075 // Last longword. In the case where length == 4 we compare the
6076 // same longword twice, but that's still faster than another
6077 // conditional branch.
6078 // cnt1 could be 0, -1, -2, -3, -4 for chars; -4 only happens when
6079 // length == 4.
6080 ldr(tmp1, Address(a1, cnt1));
6081 ldr(tmp2, Address(a2, cnt1));
6082 eor(tmp2, tmp1, tmp2);
6083 cbnz(tmp2, DONE);
6084 b(SAME);
6085
6086 bind(SHORT);
6087 Label TAIL03, TAIL01;
6088
6089 tbz(cnt1, 2, TAIL03); // 0-7 bytes left.
6090 {
6091 ldrw(tmp1, Address(post(a1, 4)));
6092 ldrw(tmp2, Address(post(a2, 4)));
6093 eorw(tmp1, tmp1, tmp2);
6094 cbnzw(tmp1, DONE);
6095 }
6096 bind(TAIL03);
6097 tbz(cnt1, 1, TAIL01); // 0-3 bytes left.
6098 {
6099 ldrh(tmp1, Address(post(a1, 2)));
6100 ldrh(tmp2, Address(post(a2, 2)));
6101 eorw(tmp1, tmp1, tmp2);
6102 cbnzw(tmp1, DONE);
6103 }
6104 bind(TAIL01);
6105 tbz(cnt1, 0, SAME); // 0-1 bytes left.
6106 {
6107 ldrb(tmp1, a1);
6108 ldrb(tmp2, a2);
6109 eorw(tmp1, tmp1, tmp2);
6110 cbnzw(tmp1, DONE);
6111 }
6112 // Arrays are equal.
6113 bind(SAME);
6114 mov(result, true);
6115
6116 // That's it.
6117 bind(DONE);
6118 BLOCK_COMMENT("} string_equals");
6119 }
6120
6121
6122 // The size of the blocks erased by the zero_blocks stub. We must
6123 // handle anything smaller than this ourselves in zero_words().
6124 const int MacroAssembler::zero_words_block_size = 8;
6125
6126 // zero_words() is used by C2 ClearArray patterns and by
6127 // C1_MacroAssembler. It is as small as possible, handling small word
6128 // counts locally and delegating anything larger to the zero_blocks
6129 // stub. It is expanded many times in compiled code, so it is
6130 // important to keep it short.
6131
6132 // ptr: Address of a buffer to be zeroed.
6133 // cnt: Count in HeapWords.
6134 //
6135 // ptr, cnt, rscratch1, and rscratch2 are clobbered.
6136 address MacroAssembler::zero_words(Register ptr, Register cnt)
6137 {
6138 assert(is_power_of_2(zero_words_block_size), "adjust this");
6139
6140 BLOCK_COMMENT("zero_words {");
6141 assert(ptr == r10 && cnt == r11, "mismatch in register usage");
6142 RuntimeAddress zero_blocks = RuntimeAddress(StubRoutines::aarch64::zero_blocks());
6143 assert(zero_blocks.target() != nullptr, "zero_blocks stub has not been generated");
6144
6145 subs(rscratch1, cnt, zero_words_block_size);
6146 Label around;
6147 br(LO, around);
6148 {
6149 RuntimeAddress zero_blocks = RuntimeAddress(StubRoutines::aarch64::zero_blocks());
6150 assert(zero_blocks.target() != nullptr, "zero_blocks stub has not been generated");
6151 // Make sure this is a C2 compilation. C1 allocates space only for
6152 // trampoline stubs generated by Call LIR ops, and in any case it
6153 // makes sense for a C1 compilation task to proceed as quickly as
6154 // possible.
6155 CompileTask* task;
6156 if (StubRoutines::aarch64::complete()
6157 && Thread::current()->is_Compiler_thread()
6158 && (task = ciEnv::current()->task())
6159 && is_c2_compile(task->comp_level())) {
6160 address tpc = trampoline_call(zero_blocks);
6161 if (tpc == nullptr) {
6162 DEBUG_ONLY(reset_labels(around));
6163 return nullptr;
6164 }
6165 } else {
6166 far_call(zero_blocks);
6167 }
6168 }
6169 bind(around);
6170
6171 // We have a few words left to do. zero_blocks has adjusted r10 and r11
6172 // for us.
6173 for (int i = zero_words_block_size >> 1; i > 1; i >>= 1) {
6174 Label l;
6175 tbz(cnt, exact_log2(i), l);
6176 for (int j = 0; j < i; j += 2) {
6177 stp(zr, zr, post(ptr, 2 * BytesPerWord));
6178 }
6179 bind(l);
6180 }
6181 {
6182 Label l;
6183 tbz(cnt, 0, l);
6184 str(zr, Address(ptr));
6185 bind(l);
6186 }
6187
6188 BLOCK_COMMENT("} zero_words");
6189 return pc();
6190 }
6191
6192 // base: Address of a buffer to be zeroed, 8 bytes aligned.
6193 // cnt: Immediate count in HeapWords.
6194 //
6195 // r10, r11, rscratch1, and rscratch2 are clobbered.
6196 address MacroAssembler::zero_words(Register base, uint64_t cnt)
6197 {
6198 assert(wordSize <= BlockZeroingLowLimit,
6199 "increase BlockZeroingLowLimit");
6200 address result = nullptr;
6201 if (cnt <= (uint64_t)BlockZeroingLowLimit / BytesPerWord) {
6202 #ifndef PRODUCT
6203 {
6204 char buf[64];
6205 snprintf(buf, sizeof buf, "zero_words (count = %" PRIu64 ") {", cnt);
6206 BLOCK_COMMENT(buf);
6207 }
6208 #endif
6209 if (cnt >= 16) {
6210 uint64_t loops = cnt/16;
6211 if (loops > 1) {
6212 mov(rscratch2, loops - 1);
6213 }
6214 {
6215 Label loop;
6216 bind(loop);
6217 for (int i = 0; i < 16; i += 2) {
6218 stp(zr, zr, Address(base, i * BytesPerWord));
6219 }
6220 add(base, base, 16 * BytesPerWord);
6221 if (loops > 1) {
6222 subs(rscratch2, rscratch2, 1);
6223 br(GE, loop);
6224 }
6225 }
6226 }
6227 cnt %= 16;
6228 int i = cnt & 1; // store any odd word to start
6229 if (i) str(zr, Address(base));
6230 for (; i < (int)cnt; i += 2) {
6231 stp(zr, zr, Address(base, i * wordSize));
6232 }
6233 BLOCK_COMMENT("} zero_words");
6234 result = pc();
6235 } else {
6236 mov(r10, base); mov(r11, cnt);
6237 result = zero_words(r10, r11);
6238 }
6239 return result;
6240 }
6241
6242 // Zero blocks of memory by using DC ZVA.
6243 //
6244 // Aligns the base address first sufficiently for DC ZVA, then uses
6245 // DC ZVA repeatedly for every full block. cnt is the size to be
6246 // zeroed in HeapWords. Returns the count of words left to be zeroed
6247 // in cnt.
6248 //
6249 // NOTE: This is intended to be used in the zero_blocks() stub. If
6250 // you want to use it elsewhere, note that cnt must be >= 2*zva_length.
6251 void MacroAssembler::zero_dcache_blocks(Register base, Register cnt) {
6252 Register tmp = rscratch1;
6253 Register tmp2 = rscratch2;
6254 int zva_length = VM_Version::zva_length();
6255 Label initial_table_end, loop_zva;
6256 Label fini;
6257
6258 // Base must be 16 byte aligned. If not just return and let caller handle it
6259 tst(base, 0x0f);
6260 br(Assembler::NE, fini);
6261 // Align base with ZVA length.
6262 neg(tmp, base);
6263 andr(tmp, tmp, zva_length - 1);
6264
6265 // tmp: the number of bytes to be filled to align the base with ZVA length.
6266 add(base, base, tmp);
6267 sub(cnt, cnt, tmp, Assembler::ASR, 3);
6268 adr(tmp2, initial_table_end);
6269 sub(tmp2, tmp2, tmp, Assembler::LSR, 2);
6270 br(tmp2);
6271
6272 for (int i = -zva_length + 16; i < 0; i += 16)
6273 stp(zr, zr, Address(base, i));
6274 bind(initial_table_end);
6275
6276 sub(cnt, cnt, zva_length >> 3);
6277 bind(loop_zva);
6278 dc(Assembler::ZVA, base);
6279 subs(cnt, cnt, zva_length >> 3);
6280 add(base, base, zva_length);
6281 br(Assembler::GE, loop_zva);
6282 add(cnt, cnt, zva_length >> 3); // count not zeroed by DC ZVA
6283 bind(fini);
6284 }
6285
6286 // base: Address of a buffer to be filled, 8 bytes aligned.
6287 // cnt: Count in 8-byte unit.
6288 // value: Value to be filled with.
6289 // base will point to the end of the buffer after filling.
6290 void MacroAssembler::fill_words(Register base, Register cnt, Register value)
6291 {
6292 // Algorithm:
6293 //
6294 // if (cnt == 0) {
6295 // return;
6296 // }
6297 // if ((p & 8) != 0) {
6298 // *p++ = v;
6299 // }
6300 //
6301 // scratch1 = cnt & 14;
6302 // cnt -= scratch1;
6303 // p += scratch1;
6304 // switch (scratch1 / 2) {
6305 // do {
6306 // cnt -= 16;
6307 // p[-16] = v;
6308 // p[-15] = v;
6309 // case 7:
6310 // p[-14] = v;
6311 // p[-13] = v;
6312 // case 6:
6313 // p[-12] = v;
6314 // p[-11] = v;
6315 // // ...
6316 // case 1:
6317 // p[-2] = v;
6318 // p[-1] = v;
6319 // case 0:
6320 // p += 16;
6321 // } while (cnt);
6322 // }
6323 // if ((cnt & 1) == 1) {
6324 // *p++ = v;
6325 // }
6326
6327 assert_different_registers(base, cnt, value, rscratch1, rscratch2);
6328
6329 Label fini, skip, entry, loop;
6330 const int unroll = 8; // Number of stp instructions we'll unroll
6331
6332 cbz(cnt, fini);
6333 tbz(base, 3, skip);
6334 str(value, Address(post(base, 8)));
6335 sub(cnt, cnt, 1);
6336 bind(skip);
6337
6338 andr(rscratch1, cnt, (unroll-1) * 2);
6339 sub(cnt, cnt, rscratch1);
6340 add(base, base, rscratch1, Assembler::LSL, 3);
6341 adr(rscratch2, entry);
6342 sub(rscratch2, rscratch2, rscratch1, Assembler::LSL, 1);
6343 br(rscratch2);
6344
6345 bind(loop);
6346 add(base, base, unroll * 16);
6347 for (int i = -unroll; i < 0; i++)
6348 stp(value, value, Address(base, i * 16));
6349 bind(entry);
6350 subs(cnt, cnt, unroll * 2);
6351 br(Assembler::GE, loop);
6352
6353 tbz(cnt, 0, fini);
6354 str(value, Address(post(base, 8)));
6355 bind(fini);
6356 }
6357
6358 // Intrinsic for
6359 //
6360 // - sun/nio/cs/ISO_8859_1$Encoder.implEncodeISOArray
6361 // return the number of characters copied.
6362 // - java/lang/StringUTF16.compress
6363 // return index of non-latin1 character if copy fails, otherwise 'len'.
6364 //
6365 // This version always returns the number of characters copied, and does not
6366 // clobber the 'len' register. A successful copy will complete with the post-
6367 // condition: 'res' == 'len', while an unsuccessful copy will exit with the
6368 // post-condition: 0 <= 'res' < 'len'.
6369 //
6370 // NOTE: Attempts to use 'ld2' (and 'umaxv' in the ISO part) has proven to
6371 // degrade performance (on Ampere Altra - Neoverse N1), to an extent
6372 // beyond the acceptable, even though the footprint would be smaller.
6373 // Using 'umaxv' in the ASCII-case comes with a small penalty but does
6374 // avoid additional bloat.
6375 //
6376 // Clobbers: src, dst, res, rscratch1, rscratch2, rflags
6377 void MacroAssembler::encode_iso_array(Register src, Register dst,
6378 Register len, Register res, bool ascii,
6379 FloatRegister vtmp0, FloatRegister vtmp1,
6380 FloatRegister vtmp2, FloatRegister vtmp3,
6381 FloatRegister vtmp4, FloatRegister vtmp5)
6382 {
6383 Register cnt = res;
6384 Register max = rscratch1;
6385 Register chk = rscratch2;
6386
6387 prfm(Address(src), PLDL1STRM);
6388 movw(cnt, len);
6389
6390 #define ASCII(insn) do { if (ascii) { insn; } } while (0)
6391
6392 Label LOOP_32, DONE_32, FAIL_32;
6393
6394 BIND(LOOP_32);
6395 {
6396 cmpw(cnt, 32);
6397 br(LT, DONE_32);
6398 ld1(vtmp0, vtmp1, vtmp2, vtmp3, T8H, Address(post(src, 64)));
6399 // Extract lower bytes.
6400 FloatRegister vlo0 = vtmp4;
6401 FloatRegister vlo1 = vtmp5;
6402 uzp1(vlo0, T16B, vtmp0, vtmp1);
6403 uzp1(vlo1, T16B, vtmp2, vtmp3);
6404 // Merge bits...
6405 orr(vtmp0, T16B, vtmp0, vtmp1);
6406 orr(vtmp2, T16B, vtmp2, vtmp3);
6407 // Extract merged upper bytes.
6408 FloatRegister vhix = vtmp0;
6409 uzp2(vhix, T16B, vtmp0, vtmp2);
6410 // ISO-check on hi-parts (all zero).
6411 // ASCII-check on lo-parts (no sign).
6412 FloatRegister vlox = vtmp1; // Merge lower bytes.
6413 ASCII(orr(vlox, T16B, vlo0, vlo1));
6414 umov(chk, vhix, D, 1); ASCII(cm(LT, vlox, T16B, vlox));
6415 fmovd(max, vhix); ASCII(umaxv(vlox, T16B, vlox));
6416 orr(chk, chk, max); ASCII(umov(max, vlox, B, 0));
6417 ASCII(orr(chk, chk, max));
6418 cbnz(chk, FAIL_32);
6419 subw(cnt, cnt, 32);
6420 st1(vlo0, vlo1, T16B, Address(post(dst, 32)));
6421 b(LOOP_32);
6422 }
6423 BIND(FAIL_32);
6424 sub(src, src, 64);
6425 BIND(DONE_32);
6426
6427 Label LOOP_8, SKIP_8;
6428
6429 BIND(LOOP_8);
6430 {
6431 cmpw(cnt, 8);
6432 br(LT, SKIP_8);
6433 FloatRegister vhi = vtmp0;
6434 FloatRegister vlo = vtmp1;
6435 ld1(vtmp3, T8H, src);
6436 uzp1(vlo, T16B, vtmp3, vtmp3);
6437 uzp2(vhi, T16B, vtmp3, vtmp3);
6438 // ISO-check on hi-parts (all zero).
6439 // ASCII-check on lo-parts (no sign).
6440 ASCII(cm(LT, vtmp2, T16B, vlo));
6441 fmovd(chk, vhi); ASCII(umaxv(vtmp2, T16B, vtmp2));
6442 ASCII(umov(max, vtmp2, B, 0));
6443 ASCII(orr(chk, chk, max));
6444 cbnz(chk, SKIP_8);
6445
6446 strd(vlo, Address(post(dst, 8)));
6447 subw(cnt, cnt, 8);
6448 add(src, src, 16);
6449 b(LOOP_8);
6450 }
6451 BIND(SKIP_8);
6452
6453 #undef ASCII
6454
6455 Label LOOP, DONE;
6456
6457 cbz(cnt, DONE);
6458 BIND(LOOP);
6459 {
6460 Register chr = rscratch1;
6461 ldrh(chr, Address(post(src, 2)));
6462 tst(chr, ascii ? 0xff80 : 0xff00);
6463 br(NE, DONE);
6464 strb(chr, Address(post(dst, 1)));
6465 subs(cnt, cnt, 1);
6466 br(GT, LOOP);
6467 }
6468 BIND(DONE);
6469 // Return index where we stopped.
6470 subw(res, len, cnt);
6471 }
6472
6473 // Inflate byte[] array to char[].
6474 // Clobbers: src, dst, len, rflags, rscratch1, v0-v6
6475 address MacroAssembler::byte_array_inflate(Register src, Register dst, Register len,
6476 FloatRegister vtmp1, FloatRegister vtmp2,
6477 FloatRegister vtmp3, Register tmp4) {
6478 Label big, done, after_init, to_stub;
6479
6480 assert_different_registers(src, dst, len, tmp4, rscratch1);
6481
6482 fmovd(vtmp1, 0.0);
6483 lsrw(tmp4, len, 3);
6484 bind(after_init);
6485 cbnzw(tmp4, big);
6486 // Short string: less than 8 bytes.
6487 {
6488 Label loop, tiny;
6489
6490 cmpw(len, 4);
6491 br(LT, tiny);
6492 // Use SIMD to do 4 bytes.
6493 ldrs(vtmp2, post(src, 4));
6494 zip1(vtmp3, T8B, vtmp2, vtmp1);
6495 subw(len, len, 4);
6496 strd(vtmp3, post(dst, 8));
6497
6498 cbzw(len, done);
6499
6500 // Do the remaining bytes by steam.
6501 bind(loop);
6502 ldrb(tmp4, post(src, 1));
6503 strh(tmp4, post(dst, 2));
6504 subw(len, len, 1);
6505
6506 bind(tiny);
6507 cbnz(len, loop);
6508
6509 b(done);
6510 }
6511
6512 if (SoftwarePrefetchHintDistance >= 0) {
6513 bind(to_stub);
6514 RuntimeAddress stub = RuntimeAddress(StubRoutines::aarch64::large_byte_array_inflate());
6515 assert(stub.target() != nullptr, "large_byte_array_inflate stub has not been generated");
6516 address tpc = trampoline_call(stub);
6517 if (tpc == nullptr) {
6518 DEBUG_ONLY(reset_labels(big, done));
6519 postcond(pc() == badAddress);
6520 return nullptr;
6521 }
6522 b(after_init);
6523 }
6524
6525 // Unpack the bytes 8 at a time.
6526 bind(big);
6527 {
6528 Label loop, around, loop_last, loop_start;
6529
6530 if (SoftwarePrefetchHintDistance >= 0) {
6531 const int large_loop_threshold = (64 + 16)/8;
6532 ldrd(vtmp2, post(src, 8));
6533 andw(len, len, 7);
6534 cmp(tmp4, (u1)large_loop_threshold);
6535 br(GE, to_stub);
6536 b(loop_start);
6537
6538 bind(loop);
6539 ldrd(vtmp2, post(src, 8));
6540 bind(loop_start);
6541 subs(tmp4, tmp4, 1);
6542 br(EQ, loop_last);
6543 zip1(vtmp2, T16B, vtmp2, vtmp1);
6544 ldrd(vtmp3, post(src, 8));
6545 st1(vtmp2, T8H, post(dst, 16));
6546 subs(tmp4, tmp4, 1);
6547 zip1(vtmp3, T16B, vtmp3, vtmp1);
6548 st1(vtmp3, T8H, post(dst, 16));
6549 br(NE, loop);
6550 b(around);
6551 bind(loop_last);
6552 zip1(vtmp2, T16B, vtmp2, vtmp1);
6553 st1(vtmp2, T8H, post(dst, 16));
6554 bind(around);
6555 cbz(len, done);
6556 } else {
6557 andw(len, len, 7);
6558 bind(loop);
6559 ldrd(vtmp2, post(src, 8));
6560 sub(tmp4, tmp4, 1);
6561 zip1(vtmp3, T16B, vtmp2, vtmp1);
6562 st1(vtmp3, T8H, post(dst, 16));
6563 cbnz(tmp4, loop);
6564 }
6565 }
6566
6567 // Do the tail of up to 8 bytes.
6568 add(src, src, len);
6569 ldrd(vtmp3, Address(src, -8));
6570 add(dst, dst, len, ext::uxtw, 1);
6571 zip1(vtmp3, T16B, vtmp3, vtmp1);
6572 strq(vtmp3, Address(dst, -16));
6573
6574 bind(done);
6575 postcond(pc() != badAddress);
6576 return pc();
6577 }
6578
6579 // Compress char[] array to byte[].
6580 // Intrinsic for java.lang.StringUTF16.compress(char[] src, int srcOff, byte[] dst, int dstOff, int len)
6581 // Return the array length if every element in array can be encoded,
6582 // otherwise, the index of first non-latin1 (> 0xff) character.
6583 void MacroAssembler::char_array_compress(Register src, Register dst, Register len,
6584 Register res,
6585 FloatRegister tmp0, FloatRegister tmp1,
6586 FloatRegister tmp2, FloatRegister tmp3,
6587 FloatRegister tmp4, FloatRegister tmp5) {
6588 encode_iso_array(src, dst, len, res, false, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5);
6589 }
6590
6591 // java.math.round(double a)
6592 // Returns the closest long to the argument, with ties rounding to
6593 // positive infinity. This requires some fiddling for corner
6594 // cases. We take care to avoid double rounding in e.g. (jlong)(a + 0.5).
6595 void MacroAssembler::java_round_double(Register dst, FloatRegister src,
6596 FloatRegister ftmp) {
6597 Label DONE;
6598 BLOCK_COMMENT("java_round_double: { ");
6599 fmovd(rscratch1, src);
6600 // Use RoundToNearestTiesAway unless src small and -ve.
6601 fcvtasd(dst, src);
6602 // Test if src >= 0 || abs(src) >= 0x1.0p52
6603 eor(rscratch1, rscratch1, UCONST64(1) << 63); // flip sign bit
6604 mov(rscratch2, julong_cast(0x1.0p52));
6605 cmp(rscratch1, rscratch2);
6606 br(HS, DONE); {
6607 // src < 0 && abs(src) < 0x1.0p52
6608 // src may have a fractional part, so add 0.5
6609 fmovd(ftmp, 0.5);
6610 faddd(ftmp, src, ftmp);
6611 // Convert double to jlong, use RoundTowardsNegative
6612 fcvtmsd(dst, ftmp);
6613 }
6614 bind(DONE);
6615 BLOCK_COMMENT("} java_round_double");
6616 }
6617
6618 void MacroAssembler::java_round_float(Register dst, FloatRegister src,
6619 FloatRegister ftmp) {
6620 Label DONE;
6621 BLOCK_COMMENT("java_round_float: { ");
6622 fmovs(rscratch1, src);
6623 // Use RoundToNearestTiesAway unless src small and -ve.
6624 fcvtassw(dst, src);
6625 // Test if src >= 0 || abs(src) >= 0x1.0p23
6626 eor(rscratch1, rscratch1, 0x80000000); // flip sign bit
6627 mov(rscratch2, jint_cast(0x1.0p23f));
6628 cmp(rscratch1, rscratch2);
6629 br(HS, DONE); {
6630 // src < 0 && |src| < 0x1.0p23
6631 // src may have a fractional part, so add 0.5
6632 fmovs(ftmp, 0.5f);
6633 fadds(ftmp, src, ftmp);
6634 // Convert float to jint, use RoundTowardsNegative
6635 fcvtmssw(dst, ftmp);
6636 }
6637 bind(DONE);
6638 BLOCK_COMMENT("} java_round_float");
6639 }
6640
6641 // get_thread() can be called anywhere inside generated code so we
6642 // need to save whatever non-callee save context might get clobbered
6643 // by the call to JavaThread::aarch64_get_thread_helper() or, indeed,
6644 // the call setup code.
6645 //
6646 // On Linux, aarch64_get_thread_helper() clobbers only r0, r1, and flags.
6647 // On other systems, the helper is a usual C function.
6648 //
6649 void MacroAssembler::get_thread(Register dst) {
6650 RegSet saved_regs =
6651 LINUX_ONLY(RegSet::range(r0, r1) + lr - dst)
6652 NOT_LINUX (RegSet::range(r0, r17) + lr - dst);
6653
6654 protect_return_address();
6655 push(saved_regs, sp);
6656
6657 mov(lr, CAST_FROM_FN_PTR(address, JavaThread::aarch64_get_thread_helper));
6658 blr(lr);
6659 if (dst != c_rarg0) {
6660 mov(dst, c_rarg0);
6661 }
6662
6663 pop(saved_regs, sp);
6664 authenticate_return_address();
6665 }
6666
6667 void MacroAssembler::cache_wb(Address line) {
6668 assert(line.getMode() == Address::base_plus_offset, "mode should be base_plus_offset");
6669 assert(line.index() == noreg, "index should be noreg");
6670 assert(line.offset() == 0, "offset should be 0");
6671 // would like to assert this
6672 // assert(line._ext.shift == 0, "shift should be zero");
6673 if (VM_Version::supports_dcpop()) {
6674 // writeback using clear virtual address to point of persistence
6675 dc(Assembler::CVAP, line.base());
6676 } else {
6677 // no need to generate anything as Unsafe.writebackMemory should
6678 // never invoke this stub
6679 }
6680 }
6681
6682 void MacroAssembler::cache_wbsync(bool is_pre) {
6683 // we only need a barrier post sync
6684 if (!is_pre) {
6685 membar(Assembler::AnyAny);
6686 }
6687 }
6688
6689 void MacroAssembler::verify_sve_vector_length(Register tmp) {
6690 if (!UseSVE || VM_Version::get_max_supported_sve_vector_length() == FloatRegister::sve_vl_min) {
6691 return;
6692 }
6693 // Make sure that native code does not change SVE vector length.
6694 Label verify_ok;
6695 movw(tmp, zr);
6696 sve_inc(tmp, B);
6697 subsw(zr, tmp, VM_Version::get_initial_sve_vector_length());
6698 br(EQ, verify_ok);
6699 stop("Error: SVE vector length has changed since jvm startup");
6700 bind(verify_ok);
6701 }
6702
6703 void MacroAssembler::verify_ptrue() {
6704 Label verify_ok;
6705 if (!UseSVE) {
6706 return;
6707 }
6708 sve_cntp(rscratch1, B, ptrue, ptrue); // get true elements count.
6709 sve_dec(rscratch1, B);
6710 cbz(rscratch1, verify_ok);
6711 stop("Error: the preserved predicate register (p7) elements are not all true");
6712 bind(verify_ok);
6713 }
6714
6715 void MacroAssembler::safepoint_isb() {
6716 isb();
6717 #ifndef PRODUCT
6718 if (VerifyCrossModifyFence) {
6719 // Clear the thread state.
6720 strb(zr, Address(rthread, in_bytes(JavaThread::requires_cross_modify_fence_offset())));
6721 }
6722 #endif
6723 }
6724
6725 #ifndef PRODUCT
6726 void MacroAssembler::verify_cross_modify_fence_not_required() {
6727 if (VerifyCrossModifyFence) {
6728 // Check if thread needs a cross modify fence.
6729 ldrb(rscratch1, Address(rthread, in_bytes(JavaThread::requires_cross_modify_fence_offset())));
6730 Label fence_not_required;
6731 cbz(rscratch1, fence_not_required);
6732 // If it does then fail.
6733 lea(rscratch1, RuntimeAddress(CAST_FROM_FN_PTR(address, JavaThread::verify_cross_modify_fence_failure)));
6734 mov(c_rarg0, rthread);
6735 blr(rscratch1);
6736 bind(fence_not_required);
6737 }
6738 }
6739 #endif
6740
6741 void MacroAssembler::spin_wait() {
6742 for (int i = 0; i < VM_Version::spin_wait_desc().inst_count(); ++i) {
6743 switch (VM_Version::spin_wait_desc().inst()) {
6744 case SpinWait::NOP:
6745 nop();
6746 break;
6747 case SpinWait::ISB:
6748 isb();
6749 break;
6750 case SpinWait::YIELD:
6751 yield();
6752 break;
6753 default:
6754 ShouldNotReachHere();
6755 }
6756 }
6757 }
6758
6759 // Stack frame creation/removal
6760
6761 void MacroAssembler::enter(bool strip_ret_addr) {
6762 if (strip_ret_addr) {
6763 // Addresses can only be signed once. If there are multiple nested frames being created
6764 // in the same function, then the return address needs stripping first.
6765 strip_return_address();
6766 }
6767 protect_return_address();
6768 stp(rfp, lr, Address(pre(sp, -2 * wordSize)));
6769 mov(rfp, sp);
6770 }
6771
6772 void MacroAssembler::leave() {
6773 mov(sp, rfp);
6774 ldp(rfp, lr, Address(post(sp, 2 * wordSize)));
6775 authenticate_return_address();
6776 }
6777
6778 // ROP Protection
6779 // Use the AArch64 PAC feature to add ROP protection for generated code. Use whenever creating/
6780 // destroying stack frames or whenever directly loading/storing the LR to memory.
6781 // If ROP protection is not set then these functions are no-ops.
6782 // For more details on PAC see pauth_aarch64.hpp.
6783
6784 // Sign the LR. Use during construction of a stack frame, before storing the LR to memory.
6785 // Uses value zero as the modifier.
6786 //
6787 void MacroAssembler::protect_return_address() {
6788 if (VM_Version::use_rop_protection()) {
6789 check_return_address();
6790 paciaz();
6791 }
6792 }
6793
6794 // Sign the return value in the given register. Use before updating the LR in the existing stack
6795 // frame for the current function.
6796 // Uses value zero as the modifier.
6797 //
6798 void MacroAssembler::protect_return_address(Register return_reg) {
6799 if (VM_Version::use_rop_protection()) {
6800 check_return_address(return_reg);
6801 paciza(return_reg);
6802 }
6803 }
6804
6805 // Authenticate the LR. Use before function return, after restoring FP and loading LR from memory.
6806 // Uses value zero as the modifier.
6807 //
6808 void MacroAssembler::authenticate_return_address() {
6809 if (VM_Version::use_rop_protection()) {
6810 autiaz();
6811 check_return_address();
6812 }
6813 }
6814
6815 // Authenticate the return value in the given register. Use before updating the LR in the existing
6816 // stack frame for the current function.
6817 // Uses value zero as the modifier.
6818 //
6819 void MacroAssembler::authenticate_return_address(Register return_reg) {
6820 if (VM_Version::use_rop_protection()) {
6821 autiza(return_reg);
6822 check_return_address(return_reg);
6823 }
6824 }
6825
6826 // Strip any PAC data from LR without performing any authentication. Use with caution - only if
6827 // there is no guaranteed way of authenticating the LR.
6828 //
6829 void MacroAssembler::strip_return_address() {
6830 if (VM_Version::use_rop_protection()) {
6831 xpaclri();
6832 }
6833 }
6834
6835 #ifndef PRODUCT
6836 // PAC failures can be difficult to debug. After an authentication failure, a segfault will only
6837 // occur when the pointer is used - ie when the program returns to the invalid LR. At this point
6838 // it is difficult to debug back to the callee function.
6839 // This function simply loads from the address in the given register.
6840 // Use directly after authentication to catch authentication failures.
6841 // Also use before signing to check that the pointer is valid and hasn't already been signed.
6842 //
6843 void MacroAssembler::check_return_address(Register return_reg) {
6844 if (VM_Version::use_rop_protection()) {
6845 ldr(zr, Address(return_reg));
6846 }
6847 }
6848 #endif
6849
6850 // The java_calling_convention describes stack locations as ideal slots on
6851 // a frame with no abi restrictions. Since we must observe abi restrictions
6852 // (like the placement of the register window) the slots must be biased by
6853 // the following value.
6854 static int reg2offset_in(VMReg r) {
6855 // Account for saved rfp and lr
6856 // This should really be in_preserve_stack_slots
6857 return (r->reg2stack() + 4) * VMRegImpl::stack_slot_size;
6858 }
6859
6860 static int reg2offset_out(VMReg r) {
6861 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
6862 }
6863
6864 // On 64bit we will store integer like items to the stack as
6865 // 64bits items (AArch64 ABI) even though java would only store
6866 // 32bits for a parameter. On 32bit it will simply be 32bits
6867 // So this routine will do 32->32 on 32bit and 32->64 on 64bit
6868 void MacroAssembler::move32_64(VMRegPair src, VMRegPair dst, Register tmp) {
6869 if (src.first()->is_stack()) {
6870 if (dst.first()->is_stack()) {
6871 // stack to stack
6872 ldr(tmp, Address(rfp, reg2offset_in(src.first())));
6873 str(tmp, Address(sp, reg2offset_out(dst.first())));
6874 } else {
6875 // stack to reg
6876 ldrsw(dst.first()->as_Register(), Address(rfp, reg2offset_in(src.first())));
6877 }
6878 } else if (dst.first()->is_stack()) {
6879 // reg to stack
6880 str(src.first()->as_Register(), Address(sp, reg2offset_out(dst.first())));
6881 } else {
6882 if (dst.first() != src.first()) {
6883 sxtw(dst.first()->as_Register(), src.first()->as_Register());
6884 }
6885 }
6886 }
6887
6888 // An oop arg. Must pass a handle not the oop itself
6889 void MacroAssembler::object_move(
6890 OopMap* map,
6891 int oop_handle_offset,
6892 int framesize_in_slots,
6893 VMRegPair src,
6894 VMRegPair dst,
6895 bool is_receiver,
6896 int* receiver_offset) {
6897
6898 // must pass a handle. First figure out the location we use as a handle
6899
6900 Register rHandle = dst.first()->is_stack() ? rscratch2 : dst.first()->as_Register();
6901
6902 // See if oop is null if it is we need no handle
6903
6904 if (src.first()->is_stack()) {
6905
6906 // Oop is already on the stack as an argument
6907 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots();
6908 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots));
6909 if (is_receiver) {
6910 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size;
6911 }
6912
6913 ldr(rscratch1, Address(rfp, reg2offset_in(src.first())));
6914 lea(rHandle, Address(rfp, reg2offset_in(src.first())));
6915 // conditionally move a null
6916 cmp(rscratch1, zr);
6917 csel(rHandle, zr, rHandle, Assembler::EQ);
6918 } else {
6919
6920 // Oop is in an a register we must store it to the space we reserve
6921 // on the stack for oop_handles and pass a handle if oop is non-null
6922
6923 const Register rOop = src.first()->as_Register();
6924 int oop_slot;
6925 if (rOop == j_rarg0)
6926 oop_slot = 0;
6927 else if (rOop == j_rarg1)
6928 oop_slot = 1;
6929 else if (rOop == j_rarg2)
6930 oop_slot = 2;
6931 else if (rOop == j_rarg3)
6932 oop_slot = 3;
6933 else if (rOop == j_rarg4)
6934 oop_slot = 4;
6935 else if (rOop == j_rarg5)
6936 oop_slot = 5;
6937 else if (rOop == j_rarg6)
6938 oop_slot = 6;
6939 else {
6940 assert(rOop == j_rarg7, "wrong register");
6941 oop_slot = 7;
6942 }
6943
6944 oop_slot = oop_slot * VMRegImpl::slots_per_word + oop_handle_offset;
6945 int offset = oop_slot*VMRegImpl::stack_slot_size;
6946
6947 map->set_oop(VMRegImpl::stack2reg(oop_slot));
6948 // Store oop in handle area, may be null
6949 str(rOop, Address(sp, offset));
6950 if (is_receiver) {
6951 *receiver_offset = offset;
6952 }
6953
6954 cmp(rOop, zr);
6955 lea(rHandle, Address(sp, offset));
6956 // conditionally move a null
6957 csel(rHandle, zr, rHandle, Assembler::EQ);
6958 }
6959
6960 // If arg is on the stack then place it otherwise it is already in correct reg.
6961 if (dst.first()->is_stack()) {
6962 str(rHandle, Address(sp, reg2offset_out(dst.first())));
6963 }
6964 }
6965
6966 // A float arg may have to do float reg int reg conversion
6967 void MacroAssembler::float_move(VMRegPair src, VMRegPair dst, Register tmp) {
6968 if (src.first()->is_stack()) {
6969 if (dst.first()->is_stack()) {
6970 ldrw(tmp, Address(rfp, reg2offset_in(src.first())));
6971 strw(tmp, Address(sp, reg2offset_out(dst.first())));
6972 } else {
6973 ldrs(dst.first()->as_FloatRegister(), Address(rfp, reg2offset_in(src.first())));
6974 }
6975 } else if (src.first() != dst.first()) {
6976 if (src.is_single_phys_reg() && dst.is_single_phys_reg())
6977 fmovs(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister());
6978 else
6979 strs(src.first()->as_FloatRegister(), Address(sp, reg2offset_out(dst.first())));
6980 }
6981 }
6982
6983 // A long move
6984 void MacroAssembler::long_move(VMRegPair src, VMRegPair dst, Register tmp) {
6985 if (src.first()->is_stack()) {
6986 if (dst.first()->is_stack()) {
6987 // stack to stack
6988 ldr(tmp, Address(rfp, reg2offset_in(src.first())));
6989 str(tmp, Address(sp, reg2offset_out(dst.first())));
6990 } else {
6991 // stack to reg
6992 ldr(dst.first()->as_Register(), Address(rfp, reg2offset_in(src.first())));
6993 }
6994 } else if (dst.first()->is_stack()) {
6995 // reg to stack
6996 // Do we really have to sign extend???
6997 // __ movslq(src.first()->as_Register(), src.first()->as_Register());
6998 str(src.first()->as_Register(), Address(sp, reg2offset_out(dst.first())));
6999 } else {
7000 if (dst.first() != src.first()) {
7001 mov(dst.first()->as_Register(), src.first()->as_Register());
7002 }
7003 }
7004 }
7005
7006
7007 // A double move
7008 void MacroAssembler::double_move(VMRegPair src, VMRegPair dst, Register tmp) {
7009 if (src.first()->is_stack()) {
7010 if (dst.first()->is_stack()) {
7011 ldr(tmp, Address(rfp, reg2offset_in(src.first())));
7012 str(tmp, Address(sp, reg2offset_out(dst.first())));
7013 } else {
7014 ldrd(dst.first()->as_FloatRegister(), Address(rfp, reg2offset_in(src.first())));
7015 }
7016 } else if (src.first() != dst.first()) {
7017 if (src.is_single_phys_reg() && dst.is_single_phys_reg())
7018 fmovd(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister());
7019 else
7020 strd(src.first()->as_FloatRegister(), Address(sp, reg2offset_out(dst.first())));
7021 }
7022 }
7023
7024 // Implements lightweight-locking.
7025 //
7026 // - obj: the object to be locked
7027 // - t1, t2, t3: temporary registers, will be destroyed
7028 // - slow: branched to if locking fails, absolute offset may larger than 32KB (imm14 encoding).
7029 void MacroAssembler::lightweight_lock(Register basic_lock, Register obj, Register t1, Register t2, Register t3, Label& slow) {
7030 assert(LockingMode == LM_LIGHTWEIGHT, "only used with new lightweight locking");
7031 assert_different_registers(basic_lock, obj, t1, t2, t3, rscratch1);
7032
7033 Label push;
7034 const Register top = t1;
7035 const Register mark = t2;
7036 const Register t = t3;
7037
7038 // Preload the markWord. It is important that this is the first
7039 // instruction emitted as it is part of C1's null check semantics.
7040 ldr(mark, Address(obj, oopDesc::mark_offset_in_bytes()));
7041
7042 if (UseObjectMonitorTable) {
7043 // Clear cache in case fast locking succeeds or we need to take the slow-path.
7044 str(zr, Address(basic_lock, BasicObjectLock::lock_offset() + in_ByteSize((BasicLock::object_monitor_cache_offset_in_bytes()))));
7045 }
7046
7047 if (DiagnoseSyncOnValueBasedClasses != 0) {
7048 load_klass(t1, obj);
7049 ldrb(t1, Address(t1, Klass::misc_flags_offset()));
7050 tst(t1, KlassFlags::_misc_is_value_based_class);
7051 br(Assembler::NE, slow);
7052 }
7053
7054 // Check if the lock-stack is full.
7055 ldrw(top, Address(rthread, JavaThread::lock_stack_top_offset()));
7056 cmpw(top, (unsigned)LockStack::end_offset());
7057 br(Assembler::GE, slow);
7058
7059 // Check for recursion.
7060 subw(t, top, oopSize);
7061 ldr(t, Address(rthread, t));
7062 cmp(obj, t);
7063 br(Assembler::EQ, push);
7064
7065 // Check header for monitor (0b10).
7066 tst(mark, markWord::monitor_value);
7067 br(Assembler::NE, slow);
7068
7069 // Try to lock. Transition lock bits 0b01 => 0b00
7070 assert(oopDesc::mark_offset_in_bytes() == 0, "required to avoid lea");
7071 orr(mark, mark, markWord::unlocked_value);
7072 eor(t, mark, markWord::unlocked_value);
7073 cmpxchg(/*addr*/ obj, /*expected*/ mark, /*new*/ t, Assembler::xword,
7074 /*acquire*/ true, /*release*/ false, /*weak*/ false, noreg);
7075 br(Assembler::NE, slow);
7076
7077 bind(push);
7078 // After successful lock, push object on lock-stack.
7079 str(obj, Address(rthread, top));
7080 addw(top, top, oopSize);
7081 strw(top, Address(rthread, JavaThread::lock_stack_top_offset()));
7082 }
7083
7084 // Implements lightweight-unlocking.
7085 //
7086 // - obj: the object to be unlocked
7087 // - t1, t2, t3: temporary registers
7088 // - slow: branched to if unlocking fails, absolute offset may larger than 32KB (imm14 encoding).
7089 void MacroAssembler::lightweight_unlock(Register obj, Register t1, Register t2, Register t3, Label& slow) {
7090 assert(LockingMode == LM_LIGHTWEIGHT, "only used with new lightweight locking");
7091 // cmpxchg clobbers rscratch1.
7092 assert_different_registers(obj, t1, t2, t3, rscratch1);
7093
7094 #ifdef ASSERT
7095 {
7096 // Check for lock-stack underflow.
7097 Label stack_ok;
7098 ldrw(t1, Address(rthread, JavaThread::lock_stack_top_offset()));
7099 cmpw(t1, (unsigned)LockStack::start_offset());
7100 br(Assembler::GE, stack_ok);
7101 STOP("Lock-stack underflow");
7102 bind(stack_ok);
7103 }
7104 #endif
7105
7106 Label unlocked, push_and_slow;
7107 const Register top = t1;
7108 const Register mark = t2;
7109 const Register t = t3;
7110
7111 // Check if obj is top of lock-stack.
7112 ldrw(top, Address(rthread, JavaThread::lock_stack_top_offset()));
7113 subw(top, top, oopSize);
7114 ldr(t, Address(rthread, top));
7115 cmp(obj, t);
7116 br(Assembler::NE, slow);
7117
7118 // Pop lock-stack.
7119 DEBUG_ONLY(str(zr, Address(rthread, top));)
7120 strw(top, Address(rthread, JavaThread::lock_stack_top_offset()));
7121
7122 // Check if recursive.
7123 subw(t, top, oopSize);
7124 ldr(t, Address(rthread, t));
7125 cmp(obj, t);
7126 br(Assembler::EQ, unlocked);
7127
7128 // Not recursive. Check header for monitor (0b10).
7129 ldr(mark, Address(obj, oopDesc::mark_offset_in_bytes()));
7130 tbnz(mark, log2i_exact(markWord::monitor_value), push_and_slow);
7131
7132 #ifdef ASSERT
7133 // Check header not unlocked (0b01).
7134 Label not_unlocked;
7135 tbz(mark, log2i_exact(markWord::unlocked_value), not_unlocked);
7136 stop("lightweight_unlock already unlocked");
7137 bind(not_unlocked);
7138 #endif
7139
7140 // Try to unlock. Transition lock bits 0b00 => 0b01
7141 assert(oopDesc::mark_offset_in_bytes() == 0, "required to avoid lea");
7142 orr(t, mark, markWord::unlocked_value);
7143 cmpxchg(obj, mark, t, Assembler::xword,
7144 /*acquire*/ false, /*release*/ true, /*weak*/ false, noreg);
7145 br(Assembler::EQ, unlocked);
7146
7147 bind(push_and_slow);
7148 // Restore lock-stack and handle the unlock in runtime.
7149 DEBUG_ONLY(str(obj, Address(rthread, top));)
7150 addw(top, top, oopSize);
7151 strw(top, Address(rthread, JavaThread::lock_stack_top_offset()));
7152 b(slow);
7153
7154 bind(unlocked);
7155 }