1 /*
2 * Copyright (c) 2003, 2025, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2014, 2025, 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/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "asm/register.hpp"
29 #include "atomic_aarch64.hpp"
30 #include "compiler/oopMap.hpp"
31 #include "gc/shared/barrierSet.hpp"
32 #include "gc/shared/barrierSetAssembler.hpp"
33 #include "gc/shared/gc_globals.hpp"
34 #include "gc/shared/tlab_globals.hpp"
35 #include "interpreter/interpreter.hpp"
36 #include "memory/universe.hpp"
37 #include "nativeInst_aarch64.hpp"
38 #include "oops/instanceOop.hpp"
39 #include "oops/method.hpp"
40 #include "oops/objArrayKlass.hpp"
41 #include "oops/oop.inline.hpp"
42 #include "prims/methodHandles.hpp"
43 #include "prims/upcallLinker.hpp"
44 #include "runtime/arguments.hpp"
45 #include "runtime/atomic.hpp"
46 #include "runtime/continuation.hpp"
47 #include "runtime/continuationEntry.inline.hpp"
48 #include "runtime/frame.inline.hpp"
49 #include "runtime/handles.inline.hpp"
50 #include "runtime/javaThread.hpp"
51 #include "runtime/sharedRuntime.hpp"
52 #include "runtime/stubCodeGenerator.hpp"
53 #include "runtime/stubRoutines.hpp"
54 #include "utilities/align.hpp"
55 #include "utilities/checkedCast.hpp"
56 #include "utilities/debug.hpp"
57 #include "utilities/globalDefinitions.hpp"
58 #include "utilities/intpow.hpp"
59 #include "utilities/powerOfTwo.hpp"
60 #ifdef COMPILER2
61 #include "opto/runtime.hpp"
62 #endif
63 #if INCLUDE_ZGC
64 #include "gc/z/zThreadLocalData.hpp"
65 #endif
66
67 // Declaration and definition of StubGenerator (no .hpp file).
68 // For a more detailed description of the stub routine structure
69 // see the comment in stubRoutines.hpp
70
71 #undef __
72 #define __ _masm->
73
74 #ifdef PRODUCT
75 #define BLOCK_COMMENT(str) /* nothing */
76 #else
77 #define BLOCK_COMMENT(str) __ block_comment(str)
78 #endif
79
80 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
81
82 // Stub Code definitions
83
84 class StubGenerator: public StubCodeGenerator {
85 private:
86
87 #ifdef PRODUCT
88 #define inc_counter_np(counter) ((void)0)
89 #else
90 void inc_counter_np_(uint& counter) {
91 __ incrementw(ExternalAddress((address)&counter));
92 }
93 #define inc_counter_np(counter) \
94 BLOCK_COMMENT("inc_counter " #counter); \
95 inc_counter_np_(counter);
96 #endif
97
98 // Call stubs are used to call Java from C
99 //
100 // Arguments:
101 // c_rarg0: call wrapper address address
102 // c_rarg1: result address
103 // c_rarg2: result type BasicType
104 // c_rarg3: method Method*
105 // c_rarg4: (interpreter) entry point address
106 // c_rarg5: parameters intptr_t*
107 // c_rarg6: parameter size (in words) int
108 // c_rarg7: thread Thread*
109 //
110 // There is no return from the stub itself as any Java result
111 // is written to result
112 //
113 // we save r30 (lr) as the return PC at the base of the frame and
114 // link r29 (fp) below it as the frame pointer installing sp (r31)
115 // into fp.
116 //
117 // we save r0-r7, which accounts for all the c arguments.
118 //
119 // TODO: strictly do we need to save them all? they are treated as
120 // volatile by C so could we omit saving the ones we are going to
121 // place in global registers (thread? method?) or those we only use
122 // during setup of the Java call?
123 //
124 // we don't need to save r8 which C uses as an indirect result location
125 // return register.
126 //
127 // we don't need to save r9-r15 which both C and Java treat as
128 // volatile
129 //
130 // we don't need to save r16-18 because Java does not use them
131 //
132 // we save r19-r28 which Java uses as scratch registers and C
133 // expects to be callee-save
134 //
135 // we save the bottom 64 bits of each value stored in v8-v15; it is
136 // the responsibility of the caller to preserve larger values.
137 //
138 // so the stub frame looks like this when we enter Java code
139 //
140 // [ return_from_Java ] <--- sp
141 // [ argument word n ]
142 // ...
143 // -29 [ argument word 1 ]
144 // -28 [ saved Floating-point Control Register ]
145 // -26 [ saved v15 ] <--- sp_after_call
146 // -25 [ saved v14 ]
147 // -24 [ saved v13 ]
148 // -23 [ saved v12 ]
149 // -22 [ saved v11 ]
150 // -21 [ saved v10 ]
151 // -20 [ saved v9 ]
152 // -19 [ saved v8 ]
153 // -18 [ saved r28 ]
154 // -17 [ saved r27 ]
155 // -16 [ saved r26 ]
156 // -15 [ saved r25 ]
157 // -14 [ saved r24 ]
158 // -13 [ saved r23 ]
159 // -12 [ saved r22 ]
160 // -11 [ saved r21 ]
161 // -10 [ saved r20 ]
162 // -9 [ saved r19 ]
163 // -8 [ call wrapper (r0) ]
164 // -7 [ result (r1) ]
165 // -6 [ result type (r2) ]
166 // -5 [ method (r3) ]
167 // -4 [ entry point (r4) ]
168 // -3 [ parameters (r5) ]
169 // -2 [ parameter size (r6) ]
170 // -1 [ thread (r7) ]
171 // 0 [ saved fp (r29) ] <--- fp == saved sp (r31)
172 // 1 [ saved lr (r30) ]
173
174 // Call stub stack layout word offsets from fp
175 enum call_stub_layout {
176 sp_after_call_off = -28,
177
178 fpcr_off = sp_after_call_off,
179 d15_off = -26,
180 d13_off = -24,
181 d11_off = -22,
182 d9_off = -20,
183
184 r28_off = -18,
185 r26_off = -16,
186 r24_off = -14,
187 r22_off = -12,
188 r20_off = -10,
189 call_wrapper_off = -8,
190 result_off = -7,
191 result_type_off = -6,
192 method_off = -5,
193 entry_point_off = -4,
194 parameter_size_off = -2,
195 thread_off = -1,
196 fp_f = 0,
197 retaddr_off = 1,
198 };
199
200 address generate_call_stub(address& return_address) {
201 assert((int)frame::entry_frame_after_call_words == -(int)sp_after_call_off + 1 &&
202 (int)frame::entry_frame_call_wrapper_offset == (int)call_wrapper_off,
203 "adjust this code");
204
205 StubGenStubId stub_id = StubGenStubId::call_stub_id;
206 StubCodeMark mark(this, stub_id);
207 address start = __ pc();
208
209 const Address sp_after_call (rfp, sp_after_call_off * wordSize);
210
211 const Address fpcr_save (rfp, fpcr_off * wordSize);
212 const Address call_wrapper (rfp, call_wrapper_off * wordSize);
213 const Address result (rfp, result_off * wordSize);
214 const Address result_type (rfp, result_type_off * wordSize);
215 const Address method (rfp, method_off * wordSize);
216 const Address entry_point (rfp, entry_point_off * wordSize);
217 const Address parameter_size(rfp, parameter_size_off * wordSize);
218
219 const Address thread (rfp, thread_off * wordSize);
220
221 const Address d15_save (rfp, d15_off * wordSize);
222 const Address d13_save (rfp, d13_off * wordSize);
223 const Address d11_save (rfp, d11_off * wordSize);
224 const Address d9_save (rfp, d9_off * wordSize);
225
226 const Address r28_save (rfp, r28_off * wordSize);
227 const Address r26_save (rfp, r26_off * wordSize);
228 const Address r24_save (rfp, r24_off * wordSize);
229 const Address r22_save (rfp, r22_off * wordSize);
230 const Address r20_save (rfp, r20_off * wordSize);
231
232 // stub code
233
234 address aarch64_entry = __ pc();
235
236 // set up frame and move sp to end of save area
237 __ enter();
238 __ sub(sp, rfp, -sp_after_call_off * wordSize);
239
240 // save register parameters and Java scratch/global registers
241 // n.b. we save thread even though it gets installed in
242 // rthread because we want to sanity check rthread later
243 __ str(c_rarg7, thread);
244 __ strw(c_rarg6, parameter_size);
245 __ stp(c_rarg4, c_rarg5, entry_point);
246 __ stp(c_rarg2, c_rarg3, result_type);
247 __ stp(c_rarg0, c_rarg1, call_wrapper);
248
249 __ stp(r20, r19, r20_save);
250 __ stp(r22, r21, r22_save);
251 __ stp(r24, r23, r24_save);
252 __ stp(r26, r25, r26_save);
253 __ stp(r28, r27, r28_save);
254
255 __ stpd(v9, v8, d9_save);
256 __ stpd(v11, v10, d11_save);
257 __ stpd(v13, v12, d13_save);
258 __ stpd(v15, v14, d15_save);
259
260 __ get_fpcr(rscratch1);
261 __ str(rscratch1, fpcr_save);
262 // Set FPCR to the state we need. We do want Round to Nearest. We
263 // don't want non-IEEE rounding modes or floating-point traps.
264 __ bfi(rscratch1, zr, 22, 4); // Clear DN, FZ, and Rmode
265 __ bfi(rscratch1, zr, 8, 5); // Clear exception-control bits (8-12)
266 __ set_fpcr(rscratch1);
267
268 // install Java thread in global register now we have saved
269 // whatever value it held
270 __ mov(rthread, c_rarg7);
271 // And method
272 __ mov(rmethod, c_rarg3);
273
274 // set up the heapbase register
275 __ reinit_heapbase();
276
277 #ifdef ASSERT
278 // make sure we have no pending exceptions
279 {
280 Label L;
281 __ ldr(rscratch1, Address(rthread, in_bytes(Thread::pending_exception_offset())));
282 __ cmp(rscratch1, (u1)NULL_WORD);
283 __ br(Assembler::EQ, L);
284 __ stop("StubRoutines::call_stub: entered with pending exception");
285 __ BIND(L);
286 }
287 #endif
288 // pass parameters if any
289 __ mov(esp, sp);
290 __ sub(rscratch1, sp, c_rarg6, ext::uxtw, LogBytesPerWord); // Move SP out of the way
291 __ andr(sp, rscratch1, -2 * wordSize);
292
293 BLOCK_COMMENT("pass parameters if any");
294 Label parameters_done;
295 // parameter count is still in c_rarg6
296 // and parameter pointer identifying param 1 is in c_rarg5
297 __ cbzw(c_rarg6, parameters_done);
298
299 address loop = __ pc();
300 __ ldr(rscratch1, Address(__ post(c_rarg5, wordSize)));
301 __ subsw(c_rarg6, c_rarg6, 1);
302 __ push(rscratch1);
303 __ br(Assembler::GT, loop);
304
305 __ BIND(parameters_done);
306
307 // call Java entry -- passing methdoOop, and current sp
308 // rmethod: Method*
309 // r19_sender_sp: sender sp
310 BLOCK_COMMENT("call Java function");
311 __ mov(r19_sender_sp, sp);
312 __ blr(c_rarg4);
313
314 // we do this here because the notify will already have been done
315 // if we get to the next instruction via an exception
316 //
317 // n.b. adding this instruction here affects the calculation of
318 // whether or not a routine returns to the call stub (used when
319 // doing stack walks) since the normal test is to check the return
320 // pc against the address saved below. so we may need to allow for
321 // this extra instruction in the check.
322
323 // save current address for use by exception handling code
324
325 return_address = __ pc();
326
327 // store result depending on type (everything that is not
328 // T_OBJECT, T_LONG, T_FLOAT or T_DOUBLE is treated as T_INT)
329 // n.b. this assumes Java returns an integral result in r0
330 // and a floating result in j_farg0
331 __ ldr(j_rarg2, result);
332 Label is_long, is_float, is_double, exit;
333 __ ldr(j_rarg1, result_type);
334 __ cmp(j_rarg1, (u1)T_OBJECT);
335 __ br(Assembler::EQ, is_long);
336 __ cmp(j_rarg1, (u1)T_LONG);
337 __ br(Assembler::EQ, is_long);
338 __ cmp(j_rarg1, (u1)T_FLOAT);
339 __ br(Assembler::EQ, is_float);
340 __ cmp(j_rarg1, (u1)T_DOUBLE);
341 __ br(Assembler::EQ, is_double);
342
343 // handle T_INT case
344 __ strw(r0, Address(j_rarg2));
345
346 __ BIND(exit);
347
348 // pop parameters
349 __ sub(esp, rfp, -sp_after_call_off * wordSize);
350
351 #ifdef ASSERT
352 // verify that threads correspond
353 {
354 Label L, S;
355 __ ldr(rscratch1, thread);
356 __ cmp(rthread, rscratch1);
357 __ br(Assembler::NE, S);
358 __ get_thread(rscratch1);
359 __ cmp(rthread, rscratch1);
360 __ br(Assembler::EQ, L);
361 __ BIND(S);
362 __ stop("StubRoutines::call_stub: threads must correspond");
363 __ BIND(L);
364 }
365 #endif
366
367 __ pop_cont_fastpath(rthread);
368
369 // restore callee-save registers
370 __ ldpd(v15, v14, d15_save);
371 __ ldpd(v13, v12, d13_save);
372 __ ldpd(v11, v10, d11_save);
373 __ ldpd(v9, v8, d9_save);
374
375 __ ldp(r28, r27, r28_save);
376 __ ldp(r26, r25, r26_save);
377 __ ldp(r24, r23, r24_save);
378 __ ldp(r22, r21, r22_save);
379 __ ldp(r20, r19, r20_save);
380
381 // restore fpcr
382 __ ldr(rscratch1, fpcr_save);
383 __ set_fpcr(rscratch1);
384
385 __ ldp(c_rarg0, c_rarg1, call_wrapper);
386 __ ldrw(c_rarg2, result_type);
387 __ ldr(c_rarg3, method);
388 __ ldp(c_rarg4, c_rarg5, entry_point);
389 __ ldp(c_rarg6, c_rarg7, parameter_size);
390
391 // leave frame and return to caller
392 __ leave();
393 __ ret(lr);
394
395 // handle return types different from T_INT
396
397 __ BIND(is_long);
398 __ str(r0, Address(j_rarg2, 0));
399 __ br(Assembler::AL, exit);
400
401 __ BIND(is_float);
402 __ strs(j_farg0, Address(j_rarg2, 0));
403 __ br(Assembler::AL, exit);
404
405 __ BIND(is_double);
406 __ strd(j_farg0, Address(j_rarg2, 0));
407 __ br(Assembler::AL, exit);
408
409 return start;
410 }
411
412 // Return point for a Java call if there's an exception thrown in
413 // Java code. The exception is caught and transformed into a
414 // pending exception stored in JavaThread that can be tested from
415 // within the VM.
416 //
417 // Note: Usually the parameters are removed by the callee. In case
418 // of an exception crossing an activation frame boundary, that is
419 // not the case if the callee is compiled code => need to setup the
420 // rsp.
421 //
422 // r0: exception oop
423
424 address generate_catch_exception() {
425 StubGenStubId stub_id = StubGenStubId::catch_exception_id;
426 StubCodeMark mark(this, stub_id);
427 address start = __ pc();
428
429 // same as in generate_call_stub():
430 const Address sp_after_call(rfp, sp_after_call_off * wordSize);
431 const Address thread (rfp, thread_off * wordSize);
432
433 #ifdef ASSERT
434 // verify that threads correspond
435 {
436 Label L, S;
437 __ ldr(rscratch1, thread);
438 __ cmp(rthread, rscratch1);
439 __ br(Assembler::NE, S);
440 __ get_thread(rscratch1);
441 __ cmp(rthread, rscratch1);
442 __ br(Assembler::EQ, L);
443 __ bind(S);
444 __ stop("StubRoutines::catch_exception: threads must correspond");
445 __ bind(L);
446 }
447 #endif
448
449 // set pending exception
450 __ verify_oop(r0);
451
452 __ str(r0, Address(rthread, Thread::pending_exception_offset()));
453 __ mov(rscratch1, (address)__FILE__);
454 __ str(rscratch1, Address(rthread, Thread::exception_file_offset()));
455 __ movw(rscratch1, (int)__LINE__);
456 __ strw(rscratch1, Address(rthread, Thread::exception_line_offset()));
457
458 // complete return to VM
459 assert(StubRoutines::_call_stub_return_address != nullptr,
460 "_call_stub_return_address must have been generated before");
461 __ b(StubRoutines::_call_stub_return_address);
462
463 return start;
464 }
465
466 // Continuation point for runtime calls returning with a pending
467 // exception. The pending exception check happened in the runtime
468 // or native call stub. The pending exception in Thread is
469 // converted into a Java-level exception.
470 //
471 // Contract with Java-level exception handlers:
472 // r0: exception
473 // r3: throwing pc
474 //
475 // NOTE: At entry of this stub, exception-pc must be in LR !!
476
477 // NOTE: this is always used as a jump target within generated code
478 // so it just needs to be generated code with no x86 prolog
479
480 address generate_forward_exception() {
481 StubGenStubId stub_id = StubGenStubId::forward_exception_id;
482 StubCodeMark mark(this, stub_id);
483 address start = __ pc();
484
485 // Upon entry, LR points to the return address returning into
486 // Java (interpreted or compiled) code; i.e., the return address
487 // becomes the throwing pc.
488 //
489 // Arguments pushed before the runtime call are still on the stack
490 // but the exception handler will reset the stack pointer ->
491 // ignore them. A potential result in registers can be ignored as
492 // well.
493
494 #ifdef ASSERT
495 // make sure this code is only executed if there is a pending exception
496 {
497 Label L;
498 __ ldr(rscratch1, Address(rthread, Thread::pending_exception_offset()));
499 __ cbnz(rscratch1, L);
500 __ stop("StubRoutines::forward exception: no pending exception (1)");
501 __ bind(L);
502 }
503 #endif
504
505 // compute exception handler into r19
506
507 // call the VM to find the handler address associated with the
508 // caller address. pass thread in r0 and caller pc (ret address)
509 // in r1. n.b. the caller pc is in lr, unlike x86 where it is on
510 // the stack.
511 __ mov(c_rarg1, lr);
512 // lr will be trashed by the VM call so we move it to R19
513 // (callee-saved) because we also need to pass it to the handler
514 // returned by this call.
515 __ mov(r19, lr);
516 BLOCK_COMMENT("call exception_handler_for_return_address");
517 __ call_VM_leaf(CAST_FROM_FN_PTR(address,
518 SharedRuntime::exception_handler_for_return_address),
519 rthread, c_rarg1);
520 // Reinitialize the ptrue predicate register, in case the external runtime
521 // call clobbers ptrue reg, as we may return to SVE compiled code.
522 __ reinitialize_ptrue();
523
524 // we should not really care that lr is no longer the callee
525 // address. we saved the value the handler needs in r19 so we can
526 // just copy it to r3. however, the C2 handler will push its own
527 // frame and then calls into the VM and the VM code asserts that
528 // the PC for the frame above the handler belongs to a compiled
529 // Java method. So, we restore lr here to satisfy that assert.
530 __ mov(lr, r19);
531 // setup r0 & r3 & clear pending exception
532 __ mov(r3, r19);
533 __ mov(r19, r0);
534 __ ldr(r0, Address(rthread, Thread::pending_exception_offset()));
535 __ str(zr, Address(rthread, Thread::pending_exception_offset()));
536
537 #ifdef ASSERT
538 // make sure exception is set
539 {
540 Label L;
541 __ cbnz(r0, L);
542 __ stop("StubRoutines::forward exception: no pending exception (2)");
543 __ bind(L);
544 }
545 #endif
546
547 // continue at exception handler
548 // r0: exception
549 // r3: throwing pc
550 // r19: exception handler
551 __ verify_oop(r0);
552 __ br(r19);
553
554 return start;
555 }
556
557 // Non-destructive plausibility checks for oops
558 //
559 // Arguments:
560 // r0: oop to verify
561 // rscratch1: error message
562 //
563 // Stack after saving c_rarg3:
564 // [tos + 0]: saved c_rarg3
565 // [tos + 1]: saved c_rarg2
566 // [tos + 2]: saved lr
567 // [tos + 3]: saved rscratch2
568 // [tos + 4]: saved r0
569 // [tos + 5]: saved rscratch1
570 address generate_verify_oop() {
571 StubGenStubId stub_id = StubGenStubId::verify_oop_id;
572 StubCodeMark mark(this, stub_id);
573 address start = __ pc();
574
575 Label exit, error;
576
577 // save c_rarg2 and c_rarg3
578 __ stp(c_rarg3, c_rarg2, Address(__ pre(sp, -16)));
579
580 // __ incrementl(ExternalAddress((address) StubRoutines::verify_oop_count_addr()));
581 __ lea(c_rarg2, ExternalAddress((address) StubRoutines::verify_oop_count_addr()));
582 __ ldr(c_rarg3, Address(c_rarg2));
583 __ add(c_rarg3, c_rarg3, 1);
584 __ str(c_rarg3, Address(c_rarg2));
585
586 // object is in r0
587 // make sure object is 'reasonable'
588 __ cbz(r0, exit); // if obj is null it is OK
589
590 BarrierSetAssembler* bs_asm = BarrierSet::barrier_set()->barrier_set_assembler();
591 bs_asm->check_oop(_masm, r0, c_rarg2, c_rarg3, error);
592
593 // return if everything seems ok
594 __ bind(exit);
595
596 __ ldp(c_rarg3, c_rarg2, Address(__ post(sp, 16)));
597 __ ret(lr);
598
599 // handle errors
600 __ bind(error);
601 __ ldp(c_rarg3, c_rarg2, Address(__ post(sp, 16)));
602
603 __ push(RegSet::range(r0, r29), sp);
604 // debug(char* msg, int64_t pc, int64_t regs[])
605 __ mov(c_rarg0, rscratch1); // pass address of error message
606 __ mov(c_rarg1, lr); // pass return address
607 __ mov(c_rarg2, sp); // pass address of regs on stack
608 #ifndef PRODUCT
609 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
610 #endif
611 BLOCK_COMMENT("call MacroAssembler::debug");
612 __ mov(rscratch1, CAST_FROM_FN_PTR(address, MacroAssembler::debug64));
613 __ blr(rscratch1);
614 __ hlt(0);
615
616 return start;
617 }
618
619 // Generate indices for iota vector.
620 address generate_iota_indices(StubGenStubId stub_id) {
621 __ align(CodeEntryAlignment);
622 StubCodeMark mark(this, stub_id);
623 address start = __ pc();
624 // B
625 __ emit_data64(0x0706050403020100, relocInfo::none);
626 __ emit_data64(0x0F0E0D0C0B0A0908, relocInfo::none);
627 // H
628 __ emit_data64(0x0003000200010000, relocInfo::none);
629 __ emit_data64(0x0007000600050004, relocInfo::none);
630 // S
631 __ emit_data64(0x0000000100000000, relocInfo::none);
632 __ emit_data64(0x0000000300000002, relocInfo::none);
633 // D
634 __ emit_data64(0x0000000000000000, relocInfo::none);
635 __ emit_data64(0x0000000000000001, relocInfo::none);
636 // S - FP
637 __ emit_data64(0x3F80000000000000, relocInfo::none); // 0.0f, 1.0f
638 __ emit_data64(0x4040000040000000, relocInfo::none); // 2.0f, 3.0f
639 // D - FP
640 __ emit_data64(0x0000000000000000, relocInfo::none); // 0.0d
641 __ emit_data64(0x3FF0000000000000, relocInfo::none); // 1.0d
642 return start;
643 }
644
645 // The inner part of zero_words(). This is the bulk operation,
646 // zeroing words in blocks, possibly using DC ZVA to do it. The
647 // caller is responsible for zeroing the last few words.
648 //
649 // Inputs:
650 // r10: the HeapWord-aligned base address of an array to zero.
651 // r11: the count in HeapWords, r11 > 0.
652 //
653 // Returns r10 and r11, adjusted for the caller to clear.
654 // r10: the base address of the tail of words left to clear.
655 // r11: the number of words in the tail.
656 // r11 < MacroAssembler::zero_words_block_size.
657
658 address generate_zero_blocks() {
659 Label done;
660 Label base_aligned;
661
662 Register base = r10, cnt = r11;
663
664 __ align(CodeEntryAlignment);
665 StubGenStubId stub_id = StubGenStubId::zero_blocks_id;
666 StubCodeMark mark(this, stub_id);
667 address start = __ pc();
668
669 if (UseBlockZeroing) {
670 int zva_length = VM_Version::zva_length();
671
672 // Ensure ZVA length can be divided by 16. This is required by
673 // the subsequent operations.
674 assert (zva_length % 16 == 0, "Unexpected ZVA Length");
675
676 __ tbz(base, 3, base_aligned);
677 __ str(zr, Address(__ post(base, 8)));
678 __ sub(cnt, cnt, 1);
679 __ bind(base_aligned);
680
681 // Ensure count >= zva_length * 2 so that it still deserves a zva after
682 // alignment.
683 Label small;
684 int low_limit = MAX2(zva_length * 2, (int)BlockZeroingLowLimit);
685 __ subs(rscratch1, cnt, low_limit >> 3);
686 __ br(Assembler::LT, small);
687 __ zero_dcache_blocks(base, cnt);
688 __ bind(small);
689 }
690
691 {
692 // Number of stp instructions we'll unroll
693 const int unroll =
694 MacroAssembler::zero_words_block_size / 2;
695 // Clear the remaining blocks.
696 Label loop;
697 __ subs(cnt, cnt, unroll * 2);
698 __ br(Assembler::LT, done);
699 __ bind(loop);
700 for (int i = 0; i < unroll; i++)
701 __ stp(zr, zr, __ post(base, 16));
702 __ subs(cnt, cnt, unroll * 2);
703 __ br(Assembler::GE, loop);
704 __ bind(done);
705 __ add(cnt, cnt, unroll * 2);
706 }
707
708 __ ret(lr);
709
710 return start;
711 }
712
713
714 typedef enum {
715 copy_forwards = 1,
716 copy_backwards = -1
717 } copy_direction;
718
719 // Helper object to reduce noise when telling the GC barriers how to perform loads and stores
720 // for arraycopy stubs.
721 class ArrayCopyBarrierSetHelper : StackObj {
722 BarrierSetAssembler* _bs_asm;
723 MacroAssembler* _masm;
724 DecoratorSet _decorators;
725 BasicType _type;
726 Register _gct1;
727 Register _gct2;
728 Register _gct3;
729 FloatRegister _gcvt1;
730 FloatRegister _gcvt2;
731 FloatRegister _gcvt3;
732
733 public:
734 ArrayCopyBarrierSetHelper(MacroAssembler* masm,
735 DecoratorSet decorators,
736 BasicType type,
737 Register gct1,
738 Register gct2,
739 Register gct3,
740 FloatRegister gcvt1,
741 FloatRegister gcvt2,
742 FloatRegister gcvt3)
743 : _bs_asm(BarrierSet::barrier_set()->barrier_set_assembler()),
744 _masm(masm),
745 _decorators(decorators),
746 _type(type),
747 _gct1(gct1),
748 _gct2(gct2),
749 _gct3(gct3),
750 _gcvt1(gcvt1),
751 _gcvt2(gcvt2),
752 _gcvt3(gcvt3) {
753 }
754
755 void copy_load_at_32(FloatRegister dst1, FloatRegister dst2, Address src) {
756 _bs_asm->copy_load_at(_masm, _decorators, _type, 32,
757 dst1, dst2, src,
758 _gct1, _gct2, _gcvt1);
759 }
760
761 void copy_store_at_32(Address dst, FloatRegister src1, FloatRegister src2) {
762 _bs_asm->copy_store_at(_masm, _decorators, _type, 32,
763 dst, src1, src2,
764 _gct1, _gct2, _gct3, _gcvt1, _gcvt2, _gcvt3);
765 }
766
767 void copy_load_at_16(Register dst1, Register dst2, Address src) {
768 _bs_asm->copy_load_at(_masm, _decorators, _type, 16,
769 dst1, dst2, src,
770 _gct1);
771 }
772
773 void copy_store_at_16(Address dst, Register src1, Register src2) {
774 _bs_asm->copy_store_at(_masm, _decorators, _type, 16,
775 dst, src1, src2,
776 _gct1, _gct2, _gct3);
777 }
778
779 void copy_load_at_8(Register dst, Address src) {
780 _bs_asm->copy_load_at(_masm, _decorators, _type, 8,
781 dst, noreg, src,
782 _gct1);
783 }
784
785 void copy_store_at_8(Address dst, Register src) {
786 _bs_asm->copy_store_at(_masm, _decorators, _type, 8,
787 dst, src, noreg,
788 _gct1, _gct2, _gct3);
789 }
790 };
791
792 // Bulk copy of blocks of 8 words.
793 //
794 // count is a count of words.
795 //
796 // Precondition: count >= 8
797 //
798 // Postconditions:
799 //
800 // The least significant bit of count contains the remaining count
801 // of words to copy. The rest of count is trash.
802 //
803 // s and d are adjusted to point to the remaining words to copy
804 //
805 void generate_copy_longs(StubGenStubId stub_id, DecoratorSet decorators, Label &start, Register s, Register d, Register count) {
806 BasicType type;
807 copy_direction direction;
808
809 switch (stub_id) {
810 case copy_byte_f_id:
811 direction = copy_forwards;
812 type = T_BYTE;
813 break;
814 case copy_byte_b_id:
815 direction = copy_backwards;
816 type = T_BYTE;
817 break;
818 case copy_oop_f_id:
819 direction = copy_forwards;
820 type = T_OBJECT;
821 break;
822 case copy_oop_b_id:
823 direction = copy_backwards;
824 type = T_OBJECT;
825 break;
826 case copy_oop_uninit_f_id:
827 direction = copy_forwards;
828 type = T_OBJECT;
829 break;
830 case copy_oop_uninit_b_id:
831 direction = copy_backwards;
832 type = T_OBJECT;
833 break;
834 default:
835 ShouldNotReachHere();
836 }
837
838 int unit = wordSize * direction;
839 int bias = (UseSIMDForMemoryOps ? 4:2) * wordSize;
840
841 const Register t0 = r3, t1 = r4, t2 = r5, t3 = r6,
842 t4 = r7, t5 = r11, t6 = r12, t7 = r13;
843 const Register stride = r14;
844 const Register gct1 = rscratch1, gct2 = rscratch2, gct3 = r10;
845 const FloatRegister gcvt1 = v6, gcvt2 = v7, gcvt3 = v16; // Note that v8-v15 are callee saved
846 ArrayCopyBarrierSetHelper bs(_masm, decorators, type, gct1, gct2, gct3, gcvt1, gcvt2, gcvt3);
847
848 assert_different_registers(rscratch1, rscratch2, t0, t1, t2, t3, t4, t5, t6, t7);
849 assert_different_registers(s, d, count, rscratch1, rscratch2);
850
851 Label again, drain;
852
853 __ align(CodeEntryAlignment);
854
855 StubCodeMark mark(this, stub_id);
856
857 __ bind(start);
858
859 Label unaligned_copy_long;
860 if (AvoidUnalignedAccesses) {
861 __ tbnz(d, 3, unaligned_copy_long);
862 }
863
864 if (direction == copy_forwards) {
865 __ sub(s, s, bias);
866 __ sub(d, d, bias);
867 }
868
869 #ifdef ASSERT
870 // Make sure we are never given < 8 words
871 {
872 Label L;
873 __ cmp(count, (u1)8);
874 __ br(Assembler::GE, L);
875 __ stop("genrate_copy_longs called with < 8 words");
876 __ bind(L);
877 }
878 #endif
879
880 // Fill 8 registers
881 if (UseSIMDForMemoryOps) {
882 bs.copy_load_at_32(v0, v1, Address(s, 4 * unit));
883 bs.copy_load_at_32(v2, v3, Address(__ pre(s, 8 * unit)));
884 } else {
885 bs.copy_load_at_16(t0, t1, Address(s, 2 * unit));
886 bs.copy_load_at_16(t2, t3, Address(s, 4 * unit));
887 bs.copy_load_at_16(t4, t5, Address(s, 6 * unit));
888 bs.copy_load_at_16(t6, t7, Address(__ pre(s, 8 * unit)));
889 }
890
891 __ subs(count, count, 16);
892 __ br(Assembler::LO, drain);
893
894 int prefetch = PrefetchCopyIntervalInBytes;
895 bool use_stride = false;
896 if (direction == copy_backwards) {
897 use_stride = prefetch > 256;
898 prefetch = -prefetch;
899 if (use_stride) __ mov(stride, prefetch);
900 }
901
902 __ bind(again);
903
904 if (PrefetchCopyIntervalInBytes > 0)
905 __ prfm(use_stride ? Address(s, stride) : Address(s, prefetch), PLDL1KEEP);
906
907 if (UseSIMDForMemoryOps) {
908 bs.copy_store_at_32(Address(d, 4 * unit), v0, v1);
909 bs.copy_load_at_32(v0, v1, Address(s, 4 * unit));
910 bs.copy_store_at_32(Address(__ pre(d, 8 * unit)), v2, v3);
911 bs.copy_load_at_32(v2, v3, Address(__ pre(s, 8 * unit)));
912 } else {
913 bs.copy_store_at_16(Address(d, 2 * unit), t0, t1);
914 bs.copy_load_at_16(t0, t1, Address(s, 2 * unit));
915 bs.copy_store_at_16(Address(d, 4 * unit), t2, t3);
916 bs.copy_load_at_16(t2, t3, Address(s, 4 * unit));
917 bs.copy_store_at_16(Address(d, 6 * unit), t4, t5);
918 bs.copy_load_at_16(t4, t5, Address(s, 6 * unit));
919 bs.copy_store_at_16(Address(__ pre(d, 8 * unit)), t6, t7);
920 bs.copy_load_at_16(t6, t7, Address(__ pre(s, 8 * unit)));
921 }
922
923 __ subs(count, count, 8);
924 __ br(Assembler::HS, again);
925
926 // Drain
927 __ bind(drain);
928 if (UseSIMDForMemoryOps) {
929 bs.copy_store_at_32(Address(d, 4 * unit), v0, v1);
930 bs.copy_store_at_32(Address(__ pre(d, 8 * unit)), v2, v3);
931 } else {
932 bs.copy_store_at_16(Address(d, 2 * unit), t0, t1);
933 bs.copy_store_at_16(Address(d, 4 * unit), t2, t3);
934 bs.copy_store_at_16(Address(d, 6 * unit), t4, t5);
935 bs.copy_store_at_16(Address(__ pre(d, 8 * unit)), t6, t7);
936 }
937
938 {
939 Label L1, L2;
940 __ tbz(count, exact_log2(4), L1);
941 if (UseSIMDForMemoryOps) {
942 bs.copy_load_at_32(v0, v1, Address(__ pre(s, 4 * unit)));
943 bs.copy_store_at_32(Address(__ pre(d, 4 * unit)), v0, v1);
944 } else {
945 bs.copy_load_at_16(t0, t1, Address(s, 2 * unit));
946 bs.copy_load_at_16(t2, t3, Address(__ pre(s, 4 * unit)));
947 bs.copy_store_at_16(Address(d, 2 * unit), t0, t1);
948 bs.copy_store_at_16(Address(__ pre(d, 4 * unit)), t2, t3);
949 }
950 __ bind(L1);
951
952 if (direction == copy_forwards) {
953 __ add(s, s, bias);
954 __ add(d, d, bias);
955 }
956
957 __ tbz(count, 1, L2);
958 bs.copy_load_at_16(t0, t1, Address(__ adjust(s, 2 * unit, direction == copy_backwards)));
959 bs.copy_store_at_16(Address(__ adjust(d, 2 * unit, direction == copy_backwards)), t0, t1);
960 __ bind(L2);
961 }
962
963 __ ret(lr);
964
965 if (AvoidUnalignedAccesses) {
966 Label drain, again;
967 // Register order for storing. Order is different for backward copy.
968
969 __ bind(unaligned_copy_long);
970
971 // source address is even aligned, target odd aligned
972 //
973 // when forward copying word pairs we read long pairs at offsets
974 // {0, 2, 4, 6} (in long words). when backwards copying we read
975 // long pairs at offsets {-2, -4, -6, -8}. We adjust the source
976 // address by -2 in the forwards case so we can compute the
977 // source offsets for both as {2, 4, 6, 8} * unit where unit = 1
978 // or -1.
979 //
980 // when forward copying we need to store 1 word, 3 pairs and
981 // then 1 word at offsets {0, 1, 3, 5, 7}. Rather than use a
982 // zero offset We adjust the destination by -1 which means we
983 // have to use offsets { 1, 2, 4, 6, 8} * unit for the stores.
984 //
985 // When backwards copyng we need to store 1 word, 3 pairs and
986 // then 1 word at offsets {-1, -3, -5, -7, -8} i.e. we use
987 // offsets {1, 3, 5, 7, 8} * unit.
988
989 if (direction == copy_forwards) {
990 __ sub(s, s, 16);
991 __ sub(d, d, 8);
992 }
993
994 // Fill 8 registers
995 //
996 // for forwards copy s was offset by -16 from the original input
997 // value of s so the register contents are at these offsets
998 // relative to the 64 bit block addressed by that original input
999 // and so on for each successive 64 byte block when s is updated
1000 //
1001 // t0 at offset 0, t1 at offset 8
1002 // t2 at offset 16, t3 at offset 24
1003 // t4 at offset 32, t5 at offset 40
1004 // t6 at offset 48, t7 at offset 56
1005
1006 // for backwards copy s was not offset so the register contents
1007 // are at these offsets into the preceding 64 byte block
1008 // relative to that original input and so on for each successive
1009 // preceding 64 byte block when s is updated. this explains the
1010 // slightly counter-intuitive looking pattern of register usage
1011 // in the stp instructions for backwards copy.
1012 //
1013 // t0 at offset -16, t1 at offset -8
1014 // t2 at offset -32, t3 at offset -24
1015 // t4 at offset -48, t5 at offset -40
1016 // t6 at offset -64, t7 at offset -56
1017
1018 bs.copy_load_at_16(t0, t1, Address(s, 2 * unit));
1019 bs.copy_load_at_16(t2, t3, Address(s, 4 * unit));
1020 bs.copy_load_at_16(t4, t5, Address(s, 6 * unit));
1021 bs.copy_load_at_16(t6, t7, Address(__ pre(s, 8 * unit)));
1022
1023 __ subs(count, count, 16);
1024 __ br(Assembler::LO, drain);
1025
1026 int prefetch = PrefetchCopyIntervalInBytes;
1027 bool use_stride = false;
1028 if (direction == copy_backwards) {
1029 use_stride = prefetch > 256;
1030 prefetch = -prefetch;
1031 if (use_stride) __ mov(stride, prefetch);
1032 }
1033
1034 __ bind(again);
1035
1036 if (PrefetchCopyIntervalInBytes > 0)
1037 __ prfm(use_stride ? Address(s, stride) : Address(s, prefetch), PLDL1KEEP);
1038
1039 if (direction == copy_forwards) {
1040 // allowing for the offset of -8 the store instructions place
1041 // registers into the target 64 bit block at the following
1042 // offsets
1043 //
1044 // t0 at offset 0
1045 // t1 at offset 8, t2 at offset 16
1046 // t3 at offset 24, t4 at offset 32
1047 // t5 at offset 40, t6 at offset 48
1048 // t7 at offset 56
1049
1050 bs.copy_store_at_8(Address(d, 1 * unit), t0);
1051 bs.copy_store_at_16(Address(d, 2 * unit), t1, t2);
1052 bs.copy_load_at_16(t0, t1, Address(s, 2 * unit));
1053 bs.copy_store_at_16(Address(d, 4 * unit), t3, t4);
1054 bs.copy_load_at_16(t2, t3, Address(s, 4 * unit));
1055 bs.copy_store_at_16(Address(d, 6 * unit), t5, t6);
1056 bs.copy_load_at_16(t4, t5, Address(s, 6 * unit));
1057 bs.copy_store_at_8(Address(__ pre(d, 8 * unit)), t7);
1058 bs.copy_load_at_16(t6, t7, Address(__ pre(s, 8 * unit)));
1059 } else {
1060 // d was not offset when we started so the registers are
1061 // written into the 64 bit block preceding d with the following
1062 // offsets
1063 //
1064 // t1 at offset -8
1065 // t3 at offset -24, t0 at offset -16
1066 // t5 at offset -48, t2 at offset -32
1067 // t7 at offset -56, t4 at offset -48
1068 // t6 at offset -64
1069 //
1070 // note that this matches the offsets previously noted for the
1071 // loads
1072
1073 bs.copy_store_at_8(Address(d, 1 * unit), t1);
1074 bs.copy_store_at_16(Address(d, 3 * unit), t3, t0);
1075 bs.copy_load_at_16(t0, t1, Address(s, 2 * unit));
1076 bs.copy_store_at_16(Address(d, 5 * unit), t5, t2);
1077 bs.copy_load_at_16(t2, t3, Address(s, 4 * unit));
1078 bs.copy_store_at_16(Address(d, 7 * unit), t7, t4);
1079 bs.copy_load_at_16(t4, t5, Address(s, 6 * unit));
1080 bs.copy_store_at_8(Address(__ pre(d, 8 * unit)), t6);
1081 bs.copy_load_at_16(t6, t7, Address(__ pre(s, 8 * unit)));
1082 }
1083
1084 __ subs(count, count, 8);
1085 __ br(Assembler::HS, again);
1086
1087 // Drain
1088 //
1089 // this uses the same pattern of offsets and register arguments
1090 // as above
1091 __ bind(drain);
1092 if (direction == copy_forwards) {
1093 bs.copy_store_at_8(Address(d, 1 * unit), t0);
1094 bs.copy_store_at_16(Address(d, 2 * unit), t1, t2);
1095 bs.copy_store_at_16(Address(d, 4 * unit), t3, t4);
1096 bs.copy_store_at_16(Address(d, 6 * unit), t5, t6);
1097 bs.copy_store_at_8(Address(__ pre(d, 8 * unit)), t7);
1098 } else {
1099 bs.copy_store_at_8(Address(d, 1 * unit), t1);
1100 bs.copy_store_at_16(Address(d, 3 * unit), t3, t0);
1101 bs.copy_store_at_16(Address(d, 5 * unit), t5, t2);
1102 bs.copy_store_at_16(Address(d, 7 * unit), t7, t4);
1103 bs.copy_store_at_8(Address(__ pre(d, 8 * unit)), t6);
1104 }
1105 // now we need to copy any remaining part block which may
1106 // include a 4 word block subblock and/or a 2 word subblock.
1107 // bits 2 and 1 in the count are the tell-tale for whether we
1108 // have each such subblock
1109 {
1110 Label L1, L2;
1111 __ tbz(count, exact_log2(4), L1);
1112 // this is the same as above but copying only 4 longs hence
1113 // with only one intervening stp between the str instructions
1114 // but note that the offsets and registers still follow the
1115 // same pattern
1116 bs.copy_load_at_16(t0, t1, Address(s, 2 * unit));
1117 bs.copy_load_at_16(t2, t3, Address(__ pre(s, 4 * unit)));
1118 if (direction == copy_forwards) {
1119 bs.copy_store_at_8(Address(d, 1 * unit), t0);
1120 bs.copy_store_at_16(Address(d, 2 * unit), t1, t2);
1121 bs.copy_store_at_8(Address(__ pre(d, 4 * unit)), t3);
1122 } else {
1123 bs.copy_store_at_8(Address(d, 1 * unit), t1);
1124 bs.copy_store_at_16(Address(d, 3 * unit), t3, t0);
1125 bs.copy_store_at_8(Address(__ pre(d, 4 * unit)), t2);
1126 }
1127 __ bind(L1);
1128
1129 __ tbz(count, 1, L2);
1130 // this is the same as above but copying only 2 longs hence
1131 // there is no intervening stp between the str instructions
1132 // but note that the offset and register patterns are still
1133 // the same
1134 bs.copy_load_at_16(t0, t1, Address(__ pre(s, 2 * unit)));
1135 if (direction == copy_forwards) {
1136 bs.copy_store_at_8(Address(d, 1 * unit), t0);
1137 bs.copy_store_at_8(Address(__ pre(d, 2 * unit)), t1);
1138 } else {
1139 bs.copy_store_at_8(Address(d, 1 * unit), t1);
1140 bs.copy_store_at_8(Address(__ pre(d, 2 * unit)), t0);
1141 }
1142 __ bind(L2);
1143
1144 // for forwards copy we need to re-adjust the offsets we
1145 // applied so that s and d are follow the last words written
1146
1147 if (direction == copy_forwards) {
1148 __ add(s, s, 16);
1149 __ add(d, d, 8);
1150 }
1151
1152 }
1153
1154 __ ret(lr);
1155 }
1156 }
1157
1158 // Small copy: less than 16 bytes.
1159 //
1160 // NB: Ignores all of the bits of count which represent more than 15
1161 // bytes, so a caller doesn't have to mask them.
1162
1163 void copy_memory_small(DecoratorSet decorators, BasicType type, Register s, Register d, Register count, int step) {
1164 bool is_backwards = step < 0;
1165 size_t granularity = uabs(step);
1166 int direction = is_backwards ? -1 : 1;
1167
1168 Label Lword, Lint, Lshort, Lbyte;
1169
1170 assert(granularity
1171 && granularity <= sizeof (jlong), "Impossible granularity in copy_memory_small");
1172
1173 const Register t0 = r3;
1174 const Register gct1 = rscratch1, gct2 = rscratch2, gct3 = r10;
1175 ArrayCopyBarrierSetHelper bs(_masm, decorators, type, gct1, gct2, gct3, fnoreg, fnoreg, fnoreg);
1176
1177 // ??? I don't know if this bit-test-and-branch is the right thing
1178 // to do. It does a lot of jumping, resulting in several
1179 // mispredicted branches. It might make more sense to do this
1180 // with something like Duff's device with a single computed branch.
1181
1182 __ tbz(count, 3 - exact_log2(granularity), Lword);
1183 bs.copy_load_at_8(t0, Address(__ adjust(s, direction * wordSize, is_backwards)));
1184 bs.copy_store_at_8(Address(__ adjust(d, direction * wordSize, is_backwards)), t0);
1185 __ bind(Lword);
1186
1187 if (granularity <= sizeof (jint)) {
1188 __ tbz(count, 2 - exact_log2(granularity), Lint);
1189 __ ldrw(t0, Address(__ adjust(s, sizeof (jint) * direction, is_backwards)));
1190 __ strw(t0, Address(__ adjust(d, sizeof (jint) * direction, is_backwards)));
1191 __ bind(Lint);
1192 }
1193
1194 if (granularity <= sizeof (jshort)) {
1195 __ tbz(count, 1 - exact_log2(granularity), Lshort);
1196 __ ldrh(t0, Address(__ adjust(s, sizeof (jshort) * direction, is_backwards)));
1197 __ strh(t0, Address(__ adjust(d, sizeof (jshort) * direction, is_backwards)));
1198 __ bind(Lshort);
1199 }
1200
1201 if (granularity <= sizeof (jbyte)) {
1202 __ tbz(count, 0, Lbyte);
1203 __ ldrb(t0, Address(__ adjust(s, sizeof (jbyte) * direction, is_backwards)));
1204 __ strb(t0, Address(__ adjust(d, sizeof (jbyte) * direction, is_backwards)));
1205 __ bind(Lbyte);
1206 }
1207 }
1208
1209 Label copy_f, copy_b;
1210 Label copy_obj_f, copy_obj_b;
1211 Label copy_obj_uninit_f, copy_obj_uninit_b;
1212
1213 // All-singing all-dancing memory copy.
1214 //
1215 // Copy count units of memory from s to d. The size of a unit is
1216 // step, which can be positive or negative depending on the direction
1217 // of copy. If is_aligned is false, we align the source address.
1218 //
1219
1220 void copy_memory(DecoratorSet decorators, BasicType type, bool is_aligned,
1221 Register s, Register d, Register count, int step) {
1222 copy_direction direction = step < 0 ? copy_backwards : copy_forwards;
1223 bool is_backwards = step < 0;
1224 unsigned int granularity = uabs(step);
1225 const Register t0 = r3, t1 = r4;
1226
1227 // <= 80 (or 96 for SIMD) bytes do inline. Direction doesn't matter because we always
1228 // load all the data before writing anything
1229 Label copy4, copy8, copy16, copy32, copy80, copy_big, finish;
1230 const Register t2 = r5, t3 = r6, t4 = r7, t5 = r11;
1231 const Register t6 = r12, t7 = r13, t8 = r14, t9 = r15;
1232 const Register send = r17, dend = r16;
1233 const Register gct1 = rscratch1, gct2 = rscratch2, gct3 = r10;
1234 const FloatRegister gcvt1 = v6, gcvt2 = v7, gcvt3 = v16; // Note that v8-v15 are callee saved
1235 ArrayCopyBarrierSetHelper bs(_masm, decorators, type, gct1, gct2, gct3, gcvt1, gcvt2, gcvt3);
1236
1237 if (PrefetchCopyIntervalInBytes > 0)
1238 __ prfm(Address(s, 0), PLDL1KEEP);
1239 __ cmp(count, u1((UseSIMDForMemoryOps ? 96:80)/granularity));
1240 __ br(Assembler::HI, copy_big);
1241
1242 __ lea(send, Address(s, count, Address::lsl(exact_log2(granularity))));
1243 __ lea(dend, Address(d, count, Address::lsl(exact_log2(granularity))));
1244
1245 __ cmp(count, u1(16/granularity));
1246 __ br(Assembler::LS, copy16);
1247
1248 __ cmp(count, u1(64/granularity));
1249 __ br(Assembler::HI, copy80);
1250
1251 __ cmp(count, u1(32/granularity));
1252 __ br(Assembler::LS, copy32);
1253
1254 // 33..64 bytes
1255 if (UseSIMDForMemoryOps) {
1256 bs.copy_load_at_32(v0, v1, Address(s, 0));
1257 bs.copy_load_at_32(v2, v3, Address(send, -32));
1258 bs.copy_store_at_32(Address(d, 0), v0, v1);
1259 bs.copy_store_at_32(Address(dend, -32), v2, v3);
1260 } else {
1261 bs.copy_load_at_16(t0, t1, Address(s, 0));
1262 bs.copy_load_at_16(t2, t3, Address(s, 16));
1263 bs.copy_load_at_16(t4, t5, Address(send, -32));
1264 bs.copy_load_at_16(t6, t7, Address(send, -16));
1265
1266 bs.copy_store_at_16(Address(d, 0), t0, t1);
1267 bs.copy_store_at_16(Address(d, 16), t2, t3);
1268 bs.copy_store_at_16(Address(dend, -32), t4, t5);
1269 bs.copy_store_at_16(Address(dend, -16), t6, t7);
1270 }
1271 __ b(finish);
1272
1273 // 17..32 bytes
1274 __ bind(copy32);
1275 bs.copy_load_at_16(t0, t1, Address(s, 0));
1276 bs.copy_load_at_16(t6, t7, Address(send, -16));
1277
1278 bs.copy_store_at_16(Address(d, 0), t0, t1);
1279 bs.copy_store_at_16(Address(dend, -16), t6, t7);
1280 __ b(finish);
1281
1282 // 65..80/96 bytes
1283 // (96 bytes if SIMD because we do 32 byes per instruction)
1284 __ bind(copy80);
1285 if (UseSIMDForMemoryOps) {
1286 bs.copy_load_at_32(v0, v1, Address(s, 0));
1287 bs.copy_load_at_32(v2, v3, Address(s, 32));
1288 // Unaligned pointers can be an issue for copying.
1289 // The issue has more chances to happen when granularity of data is
1290 // less than 4(sizeof(jint)). Pointers for arrays of jint are at least
1291 // 4 byte aligned. Pointers for arrays of jlong are 8 byte aligned.
1292 // The most performance drop has been seen for the range 65-80 bytes.
1293 // For such cases using the pair of ldp/stp instead of the third pair of
1294 // ldpq/stpq fixes the performance issue.
1295 if (granularity < sizeof (jint)) {
1296 Label copy96;
1297 __ cmp(count, u1(80/granularity));
1298 __ br(Assembler::HI, copy96);
1299 bs.copy_load_at_16(t0, t1, Address(send, -16));
1300
1301 bs.copy_store_at_32(Address(d, 0), v0, v1);
1302 bs.copy_store_at_32(Address(d, 32), v2, v3);
1303
1304 bs.copy_store_at_16(Address(dend, -16), t0, t1);
1305 __ b(finish);
1306
1307 __ bind(copy96);
1308 }
1309 bs.copy_load_at_32(v4, v5, Address(send, -32));
1310
1311 bs.copy_store_at_32(Address(d, 0), v0, v1);
1312 bs.copy_store_at_32(Address(d, 32), v2, v3);
1313
1314 bs.copy_store_at_32(Address(dend, -32), v4, v5);
1315 } else {
1316 bs.copy_load_at_16(t0, t1, Address(s, 0));
1317 bs.copy_load_at_16(t2, t3, Address(s, 16));
1318 bs.copy_load_at_16(t4, t5, Address(s, 32));
1319 bs.copy_load_at_16(t6, t7, Address(s, 48));
1320 bs.copy_load_at_16(t8, t9, Address(send, -16));
1321
1322 bs.copy_store_at_16(Address(d, 0), t0, t1);
1323 bs.copy_store_at_16(Address(d, 16), t2, t3);
1324 bs.copy_store_at_16(Address(d, 32), t4, t5);
1325 bs.copy_store_at_16(Address(d, 48), t6, t7);
1326 bs.copy_store_at_16(Address(dend, -16), t8, t9);
1327 }
1328 __ b(finish);
1329
1330 // 0..16 bytes
1331 __ bind(copy16);
1332 __ cmp(count, u1(8/granularity));
1333 __ br(Assembler::LO, copy8);
1334
1335 // 8..16 bytes
1336 bs.copy_load_at_8(t0, Address(s, 0));
1337 bs.copy_load_at_8(t1, Address(send, -8));
1338 bs.copy_store_at_8(Address(d, 0), t0);
1339 bs.copy_store_at_8(Address(dend, -8), t1);
1340 __ b(finish);
1341
1342 if (granularity < 8) {
1343 // 4..7 bytes
1344 __ bind(copy8);
1345 __ tbz(count, 2 - exact_log2(granularity), copy4);
1346 __ ldrw(t0, Address(s, 0));
1347 __ ldrw(t1, Address(send, -4));
1348 __ strw(t0, Address(d, 0));
1349 __ strw(t1, Address(dend, -4));
1350 __ b(finish);
1351 if (granularity < 4) {
1352 // 0..3 bytes
1353 __ bind(copy4);
1354 __ cbz(count, finish); // get rid of 0 case
1355 if (granularity == 2) {
1356 __ ldrh(t0, Address(s, 0));
1357 __ strh(t0, Address(d, 0));
1358 } else { // granularity == 1
1359 // Now 1..3 bytes. Handle the 1 and 2 byte case by copying
1360 // the first and last byte.
1361 // Handle the 3 byte case by loading and storing base + count/2
1362 // (count == 1 (s+0)->(d+0), count == 2,3 (s+1) -> (d+1))
1363 // This does means in the 1 byte case we load/store the same
1364 // byte 3 times.
1365 __ lsr(count, count, 1);
1366 __ ldrb(t0, Address(s, 0));
1367 __ ldrb(t1, Address(send, -1));
1368 __ ldrb(t2, Address(s, count));
1369 __ strb(t0, Address(d, 0));
1370 __ strb(t1, Address(dend, -1));
1371 __ strb(t2, Address(d, count));
1372 }
1373 __ b(finish);
1374 }
1375 }
1376
1377 __ bind(copy_big);
1378 if (is_backwards) {
1379 __ lea(s, Address(s, count, Address::lsl(exact_log2(-step))));
1380 __ lea(d, Address(d, count, Address::lsl(exact_log2(-step))));
1381 }
1382
1383 // Now we've got the small case out of the way we can align the
1384 // source address on a 2-word boundary.
1385
1386 // Here we will materialize a count in r15, which is used by copy_memory_small
1387 // and the various generate_copy_longs stubs that we use for 2 word aligned bytes.
1388 // Up until here, we have used t9, which aliases r15, but from here on, that register
1389 // can not be used as a temp register, as it contains the count.
1390
1391 Label aligned;
1392
1393 if (is_aligned) {
1394 // We may have to adjust by 1 word to get s 2-word-aligned.
1395 __ tbz(s, exact_log2(wordSize), aligned);
1396 bs.copy_load_at_8(t0, Address(__ adjust(s, direction * wordSize, is_backwards)));
1397 bs.copy_store_at_8(Address(__ adjust(d, direction * wordSize, is_backwards)), t0);
1398 __ sub(count, count, wordSize/granularity);
1399 } else {
1400 if (is_backwards) {
1401 __ andr(r15, s, 2 * wordSize - 1);
1402 } else {
1403 __ neg(r15, s);
1404 __ andr(r15, r15, 2 * wordSize - 1);
1405 }
1406 // r15 is the byte adjustment needed to align s.
1407 __ cbz(r15, aligned);
1408 int shift = exact_log2(granularity);
1409 if (shift > 0) {
1410 __ lsr(r15, r15, shift);
1411 }
1412 __ sub(count, count, r15);
1413
1414 #if 0
1415 // ?? This code is only correct for a disjoint copy. It may or
1416 // may not make sense to use it in that case.
1417
1418 // Copy the first pair; s and d may not be aligned.
1419 __ ldp(t0, t1, Address(s, is_backwards ? -2 * wordSize : 0));
1420 __ stp(t0, t1, Address(d, is_backwards ? -2 * wordSize : 0));
1421
1422 // Align s and d, adjust count
1423 if (is_backwards) {
1424 __ sub(s, s, r15);
1425 __ sub(d, d, r15);
1426 } else {
1427 __ add(s, s, r15);
1428 __ add(d, d, r15);
1429 }
1430 #else
1431 copy_memory_small(decorators, type, s, d, r15, step);
1432 #endif
1433 }
1434
1435 __ bind(aligned);
1436
1437 // s is now 2-word-aligned.
1438
1439 // We have a count of units and some trailing bytes. Adjust the
1440 // count and do a bulk copy of words. If the shift is zero
1441 // perform a move instead to benefit from zero latency moves.
1442 int shift = exact_log2(wordSize/granularity);
1443 if (shift > 0) {
1444 __ lsr(r15, count, shift);
1445 } else {
1446 __ mov(r15, count);
1447 }
1448 if (direction == copy_forwards) {
1449 if (type != T_OBJECT) {
1450 __ bl(copy_f);
1451 } else if ((decorators & IS_DEST_UNINITIALIZED) != 0) {
1452 __ bl(copy_obj_uninit_f);
1453 } else {
1454 __ bl(copy_obj_f);
1455 }
1456 } else {
1457 if (type != T_OBJECT) {
1458 __ bl(copy_b);
1459 } else if ((decorators & IS_DEST_UNINITIALIZED) != 0) {
1460 __ bl(copy_obj_uninit_b);
1461 } else {
1462 __ bl(copy_obj_b);
1463 }
1464 }
1465
1466 // And the tail.
1467 copy_memory_small(decorators, type, s, d, count, step);
1468
1469 if (granularity >= 8) __ bind(copy8);
1470 if (granularity >= 4) __ bind(copy4);
1471 __ bind(finish);
1472 }
1473
1474
1475 void clobber_registers() {
1476 #ifdef ASSERT
1477 RegSet clobbered
1478 = MacroAssembler::call_clobbered_gp_registers() - rscratch1;
1479 __ mov(rscratch1, (uint64_t)0xdeadbeef);
1480 __ orr(rscratch1, rscratch1, rscratch1, Assembler::LSL, 32);
1481 for (RegSetIterator<Register> it = clobbered.begin(); *it != noreg; ++it) {
1482 __ mov(*it, rscratch1);
1483 }
1484 #endif
1485
1486 }
1487
1488 // Scan over array at a for count oops, verifying each one.
1489 // Preserves a and count, clobbers rscratch1 and rscratch2.
1490 void verify_oop_array (int size, Register a, Register count, Register temp) {
1491 Label loop, end;
1492 __ mov(rscratch1, a);
1493 __ mov(rscratch2, zr);
1494 __ bind(loop);
1495 __ cmp(rscratch2, count);
1496 __ br(Assembler::HS, end);
1497 if (size == wordSize) {
1498 __ ldr(temp, Address(a, rscratch2, Address::lsl(exact_log2(size))));
1499 __ verify_oop(temp);
1500 } else {
1501 __ ldrw(temp, Address(a, rscratch2, Address::lsl(exact_log2(size))));
1502 __ decode_heap_oop(temp); // calls verify_oop
1503 }
1504 __ add(rscratch2, rscratch2, 1);
1505 __ b(loop);
1506 __ bind(end);
1507 }
1508
1509 // Arguments:
1510 // stub_id - is used to name the stub and identify all details of
1511 // how to perform the copy.
1512 //
1513 // entry - is assigned to the stub's post push entry point unless
1514 // it is null
1515 //
1516 // Inputs:
1517 // c_rarg0 - source array address
1518 // c_rarg1 - destination array address
1519 // c_rarg2 - element count, treated as ssize_t, can be zero
1520 //
1521 // If 'from' and/or 'to' are aligned on 4-byte boundaries, we let
1522 // the hardware handle it. The two dwords within qwords that span
1523 // cache line boundaries will still be loaded and stored atomically.
1524 //
1525 // Side Effects: entry is set to the (post push) entry point so it
1526 // can be used by the corresponding conjoint copy
1527 // method
1528 //
1529 address generate_disjoint_copy(StubGenStubId stub_id, address *entry) {
1530 Register s = c_rarg0, d = c_rarg1, count = c_rarg2;
1531 RegSet saved_reg = RegSet::of(s, d, count);
1532 int size;
1533 bool aligned;
1534 bool is_oop;
1535 bool dest_uninitialized;
1536 switch (stub_id) {
1537 case jbyte_disjoint_arraycopy_id:
1538 size = sizeof(jbyte);
1539 aligned = false;
1540 is_oop = false;
1541 dest_uninitialized = false;
1542 break;
1543 case arrayof_jbyte_disjoint_arraycopy_id:
1544 size = sizeof(jbyte);
1545 aligned = true;
1546 is_oop = false;
1547 dest_uninitialized = false;
1548 break;
1549 case jshort_disjoint_arraycopy_id:
1550 size = sizeof(jshort);
1551 aligned = false;
1552 is_oop = false;
1553 dest_uninitialized = false;
1554 break;
1555 case arrayof_jshort_disjoint_arraycopy_id:
1556 size = sizeof(jshort);
1557 aligned = true;
1558 is_oop = false;
1559 dest_uninitialized = false;
1560 break;
1561 case jint_disjoint_arraycopy_id:
1562 size = sizeof(jint);
1563 aligned = false;
1564 is_oop = false;
1565 dest_uninitialized = false;
1566 break;
1567 case arrayof_jint_disjoint_arraycopy_id:
1568 size = sizeof(jint);
1569 aligned = true;
1570 is_oop = false;
1571 dest_uninitialized = false;
1572 break;
1573 case jlong_disjoint_arraycopy_id:
1574 // since this is always aligned we can (should!) use the same
1575 // stub as for case arrayof_jlong_disjoint_arraycopy
1576 ShouldNotReachHere();
1577 break;
1578 case arrayof_jlong_disjoint_arraycopy_id:
1579 size = sizeof(jlong);
1580 aligned = true;
1581 is_oop = false;
1582 dest_uninitialized = false;
1583 break;
1584 case oop_disjoint_arraycopy_id:
1585 size = UseCompressedOops ? sizeof (jint) : sizeof (jlong);
1586 aligned = !UseCompressedOops;
1587 is_oop = true;
1588 dest_uninitialized = false;
1589 break;
1590 case arrayof_oop_disjoint_arraycopy_id:
1591 size = UseCompressedOops ? sizeof (jint) : sizeof (jlong);
1592 aligned = !UseCompressedOops;
1593 is_oop = true;
1594 dest_uninitialized = false;
1595 break;
1596 case oop_disjoint_arraycopy_uninit_id:
1597 size = UseCompressedOops ? sizeof (jint) : sizeof (jlong);
1598 aligned = !UseCompressedOops;
1599 is_oop = true;
1600 dest_uninitialized = true;
1601 break;
1602 case arrayof_oop_disjoint_arraycopy_uninit_id:
1603 size = UseCompressedOops ? sizeof (jint) : sizeof (jlong);
1604 aligned = !UseCompressedOops;
1605 is_oop = true;
1606 dest_uninitialized = true;
1607 break;
1608 default:
1609 ShouldNotReachHere();
1610 break;
1611 }
1612
1613 __ align(CodeEntryAlignment);
1614 StubCodeMark mark(this, stub_id);
1615 address start = __ pc();
1616 __ enter();
1617
1618 if (entry != nullptr) {
1619 *entry = __ pc();
1620 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1621 BLOCK_COMMENT("Entry:");
1622 }
1623
1624 DecoratorSet decorators = IN_HEAP | IS_ARRAY | ARRAYCOPY_DISJOINT;
1625 if (dest_uninitialized) {
1626 decorators |= IS_DEST_UNINITIALIZED;
1627 }
1628 if (aligned) {
1629 decorators |= ARRAYCOPY_ALIGNED;
1630 }
1631
1632 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler();
1633 bs->arraycopy_prologue(_masm, decorators, is_oop, s, d, count, saved_reg);
1634
1635 if (is_oop) {
1636 // save regs before copy_memory
1637 __ push(RegSet::of(d, count), sp);
1638 }
1639 {
1640 // UnsafeMemoryAccess page error: continue after unsafe access
1641 bool add_entry = !is_oop && (!aligned || sizeof(jlong) == size);
1642 UnsafeMemoryAccessMark umam(this, add_entry, true);
1643 copy_memory(decorators, is_oop ? T_OBJECT : T_BYTE, aligned, s, d, count, size);
1644 }
1645
1646 if (is_oop) {
1647 __ pop(RegSet::of(d, count), sp);
1648 if (VerifyOops)
1649 verify_oop_array(size, d, count, r16);
1650 }
1651
1652 bs->arraycopy_epilogue(_masm, decorators, is_oop, d, count, rscratch1, RegSet());
1653
1654 __ leave();
1655 __ mov(r0, zr); // return 0
1656 __ ret(lr);
1657 return start;
1658 }
1659
1660 // Arguments:
1661 // stub_id - is used to name the stub and identify all details of
1662 // how to perform the copy.
1663 //
1664 // nooverlap_target - identifes the (post push) entry for the
1665 // corresponding disjoint copy routine which can be
1666 // jumped to if the ranges do not actually overlap
1667 //
1668 // entry - is assigned to the stub's post push entry point unless
1669 // it is null
1670 //
1671 //
1672 // Inputs:
1673 // c_rarg0 - source array address
1674 // c_rarg1 - destination array address
1675 // c_rarg2 - element count, treated as ssize_t, can be zero
1676 //
1677 // If 'from' and/or 'to' are aligned on 4-byte boundaries, we let
1678 // the hardware handle it. The two dwords within qwords that span
1679 // cache line boundaries will still be loaded and stored atomically.
1680 //
1681 // Side Effects:
1682 // entry is set to the no-overlap entry point so it can be used by
1683 // some other conjoint copy method
1684 //
1685 address generate_conjoint_copy(StubGenStubId stub_id, address nooverlap_target, address *entry) {
1686 Register s = c_rarg0, d = c_rarg1, count = c_rarg2;
1687 RegSet saved_regs = RegSet::of(s, d, count);
1688 int size;
1689 bool aligned;
1690 bool is_oop;
1691 bool dest_uninitialized;
1692 switch (stub_id) {
1693 case jbyte_arraycopy_id:
1694 size = sizeof(jbyte);
1695 aligned = false;
1696 is_oop = false;
1697 dest_uninitialized = false;
1698 break;
1699 case arrayof_jbyte_arraycopy_id:
1700 size = sizeof(jbyte);
1701 aligned = true;
1702 is_oop = false;
1703 dest_uninitialized = false;
1704 break;
1705 case jshort_arraycopy_id:
1706 size = sizeof(jshort);
1707 aligned = false;
1708 is_oop = false;
1709 dest_uninitialized = false;
1710 break;
1711 case arrayof_jshort_arraycopy_id:
1712 size = sizeof(jshort);
1713 aligned = true;
1714 is_oop = false;
1715 dest_uninitialized = false;
1716 break;
1717 case jint_arraycopy_id:
1718 size = sizeof(jint);
1719 aligned = false;
1720 is_oop = false;
1721 dest_uninitialized = false;
1722 break;
1723 case arrayof_jint_arraycopy_id:
1724 size = sizeof(jint);
1725 aligned = true;
1726 is_oop = false;
1727 dest_uninitialized = false;
1728 break;
1729 case jlong_arraycopy_id:
1730 // since this is always aligned we can (should!) use the same
1731 // stub as for case arrayof_jlong_disjoint_arraycopy
1732 ShouldNotReachHere();
1733 break;
1734 case arrayof_jlong_arraycopy_id:
1735 size = sizeof(jlong);
1736 aligned = true;
1737 is_oop = false;
1738 dest_uninitialized = false;
1739 break;
1740 case oop_arraycopy_id:
1741 size = UseCompressedOops ? sizeof (jint) : sizeof (jlong);
1742 aligned = !UseCompressedOops;
1743 is_oop = true;
1744 dest_uninitialized = false;
1745 break;
1746 case arrayof_oop_arraycopy_id:
1747 size = UseCompressedOops ? sizeof (jint) : sizeof (jlong);
1748 aligned = !UseCompressedOops;
1749 is_oop = true;
1750 dest_uninitialized = false;
1751 break;
1752 case oop_arraycopy_uninit_id:
1753 size = UseCompressedOops ? sizeof (jint) : sizeof (jlong);
1754 aligned = !UseCompressedOops;
1755 is_oop = true;
1756 dest_uninitialized = true;
1757 break;
1758 case arrayof_oop_arraycopy_uninit_id:
1759 size = UseCompressedOops ? sizeof (jint) : sizeof (jlong);
1760 aligned = !UseCompressedOops;
1761 is_oop = true;
1762 dest_uninitialized = true;
1763 break;
1764 default:
1765 ShouldNotReachHere();
1766 }
1767
1768 StubCodeMark mark(this, stub_id);
1769 address start = __ pc();
1770 __ enter();
1771
1772 if (entry != nullptr) {
1773 *entry = __ pc();
1774 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1775 BLOCK_COMMENT("Entry:");
1776 }
1777
1778 // use fwd copy when (d-s) above_equal (count*size)
1779 __ sub(rscratch1, d, s);
1780 __ cmp(rscratch1, count, Assembler::LSL, exact_log2(size));
1781 __ br(Assembler::HS, nooverlap_target);
1782
1783 DecoratorSet decorators = IN_HEAP | IS_ARRAY;
1784 if (dest_uninitialized) {
1785 decorators |= IS_DEST_UNINITIALIZED;
1786 }
1787 if (aligned) {
1788 decorators |= ARRAYCOPY_ALIGNED;
1789 }
1790
1791 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler();
1792 bs->arraycopy_prologue(_masm, decorators, is_oop, s, d, count, saved_regs);
1793
1794 if (is_oop) {
1795 // save regs before copy_memory
1796 __ push(RegSet::of(d, count), sp);
1797 }
1798 {
1799 // UnsafeMemoryAccess page error: continue after unsafe access
1800 bool add_entry = !is_oop && (!aligned || sizeof(jlong) == size);
1801 UnsafeMemoryAccessMark umam(this, add_entry, true);
1802 copy_memory(decorators, is_oop ? T_OBJECT : T_BYTE, aligned, s, d, count, -size);
1803 }
1804 if (is_oop) {
1805 __ pop(RegSet::of(d, count), sp);
1806 if (VerifyOops)
1807 verify_oop_array(size, d, count, r16);
1808 }
1809 bs->arraycopy_epilogue(_masm, decorators, is_oop, d, count, rscratch1, RegSet());
1810 __ leave();
1811 __ mov(r0, zr); // return 0
1812 __ ret(lr);
1813 return start;
1814 }
1815
1816 // Helper for generating a dynamic type check.
1817 // Smashes rscratch1, rscratch2.
1818 void generate_type_check(Register sub_klass,
1819 Register super_check_offset,
1820 Register super_klass,
1821 Register temp1,
1822 Register temp2,
1823 Register result,
1824 Label& L_success) {
1825 assert_different_registers(sub_klass, super_check_offset, super_klass);
1826
1827 BLOCK_COMMENT("type_check:");
1828
1829 Label L_miss;
1830
1831 __ check_klass_subtype_fast_path(sub_klass, super_klass, noreg, &L_success, &L_miss, nullptr,
1832 super_check_offset);
1833 __ check_klass_subtype_slow_path(sub_klass, super_klass, temp1, temp2, &L_success, nullptr);
1834
1835 // Fall through on failure!
1836 __ BIND(L_miss);
1837 }
1838
1839 //
1840 // Generate checkcasting array copy stub
1841 //
1842 // Input:
1843 // c_rarg0 - source array address
1844 // c_rarg1 - destination array address
1845 // c_rarg2 - element count, treated as ssize_t, can be zero
1846 // c_rarg3 - size_t ckoff (super_check_offset)
1847 // c_rarg4 - oop ckval (super_klass)
1848 //
1849 // Output:
1850 // r0 == 0 - success
1851 // r0 == -1^K - failure, where K is partial transfer count
1852 //
1853 address generate_checkcast_copy(StubGenStubId stub_id, address *entry) {
1854 bool dest_uninitialized;
1855 switch (stub_id) {
1856 case checkcast_arraycopy_id:
1857 dest_uninitialized = false;
1858 break;
1859 case checkcast_arraycopy_uninit_id:
1860 dest_uninitialized = true;
1861 break;
1862 default:
1863 ShouldNotReachHere();
1864 }
1865
1866 Label L_load_element, L_store_element, L_do_card_marks, L_done, L_done_pop;
1867
1868 // Input registers (after setup_arg_regs)
1869 const Register from = c_rarg0; // source array address
1870 const Register to = c_rarg1; // destination array address
1871 const Register count = c_rarg2; // elementscount
1872 const Register ckoff = c_rarg3; // super_check_offset
1873 const Register ckval = c_rarg4; // super_klass
1874
1875 RegSet wb_pre_saved_regs = RegSet::range(c_rarg0, c_rarg4);
1876 RegSet wb_post_saved_regs = RegSet::of(count);
1877
1878 // Registers used as temps (r19, r20, r21, r22 are save-on-entry)
1879 const Register copied_oop = r22; // actual oop copied
1880 const Register count_save = r21; // orig elementscount
1881 const Register start_to = r20; // destination array start address
1882 const Register r19_klass = r19; // oop._klass
1883
1884 // Registers used as gc temps (r5, r6, r7 are save-on-call)
1885 const Register gct1 = r5, gct2 = r6, gct3 = r7;
1886
1887 //---------------------------------------------------------------
1888 // Assembler stub will be used for this call to arraycopy
1889 // if the two arrays are subtypes of Object[] but the
1890 // destination array type is not equal to or a supertype
1891 // of the source type. Each element must be separately
1892 // checked.
1893
1894 assert_different_registers(from, to, count, ckoff, ckval, start_to,
1895 copied_oop, r19_klass, count_save);
1896
1897 __ align(CodeEntryAlignment);
1898 StubCodeMark mark(this, stub_id);
1899 address start = __ pc();
1900
1901 __ enter(); // required for proper stackwalking of RuntimeStub frame
1902
1903 #ifdef ASSERT
1904 // caller guarantees that the arrays really are different
1905 // otherwise, we would have to make conjoint checks
1906 { Label L;
1907 __ b(L); // conjoint check not yet implemented
1908 __ stop("checkcast_copy within a single array");
1909 __ bind(L);
1910 }
1911 #endif //ASSERT
1912
1913 // Caller of this entry point must set up the argument registers.
1914 if (entry != nullptr) {
1915 *entry = __ pc();
1916 BLOCK_COMMENT("Entry:");
1917 }
1918
1919 // Empty array: Nothing to do.
1920 __ cbz(count, L_done);
1921 __ push(RegSet::of(r19, r20, r21, r22), sp);
1922
1923 #ifdef ASSERT
1924 BLOCK_COMMENT("assert consistent ckoff/ckval");
1925 // The ckoff and ckval must be mutually consistent,
1926 // even though caller generates both.
1927 { Label L;
1928 int sco_offset = in_bytes(Klass::super_check_offset_offset());
1929 __ ldrw(start_to, Address(ckval, sco_offset));
1930 __ cmpw(ckoff, start_to);
1931 __ br(Assembler::EQ, L);
1932 __ stop("super_check_offset inconsistent");
1933 __ bind(L);
1934 }
1935 #endif //ASSERT
1936
1937 DecoratorSet decorators = IN_HEAP | IS_ARRAY | ARRAYCOPY_CHECKCAST | ARRAYCOPY_DISJOINT;
1938 bool is_oop = true;
1939 int element_size = UseCompressedOops ? 4 : 8;
1940 if (dest_uninitialized) {
1941 decorators |= IS_DEST_UNINITIALIZED;
1942 }
1943
1944 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler();
1945 bs->arraycopy_prologue(_masm, decorators, is_oop, from, to, count, wb_pre_saved_regs);
1946
1947 // save the original count
1948 __ mov(count_save, count);
1949
1950 // Copy from low to high addresses
1951 __ mov(start_to, to); // Save destination array start address
1952 __ b(L_load_element);
1953
1954 // ======== begin loop ========
1955 // (Loop is rotated; its entry is L_load_element.)
1956 // Loop control:
1957 // for (; count != 0; count--) {
1958 // copied_oop = load_heap_oop(from++);
1959 // ... generate_type_check ...;
1960 // store_heap_oop(to++, copied_oop);
1961 // }
1962 __ align(OptoLoopAlignment);
1963
1964 __ BIND(L_store_element);
1965 bs->copy_store_at(_masm, decorators, T_OBJECT, element_size,
1966 __ post(to, element_size), copied_oop, noreg,
1967 gct1, gct2, gct3);
1968 __ sub(count, count, 1);
1969 __ cbz(count, L_do_card_marks);
1970
1971 // ======== loop entry is here ========
1972 __ BIND(L_load_element);
1973 bs->copy_load_at(_masm, decorators, T_OBJECT, element_size,
1974 copied_oop, noreg, __ post(from, element_size),
1975 gct1);
1976 __ cbz(copied_oop, L_store_element);
1977
1978 __ load_klass(r19_klass, copied_oop);// query the object klass
1979
1980 BLOCK_COMMENT("type_check:");
1981 generate_type_check(/*sub_klass*/r19_klass,
1982 /*super_check_offset*/ckoff,
1983 /*super_klass*/ckval,
1984 /*r_array_base*/gct1,
1985 /*temp2*/gct2,
1986 /*result*/r10, L_store_element);
1987
1988 // Fall through on failure!
1989
1990 // ======== end loop ========
1991
1992 // It was a real error; we must depend on the caller to finish the job.
1993 // Register count = remaining oops, count_orig = total oops.
1994 // Emit GC store barriers for the oops we have copied and report
1995 // their number to the caller.
1996
1997 __ subs(count, count_save, count); // K = partially copied oop count
1998 __ eon(count, count, zr); // report (-1^K) to caller
1999 __ br(Assembler::EQ, L_done_pop);
2000
2001 __ BIND(L_do_card_marks);
2002 bs->arraycopy_epilogue(_masm, decorators, is_oop, start_to, count_save, rscratch1, wb_post_saved_regs);
2003
2004 __ bind(L_done_pop);
2005 __ pop(RegSet::of(r19, r20, r21, r22), sp);
2006 inc_counter_np(SharedRuntime::_checkcast_array_copy_ctr);
2007
2008 __ bind(L_done);
2009 __ mov(r0, count);
2010 __ leave();
2011 __ ret(lr);
2012
2013 return start;
2014 }
2015
2016 // Perform range checks on the proposed arraycopy.
2017 // Kills temp, but nothing else.
2018 // Also, clean the sign bits of src_pos and dst_pos.
2019 void arraycopy_range_checks(Register src, // source array oop (c_rarg0)
2020 Register src_pos, // source position (c_rarg1)
2021 Register dst, // destination array oo (c_rarg2)
2022 Register dst_pos, // destination position (c_rarg3)
2023 Register length,
2024 Register temp,
2025 Label& L_failed) {
2026 BLOCK_COMMENT("arraycopy_range_checks:");
2027
2028 assert_different_registers(rscratch1, temp);
2029
2030 // if (src_pos + length > arrayOop(src)->length()) FAIL;
2031 __ ldrw(rscratch1, Address(src, arrayOopDesc::length_offset_in_bytes()));
2032 __ addw(temp, length, src_pos);
2033 __ cmpw(temp, rscratch1);
2034 __ br(Assembler::HI, L_failed);
2035
2036 // if (dst_pos + length > arrayOop(dst)->length()) FAIL;
2037 __ ldrw(rscratch1, Address(dst, arrayOopDesc::length_offset_in_bytes()));
2038 __ addw(temp, length, dst_pos);
2039 __ cmpw(temp, rscratch1);
2040 __ br(Assembler::HI, L_failed);
2041
2042 // Have to clean up high 32 bits of 'src_pos' and 'dst_pos'.
2043 __ movw(src_pos, src_pos);
2044 __ movw(dst_pos, dst_pos);
2045
2046 BLOCK_COMMENT("arraycopy_range_checks done");
2047 }
2048
2049 // These stubs get called from some dumb test routine.
2050 // I'll write them properly when they're called from
2051 // something that's actually doing something.
2052 static void fake_arraycopy_stub(address src, address dst, int count) {
2053 assert(count == 0, "huh?");
2054 }
2055
2056
2057 //
2058 // Generate 'unsafe' array copy stub
2059 // Though just as safe as the other stubs, it takes an unscaled
2060 // size_t argument instead of an element count.
2061 //
2062 // Input:
2063 // c_rarg0 - source array address
2064 // c_rarg1 - destination array address
2065 // c_rarg2 - byte count, treated as ssize_t, can be zero
2066 //
2067 // Examines the alignment of the operands and dispatches
2068 // to a long, int, short, or byte copy loop.
2069 //
2070 address generate_unsafe_copy(address byte_copy_entry,
2071 address short_copy_entry,
2072 address int_copy_entry,
2073 address long_copy_entry) {
2074 StubGenStubId stub_id = StubGenStubId::unsafe_arraycopy_id;
2075
2076 Label L_long_aligned, L_int_aligned, L_short_aligned;
2077 Register s = c_rarg0, d = c_rarg1, count = c_rarg2;
2078
2079 __ align(CodeEntryAlignment);
2080 StubCodeMark mark(this, stub_id);
2081 address start = __ pc();
2082 __ enter(); // required for proper stackwalking of RuntimeStub frame
2083
2084 // bump this on entry, not on exit:
2085 inc_counter_np(SharedRuntime::_unsafe_array_copy_ctr);
2086
2087 __ orr(rscratch1, s, d);
2088 __ orr(rscratch1, rscratch1, count);
2089
2090 __ andr(rscratch1, rscratch1, BytesPerLong-1);
2091 __ cbz(rscratch1, L_long_aligned);
2092 __ andr(rscratch1, rscratch1, BytesPerInt-1);
2093 __ cbz(rscratch1, L_int_aligned);
2094 __ tbz(rscratch1, 0, L_short_aligned);
2095 __ b(RuntimeAddress(byte_copy_entry));
2096
2097 __ BIND(L_short_aligned);
2098 __ lsr(count, count, LogBytesPerShort); // size => short_count
2099 __ b(RuntimeAddress(short_copy_entry));
2100 __ BIND(L_int_aligned);
2101 __ lsr(count, count, LogBytesPerInt); // size => int_count
2102 __ b(RuntimeAddress(int_copy_entry));
2103 __ BIND(L_long_aligned);
2104 __ lsr(count, count, LogBytesPerLong); // size => long_count
2105 __ b(RuntimeAddress(long_copy_entry));
2106
2107 return start;
2108 }
2109
2110 //
2111 // Generate generic array copy stubs
2112 //
2113 // Input:
2114 // c_rarg0 - src oop
2115 // c_rarg1 - src_pos (32-bits)
2116 // c_rarg2 - dst oop
2117 // c_rarg3 - dst_pos (32-bits)
2118 // c_rarg4 - element count (32-bits)
2119 //
2120 // Output:
2121 // r0 == 0 - success
2122 // r0 == -1^K - failure, where K is partial transfer count
2123 //
2124 address generate_generic_copy(address byte_copy_entry, address short_copy_entry,
2125 address int_copy_entry, address oop_copy_entry,
2126 address long_copy_entry, address checkcast_copy_entry) {
2127 StubGenStubId stub_id = StubGenStubId::generic_arraycopy_id;
2128
2129 Label L_failed, L_objArray;
2130 Label L_copy_bytes, L_copy_shorts, L_copy_ints, L_copy_longs;
2131
2132 // Input registers
2133 const Register src = c_rarg0; // source array oop
2134 const Register src_pos = c_rarg1; // source position
2135 const Register dst = c_rarg2; // destination array oop
2136 const Register dst_pos = c_rarg3; // destination position
2137 const Register length = c_rarg4;
2138
2139
2140 // Registers used as temps
2141 const Register dst_klass = c_rarg5;
2142
2143 __ align(CodeEntryAlignment);
2144
2145 StubCodeMark mark(this, stub_id);
2146
2147 address start = __ pc();
2148
2149 __ enter(); // required for proper stackwalking of RuntimeStub frame
2150
2151 // bump this on entry, not on exit:
2152 inc_counter_np(SharedRuntime::_generic_array_copy_ctr);
2153
2154 //-----------------------------------------------------------------------
2155 // Assembler stub will be used for this call to arraycopy
2156 // if the following conditions are met:
2157 //
2158 // (1) src and dst must not be null.
2159 // (2) src_pos must not be negative.
2160 // (3) dst_pos must not be negative.
2161 // (4) length must not be negative.
2162 // (5) src klass and dst klass should be the same and not null.
2163 // (6) src and dst should be arrays.
2164 // (7) src_pos + length must not exceed length of src.
2165 // (8) dst_pos + length must not exceed length of dst.
2166 //
2167
2168 // if (src == nullptr) return -1;
2169 __ cbz(src, L_failed);
2170
2171 // if (src_pos < 0) return -1;
2172 __ tbnz(src_pos, 31, L_failed); // i.e. sign bit set
2173
2174 // if (dst == nullptr) return -1;
2175 __ cbz(dst, L_failed);
2176
2177 // if (dst_pos < 0) return -1;
2178 __ tbnz(dst_pos, 31, L_failed); // i.e. sign bit set
2179
2180 // registers used as temp
2181 const Register scratch_length = r16; // elements count to copy
2182 const Register scratch_src_klass = r17; // array klass
2183 const Register lh = r15; // layout helper
2184
2185 // if (length < 0) return -1;
2186 __ movw(scratch_length, length); // length (elements count, 32-bits value)
2187 __ tbnz(scratch_length, 31, L_failed); // i.e. sign bit set
2188
2189 __ load_klass(scratch_src_klass, src);
2190 #ifdef ASSERT
2191 // assert(src->klass() != nullptr);
2192 {
2193 BLOCK_COMMENT("assert klasses not null {");
2194 Label L1, L2;
2195 __ cbnz(scratch_src_klass, L2); // it is broken if klass is null
2196 __ bind(L1);
2197 __ stop("broken null klass");
2198 __ bind(L2);
2199 __ load_klass(rscratch1, dst);
2200 __ cbz(rscratch1, L1); // this would be broken also
2201 BLOCK_COMMENT("} assert klasses not null done");
2202 }
2203 #endif
2204
2205 // Load layout helper (32-bits)
2206 //
2207 // |array_tag| | header_size | element_type | |log2_element_size|
2208 // 32 30 24 16 8 2 0
2209 //
2210 // array_tag: typeArray = 0x3, objArray = 0x2, non-array = 0x0
2211 //
2212
2213 const int lh_offset = in_bytes(Klass::layout_helper_offset());
2214
2215 // Handle objArrays completely differently...
2216 const jint objArray_lh = Klass::array_layout_helper(T_OBJECT);
2217 __ ldrw(lh, Address(scratch_src_klass, lh_offset));
2218 __ movw(rscratch1, objArray_lh);
2219 __ eorw(rscratch2, lh, rscratch1);
2220 __ cbzw(rscratch2, L_objArray);
2221
2222 // if (src->klass() != dst->klass()) return -1;
2223 __ load_klass(rscratch2, dst);
2224 __ eor(rscratch2, rscratch2, scratch_src_klass);
2225 __ cbnz(rscratch2, L_failed);
2226
2227 // if (!src->is_Array()) return -1;
2228 __ tbz(lh, 31, L_failed); // i.e. (lh >= 0)
2229
2230 // At this point, it is known to be a typeArray (array_tag 0x3).
2231 #ifdef ASSERT
2232 {
2233 BLOCK_COMMENT("assert primitive array {");
2234 Label L;
2235 __ movw(rscratch2, Klass::_lh_array_tag_type_value << Klass::_lh_array_tag_shift);
2236 __ cmpw(lh, rscratch2);
2237 __ br(Assembler::GE, L);
2238 __ stop("must be a primitive array");
2239 __ bind(L);
2240 BLOCK_COMMENT("} assert primitive array done");
2241 }
2242 #endif
2243
2244 arraycopy_range_checks(src, src_pos, dst, dst_pos, scratch_length,
2245 rscratch2, L_failed);
2246
2247 // TypeArrayKlass
2248 //
2249 // src_addr = (src + array_header_in_bytes()) + (src_pos << log2elemsize);
2250 // dst_addr = (dst + array_header_in_bytes()) + (dst_pos << log2elemsize);
2251 //
2252
2253 const Register rscratch1_offset = rscratch1; // array offset
2254 const Register r15_elsize = lh; // element size
2255
2256 __ ubfx(rscratch1_offset, lh, Klass::_lh_header_size_shift,
2257 exact_log2(Klass::_lh_header_size_mask+1)); // array_offset
2258 __ add(src, src, rscratch1_offset); // src array offset
2259 __ add(dst, dst, rscratch1_offset); // dst array offset
2260 BLOCK_COMMENT("choose copy loop based on element size");
2261
2262 // next registers should be set before the jump to corresponding stub
2263 const Register from = c_rarg0; // source array address
2264 const Register to = c_rarg1; // destination array address
2265 const Register count = c_rarg2; // elements count
2266
2267 // 'from', 'to', 'count' registers should be set in such order
2268 // since they are the same as 'src', 'src_pos', 'dst'.
2269
2270 assert(Klass::_lh_log2_element_size_shift == 0, "fix this code");
2271
2272 // The possible values of elsize are 0-3, i.e. exact_log2(element
2273 // size in bytes). We do a simple bitwise binary search.
2274 __ BIND(L_copy_bytes);
2275 __ tbnz(r15_elsize, 1, L_copy_ints);
2276 __ tbnz(r15_elsize, 0, L_copy_shorts);
2277 __ lea(from, Address(src, src_pos));// src_addr
2278 __ lea(to, Address(dst, dst_pos));// dst_addr
2279 __ movw(count, scratch_length); // length
2280 __ b(RuntimeAddress(byte_copy_entry));
2281
2282 __ BIND(L_copy_shorts);
2283 __ lea(from, Address(src, src_pos, Address::lsl(1)));// src_addr
2284 __ lea(to, Address(dst, dst_pos, Address::lsl(1)));// dst_addr
2285 __ movw(count, scratch_length); // length
2286 __ b(RuntimeAddress(short_copy_entry));
2287
2288 __ BIND(L_copy_ints);
2289 __ tbnz(r15_elsize, 0, L_copy_longs);
2290 __ lea(from, Address(src, src_pos, Address::lsl(2)));// src_addr
2291 __ lea(to, Address(dst, dst_pos, Address::lsl(2)));// dst_addr
2292 __ movw(count, scratch_length); // length
2293 __ b(RuntimeAddress(int_copy_entry));
2294
2295 __ BIND(L_copy_longs);
2296 #ifdef ASSERT
2297 {
2298 BLOCK_COMMENT("assert long copy {");
2299 Label L;
2300 __ andw(lh, lh, Klass::_lh_log2_element_size_mask); // lh -> r15_elsize
2301 __ cmpw(r15_elsize, LogBytesPerLong);
2302 __ br(Assembler::EQ, L);
2303 __ stop("must be long copy, but elsize is wrong");
2304 __ bind(L);
2305 BLOCK_COMMENT("} assert long copy done");
2306 }
2307 #endif
2308 __ lea(from, Address(src, src_pos, Address::lsl(3)));// src_addr
2309 __ lea(to, Address(dst, dst_pos, Address::lsl(3)));// dst_addr
2310 __ movw(count, scratch_length); // length
2311 __ b(RuntimeAddress(long_copy_entry));
2312
2313 // ObjArrayKlass
2314 __ BIND(L_objArray);
2315 // live at this point: scratch_src_klass, scratch_length, src[_pos], dst[_pos]
2316
2317 Label L_plain_copy, L_checkcast_copy;
2318 // test array classes for subtyping
2319 __ load_klass(r15, dst);
2320 __ cmp(scratch_src_klass, r15); // usual case is exact equality
2321 __ br(Assembler::NE, L_checkcast_copy);
2322
2323 // Identically typed arrays can be copied without element-wise checks.
2324 arraycopy_range_checks(src, src_pos, dst, dst_pos, scratch_length,
2325 rscratch2, L_failed);
2326
2327 __ lea(from, Address(src, src_pos, Address::lsl(LogBytesPerHeapOop)));
2328 __ add(from, from, arrayOopDesc::base_offset_in_bytes(T_OBJECT));
2329 __ lea(to, Address(dst, dst_pos, Address::lsl(LogBytesPerHeapOop)));
2330 __ add(to, to, arrayOopDesc::base_offset_in_bytes(T_OBJECT));
2331 __ movw(count, scratch_length); // length
2332 __ BIND(L_plain_copy);
2333 __ b(RuntimeAddress(oop_copy_entry));
2334
2335 __ BIND(L_checkcast_copy);
2336 // live at this point: scratch_src_klass, scratch_length, r15 (dst_klass)
2337 {
2338 // Before looking at dst.length, make sure dst is also an objArray.
2339 __ ldrw(rscratch1, Address(r15, lh_offset));
2340 __ movw(rscratch2, objArray_lh);
2341 __ eorw(rscratch1, rscratch1, rscratch2);
2342 __ cbnzw(rscratch1, L_failed);
2343
2344 // It is safe to examine both src.length and dst.length.
2345 arraycopy_range_checks(src, src_pos, dst, dst_pos, scratch_length,
2346 r15, L_failed);
2347
2348 __ load_klass(dst_klass, dst); // reload
2349
2350 // Marshal the base address arguments now, freeing registers.
2351 __ lea(from, Address(src, src_pos, Address::lsl(LogBytesPerHeapOop)));
2352 __ add(from, from, arrayOopDesc::base_offset_in_bytes(T_OBJECT));
2353 __ lea(to, Address(dst, dst_pos, Address::lsl(LogBytesPerHeapOop)));
2354 __ add(to, to, arrayOopDesc::base_offset_in_bytes(T_OBJECT));
2355 __ movw(count, length); // length (reloaded)
2356 Register sco_temp = c_rarg3; // this register is free now
2357 assert_different_registers(from, to, count, sco_temp,
2358 dst_klass, scratch_src_klass);
2359 // assert_clean_int(count, sco_temp);
2360
2361 // Generate the type check.
2362 const int sco_offset = in_bytes(Klass::super_check_offset_offset());
2363 __ ldrw(sco_temp, Address(dst_klass, sco_offset));
2364
2365 // Smashes rscratch1, rscratch2
2366 generate_type_check(scratch_src_klass, sco_temp, dst_klass, /*temps*/ noreg, noreg, noreg,
2367 L_plain_copy);
2368
2369 // Fetch destination element klass from the ObjArrayKlass header.
2370 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
2371 __ ldr(dst_klass, Address(dst_klass, ek_offset));
2372 __ ldrw(sco_temp, Address(dst_klass, sco_offset));
2373
2374 // the checkcast_copy loop needs two extra arguments:
2375 assert(c_rarg3 == sco_temp, "#3 already in place");
2376 // Set up arguments for checkcast_copy_entry.
2377 __ mov(c_rarg4, dst_klass); // dst.klass.element_klass
2378 __ b(RuntimeAddress(checkcast_copy_entry));
2379 }
2380
2381 __ BIND(L_failed);
2382 __ mov(r0, -1);
2383 __ leave(); // required for proper stackwalking of RuntimeStub frame
2384 __ ret(lr);
2385
2386 return start;
2387 }
2388
2389 //
2390 // Generate stub for array fill. If "aligned" is true, the
2391 // "to" address is assumed to be heapword aligned.
2392 //
2393 // Arguments for generated stub:
2394 // to: c_rarg0
2395 // value: c_rarg1
2396 // count: c_rarg2 treated as signed
2397 //
2398 address generate_fill(StubGenStubId stub_id) {
2399 BasicType t;
2400 bool aligned;
2401
2402 switch (stub_id) {
2403 case jbyte_fill_id:
2404 t = T_BYTE;
2405 aligned = false;
2406 break;
2407 case jshort_fill_id:
2408 t = T_SHORT;
2409 aligned = false;
2410 break;
2411 case jint_fill_id:
2412 t = T_INT;
2413 aligned = false;
2414 break;
2415 case arrayof_jbyte_fill_id:
2416 t = T_BYTE;
2417 aligned = true;
2418 break;
2419 case arrayof_jshort_fill_id:
2420 t = T_SHORT;
2421 aligned = true;
2422 break;
2423 case arrayof_jint_fill_id:
2424 t = T_INT;
2425 aligned = true;
2426 break;
2427 default:
2428 ShouldNotReachHere();
2429 };
2430
2431 __ align(CodeEntryAlignment);
2432 StubCodeMark mark(this, stub_id);
2433 address start = __ pc();
2434
2435 BLOCK_COMMENT("Entry:");
2436
2437 const Register to = c_rarg0; // source array address
2438 const Register value = c_rarg1; // value
2439 const Register count = c_rarg2; // elements count
2440
2441 const Register bz_base = r10; // base for block_zero routine
2442 const Register cnt_words = r11; // temp register
2443
2444 __ enter();
2445
2446 Label L_fill_elements, L_exit1;
2447
2448 int shift = -1;
2449 switch (t) {
2450 case T_BYTE:
2451 shift = 0;
2452 __ cmpw(count, 8 >> shift); // Short arrays (< 8 bytes) fill by element
2453 __ bfi(value, value, 8, 8); // 8 bit -> 16 bit
2454 __ bfi(value, value, 16, 16); // 16 bit -> 32 bit
2455 __ br(Assembler::LO, L_fill_elements);
2456 break;
2457 case T_SHORT:
2458 shift = 1;
2459 __ cmpw(count, 8 >> shift); // Short arrays (< 8 bytes) fill by element
2460 __ bfi(value, value, 16, 16); // 16 bit -> 32 bit
2461 __ br(Assembler::LO, L_fill_elements);
2462 break;
2463 case T_INT:
2464 shift = 2;
2465 __ cmpw(count, 8 >> shift); // Short arrays (< 8 bytes) fill by element
2466 __ br(Assembler::LO, L_fill_elements);
2467 break;
2468 default: ShouldNotReachHere();
2469 }
2470
2471 // Align source address at 8 bytes address boundary.
2472 Label L_skip_align1, L_skip_align2, L_skip_align4;
2473 if (!aligned) {
2474 switch (t) {
2475 case T_BYTE:
2476 // One byte misalignment happens only for byte arrays.
2477 __ tbz(to, 0, L_skip_align1);
2478 __ strb(value, Address(__ post(to, 1)));
2479 __ subw(count, count, 1);
2480 __ bind(L_skip_align1);
2481 // Fallthrough
2482 case T_SHORT:
2483 // Two bytes misalignment happens only for byte and short (char) arrays.
2484 __ tbz(to, 1, L_skip_align2);
2485 __ strh(value, Address(__ post(to, 2)));
2486 __ subw(count, count, 2 >> shift);
2487 __ bind(L_skip_align2);
2488 // Fallthrough
2489 case T_INT:
2490 // Align to 8 bytes, we know we are 4 byte aligned to start.
2491 __ tbz(to, 2, L_skip_align4);
2492 __ strw(value, Address(__ post(to, 4)));
2493 __ subw(count, count, 4 >> shift);
2494 __ bind(L_skip_align4);
2495 break;
2496 default: ShouldNotReachHere();
2497 }
2498 }
2499
2500 //
2501 // Fill large chunks
2502 //
2503 __ lsrw(cnt_words, count, 3 - shift); // number of words
2504 __ bfi(value, value, 32, 32); // 32 bit -> 64 bit
2505 __ subw(count, count, cnt_words, Assembler::LSL, 3 - shift);
2506 if (UseBlockZeroing) {
2507 Label non_block_zeroing, rest;
2508 // If the fill value is zero we can use the fast zero_words().
2509 __ cbnz(value, non_block_zeroing);
2510 __ mov(bz_base, to);
2511 __ add(to, to, cnt_words, Assembler::LSL, LogBytesPerWord);
2512 address tpc = __ zero_words(bz_base, cnt_words);
2513 if (tpc == nullptr) {
2514 fatal("CodeCache is full at generate_fill");
2515 }
2516 __ b(rest);
2517 __ bind(non_block_zeroing);
2518 __ fill_words(to, cnt_words, value);
2519 __ bind(rest);
2520 } else {
2521 __ fill_words(to, cnt_words, value);
2522 }
2523
2524 // Remaining count is less than 8 bytes. Fill it by a single store.
2525 // Note that the total length is no less than 8 bytes.
2526 if (t == T_BYTE || t == T_SHORT) {
2527 Label L_exit1;
2528 __ cbzw(count, L_exit1);
2529 __ add(to, to, count, Assembler::LSL, shift); // points to the end
2530 __ str(value, Address(to, -8)); // overwrite some elements
2531 __ bind(L_exit1);
2532 __ leave();
2533 __ ret(lr);
2534 }
2535
2536 // Handle copies less than 8 bytes.
2537 Label L_fill_2, L_fill_4, L_exit2;
2538 __ bind(L_fill_elements);
2539 switch (t) {
2540 case T_BYTE:
2541 __ tbz(count, 0, L_fill_2);
2542 __ strb(value, Address(__ post(to, 1)));
2543 __ bind(L_fill_2);
2544 __ tbz(count, 1, L_fill_4);
2545 __ strh(value, Address(__ post(to, 2)));
2546 __ bind(L_fill_4);
2547 __ tbz(count, 2, L_exit2);
2548 __ strw(value, Address(to));
2549 break;
2550 case T_SHORT:
2551 __ tbz(count, 0, L_fill_4);
2552 __ strh(value, Address(__ post(to, 2)));
2553 __ bind(L_fill_4);
2554 __ tbz(count, 1, L_exit2);
2555 __ strw(value, Address(to));
2556 break;
2557 case T_INT:
2558 __ cbzw(count, L_exit2);
2559 __ strw(value, Address(to));
2560 break;
2561 default: ShouldNotReachHere();
2562 }
2563 __ bind(L_exit2);
2564 __ leave();
2565 __ ret(lr);
2566 return start;
2567 }
2568
2569 address generate_data_cache_writeback() {
2570 const Register line = c_rarg0; // address of line to write back
2571
2572 __ align(CodeEntryAlignment);
2573
2574 StubGenStubId stub_id = StubGenStubId::data_cache_writeback_id;
2575 StubCodeMark mark(this, stub_id);
2576
2577 address start = __ pc();
2578 __ enter();
2579 __ cache_wb(Address(line, 0));
2580 __ leave();
2581 __ ret(lr);
2582
2583 return start;
2584 }
2585
2586 address generate_data_cache_writeback_sync() {
2587 const Register is_pre = c_rarg0; // pre or post sync
2588
2589 __ align(CodeEntryAlignment);
2590
2591 StubGenStubId stub_id = StubGenStubId::data_cache_writeback_sync_id;
2592 StubCodeMark mark(this, stub_id);
2593
2594 // pre wbsync is a no-op
2595 // post wbsync translates to an sfence
2596
2597 Label skip;
2598 address start = __ pc();
2599 __ enter();
2600 __ cbnz(is_pre, skip);
2601 __ cache_wbsync(false);
2602 __ bind(skip);
2603 __ leave();
2604 __ ret(lr);
2605
2606 return start;
2607 }
2608
2609 void generate_arraycopy_stubs() {
2610 address entry;
2611 address entry_jbyte_arraycopy;
2612 address entry_jshort_arraycopy;
2613 address entry_jint_arraycopy;
2614 address entry_oop_arraycopy;
2615 address entry_jlong_arraycopy;
2616 address entry_checkcast_arraycopy;
2617
2618 generate_copy_longs(StubGenStubId::copy_byte_f_id, IN_HEAP | IS_ARRAY, copy_f, r0, r1, r15);
2619 generate_copy_longs(StubGenStubId::copy_byte_b_id, IN_HEAP | IS_ARRAY, copy_b, r0, r1, r15);
2620
2621 generate_copy_longs(StubGenStubId::copy_oop_f_id, IN_HEAP | IS_ARRAY, copy_obj_f, r0, r1, r15);
2622 generate_copy_longs(StubGenStubId::copy_oop_b_id, IN_HEAP | IS_ARRAY, copy_obj_b, r0, r1, r15);
2623
2624 generate_copy_longs(StubGenStubId::copy_oop_uninit_f_id, IN_HEAP | IS_ARRAY | IS_DEST_UNINITIALIZED, copy_obj_uninit_f, r0, r1, r15);
2625 generate_copy_longs(StubGenStubId::copy_oop_uninit_b_id, IN_HEAP | IS_ARRAY | IS_DEST_UNINITIALIZED, copy_obj_uninit_b, r0, r1, r15);
2626
2627 StubRoutines::aarch64::_zero_blocks = generate_zero_blocks();
2628
2629 //*** jbyte
2630 // Always need aligned and unaligned versions
2631 StubRoutines::_jbyte_disjoint_arraycopy = generate_disjoint_copy(StubGenStubId::jbyte_disjoint_arraycopy_id, &entry);
2632 StubRoutines::_jbyte_arraycopy = generate_conjoint_copy(StubGenStubId::jbyte_arraycopy_id, entry, &entry_jbyte_arraycopy);
2633 StubRoutines::_arrayof_jbyte_disjoint_arraycopy = generate_disjoint_copy(StubGenStubId::arrayof_jbyte_disjoint_arraycopy_id, &entry);
2634 StubRoutines::_arrayof_jbyte_arraycopy = generate_conjoint_copy(StubGenStubId::arrayof_jbyte_arraycopy_id, entry, nullptr);
2635
2636 //*** jshort
2637 // Always need aligned and unaligned versions
2638 StubRoutines::_jshort_disjoint_arraycopy = generate_disjoint_copy(StubGenStubId::jshort_disjoint_arraycopy_id, &entry);
2639 StubRoutines::_jshort_arraycopy = generate_conjoint_copy(StubGenStubId::jshort_arraycopy_id, entry, &entry_jshort_arraycopy);
2640 StubRoutines::_arrayof_jshort_disjoint_arraycopy = generate_disjoint_copy(StubGenStubId::arrayof_jshort_disjoint_arraycopy_id, &entry);
2641 StubRoutines::_arrayof_jshort_arraycopy = generate_conjoint_copy(StubGenStubId::arrayof_jshort_arraycopy_id, entry, nullptr);
2642
2643 //*** jint
2644 // Aligned versions
2645 StubRoutines::_arrayof_jint_disjoint_arraycopy = generate_disjoint_copy(StubGenStubId::arrayof_jint_disjoint_arraycopy_id, &entry);
2646 StubRoutines::_arrayof_jint_arraycopy = generate_conjoint_copy(StubGenStubId::arrayof_jint_arraycopy_id, entry, &entry_jint_arraycopy);
2647 // In 64 bit we need both aligned and unaligned versions of jint arraycopy.
2648 // entry_jint_arraycopy always points to the unaligned version
2649 StubRoutines::_jint_disjoint_arraycopy = generate_disjoint_copy(StubGenStubId::jint_disjoint_arraycopy_id, &entry);
2650 StubRoutines::_jint_arraycopy = generate_conjoint_copy(StubGenStubId::jint_arraycopy_id, entry, &entry_jint_arraycopy);
2651
2652 //*** jlong
2653 // It is always aligned
2654 StubRoutines::_arrayof_jlong_disjoint_arraycopy = generate_disjoint_copy(StubGenStubId::arrayof_jlong_disjoint_arraycopy_id, &entry);
2655 StubRoutines::_arrayof_jlong_arraycopy = generate_conjoint_copy(StubGenStubId::arrayof_jlong_arraycopy_id, entry, &entry_jlong_arraycopy);
2656 StubRoutines::_jlong_disjoint_arraycopy = StubRoutines::_arrayof_jlong_disjoint_arraycopy;
2657 StubRoutines::_jlong_arraycopy = StubRoutines::_arrayof_jlong_arraycopy;
2658
2659 //*** oops
2660 {
2661 // With compressed oops we need unaligned versions; notice that
2662 // we overwrite entry_oop_arraycopy.
2663 bool aligned = !UseCompressedOops;
2664
2665 StubRoutines::_arrayof_oop_disjoint_arraycopy
2666 = generate_disjoint_copy(StubGenStubId::arrayof_oop_disjoint_arraycopy_id, &entry);
2667 StubRoutines::_arrayof_oop_arraycopy
2668 = generate_conjoint_copy(StubGenStubId::arrayof_oop_arraycopy_id, entry, &entry_oop_arraycopy);
2669 // Aligned versions without pre-barriers
2670 StubRoutines::_arrayof_oop_disjoint_arraycopy_uninit
2671 = generate_disjoint_copy(StubGenStubId::arrayof_oop_disjoint_arraycopy_uninit_id, &entry);
2672 StubRoutines::_arrayof_oop_arraycopy_uninit
2673 = generate_conjoint_copy(StubGenStubId::arrayof_oop_arraycopy_uninit_id, entry, nullptr);
2674 }
2675
2676 StubRoutines::_oop_disjoint_arraycopy = StubRoutines::_arrayof_oop_disjoint_arraycopy;
2677 StubRoutines::_oop_arraycopy = StubRoutines::_arrayof_oop_arraycopy;
2678 StubRoutines::_oop_disjoint_arraycopy_uninit = StubRoutines::_arrayof_oop_disjoint_arraycopy_uninit;
2679 StubRoutines::_oop_arraycopy_uninit = StubRoutines::_arrayof_oop_arraycopy_uninit;
2680
2681 StubRoutines::_checkcast_arraycopy = generate_checkcast_copy(StubGenStubId::checkcast_arraycopy_id, &entry_checkcast_arraycopy);
2682 StubRoutines::_checkcast_arraycopy_uninit = generate_checkcast_copy(StubGenStubId::checkcast_arraycopy_uninit_id, nullptr);
2683
2684 StubRoutines::_unsafe_arraycopy = generate_unsafe_copy(entry_jbyte_arraycopy,
2685 entry_jshort_arraycopy,
2686 entry_jint_arraycopy,
2687 entry_jlong_arraycopy);
2688
2689 StubRoutines::_generic_arraycopy = generate_generic_copy(entry_jbyte_arraycopy,
2690 entry_jshort_arraycopy,
2691 entry_jint_arraycopy,
2692 entry_oop_arraycopy,
2693 entry_jlong_arraycopy,
2694 entry_checkcast_arraycopy);
2695
2696 StubRoutines::_jbyte_fill = generate_fill(StubGenStubId::jbyte_fill_id);
2697 StubRoutines::_jshort_fill = generate_fill(StubGenStubId::jshort_fill_id);
2698 StubRoutines::_jint_fill = generate_fill(StubGenStubId::jint_fill_id);
2699 StubRoutines::_arrayof_jbyte_fill = generate_fill(StubGenStubId::arrayof_jbyte_fill_id);
2700 StubRoutines::_arrayof_jshort_fill = generate_fill(StubGenStubId::arrayof_jshort_fill_id);
2701 StubRoutines::_arrayof_jint_fill = generate_fill(StubGenStubId::arrayof_jint_fill_id);
2702 }
2703
2704 void generate_math_stubs() { Unimplemented(); }
2705
2706 // Arguments:
2707 //
2708 // Inputs:
2709 // c_rarg0 - source byte array address
2710 // c_rarg1 - destination byte array address
2711 // c_rarg2 - K (key) in little endian int array
2712 //
2713 address generate_aescrypt_encryptBlock() {
2714 __ align(CodeEntryAlignment);
2715 StubGenStubId stub_id = StubGenStubId::aescrypt_encryptBlock_id;
2716 StubCodeMark mark(this, stub_id);
2717
2718 const Register from = c_rarg0; // source array address
2719 const Register to = c_rarg1; // destination array address
2720 const Register key = c_rarg2; // key array address
2721 const Register keylen = rscratch1;
2722
2723 address start = __ pc();
2724 __ enter();
2725
2726 __ ldrw(keylen, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
2727
2728 __ aesenc_loadkeys(key, keylen);
2729 __ aesecb_encrypt(from, to, keylen);
2730
2731 __ mov(r0, 0);
2732
2733 __ leave();
2734 __ ret(lr);
2735
2736 return start;
2737 }
2738
2739 // Arguments:
2740 //
2741 // Inputs:
2742 // c_rarg0 - source byte array address
2743 // c_rarg1 - destination byte array address
2744 // c_rarg2 - K (key) in little endian int array
2745 //
2746 address generate_aescrypt_decryptBlock() {
2747 assert(UseAES, "need AES cryptographic extension support");
2748 __ align(CodeEntryAlignment);
2749 StubGenStubId stub_id = StubGenStubId::aescrypt_decryptBlock_id;
2750 StubCodeMark mark(this, stub_id);
2751 Label L_doLast;
2752
2753 const Register from = c_rarg0; // source array address
2754 const Register to = c_rarg1; // destination array address
2755 const Register key = c_rarg2; // key array address
2756 const Register keylen = rscratch1;
2757
2758 address start = __ pc();
2759 __ enter(); // required for proper stackwalking of RuntimeStub frame
2760
2761 __ ldrw(keylen, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
2762
2763 __ aesecb_decrypt(from, to, key, keylen);
2764
2765 __ mov(r0, 0);
2766
2767 __ leave();
2768 __ ret(lr);
2769
2770 return start;
2771 }
2772
2773 // Arguments:
2774 //
2775 // Inputs:
2776 // c_rarg0 - source byte array address
2777 // c_rarg1 - destination byte array address
2778 // c_rarg2 - K (key) in little endian int array
2779 // c_rarg3 - r vector byte array address
2780 // c_rarg4 - input length
2781 //
2782 // Output:
2783 // x0 - input length
2784 //
2785 address generate_cipherBlockChaining_encryptAESCrypt() {
2786 assert(UseAES, "need AES cryptographic extension support");
2787 __ align(CodeEntryAlignment);
2788 StubGenStubId stub_id = StubGenStubId::cipherBlockChaining_encryptAESCrypt_id;
2789 StubCodeMark mark(this, stub_id);
2790
2791 Label L_loadkeys_44, L_loadkeys_52, L_aes_loop, L_rounds_44, L_rounds_52;
2792
2793 const Register from = c_rarg0; // source array address
2794 const Register to = c_rarg1; // destination array address
2795 const Register key = c_rarg2; // key array address
2796 const Register rvec = c_rarg3; // r byte array initialized from initvector array address
2797 // and left with the results of the last encryption block
2798 const Register len_reg = c_rarg4; // src len (must be multiple of blocksize 16)
2799 const Register keylen = rscratch1;
2800
2801 address start = __ pc();
2802
2803 __ enter();
2804
2805 __ movw(rscratch2, len_reg);
2806
2807 __ ldrw(keylen, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
2808
2809 __ ld1(v0, __ T16B, rvec);
2810
2811 __ cmpw(keylen, 52);
2812 __ br(Assembler::CC, L_loadkeys_44);
2813 __ br(Assembler::EQ, L_loadkeys_52);
2814
2815 __ ld1(v17, v18, __ T16B, __ post(key, 32));
2816 __ rev32(v17, __ T16B, v17);
2817 __ rev32(v18, __ T16B, v18);
2818 __ BIND(L_loadkeys_52);
2819 __ ld1(v19, v20, __ T16B, __ post(key, 32));
2820 __ rev32(v19, __ T16B, v19);
2821 __ rev32(v20, __ T16B, v20);
2822 __ BIND(L_loadkeys_44);
2823 __ ld1(v21, v22, v23, v24, __ T16B, __ post(key, 64));
2824 __ rev32(v21, __ T16B, v21);
2825 __ rev32(v22, __ T16B, v22);
2826 __ rev32(v23, __ T16B, v23);
2827 __ rev32(v24, __ T16B, v24);
2828 __ ld1(v25, v26, v27, v28, __ T16B, __ post(key, 64));
2829 __ rev32(v25, __ T16B, v25);
2830 __ rev32(v26, __ T16B, v26);
2831 __ rev32(v27, __ T16B, v27);
2832 __ rev32(v28, __ T16B, v28);
2833 __ ld1(v29, v30, v31, __ T16B, key);
2834 __ rev32(v29, __ T16B, v29);
2835 __ rev32(v30, __ T16B, v30);
2836 __ rev32(v31, __ T16B, v31);
2837
2838 __ BIND(L_aes_loop);
2839 __ ld1(v1, __ T16B, __ post(from, 16));
2840 __ eor(v0, __ T16B, v0, v1);
2841
2842 __ br(Assembler::CC, L_rounds_44);
2843 __ br(Assembler::EQ, L_rounds_52);
2844
2845 __ aese(v0, v17); __ aesmc(v0, v0);
2846 __ aese(v0, v18); __ aesmc(v0, v0);
2847 __ BIND(L_rounds_52);
2848 __ aese(v0, v19); __ aesmc(v0, v0);
2849 __ aese(v0, v20); __ aesmc(v0, v0);
2850 __ BIND(L_rounds_44);
2851 __ aese(v0, v21); __ aesmc(v0, v0);
2852 __ aese(v0, v22); __ aesmc(v0, v0);
2853 __ aese(v0, v23); __ aesmc(v0, v0);
2854 __ aese(v0, v24); __ aesmc(v0, v0);
2855 __ aese(v0, v25); __ aesmc(v0, v0);
2856 __ aese(v0, v26); __ aesmc(v0, v0);
2857 __ aese(v0, v27); __ aesmc(v0, v0);
2858 __ aese(v0, v28); __ aesmc(v0, v0);
2859 __ aese(v0, v29); __ aesmc(v0, v0);
2860 __ aese(v0, v30);
2861 __ eor(v0, __ T16B, v0, v31);
2862
2863 __ st1(v0, __ T16B, __ post(to, 16));
2864
2865 __ subw(len_reg, len_reg, 16);
2866 __ cbnzw(len_reg, L_aes_loop);
2867
2868 __ st1(v0, __ T16B, rvec);
2869
2870 __ mov(r0, rscratch2);
2871
2872 __ leave();
2873 __ ret(lr);
2874
2875 return start;
2876 }
2877
2878 // Arguments:
2879 //
2880 // Inputs:
2881 // c_rarg0 - source byte array address
2882 // c_rarg1 - destination byte array address
2883 // c_rarg2 - K (key) in little endian int array
2884 // c_rarg3 - r vector byte array address
2885 // c_rarg4 - input length
2886 //
2887 // Output:
2888 // r0 - input length
2889 //
2890 address generate_cipherBlockChaining_decryptAESCrypt() {
2891 assert(UseAES, "need AES cryptographic extension support");
2892 __ align(CodeEntryAlignment);
2893 StubGenStubId stub_id = StubGenStubId::cipherBlockChaining_decryptAESCrypt_id;
2894 StubCodeMark mark(this, stub_id);
2895
2896 Label L_loadkeys_44, L_loadkeys_52, L_aes_loop, L_rounds_44, L_rounds_52;
2897
2898 const Register from = c_rarg0; // source array address
2899 const Register to = c_rarg1; // destination array address
2900 const Register key = c_rarg2; // key array address
2901 const Register rvec = c_rarg3; // r byte array initialized from initvector array address
2902 // and left with the results of the last encryption block
2903 const Register len_reg = c_rarg4; // src len (must be multiple of blocksize 16)
2904 const Register keylen = rscratch1;
2905
2906 address start = __ pc();
2907
2908 __ enter();
2909
2910 __ movw(rscratch2, len_reg);
2911
2912 __ ldrw(keylen, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
2913
2914 __ ld1(v2, __ T16B, rvec);
2915
2916 __ ld1(v31, __ T16B, __ post(key, 16));
2917 __ rev32(v31, __ T16B, v31);
2918
2919 __ cmpw(keylen, 52);
2920 __ br(Assembler::CC, L_loadkeys_44);
2921 __ br(Assembler::EQ, L_loadkeys_52);
2922
2923 __ ld1(v17, v18, __ T16B, __ post(key, 32));
2924 __ rev32(v17, __ T16B, v17);
2925 __ rev32(v18, __ T16B, v18);
2926 __ BIND(L_loadkeys_52);
2927 __ ld1(v19, v20, __ T16B, __ post(key, 32));
2928 __ rev32(v19, __ T16B, v19);
2929 __ rev32(v20, __ T16B, v20);
2930 __ BIND(L_loadkeys_44);
2931 __ ld1(v21, v22, v23, v24, __ T16B, __ post(key, 64));
2932 __ rev32(v21, __ T16B, v21);
2933 __ rev32(v22, __ T16B, v22);
2934 __ rev32(v23, __ T16B, v23);
2935 __ rev32(v24, __ T16B, v24);
2936 __ ld1(v25, v26, v27, v28, __ T16B, __ post(key, 64));
2937 __ rev32(v25, __ T16B, v25);
2938 __ rev32(v26, __ T16B, v26);
2939 __ rev32(v27, __ T16B, v27);
2940 __ rev32(v28, __ T16B, v28);
2941 __ ld1(v29, v30, __ T16B, key);
2942 __ rev32(v29, __ T16B, v29);
2943 __ rev32(v30, __ T16B, v30);
2944
2945 __ BIND(L_aes_loop);
2946 __ ld1(v0, __ T16B, __ post(from, 16));
2947 __ orr(v1, __ T16B, v0, v0);
2948
2949 __ br(Assembler::CC, L_rounds_44);
2950 __ br(Assembler::EQ, L_rounds_52);
2951
2952 __ aesd(v0, v17); __ aesimc(v0, v0);
2953 __ aesd(v0, v18); __ aesimc(v0, v0);
2954 __ BIND(L_rounds_52);
2955 __ aesd(v0, v19); __ aesimc(v0, v0);
2956 __ aesd(v0, v20); __ aesimc(v0, v0);
2957 __ BIND(L_rounds_44);
2958 __ aesd(v0, v21); __ aesimc(v0, v0);
2959 __ aesd(v0, v22); __ aesimc(v0, v0);
2960 __ aesd(v0, v23); __ aesimc(v0, v0);
2961 __ aesd(v0, v24); __ aesimc(v0, v0);
2962 __ aesd(v0, v25); __ aesimc(v0, v0);
2963 __ aesd(v0, v26); __ aesimc(v0, v0);
2964 __ aesd(v0, v27); __ aesimc(v0, v0);
2965 __ aesd(v0, v28); __ aesimc(v0, v0);
2966 __ aesd(v0, v29); __ aesimc(v0, v0);
2967 __ aesd(v0, v30);
2968 __ eor(v0, __ T16B, v0, v31);
2969 __ eor(v0, __ T16B, v0, v2);
2970
2971 __ st1(v0, __ T16B, __ post(to, 16));
2972 __ orr(v2, __ T16B, v1, v1);
2973
2974 __ subw(len_reg, len_reg, 16);
2975 __ cbnzw(len_reg, L_aes_loop);
2976
2977 __ st1(v2, __ T16B, rvec);
2978
2979 __ mov(r0, rscratch2);
2980
2981 __ leave();
2982 __ ret(lr);
2983
2984 return start;
2985 }
2986
2987 // Big-endian 128-bit + 64-bit -> 128-bit addition.
2988 // Inputs: 128-bits. in is preserved.
2989 // The least-significant 64-bit word is in the upper dword of each vector.
2990 // inc (the 64-bit increment) is preserved. Its lower dword must be zero.
2991 // Output: result
2992 void be_add_128_64(FloatRegister result, FloatRegister in,
2993 FloatRegister inc, FloatRegister tmp) {
2994 assert_different_registers(result, tmp, inc);
2995
2996 __ addv(result, __ T2D, in, inc); // Add inc to the least-significant dword of
2997 // input
2998 __ cm(__ HI, tmp, __ T2D, inc, result);// Check for result overflowing
2999 __ ext(tmp, __ T16B, tmp, tmp, 0x08); // Swap LSD of comparison result to MSD and
3000 // MSD == 0 (must be!) to LSD
3001 __ subv(result, __ T2D, result, tmp); // Subtract -1 from MSD if there was an overflow
3002 }
3003
3004 // CTR AES crypt.
3005 // Arguments:
3006 //
3007 // Inputs:
3008 // c_rarg0 - source byte array address
3009 // c_rarg1 - destination byte array address
3010 // c_rarg2 - K (key) in little endian int array
3011 // c_rarg3 - counter vector byte array address
3012 // c_rarg4 - input length
3013 // c_rarg5 - saved encryptedCounter start
3014 // c_rarg6 - saved used length
3015 //
3016 // Output:
3017 // r0 - input length
3018 //
3019 address generate_counterMode_AESCrypt() {
3020 const Register in = c_rarg0;
3021 const Register out = c_rarg1;
3022 const Register key = c_rarg2;
3023 const Register counter = c_rarg3;
3024 const Register saved_len = c_rarg4, len = r10;
3025 const Register saved_encrypted_ctr = c_rarg5;
3026 const Register used_ptr = c_rarg6, used = r12;
3027
3028 const Register offset = r7;
3029 const Register keylen = r11;
3030
3031 const unsigned char block_size = 16;
3032 const int bulk_width = 4;
3033 // NB: bulk_width can be 4 or 8. 8 gives slightly faster
3034 // performance with larger data sizes, but it also means that the
3035 // fast path isn't used until you have at least 8 blocks, and up
3036 // to 127 bytes of data will be executed on the slow path. For
3037 // that reason, and also so as not to blow away too much icache, 4
3038 // blocks seems like a sensible compromise.
3039
3040 // Algorithm:
3041 //
3042 // if (len == 0) {
3043 // goto DONE;
3044 // }
3045 // int result = len;
3046 // do {
3047 // if (used >= blockSize) {
3048 // if (len >= bulk_width * blockSize) {
3049 // CTR_large_block();
3050 // if (len == 0)
3051 // goto DONE;
3052 // }
3053 // for (;;) {
3054 // 16ByteVector v0 = counter;
3055 // embeddedCipher.encryptBlock(v0, 0, encryptedCounter, 0);
3056 // used = 0;
3057 // if (len < blockSize)
3058 // break; /* goto NEXT */
3059 // 16ByteVector v1 = load16Bytes(in, offset);
3060 // v1 = v1 ^ encryptedCounter;
3061 // store16Bytes(out, offset);
3062 // used = blockSize;
3063 // offset += blockSize;
3064 // len -= blockSize;
3065 // if (len == 0)
3066 // goto DONE;
3067 // }
3068 // }
3069 // NEXT:
3070 // out[outOff++] = (byte)(in[inOff++] ^ encryptedCounter[used++]);
3071 // len--;
3072 // } while (len != 0);
3073 // DONE:
3074 // return result;
3075 //
3076 // CTR_large_block()
3077 // Wide bulk encryption of whole blocks.
3078
3079 __ align(CodeEntryAlignment);
3080 StubGenStubId stub_id = StubGenStubId::counterMode_AESCrypt_id;
3081 StubCodeMark mark(this, stub_id);
3082 const address start = __ pc();
3083 __ enter();
3084
3085 Label DONE, CTR_large_block, large_block_return;
3086 __ ldrw(used, Address(used_ptr));
3087 __ cbzw(saved_len, DONE);
3088
3089 __ mov(len, saved_len);
3090 __ mov(offset, 0);
3091
3092 // Compute #rounds for AES based on the length of the key array
3093 __ ldrw(keylen, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3094
3095 __ aesenc_loadkeys(key, keylen);
3096
3097 {
3098 Label L_CTR_loop, NEXT;
3099
3100 __ bind(L_CTR_loop);
3101
3102 __ cmp(used, block_size);
3103 __ br(__ LO, NEXT);
3104
3105 // Maybe we have a lot of data
3106 __ subsw(rscratch1, len, bulk_width * block_size);
3107 __ br(__ HS, CTR_large_block);
3108 __ BIND(large_block_return);
3109 __ cbzw(len, DONE);
3110
3111 // Setup the counter
3112 __ movi(v4, __ T4S, 0);
3113 __ movi(v5, __ T4S, 1);
3114 __ ins(v4, __ S, v5, 2, 2); // v4 contains { 0, 1 }
3115
3116 // 128-bit big-endian increment
3117 __ ld1(v0, __ T16B, counter);
3118 __ rev64(v16, __ T16B, v0);
3119 be_add_128_64(v16, v16, v4, /*tmp*/v5);
3120 __ rev64(v16, __ T16B, v16);
3121 __ st1(v16, __ T16B, counter);
3122 // Previous counter value is in v0
3123 // v4 contains { 0, 1 }
3124
3125 {
3126 // We have fewer than bulk_width blocks of data left. Encrypt
3127 // them one by one until there is less than a full block
3128 // remaining, being careful to save both the encrypted counter
3129 // and the counter.
3130
3131 Label inner_loop;
3132 __ bind(inner_loop);
3133 // Counter to encrypt is in v0
3134 __ aesecb_encrypt(noreg, noreg, keylen);
3135 __ st1(v0, __ T16B, saved_encrypted_ctr);
3136
3137 // Do we have a remaining full block?
3138
3139 __ mov(used, 0);
3140 __ cmp(len, block_size);
3141 __ br(__ LO, NEXT);
3142
3143 // Yes, we have a full block
3144 __ ldrq(v1, Address(in, offset));
3145 __ eor(v1, __ T16B, v1, v0);
3146 __ strq(v1, Address(out, offset));
3147 __ mov(used, block_size);
3148 __ add(offset, offset, block_size);
3149
3150 __ subw(len, len, block_size);
3151 __ cbzw(len, DONE);
3152
3153 // Increment the counter, store it back
3154 __ orr(v0, __ T16B, v16, v16);
3155 __ rev64(v16, __ T16B, v16);
3156 be_add_128_64(v16, v16, v4, /*tmp*/v5);
3157 __ rev64(v16, __ T16B, v16);
3158 __ st1(v16, __ T16B, counter); // Save the incremented counter back
3159
3160 __ b(inner_loop);
3161 }
3162
3163 __ BIND(NEXT);
3164
3165 // Encrypt a single byte, and loop.
3166 // We expect this to be a rare event.
3167 __ ldrb(rscratch1, Address(in, offset));
3168 __ ldrb(rscratch2, Address(saved_encrypted_ctr, used));
3169 __ eor(rscratch1, rscratch1, rscratch2);
3170 __ strb(rscratch1, Address(out, offset));
3171 __ add(offset, offset, 1);
3172 __ add(used, used, 1);
3173 __ subw(len, len,1);
3174 __ cbnzw(len, L_CTR_loop);
3175 }
3176
3177 __ bind(DONE);
3178 __ strw(used, Address(used_ptr));
3179 __ mov(r0, saved_len);
3180
3181 __ leave(); // required for proper stackwalking of RuntimeStub frame
3182 __ ret(lr);
3183
3184 // Bulk encryption
3185
3186 __ BIND (CTR_large_block);
3187 assert(bulk_width == 4 || bulk_width == 8, "must be");
3188
3189 if (bulk_width == 8) {
3190 __ sub(sp, sp, 4 * 16);
3191 __ st1(v12, v13, v14, v15, __ T16B, Address(sp));
3192 }
3193 __ sub(sp, sp, 4 * 16);
3194 __ st1(v8, v9, v10, v11, __ T16B, Address(sp));
3195 RegSet saved_regs = (RegSet::of(in, out, offset)
3196 + RegSet::of(saved_encrypted_ctr, used_ptr, len));
3197 __ push(saved_regs, sp);
3198 __ andr(len, len, -16 * bulk_width); // 8/4 encryptions, 16 bytes per encryption
3199 __ add(in, in, offset);
3200 __ add(out, out, offset);
3201
3202 // Keys should already be loaded into the correct registers
3203
3204 __ ld1(v0, __ T16B, counter); // v0 contains the first counter
3205 __ rev64(v16, __ T16B, v0); // v16 contains byte-reversed counter
3206
3207 // AES/CTR loop
3208 {
3209 Label L_CTR_loop;
3210 __ BIND(L_CTR_loop);
3211
3212 // Setup the counters
3213 __ movi(v8, __ T4S, 0);
3214 __ movi(v9, __ T4S, 1);
3215 __ ins(v8, __ S, v9, 2, 2); // v8 contains { 0, 1 }
3216
3217 for (int i = 0; i < bulk_width; i++) {
3218 FloatRegister v0_ofs = as_FloatRegister(v0->encoding() + i);
3219 __ rev64(v0_ofs, __ T16B, v16);
3220 be_add_128_64(v16, v16, v8, /*tmp*/v9);
3221 }
3222
3223 __ ld1(v8, v9, v10, v11, __ T16B, __ post(in, 4 * 16));
3224
3225 // Encrypt the counters
3226 __ aesecb_encrypt(noreg, noreg, keylen, v0, bulk_width);
3227
3228 if (bulk_width == 8) {
3229 __ ld1(v12, v13, v14, v15, __ T16B, __ post(in, 4 * 16));
3230 }
3231
3232 // XOR the encrypted counters with the inputs
3233 for (int i = 0; i < bulk_width; i++) {
3234 FloatRegister v0_ofs = as_FloatRegister(v0->encoding() + i);
3235 FloatRegister v8_ofs = as_FloatRegister(v8->encoding() + i);
3236 __ eor(v0_ofs, __ T16B, v0_ofs, v8_ofs);
3237 }
3238
3239 // Write the encrypted data
3240 __ st1(v0, v1, v2, v3, __ T16B, __ post(out, 4 * 16));
3241 if (bulk_width == 8) {
3242 __ st1(v4, v5, v6, v7, __ T16B, __ post(out, 4 * 16));
3243 }
3244
3245 __ subw(len, len, 16 * bulk_width);
3246 __ cbnzw(len, L_CTR_loop);
3247 }
3248
3249 // Save the counter back where it goes
3250 __ rev64(v16, __ T16B, v16);
3251 __ st1(v16, __ T16B, counter);
3252
3253 __ pop(saved_regs, sp);
3254
3255 __ ld1(v8, v9, v10, v11, __ T16B, __ post(sp, 4 * 16));
3256 if (bulk_width == 8) {
3257 __ ld1(v12, v13, v14, v15, __ T16B, __ post(sp, 4 * 16));
3258 }
3259
3260 __ andr(rscratch1, len, -16 * bulk_width);
3261 __ sub(len, len, rscratch1);
3262 __ add(offset, offset, rscratch1);
3263 __ mov(used, 16);
3264 __ strw(used, Address(used_ptr));
3265 __ b(large_block_return);
3266
3267 return start;
3268 }
3269
3270 // Vector AES Galois Counter Mode implementation. Parameters:
3271 //
3272 // in = c_rarg0
3273 // len = c_rarg1
3274 // ct = c_rarg2 - ciphertext that ghash will read (in for encrypt, out for decrypt)
3275 // out = c_rarg3
3276 // key = c_rarg4
3277 // state = c_rarg5 - GHASH.state
3278 // subkeyHtbl = c_rarg6 - powers of H
3279 // counter = c_rarg7 - 16 bytes of CTR
3280 // return - number of processed bytes
3281 address generate_galoisCounterMode_AESCrypt() {
3282 address ghash_polynomial = __ pc();
3283 __ emit_int64(0x87); // The low-order bits of the field
3284 // polynomial (i.e. p = z^7+z^2+z+1)
3285 // repeated in the low and high parts of a
3286 // 128-bit vector
3287 __ emit_int64(0x87);
3288
3289 __ align(CodeEntryAlignment);
3290 StubGenStubId stub_id = StubGenStubId::galoisCounterMode_AESCrypt_id;
3291 StubCodeMark mark(this, stub_id);
3292 address start = __ pc();
3293 __ enter();
3294
3295 const Register in = c_rarg0;
3296 const Register len = c_rarg1;
3297 const Register ct = c_rarg2;
3298 const Register out = c_rarg3;
3299 // and updated with the incremented counter in the end
3300
3301 const Register key = c_rarg4;
3302 const Register state = c_rarg5;
3303
3304 const Register subkeyHtbl = c_rarg6;
3305
3306 const Register counter = c_rarg7;
3307
3308 const Register keylen = r10;
3309 // Save state before entering routine
3310 __ sub(sp, sp, 4 * 16);
3311 __ st1(v12, v13, v14, v15, __ T16B, Address(sp));
3312 __ sub(sp, sp, 4 * 16);
3313 __ st1(v8, v9, v10, v11, __ T16B, Address(sp));
3314
3315 // __ andr(len, len, -512);
3316 __ andr(len, len, -16 * 8); // 8 encryptions, 16 bytes per encryption
3317 __ str(len, __ pre(sp, -2 * wordSize));
3318
3319 Label DONE;
3320 __ cbz(len, DONE);
3321
3322 // Compute #rounds for AES based on the length of the key array
3323 __ ldrw(keylen, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3324
3325 __ aesenc_loadkeys(key, keylen);
3326 __ ld1(v0, __ T16B, counter); // v0 contains the first counter
3327 __ rev32(v16, __ T16B, v0); // v16 contains byte-reversed counter
3328
3329 // AES/CTR loop
3330 {
3331 Label L_CTR_loop;
3332 __ BIND(L_CTR_loop);
3333
3334 // Setup the counters
3335 __ movi(v8, __ T4S, 0);
3336 __ movi(v9, __ T4S, 1);
3337 __ ins(v8, __ S, v9, 3, 3); // v8 contains { 0, 0, 0, 1 }
3338
3339 assert(v0->encoding() < v8->encoding(), "");
3340 for (int i = v0->encoding(); i < v8->encoding(); i++) {
3341 FloatRegister f = as_FloatRegister(i);
3342 __ rev32(f, __ T16B, v16);
3343 __ addv(v16, __ T4S, v16, v8);
3344 }
3345
3346 __ ld1(v8, v9, v10, v11, __ T16B, __ post(in, 4 * 16));
3347
3348 // Encrypt the counters
3349 __ aesecb_encrypt(noreg, noreg, keylen, v0, /*unrolls*/8);
3350
3351 __ ld1(v12, v13, v14, v15, __ T16B, __ post(in, 4 * 16));
3352
3353 // XOR the encrypted counters with the inputs
3354 for (int i = 0; i < 8; i++) {
3355 FloatRegister v0_ofs = as_FloatRegister(v0->encoding() + i);
3356 FloatRegister v8_ofs = as_FloatRegister(v8->encoding() + i);
3357 __ eor(v0_ofs, __ T16B, v0_ofs, v8_ofs);
3358 }
3359 __ st1(v0, v1, v2, v3, __ T16B, __ post(out, 4 * 16));
3360 __ st1(v4, v5, v6, v7, __ T16B, __ post(out, 4 * 16));
3361
3362 __ subw(len, len, 16 * 8);
3363 __ cbnzw(len, L_CTR_loop);
3364 }
3365
3366 __ rev32(v16, __ T16B, v16);
3367 __ st1(v16, __ T16B, counter);
3368
3369 __ ldr(len, Address(sp));
3370 __ lsr(len, len, exact_log2(16)); // We want the count of blocks
3371
3372 // GHASH/CTR loop
3373 __ ghash_processBlocks_wide(ghash_polynomial, state, subkeyHtbl, ct,
3374 len, /*unrolls*/4);
3375
3376 #ifdef ASSERT
3377 { Label L;
3378 __ cmp(len, (unsigned char)0);
3379 __ br(Assembler::EQ, L);
3380 __ stop("stubGenerator: abort");
3381 __ bind(L);
3382 }
3383 #endif
3384
3385 __ bind(DONE);
3386 // Return the number of bytes processed
3387 __ ldr(r0, __ post(sp, 2 * wordSize));
3388
3389 __ ld1(v8, v9, v10, v11, __ T16B, __ post(sp, 4 * 16));
3390 __ ld1(v12, v13, v14, v15, __ T16B, __ post(sp, 4 * 16));
3391
3392 __ leave(); // required for proper stackwalking of RuntimeStub frame
3393 __ ret(lr);
3394 return start;
3395 }
3396
3397 class Cached64Bytes {
3398 private:
3399 MacroAssembler *_masm;
3400 Register _regs[8];
3401
3402 public:
3403 Cached64Bytes(MacroAssembler *masm, RegSet rs): _masm(masm) {
3404 assert(rs.size() == 8, "%u registers are used to cache 16 4-byte data", rs.size());
3405 auto it = rs.begin();
3406 for (auto &r: _regs) {
3407 r = *it;
3408 ++it;
3409 }
3410 }
3411
3412 void gen_loads(Register base) {
3413 for (int i = 0; i < 8; i += 2) {
3414 __ ldp(_regs[i], _regs[i + 1], Address(base, 8 * i));
3415 }
3416 }
3417
3418 // Generate code extracting i-th unsigned word (4 bytes) from cached 64 bytes.
3419 void extract_u32(Register dest, int i) {
3420 __ ubfx(dest, _regs[i / 2], 32 * (i % 2), 32);
3421 }
3422 };
3423
3424 // Utility routines for md5.
3425 // Clobbers r10 and r11.
3426 void md5_FF(Cached64Bytes& reg_cache, Register r1, Register r2, Register r3, Register r4,
3427 int k, int s, int t) {
3428 Register rscratch3 = r10;
3429 Register rscratch4 = r11;
3430
3431 __ eorw(rscratch3, r3, r4);
3432 __ movw(rscratch2, t);
3433 __ andw(rscratch3, rscratch3, r2);
3434 __ addw(rscratch4, r1, rscratch2);
3435 reg_cache.extract_u32(rscratch1, k);
3436 __ eorw(rscratch3, rscratch3, r4);
3437 __ addw(rscratch4, rscratch4, rscratch1);
3438 __ addw(rscratch3, rscratch3, rscratch4);
3439 __ rorw(rscratch2, rscratch3, 32 - s);
3440 __ addw(r1, rscratch2, r2);
3441 }
3442
3443 void md5_GG(Cached64Bytes& reg_cache, Register r1, Register r2, Register r3, Register r4,
3444 int k, int s, int t) {
3445 Register rscratch3 = r10;
3446 Register rscratch4 = r11;
3447
3448 reg_cache.extract_u32(rscratch1, k);
3449 __ movw(rscratch2, t);
3450 __ addw(rscratch4, r1, rscratch2);
3451 __ addw(rscratch4, rscratch4, rscratch1);
3452 __ bicw(rscratch2, r3, r4);
3453 __ andw(rscratch3, r2, r4);
3454 __ addw(rscratch2, rscratch2, rscratch4);
3455 __ addw(rscratch2, rscratch2, rscratch3);
3456 __ rorw(rscratch2, rscratch2, 32 - s);
3457 __ addw(r1, rscratch2, r2);
3458 }
3459
3460 void md5_HH(Cached64Bytes& reg_cache, Register r1, Register r2, Register r3, Register r4,
3461 int k, int s, int t) {
3462 Register rscratch3 = r10;
3463 Register rscratch4 = r11;
3464
3465 __ eorw(rscratch3, r3, r4);
3466 __ movw(rscratch2, t);
3467 __ addw(rscratch4, r1, rscratch2);
3468 reg_cache.extract_u32(rscratch1, k);
3469 __ eorw(rscratch3, rscratch3, r2);
3470 __ addw(rscratch4, rscratch4, rscratch1);
3471 __ addw(rscratch3, rscratch3, rscratch4);
3472 __ rorw(rscratch2, rscratch3, 32 - s);
3473 __ addw(r1, rscratch2, r2);
3474 }
3475
3476 void md5_II(Cached64Bytes& reg_cache, Register r1, Register r2, Register r3, Register r4,
3477 int k, int s, int t) {
3478 Register rscratch3 = r10;
3479 Register rscratch4 = r11;
3480
3481 __ movw(rscratch3, t);
3482 __ ornw(rscratch2, r2, r4);
3483 __ addw(rscratch4, r1, rscratch3);
3484 reg_cache.extract_u32(rscratch1, k);
3485 __ eorw(rscratch3, rscratch2, r3);
3486 __ addw(rscratch4, rscratch4, rscratch1);
3487 __ addw(rscratch3, rscratch3, rscratch4);
3488 __ rorw(rscratch2, rscratch3, 32 - s);
3489 __ addw(r1, rscratch2, r2);
3490 }
3491
3492 // Arguments:
3493 //
3494 // Inputs:
3495 // c_rarg0 - byte[] source+offset
3496 // c_rarg1 - int[] SHA.state
3497 // c_rarg2 - int offset
3498 // c_rarg3 - int limit
3499 //
3500 address generate_md5_implCompress(StubGenStubId stub_id) {
3501 bool multi_block;
3502 switch (stub_id) {
3503 case md5_implCompress_id:
3504 multi_block = false;
3505 break;
3506 case md5_implCompressMB_id:
3507 multi_block = true;
3508 break;
3509 default:
3510 ShouldNotReachHere();
3511 }
3512 __ align(CodeEntryAlignment);
3513
3514 StubCodeMark mark(this, stub_id);
3515 address start = __ pc();
3516
3517 Register buf = c_rarg0;
3518 Register state = c_rarg1;
3519 Register ofs = c_rarg2;
3520 Register limit = c_rarg3;
3521 Register a = r4;
3522 Register b = r5;
3523 Register c = r6;
3524 Register d = r7;
3525 Register rscratch3 = r10;
3526 Register rscratch4 = r11;
3527
3528 Register state_regs[2] = { r12, r13 };
3529 RegSet saved_regs = RegSet::range(r16, r22) - r18_tls;
3530 Cached64Bytes reg_cache(_masm, RegSet::of(r14, r15) + saved_regs); // using 8 registers
3531
3532 __ push(saved_regs, sp);
3533
3534 __ ldp(state_regs[0], state_regs[1], Address(state));
3535 __ ubfx(a, state_regs[0], 0, 32);
3536 __ ubfx(b, state_regs[0], 32, 32);
3537 __ ubfx(c, state_regs[1], 0, 32);
3538 __ ubfx(d, state_regs[1], 32, 32);
3539
3540 Label md5_loop;
3541 __ BIND(md5_loop);
3542
3543 reg_cache.gen_loads(buf);
3544
3545 // Round 1
3546 md5_FF(reg_cache, a, b, c, d, 0, 7, 0xd76aa478);
3547 md5_FF(reg_cache, d, a, b, c, 1, 12, 0xe8c7b756);
3548 md5_FF(reg_cache, c, d, a, b, 2, 17, 0x242070db);
3549 md5_FF(reg_cache, b, c, d, a, 3, 22, 0xc1bdceee);
3550 md5_FF(reg_cache, a, b, c, d, 4, 7, 0xf57c0faf);
3551 md5_FF(reg_cache, d, a, b, c, 5, 12, 0x4787c62a);
3552 md5_FF(reg_cache, c, d, a, b, 6, 17, 0xa8304613);
3553 md5_FF(reg_cache, b, c, d, a, 7, 22, 0xfd469501);
3554 md5_FF(reg_cache, a, b, c, d, 8, 7, 0x698098d8);
3555 md5_FF(reg_cache, d, a, b, c, 9, 12, 0x8b44f7af);
3556 md5_FF(reg_cache, c, d, a, b, 10, 17, 0xffff5bb1);
3557 md5_FF(reg_cache, b, c, d, a, 11, 22, 0x895cd7be);
3558 md5_FF(reg_cache, a, b, c, d, 12, 7, 0x6b901122);
3559 md5_FF(reg_cache, d, a, b, c, 13, 12, 0xfd987193);
3560 md5_FF(reg_cache, c, d, a, b, 14, 17, 0xa679438e);
3561 md5_FF(reg_cache, b, c, d, a, 15, 22, 0x49b40821);
3562
3563 // Round 2
3564 md5_GG(reg_cache, a, b, c, d, 1, 5, 0xf61e2562);
3565 md5_GG(reg_cache, d, a, b, c, 6, 9, 0xc040b340);
3566 md5_GG(reg_cache, c, d, a, b, 11, 14, 0x265e5a51);
3567 md5_GG(reg_cache, b, c, d, a, 0, 20, 0xe9b6c7aa);
3568 md5_GG(reg_cache, a, b, c, d, 5, 5, 0xd62f105d);
3569 md5_GG(reg_cache, d, a, b, c, 10, 9, 0x02441453);
3570 md5_GG(reg_cache, c, d, a, b, 15, 14, 0xd8a1e681);
3571 md5_GG(reg_cache, b, c, d, a, 4, 20, 0xe7d3fbc8);
3572 md5_GG(reg_cache, a, b, c, d, 9, 5, 0x21e1cde6);
3573 md5_GG(reg_cache, d, a, b, c, 14, 9, 0xc33707d6);
3574 md5_GG(reg_cache, c, d, a, b, 3, 14, 0xf4d50d87);
3575 md5_GG(reg_cache, b, c, d, a, 8, 20, 0x455a14ed);
3576 md5_GG(reg_cache, a, b, c, d, 13, 5, 0xa9e3e905);
3577 md5_GG(reg_cache, d, a, b, c, 2, 9, 0xfcefa3f8);
3578 md5_GG(reg_cache, c, d, a, b, 7, 14, 0x676f02d9);
3579 md5_GG(reg_cache, b, c, d, a, 12, 20, 0x8d2a4c8a);
3580
3581 // Round 3
3582 md5_HH(reg_cache, a, b, c, d, 5, 4, 0xfffa3942);
3583 md5_HH(reg_cache, d, a, b, c, 8, 11, 0x8771f681);
3584 md5_HH(reg_cache, c, d, a, b, 11, 16, 0x6d9d6122);
3585 md5_HH(reg_cache, b, c, d, a, 14, 23, 0xfde5380c);
3586 md5_HH(reg_cache, a, b, c, d, 1, 4, 0xa4beea44);
3587 md5_HH(reg_cache, d, a, b, c, 4, 11, 0x4bdecfa9);
3588 md5_HH(reg_cache, c, d, a, b, 7, 16, 0xf6bb4b60);
3589 md5_HH(reg_cache, b, c, d, a, 10, 23, 0xbebfbc70);
3590 md5_HH(reg_cache, a, b, c, d, 13, 4, 0x289b7ec6);
3591 md5_HH(reg_cache, d, a, b, c, 0, 11, 0xeaa127fa);
3592 md5_HH(reg_cache, c, d, a, b, 3, 16, 0xd4ef3085);
3593 md5_HH(reg_cache, b, c, d, a, 6, 23, 0x04881d05);
3594 md5_HH(reg_cache, a, b, c, d, 9, 4, 0xd9d4d039);
3595 md5_HH(reg_cache, d, a, b, c, 12, 11, 0xe6db99e5);
3596 md5_HH(reg_cache, c, d, a, b, 15, 16, 0x1fa27cf8);
3597 md5_HH(reg_cache, b, c, d, a, 2, 23, 0xc4ac5665);
3598
3599 // Round 4
3600 md5_II(reg_cache, a, b, c, d, 0, 6, 0xf4292244);
3601 md5_II(reg_cache, d, a, b, c, 7, 10, 0x432aff97);
3602 md5_II(reg_cache, c, d, a, b, 14, 15, 0xab9423a7);
3603 md5_II(reg_cache, b, c, d, a, 5, 21, 0xfc93a039);
3604 md5_II(reg_cache, a, b, c, d, 12, 6, 0x655b59c3);
3605 md5_II(reg_cache, d, a, b, c, 3, 10, 0x8f0ccc92);
3606 md5_II(reg_cache, c, d, a, b, 10, 15, 0xffeff47d);
3607 md5_II(reg_cache, b, c, d, a, 1, 21, 0x85845dd1);
3608 md5_II(reg_cache, a, b, c, d, 8, 6, 0x6fa87e4f);
3609 md5_II(reg_cache, d, a, b, c, 15, 10, 0xfe2ce6e0);
3610 md5_II(reg_cache, c, d, a, b, 6, 15, 0xa3014314);
3611 md5_II(reg_cache, b, c, d, a, 13, 21, 0x4e0811a1);
3612 md5_II(reg_cache, a, b, c, d, 4, 6, 0xf7537e82);
3613 md5_II(reg_cache, d, a, b, c, 11, 10, 0xbd3af235);
3614 md5_II(reg_cache, c, d, a, b, 2, 15, 0x2ad7d2bb);
3615 md5_II(reg_cache, b, c, d, a, 9, 21, 0xeb86d391);
3616
3617 __ addw(a, state_regs[0], a);
3618 __ ubfx(rscratch2, state_regs[0], 32, 32);
3619 __ addw(b, rscratch2, b);
3620 __ addw(c, state_regs[1], c);
3621 __ ubfx(rscratch4, state_regs[1], 32, 32);
3622 __ addw(d, rscratch4, d);
3623
3624 __ orr(state_regs[0], a, b, Assembler::LSL, 32);
3625 __ orr(state_regs[1], c, d, Assembler::LSL, 32);
3626
3627 if (multi_block) {
3628 __ add(buf, buf, 64);
3629 __ add(ofs, ofs, 64);
3630 __ cmp(ofs, limit);
3631 __ br(Assembler::LE, md5_loop);
3632 __ mov(c_rarg0, ofs); // return ofs
3633 }
3634
3635 // write hash values back in the correct order
3636 __ stp(state_regs[0], state_regs[1], Address(state));
3637
3638 __ pop(saved_regs, sp);
3639
3640 __ ret(lr);
3641
3642 return start;
3643 }
3644
3645 // Arguments:
3646 //
3647 // Inputs:
3648 // c_rarg0 - byte[] source+offset
3649 // c_rarg1 - int[] SHA.state
3650 // c_rarg2 - int offset
3651 // c_rarg3 - int limit
3652 //
3653 address generate_sha1_implCompress(StubGenStubId stub_id) {
3654 bool multi_block;
3655 switch (stub_id) {
3656 case sha1_implCompress_id:
3657 multi_block = false;
3658 break;
3659 case sha1_implCompressMB_id:
3660 multi_block = true;
3661 break;
3662 default:
3663 ShouldNotReachHere();
3664 }
3665
3666 __ align(CodeEntryAlignment);
3667
3668 StubCodeMark mark(this, stub_id);
3669 address start = __ pc();
3670
3671 Register buf = c_rarg0;
3672 Register state = c_rarg1;
3673 Register ofs = c_rarg2;
3674 Register limit = c_rarg3;
3675
3676 Label keys;
3677 Label sha1_loop;
3678
3679 // load the keys into v0..v3
3680 __ adr(rscratch1, keys);
3681 __ ld4r(v0, v1, v2, v3, __ T4S, Address(rscratch1));
3682 // load 5 words state into v6, v7
3683 __ ldrq(v6, Address(state, 0));
3684 __ ldrs(v7, Address(state, 16));
3685
3686
3687 __ BIND(sha1_loop);
3688 // load 64 bytes of data into v16..v19
3689 __ ld1(v16, v17, v18, v19, __ T4S, multi_block ? __ post(buf, 64) : buf);
3690 __ rev32(v16, __ T16B, v16);
3691 __ rev32(v17, __ T16B, v17);
3692 __ rev32(v18, __ T16B, v18);
3693 __ rev32(v19, __ T16B, v19);
3694
3695 // do the sha1
3696 __ addv(v4, __ T4S, v16, v0);
3697 __ orr(v20, __ T16B, v6, v6);
3698
3699 FloatRegister d0 = v16;
3700 FloatRegister d1 = v17;
3701 FloatRegister d2 = v18;
3702 FloatRegister d3 = v19;
3703
3704 for (int round = 0; round < 20; round++) {
3705 FloatRegister tmp1 = (round & 1) ? v4 : v5;
3706 FloatRegister tmp2 = (round & 1) ? v21 : v22;
3707 FloatRegister tmp3 = round ? ((round & 1) ? v22 : v21) : v7;
3708 FloatRegister tmp4 = (round & 1) ? v5 : v4;
3709 FloatRegister key = (round < 4) ? v0 : ((round < 9) ? v1 : ((round < 14) ? v2 : v3));
3710
3711 if (round < 16) __ sha1su0(d0, __ T4S, d1, d2);
3712 if (round < 19) __ addv(tmp1, __ T4S, d1, key);
3713 __ sha1h(tmp2, __ T4S, v20);
3714 if (round < 5)
3715 __ sha1c(v20, __ T4S, tmp3, tmp4);
3716 else if (round < 10 || round >= 15)
3717 __ sha1p(v20, __ T4S, tmp3, tmp4);
3718 else
3719 __ sha1m(v20, __ T4S, tmp3, tmp4);
3720 if (round < 16) __ sha1su1(d0, __ T4S, d3);
3721
3722 tmp1 = d0; d0 = d1; d1 = d2; d2 = d3; d3 = tmp1;
3723 }
3724
3725 __ addv(v7, __ T2S, v7, v21);
3726 __ addv(v6, __ T4S, v6, v20);
3727
3728 if (multi_block) {
3729 __ add(ofs, ofs, 64);
3730 __ cmp(ofs, limit);
3731 __ br(Assembler::LE, sha1_loop);
3732 __ mov(c_rarg0, ofs); // return ofs
3733 }
3734
3735 __ strq(v6, Address(state, 0));
3736 __ strs(v7, Address(state, 16));
3737
3738 __ ret(lr);
3739
3740 __ bind(keys);
3741 __ emit_int32(0x5a827999);
3742 __ emit_int32(0x6ed9eba1);
3743 __ emit_int32(0x8f1bbcdc);
3744 __ emit_int32(0xca62c1d6);
3745
3746 return start;
3747 }
3748
3749
3750 // Arguments:
3751 //
3752 // Inputs:
3753 // c_rarg0 - byte[] source+offset
3754 // c_rarg1 - int[] SHA.state
3755 // c_rarg2 - int offset
3756 // c_rarg3 - int limit
3757 //
3758 address generate_sha256_implCompress(StubGenStubId stub_id) {
3759 bool multi_block;
3760 switch (stub_id) {
3761 case sha256_implCompress_id:
3762 multi_block = false;
3763 break;
3764 case sha256_implCompressMB_id:
3765 multi_block = true;
3766 break;
3767 default:
3768 ShouldNotReachHere();
3769 }
3770
3771 static const uint32_t round_consts[64] = {
3772 0x428a2f98, 0x71374491, 0xb5c0fbcf, 0xe9b5dba5,
3773 0x3956c25b, 0x59f111f1, 0x923f82a4, 0xab1c5ed5,
3774 0xd807aa98, 0x12835b01, 0x243185be, 0x550c7dc3,
3775 0x72be5d74, 0x80deb1fe, 0x9bdc06a7, 0xc19bf174,
3776 0xe49b69c1, 0xefbe4786, 0x0fc19dc6, 0x240ca1cc,
3777 0x2de92c6f, 0x4a7484aa, 0x5cb0a9dc, 0x76f988da,
3778 0x983e5152, 0xa831c66d, 0xb00327c8, 0xbf597fc7,
3779 0xc6e00bf3, 0xd5a79147, 0x06ca6351, 0x14292967,
3780 0x27b70a85, 0x2e1b2138, 0x4d2c6dfc, 0x53380d13,
3781 0x650a7354, 0x766a0abb, 0x81c2c92e, 0x92722c85,
3782 0xa2bfe8a1, 0xa81a664b, 0xc24b8b70, 0xc76c51a3,
3783 0xd192e819, 0xd6990624, 0xf40e3585, 0x106aa070,
3784 0x19a4c116, 0x1e376c08, 0x2748774c, 0x34b0bcb5,
3785 0x391c0cb3, 0x4ed8aa4a, 0x5b9cca4f, 0x682e6ff3,
3786 0x748f82ee, 0x78a5636f, 0x84c87814, 0x8cc70208,
3787 0x90befffa, 0xa4506ceb, 0xbef9a3f7, 0xc67178f2,
3788 };
3789
3790 __ align(CodeEntryAlignment);
3791
3792 StubCodeMark mark(this, stub_id);
3793 address start = __ pc();
3794
3795 Register buf = c_rarg0;
3796 Register state = c_rarg1;
3797 Register ofs = c_rarg2;
3798 Register limit = c_rarg3;
3799
3800 Label sha1_loop;
3801
3802 __ stpd(v8, v9, __ pre(sp, -32));
3803 __ stpd(v10, v11, Address(sp, 16));
3804
3805 // dga == v0
3806 // dgb == v1
3807 // dg0 == v2
3808 // dg1 == v3
3809 // dg2 == v4
3810 // t0 == v6
3811 // t1 == v7
3812
3813 // load 16 keys to v16..v31
3814 __ lea(rscratch1, ExternalAddress((address)round_consts));
3815 __ ld1(v16, v17, v18, v19, __ T4S, __ post(rscratch1, 64));
3816 __ ld1(v20, v21, v22, v23, __ T4S, __ post(rscratch1, 64));
3817 __ ld1(v24, v25, v26, v27, __ T4S, __ post(rscratch1, 64));
3818 __ ld1(v28, v29, v30, v31, __ T4S, rscratch1);
3819
3820 // load 8 words (256 bits) state
3821 __ ldpq(v0, v1, state);
3822
3823 __ BIND(sha1_loop);
3824 // load 64 bytes of data into v8..v11
3825 __ ld1(v8, v9, v10, v11, __ T4S, multi_block ? __ post(buf, 64) : buf);
3826 __ rev32(v8, __ T16B, v8);
3827 __ rev32(v9, __ T16B, v9);
3828 __ rev32(v10, __ T16B, v10);
3829 __ rev32(v11, __ T16B, v11);
3830
3831 __ addv(v6, __ T4S, v8, v16);
3832 __ orr(v2, __ T16B, v0, v0);
3833 __ orr(v3, __ T16B, v1, v1);
3834
3835 FloatRegister d0 = v8;
3836 FloatRegister d1 = v9;
3837 FloatRegister d2 = v10;
3838 FloatRegister d3 = v11;
3839
3840
3841 for (int round = 0; round < 16; round++) {
3842 FloatRegister tmp1 = (round & 1) ? v6 : v7;
3843 FloatRegister tmp2 = (round & 1) ? v7 : v6;
3844 FloatRegister tmp3 = (round & 1) ? v2 : v4;
3845 FloatRegister tmp4 = (round & 1) ? v4 : v2;
3846
3847 if (round < 12) __ sha256su0(d0, __ T4S, d1);
3848 __ orr(v4, __ T16B, v2, v2);
3849 if (round < 15)
3850 __ addv(tmp1, __ T4S, d1, as_FloatRegister(round + 17));
3851 __ sha256h(v2, __ T4S, v3, tmp2);
3852 __ sha256h2(v3, __ T4S, v4, tmp2);
3853 if (round < 12) __ sha256su1(d0, __ T4S, d2, d3);
3854
3855 tmp1 = d0; d0 = d1; d1 = d2; d2 = d3; d3 = tmp1;
3856 }
3857
3858 __ addv(v0, __ T4S, v0, v2);
3859 __ addv(v1, __ T4S, v1, v3);
3860
3861 if (multi_block) {
3862 __ add(ofs, ofs, 64);
3863 __ cmp(ofs, limit);
3864 __ br(Assembler::LE, sha1_loop);
3865 __ mov(c_rarg0, ofs); // return ofs
3866 }
3867
3868 __ ldpd(v10, v11, Address(sp, 16));
3869 __ ldpd(v8, v9, __ post(sp, 32));
3870
3871 __ stpq(v0, v1, state);
3872
3873 __ ret(lr);
3874
3875 return start;
3876 }
3877
3878 // Double rounds for sha512.
3879 void sha512_dround(int dr,
3880 FloatRegister vi0, FloatRegister vi1,
3881 FloatRegister vi2, FloatRegister vi3,
3882 FloatRegister vi4, FloatRegister vrc0,
3883 FloatRegister vrc1, FloatRegister vin0,
3884 FloatRegister vin1, FloatRegister vin2,
3885 FloatRegister vin3, FloatRegister vin4) {
3886 if (dr < 36) {
3887 __ ld1(vrc1, __ T2D, __ post(rscratch2, 16));
3888 }
3889 __ addv(v5, __ T2D, vrc0, vin0);
3890 __ ext(v6, __ T16B, vi2, vi3, 8);
3891 __ ext(v5, __ T16B, v5, v5, 8);
3892 __ ext(v7, __ T16B, vi1, vi2, 8);
3893 __ addv(vi3, __ T2D, vi3, v5);
3894 if (dr < 32) {
3895 __ ext(v5, __ T16B, vin3, vin4, 8);
3896 __ sha512su0(vin0, __ T2D, vin1);
3897 }
3898 __ sha512h(vi3, __ T2D, v6, v7);
3899 if (dr < 32) {
3900 __ sha512su1(vin0, __ T2D, vin2, v5);
3901 }
3902 __ addv(vi4, __ T2D, vi1, vi3);
3903 __ sha512h2(vi3, __ T2D, vi1, vi0);
3904 }
3905
3906 // Arguments:
3907 //
3908 // Inputs:
3909 // c_rarg0 - byte[] source+offset
3910 // c_rarg1 - int[] SHA.state
3911 // c_rarg2 - int offset
3912 // c_rarg3 - int limit
3913 //
3914 address generate_sha512_implCompress(StubGenStubId stub_id) {
3915 bool multi_block;
3916 switch (stub_id) {
3917 case sha512_implCompress_id:
3918 multi_block = false;
3919 break;
3920 case sha512_implCompressMB_id:
3921 multi_block = true;
3922 break;
3923 default:
3924 ShouldNotReachHere();
3925 }
3926
3927 static const uint64_t round_consts[80] = {
3928 0x428A2F98D728AE22L, 0x7137449123EF65CDL, 0xB5C0FBCFEC4D3B2FL,
3929 0xE9B5DBA58189DBBCL, 0x3956C25BF348B538L, 0x59F111F1B605D019L,
3930 0x923F82A4AF194F9BL, 0xAB1C5ED5DA6D8118L, 0xD807AA98A3030242L,
3931 0x12835B0145706FBEL, 0x243185BE4EE4B28CL, 0x550C7DC3D5FFB4E2L,
3932 0x72BE5D74F27B896FL, 0x80DEB1FE3B1696B1L, 0x9BDC06A725C71235L,
3933 0xC19BF174CF692694L, 0xE49B69C19EF14AD2L, 0xEFBE4786384F25E3L,
3934 0x0FC19DC68B8CD5B5L, 0x240CA1CC77AC9C65L, 0x2DE92C6F592B0275L,
3935 0x4A7484AA6EA6E483L, 0x5CB0A9DCBD41FBD4L, 0x76F988DA831153B5L,
3936 0x983E5152EE66DFABL, 0xA831C66D2DB43210L, 0xB00327C898FB213FL,
3937 0xBF597FC7BEEF0EE4L, 0xC6E00BF33DA88FC2L, 0xD5A79147930AA725L,
3938 0x06CA6351E003826FL, 0x142929670A0E6E70L, 0x27B70A8546D22FFCL,
3939 0x2E1B21385C26C926L, 0x4D2C6DFC5AC42AEDL, 0x53380D139D95B3DFL,
3940 0x650A73548BAF63DEL, 0x766A0ABB3C77B2A8L, 0x81C2C92E47EDAEE6L,
3941 0x92722C851482353BL, 0xA2BFE8A14CF10364L, 0xA81A664BBC423001L,
3942 0xC24B8B70D0F89791L, 0xC76C51A30654BE30L, 0xD192E819D6EF5218L,
3943 0xD69906245565A910L, 0xF40E35855771202AL, 0x106AA07032BBD1B8L,
3944 0x19A4C116B8D2D0C8L, 0x1E376C085141AB53L, 0x2748774CDF8EEB99L,
3945 0x34B0BCB5E19B48A8L, 0x391C0CB3C5C95A63L, 0x4ED8AA4AE3418ACBL,
3946 0x5B9CCA4F7763E373L, 0x682E6FF3D6B2B8A3L, 0x748F82EE5DEFB2FCL,
3947 0x78A5636F43172F60L, 0x84C87814A1F0AB72L, 0x8CC702081A6439ECL,
3948 0x90BEFFFA23631E28L, 0xA4506CEBDE82BDE9L, 0xBEF9A3F7B2C67915L,
3949 0xC67178F2E372532BL, 0xCA273ECEEA26619CL, 0xD186B8C721C0C207L,
3950 0xEADA7DD6CDE0EB1EL, 0xF57D4F7FEE6ED178L, 0x06F067AA72176FBAL,
3951 0x0A637DC5A2C898A6L, 0x113F9804BEF90DAEL, 0x1B710B35131C471BL,
3952 0x28DB77F523047D84L, 0x32CAAB7B40C72493L, 0x3C9EBE0A15C9BEBCL,
3953 0x431D67C49C100D4CL, 0x4CC5D4BECB3E42B6L, 0x597F299CFC657E2AL,
3954 0x5FCB6FAB3AD6FAECL, 0x6C44198C4A475817L
3955 };
3956
3957 __ align(CodeEntryAlignment);
3958
3959 StubCodeMark mark(this, stub_id);
3960 address start = __ pc();
3961
3962 Register buf = c_rarg0;
3963 Register state = c_rarg1;
3964 Register ofs = c_rarg2;
3965 Register limit = c_rarg3;
3966
3967 __ stpd(v8, v9, __ pre(sp, -64));
3968 __ stpd(v10, v11, Address(sp, 16));
3969 __ stpd(v12, v13, Address(sp, 32));
3970 __ stpd(v14, v15, Address(sp, 48));
3971
3972 Label sha512_loop;
3973
3974 // load state
3975 __ ld1(v8, v9, v10, v11, __ T2D, state);
3976
3977 // load first 4 round constants
3978 __ lea(rscratch1, ExternalAddress((address)round_consts));
3979 __ ld1(v20, v21, v22, v23, __ T2D, __ post(rscratch1, 64));
3980
3981 __ BIND(sha512_loop);
3982 // load 128B of data into v12..v19
3983 __ ld1(v12, v13, v14, v15, __ T2D, __ post(buf, 64));
3984 __ ld1(v16, v17, v18, v19, __ T2D, __ post(buf, 64));
3985 __ rev64(v12, __ T16B, v12);
3986 __ rev64(v13, __ T16B, v13);
3987 __ rev64(v14, __ T16B, v14);
3988 __ rev64(v15, __ T16B, v15);
3989 __ rev64(v16, __ T16B, v16);
3990 __ rev64(v17, __ T16B, v17);
3991 __ rev64(v18, __ T16B, v18);
3992 __ rev64(v19, __ T16B, v19);
3993
3994 __ mov(rscratch2, rscratch1);
3995
3996 __ mov(v0, __ T16B, v8);
3997 __ mov(v1, __ T16B, v9);
3998 __ mov(v2, __ T16B, v10);
3999 __ mov(v3, __ T16B, v11);
4000
4001 sha512_dround( 0, v0, v1, v2, v3, v4, v20, v24, v12, v13, v19, v16, v17);
4002 sha512_dround( 1, v3, v0, v4, v2, v1, v21, v25, v13, v14, v12, v17, v18);
4003 sha512_dround( 2, v2, v3, v1, v4, v0, v22, v26, v14, v15, v13, v18, v19);
4004 sha512_dround( 3, v4, v2, v0, v1, v3, v23, v27, v15, v16, v14, v19, v12);
4005 sha512_dround( 4, v1, v4, v3, v0, v2, v24, v28, v16, v17, v15, v12, v13);
4006 sha512_dround( 5, v0, v1, v2, v3, v4, v25, v29, v17, v18, v16, v13, v14);
4007 sha512_dround( 6, v3, v0, v4, v2, v1, v26, v30, v18, v19, v17, v14, v15);
4008 sha512_dround( 7, v2, v3, v1, v4, v0, v27, v31, v19, v12, v18, v15, v16);
4009 sha512_dround( 8, v4, v2, v0, v1, v3, v28, v24, v12, v13, v19, v16, v17);
4010 sha512_dround( 9, v1, v4, v3, v0, v2, v29, v25, v13, v14, v12, v17, v18);
4011 sha512_dround(10, v0, v1, v2, v3, v4, v30, v26, v14, v15, v13, v18, v19);
4012 sha512_dround(11, v3, v0, v4, v2, v1, v31, v27, v15, v16, v14, v19, v12);
4013 sha512_dround(12, v2, v3, v1, v4, v0, v24, v28, v16, v17, v15, v12, v13);
4014 sha512_dround(13, v4, v2, v0, v1, v3, v25, v29, v17, v18, v16, v13, v14);
4015 sha512_dround(14, v1, v4, v3, v0, v2, v26, v30, v18, v19, v17, v14, v15);
4016 sha512_dround(15, v0, v1, v2, v3, v4, v27, v31, v19, v12, v18, v15, v16);
4017 sha512_dround(16, v3, v0, v4, v2, v1, v28, v24, v12, v13, v19, v16, v17);
4018 sha512_dround(17, v2, v3, v1, v4, v0, v29, v25, v13, v14, v12, v17, v18);
4019 sha512_dround(18, v4, v2, v0, v1, v3, v30, v26, v14, v15, v13, v18, v19);
4020 sha512_dround(19, v1, v4, v3, v0, v2, v31, v27, v15, v16, v14, v19, v12);
4021 sha512_dround(20, v0, v1, v2, v3, v4, v24, v28, v16, v17, v15, v12, v13);
4022 sha512_dround(21, v3, v0, v4, v2, v1, v25, v29, v17, v18, v16, v13, v14);
4023 sha512_dround(22, v2, v3, v1, v4, v0, v26, v30, v18, v19, v17, v14, v15);
4024 sha512_dround(23, v4, v2, v0, v1, v3, v27, v31, v19, v12, v18, v15, v16);
4025 sha512_dround(24, v1, v4, v3, v0, v2, v28, v24, v12, v13, v19, v16, v17);
4026 sha512_dround(25, v0, v1, v2, v3, v4, v29, v25, v13, v14, v12, v17, v18);
4027 sha512_dround(26, v3, v0, v4, v2, v1, v30, v26, v14, v15, v13, v18, v19);
4028 sha512_dround(27, v2, v3, v1, v4, v0, v31, v27, v15, v16, v14, v19, v12);
4029 sha512_dround(28, v4, v2, v0, v1, v3, v24, v28, v16, v17, v15, v12, v13);
4030 sha512_dround(29, v1, v4, v3, v0, v2, v25, v29, v17, v18, v16, v13, v14);
4031 sha512_dround(30, v0, v1, v2, v3, v4, v26, v30, v18, v19, v17, v14, v15);
4032 sha512_dround(31, v3, v0, v4, v2, v1, v27, v31, v19, v12, v18, v15, v16);
4033 sha512_dround(32, v2, v3, v1, v4, v0, v28, v24, v12, v0, v0, v0, v0);
4034 sha512_dround(33, v4, v2, v0, v1, v3, v29, v25, v13, v0, v0, v0, v0);
4035 sha512_dround(34, v1, v4, v3, v0, v2, v30, v26, v14, v0, v0, v0, v0);
4036 sha512_dround(35, v0, v1, v2, v3, v4, v31, v27, v15, v0, v0, v0, v0);
4037 sha512_dround(36, v3, v0, v4, v2, v1, v24, v0, v16, v0, v0, v0, v0);
4038 sha512_dround(37, v2, v3, v1, v4, v0, v25, v0, v17, v0, v0, v0, v0);
4039 sha512_dround(38, v4, v2, v0, v1, v3, v26, v0, v18, v0, v0, v0, v0);
4040 sha512_dround(39, v1, v4, v3, v0, v2, v27, v0, v19, v0, v0, v0, v0);
4041
4042 __ addv(v8, __ T2D, v8, v0);
4043 __ addv(v9, __ T2D, v9, v1);
4044 __ addv(v10, __ T2D, v10, v2);
4045 __ addv(v11, __ T2D, v11, v3);
4046
4047 if (multi_block) {
4048 __ add(ofs, ofs, 128);
4049 __ cmp(ofs, limit);
4050 __ br(Assembler::LE, sha512_loop);
4051 __ mov(c_rarg0, ofs); // return ofs
4052 }
4053
4054 __ st1(v8, v9, v10, v11, __ T2D, state);
4055
4056 __ ldpd(v14, v15, Address(sp, 48));
4057 __ ldpd(v12, v13, Address(sp, 32));
4058 __ ldpd(v10, v11, Address(sp, 16));
4059 __ ldpd(v8, v9, __ post(sp, 64));
4060
4061 __ ret(lr);
4062
4063 return start;
4064 }
4065
4066 // Execute one round of keccak of two computations in parallel.
4067 // One of the states should be loaded into the lower halves of
4068 // the vector registers v0-v24, the other should be loaded into
4069 // the upper halves of those registers. The ld1r instruction loads
4070 // the round constant into both halves of register v31.
4071 // Intermediate results c0...c5 and d0...d5 are computed
4072 // in registers v25...v30.
4073 // All vector instructions that are used operate on both register
4074 // halves in parallel.
4075 // If only a single computation is needed, one can only load the lower halves.
4076 void keccak_round(Register rscratch1) {
4077 __ eor3(v29, __ T16B, v4, v9, v14); // c4 = a4 ^ a9 ^ a14
4078 __ eor3(v26, __ T16B, v1, v6, v11); // c1 = a1 ^ a16 ^ a11
4079 __ eor3(v28, __ T16B, v3, v8, v13); // c3 = a3 ^ a8 ^a13
4080 __ eor3(v25, __ T16B, v0, v5, v10); // c0 = a0 ^ a5 ^ a10
4081 __ eor3(v27, __ T16B, v2, v7, v12); // c2 = a2 ^ a7 ^ a12
4082 __ eor3(v29, __ T16B, v29, v19, v24); // c4 ^= a19 ^ a24
4083 __ eor3(v26, __ T16B, v26, v16, v21); // c1 ^= a16 ^ a21
4084 __ eor3(v28, __ T16B, v28, v18, v23); // c3 ^= a18 ^ a23
4085 __ eor3(v25, __ T16B, v25, v15, v20); // c0 ^= a15 ^ a20
4086 __ eor3(v27, __ T16B, v27, v17, v22); // c2 ^= a17 ^ a22
4087
4088 __ rax1(v30, __ T2D, v29, v26); // d0 = c4 ^ rol(c1, 1)
4089 __ rax1(v26, __ T2D, v26, v28); // d2 = c1 ^ rol(c3, 1)
4090 __ rax1(v28, __ T2D, v28, v25); // d4 = c3 ^ rol(c0, 1)
4091 __ rax1(v25, __ T2D, v25, v27); // d1 = c0 ^ rol(c2, 1)
4092 __ rax1(v27, __ T2D, v27, v29); // d3 = c2 ^ rol(c4, 1)
4093
4094 __ eor(v0, __ T16B, v0, v30); // a0 = a0 ^ d0
4095 __ xar(v29, __ T2D, v1, v25, (64 - 1)); // a10' = rol((a1^d1), 1)
4096 __ xar(v1, __ T2D, v6, v25, (64 - 44)); // a1 = rol(a6^d1), 44)
4097 __ xar(v6, __ T2D, v9, v28, (64 - 20)); // a6 = rol((a9^d4), 20)
4098 __ xar(v9, __ T2D, v22, v26, (64 - 61)); // a9 = rol((a22^d2), 61)
4099 __ xar(v22, __ T2D, v14, v28, (64 - 39)); // a22 = rol((a14^d4), 39)
4100 __ xar(v14, __ T2D, v20, v30, (64 - 18)); // a14 = rol((a20^d0), 18)
4101 __ xar(v31, __ T2D, v2, v26, (64 - 62)); // a20' = rol((a2^d2), 62)
4102 __ xar(v2, __ T2D, v12, v26, (64 - 43)); // a2 = rol((a12^d2), 43)
4103 __ xar(v12, __ T2D, v13, v27, (64 - 25)); // a12 = rol((a13^d3), 25)
4104 __ xar(v13, __ T2D, v19, v28, (64 - 8)); // a13 = rol((a19^d4), 8)
4105 __ xar(v19, __ T2D, v23, v27, (64 - 56)); // a19 = rol((a23^d3), 56)
4106 __ xar(v23, __ T2D, v15, v30, (64 - 41)); // a23 = rol((a15^d0), 41)
4107 __ xar(v15, __ T2D, v4, v28, (64 - 27)); // a15 = rol((a4^d4), 27)
4108 __ xar(v28, __ T2D, v24, v28, (64 - 14)); // a4' = rol((a24^d4), 14)
4109 __ xar(v24, __ T2D, v21, v25, (64 - 2)); // a24 = rol((a21^d1), 2)
4110 __ xar(v8, __ T2D, v8, v27, (64 - 55)); // a21' = rol((a8^d3), 55)
4111 __ xar(v4, __ T2D, v16, v25, (64 - 45)); // a8' = rol((a16^d1), 45)
4112 __ xar(v16, __ T2D, v5, v30, (64 - 36)); // a16 = rol((a5^d0), 36)
4113 __ xar(v5, __ T2D, v3, v27, (64 - 28)); // a5 = rol((a3^d3), 28)
4114 __ xar(v27, __ T2D, v18, v27, (64 - 21)); // a3' = rol((a18^d3), 21)
4115 __ xar(v3, __ T2D, v17, v26, (64 - 15)); // a18' = rol((a17^d2), 15)
4116 __ xar(v25, __ T2D, v11, v25, (64 - 10)); // a17' = rol((a11^d1), 10)
4117 __ xar(v26, __ T2D, v7, v26, (64 - 6)); // a11' = rol((a7^d2), 6)
4118 __ xar(v30, __ T2D, v10, v30, (64 - 3)); // a7' = rol((a10^d0), 3)
4119
4120 __ bcax(v20, __ T16B, v31, v22, v8); // a20 = a20' ^ (~a21 & a22')
4121 __ bcax(v21, __ T16B, v8, v23, v22); // a21 = a21' ^ (~a22 & a23)
4122 __ bcax(v22, __ T16B, v22, v24, v23); // a22 = a22 ^ (~a23 & a24)
4123 __ bcax(v23, __ T16B, v23, v31, v24); // a23 = a23 ^ (~a24 & a20')
4124 __ bcax(v24, __ T16B, v24, v8, v31); // a24 = a24 ^ (~a20' & a21')
4125
4126 __ ld1r(v31, __ T2D, __ post(rscratch1, 8)); // rc = round_constants[i]
4127
4128 __ bcax(v17, __ T16B, v25, v19, v3); // a17 = a17' ^ (~a18' & a19)
4129 __ bcax(v18, __ T16B, v3, v15, v19); // a18 = a18' ^ (~a19 & a15')
4130 __ bcax(v19, __ T16B, v19, v16, v15); // a19 = a19 ^ (~a15 & a16)
4131 __ bcax(v15, __ T16B, v15, v25, v16); // a15 = a15 ^ (~a16 & a17')
4132 __ bcax(v16, __ T16B, v16, v3, v25); // a16 = a16 ^ (~a17' & a18')
4133
4134 __ bcax(v10, __ T16B, v29, v12, v26); // a10 = a10' ^ (~a11' & a12)
4135 __ bcax(v11, __ T16B, v26, v13, v12); // a11 = a11' ^ (~a12 & a13)
4136 __ bcax(v12, __ T16B, v12, v14, v13); // a12 = a12 ^ (~a13 & a14)
4137 __ bcax(v13, __ T16B, v13, v29, v14); // a13 = a13 ^ (~a14 & a10')
4138 __ bcax(v14, __ T16B, v14, v26, v29); // a14 = a14 ^ (~a10' & a11')
4139
4140 __ bcax(v7, __ T16B, v30, v9, v4); // a7 = a7' ^ (~a8' & a9)
4141 __ bcax(v8, __ T16B, v4, v5, v9); // a8 = a8' ^ (~a9 & a5)
4142 __ bcax(v9, __ T16B, v9, v6, v5); // a9 = a9 ^ (~a5 & a6)
4143 __ bcax(v5, __ T16B, v5, v30, v6); // a5 = a5 ^ (~a6 & a7)
4144 __ bcax(v6, __ T16B, v6, v4, v30); // a6 = a6 ^ (~a7 & a8')
4145
4146 __ bcax(v3, __ T16B, v27, v0, v28); // a3 = a3' ^ (~a4' & a0)
4147 __ bcax(v4, __ T16B, v28, v1, v0); // a4 = a4' ^ (~a0 & a1)
4148 __ bcax(v0, __ T16B, v0, v2, v1); // a0 = a0 ^ (~a1 & a2)
4149 __ bcax(v1, __ T16B, v1, v27, v2); // a1 = a1 ^ (~a2 & a3)
4150 __ bcax(v2, __ T16B, v2, v28, v27); // a2 = a2 ^ (~a3 & a4')
4151
4152 __ eor(v0, __ T16B, v0, v31); // a0 = a0 ^ rc
4153 }
4154
4155 // Arguments:
4156 //
4157 // Inputs:
4158 // c_rarg0 - byte[] source+offset
4159 // c_rarg1 - byte[] SHA.state
4160 // c_rarg2 - int block_size
4161 // c_rarg3 - int offset
4162 // c_rarg4 - int limit
4163 //
4164 address generate_sha3_implCompress(StubGenStubId stub_id) {
4165 bool multi_block;
4166 switch (stub_id) {
4167 case sha3_implCompress_id:
4168 multi_block = false;
4169 break;
4170 case sha3_implCompressMB_id:
4171 multi_block = true;
4172 break;
4173 default:
4174 ShouldNotReachHere();
4175 }
4176
4177 static const uint64_t round_consts[24] = {
4178 0x0000000000000001L, 0x0000000000008082L, 0x800000000000808AL,
4179 0x8000000080008000L, 0x000000000000808BL, 0x0000000080000001L,
4180 0x8000000080008081L, 0x8000000000008009L, 0x000000000000008AL,
4181 0x0000000000000088L, 0x0000000080008009L, 0x000000008000000AL,
4182 0x000000008000808BL, 0x800000000000008BL, 0x8000000000008089L,
4183 0x8000000000008003L, 0x8000000000008002L, 0x8000000000000080L,
4184 0x000000000000800AL, 0x800000008000000AL, 0x8000000080008081L,
4185 0x8000000000008080L, 0x0000000080000001L, 0x8000000080008008L
4186 };
4187
4188 __ align(CodeEntryAlignment);
4189
4190 StubCodeMark mark(this, stub_id);
4191 address start = __ pc();
4192
4193 Register buf = c_rarg0;
4194 Register state = c_rarg1;
4195 Register block_size = c_rarg2;
4196 Register ofs = c_rarg3;
4197 Register limit = c_rarg4;
4198
4199 Label sha3_loop, rounds24_loop;
4200 Label sha3_512_or_sha3_384, shake128;
4201
4202 __ stpd(v8, v9, __ pre(sp, -64));
4203 __ stpd(v10, v11, Address(sp, 16));
4204 __ stpd(v12, v13, Address(sp, 32));
4205 __ stpd(v14, v15, Address(sp, 48));
4206
4207 // load state
4208 __ add(rscratch1, state, 32);
4209 __ ld1(v0, v1, v2, v3, __ T1D, state);
4210 __ ld1(v4, v5, v6, v7, __ T1D, __ post(rscratch1, 32));
4211 __ ld1(v8, v9, v10, v11, __ T1D, __ post(rscratch1, 32));
4212 __ ld1(v12, v13, v14, v15, __ T1D, __ post(rscratch1, 32));
4213 __ ld1(v16, v17, v18, v19, __ T1D, __ post(rscratch1, 32));
4214 __ ld1(v20, v21, v22, v23, __ T1D, __ post(rscratch1, 32));
4215 __ ld1(v24, __ T1D, rscratch1);
4216
4217 __ BIND(sha3_loop);
4218
4219 // 24 keccak rounds
4220 __ movw(rscratch2, 24);
4221
4222 // load round_constants base
4223 __ lea(rscratch1, ExternalAddress((address) round_consts));
4224
4225 // load input
4226 __ ld1(v25, v26, v27, v28, __ T8B, __ post(buf, 32));
4227 __ ld1(v29, v30, v31, __ T8B, __ post(buf, 24));
4228 __ eor(v0, __ T8B, v0, v25);
4229 __ eor(v1, __ T8B, v1, v26);
4230 __ eor(v2, __ T8B, v2, v27);
4231 __ eor(v3, __ T8B, v3, v28);
4232 __ eor(v4, __ T8B, v4, v29);
4233 __ eor(v5, __ T8B, v5, v30);
4234 __ eor(v6, __ T8B, v6, v31);
4235
4236 // block_size == 72, SHA3-512; block_size == 104, SHA3-384
4237 __ tbz(block_size, 7, sha3_512_or_sha3_384);
4238
4239 __ ld1(v25, v26, v27, v28, __ T8B, __ post(buf, 32));
4240 __ ld1(v29, v30, v31, __ T8B, __ post(buf, 24));
4241 __ eor(v7, __ T8B, v7, v25);
4242 __ eor(v8, __ T8B, v8, v26);
4243 __ eor(v9, __ T8B, v9, v27);
4244 __ eor(v10, __ T8B, v10, v28);
4245 __ eor(v11, __ T8B, v11, v29);
4246 __ eor(v12, __ T8B, v12, v30);
4247 __ eor(v13, __ T8B, v13, v31);
4248
4249 __ ld1(v25, v26, v27, __ T8B, __ post(buf, 24));
4250 __ eor(v14, __ T8B, v14, v25);
4251 __ eor(v15, __ T8B, v15, v26);
4252 __ eor(v16, __ T8B, v16, v27);
4253
4254 // block_size == 136, bit4 == 0 and bit5 == 0, SHA3-256 or SHAKE256
4255 __ andw(c_rarg5, block_size, 48);
4256 __ cbzw(c_rarg5, rounds24_loop);
4257
4258 __ tbnz(block_size, 5, shake128);
4259 // block_size == 144, bit5 == 0, SHA3-224
4260 __ ldrd(v28, __ post(buf, 8));
4261 __ eor(v17, __ T8B, v17, v28);
4262 __ b(rounds24_loop);
4263
4264 __ BIND(shake128);
4265 __ ld1(v28, v29, v30, v31, __ T8B, __ post(buf, 32));
4266 __ eor(v17, __ T8B, v17, v28);
4267 __ eor(v18, __ T8B, v18, v29);
4268 __ eor(v19, __ T8B, v19, v30);
4269 __ eor(v20, __ T8B, v20, v31);
4270 __ b(rounds24_loop); // block_size == 168, SHAKE128
4271
4272 __ BIND(sha3_512_or_sha3_384);
4273 __ ld1(v25, v26, __ T8B, __ post(buf, 16));
4274 __ eor(v7, __ T8B, v7, v25);
4275 __ eor(v8, __ T8B, v8, v26);
4276 __ tbz(block_size, 5, rounds24_loop); // SHA3-512
4277
4278 // SHA3-384
4279 __ ld1(v27, v28, v29, v30, __ T8B, __ post(buf, 32));
4280 __ eor(v9, __ T8B, v9, v27);
4281 __ eor(v10, __ T8B, v10, v28);
4282 __ eor(v11, __ T8B, v11, v29);
4283 __ eor(v12, __ T8B, v12, v30);
4284
4285 __ BIND(rounds24_loop);
4286 __ subw(rscratch2, rscratch2, 1);
4287
4288 keccak_round(rscratch1);
4289
4290 __ cbnzw(rscratch2, rounds24_loop);
4291
4292 if (multi_block) {
4293 __ add(ofs, ofs, block_size);
4294 __ cmp(ofs, limit);
4295 __ br(Assembler::LE, sha3_loop);
4296 __ mov(c_rarg0, ofs); // return ofs
4297 }
4298
4299 __ st1(v0, v1, v2, v3, __ T1D, __ post(state, 32));
4300 __ st1(v4, v5, v6, v7, __ T1D, __ post(state, 32));
4301 __ st1(v8, v9, v10, v11, __ T1D, __ post(state, 32));
4302 __ st1(v12, v13, v14, v15, __ T1D, __ post(state, 32));
4303 __ st1(v16, v17, v18, v19, __ T1D, __ post(state, 32));
4304 __ st1(v20, v21, v22, v23, __ T1D, __ post(state, 32));
4305 __ st1(v24, __ T1D, state);
4306
4307 // restore callee-saved registers
4308 __ ldpd(v14, v15, Address(sp, 48));
4309 __ ldpd(v12, v13, Address(sp, 32));
4310 __ ldpd(v10, v11, Address(sp, 16));
4311 __ ldpd(v8, v9, __ post(sp, 64));
4312
4313 __ ret(lr);
4314
4315 return start;
4316 }
4317
4318 // Inputs:
4319 // c_rarg0 - long[] state0
4320 // c_rarg1 - long[] state1
4321 address generate_double_keccak() {
4322 static const uint64_t round_consts[24] = {
4323 0x0000000000000001L, 0x0000000000008082L, 0x800000000000808AL,
4324 0x8000000080008000L, 0x000000000000808BL, 0x0000000080000001L,
4325 0x8000000080008081L, 0x8000000000008009L, 0x000000000000008AL,
4326 0x0000000000000088L, 0x0000000080008009L, 0x000000008000000AL,
4327 0x000000008000808BL, 0x800000000000008BL, 0x8000000000008089L,
4328 0x8000000000008003L, 0x8000000000008002L, 0x8000000000000080L,
4329 0x000000000000800AL, 0x800000008000000AL, 0x8000000080008081L,
4330 0x8000000000008080L, 0x0000000080000001L, 0x8000000080008008L
4331 };
4332
4333 // Implements the double_keccak() method of the
4334 // sun.secyrity.provider.SHA3Parallel class
4335 __ align(CodeEntryAlignment);
4336 StubCodeMark mark(this, "StubRoutines", "double_keccak");
4337 address start = __ pc();
4338 __ enter();
4339
4340 Register state0 = c_rarg0;
4341 Register state1 = c_rarg1;
4342
4343 Label rounds24_loop;
4344
4345 // save callee-saved registers
4346 __ stpd(v8, v9, __ pre(sp, -64));
4347 __ stpd(v10, v11, Address(sp, 16));
4348 __ stpd(v12, v13, Address(sp, 32));
4349 __ stpd(v14, v15, Address(sp, 48));
4350
4351 // load states
4352 __ add(rscratch1, state0, 32);
4353 __ ld4(v0, v1, v2, v3, __ D, 0, state0);
4354 __ ld4(v4, v5, v6, v7, __ D, 0, __ post(rscratch1, 32));
4355 __ ld4(v8, v9, v10, v11, __ D, 0, __ post(rscratch1, 32));
4356 __ ld4(v12, v13, v14, v15, __ D, 0, __ post(rscratch1, 32));
4357 __ ld4(v16, v17, v18, v19, __ D, 0, __ post(rscratch1, 32));
4358 __ ld4(v20, v21, v22, v23, __ D, 0, __ post(rscratch1, 32));
4359 __ ld1(v24, __ D, 0, rscratch1);
4360 __ add(rscratch1, state1, 32);
4361 __ ld4(v0, v1, v2, v3, __ D, 1, state1);
4362 __ ld4(v4, v5, v6, v7, __ D, 1, __ post(rscratch1, 32));
4363 __ ld4(v8, v9, v10, v11, __ D, 1, __ post(rscratch1, 32));
4364 __ ld4(v12, v13, v14, v15, __ D, 1, __ post(rscratch1, 32));
4365 __ ld4(v16, v17, v18, v19, __ D, 1, __ post(rscratch1, 32));
4366 __ ld4(v20, v21, v22, v23, __ D, 1, __ post(rscratch1, 32));
4367 __ ld1(v24, __ D, 1, rscratch1);
4368
4369 // 24 keccak rounds
4370 __ movw(rscratch2, 24);
4371
4372 // load round_constants base
4373 __ lea(rscratch1, ExternalAddress((address) round_consts));
4374
4375 __ BIND(rounds24_loop);
4376 __ subw(rscratch2, rscratch2, 1);
4377 keccak_round(rscratch1);
4378 __ cbnzw(rscratch2, rounds24_loop);
4379
4380 __ st4(v0, v1, v2, v3, __ D, 0, __ post(state0, 32));
4381 __ st4(v4, v5, v6, v7, __ D, 0, __ post(state0, 32));
4382 __ st4(v8, v9, v10, v11, __ D, 0, __ post(state0, 32));
4383 __ st4(v12, v13, v14, v15, __ D, 0, __ post(state0, 32));
4384 __ st4(v16, v17, v18, v19, __ D, 0, __ post(state0, 32));
4385 __ st4(v20, v21, v22, v23, __ D, 0, __ post(state0, 32));
4386 __ st1(v24, __ D, 0, state0);
4387 __ st4(v0, v1, v2, v3, __ D, 1, __ post(state1, 32));
4388 __ st4(v4, v5, v6, v7, __ D, 1, __ post(state1, 32));
4389 __ st4(v8, v9, v10, v11, __ D, 1, __ post(state1, 32));
4390 __ st4(v12, v13, v14, v15, __ D, 1, __ post(state1, 32));
4391 __ st4(v16, v17, v18, v19, __ D, 1, __ post(state1, 32));
4392 __ st4(v20, v21, v22, v23, __ D, 1, __ post(state1, 32));
4393 __ st1(v24, __ D, 1, state1);
4394
4395 // restore callee-saved vector registers
4396 __ ldpd(v14, v15, Address(sp, 48));
4397 __ ldpd(v12, v13, Address(sp, 32));
4398 __ ldpd(v10, v11, Address(sp, 16));
4399 __ ldpd(v8, v9, __ post(sp, 64));
4400
4401 __ leave(); // required for proper stackwalking of RuntimeStub frame
4402 __ mov(r0, zr); // return 0
4403 __ ret(lr);
4404
4405 return start;
4406 }
4407
4408 // ChaCha20 block function. This version parallelizes the 32-bit
4409 // state elements on each of 16 vectors, producing 4 blocks of
4410 // keystream at a time.
4411 //
4412 // state (int[16]) = c_rarg0
4413 // keystream (byte[256]) = c_rarg1
4414 // return - number of bytes of produced keystream (always 256)
4415 //
4416 // This implementation takes each 32-bit integer from the state
4417 // array and broadcasts it across all 4 32-bit lanes of a vector register
4418 // (e.g. state[0] is replicated on all 4 lanes of v4, state[1] to all 4 lanes
4419 // of v5, etc.). Once all 16 elements have been broadcast onto 16 vectors,
4420 // the quarter round schedule is implemented as outlined in RFC 7539 section
4421 // 2.3. However, instead of sequentially processing the 3 quarter round
4422 // operations represented by one QUARTERROUND function, we instead stack all
4423 // the adds, xors and left-rotations from the first 4 quarter rounds together
4424 // and then do the same for the second set of 4 quarter rounds. This removes
4425 // some latency that would otherwise be incurred by waiting for an add to
4426 // complete before performing an xor (which depends on the result of the
4427 // add), etc. An adjustment happens between the first and second groups of 4
4428 // quarter rounds, but this is done only in the inputs to the macro functions
4429 // that generate the assembly instructions - these adjustments themselves are
4430 // not part of the resulting assembly.
4431 // The 4 registers v0-v3 are used during the quarter round operations as
4432 // scratch registers. Once the 20 rounds are complete, these 4 scratch
4433 // registers become the vectors involved in adding the start state back onto
4434 // the post-QR working state. After the adds are complete, each of the 16
4435 // vectors write their first lane back to the keystream buffer, followed
4436 // by the second lane from all vectors and so on.
4437 address generate_chacha20Block_blockpar() {
4438 Label L_twoRounds, L_cc20_const;
4439 // The constant data is broken into two 128-bit segments to be loaded
4440 // onto FloatRegisters. The first 128 bits are a counter add overlay
4441 // that adds +0/+1/+2/+3 to the vector holding replicated state[12].
4442 // The second 128-bits is a table constant used for 8-bit left rotations.
4443 __ BIND(L_cc20_const);
4444 __ emit_int64(0x0000000100000000UL);
4445 __ emit_int64(0x0000000300000002UL);
4446 __ emit_int64(0x0605040702010003UL);
4447 __ emit_int64(0x0E0D0C0F0A09080BUL);
4448
4449 __ align(CodeEntryAlignment);
4450 StubGenStubId stub_id = StubGenStubId::chacha20Block_id;
4451 StubCodeMark mark(this, stub_id);
4452 address start = __ pc();
4453 __ enter();
4454
4455 int i, j;
4456 const Register state = c_rarg0;
4457 const Register keystream = c_rarg1;
4458 const Register loopCtr = r10;
4459 const Register tmpAddr = r11;
4460 const FloatRegister ctrAddOverlay = v28;
4461 const FloatRegister lrot8Tbl = v29;
4462
4463 // Organize SIMD registers in an array that facilitates
4464 // putting repetitive opcodes into loop structures. It is
4465 // important that each grouping of 4 registers is monotonically
4466 // increasing to support the requirements of multi-register
4467 // instructions (e.g. ld4r, st4, etc.)
4468 const FloatRegister workSt[16] = {
4469 v4, v5, v6, v7, v16, v17, v18, v19,
4470 v20, v21, v22, v23, v24, v25, v26, v27
4471 };
4472
4473 // Pull in constant data. The first 16 bytes are the add overlay
4474 // which is applied to the vector holding the counter (state[12]).
4475 // The second 16 bytes is the index register for the 8-bit left
4476 // rotation tbl instruction.
4477 __ adr(tmpAddr, L_cc20_const);
4478 __ ldpq(ctrAddOverlay, lrot8Tbl, Address(tmpAddr));
4479
4480 // Load from memory and interlace across 16 SIMD registers,
4481 // With each word from memory being broadcast to all lanes of
4482 // each successive SIMD register.
4483 // Addr(0) -> All lanes in workSt[i]
4484 // Addr(4) -> All lanes workSt[i + 1], etc.
4485 __ mov(tmpAddr, state);
4486 for (i = 0; i < 16; i += 4) {
4487 __ ld4r(workSt[i], workSt[i + 1], workSt[i + 2], workSt[i + 3], __ T4S,
4488 __ post(tmpAddr, 16));
4489 }
4490 __ addv(workSt[12], __ T4S, workSt[12], ctrAddOverlay); // Add ctr overlay
4491
4492 // Before entering the loop, create 5 4-register arrays. These
4493 // will hold the 4 registers that represent the a/b/c/d fields
4494 // in the quarter round operation. For instance the "b" field
4495 // for the first 4 quarter round operations is the set of v16/v17/v18/v19,
4496 // but in the second 4 quarter rounds it gets adjusted to v17/v18/v19/v16
4497 // since it is part of a diagonal organization. The aSet and scratch
4498 // register sets are defined at declaration time because they do not change
4499 // organization at any point during the 20-round processing.
4500 FloatRegister aSet[4] = { v4, v5, v6, v7 };
4501 FloatRegister bSet[4];
4502 FloatRegister cSet[4];
4503 FloatRegister dSet[4];
4504 FloatRegister scratch[4] = { v0, v1, v2, v3 };
4505
4506 // Set up the 10 iteration loop and perform all 8 quarter round ops
4507 __ mov(loopCtr, 10);
4508 __ BIND(L_twoRounds);
4509
4510 // Set to columnar organization and do the following 4 quarter-rounds:
4511 // QUARTERROUND(0, 4, 8, 12)
4512 // QUARTERROUND(1, 5, 9, 13)
4513 // QUARTERROUND(2, 6, 10, 14)
4514 // QUARTERROUND(3, 7, 11, 15)
4515 __ cc20_set_qr_registers(bSet, workSt, 4, 5, 6, 7);
4516 __ cc20_set_qr_registers(cSet, workSt, 8, 9, 10, 11);
4517 __ cc20_set_qr_registers(dSet, workSt, 12, 13, 14, 15);
4518
4519 __ cc20_qr_add4(aSet, bSet); // a += b
4520 __ cc20_qr_xor4(dSet, aSet, dSet); // d ^= a
4521 __ cc20_qr_lrot4(dSet, dSet, 16, lrot8Tbl); // d <<<= 16
4522
4523 __ cc20_qr_add4(cSet, dSet); // c += d
4524 __ cc20_qr_xor4(bSet, cSet, scratch); // b ^= c (scratch)
4525 __ cc20_qr_lrot4(scratch, bSet, 12, lrot8Tbl); // b <<<= 12
4526
4527 __ cc20_qr_add4(aSet, bSet); // a += b
4528 __ cc20_qr_xor4(dSet, aSet, dSet); // d ^= a
4529 __ cc20_qr_lrot4(dSet, dSet, 8, lrot8Tbl); // d <<<= 8
4530
4531 __ cc20_qr_add4(cSet, dSet); // c += d
4532 __ cc20_qr_xor4(bSet, cSet, scratch); // b ^= c (scratch)
4533 __ cc20_qr_lrot4(scratch, bSet, 7, lrot8Tbl); // b <<<= 12
4534
4535 // Set to diagonal organization and do the next 4 quarter-rounds:
4536 // QUARTERROUND(0, 5, 10, 15)
4537 // QUARTERROUND(1, 6, 11, 12)
4538 // QUARTERROUND(2, 7, 8, 13)
4539 // QUARTERROUND(3, 4, 9, 14)
4540 __ cc20_set_qr_registers(bSet, workSt, 5, 6, 7, 4);
4541 __ cc20_set_qr_registers(cSet, workSt, 10, 11, 8, 9);
4542 __ cc20_set_qr_registers(dSet, workSt, 15, 12, 13, 14);
4543
4544 __ cc20_qr_add4(aSet, bSet); // a += b
4545 __ cc20_qr_xor4(dSet, aSet, dSet); // d ^= a
4546 __ cc20_qr_lrot4(dSet, dSet, 16, lrot8Tbl); // d <<<= 16
4547
4548 __ cc20_qr_add4(cSet, dSet); // c += d
4549 __ cc20_qr_xor4(bSet, cSet, scratch); // b ^= c (scratch)
4550 __ cc20_qr_lrot4(scratch, bSet, 12, lrot8Tbl); // b <<<= 12
4551
4552 __ cc20_qr_add4(aSet, bSet); // a += b
4553 __ cc20_qr_xor4(dSet, aSet, dSet); // d ^= a
4554 __ cc20_qr_lrot4(dSet, dSet, 8, lrot8Tbl); // d <<<= 8
4555
4556 __ cc20_qr_add4(cSet, dSet); // c += d
4557 __ cc20_qr_xor4(bSet, cSet, scratch); // b ^= c (scratch)
4558 __ cc20_qr_lrot4(scratch, bSet, 7, lrot8Tbl); // b <<<= 12
4559
4560 // Decrement and iterate
4561 __ sub(loopCtr, loopCtr, 1);
4562 __ cbnz(loopCtr, L_twoRounds);
4563
4564 __ mov(tmpAddr, state);
4565
4566 // Add the starting state back to the post-loop keystream
4567 // state. We read/interlace the state array from memory into
4568 // 4 registers similar to what we did in the beginning. Then
4569 // add the counter overlay onto workSt[12] at the end.
4570 for (i = 0; i < 16; i += 4) {
4571 __ ld4r(v0, v1, v2, v3, __ T4S, __ post(tmpAddr, 16));
4572 __ addv(workSt[i], __ T4S, workSt[i], v0);
4573 __ addv(workSt[i + 1], __ T4S, workSt[i + 1], v1);
4574 __ addv(workSt[i + 2], __ T4S, workSt[i + 2], v2);
4575 __ addv(workSt[i + 3], __ T4S, workSt[i + 3], v3);
4576 }
4577 __ addv(workSt[12], __ T4S, workSt[12], ctrAddOverlay); // Add ctr overlay
4578
4579 // Write working state into the keystream buffer. This is accomplished
4580 // by taking the lane "i" from each of the four vectors and writing
4581 // it to consecutive 4-byte offsets, then post-incrementing by 16 and
4582 // repeating with the next 4 vectors until all 16 vectors have been used.
4583 // Then move to the next lane and repeat the process until all lanes have
4584 // been written.
4585 for (i = 0; i < 4; i++) {
4586 for (j = 0; j < 16; j += 4) {
4587 __ st4(workSt[j], workSt[j + 1], workSt[j + 2], workSt[j + 3], __ S, i,
4588 __ post(keystream, 16));
4589 }
4590 }
4591
4592 __ mov(r0, 256); // Return length of output keystream
4593 __ leave();
4594 __ ret(lr);
4595
4596 return start;
4597 }
4598
4599 // Helpers to schedule parallel operation bundles across vector
4600 // register sequences of size 2, 4 or 8.
4601
4602 // Implement various primitive computations across vector sequences
4603
4604 template<int N>
4605 void vs_addv(const VSeq<N>& v, Assembler::SIMD_Arrangement T,
4606 const VSeq<N>& v1, const VSeq<N>& v2) {
4607 // output must not be constant
4608 assert(N == 1 || !v.is_constant(), "cannot output multiple values to a constant vector");
4609 // output cannot overwrite pending inputs
4610 assert(!vs_write_before_read(v, v1), "output overwrites input");
4611 assert(!vs_write_before_read(v, v2), "output overwrites input");
4612 for (int i = 0; i < N; i++) {
4613 __ addv(v[i], T, v1[i], v2[i]);
4614 }
4615 }
4616
4617 template<int N>
4618 void vs_subv(const VSeq<N>& v, Assembler::SIMD_Arrangement T,
4619 const VSeq<N>& v1, const VSeq<N>& v2) {
4620 // output must not be constant
4621 assert(N == 1 || !v.is_constant(), "cannot output multiple values to a constant vector");
4622 // output cannot overwrite pending inputs
4623 assert(!vs_write_before_read(v, v1), "output overwrites input");
4624 assert(!vs_write_before_read(v, v2), "output overwrites input");
4625 for (int i = 0; i < N; i++) {
4626 __ subv(v[i], T, v1[i], v2[i]);
4627 }
4628 }
4629
4630 template<int N>
4631 void vs_mulv(const VSeq<N>& v, Assembler::SIMD_Arrangement T,
4632 const VSeq<N>& v1, const VSeq<N>& v2) {
4633 // output must not be constant
4634 assert(N == 1 || !v.is_constant(), "cannot output multiple values to a constant vector");
4635 // output cannot overwrite pending inputs
4636 assert(!vs_write_before_read(v, v1), "output overwrites input");
4637 assert(!vs_write_before_read(v, v2), "output overwrites input");
4638 for (int i = 0; i < N; i++) {
4639 __ mulv(v[i], T, v1[i], v2[i]);
4640 }
4641 }
4642
4643 template<int N>
4644 void vs_negr(const VSeq<N>& v, Assembler::SIMD_Arrangement T, const VSeq<N>& v1) {
4645 // output must not be constant
4646 assert(N == 1 || !v.is_constant(), "cannot output multiple values to a constant vector");
4647 // output cannot overwrite pending inputs
4648 assert(!vs_write_before_read(v, v1), "output overwrites input");
4649 for (int i = 0; i < N; i++) {
4650 __ negr(v[i], T, v1[i]);
4651 }
4652 }
4653
4654 template<int N>
4655 void vs_sshr(const VSeq<N>& v, Assembler::SIMD_Arrangement T,
4656 const VSeq<N>& v1, int shift) {
4657 // output must not be constant
4658 assert(N == 1 || !v.is_constant(), "cannot output multiple values to a constant vector");
4659 // output cannot overwrite pending inputs
4660 assert(!vs_write_before_read(v, v1), "output overwrites input");
4661 for (int i = 0; i < N; i++) {
4662 __ sshr(v[i], T, v1[i], shift);
4663 }
4664 }
4665
4666 template<int N>
4667 void vs_andr(const VSeq<N>& v, const VSeq<N>& v1, const VSeq<N>& v2) {
4668 // output must not be constant
4669 assert(N == 1 || !v.is_constant(), "cannot output multiple values to a constant vector");
4670 // output cannot overwrite pending inputs
4671 assert(!vs_write_before_read(v, v1), "output overwrites input");
4672 assert(!vs_write_before_read(v, v2), "output overwrites input");
4673 for (int i = 0; i < N; i++) {
4674 __ andr(v[i], __ T16B, v1[i], v2[i]);
4675 }
4676 }
4677
4678 template<int N>
4679 void vs_orr(const VSeq<N>& v, const VSeq<N>& v1, const VSeq<N>& v2) {
4680 // output must not be constant
4681 assert(N == 1 || !v.is_constant(), "cannot output multiple values to a constant vector");
4682 // output cannot overwrite pending inputs
4683 assert(!vs_write_before_read(v, v1), "output overwrites input");
4684 assert(!vs_write_before_read(v, v2), "output overwrites input");
4685 for (int i = 0; i < N; i++) {
4686 __ orr(v[i], __ T16B, v1[i], v2[i]);
4687 }
4688 }
4689
4690 template<int N>
4691 void vs_notr(const VSeq<N>& v, const VSeq<N>& v1) {
4692 // output must not be constant
4693 assert(N == 1 || !v.is_constant(), "cannot output multiple values to a constant vector");
4694 // output cannot overwrite pending inputs
4695 assert(!vs_write_before_read(v, v1), "output overwrites input");
4696 for (int i = 0; i < N; i++) {
4697 __ notr(v[i], __ T16B, v1[i]);
4698 }
4699 }
4700
4701 template<int N>
4702 void vs_sqdmulh(const VSeq<N>& v, Assembler::SIMD_Arrangement T, const VSeq<N>& v1, const VSeq<N>& v2) {
4703 // output must not be constant
4704 assert(N == 1 || !v.is_constant(), "cannot output multiple values to a constant vector");
4705 // output cannot overwrite pending inputs
4706 assert(!vs_write_before_read(v, v1), "output overwrites input");
4707 assert(!vs_write_before_read(v, v2), "output overwrites input");
4708 for (int i = 0; i < N; i++) {
4709 __ sqdmulh(v[i], T, v1[i], v2[i]);
4710 }
4711 }
4712
4713 template<int N>
4714 void vs_mlsv(const VSeq<N>& v, Assembler::SIMD_Arrangement T, const VSeq<N>& v1, VSeq<N>& v2) {
4715 // output must not be constant
4716 assert(N == 1 || !v.is_constant(), "cannot output multiple values to a constant vector");
4717 // output cannot overwrite pending inputs
4718 assert(!vs_write_before_read(v, v1), "output overwrites input");
4719 assert(!vs_write_before_read(v, v2), "output overwrites input");
4720 for (int i = 0; i < N; i++) {
4721 __ mlsv(v[i], T, v1[i], v2[i]);
4722 }
4723 }
4724
4725 // load N/2 successive pairs of quadword values from memory in order
4726 // into N successive vector registers of the sequence via the
4727 // address supplied in base.
4728 template<int N>
4729 void vs_ldpq(const VSeq<N>& v, Register base) {
4730 for (int i = 0; i < N; i += 2) {
4731 __ ldpq(v[i], v[i+1], Address(base, 32 * i));
4732 }
4733 }
4734
4735 // load N/2 successive pairs of quadword values from memory in order
4736 // into N vector registers of the sequence via the address supplied
4737 // in base using post-increment addressing
4738 template<int N>
4739 void vs_ldpq_post(const VSeq<N>& v, Register base) {
4740 static_assert((N & (N - 1)) == 0, "sequence length must be even");
4741 for (int i = 0; i < N; i += 2) {
4742 __ ldpq(v[i], v[i+1], __ post(base, 32));
4743 }
4744 }
4745
4746 // store N successive vector registers of the sequence into N/2
4747 // successive pairs of quadword memory locations via the address
4748 // supplied in base using post-increment addressing
4749 template<int N>
4750 void vs_stpq_post(const VSeq<N>& v, Register base) {
4751 static_assert((N & (N - 1)) == 0, "sequence length must be even");
4752 for (int i = 0; i < N; i += 2) {
4753 __ stpq(v[i], v[i+1], __ post(base, 32));
4754 }
4755 }
4756
4757 // load N/2 pairs of quadword values from memory de-interleaved into
4758 // N vector registers 2 at a time via the address supplied in base
4759 // using post-increment addressing.
4760 template<int N>
4761 void vs_ld2_post(const VSeq<N>& v, Assembler::SIMD_Arrangement T, Register base) {
4762 static_assert((N & (N - 1)) == 0, "sequence length must be even");
4763 for (int i = 0; i < N; i += 2) {
4764 __ ld2(v[i], v[i+1], T, __ post(base, 32));
4765 }
4766 }
4767
4768 // store N vector registers interleaved into N/2 pairs of quadword
4769 // memory locations via the address supplied in base using
4770 // post-increment addressing.
4771 template<int N>
4772 void vs_st2_post(const VSeq<N>& v, Assembler::SIMD_Arrangement T, Register base) {
4773 static_assert((N & (N - 1)) == 0, "sequence length must be even");
4774 for (int i = 0; i < N; i += 2) {
4775 __ st2(v[i], v[i+1], T, __ post(base, 32));
4776 }
4777 }
4778
4779 // load N quadword values from memory de-interleaved into N vector
4780 // registers 3 elements at a time via the address supplied in base.
4781 template<int N>
4782 void vs_ld3(const VSeq<N>& v, Assembler::SIMD_Arrangement T, Register base) {
4783 static_assert(N == ((N / 3) * 3), "sequence length must be multiple of 3");
4784 for (int i = 0; i < N; i += 3) {
4785 __ ld3(v[i], v[i+1], v[i+2], T, base);
4786 }
4787 }
4788
4789 // load N quadword values from memory de-interleaved into N vector
4790 // registers 3 elements at a time via the address supplied in base
4791 // using post-increment addressing.
4792 template<int N>
4793 void vs_ld3_post(const VSeq<N>& v, Assembler::SIMD_Arrangement T, Register base) {
4794 static_assert(N == ((N / 3) * 3), "sequence length must be multiple of 3");
4795 for (int i = 0; i < N; i += 3) {
4796 __ ld3(v[i], v[i+1], v[i+2], T, __ post(base, 48));
4797 }
4798 }
4799
4800 // load N/2 pairs of quadword values from memory into N vector
4801 // registers via the address supplied in base with each pair indexed
4802 // using the the start offset plus the corresponding entry in the
4803 // offsets array
4804 template<int N>
4805 void vs_ldpq_indexed(const VSeq<N>& v, Register base, int start, int (&offsets)[N/2]) {
4806 for (int i = 0; i < N/2; i++) {
4807 __ ldpq(v[2*i], v[2*i+1], Address(base, start + offsets[i]));
4808 }
4809 }
4810
4811 // store N vector registers into N/2 pairs of quadword memory
4812 // locations via the address supplied in base with each pair indexed
4813 // using the the start offset plus the corresponding entry in the
4814 // offsets array
4815 template<int N>
4816 void vs_stpq_indexed(const VSeq<N>& v, Register base, int start, int offsets[N/2]) {
4817 for (int i = 0; i < N/2; i++) {
4818 __ stpq(v[2*i], v[2*i+1], Address(base, start + offsets[i]));
4819 }
4820 }
4821
4822 // load N single quadword values from memory into N vector registers
4823 // via the address supplied in base with each value indexed using
4824 // the the start offset plus the corresponding entry in the offsets
4825 // array
4826 template<int N>
4827 void vs_ldr_indexed(const VSeq<N>& v, Assembler::SIMD_RegVariant T, Register base,
4828 int start, int (&offsets)[N]) {
4829 for (int i = 0; i < N; i++) {
4830 __ ldr(v[i], T, Address(base, start + offsets[i]));
4831 }
4832 }
4833
4834 // store N vector registers into N single quadword memory locations
4835 // via the address supplied in base with each value indexed using
4836 // the the start offset plus the corresponding entry in the offsets
4837 // array
4838 template<int N>
4839 void vs_str_indexed(const VSeq<N>& v, Assembler::SIMD_RegVariant T, Register base,
4840 int start, int (&offsets)[N]) {
4841 for (int i = 0; i < N; i++) {
4842 __ str(v[i], T, Address(base, start + offsets[i]));
4843 }
4844 }
4845
4846 // load N/2 pairs of quadword values from memory de-interleaved into
4847 // N vector registers 2 at a time via the address supplied in base
4848 // with each pair indexed using the the start offset plus the
4849 // corresponding entry in the offsets array
4850 template<int N>
4851 void vs_ld2_indexed(const VSeq<N>& v, Assembler::SIMD_Arrangement T, Register base,
4852 Register tmp, int start, int (&offsets)[N/2]) {
4853 for (int i = 0; i < N/2; i++) {
4854 __ add(tmp, base, start + offsets[i]);
4855 __ ld2(v[2*i], v[2*i+1], T, tmp);
4856 }
4857 }
4858
4859 // store N vector registers 2 at a time interleaved into N/2 pairs
4860 // of quadword memory locations via the address supplied in base
4861 // with each pair indexed using the the start offset plus the
4862 // corresponding entry in the offsets array
4863 template<int N>
4864 void vs_st2_indexed(const VSeq<N>& v, Assembler::SIMD_Arrangement T, Register base,
4865 Register tmp, int start, int (&offsets)[N/2]) {
4866 for (int i = 0; i < N/2; i++) {
4867 __ add(tmp, base, start + offsets[i]);
4868 __ st2(v[2*i], v[2*i+1], T, tmp);
4869 }
4870 }
4871
4872 // Helper routines for various flavours of Montgomery multiply
4873
4874 // Perform 16 32-bit (4x4S) or 32 16-bit (4 x 8H) Montgomery
4875 // multiplications in parallel
4876 //
4877
4878 // See the montMul() method of the sun.security.provider.ML_DSA
4879 // class.
4880 //
4881 // Computes 4x4S results or 8x8H results
4882 // a = b * c * 2^MONT_R_BITS mod MONT_Q
4883 // Inputs: vb, vc - 4x4S or 4x8H vector register sequences
4884 // vq - 2x4S or 2x8H constants <MONT_Q, MONT_Q_INV_MOD_R>
4885 // Temps: vtmp - 4x4S or 4x8H vector sequence trashed after call
4886 // Outputs: va - 4x4S or 4x8H vector register sequences
4887 // vb, vc, vtmp and vq must all be disjoint
4888 // va must be disjoint from all other inputs/temps or must equal vc
4889 // va must have a non-zero delta i.e. it must not be a constant vseq.
4890 // n.b. MONT_R_BITS is 16 or 32, so the right shift by it is implicit.
4891 void vs_montmul4(const VSeq<4>& va, const VSeq<4>& vb, const VSeq<4>& vc,
4892 Assembler::SIMD_Arrangement T,
4893 const VSeq<4>& vtmp, const VSeq<2>& vq) {
4894 assert (T == __ T4S || T == __ T8H, "invalid arrangement for montmul");
4895 assert(vs_disjoint(vb, vc), "vb and vc overlap");
4896 assert(vs_disjoint(vb, vq), "vb and vq overlap");
4897 assert(vs_disjoint(vb, vtmp), "vb and vtmp overlap");
4898
4899 assert(vs_disjoint(vc, vq), "vc and vq overlap");
4900 assert(vs_disjoint(vc, vtmp), "vc and vtmp overlap");
4901
4902 assert(vs_disjoint(vq, vtmp), "vq and vtmp overlap");
4903
4904 assert(vs_disjoint(va, vc) || vs_same(va, vc), "va and vc neither disjoint nor equal");
4905 assert(vs_disjoint(va, vb), "va and vb overlap");
4906 assert(vs_disjoint(va, vq), "va and vq overlap");
4907 assert(vs_disjoint(va, vtmp), "va and vtmp overlap");
4908 assert(!va.is_constant(), "output vector must identify 4 different registers");
4909
4910 // schedule 4 streams of instructions across the vector sequences
4911 for (int i = 0; i < 4; i++) {
4912 __ sqdmulh(vtmp[i], T, vb[i], vc[i]); // aHigh = hi32(2 * b * c)
4913 __ mulv(va[i], T, vb[i], vc[i]); // aLow = lo32(b * c)
4914 }
4915
4916 for (int i = 0; i < 4; i++) {
4917 __ mulv(va[i], T, va[i], vq[0]); // m = aLow * qinv
4918 }
4919
4920 for (int i = 0; i < 4; i++) {
4921 __ sqdmulh(va[i], T, va[i], vq[1]); // n = hi32(2 * m * q)
4922 }
4923
4924 for (int i = 0; i < 4; i++) {
4925 __ shsubv(va[i], T, vtmp[i], va[i]); // a = (aHigh - n) / 2
4926 }
4927 }
4928
4929 // Perform 8 32-bit (4x4S) or 16 16-bit (2 x 8H) Montgomery
4930 // multiplications in parallel
4931 //
4932
4933 // See the montMul() method of the sun.security.provider.ML_DSA
4934 // class.
4935 //
4936 // Computes 4x4S results or 8x8H results
4937 // a = b * c * 2^MONT_R_BITS mod MONT_Q
4938 // Inputs: vb, vc - 4x4S or 4x8H vector register sequences
4939 // vq - 2x4S or 2x8H constants <MONT_Q, MONT_Q_INV_MOD_R>
4940 // Temps: vtmp - 4x4S or 4x8H vector sequence trashed after call
4941 // Outputs: va - 4x4S or 4x8H vector register sequences
4942 // vb, vc, vtmp and vq must all be disjoint
4943 // va must be disjoint from all other inputs/temps or must equal vc
4944 // va must have a non-zero delta i.e. it must not be a constant vseq.
4945 // n.b. MONT_R_BITS is 16 or 32, so the right shift by it is implicit.
4946 void vs_montmul2(const VSeq<2>& va, const VSeq<2>& vb, const VSeq<2>& vc,
4947 Assembler::SIMD_Arrangement T,
4948 const VSeq<2>& vtmp, const VSeq<2>& vq) {
4949 assert (T == __ T4S || T == __ T8H, "invalid arrangement for montmul");
4950 assert(vs_disjoint(vb, vc), "vb and vc overlap");
4951 assert(vs_disjoint(vb, vq), "vb and vq overlap");
4952 assert(vs_disjoint(vb, vtmp), "vb and vtmp overlap");
4953
4954 assert(vs_disjoint(vc, vq), "vc and vq overlap");
4955 assert(vs_disjoint(vc, vtmp), "vc and vtmp overlap");
4956
4957 assert(vs_disjoint(vq, vtmp), "vq and vtmp overlap");
4958
4959 assert(vs_disjoint(va, vc) || vs_same(va, vc), "va and vc neither disjoint nor equal");
4960 assert(vs_disjoint(va, vb), "va and vb overlap");
4961 assert(vs_disjoint(va, vq), "va and vq overlap");
4962 assert(vs_disjoint(va, vtmp), "va and vtmp overlap");
4963 assert(!va.is_constant(), "output vector must identify 2 different registers");
4964
4965 // schedule 2 streams of instructions across the vector sequences
4966 for (int i = 0; i < 2; i++) {
4967 __ sqdmulh(vtmp[i], T, vb[i], vc[i]); // aHigh = hi32(2 * b * c)
4968 __ mulv(va[i], T, vb[i], vc[i]); // aLow = lo32(b * c)
4969 }
4970
4971 for (int i = 0; i < 2; i++) {
4972 __ mulv(va[i], T, va[i], vq[0]); // m = aLow * qinv
4973 }
4974
4975 for (int i = 0; i < 2; i++) {
4976 __ sqdmulh(va[i], T, va[i], vq[1]); // n = hi32(2 * m * q)
4977 }
4978
4979 for (int i = 0; i < 2; i++) {
4980 __ shsubv(va[i], T, vtmp[i], va[i]); // a = (aHigh - n) / 2
4981 }
4982 }
4983
4984 // Perform 16 16-bit Montgomery multiplications in parallel.
4985 void kyber_montmul16(const VSeq<2>& va, const VSeq<2>& vb, const VSeq<2>& vc,
4986 const VSeq<2>& vtmp, const VSeq<2>& vq) {
4987 // Use the helper routine to schedule a 2x8H Montgomery multiply.
4988 // It will assert that the register use is valid
4989 vs_montmul2(va, vb, vc, __ T8H, vtmp, vq);
4990 }
4991
4992 // Perform 32 16-bit Montgomery multiplications in parallel.
4993 void kyber_montmul32(const VSeq<4>& va, const VSeq<4>& vb, const VSeq<4>& vc,
4994 const VSeq<4>& vtmp, const VSeq<2>& vq) {
4995 // Use the helper routine to schedule a 4x8H Montgomery multiply.
4996 // It will assert that the register use is valid
4997 vs_montmul4(va, vb, vc, __ T8H, vtmp, vq);
4998 }
4999
5000 // Perform 64 16-bit Montgomery multiplications in parallel.
5001 void kyber_montmul64(const VSeq<8>& va, const VSeq<8>& vb, const VSeq<8>& vc,
5002 const VSeq<4>& vtmp, const VSeq<2>& vq) {
5003 // Schedule two successive 4x8H multiplies via the montmul helper
5004 // on the front and back halves of va, vb and vc. The helper will
5005 // assert that the register use has no overlap conflicts on each
5006 // individual call but we also need to ensure that the necessary
5007 // disjoint/equality constraints are met across both calls.
5008
5009 // vb, vc, vtmp and vq must be disjoint. va must either be
5010 // disjoint from all other registers or equal vc
5011
5012 assert(vs_disjoint(vb, vc), "vb and vc overlap");
5013 assert(vs_disjoint(vb, vq), "vb and vq overlap");
5014 assert(vs_disjoint(vb, vtmp), "vb and vtmp overlap");
5015
5016 assert(vs_disjoint(vc, vq), "vc and vq overlap");
5017 assert(vs_disjoint(vc, vtmp), "vc and vtmp overlap");
5018
5019 assert(vs_disjoint(vq, vtmp), "vq and vtmp overlap");
5020
5021 assert(vs_disjoint(va, vc) || vs_same(va, vc), "va and vc neither disjoint nor equal");
5022 assert(vs_disjoint(va, vb), "va and vb overlap");
5023 assert(vs_disjoint(va, vq), "va and vq overlap");
5024 assert(vs_disjoint(va, vtmp), "va and vtmp overlap");
5025
5026 // we multiply the front and back halves of each sequence 4 at a
5027 // time because
5028 //
5029 // 1) we are currently only able to get 4-way instruction
5030 // parallelism at best
5031 //
5032 // 2) we need registers for the constants in vq and temporary
5033 // scratch registers to hold intermediate results so vtmp can only
5034 // be a VSeq<4> which means we only have 4 scratch slots
5035
5036 vs_montmul4(vs_front(va), vs_front(vb), vs_front(vc), __ T8H, vtmp, vq);
5037 vs_montmul4(vs_back(va), vs_back(vb), vs_back(vc), __ T8H, vtmp, vq);
5038 }
5039
5040 void kyber_montmul32_sub_add(const VSeq<4>& va0, const VSeq<4>& va1,
5041 const VSeq<4>& vc,
5042 const VSeq<4>& vtmp,
5043 const VSeq<2>& vq) {
5044 // compute a = montmul(a1, c)
5045 kyber_montmul32(vc, va1, vc, vtmp, vq);
5046 // ouptut a1 = a0 - a
5047 vs_subv(va1, __ T8H, va0, vc);
5048 // and a0 = a0 + a
5049 vs_addv(va0, __ T8H, va0, vc);
5050 }
5051
5052 void kyber_sub_add_montmul32(const VSeq<4>& va0, const VSeq<4>& va1,
5053 const VSeq<4>& vb,
5054 const VSeq<4>& vtmp1,
5055 const VSeq<4>& vtmp2,
5056 const VSeq<2>& vq) {
5057 // compute c = a0 - a1
5058 vs_subv(vtmp1, __ T8H, va0, va1);
5059 // output a0 = a0 + a1
5060 vs_addv(va0, __ T8H, va0, va1);
5061 // output a1 = b montmul c
5062 kyber_montmul32(va1, vtmp1, vb, vtmp2, vq);
5063 }
5064
5065 void load64shorts(const VSeq<8>& v, Register shorts) {
5066 vs_ldpq_post(v, shorts);
5067 }
5068
5069 void load32shorts(const VSeq<4>& v, Register shorts) {
5070 vs_ldpq_post(v, shorts);
5071 }
5072
5073 void store64shorts(VSeq<8> v, Register tmpAddr) {
5074 vs_stpq_post(v, tmpAddr);
5075 }
5076
5077 // Kyber NTT function.
5078 // Implements
5079 // static int implKyberNtt(short[] poly, short[] ntt_zetas) {}
5080 //
5081 // coeffs (short[256]) = c_rarg0
5082 // ntt_zetas (short[256]) = c_rarg1
5083 address generate_kyberNtt() {
5084
5085 __ align(CodeEntryAlignment);
5086 StubGenStubId stub_id = StubGenStubId::kyberNtt_id;
5087 StubCodeMark mark(this, stub_id);
5088 address start = __ pc();
5089 __ enter();
5090
5091 const Register coeffs = c_rarg0;
5092 const Register zetas = c_rarg1;
5093
5094 const Register kyberConsts = r10;
5095 const Register tmpAddr = r11;
5096
5097 VSeq<8> vs1(0), vs2(16), vs3(24); // 3 sets of 8x8H inputs/outputs
5098 VSeq<4> vtmp = vs_front(vs3); // n.b. tmp registers overlap vs3
5099 VSeq<2> vq(30); // n.b. constants overlap vs3
5100
5101 __ lea(kyberConsts, ExternalAddress((address) StubRoutines::aarch64::_kyberConsts));
5102 // load the montmul constants
5103 vs_ldpq(vq, kyberConsts);
5104
5105 // Each level corresponds to an iteration of the outermost loop of the
5106 // Java method seilerNTT(int[] coeffs). There are some differences
5107 // from what is done in the seilerNTT() method, though:
5108 // 1. The computation is using 16-bit signed values, we do not convert them
5109 // to ints here.
5110 // 2. The zetas are delivered in a bigger array, 128 zetas are stored in
5111 // this array for each level, it is easier that way to fill up the vector
5112 // registers.
5113 // 3. In the seilerNTT() method we use R = 2^20 for the Montgomery
5114 // multiplications (this is because that way there should not be any
5115 // overflow during the inverse NTT computation), here we usr R = 2^16 so
5116 // that we can use the 16-bit arithmetic in the vector unit.
5117 //
5118 // On each level, we fill up the vector registers in such a way that the
5119 // array elements that need to be multiplied by the zetas go into one
5120 // set of vector registers while the corresponding ones that don't need to
5121 // be multiplied, go into another set.
5122 // We can do 32 Montgomery multiplications in parallel, using 12 vector
5123 // registers interleaving the steps of 4 identical computations,
5124 // each done on 8 16-bit values per register.
5125
5126 // At levels 0-3 the coefficients multiplied by or added/subtracted
5127 // to the zetas occur in discrete blocks whose size is some multiple
5128 // of 32.
5129
5130 // level 0
5131 __ add(tmpAddr, coeffs, 256);
5132 load64shorts(vs1, tmpAddr);
5133 load64shorts(vs2, zetas);
5134 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5135 __ add(tmpAddr, coeffs, 0);
5136 load64shorts(vs1, tmpAddr);
5137 vs_subv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5138 vs_addv(vs1, __ T8H, vs1, vs2);
5139 __ add(tmpAddr, coeffs, 0);
5140 vs_stpq_post(vs1, tmpAddr);
5141 __ add(tmpAddr, coeffs, 256);
5142 vs_stpq_post(vs3, tmpAddr);
5143 // restore montmul constants
5144 vs_ldpq(vq, kyberConsts);
5145 load64shorts(vs1, tmpAddr);
5146 load64shorts(vs2, zetas);
5147 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5148 __ add(tmpAddr, coeffs, 128);
5149 load64shorts(vs1, tmpAddr);
5150 vs_subv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5151 vs_addv(vs1, __ T8H, vs1, vs2);
5152 __ add(tmpAddr, coeffs, 128);
5153 store64shorts(vs1, tmpAddr);
5154 __ add(tmpAddr, coeffs, 384);
5155 store64shorts(vs3, tmpAddr);
5156
5157 // level 1
5158 // restore montmul constants
5159 vs_ldpq(vq, kyberConsts);
5160 __ add(tmpAddr, coeffs, 128);
5161 load64shorts(vs1, tmpAddr);
5162 load64shorts(vs2, zetas);
5163 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5164 __ add(tmpAddr, coeffs, 0);
5165 load64shorts(vs1, tmpAddr);
5166 vs_subv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5167 vs_addv(vs1, __ T8H, vs1, vs2);
5168 __ add(tmpAddr, coeffs, 0);
5169 store64shorts(vs1, tmpAddr);
5170 store64shorts(vs3, tmpAddr);
5171 vs_ldpq(vq, kyberConsts);
5172 __ add(tmpAddr, coeffs, 384);
5173 load64shorts(vs1, tmpAddr);
5174 load64shorts(vs2, zetas);
5175 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5176 __ add(tmpAddr, coeffs, 256);
5177 load64shorts(vs1, tmpAddr);
5178 vs_subv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5179 vs_addv(vs1, __ T8H, vs1, vs2);
5180 __ add(tmpAddr, coeffs, 256);
5181 store64shorts(vs1, tmpAddr);
5182 store64shorts(vs3, tmpAddr);
5183
5184 // level 2
5185 vs_ldpq(vq, kyberConsts);
5186 int offsets1[4] = { 0, 32, 128, 160 };
5187 vs_ldpq_indexed(vs1, coeffs, 64, offsets1);
5188 load64shorts(vs2, zetas);
5189 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5190 vs_ldpq_indexed(vs1, coeffs, 0, offsets1);
5191 // kyber_subv_addv64();
5192 vs_subv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5193 vs_addv(vs1, __ T8H, vs1, vs2);
5194 __ add(tmpAddr, coeffs, 0);
5195 vs_stpq_post(vs_front(vs1), tmpAddr);
5196 vs_stpq_post(vs_front(vs3), tmpAddr);
5197 vs_stpq_post(vs_back(vs1), tmpAddr);
5198 vs_stpq_post(vs_back(vs3), tmpAddr);
5199 vs_ldpq(vq, kyberConsts);
5200 vs_ldpq_indexed(vs1, tmpAddr, 64, offsets1);
5201 load64shorts(vs2, zetas);
5202 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5203 vs_ldpq_indexed(vs1, coeffs, 256, offsets1);
5204 // kyber_subv_addv64();
5205 vs_subv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5206 vs_addv(vs1, __ T8H, vs1, vs2);
5207 __ add(tmpAddr, coeffs, 256);
5208 vs_stpq_post(vs_front(vs1), tmpAddr);
5209 vs_stpq_post(vs_front(vs3), tmpAddr);
5210 vs_stpq_post(vs_back(vs1), tmpAddr);
5211 vs_stpq_post(vs_back(vs3), tmpAddr);
5212
5213 // level 3
5214 vs_ldpq(vq, kyberConsts);
5215 int offsets2[4] = { 0, 64, 128, 192 };
5216 vs_ldpq_indexed(vs1, coeffs, 32, offsets2);
5217 load64shorts(vs2, zetas);
5218 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5219 vs_ldpq_indexed(vs1, coeffs, 0, offsets2);
5220 vs_subv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5221 vs_addv(vs1, __ T8H, vs1, vs2);
5222 vs_stpq_indexed(vs1, coeffs, 0, offsets2);
5223 vs_stpq_indexed(vs3, coeffs, 32, offsets2);
5224
5225 vs_ldpq(vq, kyberConsts);
5226 vs_ldpq_indexed(vs1, coeffs, 256 + 32, offsets2);
5227 load64shorts(vs2, zetas);
5228 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5229 vs_ldpq_indexed(vs1, coeffs, 256, offsets2);
5230 vs_subv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5231 vs_addv(vs1, __ T8H, vs1, vs2);
5232 vs_stpq_indexed(vs1, coeffs, 256, offsets2);
5233 vs_stpq_indexed(vs3, coeffs, 256 + 32, offsets2);
5234
5235 // level 4
5236 // At level 4 coefficients occur in 8 discrete blocks of size 16
5237 // so they are loaded using employing an ldr at 8 distinct offsets.
5238
5239 vs_ldpq(vq, kyberConsts);
5240 int offsets3[8] = { 0, 32, 64, 96, 128, 160, 192, 224 };
5241 vs_ldr_indexed(vs1, __ Q, coeffs, 16, offsets3);
5242 load64shorts(vs2, zetas);
5243 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5244 vs_ldr_indexed(vs1, __ Q, coeffs, 0, offsets3);
5245 vs_subv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5246 vs_addv(vs1, __ T8H, vs1, vs2);
5247 vs_str_indexed(vs1, __ Q, coeffs, 0, offsets3);
5248 vs_str_indexed(vs3, __ Q, coeffs, 16, offsets3);
5249
5250 vs_ldpq(vq, kyberConsts);
5251 vs_ldr_indexed(vs1, __ Q, coeffs, 256 + 16, offsets3);
5252 load64shorts(vs2, zetas);
5253 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5254 vs_ldr_indexed(vs1, __ Q, coeffs, 256, offsets3);
5255 vs_subv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5256 vs_addv(vs1, __ T8H, vs1, vs2);
5257 vs_str_indexed(vs1, __ Q, coeffs, 256, offsets3);
5258 vs_str_indexed(vs3, __ Q, coeffs, 256 + 16, offsets3);
5259
5260 // level 5
5261 // At level 5 related coefficients occur in discrete blocks of size 8 so
5262 // need to be loaded interleaved using an ld2 operation with arrangement 2D.
5263
5264 vs_ldpq(vq, kyberConsts);
5265 int offsets4[4] = { 0, 32, 64, 96 };
5266 vs_ld2_indexed(vs1, __ T2D, coeffs, tmpAddr, 0, offsets4);
5267 load32shorts(vs_front(vs2), zetas);
5268 kyber_montmul32_sub_add(vs_even(vs1), vs_odd(vs1), vs_front(vs2), vtmp, vq);
5269 vs_st2_indexed(vs1, __ T2D, coeffs, tmpAddr, 0, offsets4);
5270 vs_ld2_indexed(vs1, __ T2D, coeffs, tmpAddr, 128, offsets4);
5271 load32shorts(vs_front(vs2), zetas);
5272 kyber_montmul32_sub_add(vs_even(vs1), vs_odd(vs1), vs_front(vs2), vtmp, vq);
5273 vs_st2_indexed(vs1, __ T2D, coeffs, tmpAddr, 128, offsets4);
5274 vs_ld2_indexed(vs1, __ T2D, coeffs, tmpAddr, 256, offsets4);
5275 load32shorts(vs_front(vs2), zetas);
5276 kyber_montmul32_sub_add(vs_even(vs1), vs_odd(vs1), vs_front(vs2), vtmp, vq);
5277 vs_st2_indexed(vs1, __ T2D, coeffs, tmpAddr, 256, offsets4);
5278
5279 vs_ld2_indexed(vs1, __ T2D, coeffs, tmpAddr, 384, offsets4);
5280 load32shorts(vs_front(vs2), zetas);
5281 kyber_montmul32_sub_add(vs_even(vs1), vs_odd(vs1), vs_front(vs2), vtmp, vq);
5282 vs_st2_indexed(vs1, __ T2D, coeffs, tmpAddr, 384, offsets4);
5283
5284 // level 6
5285 // At level 6 related coefficients occur in discrete blocks of size 4 so
5286 // need to be loaded interleaved using an ld2 operation with arrangement 4S.
5287
5288 vs_ld2_indexed(vs1, __ T4S, coeffs, tmpAddr, 0, offsets4);
5289 load32shorts(vs_front(vs2), zetas);
5290 kyber_montmul32_sub_add(vs_even(vs1), vs_odd(vs1), vs_front(vs2), vtmp, vq);
5291 vs_st2_indexed(vs1, __ T4S, coeffs, tmpAddr, 0, offsets4);
5292 vs_ld2_indexed(vs1, __ T4S, coeffs, tmpAddr, 128, offsets4);
5293 // __ ldpq(v18, v19, __ post(zetas, 32));
5294 load32shorts(vs_front(vs2), zetas);
5295 kyber_montmul32_sub_add(vs_even(vs1), vs_odd(vs1), vs_front(vs2), vtmp, vq);
5296 vs_st2_indexed(vs1, __ T4S, coeffs, tmpAddr, 128, offsets4);
5297
5298 vs_ld2_indexed(vs1, __ T4S, coeffs, tmpAddr, 256, offsets4);
5299 load32shorts(vs_front(vs2), zetas);
5300 kyber_montmul32_sub_add(vs_even(vs1), vs_odd(vs1), vs_front(vs2), vtmp, vq);
5301 vs_st2_indexed(vs1, __ T4S, coeffs, tmpAddr, 256, offsets4);
5302
5303 vs_ld2_indexed(vs1, __ T4S, coeffs, tmpAddr, 384, offsets4);
5304 load32shorts(vs_front(vs2), zetas);
5305 kyber_montmul32_sub_add(vs_even(vs1), vs_odd(vs1), vs_front(vs2), vtmp, vq);
5306 vs_st2_indexed(vs1, __ T4S, coeffs, tmpAddr, 384, offsets4);
5307
5308 __ leave(); // required for proper stackwalking of RuntimeStub frame
5309 __ mov(r0, zr); // return 0
5310 __ ret(lr);
5311
5312 return start;
5313 }
5314
5315 // Kyber Inverse NTT function
5316 // Implements
5317 // static int implKyberInverseNtt(short[] poly, short[] zetas) {}
5318 //
5319 // coeffs (short[256]) = c_rarg0
5320 // ntt_zetas (short[256]) = c_rarg1
5321 address generate_kyberInverseNtt() {
5322
5323 __ align(CodeEntryAlignment);
5324 StubGenStubId stub_id = StubGenStubId::kyberInverseNtt_id;
5325 StubCodeMark mark(this, stub_id);
5326 address start = __ pc();
5327 __ enter();
5328
5329 const Register coeffs = c_rarg0;
5330 const Register zetas = c_rarg1;
5331
5332 const Register kyberConsts = r10;
5333 const Register tmpAddr = r11;
5334 const Register tmpAddr2 = c_rarg2;
5335
5336 VSeq<8> vs1(0), vs2(16), vs3(24); // 3 sets of 8x8H inputs/outputs
5337 VSeq<4> vtmp = vs_front(vs3); // n.b. tmp registers overlap vs3
5338 VSeq<2> vq(30); // n.b. constants overlap vs3
5339
5340 __ lea(kyberConsts,
5341 ExternalAddress((address) StubRoutines::aarch64::_kyberConsts));
5342
5343 // level 0
5344 // At level 0 related coefficients occur in discrete blocks of size 4 so
5345 // need to be loaded interleaved using an ld2 operation with arrangement 4S.
5346
5347 vs_ldpq(vq, kyberConsts);
5348 int offsets4[4] = { 0, 32, 64, 96 };
5349 vs_ld2_indexed(vs1, __ T4S, coeffs, tmpAddr, 0, offsets4);
5350 load32shorts(vs_front(vs2), zetas);
5351 kyber_sub_add_montmul32(vs_even(vs1), vs_odd(vs1),
5352 vs_front(vs2), vs_back(vs2), vtmp, vq);
5353 vs_st2_indexed(vs1, __ T4S, coeffs, tmpAddr, 0, offsets4);
5354 vs_ld2_indexed(vs1, __ T4S, coeffs, tmpAddr, 128, offsets4);
5355 load32shorts(vs_front(vs2), zetas);
5356 kyber_sub_add_montmul32(vs_even(vs1), vs_odd(vs1),
5357 vs_front(vs2), vs_back(vs2), vtmp, vq);
5358 vs_st2_indexed(vs1, __ T4S, coeffs, tmpAddr, 128, offsets4);
5359 vs_ld2_indexed(vs1, __ T4S, coeffs, tmpAddr, 256, offsets4);
5360 load32shorts(vs_front(vs2), zetas);
5361 kyber_sub_add_montmul32(vs_even(vs1), vs_odd(vs1),
5362 vs_front(vs2), vs_back(vs2), vtmp, vq);
5363 vs_st2_indexed(vs1, __ T4S, coeffs, tmpAddr, 256, offsets4);
5364 vs_ld2_indexed(vs1, __ T4S, coeffs, tmpAddr, 384, offsets4);
5365 load32shorts(vs_front(vs2), zetas);
5366 kyber_sub_add_montmul32(vs_even(vs1), vs_odd(vs1),
5367 vs_front(vs2), vs_back(vs2), vtmp, vq);
5368 vs_st2_indexed(vs1, __ T4S, coeffs, tmpAddr, 384, offsets4);
5369
5370 // level 1
5371 // At level 1 related coefficients occur in discrete blocks of size 8 so
5372 // need to be loaded interleaved using an ld2 operation with arrangement 2D.
5373
5374 vs_ld2_indexed(vs1, __ T2D, coeffs, tmpAddr, 0, offsets4);
5375 load32shorts(vs_front(vs2), zetas);
5376 kyber_sub_add_montmul32(vs_even(vs1), vs_odd(vs1),
5377 vs_front(vs2), vs_back(vs2), vtmp, vq);
5378 vs_st2_indexed(vs1, __ T2D, coeffs, tmpAddr, 0, offsets4);
5379 vs_ld2_indexed(vs1, __ T2D, coeffs, tmpAddr, 128, offsets4);
5380 load32shorts(vs_front(vs2), zetas);
5381 kyber_sub_add_montmul32(vs_even(vs1), vs_odd(vs1),
5382 vs_front(vs2), vs_back(vs2), vtmp, vq);
5383 vs_st2_indexed(vs1, __ T2D, coeffs, tmpAddr, 128, offsets4);
5384
5385 vs_ld2_indexed(vs1, __ T2D, coeffs, tmpAddr, 256, offsets4);
5386 load32shorts(vs_front(vs2), zetas);
5387 kyber_sub_add_montmul32(vs_even(vs1), vs_odd(vs1),
5388 vs_front(vs2), vs_back(vs2), vtmp, vq);
5389 vs_st2_indexed(vs1, __ T2D, coeffs, tmpAddr, 256, offsets4);
5390 vs_ld2_indexed(vs1, __ T2D, coeffs, tmpAddr, 384, offsets4);
5391 load32shorts(vs_front(vs2), zetas);
5392 kyber_sub_add_montmul32(vs_even(vs1), vs_odd(vs1),
5393 vs_front(vs2), vs_back(vs2), vtmp, vq);
5394 vs_st2_indexed(vs1, __ T2D, coeffs, tmpAddr, 384, offsets4);
5395
5396 // level 2
5397 // At level 2 coefficients occur in 8 discrete blocks of size 16
5398 // so they are loaded using employing an ldr at 8 distinct offsets.
5399
5400 int offsets3[8] = { 0, 32, 64, 96, 128, 160, 192, 224 };
5401 vs_ldr_indexed(vs1, __ Q, coeffs, 0, offsets3);
5402 vs_ldr_indexed(vs2, __ Q, coeffs, 16, offsets3);
5403 vs_addv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5404 vs_subv(vs1, __ T8H, vs1, vs2);
5405 vs_str_indexed(vs3, __ Q, coeffs, 0, offsets3);
5406 load64shorts(vs2, zetas);
5407 vs_ldpq(vq, kyberConsts);
5408 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5409 vs_str_indexed(vs2, __ Q, coeffs, 16, offsets3);
5410
5411 vs_ldr_indexed(vs1, __ Q, coeffs, 256, offsets3);
5412 vs_ldr_indexed(vs2, __ Q, coeffs, 256 + 16, offsets3);
5413 vs_addv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5414 vs_subv(vs1, __ T8H, vs1, vs2);
5415 vs_str_indexed(vs3, __ Q, coeffs, 256, offsets3);
5416 load64shorts(vs2, zetas);
5417 vs_ldpq(vq, kyberConsts);
5418 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5419 vs_str_indexed(vs2, __ Q, coeffs, 256 + 16, offsets3);
5420
5421 // Barrett reduction at indexes where overflow may happen
5422
5423 // load q and the multiplier for the Barrett reduction
5424 __ add(tmpAddr, kyberConsts, 16);
5425 vs_ldpq(vq, tmpAddr);
5426
5427 VSeq<8> vq1 = VSeq<8>(vq[0], 0); // 2 constant 8 sequences
5428 VSeq<8> vq2 = VSeq<8>(vq[1], 0); // for above two kyber constants
5429 VSeq<8> vq3 = VSeq<8>(v29, 0); // 3rd sequence for const montmul
5430 vs_ldr_indexed(vs1, __ Q, coeffs, 0, offsets3);
5431 vs_sqdmulh(vs2, __ T8H, vs1, vq2);
5432 vs_sshr(vs2, __ T8H, vs2, 11);
5433 vs_mlsv(vs1, __ T8H, vs2, vq1);
5434 vs_str_indexed(vs1, __ Q, coeffs, 0, offsets3);
5435 vs_ldr_indexed(vs1, __ Q, coeffs, 256, offsets3);
5436 vs_sqdmulh(vs2, __ T8H, vs1, vq2);
5437 vs_sshr(vs2, __ T8H, vs2, 11);
5438 vs_mlsv(vs1, __ T8H, vs2, vq1);
5439 vs_str_indexed(vs1, __ Q, coeffs, 256, offsets3);
5440
5441 // level 3
5442 // From level 3 upwards coefficients occur in discrete blocks whose size is
5443 // some multiple of 32 so can be loaded using ldpq and suitable indexes.
5444
5445 int offsets2[4] = { 0, 64, 128, 192 };
5446 vs_ldpq_indexed(vs1, coeffs, 0, offsets2);
5447 vs_ldpq_indexed(vs2, coeffs, 32, offsets2);
5448 vs_addv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5449 vs_subv(vs1, __ T8H, vs1, vs2);
5450 vs_stpq_indexed(vs3, coeffs, 0, offsets2);
5451 load64shorts(vs2, zetas);
5452 vs_ldpq(vq, kyberConsts);
5453 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5454 vs_stpq_indexed(vs2, coeffs, 32, offsets2);
5455
5456 vs_ldpq_indexed(vs1, coeffs, 256, offsets2);
5457 vs_ldpq_indexed(vs2, coeffs, 256 + 32, offsets2);
5458 vs_addv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5459 vs_subv(vs1, __ T8H, vs1, vs2);
5460 vs_stpq_indexed(vs3, coeffs, 256, offsets2);
5461 load64shorts(vs2, zetas);
5462 vs_ldpq(vq, kyberConsts);
5463 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5464 vs_stpq_indexed(vs2, coeffs, 256 + 32, offsets2);
5465
5466 // level 4
5467
5468 int offsets1[4] = { 0, 32, 128, 160 };
5469 vs_ldpq_indexed(vs1, coeffs, 0, offsets1);
5470 vs_ldpq_indexed(vs2, coeffs, 64, offsets1);
5471 vs_addv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5472 vs_subv(vs1, __ T8H, vs1, vs2);
5473 vs_stpq_indexed(vs3, coeffs, 0, offsets1);
5474 load64shorts(vs2, zetas);
5475 vs_ldpq(vq, kyberConsts);
5476 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5477 vs_stpq_indexed(vs2, coeffs, 64, offsets1);
5478
5479 vs_ldpq_indexed(vs1, coeffs, 256, offsets1);
5480 vs_ldpq_indexed(vs2, coeffs, 256 + 64, offsets1);
5481 vs_addv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5482 vs_subv(vs1, __ T8H, vs1, vs2);
5483 vs_stpq_indexed(vs3, coeffs, 256, offsets1);
5484 load64shorts(vs2, zetas);
5485 vs_ldpq(vq, kyberConsts);
5486 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5487 vs_stpq_indexed(vs2, coeffs, 256 + 64, offsets1);
5488
5489 // level 5
5490
5491 __ add(tmpAddr, coeffs, 0);
5492 load64shorts(vs1, tmpAddr);
5493 __ add(tmpAddr, coeffs, 128);
5494 load64shorts(vs2, tmpAddr);
5495 vs_addv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5496 vs_subv(vs1, __ T8H, vs1, vs2);
5497 __ add(tmpAddr, coeffs, 0);
5498 store64shorts(vs3, tmpAddr);
5499 load64shorts(vs2, zetas);
5500 vs_ldpq(vq, kyberConsts);
5501 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5502 __ add(tmpAddr, coeffs, 128);
5503 store64shorts(vs2, tmpAddr);
5504
5505 load64shorts(vs1, tmpAddr);
5506 __ add(tmpAddr, coeffs, 384);
5507 load64shorts(vs2, tmpAddr);
5508 vs_addv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5509 vs_subv(vs1, __ T8H, vs1, vs2);
5510 __ add(tmpAddr, coeffs, 256);
5511 store64shorts(vs3, tmpAddr);
5512 load64shorts(vs2, zetas);
5513 vs_ldpq(vq, kyberConsts);
5514 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5515 __ add(tmpAddr, coeffs, 384);
5516 store64shorts(vs2, tmpAddr);
5517
5518 // Barrett reduction at indexes where overflow may happen
5519
5520 // load q and the multiplier for the Barrett reduction
5521 __ add(tmpAddr, kyberConsts, 16);
5522 vs_ldpq(vq, tmpAddr);
5523
5524 int offsets0[2] = { 0, 256 };
5525 vs_ldpq_indexed(vs_front(vs1), coeffs, 0, offsets0);
5526 vs_sqdmulh(vs2, __ T8H, vs1, vq2);
5527 vs_sshr(vs2, __ T8H, vs2, 11);
5528 vs_mlsv(vs1, __ T8H, vs2, vq1);
5529 vs_stpq_indexed(vs_front(vs1), coeffs, 0, offsets0);
5530
5531 // level 6
5532
5533 __ add(tmpAddr, coeffs, 0);
5534 load64shorts(vs1, tmpAddr);
5535 __ add(tmpAddr, coeffs, 256);
5536 load64shorts(vs2, tmpAddr);
5537 vs_addv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5538 vs_subv(vs1, __ T8H, vs1, vs2);
5539 __ add(tmpAddr, coeffs, 0);
5540 store64shorts(vs3, tmpAddr);
5541 load64shorts(vs2, zetas);
5542 vs_ldpq(vq, kyberConsts);
5543 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5544 __ add(tmpAddr, coeffs, 256);
5545 store64shorts(vs2, tmpAddr);
5546
5547 __ add(tmpAddr, coeffs, 128);
5548 load64shorts(vs1, tmpAddr);
5549 __ add(tmpAddr, coeffs, 384);
5550 load64shorts(vs2, tmpAddr);
5551 vs_addv(vs3, __ T8H, vs1, vs2); // n.b. trashes vq
5552 vs_subv(vs1, __ T8H, vs1, vs2);
5553 __ add(tmpAddr, coeffs, 128);
5554 store64shorts(vs3, tmpAddr);
5555 load64shorts(vs2, zetas);
5556 vs_ldpq(vq, kyberConsts);
5557 kyber_montmul64(vs2, vs1, vs2, vtmp, vq);
5558 __ add(tmpAddr, coeffs, 384);
5559 store64shorts(vs2, tmpAddr);
5560
5561 // multiply by 2^-n
5562
5563 // load toMont(2^-n mod q)
5564 __ add(tmpAddr, kyberConsts, 48);
5565 __ ldr(v29, __ Q, tmpAddr);
5566
5567 vs_ldpq(vq, kyberConsts);
5568 __ add(tmpAddr, coeffs, 0);
5569 load64shorts(vs1, tmpAddr);
5570 kyber_montmul64(vs2, vs1, vq3, vtmp, vq);
5571 __ add(tmpAddr, coeffs, 0);
5572 store64shorts(vs2, tmpAddr);
5573
5574 // now tmpAddr contains coeffs + 128 because store64shorts adjusted it so
5575 load64shorts(vs1, tmpAddr);
5576 kyber_montmul64(vs2, vs1, vq3, vtmp, vq);
5577 __ add(tmpAddr, coeffs, 128);
5578 store64shorts(vs2, tmpAddr);
5579
5580 // now tmpAddr contains coeffs + 256
5581 load64shorts(vs1, tmpAddr);
5582 kyber_montmul64(vs2, vs1, vq3, vtmp, vq);
5583 __ add(tmpAddr, coeffs, 256);
5584 store64shorts(vs2, tmpAddr);
5585
5586 // now tmpAddr contains coeffs + 384
5587 load64shorts(vs1, tmpAddr);
5588 kyber_montmul64(vs2, vs1, vq3, vtmp, vq);
5589 __ add(tmpAddr, coeffs, 384);
5590 store64shorts(vs2, tmpAddr);
5591
5592 __ leave(); // required for proper stackwalking of RuntimeStub frame
5593 __ mov(r0, zr); // return 0
5594 __ ret(lr);
5595
5596 return start;
5597 }
5598
5599 // Kyber multiply polynomials in the NTT domain.
5600 // Implements
5601 // static int implKyberNttMult(
5602 // short[] result, short[] ntta, short[] nttb, short[] zetas) {}
5603 //
5604 // result (short[256]) = c_rarg0
5605 // ntta (short[256]) = c_rarg1
5606 // nttb (short[256]) = c_rarg2
5607 // zetas (short[128]) = c_rarg3
5608 address generate_kyberNttMult() {
5609
5610 __ align(CodeEntryAlignment);
5611 StubGenStubId stub_id = StubGenStubId::kyberNttMult_id;
5612 StubCodeMark mark(this, stub_id);
5613 address start = __ pc();
5614 __ enter();
5615
5616 const Register result = c_rarg0;
5617 const Register ntta = c_rarg1;
5618 const Register nttb = c_rarg2;
5619 const Register zetas = c_rarg3;
5620
5621 const Register kyberConsts = r10;
5622 const Register limit = r11;
5623
5624 VSeq<4> vs1(0), vs2(4); // 4 sets of 8x8H inputs/outputs/tmps
5625 VSeq<4> vs3(16), vs4(20);
5626 VSeq<2> vq(30); // pair of constants for montmul: q, qinv
5627 VSeq<2> vz(28); // pair of zetas
5628 VSeq<4> vc(27, 0); // constant sequence for montmul: montRSquareModQ
5629
5630 __ lea(kyberConsts,
5631 ExternalAddress((address) StubRoutines::aarch64::_kyberConsts));
5632
5633 Label kyberNttMult_loop;
5634
5635 __ add(limit, result, 512);
5636
5637 // load q and qinv
5638 vs_ldpq(vq, kyberConsts);
5639
5640 // load R^2 mod q (to convert back from Montgomery representation)
5641 __ add(kyberConsts, kyberConsts, 64);
5642 __ ldr(v27, __ Q, kyberConsts);
5643
5644 __ BIND(kyberNttMult_loop);
5645
5646 // load 16 zetas
5647 vs_ldpq_post(vz, zetas);
5648
5649 // load 2 sets of 32 coefficients from the two input arrays
5650 // interleaved as shorts. i.e. pairs of shorts adjacent in memory
5651 // are striped across pairs of vector registers
5652 vs_ld2_post(vs_front(vs1), __ T8H, ntta); // <a0, a1> x 8H
5653 vs_ld2_post(vs_back(vs1), __ T8H, nttb); // <b0, b1> x 8H
5654 vs_ld2_post(vs_front(vs4), __ T8H, ntta); // <a2, a3> x 8H
5655 vs_ld2_post(vs_back(vs4), __ T8H, nttb); // <b2, b3> x 8H
5656
5657 // compute 4 montmul cross-products for pairs (a0,a1) and (b0,b1)
5658 // i.e. montmul the first and second halves of vs1 in order and
5659 // then with one sequence reversed storing the two results in vs3
5660 //
5661 // vs3[0] <- montmul(a0, b0)
5662 // vs3[1] <- montmul(a1, b1)
5663 // vs3[2] <- montmul(a0, b1)
5664 // vs3[3] <- montmul(a1, b0)
5665 kyber_montmul16(vs_front(vs3), vs_front(vs1), vs_back(vs1), vs_front(vs2), vq);
5666 kyber_montmul16(vs_back(vs3),
5667 vs_front(vs1), vs_reverse(vs_back(vs1)), vs_back(vs2), vq);
5668
5669 // compute 4 montmul cross-products for pairs (a2,a3) and (b2,b3)
5670 // i.e. montmul the first and second halves of vs4 in order and
5671 // then with one sequence reversed storing the two results in vs1
5672 //
5673 // vs1[0] <- montmul(a2, b2)
5674 // vs1[1] <- montmul(a3, b3)
5675 // vs1[2] <- montmul(a2, b3)
5676 // vs1[3] <- montmul(a3, b2)
5677 kyber_montmul16(vs_front(vs1), vs_front(vs4), vs_back(vs4), vs_front(vs2), vq);
5678 kyber_montmul16(vs_back(vs1),
5679 vs_front(vs4), vs_reverse(vs_back(vs4)), vs_back(vs2), vq);
5680
5681 // montmul result 2 of each cross-product i.e. (a1*b1, a3*b3) by a zeta.
5682 // We can schedule two montmuls at a time if we use a suitable vector
5683 // sequence <vs3[1], vs1[1]>.
5684 int delta = vs1[1]->encoding() - vs3[1]->encoding();
5685 VSeq<2> vs5(vs3[1], delta);
5686
5687 // vs3[1] <- montmul(montmul(a1, b1), z0)
5688 // vs1[1] <- montmul(montmul(a3, b3), z1)
5689 kyber_montmul16(vs5, vz, vs5, vs_front(vs2), vq);
5690
5691 // add results in pairs storing in vs3
5692 // vs3[0] <- montmul(a0, b0) + montmul(montmul(a1, b1), z0);
5693 // vs3[1] <- montmul(a0, b1) + montmul(a1, b0);
5694 vs_addv(vs_front(vs3), __ T8H, vs_even(vs3), vs_odd(vs3));
5695
5696 // vs3[2] <- montmul(a2, b2) + montmul(montmul(a3, b3), z1);
5697 // vs3[3] <- montmul(a2, b3) + montmul(a3, b2);
5698 vs_addv(vs_back(vs3), __ T8H, vs_even(vs1), vs_odd(vs1));
5699
5700 // vs1 <- montmul(vs3, montRSquareModQ)
5701 kyber_montmul32(vs1, vs3, vc, vs2, vq);
5702
5703 // store back the two pairs of result vectors de-interleaved as 8H elements
5704 // i.e. storing each pairs of shorts striped across a register pair adjacent
5705 // in memory
5706 vs_st2_post(vs1, __ T8H, result);
5707
5708 __ cmp(result, limit);
5709 __ br(Assembler::NE, kyberNttMult_loop);
5710
5711 __ leave(); // required for proper stackwalking of RuntimeStub frame
5712 __ mov(r0, zr); // return 0
5713 __ ret(lr);
5714
5715 return start;
5716 }
5717
5718 // Kyber add 2 polynomials.
5719 // Implements
5720 // static int implKyberAddPoly(short[] result, short[] a, short[] b) {}
5721 //
5722 // result (short[256]) = c_rarg0
5723 // a (short[256]) = c_rarg1
5724 // b (short[256]) = c_rarg2
5725 address generate_kyberAddPoly_2() {
5726
5727 __ align(CodeEntryAlignment);
5728 StubGenStubId stub_id = StubGenStubId::kyberAddPoly_2_id;
5729 StubCodeMark mark(this, stub_id);
5730 address start = __ pc();
5731 __ enter();
5732
5733 const Register result = c_rarg0;
5734 const Register a = c_rarg1;
5735 const Register b = c_rarg2;
5736
5737 const Register kyberConsts = r11;
5738
5739 // We sum 256 sets of values in total i.e. 32 x 8H quadwords.
5740 // So, we can load, add and store the data in 3 groups of 11,
5741 // 11 and 10 at a time i.e. we need to map sets of 10 or 11
5742 // registers. A further constraint is that the mapping needs
5743 // to skip callee saves. So, we allocate the register
5744 // sequences using two 8 sequences, two 2 sequences and two
5745 // single registers.
5746 VSeq<8> vs1_1(0);
5747 VSeq<2> vs1_2(16);
5748 FloatRegister vs1_3 = v28;
5749 VSeq<8> vs2_1(18);
5750 VSeq<2> vs2_2(26);
5751 FloatRegister vs2_3 = v29;
5752
5753 // two constant vector sequences
5754 VSeq<8> vc_1(31, 0);
5755 VSeq<2> vc_2(31, 0);
5756
5757 FloatRegister vc_3 = v31;
5758 __ lea(kyberConsts,
5759 ExternalAddress((address) StubRoutines::aarch64::_kyberConsts));
5760
5761 __ ldr(vc_3, __ Q, Address(kyberConsts, 16)); // q
5762 for (int i = 0; i < 3; i++) {
5763 // load 80 or 88 values from a into vs1_1/2/3
5764 vs_ldpq_post(vs1_1, a);
5765 vs_ldpq_post(vs1_2, a);
5766 if (i < 2) {
5767 __ ldr(vs1_3, __ Q, __ post(a, 16));
5768 }
5769 // load 80 or 88 values from b into vs2_1/2/3
5770 vs_ldpq_post(vs2_1, b);
5771 vs_ldpq_post(vs2_2, b);
5772 if (i < 2) {
5773 __ ldr(vs2_3, __ Q, __ post(b, 16));
5774 }
5775 // sum 80 or 88 values across vs1 and vs2 into vs1
5776 vs_addv(vs1_1, __ T8H, vs1_1, vs2_1);
5777 vs_addv(vs1_2, __ T8H, vs1_2, vs2_2);
5778 if (i < 2) {
5779 __ addv(vs1_3, __ T8H, vs1_3, vs2_3);
5780 }
5781 // add constant to all 80 or 88 results
5782 vs_addv(vs1_1, __ T8H, vs1_1, vc_1);
5783 vs_addv(vs1_2, __ T8H, vs1_2, vc_2);
5784 if (i < 2) {
5785 __ addv(vs1_3, __ T8H, vs1_3, vc_3);
5786 }
5787 // store 80 or 88 values
5788 vs_stpq_post(vs1_1, result);
5789 vs_stpq_post(vs1_2, result);
5790 if (i < 2) {
5791 __ str(vs1_3, __ Q, __ post(result, 16));
5792 }
5793 }
5794
5795 __ leave(); // required for proper stackwalking of RuntimeStub frame
5796 __ mov(r0, zr); // return 0
5797 __ ret(lr);
5798
5799 return start;
5800 }
5801
5802 // Kyber add 3 polynomials.
5803 // Implements
5804 // static int implKyberAddPoly(short[] result, short[] a, short[] b, short[] c) {}
5805 //
5806 // result (short[256]) = c_rarg0
5807 // a (short[256]) = c_rarg1
5808 // b (short[256]) = c_rarg2
5809 // c (short[256]) = c_rarg3
5810 address generate_kyberAddPoly_3() {
5811
5812 __ align(CodeEntryAlignment);
5813 StubGenStubId stub_id = StubGenStubId::kyberAddPoly_3_id;
5814 StubCodeMark mark(this, stub_id);
5815 address start = __ pc();
5816 __ enter();
5817
5818 const Register result = c_rarg0;
5819 const Register a = c_rarg1;
5820 const Register b = c_rarg2;
5821 const Register c = c_rarg3;
5822
5823 const Register kyberConsts = r11;
5824
5825 // As above we sum 256 sets of values in total i.e. 32 x 8H
5826 // quadwords. So, we can load, add and store the data in 3
5827 // groups of 11, 11 and 10 at a time i.e. we need to map sets
5828 // of 10 or 11 registers. A further constraint is that the
5829 // mapping needs to skip callee saves. So, we allocate the
5830 // register sequences using two 8 sequences, two 2 sequences
5831 // and two single registers.
5832 VSeq<8> vs1_1(0);
5833 VSeq<2> vs1_2(16);
5834 FloatRegister vs1_3 = v28;
5835 VSeq<8> vs2_1(18);
5836 VSeq<2> vs2_2(26);
5837 FloatRegister vs2_3 = v29;
5838
5839 // two constant vector sequences
5840 VSeq<8> vc_1(31, 0);
5841 VSeq<2> vc_2(31, 0);
5842
5843 FloatRegister vc_3 = v31;
5844
5845 __ lea(kyberConsts,
5846 ExternalAddress((address) StubRoutines::aarch64::_kyberConsts));
5847
5848 __ ldr(vc_3, __ Q, Address(kyberConsts, 16)); // q
5849 for (int i = 0; i < 3; i++) {
5850 // load 80 or 88 values from a into vs1_1/2/3
5851 vs_ldpq_post(vs1_1, a);
5852 vs_ldpq_post(vs1_2, a);
5853 if (i < 2) {
5854 __ ldr(vs1_3, __ Q, __ post(a, 16));
5855 }
5856 // load 80 or 88 values from b into vs2_1/2/3
5857 vs_ldpq_post(vs2_1, b);
5858 vs_ldpq_post(vs2_2, b);
5859 if (i < 2) {
5860 __ ldr(vs2_3, __ Q, __ post(b, 16));
5861 }
5862 // sum 80 or 88 values across vs1 and vs2 into vs1
5863 vs_addv(vs1_1, __ T8H, vs1_1, vs2_1);
5864 vs_addv(vs1_2, __ T8H, vs1_2, vs2_2);
5865 if (i < 2) {
5866 __ addv(vs1_3, __ T8H, vs1_3, vs2_3);
5867 }
5868 // load 80 or 88 values from c into vs2_1/2/3
5869 vs_ldpq_post(vs2_1, c);
5870 vs_ldpq_post(vs2_2, c);
5871 if (i < 2) {
5872 __ ldr(vs2_3, __ Q, __ post(c, 16));
5873 }
5874 // sum 80 or 88 values across vs1 and vs2 into vs1
5875 vs_addv(vs1_1, __ T8H, vs1_1, vs2_1);
5876 vs_addv(vs1_2, __ T8H, vs1_2, vs2_2);
5877 if (i < 2) {
5878 __ addv(vs1_3, __ T8H, vs1_3, vs2_3);
5879 }
5880 // add constant to all 80 or 88 results
5881 vs_addv(vs1_1, __ T8H, vs1_1, vc_1);
5882 vs_addv(vs1_2, __ T8H, vs1_2, vc_2);
5883 if (i < 2) {
5884 __ addv(vs1_3, __ T8H, vs1_3, vc_3);
5885 }
5886 // store 80 or 88 values
5887 vs_stpq_post(vs1_1, result);
5888 vs_stpq_post(vs1_2, result);
5889 if (i < 2) {
5890 __ str(vs1_3, __ Q, __ post(result, 16));
5891 }
5892 }
5893
5894 __ leave(); // required for proper stackwalking of RuntimeStub frame
5895 __ mov(r0, zr); // return 0
5896 __ ret(lr);
5897
5898 return start;
5899 }
5900
5901 // Kyber parse XOF output to polynomial coefficient candidates
5902 // or decodePoly(12, ...).
5903 // Implements
5904 // static int implKyber12To16(
5905 // byte[] condensed, int index, short[] parsed, int parsedLength) {}
5906 //
5907 // (parsedLength or (parsedLength - 48) must be divisible by 64.)
5908 //
5909 // condensed (byte[]) = c_rarg0
5910 // condensedIndex = c_rarg1
5911 // parsed (short[112 or 256]) = c_rarg2
5912 // parsedLength (112 or 256) = c_rarg3
5913 address generate_kyber12To16() {
5914 Label L_F00, L_loop, L_end;
5915
5916 __ BIND(L_F00);
5917 __ emit_int64(0x0f000f000f000f00);
5918 __ emit_int64(0x0f000f000f000f00);
5919
5920 __ align(CodeEntryAlignment);
5921 StubGenStubId stub_id = StubGenStubId::kyber12To16_id;
5922 StubCodeMark mark(this, stub_id);
5923 address start = __ pc();
5924 __ enter();
5925
5926 const Register condensed = c_rarg0;
5927 const Register condensedOffs = c_rarg1;
5928 const Register parsed = c_rarg2;
5929 const Register parsedLength = c_rarg3;
5930
5931 const Register tmpAddr = r11;
5932
5933 // Data is input 96 bytes at a time i.e. in groups of 6 x 16B
5934 // quadwords so we need a 6 vector sequence for the inputs.
5935 // Parsing produces 64 shorts, employing two 8 vector
5936 // sequences to store and combine the intermediate data.
5937 VSeq<6> vin(24);
5938 VSeq<8> va(0), vb(16);
5939
5940 __ adr(tmpAddr, L_F00);
5941 __ ldr(v31, __ Q, tmpAddr); // 8H times 0x0f00
5942 __ add(condensed, condensed, condensedOffs);
5943
5944 __ BIND(L_loop);
5945 // load 96 (6 x 16B) byte values
5946 vs_ld3_post(vin, __ T16B, condensed);
5947
5948 // The front half of sequence vin (vin[0], vin[1] and vin[2])
5949 // holds 48 (16x3) contiguous bytes from memory striped
5950 // horizontally across each of the 16 byte lanes. Equivalently,
5951 // that is 16 pairs of 12-bit integers. Likewise the back half
5952 // holds the next 48 bytes in the same arrangement.
5953
5954 // Each vector in the front half can also be viewed as a vertical
5955 // strip across the 16 pairs of 12 bit integers. Each byte in
5956 // vin[0] stores the low 8 bits of the first int in a pair. Each
5957 // byte in vin[1] stores the high 4 bits of the first int and the
5958 // low 4 bits of the second int. Each byte in vin[2] stores the
5959 // high 8 bits of the second int. Likewise the vectors in second
5960 // half.
5961
5962 // Converting the data to 16-bit shorts requires first of all
5963 // expanding each of the 6 x 16B vectors into 6 corresponding
5964 // pairs of 8H vectors. Mask, shift and add operations on the
5965 // resulting vector pairs can be used to combine 4 and 8 bit
5966 // parts of related 8H vector elements.
5967 //
5968 // The middle vectors (vin[2] and vin[5]) are actually expanded
5969 // twice, one copy manipulated to provide the lower 4 bits
5970 // belonging to the first short in a pair and another copy
5971 // manipulated to provide the higher 4 bits belonging to the
5972 // second short in a pair. This is why the the vector sequences va
5973 // and vb used to hold the expanded 8H elements are of length 8.
5974
5975 // Expand vin[0] into va[0:1], and vin[1] into va[2:3] and va[4:5]
5976 // n.b. target elements 2 and 3 duplicate elements 4 and 5
5977 __ ushll(va[0], __ T8H, vin[0], __ T8B, 0);
5978 __ ushll2(va[1], __ T8H, vin[0], __ T16B, 0);
5979 __ ushll(va[2], __ T8H, vin[1], __ T8B, 0);
5980 __ ushll2(va[3], __ T8H, vin[1], __ T16B, 0);
5981 __ ushll(va[4], __ T8H, vin[1], __ T8B, 0);
5982 __ ushll2(va[5], __ T8H, vin[1], __ T16B, 0);
5983
5984 // likewise expand vin[3] into vb[0:1], and vin[4] into vb[2:3]
5985 // and vb[4:5]
5986 __ ushll(vb[0], __ T8H, vin[3], __ T8B, 0);
5987 __ ushll2(vb[1], __ T8H, vin[3], __ T16B, 0);
5988 __ ushll(vb[2], __ T8H, vin[4], __ T8B, 0);
5989 __ ushll2(vb[3], __ T8H, vin[4], __ T16B, 0);
5990 __ ushll(vb[4], __ T8H, vin[4], __ T8B, 0);
5991 __ ushll2(vb[5], __ T8H, vin[4], __ T16B, 0);
5992
5993 // shift lo byte of copy 1 of the middle stripe into the high byte
5994 __ shl(va[2], __ T8H, va[2], 8);
5995 __ shl(va[3], __ T8H, va[3], 8);
5996 __ shl(vb[2], __ T8H, vb[2], 8);
5997 __ shl(vb[3], __ T8H, vb[3], 8);
5998
5999 // expand vin[2] into va[6:7] and vin[5] into vb[6:7] but this
6000 // time pre-shifted by 4 to ensure top bits of input 12-bit int
6001 // are in bit positions [4..11].
6002 __ ushll(va[6], __ T8H, vin[2], __ T8B, 4);
6003 __ ushll2(va[7], __ T8H, vin[2], __ T16B, 4);
6004 __ ushll(vb[6], __ T8H, vin[5], __ T8B, 4);
6005 __ ushll2(vb[7], __ T8H, vin[5], __ T16B, 4);
6006
6007 // mask hi 4 bits of the 1st 12-bit int in a pair from copy1 and
6008 // shift lo 4 bits of the 2nd 12-bit int in a pair to the bottom of
6009 // copy2
6010 __ andr(va[2], __ T16B, va[2], v31);
6011 __ andr(va[3], __ T16B, va[3], v31);
6012 __ ushr(va[4], __ T8H, va[4], 4);
6013 __ ushr(va[5], __ T8H, va[5], 4);
6014 __ andr(vb[2], __ T16B, vb[2], v31);
6015 __ andr(vb[3], __ T16B, vb[3], v31);
6016 __ ushr(vb[4], __ T8H, vb[4], 4);
6017 __ ushr(vb[5], __ T8H, vb[5], 4);
6018
6019 // sum hi 4 bits and lo 8 bits of the 1st 12-bit int in each pair and
6020 // hi 8 bits plus lo 4 bits of the 2nd 12-bit int in each pair
6021 // n.b. the ordering ensures: i) inputs are consumed before they
6022 // are overwritten ii) the order of 16-bit results across successive
6023 // pairs of vectors in va and then vb reflects the order of the
6024 // corresponding 12-bit inputs
6025 __ addv(va[0], __ T8H, va[0], va[2]);
6026 __ addv(va[2], __ T8H, va[1], va[3]);
6027 __ addv(va[1], __ T8H, va[4], va[6]);
6028 __ addv(va[3], __ T8H, va[5], va[7]);
6029 __ addv(vb[0], __ T8H, vb[0], vb[2]);
6030 __ addv(vb[2], __ T8H, vb[1], vb[3]);
6031 __ addv(vb[1], __ T8H, vb[4], vb[6]);
6032 __ addv(vb[3], __ T8H, vb[5], vb[7]);
6033
6034 // store 64 results interleaved as shorts
6035 vs_st2_post(vs_front(va), __ T8H, parsed);
6036 vs_st2_post(vs_front(vb), __ T8H, parsed);
6037
6038 __ sub(parsedLength, parsedLength, 64);
6039 __ cmp(parsedLength, (u1)64);
6040 __ br(Assembler::GE, L_loop);
6041 __ cbz(parsedLength, L_end);
6042
6043 // if anything is left it should be a final 72 bytes of input
6044 // i.e. a final 48 12-bit values. so we handle this by loading
6045 // 48 bytes into all 16B lanes of front(vin) and only 24
6046 // bytes into the lower 8B lane of back(vin)
6047 vs_ld3_post(vs_front(vin), __ T16B, condensed);
6048 vs_ld3(vs_back(vin), __ T8B, condensed);
6049
6050 // Expand vin[0] into va[0:1], and vin[1] into va[2:3] and va[4:5]
6051 // n.b. target elements 2 and 3 of va duplicate elements 4 and
6052 // 5 and target element 2 of vb duplicates element 4.
6053 __ ushll(va[0], __ T8H, vin[0], __ T8B, 0);
6054 __ ushll2(va[1], __ T8H, vin[0], __ T16B, 0);
6055 __ ushll(va[2], __ T8H, vin[1], __ T8B, 0);
6056 __ ushll2(va[3], __ T8H, vin[1], __ T16B, 0);
6057 __ ushll(va[4], __ T8H, vin[1], __ T8B, 0);
6058 __ ushll2(va[5], __ T8H, vin[1], __ T16B, 0);
6059
6060 // This time expand just the lower 8 lanes
6061 __ ushll(vb[0], __ T8H, vin[3], __ T8B, 0);
6062 __ ushll(vb[2], __ T8H, vin[4], __ T8B, 0);
6063 __ ushll(vb[4], __ T8H, vin[4], __ T8B, 0);
6064
6065 // shift lo byte of copy 1 of the middle stripe into the high byte
6066 __ shl(va[2], __ T8H, va[2], 8);
6067 __ shl(va[3], __ T8H, va[3], 8);
6068 __ shl(vb[2], __ T8H, vb[2], 8);
6069
6070 // expand vin[2] into va[6:7] and lower 8 lanes of vin[5] into
6071 // vb[6] pre-shifted by 4 to ensure top bits of the input 12-bit
6072 // int are in bit positions [4..11].
6073 __ ushll(va[6], __ T8H, vin[2], __ T8B, 4);
6074 __ ushll2(va[7], __ T8H, vin[2], __ T16B, 4);
6075 __ ushll(vb[6], __ T8H, vin[5], __ T8B, 4);
6076
6077 // mask hi 4 bits of each 1st 12-bit int in pair from copy1 and
6078 // shift lo 4 bits of each 2nd 12-bit int in pair to bottom of
6079 // copy2
6080 __ andr(va[2], __ T16B, va[2], v31);
6081 __ andr(va[3], __ T16B, va[3], v31);
6082 __ ushr(va[4], __ T8H, va[4], 4);
6083 __ ushr(va[5], __ T8H, va[5], 4);
6084 __ andr(vb[2], __ T16B, vb[2], v31);
6085 __ ushr(vb[4], __ T8H, vb[4], 4);
6086
6087
6088
6089 // sum hi 4 bits and lo 8 bits of each 1st 12-bit int in pair and
6090 // hi 8 bits plus lo 4 bits of each 2nd 12-bit int in pair
6091
6092 // n.b. ordering ensures: i) inputs are consumed before they are
6093 // overwritten ii) order of 16-bit results across succsessive
6094 // pairs of vectors in va and then lower half of vb reflects order
6095 // of corresponding 12-bit inputs
6096 __ addv(va[0], __ T8H, va[0], va[2]);
6097 __ addv(va[2], __ T8H, va[1], va[3]);
6098 __ addv(va[1], __ T8H, va[4], va[6]);
6099 __ addv(va[3], __ T8H, va[5], va[7]);
6100 __ addv(vb[0], __ T8H, vb[0], vb[2]);
6101 __ addv(vb[1], __ T8H, vb[4], vb[6]);
6102
6103 // store 48 results interleaved as shorts
6104 vs_st2_post(vs_front(va), __ T8H, parsed);
6105 vs_st2_post(vs_front(vs_front(vb)), __ T8H, parsed);
6106
6107 __ BIND(L_end);
6108
6109 __ leave(); // required for proper stackwalking of RuntimeStub frame
6110 __ mov(r0, zr); // return 0
6111 __ ret(lr);
6112
6113 return start;
6114 }
6115
6116 // Kyber Barrett reduce function.
6117 // Implements
6118 // static int implKyberBarrettReduce(short[] coeffs) {}
6119 //
6120 // coeffs (short[256]) = c_rarg0
6121 address generate_kyberBarrettReduce() {
6122
6123 __ align(CodeEntryAlignment);
6124 StubGenStubId stub_id = StubGenStubId::kyberBarrettReduce_id;
6125 StubCodeMark mark(this, stub_id);
6126 address start = __ pc();
6127 __ enter();
6128
6129 const Register coeffs = c_rarg0;
6130
6131 const Register kyberConsts = r10;
6132 const Register result = r11;
6133
6134 // As above we process 256 sets of values in total i.e. 32 x
6135 // 8H quadwords. So, we can load, add and store the data in 3
6136 // groups of 11, 11 and 10 at a time i.e. we need to map sets
6137 // of 10 or 11 registers. A further constraint is that the
6138 // mapping needs to skip callee saves. So, we allocate the
6139 // register sequences using two 8 sequences, two 2 sequences
6140 // and two single registers.
6141 VSeq<8> vs1_1(0);
6142 VSeq<2> vs1_2(16);
6143 FloatRegister vs1_3 = v28;
6144 VSeq<8> vs2_1(18);
6145 VSeq<2> vs2_2(26);
6146 FloatRegister vs2_3 = v29;
6147
6148 // we also need a pair of corresponding constant sequences
6149
6150 VSeq<8> vc1_1(30, 0);
6151 VSeq<2> vc1_2(30, 0);
6152 FloatRegister vc1_3 = v30; // for kyber_q
6153
6154 VSeq<8> vc2_1(31, 0);
6155 VSeq<2> vc2_2(31, 0);
6156 FloatRegister vc2_3 = v31; // for kyberBarrettMultiplier
6157
6158 __ add(result, coeffs, 0);
6159 __ lea(kyberConsts,
6160 ExternalAddress((address) StubRoutines::aarch64::_kyberConsts));
6161
6162 // load q and the multiplier for the Barrett reduction
6163 __ add(kyberConsts, kyberConsts, 16);
6164 __ ldpq(vc1_3, vc2_3, kyberConsts);
6165
6166 for (int i = 0; i < 3; i++) {
6167 // load 80 or 88 coefficients
6168 vs_ldpq_post(vs1_1, coeffs);
6169 vs_ldpq_post(vs1_2, coeffs);
6170 if (i < 2) {
6171 __ ldr(vs1_3, __ Q, __ post(coeffs, 16));
6172 }
6173
6174 // vs2 <- (2 * vs1 * kyberBarrettMultiplier) >> 16
6175 vs_sqdmulh(vs2_1, __ T8H, vs1_1, vc2_1);
6176 vs_sqdmulh(vs2_2, __ T8H, vs1_2, vc2_2);
6177 if (i < 2) {
6178 __ sqdmulh(vs2_3, __ T8H, vs1_3, vc2_3);
6179 }
6180
6181 // vs2 <- (vs1 * kyberBarrettMultiplier) >> 26
6182 vs_sshr(vs2_1, __ T8H, vs2_1, 11);
6183 vs_sshr(vs2_2, __ T8H, vs2_2, 11);
6184 if (i < 2) {
6185 __ sshr(vs2_3, __ T8H, vs2_3, 11);
6186 }
6187
6188 // vs1 <- vs1 - vs2 * kyber_q
6189 vs_mlsv(vs1_1, __ T8H, vs2_1, vc1_1);
6190 vs_mlsv(vs1_2, __ T8H, vs2_2, vc1_2);
6191 if (i < 2) {
6192 __ mlsv(vs1_3, __ T8H, vs2_3, vc1_3);
6193 }
6194
6195 vs_stpq_post(vs1_1, result);
6196 vs_stpq_post(vs1_2, result);
6197 if (i < 2) {
6198 __ str(vs1_3, __ Q, __ post(result, 16));
6199 }
6200 }
6201
6202 __ leave(); // required for proper stackwalking of RuntimeStub frame
6203 __ mov(r0, zr); // return 0
6204 __ ret(lr);
6205
6206 return start;
6207 }
6208
6209
6210 // Dilithium-specific montmul helper routines that generate parallel
6211 // code for, respectively, a single 4x4s vector sequence montmul or
6212 // two such multiplies in a row.
6213
6214 // Perform 16 32-bit Montgomery multiplications in parallel
6215 void dilithium_montmul16(const VSeq<4>& va, const VSeq<4>& vb, const VSeq<4>& vc,
6216 const VSeq<4>& vtmp, const VSeq<2>& vq) {
6217 // Use the helper routine to schedule a 4x4S Montgomery multiply.
6218 // It will assert that the register use is valid
6219 vs_montmul4(va, vb, vc, __ T4S, vtmp, vq);
6220 }
6221
6222 // Perform 2x16 32-bit Montgomery multiplications in parallel
6223 void dilithium_montmul32(const VSeq<8>& va, const VSeq<8>& vb, const VSeq<8>& vc,
6224 const VSeq<4>& vtmp, const VSeq<2>& vq) {
6225 // Schedule two successive 4x4S multiplies via the montmul helper
6226 // on the front and back halves of va, vb and vc. The helper will
6227 // assert that the register use has no overlap conflicts on each
6228 // individual call but we also need to ensure that the necessary
6229 // disjoint/equality constraints are met across both calls.
6230
6231 // vb, vc, vtmp and vq must be disjoint. va must either be
6232 // disjoint from all other registers or equal vc
6233
6234 assert(vs_disjoint(vb, vc), "vb and vc overlap");
6235 assert(vs_disjoint(vb, vq), "vb and vq overlap");
6236 assert(vs_disjoint(vb, vtmp), "vb and vtmp overlap");
6237
6238 assert(vs_disjoint(vc, vq), "vc and vq overlap");
6239 assert(vs_disjoint(vc, vtmp), "vc and vtmp overlap");
6240
6241 assert(vs_disjoint(vq, vtmp), "vq and vtmp overlap");
6242
6243 assert(vs_disjoint(va, vc) || vs_same(va, vc), "va and vc neither disjoint nor equal");
6244 assert(vs_disjoint(va, vb), "va and vb overlap");
6245 assert(vs_disjoint(va, vq), "va and vq overlap");
6246 assert(vs_disjoint(va, vtmp), "va and vtmp overlap");
6247
6248 // We multiply the front and back halves of each sequence 4 at a
6249 // time because
6250 //
6251 // 1) we are currently only able to get 4-way instruction
6252 // parallelism at best
6253 //
6254 // 2) we need registers for the constants in vq and temporary
6255 // scratch registers to hold intermediate results so vtmp can only
6256 // be a VSeq<4> which means we only have 4 scratch slots.
6257
6258 vs_montmul4(vs_front(va), vs_front(vb), vs_front(vc), __ T4S, vtmp, vq);
6259 vs_montmul4(vs_back(va), vs_back(vb), vs_back(vc), __ T4S, vtmp, vq);
6260 }
6261
6262 // Perform combined montmul then add/sub on 4x4S vectors.
6263 void dilithium_montmul16_sub_add(
6264 const VSeq<4>& va0, const VSeq<4>& va1, const VSeq<4>& vc,
6265 const VSeq<4>& vtmp, const VSeq<2>& vq) {
6266 // compute a = montmul(a1, c)
6267 dilithium_montmul16(vc, va1, vc, vtmp, vq);
6268 // ouptut a1 = a0 - a
6269 vs_subv(va1, __ T4S, va0, vc);
6270 // and a0 = a0 + a
6271 vs_addv(va0, __ T4S, va0, vc);
6272 }
6273
6274 // Perform combined add/sub then montul on 4x4S vectors.
6275 void dilithium_sub_add_montmul16(
6276 const VSeq<4>& va0, const VSeq<4>& va1, const VSeq<4>& vb,
6277 const VSeq<4>& vtmp1, const VSeq<4>& vtmp2, const VSeq<2>& vq) {
6278 // compute c = a0 - a1
6279 vs_subv(vtmp1, __ T4S, va0, va1);
6280 // output a0 = a0 + a1
6281 vs_addv(va0, __ T4S, va0, va1);
6282 // output a1 = b montmul c
6283 dilithium_montmul16(va1, vtmp1, vb, vtmp2, vq);
6284 }
6285
6286 // At these levels, the indices that correspond to the 'j's (and 'j+l's)
6287 // in the Java implementation come in sequences of at least 8, so we
6288 // can use ldpq to collect the corresponding data into pairs of vector
6289 // registers.
6290 // We collect the coefficients corresponding to the 'j+l' indexes into
6291 // the vector registers v0-v7, the zetas into the vector registers v16-v23
6292 // then we do the (Montgomery) multiplications by the zetas in parallel
6293 // into v16-v23, load the coeffs corresponding to the 'j' indexes into
6294 // v0-v7, then do the additions into v24-v31 and the subtractions into
6295 // v0-v7 and finally save the results back to the coeffs array.
6296 void dilithiumNttLevel0_4(const Register dilithiumConsts,
6297 const Register coeffs, const Register zetas) {
6298 int c1 = 0;
6299 int c2 = 512;
6300 int startIncr;
6301 // don't use callee save registers v8 - v15
6302 VSeq<8> vs1(0), vs2(16), vs3(24); // 3 sets of 8x4s inputs/outputs
6303 VSeq<4> vtmp = vs_front(vs3); // n.b. tmp registers overlap vs3
6304 VSeq<2> vq(30); // n.b. constants overlap vs3
6305 int offsets[4] = { 0, 32, 64, 96 };
6306
6307 for (int level = 0; level < 5; level++) {
6308 int c1Start = c1;
6309 int c2Start = c2;
6310 if (level == 3) {
6311 offsets[1] = 32;
6312 offsets[2] = 128;
6313 offsets[3] = 160;
6314 } else if (level == 4) {
6315 offsets[1] = 64;
6316 offsets[2] = 128;
6317 offsets[3] = 192;
6318 }
6319
6320 // For levels 1 - 4 we simply load 2 x 4 adjacent values at a
6321 // time at 4 different offsets and multiply them in order by the
6322 // next set of input values. So we employ indexed load and store
6323 // pair instructions with arrangement 4S.
6324 for (int i = 0; i < 4; i++) {
6325 // reload q and qinv
6326 vs_ldpq(vq, dilithiumConsts); // qInv, q
6327 // load 8x4S coefficients via second start pos == c2
6328 vs_ldpq_indexed(vs1, coeffs, c2Start, offsets);
6329 // load next 8x4S inputs == b
6330 vs_ldpq_post(vs2, zetas);
6331 // compute a == c2 * b mod MONT_Q
6332 dilithium_montmul32(vs2, vs1, vs2, vtmp, vq);
6333 // load 8x4s coefficients via first start pos == c1
6334 vs_ldpq_indexed(vs1, coeffs, c1Start, offsets);
6335 // compute a1 = c1 + a
6336 vs_addv(vs3, __ T4S, vs1, vs2);
6337 // compute a2 = c1 - a
6338 vs_subv(vs1, __ T4S, vs1, vs2);
6339 // output a1 and a2
6340 vs_stpq_indexed(vs3, coeffs, c1Start, offsets);
6341 vs_stpq_indexed(vs1, coeffs, c2Start, offsets);
6342
6343 int k = 4 * level + i;
6344
6345 if (k > 7) {
6346 startIncr = 256;
6347 } else if (k == 5) {
6348 startIncr = 384;
6349 } else {
6350 startIncr = 128;
6351 }
6352
6353 c1Start += startIncr;
6354 c2Start += startIncr;
6355 }
6356
6357 c2 /= 2;
6358 }
6359 }
6360
6361 // Dilithium NTT function except for the final "normalization" to |coeff| < Q.
6362 // Implements the method
6363 // static int implDilithiumAlmostNtt(int[] coeffs, int zetas[]) {}
6364 // of the Java class sun.security.provider
6365 //
6366 // coeffs (int[256]) = c_rarg0
6367 // zetas (int[256]) = c_rarg1
6368 address generate_dilithiumAlmostNtt() {
6369
6370 __ align(CodeEntryAlignment);
6371 StubGenStubId stub_id = StubGenStubId::dilithiumAlmostNtt_id;
6372 StubCodeMark mark(this, stub_id);
6373 address start = __ pc();
6374 __ enter();
6375
6376 const Register coeffs = c_rarg0;
6377 const Register zetas = c_rarg1;
6378
6379 const Register tmpAddr = r9;
6380 const Register dilithiumConsts = r10;
6381 const Register result = r11;
6382 // don't use callee save registers v8 - v15
6383 VSeq<8> vs1(0), vs2(16), vs3(24); // 3 sets of 8x4s inputs/outputs
6384 VSeq<4> vtmp = vs_front(vs3); // n.b. tmp registers overlap vs3
6385 VSeq<2> vq(30); // n.b. constants overlap vs3
6386 int offsets[4] = { 0, 32, 64, 96};
6387 int offsets1[8] = { 16, 48, 80, 112, 144, 176, 208, 240 };
6388 int offsets2[8] = { 0, 32, 64, 96, 128, 160, 192, 224 };
6389 __ add(result, coeffs, 0);
6390 __ lea(dilithiumConsts,
6391 ExternalAddress((address) StubRoutines::aarch64::_dilithiumConsts));
6392
6393 // Each level represents one iteration of the outer for loop of the Java version.
6394
6395 // level 0-4
6396 dilithiumNttLevel0_4(dilithiumConsts, coeffs, zetas);
6397
6398 // level 5
6399
6400 // At level 5 the coefficients we need to combine with the zetas
6401 // are grouped in memory in blocks of size 4. So, for both sets of
6402 // coefficients we load 4 adjacent values at 8 different offsets
6403 // using an indexed ldr with register variant Q and multiply them
6404 // in sequence order by the next set of inputs. Likewise we store
6405 // the resuls using an indexed str with register variant Q.
6406 for (int i = 0; i < 1024; i += 256) {
6407 // reload constants q, qinv each iteration as they get clobbered later
6408 vs_ldpq(vq, dilithiumConsts); // qInv, q
6409 // load 32 (8x4S) coefficients via first offsets = c1
6410 vs_ldr_indexed(vs1, __ Q, coeffs, i, offsets1);
6411 // load next 32 (8x4S) inputs = b
6412 vs_ldpq_post(vs2, zetas);
6413 // a = b montul c1
6414 dilithium_montmul32(vs2, vs1, vs2, vtmp, vq);
6415 // load 32 (8x4S) coefficients via second offsets = c2
6416 vs_ldr_indexed(vs1, __ Q, coeffs, i, offsets2);
6417 // add/sub with result of multiply
6418 vs_addv(vs3, __ T4S, vs1, vs2); // a1 = a - c2
6419 vs_subv(vs1, __ T4S, vs1, vs2); // a0 = a + c1
6420 // write back new coefficients using same offsets
6421 vs_str_indexed(vs3, __ Q, coeffs, i, offsets2);
6422 vs_str_indexed(vs1, __ Q, coeffs, i, offsets1);
6423 }
6424
6425 // level 6
6426 // At level 6 the coefficients we need to combine with the zetas
6427 // are grouped in memory in pairs, the first two being montmul
6428 // inputs and the second add/sub inputs. We can still implement
6429 // the montmul+sub+add using 4-way parallelism but only if we
6430 // combine the coefficients with the zetas 16 at a time. We load 8
6431 // adjacent values at 4 different offsets using an ld2 load with
6432 // arrangement 2D. That interleaves the lower and upper halves of
6433 // each pair of quadwords into successive vector registers. We
6434 // then need to montmul the 4 even elements of the coefficients
6435 // register sequence by the zetas in order and then add/sub the 4
6436 // odd elements of the coefficients register sequence. We use an
6437 // equivalent st2 operation to store the results back into memory
6438 // de-interleaved.
6439 for (int i = 0; i < 1024; i += 128) {
6440 // reload constants q, qinv each iteration as they get clobbered later
6441 vs_ldpq(vq, dilithiumConsts); // qInv, q
6442 // load interleaved 16 (4x2D) coefficients via offsets
6443 vs_ld2_indexed(vs1, __ T2D, coeffs, tmpAddr, i, offsets);
6444 // load next 16 (4x4S) inputs
6445 vs_ldpq_post(vs_front(vs2), zetas);
6446 // mont multiply odd elements of vs1 by vs2 and add/sub into odds/evens
6447 dilithium_montmul16_sub_add(vs_even(vs1), vs_odd(vs1),
6448 vs_front(vs2), vtmp, vq);
6449 // store interleaved 16 (4x2D) coefficients via offsets
6450 vs_st2_indexed(vs1, __ T2D, coeffs, tmpAddr, i, offsets);
6451 }
6452
6453 // level 7
6454 // At level 7 the coefficients we need to combine with the zetas
6455 // occur singly with montmul inputs alterating with add/sub
6456 // inputs. Once again we can use 4-way parallelism to combine 16
6457 // zetas at a time. However, we have to load 8 adjacent values at
6458 // 4 different offsets using an ld2 load with arrangement 4S. That
6459 // interleaves the the odd words of each pair into one
6460 // coefficients vector register and the even words of the pair
6461 // into the next register. We then need to montmul the 4 even
6462 // elements of the coefficients register sequence by the zetas in
6463 // order and then add/sub the 4 odd elements of the coefficients
6464 // register sequence. We use an equivalent st2 operation to store
6465 // the results back into memory de-interleaved.
6466
6467 for (int i = 0; i < 1024; i += 128) {
6468 // reload constants q, qinv each iteration as they get clobbered later
6469 vs_ldpq(vq, dilithiumConsts); // qInv, q
6470 // load interleaved 16 (4x4S) coefficients via offsets
6471 vs_ld2_indexed(vs1, __ T4S, coeffs, tmpAddr, i, offsets);
6472 // load next 16 (4x4S) inputs
6473 vs_ldpq_post(vs_front(vs2), zetas);
6474 // mont multiply odd elements of vs1 by vs2 and add/sub into odds/evens
6475 dilithium_montmul16_sub_add(vs_even(vs1), vs_odd(vs1),
6476 vs_front(vs2), vtmp, vq);
6477 // store interleaved 16 (4x4S) coefficients via offsets
6478 vs_st2_indexed(vs1, __ T4S, coeffs, tmpAddr, i, offsets);
6479 }
6480 __ leave(); // required for proper stackwalking of RuntimeStub frame
6481 __ mov(r0, zr); // return 0
6482 __ ret(lr);
6483
6484 return start;
6485 }
6486
6487 // At these levels, the indices that correspond to the 'j's (and 'j+l's)
6488 // in the Java implementation come in sequences of at least 8, so we
6489 // can use ldpq to collect the corresponding data into pairs of vector
6490 // registers
6491 // We collect the coefficients that correspond to the 'j's into vs1
6492 // the coefficiets that correspond to the 'j+l's into vs2 then
6493 // do the additions into vs3 and the subtractions into vs1 then
6494 // save the result of the additions, load the zetas into vs2
6495 // do the (Montgomery) multiplications by zeta in parallel into vs2
6496 // finally save the results back to the coeffs array
6497 void dilithiumInverseNttLevel3_7(const Register dilithiumConsts,
6498 const Register coeffs, const Register zetas) {
6499 int c1 = 0;
6500 int c2 = 32;
6501 int startIncr;
6502 int offsets[4];
6503 VSeq<8> vs1(0), vs2(16), vs3(24); // 3 sets of 8x4s inputs/outputs
6504 VSeq<4> vtmp = vs_front(vs3); // n.b. tmp registers overlap vs3
6505 VSeq<2> vq(30); // n.b. constants overlap vs3
6506
6507 offsets[0] = 0;
6508
6509 for (int level = 3; level < 8; level++) {
6510 int c1Start = c1;
6511 int c2Start = c2;
6512 if (level == 3) {
6513 offsets[1] = 64;
6514 offsets[2] = 128;
6515 offsets[3] = 192;
6516 } else if (level == 4) {
6517 offsets[1] = 32;
6518 offsets[2] = 128;
6519 offsets[3] = 160;
6520 } else {
6521 offsets[1] = 32;
6522 offsets[2] = 64;
6523 offsets[3] = 96;
6524 }
6525
6526 // For levels 3 - 7 we simply load 2 x 4 adjacent values at a
6527 // time at 4 different offsets and multiply them in order by the
6528 // next set of input values. So we employ indexed load and store
6529 // pair instructions with arrangement 4S.
6530 for (int i = 0; i < 4; i++) {
6531 // load v1 32 (8x4S) coefficients relative to first start index
6532 vs_ldpq_indexed(vs1, coeffs, c1Start, offsets);
6533 // load v2 32 (8x4S) coefficients relative to second start index
6534 vs_ldpq_indexed(vs2, coeffs, c2Start, offsets);
6535 // a0 = v1 + v2 -- n.b. clobbers vqs
6536 vs_addv(vs3, __ T4S, vs1, vs2);
6537 // a1 = v1 - v2
6538 vs_subv(vs1, __ T4S, vs1, vs2);
6539 // save a1 relative to first start index
6540 vs_stpq_indexed(vs3, coeffs, c1Start, offsets);
6541 // load constants q, qinv each iteration as they get clobbered above
6542 vs_ldpq(vq, dilithiumConsts); // qInv, q
6543 // load b next 32 (8x4S) inputs
6544 vs_ldpq_post(vs2, zetas);
6545 // a = a1 montmul b
6546 dilithium_montmul32(vs2, vs1, vs2, vtmp, vq);
6547 // save a relative to second start index
6548 vs_stpq_indexed(vs2, coeffs, c2Start, offsets);
6549
6550 int k = 4 * level + i;
6551
6552 if (k < 24) {
6553 startIncr = 256;
6554 } else if (k == 25) {
6555 startIncr = 384;
6556 } else {
6557 startIncr = 128;
6558 }
6559
6560 c1Start += startIncr;
6561 c2Start += startIncr;
6562 }
6563
6564 c2 *= 2;
6565 }
6566 }
6567
6568 // Dilithium Inverse NTT function except the final mod Q division by 2^256.
6569 // Implements the method
6570 // static int implDilithiumAlmostInverseNtt(int[] coeffs, int[] zetas) {} of
6571 // the sun.security.provider.ML_DSA class.
6572 //
6573 // coeffs (int[256]) = c_rarg0
6574 // zetas (int[256]) = c_rarg1
6575 address generate_dilithiumAlmostInverseNtt() {
6576
6577 __ align(CodeEntryAlignment);
6578 StubGenStubId stub_id = StubGenStubId::dilithiumAlmostInverseNtt_id;
6579 StubCodeMark mark(this, stub_id);
6580 address start = __ pc();
6581 __ enter();
6582
6583 const Register coeffs = c_rarg0;
6584 const Register zetas = c_rarg1;
6585
6586 const Register tmpAddr = r9;
6587 const Register dilithiumConsts = r10;
6588 const Register result = r11;
6589 VSeq<8> vs1(0), vs2(16), vs3(24); // 3 sets of 8x4s inputs/outputs
6590 VSeq<4> vtmp = vs_front(vs3); // n.b. tmp registers overlap vs3
6591 VSeq<2> vq(30); // n.b. constants overlap vs3
6592 int offsets[4] = { 0, 32, 64, 96 };
6593 int offsets1[8] = { 0, 32, 64, 96, 128, 160, 192, 224 };
6594 int offsets2[8] = { 16, 48, 80, 112, 144, 176, 208, 240 };
6595
6596 __ add(result, coeffs, 0);
6597 __ lea(dilithiumConsts,
6598 ExternalAddress((address) StubRoutines::aarch64::_dilithiumConsts));
6599
6600 // Each level represents one iteration of the outer for loop of the Java version
6601
6602 // level 0
6603 // At level 0 we need to interleave adjacent quartets of
6604 // coefficients before we multiply and add/sub by the next 16
6605 // zetas just as we did for level 7 in the multiply code. So we
6606 // load and store the values using an ld2/st2 with arrangement 4S.
6607 for (int i = 0; i < 1024; i += 128) {
6608 // load constants q, qinv
6609 // n.b. this can be moved out of the loop as they do not get
6610 // clobbered by first two loops
6611 vs_ldpq(vq, dilithiumConsts); // qInv, q
6612 // a0/a1 load interleaved 32 (8x4S) coefficients
6613 vs_ld2_indexed(vs1, __ T4S, coeffs, tmpAddr, i, offsets);
6614 // b load next 32 (8x4S) inputs
6615 vs_ldpq_post(vs_front(vs2), zetas);
6616 // compute in parallel (a0, a1) = (a0 + a1, (a0 - a1) montmul b)
6617 // n.b. second half of vs2 provides temporary register storage
6618 dilithium_sub_add_montmul16(vs_even(vs1), vs_odd(vs1),
6619 vs_front(vs2), vs_back(vs2), vtmp, vq);
6620 // a0/a1 store interleaved 32 (8x4S) coefficients
6621 vs_st2_indexed(vs1, __ T4S, coeffs, tmpAddr, i, offsets);
6622 }
6623
6624 // level 1
6625 // At level 1 we need to interleave pairs of adjacent pairs of
6626 // coefficients before we multiply by the next 16 zetas just as we
6627 // did for level 6 in the multiply code. So we load and store the
6628 // values an ld2/st2 with arrangement 2D.
6629 for (int i = 0; i < 1024; i += 128) {
6630 // a0/a1 load interleaved 32 (8x2D) coefficients
6631 vs_ld2_indexed(vs1, __ T2D, coeffs, tmpAddr, i, offsets);
6632 // b load next 16 (4x4S) inputs
6633 vs_ldpq_post(vs_front(vs2), zetas);
6634 // compute in parallel (a0, a1) = (a0 + a1, (a0 - a1) montmul b)
6635 // n.b. second half of vs2 provides temporary register storage
6636 dilithium_sub_add_montmul16(vs_even(vs1), vs_odd(vs1),
6637 vs_front(vs2), vs_back(vs2), vtmp, vq);
6638 // a0/a1 store interleaved 32 (8x2D) coefficients
6639 vs_st2_indexed(vs1, __ T2D, coeffs, tmpAddr, i, offsets);
6640 }
6641
6642 // level 2
6643 // At level 2 coefficients come in blocks of 4. So, we load 4
6644 // adjacent coefficients at 8 distinct offsets for both the first
6645 // and second coefficient sequences, using an ldr with register
6646 // variant Q then combine them with next set of 32 zetas. Likewise
6647 // we store the results using an str with register variant Q.
6648 for (int i = 0; i < 1024; i += 256) {
6649 // c0 load 32 (8x4S) coefficients via first offsets
6650 vs_ldr_indexed(vs1, __ Q, coeffs, i, offsets1);
6651 // c1 load 32 (8x4S) coefficients via second offsets
6652 vs_ldr_indexed(vs2, __ Q,coeffs, i, offsets2);
6653 // a0 = c0 + c1 n.b. clobbers vq which overlaps vs3
6654 vs_addv(vs3, __ T4S, vs1, vs2);
6655 // c = c0 - c1
6656 vs_subv(vs1, __ T4S, vs1, vs2);
6657 // store a0 32 (8x4S) coefficients via first offsets
6658 vs_str_indexed(vs3, __ Q, coeffs, i, offsets1);
6659 // b load 32 (8x4S) next inputs
6660 vs_ldpq_post(vs2, zetas);
6661 // reload constants q, qinv -- they were clobbered earlier
6662 vs_ldpq(vq, dilithiumConsts); // qInv, q
6663 // compute a1 = b montmul c
6664 dilithium_montmul32(vs2, vs1, vs2, vtmp, vq);
6665 // store a1 32 (8x4S) coefficients via second offsets
6666 vs_str_indexed(vs2, __ Q, coeffs, i, offsets2);
6667 }
6668
6669 // level 3-7
6670 dilithiumInverseNttLevel3_7(dilithiumConsts, coeffs, zetas);
6671
6672 __ leave(); // required for proper stackwalking of RuntimeStub frame
6673 __ mov(r0, zr); // return 0
6674 __ ret(lr);
6675
6676 return start;
6677 }
6678
6679 // Dilithium multiply polynomials in the NTT domain.
6680 // Straightforward implementation of the method
6681 // static int implDilithiumNttMult(
6682 // int[] result, int[] ntta, int[] nttb {} of
6683 // the sun.security.provider.ML_DSA class.
6684 //
6685 // result (int[256]) = c_rarg0
6686 // poly1 (int[256]) = c_rarg1
6687 // poly2 (int[256]) = c_rarg2
6688 address generate_dilithiumNttMult() {
6689
6690 __ align(CodeEntryAlignment);
6691 StubGenStubId stub_id = StubGenStubId::dilithiumNttMult_id;
6692 StubCodeMark mark(this, stub_id);
6693 address start = __ pc();
6694 __ enter();
6695
6696 Label L_loop;
6697
6698 const Register result = c_rarg0;
6699 const Register poly1 = c_rarg1;
6700 const Register poly2 = c_rarg2;
6701
6702 const Register dilithiumConsts = r10;
6703 const Register len = r11;
6704
6705 VSeq<8> vs1(0), vs2(16), vs3(24); // 3 sets of 8x4s inputs/outputs
6706 VSeq<4> vtmp = vs_front(vs3); // n.b. tmp registers overlap vs3
6707 VSeq<2> vq(30); // n.b. constants overlap vs3
6708 VSeq<8> vrsquare(29, 0); // for montmul by constant RSQUARE
6709
6710 __ lea(dilithiumConsts,
6711 ExternalAddress((address) StubRoutines::aarch64::_dilithiumConsts));
6712
6713 // load constants q, qinv
6714 vs_ldpq(vq, dilithiumConsts); // qInv, q
6715 // load constant rSquare into v29
6716 __ ldr(v29, __ Q, Address(dilithiumConsts, 48)); // rSquare
6717
6718 __ mov(len, zr);
6719 __ add(len, len, 1024);
6720
6721 __ BIND(L_loop);
6722
6723 // b load 32 (8x4S) next inputs from poly1
6724 vs_ldpq_post(vs1, poly1);
6725 // c load 32 (8x4S) next inputs from poly2
6726 vs_ldpq_post(vs2, poly2);
6727 // compute a = b montmul c
6728 dilithium_montmul32(vs2, vs1, vs2, vtmp, vq);
6729 // compute a = rsquare montmul a
6730 dilithium_montmul32(vs2, vrsquare, vs2, vtmp, vq);
6731 // save a 32 (8x4S) results
6732 vs_stpq_post(vs2, result);
6733
6734 __ sub(len, len, 128);
6735 __ cmp(len, (u1)128);
6736 __ br(Assembler::GE, L_loop);
6737
6738 __ leave(); // required for proper stackwalking of RuntimeStub frame
6739 __ mov(r0, zr); // return 0
6740 __ ret(lr);
6741
6742 return start;
6743 }
6744
6745 // Dilithium Motgomery multiply an array by a constant.
6746 // A straightforward implementation of the method
6747 // static int implDilithiumMontMulByConstant(int[] coeffs, int constant) {}
6748 // of the sun.security.provider.MLDSA class
6749 //
6750 // coeffs (int[256]) = c_rarg0
6751 // constant (int) = c_rarg1
6752 address generate_dilithiumMontMulByConstant() {
6753
6754 __ align(CodeEntryAlignment);
6755 StubGenStubId stub_id = StubGenStubId::dilithiumMontMulByConstant_id;
6756 StubCodeMark mark(this, stub_id);
6757 address start = __ pc();
6758 __ enter();
6759
6760 Label L_loop;
6761
6762 const Register coeffs = c_rarg0;
6763 const Register constant = c_rarg1;
6764
6765 const Register dilithiumConsts = r10;
6766 const Register result = r11;
6767 const Register len = r12;
6768
6769 VSeq<8> vs1(0), vs2(16), vs3(24); // 3 sets of 8x4s inputs/outputs
6770 VSeq<4> vtmp = vs_front(vs3); // n.b. tmp registers overlap vs3
6771 VSeq<2> vq(30); // n.b. constants overlap vs3
6772 VSeq<8> vconst(29, 0); // for montmul by constant
6773
6774 // results track inputs
6775 __ add(result, coeffs, 0);
6776 __ lea(dilithiumConsts,
6777 ExternalAddress((address) StubRoutines::aarch64::_dilithiumConsts));
6778
6779 // load constants q, qinv -- they do not get clobbered by first two loops
6780 vs_ldpq(vq, dilithiumConsts); // qInv, q
6781 // copy caller supplied constant across vconst
6782 __ dup(vconst[0], __ T4S, constant);
6783 __ mov(len, zr);
6784 __ add(len, len, 1024);
6785
6786 __ BIND(L_loop);
6787
6788 // load next 32 inputs
6789 vs_ldpq_post(vs2, coeffs);
6790 // mont mul by constant
6791 dilithium_montmul32(vs2, vconst, vs2, vtmp, vq);
6792 // write next 32 results
6793 vs_stpq_post(vs2, result);
6794
6795 __ sub(len, len, 128);
6796 __ cmp(len, (u1)128);
6797 __ br(Assembler::GE, L_loop);
6798
6799 __ leave(); // required for proper stackwalking of RuntimeStub frame
6800 __ mov(r0, zr); // return 0
6801 __ ret(lr);
6802
6803 return start;
6804 }
6805
6806 // Dilithium decompose poly.
6807 // Implements the method
6808 // static int implDilithiumDecomposePoly(int[] coeffs, int constant) {}
6809 // of the sun.security.provider.ML_DSA class
6810 //
6811 // input (int[256]) = c_rarg0
6812 // lowPart (int[256]) = c_rarg1
6813 // highPart (int[256]) = c_rarg2
6814 // twoGamma2 (int) = c_rarg3
6815 // multiplier (int) = c_rarg4
6816 address generate_dilithiumDecomposePoly() {
6817
6818 __ align(CodeEntryAlignment);
6819 StubGenStubId stub_id = StubGenStubId::dilithiumDecomposePoly_id;
6820 StubCodeMark mark(this, stub_id);
6821 address start = __ pc();
6822 Label L_loop;
6823
6824 const Register input = c_rarg0;
6825 const Register lowPart = c_rarg1;
6826 const Register highPart = c_rarg2;
6827 const Register twoGamma2 = c_rarg3;
6828 const Register multiplier = c_rarg4;
6829
6830 const Register len = r9;
6831 const Register dilithiumConsts = r10;
6832 const Register tmp = r11;
6833
6834 // 6 independent sets of 4x4s values
6835 VSeq<4> vs1(0), vs2(4), vs3(8);
6836 VSeq<4> vs4(12), vs5(16), vtmp(20);
6837
6838 // 7 constants for cross-multiplying
6839 VSeq<4> one(25, 0);
6840 VSeq<4> qminus1(26, 0);
6841 VSeq<4> g2(27, 0);
6842 VSeq<4> twog2(28, 0);
6843 VSeq<4> mult(29, 0);
6844 VSeq<4> q(30, 0);
6845 VSeq<4> qadd(31, 0);
6846
6847 __ enter();
6848
6849 __ lea(dilithiumConsts,
6850 ExternalAddress((address) StubRoutines::aarch64::_dilithiumConsts));
6851
6852 // save callee-saved registers
6853 __ stpd(v8, v9, __ pre(sp, -64));
6854 __ stpd(v10, v11, Address(sp, 16));
6855 __ stpd(v12, v13, Address(sp, 32));
6856 __ stpd(v14, v15, Address(sp, 48));
6857
6858 // populate constant registers
6859 __ mov(tmp, zr);
6860 __ add(tmp, tmp, 1);
6861 __ dup(one[0], __ T4S, tmp); // 1
6862 __ ldr(q[0], __ Q, Address(dilithiumConsts, 16)); // q
6863 __ ldr(qadd[0], __ Q, Address(dilithiumConsts, 64)); // addend for mod q reduce
6864 __ dup(twog2[0], __ T4S, twoGamma2); // 2 * gamma2
6865 __ dup(mult[0], __ T4S, multiplier); // multiplier for mod 2 * gamma reduce
6866 __ subv(qminus1[0], __ T4S, v30, v25); // q - 1
6867 __ sshr(g2[0], __ T4S, v28, 1); // gamma2
6868
6869 __ mov(len, zr);
6870 __ add(len, len, 1024);
6871
6872 __ BIND(L_loop);
6873
6874 // load next 4x4S inputs interleaved: rplus --> vs1
6875 __ ld4(vs1[0], vs1[1], vs1[2], vs1[3], __ T4S, __ post(input, 64));
6876
6877 // rplus = rplus - ((rplus + qadd) >> 23) * q
6878 vs_addv(vtmp, __ T4S, vs1, qadd);
6879 vs_sshr(vtmp, __ T4S, vtmp, 23);
6880 vs_mulv(vtmp, __ T4S, vtmp, q);
6881 vs_subv(vs1, __ T4S, vs1, vtmp);
6882
6883 // rplus = rplus + ((rplus >> 31) & dilithium_q);
6884 vs_sshr(vtmp, __ T4S, vs1, 31);
6885 vs_andr(vtmp, vtmp, q);
6886 vs_addv(vs1, __ T4S, vs1, vtmp);
6887
6888 // quotient --> vs2
6889 // int quotient = (rplus * multiplier) >> 22;
6890 vs_mulv(vtmp, __ T4S, vs1, mult);
6891 vs_sshr(vs2, __ T4S, vtmp, 22);
6892
6893 // r0 --> vs3
6894 // int r0 = rplus - quotient * twoGamma2;
6895 vs_mulv(vtmp, __ T4S, vs2, twog2);
6896 vs_subv(vs3, __ T4S, vs1, vtmp);
6897
6898 // mask --> vs4
6899 // int mask = (twoGamma2 - r0) >> 22;
6900 vs_subv(vtmp, __ T4S, twog2, vs3);
6901 vs_sshr(vs4, __ T4S, vtmp, 22);
6902
6903 // r0 -= (mask & twoGamma2);
6904 vs_andr(vtmp, vs4, twog2);
6905 vs_subv(vs3, __ T4S, vs3, vtmp);
6906
6907 // quotient += (mask & 1);
6908 vs_andr(vtmp, vs4, one);
6909 vs_addv(vs2, __ T4S, vs2, vtmp);
6910
6911 // mask = (twoGamma2 / 2 - r0) >> 31;
6912 vs_subv(vtmp, __ T4S, g2, vs3);
6913 vs_sshr(vs4, __ T4S, vtmp, 31);
6914
6915 // r0 -= (mask & twoGamma2);
6916 vs_andr(vtmp, vs4, twog2);
6917 vs_subv(vs3, __ T4S, vs3, vtmp);
6918
6919 // quotient += (mask & 1);
6920 vs_andr(vtmp, vs4, one);
6921 vs_addv(vs2, __ T4S, vs2, vtmp);
6922
6923 // r1 --> vs5
6924 // int r1 = rplus - r0 - (dilithium_q - 1);
6925 vs_subv(vtmp, __ T4S, vs1, vs3);
6926 vs_subv(vs5, __ T4S, vtmp, qminus1);
6927
6928 // r1 --> vs1 (overwriting rplus)
6929 // r1 = (r1 | (-r1)) >> 31; // 0 if rplus - r0 == (dilithium_q - 1), -1 otherwise
6930 vs_negr(vtmp, __ T4S, vs5);
6931 vs_orr(vtmp, vs5, vtmp);
6932 vs_sshr(vs1, __ T4S, vtmp, 31);
6933
6934 // r0 += ~r1;
6935 vs_notr(vtmp, vs1);
6936 vs_addv(vs3, __ T4S, vs3, vtmp);
6937
6938 // r1 = r1 & quotient;
6939 vs_andr(vs1, vs2, vs1);
6940
6941 // store results inteleaved
6942 // lowPart[m] = r0;
6943 // highPart[m] = r1;
6944 __ st4(vs3[0], vs3[1], vs3[2], vs3[3], __ T4S, __ post(lowPart, 64));
6945 __ st4(vs1[0], vs1[1], vs1[2], vs1[3], __ T4S, __ post(highPart, 64));
6946
6947 __ sub(len, len, 64);
6948 __ cmp(len, (u1)64);
6949 __ br(Assembler::GE, L_loop);
6950
6951 // restore callee-saved vector registers
6952 __ ldpd(v14, v15, Address(sp, 48));
6953 __ ldpd(v12, v13, Address(sp, 32));
6954 __ ldpd(v10, v11, Address(sp, 16));
6955 __ ldpd(v8, v9, __ post(sp, 64));
6956
6957 __ leave(); // required for proper stackwalking of RuntimeStub frame
6958 __ mov(r0, zr); // return 0
6959 __ ret(lr);
6960
6961 return start;
6962 }
6963
6964 /**
6965 * Arguments:
6966 *
6967 * Inputs:
6968 * c_rarg0 - int crc
6969 * c_rarg1 - byte* buf
6970 * c_rarg2 - int length
6971 *
6972 * Output:
6973 * rax - int crc result
6974 */
6975 address generate_updateBytesCRC32() {
6976 assert(UseCRC32Intrinsics, "what are we doing here?");
6977
6978 __ align(CodeEntryAlignment);
6979 StubGenStubId stub_id = StubGenStubId::updateBytesCRC32_id;
6980 StubCodeMark mark(this, stub_id);
6981
6982 address start = __ pc();
6983
6984 const Register crc = c_rarg0; // crc
6985 const Register buf = c_rarg1; // source java byte array address
6986 const Register len = c_rarg2; // length
6987 const Register table0 = c_rarg3; // crc_table address
6988 const Register table1 = c_rarg4;
6989 const Register table2 = c_rarg5;
6990 const Register table3 = c_rarg6;
6991 const Register tmp3 = c_rarg7;
6992
6993 BLOCK_COMMENT("Entry:");
6994 __ enter(); // required for proper stackwalking of RuntimeStub frame
6995
6996 __ kernel_crc32(crc, buf, len,
6997 table0, table1, table2, table3, rscratch1, rscratch2, tmp3);
6998
6999 __ leave(); // required for proper stackwalking of RuntimeStub frame
7000 __ ret(lr);
7001
7002 return start;
7003 }
7004
7005 /**
7006 * Arguments:
7007 *
7008 * Inputs:
7009 * c_rarg0 - int crc
7010 * c_rarg1 - byte* buf
7011 * c_rarg2 - int length
7012 * c_rarg3 - int* table
7013 *
7014 * Output:
7015 * r0 - int crc result
7016 */
7017 address generate_updateBytesCRC32C() {
7018 assert(UseCRC32CIntrinsics, "what are we doing here?");
7019
7020 __ align(CodeEntryAlignment);
7021 StubGenStubId stub_id = StubGenStubId::updateBytesCRC32C_id;
7022 StubCodeMark mark(this, stub_id);
7023
7024 address start = __ pc();
7025
7026 const Register crc = c_rarg0; // crc
7027 const Register buf = c_rarg1; // source java byte array address
7028 const Register len = c_rarg2; // length
7029 const Register table0 = c_rarg3; // crc_table address
7030 const Register table1 = c_rarg4;
7031 const Register table2 = c_rarg5;
7032 const Register table3 = c_rarg6;
7033 const Register tmp3 = c_rarg7;
7034
7035 BLOCK_COMMENT("Entry:");
7036 __ enter(); // required for proper stackwalking of RuntimeStub frame
7037
7038 __ kernel_crc32c(crc, buf, len,
7039 table0, table1, table2, table3, rscratch1, rscratch2, tmp3);
7040
7041 __ leave(); // required for proper stackwalking of RuntimeStub frame
7042 __ ret(lr);
7043
7044 return start;
7045 }
7046
7047 /***
7048 * Arguments:
7049 *
7050 * Inputs:
7051 * c_rarg0 - int adler
7052 * c_rarg1 - byte* buff
7053 * c_rarg2 - int len
7054 *
7055 * Output:
7056 * c_rarg0 - int adler result
7057 */
7058 address generate_updateBytesAdler32() {
7059 __ align(CodeEntryAlignment);
7060 StubGenStubId stub_id = StubGenStubId::updateBytesAdler32_id;
7061 StubCodeMark mark(this, stub_id);
7062 address start = __ pc();
7063
7064 Label L_simple_by1_loop, L_nmax, L_nmax_loop, L_by16, L_by16_loop, L_by1_loop, L_do_mod, L_combine, L_by1;
7065
7066 // Aliases
7067 Register adler = c_rarg0;
7068 Register s1 = c_rarg0;
7069 Register s2 = c_rarg3;
7070 Register buff = c_rarg1;
7071 Register len = c_rarg2;
7072 Register nmax = r4;
7073 Register base = r5;
7074 Register count = r6;
7075 Register temp0 = rscratch1;
7076 Register temp1 = rscratch2;
7077 FloatRegister vbytes = v0;
7078 FloatRegister vs1acc = v1;
7079 FloatRegister vs2acc = v2;
7080 FloatRegister vtable = v3;
7081
7082 // Max number of bytes we can process before having to take the mod
7083 // 0x15B0 is 5552 in decimal, the largest n such that 255n(n+1)/2 + (n+1)(BASE-1) <= 2^32-1
7084 uint64_t BASE = 0xfff1;
7085 uint64_t NMAX = 0x15B0;
7086
7087 __ mov(base, BASE);
7088 __ mov(nmax, NMAX);
7089
7090 // Load accumulation coefficients for the upper 16 bits
7091 __ lea(temp0, ExternalAddress((address) StubRoutines::aarch64::_adler_table));
7092 __ ld1(vtable, __ T16B, Address(temp0));
7093
7094 // s1 is initialized to the lower 16 bits of adler
7095 // s2 is initialized to the upper 16 bits of adler
7096 __ ubfx(s2, adler, 16, 16); // s2 = ((adler >> 16) & 0xffff)
7097 __ uxth(s1, adler); // s1 = (adler & 0xffff)
7098
7099 // The pipelined loop needs at least 16 elements for 1 iteration
7100 // It does check this, but it is more effective to skip to the cleanup loop
7101 __ cmp(len, (u1)16);
7102 __ br(Assembler::HS, L_nmax);
7103 __ cbz(len, L_combine);
7104
7105 __ bind(L_simple_by1_loop);
7106 __ ldrb(temp0, Address(__ post(buff, 1)));
7107 __ add(s1, s1, temp0);
7108 __ add(s2, s2, s1);
7109 __ subs(len, len, 1);
7110 __ br(Assembler::HI, L_simple_by1_loop);
7111
7112 // s1 = s1 % BASE
7113 __ subs(temp0, s1, base);
7114 __ csel(s1, temp0, s1, Assembler::HS);
7115
7116 // s2 = s2 % BASE
7117 __ lsr(temp0, s2, 16);
7118 __ lsl(temp1, temp0, 4);
7119 __ sub(temp1, temp1, temp0);
7120 __ add(s2, temp1, s2, ext::uxth);
7121
7122 __ subs(temp0, s2, base);
7123 __ csel(s2, temp0, s2, Assembler::HS);
7124
7125 __ b(L_combine);
7126
7127 __ bind(L_nmax);
7128 __ subs(len, len, nmax);
7129 __ sub(count, nmax, 16);
7130 __ br(Assembler::LO, L_by16);
7131
7132 __ bind(L_nmax_loop);
7133
7134 generate_updateBytesAdler32_accum(s1, s2, buff, temp0, temp1,
7135 vbytes, vs1acc, vs2acc, vtable);
7136
7137 __ subs(count, count, 16);
7138 __ br(Assembler::HS, L_nmax_loop);
7139
7140 // s1 = s1 % BASE
7141 __ lsr(temp0, s1, 16);
7142 __ lsl(temp1, temp0, 4);
7143 __ sub(temp1, temp1, temp0);
7144 __ add(temp1, temp1, s1, ext::uxth);
7145
7146 __ lsr(temp0, temp1, 16);
7147 __ lsl(s1, temp0, 4);
7148 __ sub(s1, s1, temp0);
7149 __ add(s1, s1, temp1, ext:: uxth);
7150
7151 __ subs(temp0, s1, base);
7152 __ csel(s1, temp0, s1, Assembler::HS);
7153
7154 // s2 = s2 % BASE
7155 __ lsr(temp0, s2, 16);
7156 __ lsl(temp1, temp0, 4);
7157 __ sub(temp1, temp1, temp0);
7158 __ add(temp1, temp1, s2, ext::uxth);
7159
7160 __ lsr(temp0, temp1, 16);
7161 __ lsl(s2, temp0, 4);
7162 __ sub(s2, s2, temp0);
7163 __ add(s2, s2, temp1, ext:: uxth);
7164
7165 __ subs(temp0, s2, base);
7166 __ csel(s2, temp0, s2, Assembler::HS);
7167
7168 __ subs(len, len, nmax);
7169 __ sub(count, nmax, 16);
7170 __ br(Assembler::HS, L_nmax_loop);
7171
7172 __ bind(L_by16);
7173 __ adds(len, len, count);
7174 __ br(Assembler::LO, L_by1);
7175
7176 __ bind(L_by16_loop);
7177
7178 generate_updateBytesAdler32_accum(s1, s2, buff, temp0, temp1,
7179 vbytes, vs1acc, vs2acc, vtable);
7180
7181 __ subs(len, len, 16);
7182 __ br(Assembler::HS, L_by16_loop);
7183
7184 __ bind(L_by1);
7185 __ adds(len, len, 15);
7186 __ br(Assembler::LO, L_do_mod);
7187
7188 __ bind(L_by1_loop);
7189 __ ldrb(temp0, Address(__ post(buff, 1)));
7190 __ add(s1, temp0, s1);
7191 __ add(s2, s2, s1);
7192 __ subs(len, len, 1);
7193 __ br(Assembler::HS, L_by1_loop);
7194
7195 __ bind(L_do_mod);
7196 // s1 = s1 % BASE
7197 __ lsr(temp0, s1, 16);
7198 __ lsl(temp1, temp0, 4);
7199 __ sub(temp1, temp1, temp0);
7200 __ add(temp1, temp1, s1, ext::uxth);
7201
7202 __ lsr(temp0, temp1, 16);
7203 __ lsl(s1, temp0, 4);
7204 __ sub(s1, s1, temp0);
7205 __ add(s1, s1, temp1, ext:: uxth);
7206
7207 __ subs(temp0, s1, base);
7208 __ csel(s1, temp0, s1, Assembler::HS);
7209
7210 // s2 = s2 % BASE
7211 __ lsr(temp0, s2, 16);
7212 __ lsl(temp1, temp0, 4);
7213 __ sub(temp1, temp1, temp0);
7214 __ add(temp1, temp1, s2, ext::uxth);
7215
7216 __ lsr(temp0, temp1, 16);
7217 __ lsl(s2, temp0, 4);
7218 __ sub(s2, s2, temp0);
7219 __ add(s2, s2, temp1, ext:: uxth);
7220
7221 __ subs(temp0, s2, base);
7222 __ csel(s2, temp0, s2, Assembler::HS);
7223
7224 // Combine lower bits and higher bits
7225 __ bind(L_combine);
7226 __ orr(s1, s1, s2, Assembler::LSL, 16); // adler = s1 | (s2 << 16)
7227
7228 __ ret(lr);
7229
7230 return start;
7231 }
7232
7233 void generate_updateBytesAdler32_accum(Register s1, Register s2, Register buff,
7234 Register temp0, Register temp1, FloatRegister vbytes,
7235 FloatRegister vs1acc, FloatRegister vs2acc, FloatRegister vtable) {
7236 // Below is a vectorized implementation of updating s1 and s2 for 16 bytes.
7237 // We use b1, b2, ..., b16 to denote the 16 bytes loaded in each iteration.
7238 // In non-vectorized code, we update s1 and s2 as:
7239 // s1 <- s1 + b1
7240 // s2 <- s2 + s1
7241 // s1 <- s1 + b2
7242 // s2 <- s2 + b1
7243 // ...
7244 // s1 <- s1 + b16
7245 // s2 <- s2 + s1
7246 // Putting above assignments together, we have:
7247 // s1_new = s1 + b1 + b2 + ... + b16
7248 // s2_new = s2 + (s1 + b1) + (s1 + b1 + b2) + ... + (s1 + b1 + b2 + ... + b16)
7249 // = s2 + s1 * 16 + (b1 * 16 + b2 * 15 + ... + b16 * 1)
7250 // = s2 + s1 * 16 + (b1, b2, ... b16) dot (16, 15, ... 1)
7251 __ ld1(vbytes, __ T16B, Address(__ post(buff, 16)));
7252
7253 // s2 = s2 + s1 * 16
7254 __ add(s2, s2, s1, Assembler::LSL, 4);
7255
7256 // vs1acc = b1 + b2 + b3 + ... + b16
7257 // vs2acc = (b1 * 16) + (b2 * 15) + (b3 * 14) + ... + (b16 * 1)
7258 __ umullv(vs2acc, __ T8B, vtable, vbytes);
7259 __ umlalv(vs2acc, __ T16B, vtable, vbytes);
7260 __ uaddlv(vs1acc, __ T16B, vbytes);
7261 __ uaddlv(vs2acc, __ T8H, vs2acc);
7262
7263 // s1 = s1 + vs1acc, s2 = s2 + vs2acc
7264 __ fmovd(temp0, vs1acc);
7265 __ fmovd(temp1, vs2acc);
7266 __ add(s1, s1, temp0);
7267 __ add(s2, s2, temp1);
7268 }
7269
7270 /**
7271 * Arguments:
7272 *
7273 * Input:
7274 * c_rarg0 - x address
7275 * c_rarg1 - x length
7276 * c_rarg2 - y address
7277 * c_rarg3 - y length
7278 * c_rarg4 - z address
7279 */
7280 address generate_multiplyToLen() {
7281 __ align(CodeEntryAlignment);
7282 StubGenStubId stub_id = StubGenStubId::multiplyToLen_id;
7283 StubCodeMark mark(this, stub_id);
7284
7285 address start = __ pc();
7286 const Register x = r0;
7287 const Register xlen = r1;
7288 const Register y = r2;
7289 const Register ylen = r3;
7290 const Register z = r4;
7291
7292 const Register tmp0 = r5;
7293 const Register tmp1 = r10;
7294 const Register tmp2 = r11;
7295 const Register tmp3 = r12;
7296 const Register tmp4 = r13;
7297 const Register tmp5 = r14;
7298 const Register tmp6 = r15;
7299 const Register tmp7 = r16;
7300
7301 BLOCK_COMMENT("Entry:");
7302 __ enter(); // required for proper stackwalking of RuntimeStub frame
7303 __ multiply_to_len(x, xlen, y, ylen, z, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7);
7304 __ leave(); // required for proper stackwalking of RuntimeStub frame
7305 __ ret(lr);
7306
7307 return start;
7308 }
7309
7310 address generate_squareToLen() {
7311 // squareToLen algorithm for sizes 1..127 described in java code works
7312 // faster than multiply_to_len on some CPUs and slower on others, but
7313 // multiply_to_len shows a bit better overall results
7314 __ align(CodeEntryAlignment);
7315 StubGenStubId stub_id = StubGenStubId::squareToLen_id;
7316 StubCodeMark mark(this, stub_id);
7317 address start = __ pc();
7318
7319 const Register x = r0;
7320 const Register xlen = r1;
7321 const Register z = r2;
7322 const Register y = r4; // == x
7323 const Register ylen = r5; // == xlen
7324
7325 const Register tmp0 = r3;
7326 const Register tmp1 = r10;
7327 const Register tmp2 = r11;
7328 const Register tmp3 = r12;
7329 const Register tmp4 = r13;
7330 const Register tmp5 = r14;
7331 const Register tmp6 = r15;
7332 const Register tmp7 = r16;
7333
7334 RegSet spilled_regs = RegSet::of(y, ylen);
7335 BLOCK_COMMENT("Entry:");
7336 __ enter();
7337 __ push(spilled_regs, sp);
7338 __ mov(y, x);
7339 __ mov(ylen, xlen);
7340 __ multiply_to_len(x, xlen, y, ylen, z, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7);
7341 __ pop(spilled_regs, sp);
7342 __ leave();
7343 __ ret(lr);
7344 return start;
7345 }
7346
7347 address generate_mulAdd() {
7348 __ align(CodeEntryAlignment);
7349 StubGenStubId stub_id = StubGenStubId::mulAdd_id;
7350 StubCodeMark mark(this, stub_id);
7351
7352 address start = __ pc();
7353
7354 const Register out = r0;
7355 const Register in = r1;
7356 const Register offset = r2;
7357 const Register len = r3;
7358 const Register k = r4;
7359
7360 BLOCK_COMMENT("Entry:");
7361 __ enter();
7362 __ mul_add(out, in, offset, len, k);
7363 __ leave();
7364 __ ret(lr);
7365
7366 return start;
7367 }
7368
7369 // Arguments:
7370 //
7371 // Input:
7372 // c_rarg0 - newArr address
7373 // c_rarg1 - oldArr address
7374 // c_rarg2 - newIdx
7375 // c_rarg3 - shiftCount
7376 // c_rarg4 - numIter
7377 //
7378 address generate_bigIntegerRightShift() {
7379 __ align(CodeEntryAlignment);
7380 StubGenStubId stub_id = StubGenStubId::bigIntegerRightShiftWorker_id;
7381 StubCodeMark mark(this, stub_id);
7382 address start = __ pc();
7383
7384 Label ShiftSIMDLoop, ShiftTwoLoop, ShiftThree, ShiftTwo, ShiftOne, Exit;
7385
7386 Register newArr = c_rarg0;
7387 Register oldArr = c_rarg1;
7388 Register newIdx = c_rarg2;
7389 Register shiftCount = c_rarg3;
7390 Register numIter = c_rarg4;
7391 Register idx = numIter;
7392
7393 Register newArrCur = rscratch1;
7394 Register shiftRevCount = rscratch2;
7395 Register oldArrCur = r13;
7396 Register oldArrNext = r14;
7397
7398 FloatRegister oldElem0 = v0;
7399 FloatRegister oldElem1 = v1;
7400 FloatRegister newElem = v2;
7401 FloatRegister shiftVCount = v3;
7402 FloatRegister shiftVRevCount = v4;
7403
7404 __ cbz(idx, Exit);
7405
7406 __ add(newArr, newArr, newIdx, Assembler::LSL, 2);
7407
7408 // left shift count
7409 __ movw(shiftRevCount, 32);
7410 __ subw(shiftRevCount, shiftRevCount, shiftCount);
7411
7412 // numIter too small to allow a 4-words SIMD loop, rolling back
7413 __ cmp(numIter, (u1)4);
7414 __ br(Assembler::LT, ShiftThree);
7415
7416 __ dup(shiftVCount, __ T4S, shiftCount);
7417 __ dup(shiftVRevCount, __ T4S, shiftRevCount);
7418 __ negr(shiftVCount, __ T4S, shiftVCount);
7419
7420 __ BIND(ShiftSIMDLoop);
7421
7422 // Calculate the load addresses
7423 __ sub(idx, idx, 4);
7424 __ add(oldArrNext, oldArr, idx, Assembler::LSL, 2);
7425 __ add(newArrCur, newArr, idx, Assembler::LSL, 2);
7426 __ add(oldArrCur, oldArrNext, 4);
7427
7428 // Load 4 words and process
7429 __ ld1(oldElem0, __ T4S, Address(oldArrCur));
7430 __ ld1(oldElem1, __ T4S, Address(oldArrNext));
7431 __ ushl(oldElem0, __ T4S, oldElem0, shiftVCount);
7432 __ ushl(oldElem1, __ T4S, oldElem1, shiftVRevCount);
7433 __ orr(newElem, __ T16B, oldElem0, oldElem1);
7434 __ st1(newElem, __ T4S, Address(newArrCur));
7435
7436 __ cmp(idx, (u1)4);
7437 __ br(Assembler::LT, ShiftTwoLoop);
7438 __ b(ShiftSIMDLoop);
7439
7440 __ BIND(ShiftTwoLoop);
7441 __ cbz(idx, Exit);
7442 __ cmp(idx, (u1)1);
7443 __ br(Assembler::EQ, ShiftOne);
7444
7445 // Calculate the load addresses
7446 __ sub(idx, idx, 2);
7447 __ add(oldArrNext, oldArr, idx, Assembler::LSL, 2);
7448 __ add(newArrCur, newArr, idx, Assembler::LSL, 2);
7449 __ add(oldArrCur, oldArrNext, 4);
7450
7451 // Load 2 words and process
7452 __ ld1(oldElem0, __ T2S, Address(oldArrCur));
7453 __ ld1(oldElem1, __ T2S, Address(oldArrNext));
7454 __ ushl(oldElem0, __ T2S, oldElem0, shiftVCount);
7455 __ ushl(oldElem1, __ T2S, oldElem1, shiftVRevCount);
7456 __ orr(newElem, __ T8B, oldElem0, oldElem1);
7457 __ st1(newElem, __ T2S, Address(newArrCur));
7458 __ b(ShiftTwoLoop);
7459
7460 __ BIND(ShiftThree);
7461 __ tbz(idx, 1, ShiftOne);
7462 __ tbz(idx, 0, ShiftTwo);
7463 __ ldrw(r10, Address(oldArr, 12));
7464 __ ldrw(r11, Address(oldArr, 8));
7465 __ lsrvw(r10, r10, shiftCount);
7466 __ lslvw(r11, r11, shiftRevCount);
7467 __ orrw(r12, r10, r11);
7468 __ strw(r12, Address(newArr, 8));
7469
7470 __ BIND(ShiftTwo);
7471 __ ldrw(r10, Address(oldArr, 8));
7472 __ ldrw(r11, Address(oldArr, 4));
7473 __ lsrvw(r10, r10, shiftCount);
7474 __ lslvw(r11, r11, shiftRevCount);
7475 __ orrw(r12, r10, r11);
7476 __ strw(r12, Address(newArr, 4));
7477
7478 __ BIND(ShiftOne);
7479 __ ldrw(r10, Address(oldArr, 4));
7480 __ ldrw(r11, Address(oldArr));
7481 __ lsrvw(r10, r10, shiftCount);
7482 __ lslvw(r11, r11, shiftRevCount);
7483 __ orrw(r12, r10, r11);
7484 __ strw(r12, Address(newArr));
7485
7486 __ BIND(Exit);
7487 __ ret(lr);
7488
7489 return start;
7490 }
7491
7492 // Arguments:
7493 //
7494 // Input:
7495 // c_rarg0 - newArr address
7496 // c_rarg1 - oldArr address
7497 // c_rarg2 - newIdx
7498 // c_rarg3 - shiftCount
7499 // c_rarg4 - numIter
7500 //
7501 address generate_bigIntegerLeftShift() {
7502 __ align(CodeEntryAlignment);
7503 StubGenStubId stub_id = StubGenStubId::bigIntegerLeftShiftWorker_id;
7504 StubCodeMark mark(this, stub_id);
7505 address start = __ pc();
7506
7507 Label ShiftSIMDLoop, ShiftTwoLoop, ShiftThree, ShiftTwo, ShiftOne, Exit;
7508
7509 Register newArr = c_rarg0;
7510 Register oldArr = c_rarg1;
7511 Register newIdx = c_rarg2;
7512 Register shiftCount = c_rarg3;
7513 Register numIter = c_rarg4;
7514
7515 Register shiftRevCount = rscratch1;
7516 Register oldArrNext = rscratch2;
7517
7518 FloatRegister oldElem0 = v0;
7519 FloatRegister oldElem1 = v1;
7520 FloatRegister newElem = v2;
7521 FloatRegister shiftVCount = v3;
7522 FloatRegister shiftVRevCount = v4;
7523
7524 __ cbz(numIter, Exit);
7525
7526 __ add(oldArrNext, oldArr, 4);
7527 __ add(newArr, newArr, newIdx, Assembler::LSL, 2);
7528
7529 // right shift count
7530 __ movw(shiftRevCount, 32);
7531 __ subw(shiftRevCount, shiftRevCount, shiftCount);
7532
7533 // numIter too small to allow a 4-words SIMD loop, rolling back
7534 __ cmp(numIter, (u1)4);
7535 __ br(Assembler::LT, ShiftThree);
7536
7537 __ dup(shiftVCount, __ T4S, shiftCount);
7538 __ dup(shiftVRevCount, __ T4S, shiftRevCount);
7539 __ negr(shiftVRevCount, __ T4S, shiftVRevCount);
7540
7541 __ BIND(ShiftSIMDLoop);
7542
7543 // load 4 words and process
7544 __ ld1(oldElem0, __ T4S, __ post(oldArr, 16));
7545 __ ld1(oldElem1, __ T4S, __ post(oldArrNext, 16));
7546 __ ushl(oldElem0, __ T4S, oldElem0, shiftVCount);
7547 __ ushl(oldElem1, __ T4S, oldElem1, shiftVRevCount);
7548 __ orr(newElem, __ T16B, oldElem0, oldElem1);
7549 __ st1(newElem, __ T4S, __ post(newArr, 16));
7550 __ sub(numIter, numIter, 4);
7551
7552 __ cmp(numIter, (u1)4);
7553 __ br(Assembler::LT, ShiftTwoLoop);
7554 __ b(ShiftSIMDLoop);
7555
7556 __ BIND(ShiftTwoLoop);
7557 __ cbz(numIter, Exit);
7558 __ cmp(numIter, (u1)1);
7559 __ br(Assembler::EQ, ShiftOne);
7560
7561 // load 2 words and process
7562 __ ld1(oldElem0, __ T2S, __ post(oldArr, 8));
7563 __ ld1(oldElem1, __ T2S, __ post(oldArrNext, 8));
7564 __ ushl(oldElem0, __ T2S, oldElem0, shiftVCount);
7565 __ ushl(oldElem1, __ T2S, oldElem1, shiftVRevCount);
7566 __ orr(newElem, __ T8B, oldElem0, oldElem1);
7567 __ st1(newElem, __ T2S, __ post(newArr, 8));
7568 __ sub(numIter, numIter, 2);
7569 __ b(ShiftTwoLoop);
7570
7571 __ BIND(ShiftThree);
7572 __ ldrw(r10, __ post(oldArr, 4));
7573 __ ldrw(r11, __ post(oldArrNext, 4));
7574 __ lslvw(r10, r10, shiftCount);
7575 __ lsrvw(r11, r11, shiftRevCount);
7576 __ orrw(r12, r10, r11);
7577 __ strw(r12, __ post(newArr, 4));
7578 __ tbz(numIter, 1, Exit);
7579 __ tbz(numIter, 0, ShiftOne);
7580
7581 __ BIND(ShiftTwo);
7582 __ ldrw(r10, __ post(oldArr, 4));
7583 __ ldrw(r11, __ post(oldArrNext, 4));
7584 __ lslvw(r10, r10, shiftCount);
7585 __ lsrvw(r11, r11, shiftRevCount);
7586 __ orrw(r12, r10, r11);
7587 __ strw(r12, __ post(newArr, 4));
7588
7589 __ BIND(ShiftOne);
7590 __ ldrw(r10, Address(oldArr));
7591 __ ldrw(r11, Address(oldArrNext));
7592 __ lslvw(r10, r10, shiftCount);
7593 __ lsrvw(r11, r11, shiftRevCount);
7594 __ orrw(r12, r10, r11);
7595 __ strw(r12, Address(newArr));
7596
7597 __ BIND(Exit);
7598 __ ret(lr);
7599
7600 return start;
7601 }
7602
7603 address generate_count_positives(address &count_positives_long) {
7604 const u1 large_loop_size = 64;
7605 const uint64_t UPPER_BIT_MASK=0x8080808080808080;
7606 int dcache_line = VM_Version::dcache_line_size();
7607
7608 Register ary1 = r1, len = r2, result = r0;
7609
7610 __ align(CodeEntryAlignment);
7611
7612 StubGenStubId stub_id = StubGenStubId::count_positives_id;
7613 StubCodeMark mark(this, stub_id);
7614
7615 address entry = __ pc();
7616
7617 __ enter();
7618 // precondition: a copy of len is already in result
7619 // __ mov(result, len);
7620
7621 Label RET_ADJUST, RET_ADJUST_16, RET_ADJUST_LONG, RET_NO_POP, RET_LEN, ALIGNED, LOOP16, CHECK_16,
7622 LARGE_LOOP, POST_LOOP16, LEN_OVER_15, LEN_OVER_8, POST_LOOP16_LOAD_TAIL;
7623
7624 __ cmp(len, (u1)15);
7625 __ br(Assembler::GT, LEN_OVER_15);
7626 // The only case when execution falls into this code is when pointer is near
7627 // the end of memory page and we have to avoid reading next page
7628 __ add(ary1, ary1, len);
7629 __ subs(len, len, 8);
7630 __ br(Assembler::GT, LEN_OVER_8);
7631 __ ldr(rscratch2, Address(ary1, -8));
7632 __ sub(rscratch1, zr, len, __ LSL, 3); // LSL 3 is to get bits from bytes.
7633 __ lsrv(rscratch2, rscratch2, rscratch1);
7634 __ tst(rscratch2, UPPER_BIT_MASK);
7635 __ csel(result, zr, result, Assembler::NE);
7636 __ leave();
7637 __ ret(lr);
7638 __ bind(LEN_OVER_8);
7639 __ ldp(rscratch1, rscratch2, Address(ary1, -16));
7640 __ sub(len, len, 8); // no data dep., then sub can be executed while loading
7641 __ tst(rscratch2, UPPER_BIT_MASK);
7642 __ br(Assembler::NE, RET_NO_POP);
7643 __ sub(rscratch2, zr, len, __ LSL, 3); // LSL 3 is to get bits from bytes
7644 __ lsrv(rscratch1, rscratch1, rscratch2);
7645 __ tst(rscratch1, UPPER_BIT_MASK);
7646 __ bind(RET_NO_POP);
7647 __ csel(result, zr, result, Assembler::NE);
7648 __ leave();
7649 __ ret(lr);
7650
7651 Register tmp1 = r3, tmp2 = r4, tmp3 = r5, tmp4 = r6, tmp5 = r7, tmp6 = r10;
7652 const RegSet spilled_regs = RegSet::range(tmp1, tmp5) + tmp6;
7653
7654 count_positives_long = __ pc(); // 2nd entry point
7655
7656 __ enter();
7657
7658 __ bind(LEN_OVER_15);
7659 __ push(spilled_regs, sp);
7660 __ andr(rscratch2, ary1, 15); // check pointer for 16-byte alignment
7661 __ cbz(rscratch2, ALIGNED);
7662 __ ldp(tmp6, tmp1, Address(ary1));
7663 __ mov(tmp5, 16);
7664 __ sub(rscratch1, tmp5, rscratch2); // amount of bytes until aligned address
7665 __ add(ary1, ary1, rscratch1);
7666 __ orr(tmp6, tmp6, tmp1);
7667 __ tst(tmp6, UPPER_BIT_MASK);
7668 __ br(Assembler::NE, RET_ADJUST);
7669 __ sub(len, len, rscratch1);
7670
7671 __ bind(ALIGNED);
7672 __ cmp(len, large_loop_size);
7673 __ br(Assembler::LT, CHECK_16);
7674 // Perform 16-byte load as early return in pre-loop to handle situation
7675 // when initially aligned large array has negative values at starting bytes,
7676 // so LARGE_LOOP would do 4 reads instead of 1 (in worst case), which is
7677 // slower. Cases with negative bytes further ahead won't be affected that
7678 // much. In fact, it'll be faster due to early loads, less instructions and
7679 // less branches in LARGE_LOOP.
7680 __ ldp(tmp6, tmp1, Address(__ post(ary1, 16)));
7681 __ sub(len, len, 16);
7682 __ orr(tmp6, tmp6, tmp1);
7683 __ tst(tmp6, UPPER_BIT_MASK);
7684 __ br(Assembler::NE, RET_ADJUST_16);
7685 __ cmp(len, large_loop_size);
7686 __ br(Assembler::LT, CHECK_16);
7687
7688 if (SoftwarePrefetchHintDistance >= 0
7689 && SoftwarePrefetchHintDistance >= dcache_line) {
7690 // initial prefetch
7691 __ prfm(Address(ary1, SoftwarePrefetchHintDistance - dcache_line));
7692 }
7693 __ bind(LARGE_LOOP);
7694 if (SoftwarePrefetchHintDistance >= 0) {
7695 __ prfm(Address(ary1, SoftwarePrefetchHintDistance));
7696 }
7697 // Issue load instructions first, since it can save few CPU/MEM cycles, also
7698 // instead of 4 triples of "orr(...), addr(...);cbnz(...);" (for each ldp)
7699 // better generate 7 * orr(...) + 1 andr(...) + 1 cbnz(...) which saves 3
7700 // instructions per cycle and have less branches, but this approach disables
7701 // early return, thus, all 64 bytes are loaded and checked every time.
7702 __ ldp(tmp2, tmp3, Address(ary1));
7703 __ ldp(tmp4, tmp5, Address(ary1, 16));
7704 __ ldp(rscratch1, rscratch2, Address(ary1, 32));
7705 __ ldp(tmp6, tmp1, Address(ary1, 48));
7706 __ add(ary1, ary1, large_loop_size);
7707 __ sub(len, len, large_loop_size);
7708 __ orr(tmp2, tmp2, tmp3);
7709 __ orr(tmp4, tmp4, tmp5);
7710 __ orr(rscratch1, rscratch1, rscratch2);
7711 __ orr(tmp6, tmp6, tmp1);
7712 __ orr(tmp2, tmp2, tmp4);
7713 __ orr(rscratch1, rscratch1, tmp6);
7714 __ orr(tmp2, tmp2, rscratch1);
7715 __ tst(tmp2, UPPER_BIT_MASK);
7716 __ br(Assembler::NE, RET_ADJUST_LONG);
7717 __ cmp(len, large_loop_size);
7718 __ br(Assembler::GE, LARGE_LOOP);
7719
7720 __ bind(CHECK_16); // small 16-byte load pre-loop
7721 __ cmp(len, (u1)16);
7722 __ br(Assembler::LT, POST_LOOP16);
7723
7724 __ bind(LOOP16); // small 16-byte load loop
7725 __ ldp(tmp2, tmp3, Address(__ post(ary1, 16)));
7726 __ sub(len, len, 16);
7727 __ orr(tmp2, tmp2, tmp3);
7728 __ tst(tmp2, UPPER_BIT_MASK);
7729 __ br(Assembler::NE, RET_ADJUST_16);
7730 __ cmp(len, (u1)16);
7731 __ br(Assembler::GE, LOOP16); // 16-byte load loop end
7732
7733 __ bind(POST_LOOP16); // 16-byte aligned, so we can read unconditionally
7734 __ cmp(len, (u1)8);
7735 __ br(Assembler::LE, POST_LOOP16_LOAD_TAIL);
7736 __ ldr(tmp3, Address(__ post(ary1, 8)));
7737 __ tst(tmp3, UPPER_BIT_MASK);
7738 __ br(Assembler::NE, RET_ADJUST);
7739 __ sub(len, len, 8);
7740
7741 __ bind(POST_LOOP16_LOAD_TAIL);
7742 __ cbz(len, RET_LEN); // Can't shift left by 64 when len==0
7743 __ ldr(tmp1, Address(ary1));
7744 __ mov(tmp2, 64);
7745 __ sub(tmp4, tmp2, len, __ LSL, 3);
7746 __ lslv(tmp1, tmp1, tmp4);
7747 __ tst(tmp1, UPPER_BIT_MASK);
7748 __ br(Assembler::NE, RET_ADJUST);
7749 // Fallthrough
7750
7751 __ bind(RET_LEN);
7752 __ pop(spilled_regs, sp);
7753 __ leave();
7754 __ ret(lr);
7755
7756 // difference result - len is the count of guaranteed to be
7757 // positive bytes
7758
7759 __ bind(RET_ADJUST_LONG);
7760 __ add(len, len, (u1)(large_loop_size - 16));
7761 __ bind(RET_ADJUST_16);
7762 __ add(len, len, 16);
7763 __ bind(RET_ADJUST);
7764 __ pop(spilled_regs, sp);
7765 __ leave();
7766 __ sub(result, result, len);
7767 __ ret(lr);
7768
7769 return entry;
7770 }
7771
7772 void generate_large_array_equals_loop_nonsimd(int loopThreshold,
7773 bool usePrefetch, Label &NOT_EQUAL) {
7774 Register a1 = r1, a2 = r2, result = r0, cnt1 = r10, tmp1 = rscratch1,
7775 tmp2 = rscratch2, tmp3 = r3, tmp4 = r4, tmp5 = r5, tmp6 = r11,
7776 tmp7 = r12, tmp8 = r13;
7777 Label LOOP;
7778
7779 __ ldp(tmp1, tmp3, Address(__ post(a1, 2 * wordSize)));
7780 __ ldp(tmp2, tmp4, Address(__ post(a2, 2 * wordSize)));
7781 __ bind(LOOP);
7782 if (usePrefetch) {
7783 __ prfm(Address(a1, SoftwarePrefetchHintDistance));
7784 __ prfm(Address(a2, SoftwarePrefetchHintDistance));
7785 }
7786 __ ldp(tmp5, tmp7, Address(__ post(a1, 2 * wordSize)));
7787 __ eor(tmp1, tmp1, tmp2);
7788 __ eor(tmp3, tmp3, tmp4);
7789 __ ldp(tmp6, tmp8, Address(__ post(a2, 2 * wordSize)));
7790 __ orr(tmp1, tmp1, tmp3);
7791 __ cbnz(tmp1, NOT_EQUAL);
7792 __ ldp(tmp1, tmp3, Address(__ post(a1, 2 * wordSize)));
7793 __ eor(tmp5, tmp5, tmp6);
7794 __ eor(tmp7, tmp7, tmp8);
7795 __ ldp(tmp2, tmp4, Address(__ post(a2, 2 * wordSize)));
7796 __ orr(tmp5, tmp5, tmp7);
7797 __ cbnz(tmp5, NOT_EQUAL);
7798 __ ldp(tmp5, tmp7, Address(__ post(a1, 2 * wordSize)));
7799 __ eor(tmp1, tmp1, tmp2);
7800 __ eor(tmp3, tmp3, tmp4);
7801 __ ldp(tmp6, tmp8, Address(__ post(a2, 2 * wordSize)));
7802 __ orr(tmp1, tmp1, tmp3);
7803 __ cbnz(tmp1, NOT_EQUAL);
7804 __ ldp(tmp1, tmp3, Address(__ post(a1, 2 * wordSize)));
7805 __ eor(tmp5, tmp5, tmp6);
7806 __ sub(cnt1, cnt1, 8 * wordSize);
7807 __ eor(tmp7, tmp7, tmp8);
7808 __ ldp(tmp2, tmp4, Address(__ post(a2, 2 * wordSize)));
7809 // tmp6 is not used. MacroAssembler::subs is used here (rather than
7810 // cmp) because subs allows an unlimited range of immediate operand.
7811 __ subs(tmp6, cnt1, loopThreshold);
7812 __ orr(tmp5, tmp5, tmp7);
7813 __ cbnz(tmp5, NOT_EQUAL);
7814 __ br(__ GE, LOOP);
7815 // post-loop
7816 __ eor(tmp1, tmp1, tmp2);
7817 __ eor(tmp3, tmp3, tmp4);
7818 __ orr(tmp1, tmp1, tmp3);
7819 __ sub(cnt1, cnt1, 2 * wordSize);
7820 __ cbnz(tmp1, NOT_EQUAL);
7821 }
7822
7823 void generate_large_array_equals_loop_simd(int loopThreshold,
7824 bool usePrefetch, Label &NOT_EQUAL) {
7825 Register a1 = r1, a2 = r2, result = r0, cnt1 = r10, tmp1 = rscratch1,
7826 tmp2 = rscratch2;
7827 Label LOOP;
7828
7829 __ bind(LOOP);
7830 if (usePrefetch) {
7831 __ prfm(Address(a1, SoftwarePrefetchHintDistance));
7832 __ prfm(Address(a2, SoftwarePrefetchHintDistance));
7833 }
7834 __ ld1(v0, v1, v2, v3, __ T2D, Address(__ post(a1, 4 * 2 * wordSize)));
7835 __ sub(cnt1, cnt1, 8 * wordSize);
7836 __ ld1(v4, v5, v6, v7, __ T2D, Address(__ post(a2, 4 * 2 * wordSize)));
7837 __ subs(tmp1, cnt1, loopThreshold);
7838 __ eor(v0, __ T16B, v0, v4);
7839 __ eor(v1, __ T16B, v1, v5);
7840 __ eor(v2, __ T16B, v2, v6);
7841 __ eor(v3, __ T16B, v3, v7);
7842 __ orr(v0, __ T16B, v0, v1);
7843 __ orr(v1, __ T16B, v2, v3);
7844 __ orr(v0, __ T16B, v0, v1);
7845 __ umov(tmp1, v0, __ D, 0);
7846 __ umov(tmp2, v0, __ D, 1);
7847 __ orr(tmp1, tmp1, tmp2);
7848 __ cbnz(tmp1, NOT_EQUAL);
7849 __ br(__ GE, LOOP);
7850 }
7851
7852 // a1 = r1 - array1 address
7853 // a2 = r2 - array2 address
7854 // result = r0 - return value. Already contains "false"
7855 // cnt1 = r10 - amount of elements left to check, reduced by wordSize
7856 // r3-r5 are reserved temporary registers
7857 // Clobbers: v0-v7 when UseSIMDForArrayEquals, rscratch1, rscratch2
7858 address generate_large_array_equals() {
7859 Register a1 = r1, a2 = r2, result = r0, cnt1 = r10, tmp1 = rscratch1,
7860 tmp2 = rscratch2, tmp3 = r3, tmp4 = r4, tmp5 = r5, tmp6 = r11,
7861 tmp7 = r12, tmp8 = r13;
7862 Label TAIL, NOT_EQUAL, EQUAL, NOT_EQUAL_NO_POP, NO_PREFETCH_LARGE_LOOP,
7863 SMALL_LOOP, POST_LOOP;
7864 const int PRE_LOOP_SIZE = UseSIMDForArrayEquals ? 0 : 16;
7865 // calculate if at least 32 prefetched bytes are used
7866 int prefetchLoopThreshold = SoftwarePrefetchHintDistance + 32;
7867 int nonPrefetchLoopThreshold = (64 + PRE_LOOP_SIZE);
7868 RegSet spilled_regs = RegSet::range(tmp6, tmp8);
7869 assert_different_registers(a1, a2, result, cnt1, tmp1, tmp2, tmp3, tmp4,
7870 tmp5, tmp6, tmp7, tmp8);
7871
7872 __ align(CodeEntryAlignment);
7873
7874 StubGenStubId stub_id = StubGenStubId::large_array_equals_id;
7875 StubCodeMark mark(this, stub_id);
7876
7877 address entry = __ pc();
7878 __ enter();
7879 __ sub(cnt1, cnt1, wordSize); // first 8 bytes were loaded outside of stub
7880 // also advance pointers to use post-increment instead of pre-increment
7881 __ add(a1, a1, wordSize);
7882 __ add(a2, a2, wordSize);
7883 if (AvoidUnalignedAccesses) {
7884 // both implementations (SIMD/nonSIMD) are using relatively large load
7885 // instructions (ld1/ldp), which has huge penalty (up to x2 exec time)
7886 // on some CPUs in case of address is not at least 16-byte aligned.
7887 // Arrays are 8-byte aligned currently, so, we can make additional 8-byte
7888 // load if needed at least for 1st address and make if 16-byte aligned.
7889 Label ALIGNED16;
7890 __ tbz(a1, 3, ALIGNED16);
7891 __ ldr(tmp1, Address(__ post(a1, wordSize)));
7892 __ ldr(tmp2, Address(__ post(a2, wordSize)));
7893 __ sub(cnt1, cnt1, wordSize);
7894 __ eor(tmp1, tmp1, tmp2);
7895 __ cbnz(tmp1, NOT_EQUAL_NO_POP);
7896 __ bind(ALIGNED16);
7897 }
7898 if (UseSIMDForArrayEquals) {
7899 if (SoftwarePrefetchHintDistance >= 0) {
7900 __ subs(tmp1, cnt1, prefetchLoopThreshold);
7901 __ br(__ LE, NO_PREFETCH_LARGE_LOOP);
7902 generate_large_array_equals_loop_simd(prefetchLoopThreshold,
7903 /* prfm = */ true, NOT_EQUAL);
7904 __ subs(zr, cnt1, nonPrefetchLoopThreshold);
7905 __ br(__ LT, TAIL);
7906 }
7907 __ bind(NO_PREFETCH_LARGE_LOOP);
7908 generate_large_array_equals_loop_simd(nonPrefetchLoopThreshold,
7909 /* prfm = */ false, NOT_EQUAL);
7910 } else {
7911 __ push(spilled_regs, sp);
7912 if (SoftwarePrefetchHintDistance >= 0) {
7913 __ subs(tmp1, cnt1, prefetchLoopThreshold);
7914 __ br(__ LE, NO_PREFETCH_LARGE_LOOP);
7915 generate_large_array_equals_loop_nonsimd(prefetchLoopThreshold,
7916 /* prfm = */ true, NOT_EQUAL);
7917 __ subs(zr, cnt1, nonPrefetchLoopThreshold);
7918 __ br(__ LT, TAIL);
7919 }
7920 __ bind(NO_PREFETCH_LARGE_LOOP);
7921 generate_large_array_equals_loop_nonsimd(nonPrefetchLoopThreshold,
7922 /* prfm = */ false, NOT_EQUAL);
7923 }
7924 __ bind(TAIL);
7925 __ cbz(cnt1, EQUAL);
7926 __ subs(cnt1, cnt1, wordSize);
7927 __ br(__ LE, POST_LOOP);
7928 __ bind(SMALL_LOOP);
7929 __ ldr(tmp1, Address(__ post(a1, wordSize)));
7930 __ ldr(tmp2, Address(__ post(a2, wordSize)));
7931 __ subs(cnt1, cnt1, wordSize);
7932 __ eor(tmp1, tmp1, tmp2);
7933 __ cbnz(tmp1, NOT_EQUAL);
7934 __ br(__ GT, SMALL_LOOP);
7935 __ bind(POST_LOOP);
7936 __ ldr(tmp1, Address(a1, cnt1));
7937 __ ldr(tmp2, Address(a2, cnt1));
7938 __ eor(tmp1, tmp1, tmp2);
7939 __ cbnz(tmp1, NOT_EQUAL);
7940 __ bind(EQUAL);
7941 __ mov(result, true);
7942 __ bind(NOT_EQUAL);
7943 if (!UseSIMDForArrayEquals) {
7944 __ pop(spilled_regs, sp);
7945 }
7946 __ bind(NOT_EQUAL_NO_POP);
7947 __ leave();
7948 __ ret(lr);
7949 return entry;
7950 }
7951
7952 // result = r0 - return value. Contains initial hashcode value on entry.
7953 // ary = r1 - array address
7954 // cnt = r2 - elements count
7955 // Clobbers: v0-v13, rscratch1, rscratch2
7956 address generate_large_arrays_hashcode(BasicType eltype) {
7957 const Register result = r0, ary = r1, cnt = r2;
7958 const FloatRegister vdata0 = v3, vdata1 = v2, vdata2 = v1, vdata3 = v0;
7959 const FloatRegister vmul0 = v4, vmul1 = v5, vmul2 = v6, vmul3 = v7;
7960 const FloatRegister vpow = v12; // powers of 31: <31^3, ..., 31^0>
7961 const FloatRegister vpowm = v13;
7962
7963 ARRAYS_HASHCODE_REGISTERS;
7964
7965 Label SMALL_LOOP, LARGE_LOOP_PREHEADER, LARGE_LOOP, TAIL, TAIL_SHORTCUT, BR_BASE;
7966
7967 unsigned int vf; // vectorization factor
7968 bool multiply_by_halves;
7969 Assembler::SIMD_Arrangement load_arrangement;
7970 switch (eltype) {
7971 case T_BOOLEAN:
7972 case T_BYTE:
7973 load_arrangement = Assembler::T8B;
7974 multiply_by_halves = true;
7975 vf = 8;
7976 break;
7977 case T_CHAR:
7978 case T_SHORT:
7979 load_arrangement = Assembler::T8H;
7980 multiply_by_halves = true;
7981 vf = 8;
7982 break;
7983 case T_INT:
7984 load_arrangement = Assembler::T4S;
7985 multiply_by_halves = false;
7986 vf = 4;
7987 break;
7988 default:
7989 ShouldNotReachHere();
7990 }
7991
7992 // Unroll factor
7993 const unsigned uf = 4;
7994
7995 // Effective vectorization factor
7996 const unsigned evf = vf * uf;
7997
7998 __ align(CodeEntryAlignment);
7999
8000 StubGenStubId stub_id;
8001 switch (eltype) {
8002 case T_BOOLEAN:
8003 stub_id = StubGenStubId::large_arrays_hashcode_boolean_id;
8004 break;
8005 case T_BYTE:
8006 stub_id = StubGenStubId::large_arrays_hashcode_byte_id;
8007 break;
8008 case T_CHAR:
8009 stub_id = StubGenStubId::large_arrays_hashcode_char_id;
8010 break;
8011 case T_SHORT:
8012 stub_id = StubGenStubId::large_arrays_hashcode_short_id;
8013 break;
8014 case T_INT:
8015 stub_id = StubGenStubId::large_arrays_hashcode_int_id;
8016 break;
8017 default:
8018 stub_id = StubGenStubId::NO_STUBID;
8019 ShouldNotReachHere();
8020 };
8021
8022 StubCodeMark mark(this, stub_id);
8023
8024 address entry = __ pc();
8025 __ enter();
8026
8027 // Put 0-3'th powers of 31 into a single SIMD register together. The register will be used in
8028 // the SMALL and LARGE LOOPS' epilogues. The initialization is hoisted here and the register's
8029 // value shouldn't change throughout both loops.
8030 __ movw(rscratch1, intpow(31U, 3));
8031 __ mov(vpow, Assembler::S, 0, rscratch1);
8032 __ movw(rscratch1, intpow(31U, 2));
8033 __ mov(vpow, Assembler::S, 1, rscratch1);
8034 __ movw(rscratch1, intpow(31U, 1));
8035 __ mov(vpow, Assembler::S, 2, rscratch1);
8036 __ movw(rscratch1, intpow(31U, 0));
8037 __ mov(vpow, Assembler::S, 3, rscratch1);
8038
8039 __ mov(vmul0, Assembler::T16B, 0);
8040 __ mov(vmul0, Assembler::S, 3, result);
8041
8042 __ andr(rscratch2, cnt, (uf - 1) * vf);
8043 __ cbz(rscratch2, LARGE_LOOP_PREHEADER);
8044
8045 __ movw(rscratch1, intpow(31U, multiply_by_halves ? vf / 2 : vf));
8046 __ mov(vpowm, Assembler::S, 0, rscratch1);
8047
8048 // SMALL LOOP
8049 __ bind(SMALL_LOOP);
8050
8051 __ ld1(vdata0, load_arrangement, Address(__ post(ary, vf * type2aelembytes(eltype))));
8052 __ mulvs(vmul0, Assembler::T4S, vmul0, vpowm, 0);
8053 __ subsw(rscratch2, rscratch2, vf);
8054
8055 if (load_arrangement == Assembler::T8B) {
8056 // Extend 8B to 8H to be able to use vector multiply
8057 // instructions
8058 assert(load_arrangement == Assembler::T8B, "expected to extend 8B to 8H");
8059 if (is_signed_subword_type(eltype)) {
8060 __ sxtl(vdata0, Assembler::T8H, vdata0, load_arrangement);
8061 } else {
8062 __ uxtl(vdata0, Assembler::T8H, vdata0, load_arrangement);
8063 }
8064 }
8065
8066 switch (load_arrangement) {
8067 case Assembler::T4S:
8068 __ addv(vmul0, load_arrangement, vmul0, vdata0);
8069 break;
8070 case Assembler::T8B:
8071 case Assembler::T8H:
8072 assert(is_subword_type(eltype), "subword type expected");
8073 if (is_signed_subword_type(eltype)) {
8074 __ saddwv(vmul0, vmul0, Assembler::T4S, vdata0, Assembler::T4H);
8075 } else {
8076 __ uaddwv(vmul0, vmul0, Assembler::T4S, vdata0, Assembler::T4H);
8077 }
8078 break;
8079 default:
8080 __ should_not_reach_here();
8081 }
8082
8083 // Process the upper half of a vector
8084 if (load_arrangement == Assembler::T8B || load_arrangement == Assembler::T8H) {
8085 __ mulvs(vmul0, Assembler::T4S, vmul0, vpowm, 0);
8086 if (is_signed_subword_type(eltype)) {
8087 __ saddwv2(vmul0, vmul0, Assembler::T4S, vdata0, Assembler::T8H);
8088 } else {
8089 __ uaddwv2(vmul0, vmul0, Assembler::T4S, vdata0, Assembler::T8H);
8090 }
8091 }
8092
8093 __ br(Assembler::HI, SMALL_LOOP);
8094
8095 // SMALL LOOP'S EPILOQUE
8096 __ lsr(rscratch2, cnt, exact_log2(evf));
8097 __ cbnz(rscratch2, LARGE_LOOP_PREHEADER);
8098
8099 __ mulv(vmul0, Assembler::T4S, vmul0, vpow);
8100 __ addv(vmul0, Assembler::T4S, vmul0);
8101 __ umov(result, vmul0, Assembler::S, 0);
8102
8103 // TAIL
8104 __ bind(TAIL);
8105
8106 // The andr performs cnt % vf. The subtract shifted by 3 offsets past vf - 1 - (cnt % vf) pairs
8107 // of load + madd insns i.e. it only executes cnt % vf load + madd pairs.
8108 assert(is_power_of_2(vf), "can't use this value to calculate the jump target PC");
8109 __ andr(rscratch2, cnt, vf - 1);
8110 __ bind(TAIL_SHORTCUT);
8111 __ adr(rscratch1, BR_BASE);
8112 __ sub(rscratch1, rscratch1, rscratch2, ext::uxtw, 3);
8113 __ movw(rscratch2, 0x1f);
8114 __ br(rscratch1);
8115
8116 for (size_t i = 0; i < vf - 1; ++i) {
8117 __ load(rscratch1, Address(__ post(ary, type2aelembytes(eltype))),
8118 eltype);
8119 __ maddw(result, result, rscratch2, rscratch1);
8120 }
8121 __ bind(BR_BASE);
8122
8123 __ leave();
8124 __ ret(lr);
8125
8126 // LARGE LOOP
8127 __ bind(LARGE_LOOP_PREHEADER);
8128
8129 __ lsr(rscratch2, cnt, exact_log2(evf));
8130
8131 if (multiply_by_halves) {
8132 // 31^4 - multiplier between lower and upper parts of a register
8133 __ movw(rscratch1, intpow(31U, vf / 2));
8134 __ mov(vpowm, Assembler::S, 1, rscratch1);
8135 // 31^28 - remainder of the iteraion multiplier, 28 = 32 - 4
8136 __ movw(rscratch1, intpow(31U, evf - vf / 2));
8137 __ mov(vpowm, Assembler::S, 0, rscratch1);
8138 } else {
8139 // 31^16
8140 __ movw(rscratch1, intpow(31U, evf));
8141 __ mov(vpowm, Assembler::S, 0, rscratch1);
8142 }
8143
8144 __ mov(vmul3, Assembler::T16B, 0);
8145 __ mov(vmul2, Assembler::T16B, 0);
8146 __ mov(vmul1, Assembler::T16B, 0);
8147
8148 __ bind(LARGE_LOOP);
8149
8150 __ mulvs(vmul3, Assembler::T4S, vmul3, vpowm, 0);
8151 __ mulvs(vmul2, Assembler::T4S, vmul2, vpowm, 0);
8152 __ mulvs(vmul1, Assembler::T4S, vmul1, vpowm, 0);
8153 __ mulvs(vmul0, Assembler::T4S, vmul0, vpowm, 0);
8154
8155 __ ld1(vdata3, vdata2, vdata1, vdata0, load_arrangement,
8156 Address(__ post(ary, evf * type2aelembytes(eltype))));
8157
8158 if (load_arrangement == Assembler::T8B) {
8159 // Extend 8B to 8H to be able to use vector multiply
8160 // instructions
8161 assert(load_arrangement == Assembler::T8B, "expected to extend 8B to 8H");
8162 if (is_signed_subword_type(eltype)) {
8163 __ sxtl(vdata3, Assembler::T8H, vdata3, load_arrangement);
8164 __ sxtl(vdata2, Assembler::T8H, vdata2, load_arrangement);
8165 __ sxtl(vdata1, Assembler::T8H, vdata1, load_arrangement);
8166 __ sxtl(vdata0, Assembler::T8H, vdata0, load_arrangement);
8167 } else {
8168 __ uxtl(vdata3, Assembler::T8H, vdata3, load_arrangement);
8169 __ uxtl(vdata2, Assembler::T8H, vdata2, load_arrangement);
8170 __ uxtl(vdata1, Assembler::T8H, vdata1, load_arrangement);
8171 __ uxtl(vdata0, Assembler::T8H, vdata0, load_arrangement);
8172 }
8173 }
8174
8175 switch (load_arrangement) {
8176 case Assembler::T4S:
8177 __ addv(vmul3, load_arrangement, vmul3, vdata3);
8178 __ addv(vmul2, load_arrangement, vmul2, vdata2);
8179 __ addv(vmul1, load_arrangement, vmul1, vdata1);
8180 __ addv(vmul0, load_arrangement, vmul0, vdata0);
8181 break;
8182 case Assembler::T8B:
8183 case Assembler::T8H:
8184 assert(is_subword_type(eltype), "subword type expected");
8185 if (is_signed_subword_type(eltype)) {
8186 __ saddwv(vmul3, vmul3, Assembler::T4S, vdata3, Assembler::T4H);
8187 __ saddwv(vmul2, vmul2, Assembler::T4S, vdata2, Assembler::T4H);
8188 __ saddwv(vmul1, vmul1, Assembler::T4S, vdata1, Assembler::T4H);
8189 __ saddwv(vmul0, vmul0, Assembler::T4S, vdata0, Assembler::T4H);
8190 } else {
8191 __ uaddwv(vmul3, vmul3, Assembler::T4S, vdata3, Assembler::T4H);
8192 __ uaddwv(vmul2, vmul2, Assembler::T4S, vdata2, Assembler::T4H);
8193 __ uaddwv(vmul1, vmul1, Assembler::T4S, vdata1, Assembler::T4H);
8194 __ uaddwv(vmul0, vmul0, Assembler::T4S, vdata0, Assembler::T4H);
8195 }
8196 break;
8197 default:
8198 __ should_not_reach_here();
8199 }
8200
8201 // Process the upper half of a vector
8202 if (load_arrangement == Assembler::T8B || load_arrangement == Assembler::T8H) {
8203 __ mulvs(vmul3, Assembler::T4S, vmul3, vpowm, 1);
8204 __ mulvs(vmul2, Assembler::T4S, vmul2, vpowm, 1);
8205 __ mulvs(vmul1, Assembler::T4S, vmul1, vpowm, 1);
8206 __ mulvs(vmul0, Assembler::T4S, vmul0, vpowm, 1);
8207 if (is_signed_subword_type(eltype)) {
8208 __ saddwv2(vmul3, vmul3, Assembler::T4S, vdata3, Assembler::T8H);
8209 __ saddwv2(vmul2, vmul2, Assembler::T4S, vdata2, Assembler::T8H);
8210 __ saddwv2(vmul1, vmul1, Assembler::T4S, vdata1, Assembler::T8H);
8211 __ saddwv2(vmul0, vmul0, Assembler::T4S, vdata0, Assembler::T8H);
8212 } else {
8213 __ uaddwv2(vmul3, vmul3, Assembler::T4S, vdata3, Assembler::T8H);
8214 __ uaddwv2(vmul2, vmul2, Assembler::T4S, vdata2, Assembler::T8H);
8215 __ uaddwv2(vmul1, vmul1, Assembler::T4S, vdata1, Assembler::T8H);
8216 __ uaddwv2(vmul0, vmul0, Assembler::T4S, vdata0, Assembler::T8H);
8217 }
8218 }
8219
8220 __ subsw(rscratch2, rscratch2, 1);
8221 __ br(Assembler::HI, LARGE_LOOP);
8222
8223 __ mulv(vmul3, Assembler::T4S, vmul3, vpow);
8224 __ addv(vmul3, Assembler::T4S, vmul3);
8225 __ umov(result, vmul3, Assembler::S, 0);
8226
8227 __ mov(rscratch2, intpow(31U, vf));
8228
8229 __ mulv(vmul2, Assembler::T4S, vmul2, vpow);
8230 __ addv(vmul2, Assembler::T4S, vmul2);
8231 __ umov(rscratch1, vmul2, Assembler::S, 0);
8232 __ maddw(result, result, rscratch2, rscratch1);
8233
8234 __ mulv(vmul1, Assembler::T4S, vmul1, vpow);
8235 __ addv(vmul1, Assembler::T4S, vmul1);
8236 __ umov(rscratch1, vmul1, Assembler::S, 0);
8237 __ maddw(result, result, rscratch2, rscratch1);
8238
8239 __ mulv(vmul0, Assembler::T4S, vmul0, vpow);
8240 __ addv(vmul0, Assembler::T4S, vmul0);
8241 __ umov(rscratch1, vmul0, Assembler::S, 0);
8242 __ maddw(result, result, rscratch2, rscratch1);
8243
8244 __ andr(rscratch2, cnt, vf - 1);
8245 __ cbnz(rscratch2, TAIL_SHORTCUT);
8246
8247 __ leave();
8248 __ ret(lr);
8249
8250 return entry;
8251 }
8252
8253 address generate_dsin_dcos(bool isCos) {
8254 __ align(CodeEntryAlignment);
8255 StubGenStubId stub_id = (isCos ? StubGenStubId::dcos_id : StubGenStubId::dsin_id);
8256 StubCodeMark mark(this, stub_id);
8257 address start = __ pc();
8258 __ generate_dsin_dcos(isCos, (address)StubRoutines::aarch64::_npio2_hw,
8259 (address)StubRoutines::aarch64::_two_over_pi,
8260 (address)StubRoutines::aarch64::_pio2,
8261 (address)StubRoutines::aarch64::_dsin_coef,
8262 (address)StubRoutines::aarch64::_dcos_coef);
8263 return start;
8264 }
8265
8266 // code for comparing 16 characters of strings with Latin1 and Utf16 encoding
8267 void compare_string_16_x_LU(Register tmpL, Register tmpU, Label &DIFF1,
8268 Label &DIFF2) {
8269 Register cnt1 = r2, tmp2 = r11, tmp3 = r12;
8270 FloatRegister vtmp = v1, vtmpZ = v0, vtmp3 = v2;
8271
8272 __ ldrq(vtmp, Address(__ post(tmp2, 16)));
8273 __ ldr(tmpU, Address(__ post(cnt1, 8)));
8274 __ zip1(vtmp3, __ T16B, vtmp, vtmpZ);
8275 // now we have 32 bytes of characters (converted to U) in vtmp:vtmp3
8276
8277 __ fmovd(tmpL, vtmp3);
8278 __ eor(rscratch2, tmp3, tmpL);
8279 __ cbnz(rscratch2, DIFF2);
8280
8281 __ ldr(tmp3, Address(__ post(cnt1, 8)));
8282 __ umov(tmpL, vtmp3, __ D, 1);
8283 __ eor(rscratch2, tmpU, tmpL);
8284 __ cbnz(rscratch2, DIFF1);
8285
8286 __ zip2(vtmp, __ T16B, vtmp, vtmpZ);
8287 __ ldr(tmpU, Address(__ post(cnt1, 8)));
8288 __ fmovd(tmpL, vtmp);
8289 __ eor(rscratch2, tmp3, tmpL);
8290 __ cbnz(rscratch2, DIFF2);
8291
8292 __ ldr(tmp3, Address(__ post(cnt1, 8)));
8293 __ umov(tmpL, vtmp, __ D, 1);
8294 __ eor(rscratch2, tmpU, tmpL);
8295 __ cbnz(rscratch2, DIFF1);
8296 }
8297
8298 // r0 = result
8299 // r1 = str1
8300 // r2 = cnt1
8301 // r3 = str2
8302 // r4 = cnt2
8303 // r10 = tmp1
8304 // r11 = tmp2
8305 address generate_compare_long_string_different_encoding(bool isLU) {
8306 __ align(CodeEntryAlignment);
8307 StubGenStubId stub_id = (isLU ? StubGenStubId::compare_long_string_LU_id : StubGenStubId::compare_long_string_UL_id);
8308 StubCodeMark mark(this, stub_id);
8309 address entry = __ pc();
8310 Label SMALL_LOOP, TAIL, TAIL_LOAD_16, LOAD_LAST, DIFF1, DIFF2,
8311 DONE, CALCULATE_DIFFERENCE, LARGE_LOOP_PREFETCH, NO_PREFETCH,
8312 LARGE_LOOP_PREFETCH_REPEAT1, LARGE_LOOP_PREFETCH_REPEAT2;
8313 Register result = r0, str1 = r1, cnt1 = r2, str2 = r3, cnt2 = r4,
8314 tmp1 = r10, tmp2 = r11, tmp3 = r12, tmp4 = r14;
8315 FloatRegister vtmpZ = v0, vtmp = v1, vtmp3 = v2;
8316 RegSet spilled_regs = RegSet::of(tmp3, tmp4);
8317
8318 int prefetchLoopExitCondition = MAX2(64, SoftwarePrefetchHintDistance/2);
8319
8320 __ eor(vtmpZ, __ T16B, vtmpZ, vtmpZ);
8321 // cnt2 == amount of characters left to compare
8322 // Check already loaded first 4 symbols(vtmp and tmp2(LU)/tmp1(UL))
8323 __ zip1(vtmp, __ T8B, vtmp, vtmpZ);
8324 __ add(str1, str1, isLU ? wordSize/2 : wordSize);
8325 __ add(str2, str2, isLU ? wordSize : wordSize/2);
8326 __ fmovd(isLU ? tmp1 : tmp2, vtmp);
8327 __ subw(cnt2, cnt2, 8); // Already loaded 4 symbols. Last 4 is special case.
8328 __ eor(rscratch2, tmp1, tmp2);
8329 __ mov(rscratch1, tmp2);
8330 __ cbnz(rscratch2, CALCULATE_DIFFERENCE);
8331 Register tmpU = isLU ? rscratch1 : tmp1, // where to keep U for comparison
8332 tmpL = isLU ? tmp1 : rscratch1; // where to keep L for comparison
8333 __ push(spilled_regs, sp);
8334 __ mov(tmp2, isLU ? str1 : str2); // init the pointer to L next load
8335 __ mov(cnt1, isLU ? str2 : str1); // init the pointer to U next load
8336
8337 __ ldr(tmp3, Address(__ post(cnt1, 8)));
8338
8339 if (SoftwarePrefetchHintDistance >= 0) {
8340 __ subs(rscratch2, cnt2, prefetchLoopExitCondition);
8341 __ br(__ LT, NO_PREFETCH);
8342 __ bind(LARGE_LOOP_PREFETCH);
8343 __ prfm(Address(tmp2, SoftwarePrefetchHintDistance));
8344 __ mov(tmp4, 2);
8345 __ prfm(Address(cnt1, SoftwarePrefetchHintDistance));
8346 __ bind(LARGE_LOOP_PREFETCH_REPEAT1);
8347 compare_string_16_x_LU(tmpL, tmpU, DIFF1, DIFF2);
8348 __ subs(tmp4, tmp4, 1);
8349 __ br(__ GT, LARGE_LOOP_PREFETCH_REPEAT1);
8350 __ prfm(Address(cnt1, SoftwarePrefetchHintDistance));
8351 __ mov(tmp4, 2);
8352 __ bind(LARGE_LOOP_PREFETCH_REPEAT2);
8353 compare_string_16_x_LU(tmpL, tmpU, DIFF1, DIFF2);
8354 __ subs(tmp4, tmp4, 1);
8355 __ br(__ GT, LARGE_LOOP_PREFETCH_REPEAT2);
8356 __ sub(cnt2, cnt2, 64);
8357 __ subs(rscratch2, cnt2, prefetchLoopExitCondition);
8358 __ br(__ GE, LARGE_LOOP_PREFETCH);
8359 }
8360 __ cbz(cnt2, LOAD_LAST); // no characters left except last load
8361 __ bind(NO_PREFETCH);
8362 __ subs(cnt2, cnt2, 16);
8363 __ br(__ LT, TAIL);
8364 __ align(OptoLoopAlignment);
8365 __ bind(SMALL_LOOP); // smaller loop
8366 __ subs(cnt2, cnt2, 16);
8367 compare_string_16_x_LU(tmpL, tmpU, DIFF1, DIFF2);
8368 __ br(__ GE, SMALL_LOOP);
8369 __ cmn(cnt2, (u1)16);
8370 __ br(__ EQ, LOAD_LAST);
8371 __ bind(TAIL); // 1..15 characters left until last load (last 4 characters)
8372 __ add(cnt1, cnt1, cnt2, __ LSL, 1); // Address of 32 bytes before last 4 characters in UTF-16 string
8373 __ add(tmp2, tmp2, cnt2); // Address of 16 bytes before last 4 characters in Latin1 string
8374 __ ldr(tmp3, Address(cnt1, -8));
8375 compare_string_16_x_LU(tmpL, tmpU, DIFF1, DIFF2); // last 16 characters before last load
8376 __ b(LOAD_LAST);
8377 __ bind(DIFF2);
8378 __ mov(tmpU, tmp3);
8379 __ bind(DIFF1);
8380 __ pop(spilled_regs, sp);
8381 __ b(CALCULATE_DIFFERENCE);
8382 __ bind(LOAD_LAST);
8383 // Last 4 UTF-16 characters are already pre-loaded into tmp3 by compare_string_16_x_LU.
8384 // No need to load it again
8385 __ mov(tmpU, tmp3);
8386 __ pop(spilled_regs, sp);
8387
8388 // tmp2 points to the address of the last 4 Latin1 characters right now
8389 __ ldrs(vtmp, Address(tmp2));
8390 __ zip1(vtmp, __ T8B, vtmp, vtmpZ);
8391 __ fmovd(tmpL, vtmp);
8392
8393 __ eor(rscratch2, tmpU, tmpL);
8394 __ cbz(rscratch2, DONE);
8395
8396 // Find the first different characters in the longwords and
8397 // compute their difference.
8398 __ bind(CALCULATE_DIFFERENCE);
8399 __ rev(rscratch2, rscratch2);
8400 __ clz(rscratch2, rscratch2);
8401 __ andr(rscratch2, rscratch2, -16);
8402 __ lsrv(tmp1, tmp1, rscratch2);
8403 __ uxthw(tmp1, tmp1);
8404 __ lsrv(rscratch1, rscratch1, rscratch2);
8405 __ uxthw(rscratch1, rscratch1);
8406 __ subw(result, tmp1, rscratch1);
8407 __ bind(DONE);
8408 __ ret(lr);
8409 return entry;
8410 }
8411
8412 // r0 = input (float16)
8413 // v0 = result (float)
8414 // v1 = temporary float register
8415 address generate_float16ToFloat() {
8416 __ align(CodeEntryAlignment);
8417 StubGenStubId stub_id = StubGenStubId::hf2f_id;
8418 StubCodeMark mark(this, stub_id);
8419 address entry = __ pc();
8420 BLOCK_COMMENT("Entry:");
8421 __ flt16_to_flt(v0, r0, v1);
8422 __ ret(lr);
8423 return entry;
8424 }
8425
8426 // v0 = input (float)
8427 // r0 = result (float16)
8428 // v1 = temporary float register
8429 address generate_floatToFloat16() {
8430 __ align(CodeEntryAlignment);
8431 StubGenStubId stub_id = StubGenStubId::f2hf_id;
8432 StubCodeMark mark(this, stub_id);
8433 address entry = __ pc();
8434 BLOCK_COMMENT("Entry:");
8435 __ flt_to_flt16(r0, v0, v1);
8436 __ ret(lr);
8437 return entry;
8438 }
8439
8440 address generate_method_entry_barrier() {
8441 __ align(CodeEntryAlignment);
8442 StubGenStubId stub_id = StubGenStubId::method_entry_barrier_id;
8443 StubCodeMark mark(this, stub_id);
8444
8445 Label deoptimize_label;
8446
8447 address start = __ pc();
8448
8449 BarrierSetAssembler* bs_asm = BarrierSet::barrier_set()->barrier_set_assembler();
8450
8451 if (bs_asm->nmethod_patching_type() == NMethodPatchingType::conc_instruction_and_data_patch) {
8452 BarrierSetNMethod* bs_nm = BarrierSet::barrier_set()->barrier_set_nmethod();
8453 // We can get here despite the nmethod being good, if we have not
8454 // yet applied our cross modification fence (or data fence).
8455 Address thread_epoch_addr(rthread, in_bytes(bs_nm->thread_disarmed_guard_value_offset()) + 4);
8456 __ lea(rscratch2, ExternalAddress(bs_asm->patching_epoch_addr()));
8457 __ ldrw(rscratch2, rscratch2);
8458 __ strw(rscratch2, thread_epoch_addr);
8459 __ isb();
8460 __ membar(__ LoadLoad);
8461 }
8462
8463 __ set_last_Java_frame(sp, rfp, lr, rscratch1);
8464
8465 __ enter();
8466 __ add(rscratch2, sp, wordSize); // rscratch2 points to the saved lr
8467
8468 __ sub(sp, sp, 4 * wordSize); // four words for the returned {sp, fp, lr, pc}
8469
8470 __ push_call_clobbered_registers();
8471
8472 __ mov(c_rarg0, rscratch2);
8473 __ call_VM_leaf
8474 (CAST_FROM_FN_PTR
8475 (address, BarrierSetNMethod::nmethod_stub_entry_barrier), 1);
8476
8477 __ reset_last_Java_frame(true);
8478
8479 __ mov(rscratch1, r0);
8480
8481 __ pop_call_clobbered_registers();
8482
8483 __ cbnz(rscratch1, deoptimize_label);
8484
8485 __ leave();
8486 __ ret(lr);
8487
8488 __ BIND(deoptimize_label);
8489
8490 __ ldp(/* new sp */ rscratch1, rfp, Address(sp, 0 * wordSize));
8491 __ ldp(lr, /* new pc*/ rscratch2, Address(sp, 2 * wordSize));
8492
8493 __ mov(sp, rscratch1);
8494 __ br(rscratch2);
8495
8496 return start;
8497 }
8498
8499 // r0 = result
8500 // r1 = str1
8501 // r2 = cnt1
8502 // r3 = str2
8503 // r4 = cnt2
8504 // r10 = tmp1
8505 // r11 = tmp2
8506 address generate_compare_long_string_same_encoding(bool isLL) {
8507 __ align(CodeEntryAlignment);
8508 StubGenStubId stub_id = (isLL ? StubGenStubId::compare_long_string_LL_id : StubGenStubId::compare_long_string_UU_id);
8509 StubCodeMark mark(this, stub_id);
8510 address entry = __ pc();
8511 Register result = r0, str1 = r1, cnt1 = r2, str2 = r3, cnt2 = r4,
8512 tmp1 = r10, tmp2 = r11, tmp1h = rscratch1, tmp2h = rscratch2;
8513
8514 Label LARGE_LOOP_PREFETCH, LOOP_COMPARE16, DIFF, LESS16, LESS8, CAL_DIFFERENCE, LENGTH_DIFF;
8515
8516 // exit from large loop when less than 64 bytes left to read or we're about
8517 // to prefetch memory behind array border
8518 int largeLoopExitCondition = MAX2(64, SoftwarePrefetchHintDistance)/(isLL ? 1 : 2);
8519
8520 // before jumping to stub, pre-load 8 bytes already, so do comparison directly
8521 __ eor(rscratch2, tmp1, tmp2);
8522 __ cbnz(rscratch2, CAL_DIFFERENCE);
8523
8524 __ sub(cnt2, cnt2, wordSize/(isLL ? 1 : 2));
8525 // update pointers, because of previous read
8526 __ add(str1, str1, wordSize);
8527 __ add(str2, str2, wordSize);
8528 if (SoftwarePrefetchHintDistance >= 0) {
8529 __ align(OptoLoopAlignment);
8530 __ bind(LARGE_LOOP_PREFETCH);
8531 __ prfm(Address(str1, SoftwarePrefetchHintDistance));
8532 __ prfm(Address(str2, SoftwarePrefetchHintDistance));
8533
8534 for (int i = 0; i < 4; i++) {
8535 __ ldp(tmp1, tmp1h, Address(str1, i * 16));
8536 __ ldp(tmp2, tmp2h, Address(str2, i * 16));
8537 __ cmp(tmp1, tmp2);
8538 __ ccmp(tmp1h, tmp2h, 0, Assembler::EQ);
8539 __ br(Assembler::NE, DIFF);
8540 }
8541 __ sub(cnt2, cnt2, isLL ? 64 : 32);
8542 __ add(str1, str1, 64);
8543 __ add(str2, str2, 64);
8544 __ subs(rscratch2, cnt2, largeLoopExitCondition);
8545 __ br(Assembler::GE, LARGE_LOOP_PREFETCH);
8546 __ cbz(cnt2, LENGTH_DIFF); // no more chars left?
8547 }
8548
8549 __ subs(rscratch1, cnt2, isLL ? 16 : 8);
8550 __ br(Assembler::LE, LESS16);
8551 __ align(OptoLoopAlignment);
8552 __ bind(LOOP_COMPARE16);
8553 __ ldp(tmp1, tmp1h, Address(__ post(str1, 16)));
8554 __ ldp(tmp2, tmp2h, Address(__ post(str2, 16)));
8555 __ cmp(tmp1, tmp2);
8556 __ ccmp(tmp1h, tmp2h, 0, Assembler::EQ);
8557 __ br(Assembler::NE, DIFF);
8558 __ sub(cnt2, cnt2, isLL ? 16 : 8);
8559 __ subs(rscratch2, cnt2, isLL ? 16 : 8);
8560 __ br(Assembler::LT, LESS16);
8561
8562 __ ldp(tmp1, tmp1h, Address(__ post(str1, 16)));
8563 __ ldp(tmp2, tmp2h, Address(__ post(str2, 16)));
8564 __ cmp(tmp1, tmp2);
8565 __ ccmp(tmp1h, tmp2h, 0, Assembler::EQ);
8566 __ br(Assembler::NE, DIFF);
8567 __ sub(cnt2, cnt2, isLL ? 16 : 8);
8568 __ subs(rscratch2, cnt2, isLL ? 16 : 8);
8569 __ br(Assembler::GE, LOOP_COMPARE16);
8570 __ cbz(cnt2, LENGTH_DIFF);
8571
8572 __ bind(LESS16);
8573 // each 8 compare
8574 __ subs(cnt2, cnt2, isLL ? 8 : 4);
8575 __ br(Assembler::LE, LESS8);
8576 __ ldr(tmp1, Address(__ post(str1, 8)));
8577 __ ldr(tmp2, Address(__ post(str2, 8)));
8578 __ eor(rscratch2, tmp1, tmp2);
8579 __ cbnz(rscratch2, CAL_DIFFERENCE);
8580 __ sub(cnt2, cnt2, isLL ? 8 : 4);
8581
8582 __ bind(LESS8); // directly load last 8 bytes
8583 if (!isLL) {
8584 __ add(cnt2, cnt2, cnt2);
8585 }
8586 __ ldr(tmp1, Address(str1, cnt2));
8587 __ ldr(tmp2, Address(str2, cnt2));
8588 __ eor(rscratch2, tmp1, tmp2);
8589 __ cbz(rscratch2, LENGTH_DIFF);
8590 __ b(CAL_DIFFERENCE);
8591
8592 __ bind(DIFF);
8593 __ cmp(tmp1, tmp2);
8594 __ csel(tmp1, tmp1, tmp1h, Assembler::NE);
8595 __ csel(tmp2, tmp2, tmp2h, Assembler::NE);
8596 // reuse rscratch2 register for the result of eor instruction
8597 __ eor(rscratch2, tmp1, tmp2);
8598
8599 __ bind(CAL_DIFFERENCE);
8600 __ rev(rscratch2, rscratch2);
8601 __ clz(rscratch2, rscratch2);
8602 __ andr(rscratch2, rscratch2, isLL ? -8 : -16);
8603 __ lsrv(tmp1, tmp1, rscratch2);
8604 __ lsrv(tmp2, tmp2, rscratch2);
8605 if (isLL) {
8606 __ uxtbw(tmp1, tmp1);
8607 __ uxtbw(tmp2, tmp2);
8608 } else {
8609 __ uxthw(tmp1, tmp1);
8610 __ uxthw(tmp2, tmp2);
8611 }
8612 __ subw(result, tmp1, tmp2);
8613
8614 __ bind(LENGTH_DIFF);
8615 __ ret(lr);
8616 return entry;
8617 }
8618
8619 enum string_compare_mode {
8620 LL,
8621 LU,
8622 UL,
8623 UU,
8624 };
8625
8626 // The following registers are declared in aarch64.ad
8627 // r0 = result
8628 // r1 = str1
8629 // r2 = cnt1
8630 // r3 = str2
8631 // r4 = cnt2
8632 // r10 = tmp1
8633 // r11 = tmp2
8634 // z0 = ztmp1
8635 // z1 = ztmp2
8636 // p0 = pgtmp1
8637 // p1 = pgtmp2
8638 address generate_compare_long_string_sve(string_compare_mode mode) {
8639 StubGenStubId stub_id;
8640 switch (mode) {
8641 case LL: stub_id = StubGenStubId::compare_long_string_LL_id; break;
8642 case LU: stub_id = StubGenStubId::compare_long_string_LU_id; break;
8643 case UL: stub_id = StubGenStubId::compare_long_string_UL_id; break;
8644 case UU: stub_id = StubGenStubId::compare_long_string_UU_id; break;
8645 default: ShouldNotReachHere();
8646 }
8647
8648 __ align(CodeEntryAlignment);
8649 address entry = __ pc();
8650 Register result = r0, str1 = r1, cnt1 = r2, str2 = r3, cnt2 = r4,
8651 tmp1 = r10, tmp2 = r11;
8652
8653 Label LOOP, DONE, MISMATCH;
8654 Register vec_len = tmp1;
8655 Register idx = tmp2;
8656 // The minimum of the string lengths has been stored in cnt2.
8657 Register cnt = cnt2;
8658 FloatRegister ztmp1 = z0, ztmp2 = z1;
8659 PRegister pgtmp1 = p0, pgtmp2 = p1;
8660
8661 #define LOAD_PAIR(ztmp1, ztmp2, pgtmp1, src1, src2, idx) \
8662 switch (mode) { \
8663 case LL: \
8664 __ sve_ld1b(ztmp1, __ B, pgtmp1, Address(str1, idx)); \
8665 __ sve_ld1b(ztmp2, __ B, pgtmp1, Address(str2, idx)); \
8666 break; \
8667 case LU: \
8668 __ sve_ld1b(ztmp1, __ H, pgtmp1, Address(str1, idx)); \
8669 __ sve_ld1h(ztmp2, __ H, pgtmp1, Address(str2, idx, Address::lsl(1))); \
8670 break; \
8671 case UL: \
8672 __ sve_ld1h(ztmp1, __ H, pgtmp1, Address(str1, idx, Address::lsl(1))); \
8673 __ sve_ld1b(ztmp2, __ H, pgtmp1, Address(str2, idx)); \
8674 break; \
8675 case UU: \
8676 __ sve_ld1h(ztmp1, __ H, pgtmp1, Address(str1, idx, Address::lsl(1))); \
8677 __ sve_ld1h(ztmp2, __ H, pgtmp1, Address(str2, idx, Address::lsl(1))); \
8678 break; \
8679 default: \
8680 ShouldNotReachHere(); \
8681 }
8682
8683 StubCodeMark mark(this, stub_id);
8684
8685 __ mov(idx, 0);
8686 __ sve_whilelt(pgtmp1, mode == LL ? __ B : __ H, idx, cnt);
8687
8688 if (mode == LL) {
8689 __ sve_cntb(vec_len);
8690 } else {
8691 __ sve_cnth(vec_len);
8692 }
8693
8694 __ sub(rscratch1, cnt, vec_len);
8695
8696 __ bind(LOOP);
8697
8698 // main loop
8699 LOAD_PAIR(ztmp1, ztmp2, pgtmp1, src1, src2, idx);
8700 __ add(idx, idx, vec_len);
8701 // Compare strings.
8702 __ sve_cmp(Assembler::NE, pgtmp2, mode == LL ? __ B : __ H, pgtmp1, ztmp1, ztmp2);
8703 __ br(__ NE, MISMATCH);
8704 __ cmp(idx, rscratch1);
8705 __ br(__ LT, LOOP);
8706
8707 // post loop, last iteration
8708 __ sve_whilelt(pgtmp1, mode == LL ? __ B : __ H, idx, cnt);
8709
8710 LOAD_PAIR(ztmp1, ztmp2, pgtmp1, src1, src2, idx);
8711 __ sve_cmp(Assembler::NE, pgtmp2, mode == LL ? __ B : __ H, pgtmp1, ztmp1, ztmp2);
8712 __ br(__ EQ, DONE);
8713
8714 __ bind(MISMATCH);
8715
8716 // Crop the vector to find its location.
8717 __ sve_brkb(pgtmp2, pgtmp1, pgtmp2, false /* isMerge */);
8718 // Extract the first different characters of each string.
8719 __ sve_lasta(rscratch1, mode == LL ? __ B : __ H, pgtmp2, ztmp1);
8720 __ sve_lasta(rscratch2, mode == LL ? __ B : __ H, pgtmp2, ztmp2);
8721
8722 // Compute the difference of the first different characters.
8723 __ sub(result, rscratch1, rscratch2);
8724
8725 __ bind(DONE);
8726 __ ret(lr);
8727 #undef LOAD_PAIR
8728 return entry;
8729 }
8730
8731 void generate_compare_long_strings() {
8732 if (UseSVE == 0) {
8733 StubRoutines::aarch64::_compare_long_string_LL
8734 = generate_compare_long_string_same_encoding(true);
8735 StubRoutines::aarch64::_compare_long_string_UU
8736 = generate_compare_long_string_same_encoding(false);
8737 StubRoutines::aarch64::_compare_long_string_LU
8738 = generate_compare_long_string_different_encoding(true);
8739 StubRoutines::aarch64::_compare_long_string_UL
8740 = generate_compare_long_string_different_encoding(false);
8741 } else {
8742 StubRoutines::aarch64::_compare_long_string_LL
8743 = generate_compare_long_string_sve(LL);
8744 StubRoutines::aarch64::_compare_long_string_UU
8745 = generate_compare_long_string_sve(UU);
8746 StubRoutines::aarch64::_compare_long_string_LU
8747 = generate_compare_long_string_sve(LU);
8748 StubRoutines::aarch64::_compare_long_string_UL
8749 = generate_compare_long_string_sve(UL);
8750 }
8751 }
8752
8753 // R0 = result
8754 // R1 = str2
8755 // R2 = cnt1
8756 // R3 = str1
8757 // R4 = cnt2
8758 // Clobbers: rscratch1, rscratch2, v0, v1, rflags
8759 //
8760 // This generic linear code use few additional ideas, which makes it faster:
8761 // 1) we can safely keep at least 1st register of pattern(since length >= 8)
8762 // in order to skip initial loading(help in systems with 1 ld pipeline)
8763 // 2) we can use "fast" algorithm of finding single character to search for
8764 // first symbol with less branches(1 branch per each loaded register instead
8765 // of branch for each symbol), so, this is where constants like
8766 // 0x0101...01, 0x00010001...0001, 0x7f7f...7f, 0x7fff7fff...7fff comes from
8767 // 3) after loading and analyzing 1st register of source string, it can be
8768 // used to search for every 1st character entry, saving few loads in
8769 // comparison with "simplier-but-slower" implementation
8770 // 4) in order to avoid lots of push/pop operations, code below is heavily
8771 // re-using/re-initializing/compressing register values, which makes code
8772 // larger and a bit less readable, however, most of extra operations are
8773 // issued during loads or branches, so, penalty is minimal
8774 address generate_string_indexof_linear(bool str1_isL, bool str2_isL) {
8775 StubGenStubId stub_id;
8776 if (str1_isL) {
8777 if (str2_isL) {
8778 stub_id = StubGenStubId::string_indexof_linear_ll_id;
8779 } else {
8780 stub_id = StubGenStubId::string_indexof_linear_ul_id;
8781 }
8782 } else {
8783 if (str2_isL) {
8784 ShouldNotReachHere();
8785 } else {
8786 stub_id = StubGenStubId::string_indexof_linear_uu_id;
8787 }
8788 }
8789 __ align(CodeEntryAlignment);
8790 StubCodeMark mark(this, stub_id);
8791 address entry = __ pc();
8792
8793 int str1_chr_size = str1_isL ? 1 : 2;
8794 int str2_chr_size = str2_isL ? 1 : 2;
8795 int str1_chr_shift = str1_isL ? 0 : 1;
8796 int str2_chr_shift = str2_isL ? 0 : 1;
8797 bool isL = str1_isL && str2_isL;
8798 // parameters
8799 Register result = r0, str2 = r1, cnt1 = r2, str1 = r3, cnt2 = r4;
8800 // temporary registers
8801 Register tmp1 = r20, tmp2 = r21, tmp3 = r22, tmp4 = r23;
8802 RegSet spilled_regs = RegSet::range(tmp1, tmp4);
8803 // redefinitions
8804 Register ch1 = rscratch1, ch2 = rscratch2, first = tmp3;
8805
8806 __ push(spilled_regs, sp);
8807 Label L_LOOP, L_LOOP_PROCEED, L_SMALL, L_HAS_ZERO,
8808 L_HAS_ZERO_LOOP, L_CMP_LOOP, L_CMP_LOOP_NOMATCH, L_SMALL_PROCEED,
8809 L_SMALL_HAS_ZERO_LOOP, L_SMALL_CMP_LOOP_NOMATCH, L_SMALL_CMP_LOOP,
8810 L_POST_LOOP, L_CMP_LOOP_LAST_CMP, L_HAS_ZERO_LOOP_NOMATCH,
8811 L_SMALL_CMP_LOOP_LAST_CMP, L_SMALL_CMP_LOOP_LAST_CMP2,
8812 L_CMP_LOOP_LAST_CMP2, DONE, NOMATCH;
8813 // Read whole register from str1. It is safe, because length >=8 here
8814 __ ldr(ch1, Address(str1));
8815 // Read whole register from str2. It is safe, because length >=8 here
8816 __ ldr(ch2, Address(str2));
8817 __ sub(cnt2, cnt2, cnt1);
8818 __ andr(first, ch1, str1_isL ? 0xFF : 0xFFFF);
8819 if (str1_isL != str2_isL) {
8820 __ eor(v0, __ T16B, v0, v0);
8821 }
8822 __ mov(tmp1, str2_isL ? 0x0101010101010101 : 0x0001000100010001);
8823 __ mul(first, first, tmp1);
8824 // check if we have less than 1 register to check
8825 __ subs(cnt2, cnt2, wordSize/str2_chr_size - 1);
8826 if (str1_isL != str2_isL) {
8827 __ fmovd(v1, ch1);
8828 }
8829 __ br(__ LE, L_SMALL);
8830 __ eor(ch2, first, ch2);
8831 if (str1_isL != str2_isL) {
8832 __ zip1(v1, __ T16B, v1, v0);
8833 }
8834 __ sub(tmp2, ch2, tmp1);
8835 __ orr(ch2, ch2, str2_isL ? 0x7f7f7f7f7f7f7f7f : 0x7fff7fff7fff7fff);
8836 __ bics(tmp2, tmp2, ch2);
8837 if (str1_isL != str2_isL) {
8838 __ fmovd(ch1, v1);
8839 }
8840 __ br(__ NE, L_HAS_ZERO);
8841 __ subs(cnt2, cnt2, wordSize/str2_chr_size);
8842 __ add(result, result, wordSize/str2_chr_size);
8843 __ add(str2, str2, wordSize);
8844 __ br(__ LT, L_POST_LOOP);
8845 __ BIND(L_LOOP);
8846 __ ldr(ch2, Address(str2));
8847 __ eor(ch2, first, ch2);
8848 __ sub(tmp2, ch2, tmp1);
8849 __ orr(ch2, ch2, str2_isL ? 0x7f7f7f7f7f7f7f7f : 0x7fff7fff7fff7fff);
8850 __ bics(tmp2, tmp2, ch2);
8851 __ br(__ NE, L_HAS_ZERO);
8852 __ BIND(L_LOOP_PROCEED);
8853 __ subs(cnt2, cnt2, wordSize/str2_chr_size);
8854 __ add(str2, str2, wordSize);
8855 __ add(result, result, wordSize/str2_chr_size);
8856 __ br(__ GE, L_LOOP);
8857 __ BIND(L_POST_LOOP);
8858 __ subs(zr, cnt2, -wordSize/str2_chr_size); // no extra characters to check
8859 __ br(__ LE, NOMATCH);
8860 __ ldr(ch2, Address(str2));
8861 __ sub(cnt2, zr, cnt2, __ LSL, LogBitsPerByte + str2_chr_shift);
8862 __ eor(ch2, first, ch2);
8863 __ sub(tmp2, ch2, tmp1);
8864 __ orr(ch2, ch2, str2_isL ? 0x7f7f7f7f7f7f7f7f : 0x7fff7fff7fff7fff);
8865 __ mov(tmp4, -1); // all bits set
8866 __ b(L_SMALL_PROCEED);
8867 __ align(OptoLoopAlignment);
8868 __ BIND(L_SMALL);
8869 __ sub(cnt2, zr, cnt2, __ LSL, LogBitsPerByte + str2_chr_shift);
8870 __ eor(ch2, first, ch2);
8871 if (str1_isL != str2_isL) {
8872 __ zip1(v1, __ T16B, v1, v0);
8873 }
8874 __ sub(tmp2, ch2, tmp1);
8875 __ mov(tmp4, -1); // all bits set
8876 __ orr(ch2, ch2, str2_isL ? 0x7f7f7f7f7f7f7f7f : 0x7fff7fff7fff7fff);
8877 if (str1_isL != str2_isL) {
8878 __ fmovd(ch1, v1); // move converted 4 symbols
8879 }
8880 __ BIND(L_SMALL_PROCEED);
8881 __ lsrv(tmp4, tmp4, cnt2); // mask. zeroes on useless bits.
8882 __ bic(tmp2, tmp2, ch2);
8883 __ ands(tmp2, tmp2, tmp4); // clear useless bits and check
8884 __ rbit(tmp2, tmp2);
8885 __ br(__ EQ, NOMATCH);
8886 __ BIND(L_SMALL_HAS_ZERO_LOOP);
8887 __ clz(tmp4, tmp2); // potentially long. Up to 4 cycles on some cpu's
8888 __ cmp(cnt1, u1(wordSize/str2_chr_size));
8889 __ br(__ LE, L_SMALL_CMP_LOOP_LAST_CMP2);
8890 if (str2_isL) { // LL
8891 __ add(str2, str2, tmp4, __ LSR, LogBitsPerByte); // address of "index"
8892 __ ldr(ch2, Address(str2)); // read whole register of str2. Safe.
8893 __ lslv(tmp2, tmp2, tmp4); // shift off leading zeroes from match info
8894 __ add(result, result, tmp4, __ LSR, LogBitsPerByte);
8895 __ lsl(tmp2, tmp2, 1); // shift off leading "1" from match info
8896 } else {
8897 __ mov(ch2, 0xE); // all bits in byte set except last one
8898 __ andr(ch2, ch2, tmp4, __ LSR, LogBitsPerByte); // byte shift amount
8899 __ ldr(ch2, Address(str2, ch2)); // read whole register of str2. Safe.
8900 __ lslv(tmp2, tmp2, tmp4);
8901 __ add(result, result, tmp4, __ LSR, LogBitsPerByte + str2_chr_shift);
8902 __ add(str2, str2, tmp4, __ LSR, LogBitsPerByte + str2_chr_shift);
8903 __ lsl(tmp2, tmp2, 1); // shift off leading "1" from match info
8904 __ add(str2, str2, tmp4, __ LSR, LogBitsPerByte + str2_chr_shift);
8905 }
8906 __ cmp(ch1, ch2);
8907 __ mov(tmp4, wordSize/str2_chr_size);
8908 __ br(__ NE, L_SMALL_CMP_LOOP_NOMATCH);
8909 __ BIND(L_SMALL_CMP_LOOP);
8910 str1_isL ? __ ldrb(first, Address(str1, tmp4, Address::lsl(str1_chr_shift)))
8911 : __ ldrh(first, Address(str1, tmp4, Address::lsl(str1_chr_shift)));
8912 str2_isL ? __ ldrb(ch2, Address(str2, tmp4, Address::lsl(str2_chr_shift)))
8913 : __ ldrh(ch2, Address(str2, tmp4, Address::lsl(str2_chr_shift)));
8914 __ add(tmp4, tmp4, 1);
8915 __ cmp(tmp4, cnt1);
8916 __ br(__ GE, L_SMALL_CMP_LOOP_LAST_CMP);
8917 __ cmp(first, ch2);
8918 __ br(__ EQ, L_SMALL_CMP_LOOP);
8919 __ BIND(L_SMALL_CMP_LOOP_NOMATCH);
8920 __ cbz(tmp2, NOMATCH); // no more matches. exit
8921 __ clz(tmp4, tmp2);
8922 __ add(result, result, 1); // advance index
8923 __ add(str2, str2, str2_chr_size); // advance pointer
8924 __ b(L_SMALL_HAS_ZERO_LOOP);
8925 __ align(OptoLoopAlignment);
8926 __ BIND(L_SMALL_CMP_LOOP_LAST_CMP);
8927 __ cmp(first, ch2);
8928 __ br(__ NE, L_SMALL_CMP_LOOP_NOMATCH);
8929 __ b(DONE);
8930 __ align(OptoLoopAlignment);
8931 __ BIND(L_SMALL_CMP_LOOP_LAST_CMP2);
8932 if (str2_isL) { // LL
8933 __ add(str2, str2, tmp4, __ LSR, LogBitsPerByte); // address of "index"
8934 __ ldr(ch2, Address(str2)); // read whole register of str2. Safe.
8935 __ lslv(tmp2, tmp2, tmp4); // shift off leading zeroes from match info
8936 __ add(result, result, tmp4, __ LSR, LogBitsPerByte);
8937 __ lsl(tmp2, tmp2, 1); // shift off leading "1" from match info
8938 } else {
8939 __ mov(ch2, 0xE); // all bits in byte set except last one
8940 __ andr(ch2, ch2, tmp4, __ LSR, LogBitsPerByte); // byte shift amount
8941 __ ldr(ch2, Address(str2, ch2)); // read whole register of str2. Safe.
8942 __ lslv(tmp2, tmp2, tmp4);
8943 __ add(result, result, tmp4, __ LSR, LogBitsPerByte + str2_chr_shift);
8944 __ add(str2, str2, tmp4, __ LSR, LogBitsPerByte + str2_chr_shift);
8945 __ lsl(tmp2, tmp2, 1); // shift off leading "1" from match info
8946 __ add(str2, str2, tmp4, __ LSR, LogBitsPerByte + str2_chr_shift);
8947 }
8948 __ cmp(ch1, ch2);
8949 __ br(__ NE, L_SMALL_CMP_LOOP_NOMATCH);
8950 __ b(DONE);
8951 __ align(OptoLoopAlignment);
8952 __ BIND(L_HAS_ZERO);
8953 __ rbit(tmp2, tmp2);
8954 __ clz(tmp4, tmp2); // potentially long. Up to 4 cycles on some CPU's
8955 // Now, perform compression of counters(cnt2 and cnt1) into one register.
8956 // It's fine because both counters are 32bit and are not changed in this
8957 // loop. Just restore it on exit. So, cnt1 can be re-used in this loop.
8958 __ orr(cnt2, cnt2, cnt1, __ LSL, BitsPerByte * wordSize / 2);
8959 __ sub(result, result, 1);
8960 __ BIND(L_HAS_ZERO_LOOP);
8961 __ mov(cnt1, wordSize/str2_chr_size);
8962 __ cmp(cnt1, cnt2, __ LSR, BitsPerByte * wordSize / 2);
8963 __ br(__ GE, L_CMP_LOOP_LAST_CMP2); // case of 8 bytes only to compare
8964 if (str2_isL) {
8965 __ lsr(ch2, tmp4, LogBitsPerByte + str2_chr_shift); // char index
8966 __ ldr(ch2, Address(str2, ch2)); // read whole register of str2. Safe.
8967 __ lslv(tmp2, tmp2, tmp4);
8968 __ add(str2, str2, tmp4, __ LSR, LogBitsPerByte + str2_chr_shift);
8969 __ add(tmp4, tmp4, 1);
8970 __ add(result, result, tmp4, __ LSR, LogBitsPerByte + str2_chr_shift);
8971 __ lsl(tmp2, tmp2, 1);
8972 __ mov(tmp4, wordSize/str2_chr_size);
8973 } else {
8974 __ mov(ch2, 0xE);
8975 __ andr(ch2, ch2, tmp4, __ LSR, LogBitsPerByte); // byte shift amount
8976 __ ldr(ch2, Address(str2, ch2)); // read whole register of str2. Safe.
8977 __ lslv(tmp2, tmp2, tmp4);
8978 __ add(tmp4, tmp4, 1);
8979 __ add(result, result, tmp4, __ LSR, LogBitsPerByte + str2_chr_shift);
8980 __ add(str2, str2, tmp4, __ LSR, LogBitsPerByte);
8981 __ lsl(tmp2, tmp2, 1);
8982 __ mov(tmp4, wordSize/str2_chr_size);
8983 __ sub(str2, str2, str2_chr_size);
8984 }
8985 __ cmp(ch1, ch2);
8986 __ mov(tmp4, wordSize/str2_chr_size);
8987 __ br(__ NE, L_CMP_LOOP_NOMATCH);
8988 __ BIND(L_CMP_LOOP);
8989 str1_isL ? __ ldrb(cnt1, Address(str1, tmp4, Address::lsl(str1_chr_shift)))
8990 : __ ldrh(cnt1, Address(str1, tmp4, Address::lsl(str1_chr_shift)));
8991 str2_isL ? __ ldrb(ch2, Address(str2, tmp4, Address::lsl(str2_chr_shift)))
8992 : __ ldrh(ch2, Address(str2, tmp4, Address::lsl(str2_chr_shift)));
8993 __ add(tmp4, tmp4, 1);
8994 __ cmp(tmp4, cnt2, __ LSR, BitsPerByte * wordSize / 2);
8995 __ br(__ GE, L_CMP_LOOP_LAST_CMP);
8996 __ cmp(cnt1, ch2);
8997 __ br(__ EQ, L_CMP_LOOP);
8998 __ BIND(L_CMP_LOOP_NOMATCH);
8999 // here we're not matched
9000 __ cbz(tmp2, L_HAS_ZERO_LOOP_NOMATCH); // no more matches. Proceed to main loop
9001 __ clz(tmp4, tmp2);
9002 __ add(str2, str2, str2_chr_size); // advance pointer
9003 __ b(L_HAS_ZERO_LOOP);
9004 __ align(OptoLoopAlignment);
9005 __ BIND(L_CMP_LOOP_LAST_CMP);
9006 __ cmp(cnt1, ch2);
9007 __ br(__ NE, L_CMP_LOOP_NOMATCH);
9008 __ b(DONE);
9009 __ align(OptoLoopAlignment);
9010 __ BIND(L_CMP_LOOP_LAST_CMP2);
9011 if (str2_isL) {
9012 __ lsr(ch2, tmp4, LogBitsPerByte + str2_chr_shift); // char index
9013 __ ldr(ch2, Address(str2, ch2)); // read whole register of str2. Safe.
9014 __ lslv(tmp2, tmp2, tmp4);
9015 __ add(str2, str2, tmp4, __ LSR, LogBitsPerByte + str2_chr_shift);
9016 __ add(tmp4, tmp4, 1);
9017 __ add(result, result, tmp4, __ LSR, LogBitsPerByte + str2_chr_shift);
9018 __ lsl(tmp2, tmp2, 1);
9019 } else {
9020 __ mov(ch2, 0xE);
9021 __ andr(ch2, ch2, tmp4, __ LSR, LogBitsPerByte); // byte shift amount
9022 __ ldr(ch2, Address(str2, ch2)); // read whole register of str2. Safe.
9023 __ lslv(tmp2, tmp2, tmp4);
9024 __ add(tmp4, tmp4, 1);
9025 __ add(result, result, tmp4, __ LSR, LogBitsPerByte + str2_chr_shift);
9026 __ add(str2, str2, tmp4, __ LSR, LogBitsPerByte);
9027 __ lsl(tmp2, tmp2, 1);
9028 __ sub(str2, str2, str2_chr_size);
9029 }
9030 __ cmp(ch1, ch2);
9031 __ br(__ NE, L_CMP_LOOP_NOMATCH);
9032 __ b(DONE);
9033 __ align(OptoLoopAlignment);
9034 __ BIND(L_HAS_ZERO_LOOP_NOMATCH);
9035 // 1) Restore "result" index. Index was wordSize/str2_chr_size * N until
9036 // L_HAS_ZERO block. Byte octet was analyzed in L_HAS_ZERO_LOOP,
9037 // so, result was increased at max by wordSize/str2_chr_size - 1, so,
9038 // respective high bit wasn't changed. L_LOOP_PROCEED will increase
9039 // result by analyzed characters value, so, we can just reset lower bits
9040 // in result here. Clear 2 lower bits for UU/UL and 3 bits for LL
9041 // 2) restore cnt1 and cnt2 values from "compressed" cnt2
9042 // 3) advance str2 value to represent next str2 octet. result & 7/3 is
9043 // index of last analyzed substring inside current octet. So, str2 in at
9044 // respective start address. We need to advance it to next octet
9045 __ andr(tmp2, result, wordSize/str2_chr_size - 1); // symbols analyzed
9046 __ lsr(cnt1, cnt2, BitsPerByte * wordSize / 2);
9047 __ bfm(result, zr, 0, 2 - str2_chr_shift);
9048 __ sub(str2, str2, tmp2, __ LSL, str2_chr_shift); // restore str2
9049 __ movw(cnt2, cnt2);
9050 __ b(L_LOOP_PROCEED);
9051 __ align(OptoLoopAlignment);
9052 __ BIND(NOMATCH);
9053 __ mov(result, -1);
9054 __ BIND(DONE);
9055 __ pop(spilled_regs, sp);
9056 __ ret(lr);
9057 return entry;
9058 }
9059
9060 void generate_string_indexof_stubs() {
9061 StubRoutines::aarch64::_string_indexof_linear_ll = generate_string_indexof_linear(true, true);
9062 StubRoutines::aarch64::_string_indexof_linear_uu = generate_string_indexof_linear(false, false);
9063 StubRoutines::aarch64::_string_indexof_linear_ul = generate_string_indexof_linear(true, false);
9064 }
9065
9066 void inflate_and_store_2_fp_registers(bool generatePrfm,
9067 FloatRegister src1, FloatRegister src2) {
9068 Register dst = r1;
9069 __ zip1(v1, __ T16B, src1, v0);
9070 __ zip2(v2, __ T16B, src1, v0);
9071 if (generatePrfm) {
9072 __ prfm(Address(dst, SoftwarePrefetchHintDistance), PSTL1STRM);
9073 }
9074 __ zip1(v3, __ T16B, src2, v0);
9075 __ zip2(v4, __ T16B, src2, v0);
9076 __ st1(v1, v2, v3, v4, __ T16B, Address(__ post(dst, 64)));
9077 }
9078
9079 // R0 = src
9080 // R1 = dst
9081 // R2 = len
9082 // R3 = len >> 3
9083 // V0 = 0
9084 // v1 = loaded 8 bytes
9085 // Clobbers: r0, r1, r3, rscratch1, rflags, v0-v6
9086 address generate_large_byte_array_inflate() {
9087 __ align(CodeEntryAlignment);
9088 StubGenStubId stub_id = StubGenStubId::large_byte_array_inflate_id;
9089 StubCodeMark mark(this, stub_id);
9090 address entry = __ pc();
9091 Label LOOP, LOOP_START, LOOP_PRFM, LOOP_PRFM_START, DONE;
9092 Register src = r0, dst = r1, len = r2, octetCounter = r3;
9093 const int large_loop_threshold = MAX2(64, SoftwarePrefetchHintDistance)/8 + 4;
9094
9095 // do one more 8-byte read to have address 16-byte aligned in most cases
9096 // also use single store instruction
9097 __ ldrd(v2, __ post(src, 8));
9098 __ sub(octetCounter, octetCounter, 2);
9099 __ zip1(v1, __ T16B, v1, v0);
9100 __ zip1(v2, __ T16B, v2, v0);
9101 __ st1(v1, v2, __ T16B, __ post(dst, 32));
9102 __ ld1(v3, v4, v5, v6, __ T16B, Address(__ post(src, 64)));
9103 __ subs(rscratch1, octetCounter, large_loop_threshold);
9104 __ br(__ LE, LOOP_START);
9105 __ b(LOOP_PRFM_START);
9106 __ bind(LOOP_PRFM);
9107 __ ld1(v3, v4, v5, v6, __ T16B, Address(__ post(src, 64)));
9108 __ bind(LOOP_PRFM_START);
9109 __ prfm(Address(src, SoftwarePrefetchHintDistance));
9110 __ sub(octetCounter, octetCounter, 8);
9111 __ subs(rscratch1, octetCounter, large_loop_threshold);
9112 inflate_and_store_2_fp_registers(true, v3, v4);
9113 inflate_and_store_2_fp_registers(true, v5, v6);
9114 __ br(__ GT, LOOP_PRFM);
9115 __ cmp(octetCounter, (u1)8);
9116 __ br(__ LT, DONE);
9117 __ bind(LOOP);
9118 __ ld1(v3, v4, v5, v6, __ T16B, Address(__ post(src, 64)));
9119 __ bind(LOOP_START);
9120 __ sub(octetCounter, octetCounter, 8);
9121 __ cmp(octetCounter, (u1)8);
9122 inflate_and_store_2_fp_registers(false, v3, v4);
9123 inflate_and_store_2_fp_registers(false, v5, v6);
9124 __ br(__ GE, LOOP);
9125 __ bind(DONE);
9126 __ ret(lr);
9127 return entry;
9128 }
9129
9130 /**
9131 * Arguments:
9132 *
9133 * Input:
9134 * c_rarg0 - current state address
9135 * c_rarg1 - H key address
9136 * c_rarg2 - data address
9137 * c_rarg3 - number of blocks
9138 *
9139 * Output:
9140 * Updated state at c_rarg0
9141 */
9142 address generate_ghash_processBlocks() {
9143 // Bafflingly, GCM uses little-endian for the byte order, but
9144 // big-endian for the bit order. For example, the polynomial 1 is
9145 // represented as the 16-byte string 80 00 00 00 | 12 bytes of 00.
9146 //
9147 // So, we must either reverse the bytes in each word and do
9148 // everything big-endian or reverse the bits in each byte and do
9149 // it little-endian. On AArch64 it's more idiomatic to reverse
9150 // the bits in each byte (we have an instruction, RBIT, to do
9151 // that) and keep the data in little-endian bit order through the
9152 // calculation, bit-reversing the inputs and outputs.
9153
9154 StubGenStubId stub_id = StubGenStubId::ghash_processBlocks_id;
9155 StubCodeMark mark(this, stub_id);
9156 __ align(wordSize * 2);
9157 address p = __ pc();
9158 __ emit_int64(0x87); // The low-order bits of the field
9159 // polynomial (i.e. p = z^7+z^2+z+1)
9160 // repeated in the low and high parts of a
9161 // 128-bit vector
9162 __ emit_int64(0x87);
9163
9164 __ align(CodeEntryAlignment);
9165 address start = __ pc();
9166
9167 Register state = c_rarg0;
9168 Register subkeyH = c_rarg1;
9169 Register data = c_rarg2;
9170 Register blocks = c_rarg3;
9171
9172 FloatRegister vzr = v30;
9173 __ eor(vzr, __ T16B, vzr, vzr); // zero register
9174
9175 __ ldrq(v24, p); // The field polynomial
9176
9177 __ ldrq(v0, Address(state));
9178 __ ldrq(v1, Address(subkeyH));
9179
9180 __ rev64(v0, __ T16B, v0); // Bit-reverse words in state and subkeyH
9181 __ rbit(v0, __ T16B, v0);
9182 __ rev64(v1, __ T16B, v1);
9183 __ rbit(v1, __ T16B, v1);
9184
9185 __ ext(v4, __ T16B, v1, v1, 0x08); // long-swap subkeyH into v1
9186 __ eor(v4, __ T16B, v4, v1); // xor subkeyH into subkeyL (Karatsuba: (A1+A0))
9187
9188 {
9189 Label L_ghash_loop;
9190 __ bind(L_ghash_loop);
9191
9192 __ ldrq(v2, Address(__ post(data, 0x10))); // Load the data, bit
9193 // reversing each byte
9194 __ rbit(v2, __ T16B, v2);
9195 __ eor(v2, __ T16B, v0, v2); // bit-swapped data ^ bit-swapped state
9196
9197 // Multiply state in v2 by subkey in v1
9198 __ ghash_multiply(/*result_lo*/v5, /*result_hi*/v7,
9199 /*a*/v1, /*b*/v2, /*a1_xor_a0*/v4,
9200 /*temps*/v6, v3, /*reuse/clobber b*/v2);
9201 // Reduce v7:v5 by the field polynomial
9202 __ ghash_reduce(/*result*/v0, /*lo*/v5, /*hi*/v7, /*p*/v24, vzr, /*temp*/v3);
9203
9204 __ sub(blocks, blocks, 1);
9205 __ cbnz(blocks, L_ghash_loop);
9206 }
9207
9208 // The bit-reversed result is at this point in v0
9209 __ rev64(v0, __ T16B, v0);
9210 __ rbit(v0, __ T16B, v0);
9211
9212 __ st1(v0, __ T16B, state);
9213 __ ret(lr);
9214
9215 return start;
9216 }
9217
9218 address generate_ghash_processBlocks_wide() {
9219 address small = generate_ghash_processBlocks();
9220
9221 StubGenStubId stub_id = StubGenStubId::ghash_processBlocks_wide_id;
9222 StubCodeMark mark(this, stub_id);
9223 __ align(wordSize * 2);
9224 address p = __ pc();
9225 __ emit_int64(0x87); // The low-order bits of the field
9226 // polynomial (i.e. p = z^7+z^2+z+1)
9227 // repeated in the low and high parts of a
9228 // 128-bit vector
9229 __ emit_int64(0x87);
9230
9231 __ align(CodeEntryAlignment);
9232 address start = __ pc();
9233
9234 Register state = c_rarg0;
9235 Register subkeyH = c_rarg1;
9236 Register data = c_rarg2;
9237 Register blocks = c_rarg3;
9238
9239 const int unroll = 4;
9240
9241 __ cmp(blocks, (unsigned char)(unroll * 2));
9242 __ br(__ LT, small);
9243
9244 if (unroll > 1) {
9245 // Save state before entering routine
9246 __ sub(sp, sp, 4 * 16);
9247 __ st1(v12, v13, v14, v15, __ T16B, Address(sp));
9248 __ sub(sp, sp, 4 * 16);
9249 __ st1(v8, v9, v10, v11, __ T16B, Address(sp));
9250 }
9251
9252 __ ghash_processBlocks_wide(p, state, subkeyH, data, blocks, unroll);
9253
9254 if (unroll > 1) {
9255 // And restore state
9256 __ ld1(v8, v9, v10, v11, __ T16B, __ post(sp, 4 * 16));
9257 __ ld1(v12, v13, v14, v15, __ T16B, __ post(sp, 4 * 16));
9258 }
9259
9260 __ cmp(blocks, (unsigned char)0);
9261 __ br(__ GT, small);
9262
9263 __ ret(lr);
9264
9265 return start;
9266 }
9267
9268 void generate_base64_encode_simdround(Register src, Register dst,
9269 FloatRegister codec, u8 size) {
9270
9271 FloatRegister in0 = v4, in1 = v5, in2 = v6;
9272 FloatRegister out0 = v16, out1 = v17, out2 = v18, out3 = v19;
9273 FloatRegister ind0 = v20, ind1 = v21, ind2 = v22, ind3 = v23;
9274
9275 Assembler::SIMD_Arrangement arrangement = size == 16 ? __ T16B : __ T8B;
9276
9277 __ ld3(in0, in1, in2, arrangement, __ post(src, 3 * size));
9278
9279 __ ushr(ind0, arrangement, in0, 2);
9280
9281 __ ushr(ind1, arrangement, in1, 2);
9282 __ shl(in0, arrangement, in0, 6);
9283 __ orr(ind1, arrangement, ind1, in0);
9284 __ ushr(ind1, arrangement, ind1, 2);
9285
9286 __ ushr(ind2, arrangement, in2, 4);
9287 __ shl(in1, arrangement, in1, 4);
9288 __ orr(ind2, arrangement, in1, ind2);
9289 __ ushr(ind2, arrangement, ind2, 2);
9290
9291 __ shl(ind3, arrangement, in2, 2);
9292 __ ushr(ind3, arrangement, ind3, 2);
9293
9294 __ tbl(out0, arrangement, codec, 4, ind0);
9295 __ tbl(out1, arrangement, codec, 4, ind1);
9296 __ tbl(out2, arrangement, codec, 4, ind2);
9297 __ tbl(out3, arrangement, codec, 4, ind3);
9298
9299 __ st4(out0, out1, out2, out3, arrangement, __ post(dst, 4 * size));
9300 }
9301
9302 /**
9303 * Arguments:
9304 *
9305 * Input:
9306 * c_rarg0 - src_start
9307 * c_rarg1 - src_offset
9308 * c_rarg2 - src_length
9309 * c_rarg3 - dest_start
9310 * c_rarg4 - dest_offset
9311 * c_rarg5 - isURL
9312 *
9313 */
9314 address generate_base64_encodeBlock() {
9315
9316 static const char toBase64[64] = {
9317 'A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'I', 'J', 'K', 'L', 'M',
9318 'N', 'O', 'P', 'Q', 'R', 'S', 'T', 'U', 'V', 'W', 'X', 'Y', 'Z',
9319 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm',
9320 'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z',
9321 '0', '1', '2', '3', '4', '5', '6', '7', '8', '9', '+', '/'
9322 };
9323
9324 static const char toBase64URL[64] = {
9325 'A', 'B', 'C', 'D', 'E', 'F', 'G', 'H', 'I', 'J', 'K', 'L', 'M',
9326 'N', 'O', 'P', 'Q', 'R', 'S', 'T', 'U', 'V', 'W', 'X', 'Y', 'Z',
9327 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm',
9328 'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z',
9329 '0', '1', '2', '3', '4', '5', '6', '7', '8', '9', '-', '_'
9330 };
9331
9332 __ align(CodeEntryAlignment);
9333 StubGenStubId stub_id = StubGenStubId::base64_encodeBlock_id;
9334 StubCodeMark mark(this, stub_id);
9335 address start = __ pc();
9336
9337 Register src = c_rarg0; // source array
9338 Register soff = c_rarg1; // source start offset
9339 Register send = c_rarg2; // source end offset
9340 Register dst = c_rarg3; // dest array
9341 Register doff = c_rarg4; // position for writing to dest array
9342 Register isURL = c_rarg5; // Base64 or URL character set
9343
9344 // c_rarg6 and c_rarg7 are free to use as temps
9345 Register codec = c_rarg6;
9346 Register length = c_rarg7;
9347
9348 Label ProcessData, Process48B, Process24B, Process3B, SIMDExit, Exit;
9349
9350 __ add(src, src, soff);
9351 __ add(dst, dst, doff);
9352 __ sub(length, send, soff);
9353
9354 // load the codec base address
9355 __ lea(codec, ExternalAddress((address) toBase64));
9356 __ cbz(isURL, ProcessData);
9357 __ lea(codec, ExternalAddress((address) toBase64URL));
9358
9359 __ BIND(ProcessData);
9360
9361 // too short to formup a SIMD loop, roll back
9362 __ cmp(length, (u1)24);
9363 __ br(Assembler::LT, Process3B);
9364
9365 __ ld1(v0, v1, v2, v3, __ T16B, Address(codec));
9366
9367 __ BIND(Process48B);
9368 __ cmp(length, (u1)48);
9369 __ br(Assembler::LT, Process24B);
9370 generate_base64_encode_simdround(src, dst, v0, 16);
9371 __ sub(length, length, 48);
9372 __ b(Process48B);
9373
9374 __ BIND(Process24B);
9375 __ cmp(length, (u1)24);
9376 __ br(Assembler::LT, SIMDExit);
9377 generate_base64_encode_simdround(src, dst, v0, 8);
9378 __ sub(length, length, 24);
9379
9380 __ BIND(SIMDExit);
9381 __ cbz(length, Exit);
9382
9383 __ BIND(Process3B);
9384 // 3 src bytes, 24 bits
9385 __ ldrb(r10, __ post(src, 1));
9386 __ ldrb(r11, __ post(src, 1));
9387 __ ldrb(r12, __ post(src, 1));
9388 __ orrw(r11, r11, r10, Assembler::LSL, 8);
9389 __ orrw(r12, r12, r11, Assembler::LSL, 8);
9390 // codec index
9391 __ ubfmw(r15, r12, 18, 23);
9392 __ ubfmw(r14, r12, 12, 17);
9393 __ ubfmw(r13, r12, 6, 11);
9394 __ andw(r12, r12, 63);
9395 // get the code based on the codec
9396 __ ldrb(r15, Address(codec, r15, Address::uxtw(0)));
9397 __ ldrb(r14, Address(codec, r14, Address::uxtw(0)));
9398 __ ldrb(r13, Address(codec, r13, Address::uxtw(0)));
9399 __ ldrb(r12, Address(codec, r12, Address::uxtw(0)));
9400 __ strb(r15, __ post(dst, 1));
9401 __ strb(r14, __ post(dst, 1));
9402 __ strb(r13, __ post(dst, 1));
9403 __ strb(r12, __ post(dst, 1));
9404 __ sub(length, length, 3);
9405 __ cbnz(length, Process3B);
9406
9407 __ BIND(Exit);
9408 __ ret(lr);
9409
9410 return start;
9411 }
9412
9413 void generate_base64_decode_simdround(Register src, Register dst,
9414 FloatRegister codecL, FloatRegister codecH, int size, Label& Exit) {
9415
9416 FloatRegister in0 = v16, in1 = v17, in2 = v18, in3 = v19;
9417 FloatRegister out0 = v20, out1 = v21, out2 = v22;
9418
9419 FloatRegister decL0 = v23, decL1 = v24, decL2 = v25, decL3 = v26;
9420 FloatRegister decH0 = v28, decH1 = v29, decH2 = v30, decH3 = v31;
9421
9422 Label NoIllegalData, ErrorInLowerHalf, StoreLegalData;
9423
9424 Assembler::SIMD_Arrangement arrangement = size == 16 ? __ T16B : __ T8B;
9425
9426 __ ld4(in0, in1, in2, in3, arrangement, __ post(src, 4 * size));
9427
9428 // we need unsigned saturating subtract, to make sure all input values
9429 // in range [0, 63] will have 0U value in the higher half lookup
9430 __ uqsubv(decH0, __ T16B, in0, v27);
9431 __ uqsubv(decH1, __ T16B, in1, v27);
9432 __ uqsubv(decH2, __ T16B, in2, v27);
9433 __ uqsubv(decH3, __ T16B, in3, v27);
9434
9435 // lower half lookup
9436 __ tbl(decL0, arrangement, codecL, 4, in0);
9437 __ tbl(decL1, arrangement, codecL, 4, in1);
9438 __ tbl(decL2, arrangement, codecL, 4, in2);
9439 __ tbl(decL3, arrangement, codecL, 4, in3);
9440
9441 // higher half lookup
9442 __ tbx(decH0, arrangement, codecH, 4, decH0);
9443 __ tbx(decH1, arrangement, codecH, 4, decH1);
9444 __ tbx(decH2, arrangement, codecH, 4, decH2);
9445 __ tbx(decH3, arrangement, codecH, 4, decH3);
9446
9447 // combine lower and higher
9448 __ orr(decL0, arrangement, decL0, decH0);
9449 __ orr(decL1, arrangement, decL1, decH1);
9450 __ orr(decL2, arrangement, decL2, decH2);
9451 __ orr(decL3, arrangement, decL3, decH3);
9452
9453 // check illegal inputs, value larger than 63 (maximum of 6 bits)
9454 __ cm(Assembler::HI, decH0, arrangement, decL0, v27);
9455 __ cm(Assembler::HI, decH1, arrangement, decL1, v27);
9456 __ cm(Assembler::HI, decH2, arrangement, decL2, v27);
9457 __ cm(Assembler::HI, decH3, arrangement, decL3, v27);
9458 __ orr(in0, arrangement, decH0, decH1);
9459 __ orr(in1, arrangement, decH2, decH3);
9460 __ orr(in2, arrangement, in0, in1);
9461 __ umaxv(in3, arrangement, in2);
9462 __ umov(rscratch2, in3, __ B, 0);
9463
9464 // get the data to output
9465 __ shl(out0, arrangement, decL0, 2);
9466 __ ushr(out1, arrangement, decL1, 4);
9467 __ orr(out0, arrangement, out0, out1);
9468 __ shl(out1, arrangement, decL1, 4);
9469 __ ushr(out2, arrangement, decL2, 2);
9470 __ orr(out1, arrangement, out1, out2);
9471 __ shl(out2, arrangement, decL2, 6);
9472 __ orr(out2, arrangement, out2, decL3);
9473
9474 __ cbz(rscratch2, NoIllegalData);
9475
9476 // handle illegal input
9477 __ umov(r10, in2, __ D, 0);
9478 if (size == 16) {
9479 __ cbnz(r10, ErrorInLowerHalf);
9480
9481 // illegal input is in higher half, store the lower half now.
9482 __ st3(out0, out1, out2, __ T8B, __ post(dst, 24));
9483
9484 __ umov(r10, in2, __ D, 1);
9485 __ umov(r11, out0, __ D, 1);
9486 __ umov(r12, out1, __ D, 1);
9487 __ umov(r13, out2, __ D, 1);
9488 __ b(StoreLegalData);
9489
9490 __ BIND(ErrorInLowerHalf);
9491 }
9492 __ umov(r11, out0, __ D, 0);
9493 __ umov(r12, out1, __ D, 0);
9494 __ umov(r13, out2, __ D, 0);
9495
9496 __ BIND(StoreLegalData);
9497 __ tbnz(r10, 5, Exit); // 0xff indicates illegal input
9498 __ strb(r11, __ post(dst, 1));
9499 __ strb(r12, __ post(dst, 1));
9500 __ strb(r13, __ post(dst, 1));
9501 __ lsr(r10, r10, 8);
9502 __ lsr(r11, r11, 8);
9503 __ lsr(r12, r12, 8);
9504 __ lsr(r13, r13, 8);
9505 __ b(StoreLegalData);
9506
9507 __ BIND(NoIllegalData);
9508 __ st3(out0, out1, out2, arrangement, __ post(dst, 3 * size));
9509 }
9510
9511
9512 /**
9513 * Arguments:
9514 *
9515 * Input:
9516 * c_rarg0 - src_start
9517 * c_rarg1 - src_offset
9518 * c_rarg2 - src_length
9519 * c_rarg3 - dest_start
9520 * c_rarg4 - dest_offset
9521 * c_rarg5 - isURL
9522 * c_rarg6 - isMIME
9523 *
9524 */
9525 address generate_base64_decodeBlock() {
9526
9527 // The SIMD part of this Base64 decode intrinsic is based on the algorithm outlined
9528 // on http://0x80.pl/articles/base64-simd-neon.html#encoding-quadwords, in section
9529 // titled "Base64 decoding".
9530
9531 // Non-SIMD lookup tables are mostly dumped from fromBase64 array used in java.util.Base64,
9532 // except the trailing character '=' is also treated illegal value in this intrinsic. That
9533 // is java.util.Base64.fromBase64['='] = -2, while fromBase(URL)64ForNoSIMD['='] = 255 here.
9534 static const uint8_t fromBase64ForNoSIMD[256] = {
9535 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9536 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9537 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 62u, 255u, 255u, 255u, 63u,
9538 52u, 53u, 54u, 55u, 56u, 57u, 58u, 59u, 60u, 61u, 255u, 255u, 255u, 255u, 255u, 255u,
9539 255u, 0u, 1u, 2u, 3u, 4u, 5u, 6u, 7u, 8u, 9u, 10u, 11u, 12u, 13u, 14u,
9540 15u, 16u, 17u, 18u, 19u, 20u, 21u, 22u, 23u, 24u, 25u, 255u, 255u, 255u, 255u, 255u,
9541 255u, 26u, 27u, 28u, 29u, 30u, 31u, 32u, 33u, 34u, 35u, 36u, 37u, 38u, 39u, 40u,
9542 41u, 42u, 43u, 44u, 45u, 46u, 47u, 48u, 49u, 50u, 51u, 255u, 255u, 255u, 255u, 255u,
9543 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9544 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9545 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9546 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9547 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9548 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9549 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9550 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9551 };
9552
9553 static const uint8_t fromBase64URLForNoSIMD[256] = {
9554 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9555 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9556 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 62u, 255u, 255u,
9557 52u, 53u, 54u, 55u, 56u, 57u, 58u, 59u, 60u, 61u, 255u, 255u, 255u, 255u, 255u, 255u,
9558 255u, 0u, 1u, 2u, 3u, 4u, 5u, 6u, 7u, 8u, 9u, 10u, 11u, 12u, 13u, 14u,
9559 15u, 16u, 17u, 18u, 19u, 20u, 21u, 22u, 23u, 24u, 25u, 255u, 255u, 255u, 255u, 63u,
9560 255u, 26u, 27u, 28u, 29u, 30u, 31u, 32u, 33u, 34u, 35u, 36u, 37u, 38u, 39u, 40u,
9561 41u, 42u, 43u, 44u, 45u, 46u, 47u, 48u, 49u, 50u, 51u, 255u, 255u, 255u, 255u, 255u,
9562 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9563 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9564 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9565 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9566 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9567 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9568 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9569 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9570 };
9571
9572 // A legal value of base64 code is in range [0, 127]. We need two lookups
9573 // with tbl/tbx and combine them to get the decode data. The 1st table vector
9574 // lookup use tbl, out of range indices are set to 0 in destination. The 2nd
9575 // table vector lookup use tbx, out of range indices are unchanged in
9576 // destination. Input [64..126] is mapped to index [65, 127] in second lookup.
9577 // The value of index 64 is set to 0, so that we know that we already get the
9578 // decoded data with the 1st lookup.
9579 static const uint8_t fromBase64ForSIMD[128] = {
9580 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9581 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9582 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 62u, 255u, 255u, 255u, 63u,
9583 52u, 53u, 54u, 55u, 56u, 57u, 58u, 59u, 60u, 61u, 255u, 255u, 255u, 255u, 255u, 255u,
9584 0u, 255u, 0u, 1u, 2u, 3u, 4u, 5u, 6u, 7u, 8u, 9u, 10u, 11u, 12u, 13u,
9585 14u, 15u, 16u, 17u, 18u, 19u, 20u, 21u, 22u, 23u, 24u, 25u, 255u, 255u, 255u, 255u,
9586 255u, 255u, 26u, 27u, 28u, 29u, 30u, 31u, 32u, 33u, 34u, 35u, 36u, 37u, 38u, 39u,
9587 40u, 41u, 42u, 43u, 44u, 45u, 46u, 47u, 48u, 49u, 50u, 51u, 255u, 255u, 255u, 255u,
9588 };
9589
9590 static const uint8_t fromBase64URLForSIMD[128] = {
9591 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9592 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u,
9593 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 255u, 62u, 255u, 255u,
9594 52u, 53u, 54u, 55u, 56u, 57u, 58u, 59u, 60u, 61u, 255u, 255u, 255u, 255u, 255u, 255u,
9595 0u, 255u, 0u, 1u, 2u, 3u, 4u, 5u, 6u, 7u, 8u, 9u, 10u, 11u, 12u, 13u,
9596 14u, 15u, 16u, 17u, 18u, 19u, 20u, 21u, 22u, 23u, 24u, 25u, 255u, 255u, 255u, 255u,
9597 63u, 255u, 26u, 27u, 28u, 29u, 30u, 31u, 32u, 33u, 34u, 35u, 36u, 37u, 38u, 39u,
9598 40u, 41u, 42u, 43u, 44u, 45u, 46u, 47u, 48u, 49u, 50u, 51u, 255u, 255u, 255u, 255u,
9599 };
9600
9601 __ align(CodeEntryAlignment);
9602 StubGenStubId stub_id = StubGenStubId::base64_decodeBlock_id;
9603 StubCodeMark mark(this, stub_id);
9604 address start = __ pc();
9605
9606 Register src = c_rarg0; // source array
9607 Register soff = c_rarg1; // source start offset
9608 Register send = c_rarg2; // source end offset
9609 Register dst = c_rarg3; // dest array
9610 Register doff = c_rarg4; // position for writing to dest array
9611 Register isURL = c_rarg5; // Base64 or URL character set
9612 Register isMIME = c_rarg6; // Decoding MIME block - unused in this implementation
9613
9614 Register length = send; // reuse send as length of source data to process
9615
9616 Register simd_codec = c_rarg6;
9617 Register nosimd_codec = c_rarg7;
9618
9619 Label ProcessData, Process64B, Process32B, Process4B, SIMDEnter, SIMDExit, Exit;
9620
9621 __ enter();
9622
9623 __ add(src, src, soff);
9624 __ add(dst, dst, doff);
9625
9626 __ mov(doff, dst);
9627
9628 __ sub(length, send, soff);
9629 __ bfm(length, zr, 0, 1);
9630
9631 __ lea(nosimd_codec, ExternalAddress((address) fromBase64ForNoSIMD));
9632 __ cbz(isURL, ProcessData);
9633 __ lea(nosimd_codec, ExternalAddress((address) fromBase64URLForNoSIMD));
9634
9635 __ BIND(ProcessData);
9636 __ mov(rscratch1, length);
9637 __ cmp(length, (u1)144); // 144 = 80 + 64
9638 __ br(Assembler::LT, Process4B);
9639
9640 // In the MIME case, the line length cannot be more than 76
9641 // bytes (see RFC 2045). This is too short a block for SIMD
9642 // to be worthwhile, so we use non-SIMD here.
9643 __ movw(rscratch1, 79);
9644
9645 __ BIND(Process4B);
9646 __ ldrw(r14, __ post(src, 4));
9647 __ ubfxw(r10, r14, 0, 8);
9648 __ ubfxw(r11, r14, 8, 8);
9649 __ ubfxw(r12, r14, 16, 8);
9650 __ ubfxw(r13, r14, 24, 8);
9651 // get the de-code
9652 __ ldrb(r10, Address(nosimd_codec, r10, Address::uxtw(0)));
9653 __ ldrb(r11, Address(nosimd_codec, r11, Address::uxtw(0)));
9654 __ ldrb(r12, Address(nosimd_codec, r12, Address::uxtw(0)));
9655 __ ldrb(r13, Address(nosimd_codec, r13, Address::uxtw(0)));
9656 // error detection, 255u indicates an illegal input
9657 __ orrw(r14, r10, r11);
9658 __ orrw(r15, r12, r13);
9659 __ orrw(r14, r14, r15);
9660 __ tbnz(r14, 7, Exit);
9661 // recover the data
9662 __ lslw(r14, r10, 10);
9663 __ bfiw(r14, r11, 4, 6);
9664 __ bfmw(r14, r12, 2, 5);
9665 __ rev16w(r14, r14);
9666 __ bfiw(r13, r12, 6, 2);
9667 __ strh(r14, __ post(dst, 2));
9668 __ strb(r13, __ post(dst, 1));
9669 // non-simd loop
9670 __ subsw(rscratch1, rscratch1, 4);
9671 __ br(Assembler::GT, Process4B);
9672
9673 // if exiting from PreProcess80B, rscratch1 == -1;
9674 // otherwise, rscratch1 == 0.
9675 __ cbzw(rscratch1, Exit);
9676 __ sub(length, length, 80);
9677
9678 __ lea(simd_codec, ExternalAddress((address) fromBase64ForSIMD));
9679 __ cbz(isURL, SIMDEnter);
9680 __ lea(simd_codec, ExternalAddress((address) fromBase64URLForSIMD));
9681
9682 __ BIND(SIMDEnter);
9683 __ ld1(v0, v1, v2, v3, __ T16B, __ post(simd_codec, 64));
9684 __ ld1(v4, v5, v6, v7, __ T16B, Address(simd_codec));
9685 __ mov(rscratch1, 63);
9686 __ dup(v27, __ T16B, rscratch1);
9687
9688 __ BIND(Process64B);
9689 __ cmp(length, (u1)64);
9690 __ br(Assembler::LT, Process32B);
9691 generate_base64_decode_simdround(src, dst, v0, v4, 16, Exit);
9692 __ sub(length, length, 64);
9693 __ b(Process64B);
9694
9695 __ BIND(Process32B);
9696 __ cmp(length, (u1)32);
9697 __ br(Assembler::LT, SIMDExit);
9698 generate_base64_decode_simdround(src, dst, v0, v4, 8, Exit);
9699 __ sub(length, length, 32);
9700 __ b(Process32B);
9701
9702 __ BIND(SIMDExit);
9703 __ cbz(length, Exit);
9704 __ movw(rscratch1, length);
9705 __ b(Process4B);
9706
9707 __ BIND(Exit);
9708 __ sub(c_rarg0, dst, doff);
9709
9710 __ leave();
9711 __ ret(lr);
9712
9713 return start;
9714 }
9715
9716 // Support for spin waits.
9717 address generate_spin_wait() {
9718 __ align(CodeEntryAlignment);
9719 StubGenStubId stub_id = StubGenStubId::spin_wait_id;
9720 StubCodeMark mark(this, stub_id);
9721 address start = __ pc();
9722
9723 __ spin_wait();
9724 __ ret(lr);
9725
9726 return start;
9727 }
9728
9729 void generate_lookup_secondary_supers_table_stub() {
9730 StubGenStubId stub_id = StubGenStubId::lookup_secondary_supers_table_id;
9731 StubCodeMark mark(this, stub_id);
9732
9733 const Register
9734 r_super_klass = r0,
9735 r_array_base = r1,
9736 r_array_length = r2,
9737 r_array_index = r3,
9738 r_sub_klass = r4,
9739 r_bitmap = rscratch2,
9740 result = r5;
9741 const FloatRegister
9742 vtemp = v0;
9743
9744 for (int slot = 0; slot < Klass::SECONDARY_SUPERS_TABLE_SIZE; slot++) {
9745 StubRoutines::_lookup_secondary_supers_table_stubs[slot] = __ pc();
9746 Label L_success;
9747 __ enter();
9748 __ lookup_secondary_supers_table_const(r_sub_klass, r_super_klass,
9749 r_array_base, r_array_length, r_array_index,
9750 vtemp, result, slot,
9751 /*stub_is_near*/true);
9752 __ leave();
9753 __ ret(lr);
9754 }
9755 }
9756
9757 // Slow path implementation for UseSecondarySupersTable.
9758 address generate_lookup_secondary_supers_table_slow_path_stub() {
9759 StubGenStubId stub_id = StubGenStubId::lookup_secondary_supers_table_slow_path_id;
9760 StubCodeMark mark(this, stub_id);
9761
9762 address start = __ pc();
9763 const Register
9764 r_super_klass = r0, // argument
9765 r_array_base = r1, // argument
9766 temp1 = r2, // temp
9767 r_array_index = r3, // argument
9768 r_bitmap = rscratch2, // argument
9769 result = r5; // argument
9770
9771 __ lookup_secondary_supers_table_slow_path(r_super_klass, r_array_base, r_array_index, r_bitmap, temp1, result);
9772 __ ret(lr);
9773
9774 return start;
9775 }
9776
9777 #if defined (LINUX) && !defined (__ARM_FEATURE_ATOMICS)
9778
9779 // ARMv8.1 LSE versions of the atomic stubs used by Atomic::PlatformXX.
9780 //
9781 // If LSE is in use, generate LSE versions of all the stubs. The
9782 // non-LSE versions are in atomic_aarch64.S.
9783
9784 // class AtomicStubMark records the entry point of a stub and the
9785 // stub pointer which will point to it. The stub pointer is set to
9786 // the entry point when ~AtomicStubMark() is called, which must be
9787 // after ICache::invalidate_range. This ensures safe publication of
9788 // the generated code.
9789 class AtomicStubMark {
9790 address _entry_point;
9791 aarch64_atomic_stub_t *_stub;
9792 MacroAssembler *_masm;
9793 public:
9794 AtomicStubMark(MacroAssembler *masm, aarch64_atomic_stub_t *stub) {
9795 _masm = masm;
9796 __ align(32);
9797 _entry_point = __ pc();
9798 _stub = stub;
9799 }
9800 ~AtomicStubMark() {
9801 *_stub = (aarch64_atomic_stub_t)_entry_point;
9802 }
9803 };
9804
9805 // NB: For memory_order_conservative we need a trailing membar after
9806 // LSE atomic operations but not a leading membar.
9807 //
9808 // We don't need a leading membar because a clause in the Arm ARM
9809 // says:
9810 //
9811 // Barrier-ordered-before
9812 //
9813 // Barrier instructions order prior Memory effects before subsequent
9814 // Memory effects generated by the same Observer. A read or a write
9815 // RW1 is Barrier-ordered-before a read or a write RW 2 from the same
9816 // Observer if and only if RW1 appears in program order before RW 2
9817 // and [ ... ] at least one of RW 1 and RW 2 is generated by an atomic
9818 // instruction with both Acquire and Release semantics.
9819 //
9820 // All the atomic instructions {ldaddal, swapal, casal} have Acquire
9821 // and Release semantics, therefore we don't need a leading
9822 // barrier. However, there is no corresponding Barrier-ordered-after
9823 // relationship, therefore we need a trailing membar to prevent a
9824 // later store or load from being reordered with the store in an
9825 // atomic instruction.
9826 //
9827 // This was checked by using the herd7 consistency model simulator
9828 // (http://diy.inria.fr/) with this test case:
9829 //
9830 // AArch64 LseCas
9831 // { 0:X1=x; 0:X2=y; 1:X1=x; 1:X2=y; }
9832 // P0 | P1;
9833 // LDR W4, [X2] | MOV W3, #0;
9834 // DMB LD | MOV W4, #1;
9835 // LDR W3, [X1] | CASAL W3, W4, [X1];
9836 // | DMB ISH;
9837 // | STR W4, [X2];
9838 // exists
9839 // (0:X3=0 /\ 0:X4=1)
9840 //
9841 // If X3 == 0 && X4 == 1, the store to y in P1 has been reordered
9842 // with the store to x in P1. Without the DMB in P1 this may happen.
9843 //
9844 // At the time of writing we don't know of any AArch64 hardware that
9845 // reorders stores in this way, but the Reference Manual permits it.
9846
9847 void gen_cas_entry(Assembler::operand_size size,
9848 atomic_memory_order order) {
9849 Register prev = r3, ptr = c_rarg0, compare_val = c_rarg1,
9850 exchange_val = c_rarg2;
9851 bool acquire, release;
9852 switch (order) {
9853 case memory_order_relaxed:
9854 acquire = false;
9855 release = false;
9856 break;
9857 case memory_order_release:
9858 acquire = false;
9859 release = true;
9860 break;
9861 default:
9862 acquire = true;
9863 release = true;
9864 break;
9865 }
9866 __ mov(prev, compare_val);
9867 __ lse_cas(prev, exchange_val, ptr, size, acquire, release, /*not_pair*/true);
9868 if (order == memory_order_conservative) {
9869 __ membar(Assembler::StoreStore|Assembler::StoreLoad);
9870 }
9871 if (size == Assembler::xword) {
9872 __ mov(r0, prev);
9873 } else {
9874 __ movw(r0, prev);
9875 }
9876 __ ret(lr);
9877 }
9878
9879 void gen_ldadd_entry(Assembler::operand_size size, atomic_memory_order order) {
9880 Register prev = r2, addr = c_rarg0, incr = c_rarg1;
9881 // If not relaxed, then default to conservative. Relaxed is the only
9882 // case we use enough to be worth specializing.
9883 if (order == memory_order_relaxed) {
9884 __ ldadd(size, incr, prev, addr);
9885 } else {
9886 __ ldaddal(size, incr, prev, addr);
9887 __ membar(Assembler::StoreStore|Assembler::StoreLoad);
9888 }
9889 if (size == Assembler::xword) {
9890 __ mov(r0, prev);
9891 } else {
9892 __ movw(r0, prev);
9893 }
9894 __ ret(lr);
9895 }
9896
9897 void gen_swpal_entry(Assembler::operand_size size) {
9898 Register prev = r2, addr = c_rarg0, incr = c_rarg1;
9899 __ swpal(size, incr, prev, addr);
9900 __ membar(Assembler::StoreStore|Assembler::StoreLoad);
9901 if (size == Assembler::xword) {
9902 __ mov(r0, prev);
9903 } else {
9904 __ movw(r0, prev);
9905 }
9906 __ ret(lr);
9907 }
9908
9909 void generate_atomic_entry_points() {
9910 if (! UseLSE) {
9911 return;
9912 }
9913 __ align(CodeEntryAlignment);
9914 StubGenStubId stub_id = StubGenStubId::atomic_entry_points_id;
9915 StubCodeMark mark(this, stub_id);
9916 address first_entry = __ pc();
9917
9918 // ADD, memory_order_conservative
9919 AtomicStubMark mark_fetch_add_4(_masm, &aarch64_atomic_fetch_add_4_impl);
9920 gen_ldadd_entry(Assembler::word, memory_order_conservative);
9921 AtomicStubMark mark_fetch_add_8(_masm, &aarch64_atomic_fetch_add_8_impl);
9922 gen_ldadd_entry(Assembler::xword, memory_order_conservative);
9923
9924 // ADD, memory_order_relaxed
9925 AtomicStubMark mark_fetch_add_4_relaxed
9926 (_masm, &aarch64_atomic_fetch_add_4_relaxed_impl);
9927 gen_ldadd_entry(MacroAssembler::word, memory_order_relaxed);
9928 AtomicStubMark mark_fetch_add_8_relaxed
9929 (_masm, &aarch64_atomic_fetch_add_8_relaxed_impl);
9930 gen_ldadd_entry(MacroAssembler::xword, memory_order_relaxed);
9931
9932 // XCHG, memory_order_conservative
9933 AtomicStubMark mark_xchg_4(_masm, &aarch64_atomic_xchg_4_impl);
9934 gen_swpal_entry(Assembler::word);
9935 AtomicStubMark mark_xchg_8_impl(_masm, &aarch64_atomic_xchg_8_impl);
9936 gen_swpal_entry(Assembler::xword);
9937
9938 // CAS, memory_order_conservative
9939 AtomicStubMark mark_cmpxchg_1(_masm, &aarch64_atomic_cmpxchg_1_impl);
9940 gen_cas_entry(MacroAssembler::byte, memory_order_conservative);
9941 AtomicStubMark mark_cmpxchg_4(_masm, &aarch64_atomic_cmpxchg_4_impl);
9942 gen_cas_entry(MacroAssembler::word, memory_order_conservative);
9943 AtomicStubMark mark_cmpxchg_8(_masm, &aarch64_atomic_cmpxchg_8_impl);
9944 gen_cas_entry(MacroAssembler::xword, memory_order_conservative);
9945
9946 // CAS, memory_order_relaxed
9947 AtomicStubMark mark_cmpxchg_1_relaxed
9948 (_masm, &aarch64_atomic_cmpxchg_1_relaxed_impl);
9949 gen_cas_entry(MacroAssembler::byte, memory_order_relaxed);
9950 AtomicStubMark mark_cmpxchg_4_relaxed
9951 (_masm, &aarch64_atomic_cmpxchg_4_relaxed_impl);
9952 gen_cas_entry(MacroAssembler::word, memory_order_relaxed);
9953 AtomicStubMark mark_cmpxchg_8_relaxed
9954 (_masm, &aarch64_atomic_cmpxchg_8_relaxed_impl);
9955 gen_cas_entry(MacroAssembler::xword, memory_order_relaxed);
9956
9957 AtomicStubMark mark_cmpxchg_4_release
9958 (_masm, &aarch64_atomic_cmpxchg_4_release_impl);
9959 gen_cas_entry(MacroAssembler::word, memory_order_release);
9960 AtomicStubMark mark_cmpxchg_8_release
9961 (_masm, &aarch64_atomic_cmpxchg_8_release_impl);
9962 gen_cas_entry(MacroAssembler::xword, memory_order_release);
9963
9964 AtomicStubMark mark_cmpxchg_4_seq_cst
9965 (_masm, &aarch64_atomic_cmpxchg_4_seq_cst_impl);
9966 gen_cas_entry(MacroAssembler::word, memory_order_seq_cst);
9967 AtomicStubMark mark_cmpxchg_8_seq_cst
9968 (_masm, &aarch64_atomic_cmpxchg_8_seq_cst_impl);
9969 gen_cas_entry(MacroAssembler::xword, memory_order_seq_cst);
9970
9971 ICache::invalidate_range(first_entry, __ pc() - first_entry);
9972 }
9973 #endif // LINUX
9974
9975 address generate_cont_thaw(Continuation::thaw_kind kind) {
9976 bool return_barrier = Continuation::is_thaw_return_barrier(kind);
9977 bool return_barrier_exception = Continuation::is_thaw_return_barrier_exception(kind);
9978
9979 address start = __ pc();
9980
9981 if (return_barrier) {
9982 __ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset()));
9983 __ mov(sp, rscratch1);
9984 }
9985 assert_asm(_masm, (__ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset())), __ cmp(sp, rscratch1)), Assembler::EQ, "incorrect sp");
9986
9987 if (return_barrier) {
9988 // preserve possible return value from a method returning to the return barrier
9989 __ fmovd(rscratch1, v0);
9990 __ stp(rscratch1, r0, Address(__ pre(sp, -2 * wordSize)));
9991 }
9992
9993 __ movw(c_rarg1, (return_barrier ? 1 : 0));
9994 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Continuation::prepare_thaw), rthread, c_rarg1);
9995 __ mov(rscratch2, r0); // r0 contains the size of the frames to thaw, 0 if overflow or no more frames
9996
9997 if (return_barrier) {
9998 // restore return value (no safepoint in the call to thaw, so even an oop return value should be OK)
9999 __ ldp(rscratch1, r0, Address(__ post(sp, 2 * wordSize)));
10000 __ fmovd(v0, rscratch1);
10001 }
10002 assert_asm(_masm, (__ ldr(rscratch1, Address(rthread, JavaThread::cont_entry_offset())), __ cmp(sp, rscratch1)), Assembler::EQ, "incorrect sp");
10003
10004
10005 Label thaw_success;
10006 // rscratch2 contains the size of the frames to thaw, 0 if overflow or no more frames
10007 __ cbnz(rscratch2, thaw_success);
10008 __ lea(rscratch1, RuntimeAddress(SharedRuntime::throw_StackOverflowError_entry()));
10009 __ br(rscratch1);
10010 __ bind(thaw_success);
10011
10012 // make room for the thawed frames
10013 __ sub(rscratch1, sp, rscratch2);
10014 __ andr(rscratch1, rscratch1, -16); // align
10015 __ mov(sp, rscratch1);
10016
10017 if (return_barrier) {
10018 // save original return value -- again
10019 __ fmovd(rscratch1, v0);
10020 __ stp(rscratch1, r0, Address(__ pre(sp, -2 * wordSize)));
10021 }
10022
10023 // If we want, we can templatize thaw by kind, and have three different entries
10024 __ movw(c_rarg1, (uint32_t)kind);
10025
10026 __ call_VM_leaf(Continuation::thaw_entry(), rthread, c_rarg1);
10027 __ mov(rscratch2, r0); // r0 is the sp of the yielding frame
10028
10029 if (return_barrier) {
10030 // restore return value (no safepoint in the call to thaw, so even an oop return value should be OK)
10031 __ ldp(rscratch1, r0, Address(__ post(sp, 2 * wordSize)));
10032 __ fmovd(v0, rscratch1);
10033 } else {
10034 __ mov(r0, zr); // return 0 (success) from doYield
10035 }
10036
10037 // we're now on the yield frame (which is in an address above us b/c rsp has been pushed down)
10038 __ sub(sp, rscratch2, 2*wordSize); // now pointing to rfp spill
10039 __ mov(rfp, sp);
10040
10041 if (return_barrier_exception) {
10042 __ ldr(c_rarg1, Address(rfp, wordSize)); // return address
10043 __ authenticate_return_address(c_rarg1);
10044 __ verify_oop(r0);
10045 // save return value containing the exception oop in callee-saved R19
10046 __ mov(r19, r0);
10047
10048 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), rthread, c_rarg1);
10049
10050 // Reinitialize the ptrue predicate register, in case the external runtime call clobbers ptrue reg, as we may return to SVE compiled code.
10051 // __ reinitialize_ptrue();
10052
10053 // see OptoRuntime::generate_exception_blob: r0 -- exception oop, r3 -- exception pc
10054
10055 __ mov(r1, r0); // the exception handler
10056 __ mov(r0, r19); // restore return value containing the exception oop
10057 __ verify_oop(r0);
10058
10059 __ leave();
10060 __ mov(r3, lr);
10061 __ br(r1); // the exception handler
10062 } else {
10063 // We're "returning" into the topmost thawed frame; see Thaw::push_return_frame
10064 __ leave();
10065 __ ret(lr);
10066 }
10067
10068 return start;
10069 }
10070
10071 address generate_cont_thaw() {
10072 if (!Continuations::enabled()) return nullptr;
10073
10074 StubGenStubId stub_id = StubGenStubId::cont_thaw_id;
10075 StubCodeMark mark(this, stub_id);
10076 address start = __ pc();
10077 generate_cont_thaw(Continuation::thaw_top);
10078 return start;
10079 }
10080
10081 address generate_cont_returnBarrier() {
10082 if (!Continuations::enabled()) return nullptr;
10083
10084 // TODO: will probably need multiple return barriers depending on return type
10085 StubGenStubId stub_id = StubGenStubId::cont_returnBarrier_id;
10086 StubCodeMark mark(this, stub_id);
10087 address start = __ pc();
10088
10089 generate_cont_thaw(Continuation::thaw_return_barrier);
10090
10091 return start;
10092 }
10093
10094 address generate_cont_returnBarrier_exception() {
10095 if (!Continuations::enabled()) return nullptr;
10096
10097 StubGenStubId stub_id = StubGenStubId::cont_returnBarrierExc_id;
10098 StubCodeMark mark(this, stub_id);
10099 address start = __ pc();
10100
10101 generate_cont_thaw(Continuation::thaw_return_barrier_exception);
10102
10103 return start;
10104 }
10105
10106 address generate_cont_preempt_stub() {
10107 if (!Continuations::enabled()) return nullptr;
10108 StubGenStubId stub_id = StubGenStubId::cont_preempt_id;
10109 StubCodeMark mark(this, stub_id);
10110 address start = __ pc();
10111
10112 __ reset_last_Java_frame(true);
10113
10114 // Set sp to enterSpecial frame, i.e. remove all frames copied into the heap.
10115 __ ldr(rscratch2, Address(rthread, JavaThread::cont_entry_offset()));
10116 __ mov(sp, rscratch2);
10117
10118 Label preemption_cancelled;
10119 __ ldrb(rscratch1, Address(rthread, JavaThread::preemption_cancelled_offset()));
10120 __ cbnz(rscratch1, preemption_cancelled);
10121
10122 // Remove enterSpecial frame from the stack and return to Continuation.run() to unmount.
10123 SharedRuntime::continuation_enter_cleanup(_masm);
10124 __ leave();
10125 __ ret(lr);
10126
10127 // We acquired the monitor after freezing the frames so call thaw to continue execution.
10128 __ bind(preemption_cancelled);
10129 __ strb(zr, Address(rthread, JavaThread::preemption_cancelled_offset()));
10130 __ lea(rfp, Address(sp, checked_cast<int32_t>(ContinuationEntry::size())));
10131 __ lea(rscratch1, ExternalAddress(ContinuationEntry::thaw_call_pc_address()));
10132 __ ldr(rscratch1, Address(rscratch1));
10133 __ br(rscratch1);
10134
10135 return start;
10136 }
10137
10138 // In sun.security.util.math.intpoly.IntegerPolynomial1305, integers
10139 // are represented as long[5], with BITS_PER_LIMB = 26.
10140 // Pack five 26-bit limbs into three 64-bit registers.
10141 void pack_26(Register dest0, Register dest1, Register dest2, Register src) {
10142 __ ldp(dest0, rscratch1, Address(src, 0)); // 26 bits
10143 __ add(dest0, dest0, rscratch1, Assembler::LSL, 26); // 26 bits
10144 __ ldp(rscratch1, rscratch2, Address(src, 2 * sizeof (jlong)));
10145 __ add(dest0, dest0, rscratch1, Assembler::LSL, 52); // 12 bits
10146
10147 __ add(dest1, zr, rscratch1, Assembler::LSR, 12); // 14 bits
10148 __ add(dest1, dest1, rscratch2, Assembler::LSL, 14); // 26 bits
10149 __ ldr(rscratch1, Address(src, 4 * sizeof (jlong)));
10150 __ add(dest1, dest1, rscratch1, Assembler::LSL, 40); // 24 bits
10151
10152 if (dest2->is_valid()) {
10153 __ add(dest2, zr, rscratch1, Assembler::LSR, 24); // 2 bits
10154 } else {
10155 #ifdef ASSERT
10156 Label OK;
10157 __ cmp(zr, rscratch1, Assembler::LSR, 24); // 2 bits
10158 __ br(__ EQ, OK);
10159 __ stop("high bits of Poly1305 integer should be zero");
10160 __ should_not_reach_here();
10161 __ bind(OK);
10162 #endif
10163 }
10164 }
10165
10166 // As above, but return only a 128-bit integer, packed into two
10167 // 64-bit registers.
10168 void pack_26(Register dest0, Register dest1, Register src) {
10169 pack_26(dest0, dest1, noreg, src);
10170 }
10171
10172 // Multiply and multiply-accumulate unsigned 64-bit registers.
10173 void wide_mul(Register prod_lo, Register prod_hi, Register n, Register m) {
10174 __ mul(prod_lo, n, m);
10175 __ umulh(prod_hi, n, m);
10176 }
10177 void wide_madd(Register sum_lo, Register sum_hi, Register n, Register m) {
10178 wide_mul(rscratch1, rscratch2, n, m);
10179 __ adds(sum_lo, sum_lo, rscratch1);
10180 __ adc(sum_hi, sum_hi, rscratch2);
10181 }
10182
10183 // Poly1305, RFC 7539
10184
10185 // See https://loup-vaillant.fr/tutorials/poly1305-design for a
10186 // description of the tricks used to simplify and accelerate this
10187 // computation.
10188
10189 address generate_poly1305_processBlocks() {
10190 __ align(CodeEntryAlignment);
10191 StubGenStubId stub_id = StubGenStubId::poly1305_processBlocks_id;
10192 StubCodeMark mark(this, stub_id);
10193 address start = __ pc();
10194 Label here;
10195 __ enter();
10196 RegSet callee_saved = RegSet::range(r19, r28);
10197 __ push(callee_saved, sp);
10198
10199 RegSetIterator<Register> regs = (RegSet::range(c_rarg0, r28) - r18_tls - rscratch1 - rscratch2).begin();
10200
10201 // Arguments
10202 const Register input_start = *regs, length = *++regs, acc_start = *++regs, r_start = *++regs;
10203
10204 // R_n is the 128-bit randomly-generated key, packed into two
10205 // registers. The caller passes this key to us as long[5], with
10206 // BITS_PER_LIMB = 26.
10207 const Register R_0 = *++regs, R_1 = *++regs;
10208 pack_26(R_0, R_1, r_start);
10209
10210 // RR_n is (R_n >> 2) * 5
10211 const Register RR_0 = *++regs, RR_1 = *++regs;
10212 __ lsr(RR_0, R_0, 2);
10213 __ add(RR_0, RR_0, RR_0, Assembler::LSL, 2);
10214 __ lsr(RR_1, R_1, 2);
10215 __ add(RR_1, RR_1, RR_1, Assembler::LSL, 2);
10216
10217 // U_n is the current checksum
10218 const Register U_0 = *++regs, U_1 = *++regs, U_2 = *++regs;
10219 pack_26(U_0, U_1, U_2, acc_start);
10220
10221 static constexpr int BLOCK_LENGTH = 16;
10222 Label DONE, LOOP;
10223
10224 __ cmp(length, checked_cast<u1>(BLOCK_LENGTH));
10225 __ br(Assembler::LT, DONE); {
10226 __ bind(LOOP);
10227
10228 // S_n is to be the sum of U_n and the next block of data
10229 const Register S_0 = *++regs, S_1 = *++regs, S_2 = *++regs;
10230 __ ldp(S_0, S_1, __ post(input_start, 2 * wordSize));
10231 __ adds(S_0, U_0, S_0);
10232 __ adcs(S_1, U_1, S_1);
10233 __ adc(S_2, U_2, zr);
10234 __ add(S_2, S_2, 1);
10235
10236 const Register U_0HI = *++regs, U_1HI = *++regs;
10237
10238 // NB: this logic depends on some of the special properties of
10239 // Poly1305 keys. In particular, because we know that the top
10240 // four bits of R_0 and R_1 are zero, we can add together
10241 // partial products without any risk of needing to propagate a
10242 // carry out.
10243 wide_mul(U_0, U_0HI, S_0, R_0); wide_madd(U_0, U_0HI, S_1, RR_1); wide_madd(U_0, U_0HI, S_2, RR_0);
10244 wide_mul(U_1, U_1HI, S_0, R_1); wide_madd(U_1, U_1HI, S_1, R_0); wide_madd(U_1, U_1HI, S_2, RR_1);
10245 __ andr(U_2, R_0, 3);
10246 __ mul(U_2, S_2, U_2);
10247
10248 // Recycle registers S_0, S_1, S_2
10249 regs = (regs.remaining() + S_0 + S_1 + S_2).begin();
10250
10251 // Partial reduction mod 2**130 - 5
10252 __ adds(U_1, U_0HI, U_1);
10253 __ adc(U_2, U_1HI, U_2);
10254 // Sum now in U_2:U_1:U_0.
10255 // Dead: U_0HI, U_1HI.
10256 regs = (regs.remaining() + U_0HI + U_1HI).begin();
10257
10258 // U_2:U_1:U_0 += (U_2 >> 2) * 5 in two steps
10259
10260 // First, U_2:U_1:U_0 += (U_2 >> 2)
10261 __ lsr(rscratch1, U_2, 2);
10262 __ andr(U_2, U_2, (u8)3);
10263 __ adds(U_0, U_0, rscratch1);
10264 __ adcs(U_1, U_1, zr);
10265 __ adc(U_2, U_2, zr);
10266 // Second, U_2:U_1:U_0 += (U_2 >> 2) << 2
10267 __ adds(U_0, U_0, rscratch1, Assembler::LSL, 2);
10268 __ adcs(U_1, U_1, zr);
10269 __ adc(U_2, U_2, zr);
10270
10271 __ sub(length, length, checked_cast<u1>(BLOCK_LENGTH));
10272 __ cmp(length, checked_cast<u1>(BLOCK_LENGTH));
10273 __ br(~ Assembler::LT, LOOP);
10274 }
10275
10276 // Further reduce modulo 2^130 - 5
10277 __ lsr(rscratch1, U_2, 2);
10278 __ add(rscratch1, rscratch1, rscratch1, Assembler::LSL, 2); // rscratch1 = U_2 * 5
10279 __ adds(U_0, U_0, rscratch1); // U_0 += U_2 * 5
10280 __ adcs(U_1, U_1, zr);
10281 __ andr(U_2, U_2, (u1)3);
10282 __ adc(U_2, U_2, zr);
10283
10284 // Unpack the sum into five 26-bit limbs and write to memory.
10285 __ ubfiz(rscratch1, U_0, 0, 26);
10286 __ ubfx(rscratch2, U_0, 26, 26);
10287 __ stp(rscratch1, rscratch2, Address(acc_start));
10288 __ ubfx(rscratch1, U_0, 52, 12);
10289 __ bfi(rscratch1, U_1, 12, 14);
10290 __ ubfx(rscratch2, U_1, 14, 26);
10291 __ stp(rscratch1, rscratch2, Address(acc_start, 2 * sizeof (jlong)));
10292 __ ubfx(rscratch1, U_1, 40, 24);
10293 __ bfi(rscratch1, U_2, 24, 3);
10294 __ str(rscratch1, Address(acc_start, 4 * sizeof (jlong)));
10295
10296 __ bind(DONE);
10297 __ pop(callee_saved, sp);
10298 __ leave();
10299 __ ret(lr);
10300
10301 return start;
10302 }
10303
10304 // exception handler for upcall stubs
10305 address generate_upcall_stub_exception_handler() {
10306 StubGenStubId stub_id = StubGenStubId::upcall_stub_exception_handler_id;
10307 StubCodeMark mark(this, stub_id);
10308 address start = __ pc();
10309
10310 // Native caller has no idea how to handle exceptions,
10311 // so we just crash here. Up to callee to catch exceptions.
10312 __ verify_oop(r0);
10313 __ movptr(rscratch1, CAST_FROM_FN_PTR(uint64_t, UpcallLinker::handle_uncaught_exception));
10314 __ blr(rscratch1);
10315 __ should_not_reach_here();
10316
10317 return start;
10318 }
10319
10320 // load Method* target of MethodHandle
10321 // j_rarg0 = jobject receiver
10322 // rmethod = result
10323 address generate_upcall_stub_load_target() {
10324 StubGenStubId stub_id = StubGenStubId::upcall_stub_load_target_id;
10325 StubCodeMark mark(this, stub_id);
10326 address start = __ pc();
10327
10328 __ resolve_global_jobject(j_rarg0, rscratch1, rscratch2);
10329 // Load target method from receiver
10330 __ load_heap_oop(rmethod, Address(j_rarg0, java_lang_invoke_MethodHandle::form_offset()), rscratch1, rscratch2);
10331 __ load_heap_oop(rmethod, Address(rmethod, java_lang_invoke_LambdaForm::vmentry_offset()), rscratch1, rscratch2);
10332 __ load_heap_oop(rmethod, Address(rmethod, java_lang_invoke_MemberName::method_offset()), rscratch1, rscratch2);
10333 __ access_load_at(T_ADDRESS, IN_HEAP, rmethod,
10334 Address(rmethod, java_lang_invoke_ResolvedMethodName::vmtarget_offset()),
10335 noreg, noreg);
10336 __ str(rmethod, Address(rthread, JavaThread::callee_target_offset())); // just in case callee is deoptimized
10337
10338 __ ret(lr);
10339
10340 return start;
10341 }
10342
10343 #undef __
10344 #define __ masm->
10345
10346 class MontgomeryMultiplyGenerator : public MacroAssembler {
10347
10348 Register Pa_base, Pb_base, Pn_base, Pm_base, inv, Rlen, Ra, Rb, Rm, Rn,
10349 Pa, Pb, Pn, Pm, Rhi_ab, Rlo_ab, Rhi_mn, Rlo_mn, t0, t1, t2, Ri, Rj;
10350
10351 RegSet _toSave;
10352 bool _squaring;
10353
10354 public:
10355 MontgomeryMultiplyGenerator (Assembler *as, bool squaring)
10356 : MacroAssembler(as->code()), _squaring(squaring) {
10357
10358 // Register allocation
10359
10360 RegSetIterator<Register> regs = (RegSet::range(r0, r26) - r18_tls).begin();
10361 Pa_base = *regs; // Argument registers
10362 if (squaring)
10363 Pb_base = Pa_base;
10364 else
10365 Pb_base = *++regs;
10366 Pn_base = *++regs;
10367 Rlen= *++regs;
10368 inv = *++regs;
10369 Pm_base = *++regs;
10370
10371 // Working registers:
10372 Ra = *++regs; // The current digit of a, b, n, and m.
10373 Rb = *++regs;
10374 Rm = *++regs;
10375 Rn = *++regs;
10376
10377 Pa = *++regs; // Pointers to the current/next digit of a, b, n, and m.
10378 Pb = *++regs;
10379 Pm = *++regs;
10380 Pn = *++regs;
10381
10382 t0 = *++regs; // Three registers which form a
10383 t1 = *++regs; // triple-precision accumuator.
10384 t2 = *++regs;
10385
10386 Ri = *++regs; // Inner and outer loop indexes.
10387 Rj = *++regs;
10388
10389 Rhi_ab = *++regs; // Product registers: low and high parts
10390 Rlo_ab = *++regs; // of a*b and m*n.
10391 Rhi_mn = *++regs;
10392 Rlo_mn = *++regs;
10393
10394 // r19 and up are callee-saved.
10395 _toSave = RegSet::range(r19, *regs) + Pm_base;
10396 }
10397
10398 private:
10399 void save_regs() {
10400 push(_toSave, sp);
10401 }
10402
10403 void restore_regs() {
10404 pop(_toSave, sp);
10405 }
10406
10407 template <typename T>
10408 void unroll_2(Register count, T block) {
10409 Label loop, end, odd;
10410 tbnz(count, 0, odd);
10411 cbz(count, end);
10412 align(16);
10413 bind(loop);
10414 (this->*block)();
10415 bind(odd);
10416 (this->*block)();
10417 subs(count, count, 2);
10418 br(Assembler::GT, loop);
10419 bind(end);
10420 }
10421
10422 template <typename T>
10423 void unroll_2(Register count, T block, Register d, Register s, Register tmp) {
10424 Label loop, end, odd;
10425 tbnz(count, 0, odd);
10426 cbz(count, end);
10427 align(16);
10428 bind(loop);
10429 (this->*block)(d, s, tmp);
10430 bind(odd);
10431 (this->*block)(d, s, tmp);
10432 subs(count, count, 2);
10433 br(Assembler::GT, loop);
10434 bind(end);
10435 }
10436
10437 void pre1(RegisterOrConstant i) {
10438 block_comment("pre1");
10439 // Pa = Pa_base;
10440 // Pb = Pb_base + i;
10441 // Pm = Pm_base;
10442 // Pn = Pn_base + i;
10443 // Ra = *Pa;
10444 // Rb = *Pb;
10445 // Rm = *Pm;
10446 // Rn = *Pn;
10447 ldr(Ra, Address(Pa_base));
10448 ldr(Rb, Address(Pb_base, i, Address::uxtw(LogBytesPerWord)));
10449 ldr(Rm, Address(Pm_base));
10450 ldr(Rn, Address(Pn_base, i, Address::uxtw(LogBytesPerWord)));
10451 lea(Pa, Address(Pa_base));
10452 lea(Pb, Address(Pb_base, i, Address::uxtw(LogBytesPerWord)));
10453 lea(Pm, Address(Pm_base));
10454 lea(Pn, Address(Pn_base, i, Address::uxtw(LogBytesPerWord)));
10455
10456 // Zero the m*n result.
10457 mov(Rhi_mn, zr);
10458 mov(Rlo_mn, zr);
10459 }
10460
10461 // The core multiply-accumulate step of a Montgomery
10462 // multiplication. The idea is to schedule operations as a
10463 // pipeline so that instructions with long latencies (loads and
10464 // multiplies) have time to complete before their results are
10465 // used. This most benefits in-order implementations of the
10466 // architecture but out-of-order ones also benefit.
10467 void step() {
10468 block_comment("step");
10469 // MACC(Ra, Rb, t0, t1, t2);
10470 // Ra = *++Pa;
10471 // Rb = *--Pb;
10472 umulh(Rhi_ab, Ra, Rb);
10473 mul(Rlo_ab, Ra, Rb);
10474 ldr(Ra, pre(Pa, wordSize));
10475 ldr(Rb, pre(Pb, -wordSize));
10476 acc(Rhi_mn, Rlo_mn, t0, t1, t2); // The pending m*n from the
10477 // previous iteration.
10478 // MACC(Rm, Rn, t0, t1, t2);
10479 // Rm = *++Pm;
10480 // Rn = *--Pn;
10481 umulh(Rhi_mn, Rm, Rn);
10482 mul(Rlo_mn, Rm, Rn);
10483 ldr(Rm, pre(Pm, wordSize));
10484 ldr(Rn, pre(Pn, -wordSize));
10485 acc(Rhi_ab, Rlo_ab, t0, t1, t2);
10486 }
10487
10488 void post1() {
10489 block_comment("post1");
10490
10491 // MACC(Ra, Rb, t0, t1, t2);
10492 // Ra = *++Pa;
10493 // Rb = *--Pb;
10494 umulh(Rhi_ab, Ra, Rb);
10495 mul(Rlo_ab, Ra, Rb);
10496 acc(Rhi_mn, Rlo_mn, t0, t1, t2); // The pending m*n
10497 acc(Rhi_ab, Rlo_ab, t0, t1, t2);
10498
10499 // *Pm = Rm = t0 * inv;
10500 mul(Rm, t0, inv);
10501 str(Rm, Address(Pm));
10502
10503 // MACC(Rm, Rn, t0, t1, t2);
10504 // t0 = t1; t1 = t2; t2 = 0;
10505 umulh(Rhi_mn, Rm, Rn);
10506
10507 #ifndef PRODUCT
10508 // assert(m[i] * n[0] + t0 == 0, "broken Montgomery multiply");
10509 {
10510 mul(Rlo_mn, Rm, Rn);
10511 add(Rlo_mn, t0, Rlo_mn);
10512 Label ok;
10513 cbz(Rlo_mn, ok); {
10514 stop("broken Montgomery multiply");
10515 } bind(ok);
10516 }
10517 #endif
10518 // We have very carefully set things up so that
10519 // m[i]*n[0] + t0 == 0 (mod b), so we don't have to calculate
10520 // the lower half of Rm * Rn because we know the result already:
10521 // it must be -t0. t0 + (-t0) must generate a carry iff
10522 // t0 != 0. So, rather than do a mul and an adds we just set
10523 // the carry flag iff t0 is nonzero.
10524 //
10525 // mul(Rlo_mn, Rm, Rn);
10526 // adds(zr, t0, Rlo_mn);
10527 subs(zr, t0, 1); // Set carry iff t0 is nonzero
10528 adcs(t0, t1, Rhi_mn);
10529 adc(t1, t2, zr);
10530 mov(t2, zr);
10531 }
10532
10533 void pre2(RegisterOrConstant i, RegisterOrConstant len) {
10534 block_comment("pre2");
10535 // Pa = Pa_base + i-len;
10536 // Pb = Pb_base + len;
10537 // Pm = Pm_base + i-len;
10538 // Pn = Pn_base + len;
10539
10540 if (i.is_register()) {
10541 sub(Rj, i.as_register(), len);
10542 } else {
10543 mov(Rj, i.as_constant());
10544 sub(Rj, Rj, len);
10545 }
10546 // Rj == i-len
10547
10548 lea(Pa, Address(Pa_base, Rj, Address::uxtw(LogBytesPerWord)));
10549 lea(Pb, Address(Pb_base, len, Address::uxtw(LogBytesPerWord)));
10550 lea(Pm, Address(Pm_base, Rj, Address::uxtw(LogBytesPerWord)));
10551 lea(Pn, Address(Pn_base, len, Address::uxtw(LogBytesPerWord)));
10552
10553 // Ra = *++Pa;
10554 // Rb = *--Pb;
10555 // Rm = *++Pm;
10556 // Rn = *--Pn;
10557 ldr(Ra, pre(Pa, wordSize));
10558 ldr(Rb, pre(Pb, -wordSize));
10559 ldr(Rm, pre(Pm, wordSize));
10560 ldr(Rn, pre(Pn, -wordSize));
10561
10562 mov(Rhi_mn, zr);
10563 mov(Rlo_mn, zr);
10564 }
10565
10566 void post2(RegisterOrConstant i, RegisterOrConstant len) {
10567 block_comment("post2");
10568 if (i.is_constant()) {
10569 mov(Rj, i.as_constant()-len.as_constant());
10570 } else {
10571 sub(Rj, i.as_register(), len);
10572 }
10573
10574 adds(t0, t0, Rlo_mn); // The pending m*n, low part
10575
10576 // As soon as we know the least significant digit of our result,
10577 // store it.
10578 // Pm_base[i-len] = t0;
10579 str(t0, Address(Pm_base, Rj, Address::uxtw(LogBytesPerWord)));
10580
10581 // t0 = t1; t1 = t2; t2 = 0;
10582 adcs(t0, t1, Rhi_mn); // The pending m*n, high part
10583 adc(t1, t2, zr);
10584 mov(t2, zr);
10585 }
10586
10587 // A carry in t0 after Montgomery multiplication means that we
10588 // should subtract multiples of n from our result in m. We'll
10589 // keep doing that until there is no carry.
10590 void normalize(RegisterOrConstant len) {
10591 block_comment("normalize");
10592 // while (t0)
10593 // t0 = sub(Pm_base, Pn_base, t0, len);
10594 Label loop, post, again;
10595 Register cnt = t1, i = t2; // Re-use registers; we're done with them now
10596 cbz(t0, post); {
10597 bind(again); {
10598 mov(i, zr);
10599 mov(cnt, len);
10600 ldr(Rm, Address(Pm_base, i, Address::uxtw(LogBytesPerWord)));
10601 ldr(Rn, Address(Pn_base, i, Address::uxtw(LogBytesPerWord)));
10602 subs(zr, zr, zr); // set carry flag, i.e. no borrow
10603 align(16);
10604 bind(loop); {
10605 sbcs(Rm, Rm, Rn);
10606 str(Rm, Address(Pm_base, i, Address::uxtw(LogBytesPerWord)));
10607 add(i, i, 1);
10608 ldr(Rm, Address(Pm_base, i, Address::uxtw(LogBytesPerWord)));
10609 ldr(Rn, Address(Pn_base, i, Address::uxtw(LogBytesPerWord)));
10610 sub(cnt, cnt, 1);
10611 } cbnz(cnt, loop);
10612 sbc(t0, t0, zr);
10613 } cbnz(t0, again);
10614 } bind(post);
10615 }
10616
10617 // Move memory at s to d, reversing words.
10618 // Increments d to end of copied memory
10619 // Destroys tmp1, tmp2
10620 // Preserves len
10621 // Leaves s pointing to the address which was in d at start
10622 void reverse(Register d, Register s, Register len, Register tmp1, Register tmp2) {
10623 assert(tmp1->encoding() < r19->encoding(), "register corruption");
10624 assert(tmp2->encoding() < r19->encoding(), "register corruption");
10625
10626 lea(s, Address(s, len, Address::uxtw(LogBytesPerWord)));
10627 mov(tmp1, len);
10628 unroll_2(tmp1, &MontgomeryMultiplyGenerator::reverse1, d, s, tmp2);
10629 sub(s, d, len, ext::uxtw, LogBytesPerWord);
10630 }
10631 // where
10632 void reverse1(Register d, Register s, Register tmp) {
10633 ldr(tmp, pre(s, -wordSize));
10634 ror(tmp, tmp, 32);
10635 str(tmp, post(d, wordSize));
10636 }
10637
10638 void step_squaring() {
10639 // An extra ACC
10640 step();
10641 acc(Rhi_ab, Rlo_ab, t0, t1, t2);
10642 }
10643
10644 void last_squaring(RegisterOrConstant i) {
10645 Label dont;
10646 // if ((i & 1) == 0) {
10647 tbnz(i.as_register(), 0, dont); {
10648 // MACC(Ra, Rb, t0, t1, t2);
10649 // Ra = *++Pa;
10650 // Rb = *--Pb;
10651 umulh(Rhi_ab, Ra, Rb);
10652 mul(Rlo_ab, Ra, Rb);
10653 acc(Rhi_ab, Rlo_ab, t0, t1, t2);
10654 } bind(dont);
10655 }
10656
10657 void extra_step_squaring() {
10658 acc(Rhi_mn, Rlo_mn, t0, t1, t2); // The pending m*n
10659
10660 // MACC(Rm, Rn, t0, t1, t2);
10661 // Rm = *++Pm;
10662 // Rn = *--Pn;
10663 umulh(Rhi_mn, Rm, Rn);
10664 mul(Rlo_mn, Rm, Rn);
10665 ldr(Rm, pre(Pm, wordSize));
10666 ldr(Rn, pre(Pn, -wordSize));
10667 }
10668
10669 void post1_squaring() {
10670 acc(Rhi_mn, Rlo_mn, t0, t1, t2); // The pending m*n
10671
10672 // *Pm = Rm = t0 * inv;
10673 mul(Rm, t0, inv);
10674 str(Rm, Address(Pm));
10675
10676 // MACC(Rm, Rn, t0, t1, t2);
10677 // t0 = t1; t1 = t2; t2 = 0;
10678 umulh(Rhi_mn, Rm, Rn);
10679
10680 #ifndef PRODUCT
10681 // assert(m[i] * n[0] + t0 == 0, "broken Montgomery multiply");
10682 {
10683 mul(Rlo_mn, Rm, Rn);
10684 add(Rlo_mn, t0, Rlo_mn);
10685 Label ok;
10686 cbz(Rlo_mn, ok); {
10687 stop("broken Montgomery multiply");
10688 } bind(ok);
10689 }
10690 #endif
10691 // We have very carefully set things up so that
10692 // m[i]*n[0] + t0 == 0 (mod b), so we don't have to calculate
10693 // the lower half of Rm * Rn because we know the result already:
10694 // it must be -t0. t0 + (-t0) must generate a carry iff
10695 // t0 != 0. So, rather than do a mul and an adds we just set
10696 // the carry flag iff t0 is nonzero.
10697 //
10698 // mul(Rlo_mn, Rm, Rn);
10699 // adds(zr, t0, Rlo_mn);
10700 subs(zr, t0, 1); // Set carry iff t0 is nonzero
10701 adcs(t0, t1, Rhi_mn);
10702 adc(t1, t2, zr);
10703 mov(t2, zr);
10704 }
10705
10706 void acc(Register Rhi, Register Rlo,
10707 Register t0, Register t1, Register t2) {
10708 adds(t0, t0, Rlo);
10709 adcs(t1, t1, Rhi);
10710 adc(t2, t2, zr);
10711 }
10712
10713 public:
10714 /**
10715 * Fast Montgomery multiplication. The derivation of the
10716 * algorithm is in A Cryptographic Library for the Motorola
10717 * DSP56000, Dusse and Kaliski, Proc. EUROCRYPT 90, pp. 230-237.
10718 *
10719 * Arguments:
10720 *
10721 * Inputs for multiplication:
10722 * c_rarg0 - int array elements a
10723 * c_rarg1 - int array elements b
10724 * c_rarg2 - int array elements n (the modulus)
10725 * c_rarg3 - int length
10726 * c_rarg4 - int inv
10727 * c_rarg5 - int array elements m (the result)
10728 *
10729 * Inputs for squaring:
10730 * c_rarg0 - int array elements a
10731 * c_rarg1 - int array elements n (the modulus)
10732 * c_rarg2 - int length
10733 * c_rarg3 - int inv
10734 * c_rarg4 - int array elements m (the result)
10735 *
10736 */
10737 address generate_multiply() {
10738 Label argh, nothing;
10739 bind(argh);
10740 stop("MontgomeryMultiply total_allocation must be <= 8192");
10741
10742 align(CodeEntryAlignment);
10743 address entry = pc();
10744
10745 cbzw(Rlen, nothing);
10746
10747 enter();
10748
10749 // Make room.
10750 cmpw(Rlen, 512);
10751 br(Assembler::HI, argh);
10752 sub(Ra, sp, Rlen, ext::uxtw, exact_log2(4 * sizeof (jint)));
10753 andr(sp, Ra, -2 * wordSize);
10754
10755 lsrw(Rlen, Rlen, 1); // length in longwords = len/2
10756
10757 {
10758 // Copy input args, reversing as we go. We use Ra as a
10759 // temporary variable.
10760 reverse(Ra, Pa_base, Rlen, t0, t1);
10761 if (!_squaring)
10762 reverse(Ra, Pb_base, Rlen, t0, t1);
10763 reverse(Ra, Pn_base, Rlen, t0, t1);
10764 }
10765
10766 // Push all call-saved registers and also Pm_base which we'll need
10767 // at the end.
10768 save_regs();
10769
10770 #ifndef PRODUCT
10771 // assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply");
10772 {
10773 ldr(Rn, Address(Pn_base, 0));
10774 mul(Rlo_mn, Rn, inv);
10775 subs(zr, Rlo_mn, -1);
10776 Label ok;
10777 br(EQ, ok); {
10778 stop("broken inverse in Montgomery multiply");
10779 } bind(ok);
10780 }
10781 #endif
10782
10783 mov(Pm_base, Ra);
10784
10785 mov(t0, zr);
10786 mov(t1, zr);
10787 mov(t2, zr);
10788
10789 block_comment("for (int i = 0; i < len; i++) {");
10790 mov(Ri, zr); {
10791 Label loop, end;
10792 cmpw(Ri, Rlen);
10793 br(Assembler::GE, end);
10794
10795 bind(loop);
10796 pre1(Ri);
10797
10798 block_comment(" for (j = i; j; j--) {"); {
10799 movw(Rj, Ri);
10800 unroll_2(Rj, &MontgomeryMultiplyGenerator::step);
10801 } block_comment(" } // j");
10802
10803 post1();
10804 addw(Ri, Ri, 1);
10805 cmpw(Ri, Rlen);
10806 br(Assembler::LT, loop);
10807 bind(end);
10808 block_comment("} // i");
10809 }
10810
10811 block_comment("for (int i = len; i < 2*len; i++) {");
10812 mov(Ri, Rlen); {
10813 Label loop, end;
10814 cmpw(Ri, Rlen, Assembler::LSL, 1);
10815 br(Assembler::GE, end);
10816
10817 bind(loop);
10818 pre2(Ri, Rlen);
10819
10820 block_comment(" for (j = len*2-i-1; j; j--) {"); {
10821 lslw(Rj, Rlen, 1);
10822 subw(Rj, Rj, Ri);
10823 subw(Rj, Rj, 1);
10824 unroll_2(Rj, &MontgomeryMultiplyGenerator::step);
10825 } block_comment(" } // j");
10826
10827 post2(Ri, Rlen);
10828 addw(Ri, Ri, 1);
10829 cmpw(Ri, Rlen, Assembler::LSL, 1);
10830 br(Assembler::LT, loop);
10831 bind(end);
10832 }
10833 block_comment("} // i");
10834
10835 normalize(Rlen);
10836
10837 mov(Ra, Pm_base); // Save Pm_base in Ra
10838 restore_regs(); // Restore caller's Pm_base
10839
10840 // Copy our result into caller's Pm_base
10841 reverse(Pm_base, Ra, Rlen, t0, t1);
10842
10843 leave();
10844 bind(nothing);
10845 ret(lr);
10846
10847 return entry;
10848 }
10849 // In C, approximately:
10850
10851 // void
10852 // montgomery_multiply(julong Pa_base[], julong Pb_base[],
10853 // julong Pn_base[], julong Pm_base[],
10854 // julong inv, int len) {
10855 // julong t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator
10856 // julong *Pa, *Pb, *Pn, *Pm;
10857 // julong Ra, Rb, Rn, Rm;
10858
10859 // int i;
10860
10861 // assert(inv * Pn_base[0] == -1UL, "broken inverse in Montgomery multiply");
10862
10863 // for (i = 0; i < len; i++) {
10864 // int j;
10865
10866 // Pa = Pa_base;
10867 // Pb = Pb_base + i;
10868 // Pm = Pm_base;
10869 // Pn = Pn_base + i;
10870
10871 // Ra = *Pa;
10872 // Rb = *Pb;
10873 // Rm = *Pm;
10874 // Rn = *Pn;
10875
10876 // int iters = i;
10877 // for (j = 0; iters--; j++) {
10878 // assert(Ra == Pa_base[j] && Rb == Pb_base[i-j], "must be");
10879 // MACC(Ra, Rb, t0, t1, t2);
10880 // Ra = *++Pa;
10881 // Rb = *--Pb;
10882 // assert(Rm == Pm_base[j] && Rn == Pn_base[i-j], "must be");
10883 // MACC(Rm, Rn, t0, t1, t2);
10884 // Rm = *++Pm;
10885 // Rn = *--Pn;
10886 // }
10887
10888 // assert(Ra == Pa_base[i] && Rb == Pb_base[0], "must be");
10889 // MACC(Ra, Rb, t0, t1, t2);
10890 // *Pm = Rm = t0 * inv;
10891 // assert(Rm == Pm_base[i] && Rn == Pn_base[0], "must be");
10892 // MACC(Rm, Rn, t0, t1, t2);
10893
10894 // assert(t0 == 0, "broken Montgomery multiply");
10895
10896 // t0 = t1; t1 = t2; t2 = 0;
10897 // }
10898
10899 // for (i = len; i < 2*len; i++) {
10900 // int j;
10901
10902 // Pa = Pa_base + i-len;
10903 // Pb = Pb_base + len;
10904 // Pm = Pm_base + i-len;
10905 // Pn = Pn_base + len;
10906
10907 // Ra = *++Pa;
10908 // Rb = *--Pb;
10909 // Rm = *++Pm;
10910 // Rn = *--Pn;
10911
10912 // int iters = len*2-i-1;
10913 // for (j = i-len+1; iters--; j++) {
10914 // assert(Ra == Pa_base[j] && Rb == Pb_base[i-j], "must be");
10915 // MACC(Ra, Rb, t0, t1, t2);
10916 // Ra = *++Pa;
10917 // Rb = *--Pb;
10918 // assert(Rm == Pm_base[j] && Rn == Pn_base[i-j], "must be");
10919 // MACC(Rm, Rn, t0, t1, t2);
10920 // Rm = *++Pm;
10921 // Rn = *--Pn;
10922 // }
10923
10924 // Pm_base[i-len] = t0;
10925 // t0 = t1; t1 = t2; t2 = 0;
10926 // }
10927
10928 // while (t0)
10929 // t0 = sub(Pm_base, Pn_base, t0, len);
10930 // }
10931
10932 /**
10933 * Fast Montgomery squaring. This uses asymptotically 25% fewer
10934 * multiplies than Montgomery multiplication so it should be up to
10935 * 25% faster. However, its loop control is more complex and it
10936 * may actually run slower on some machines.
10937 *
10938 * Arguments:
10939 *
10940 * Inputs:
10941 * c_rarg0 - int array elements a
10942 * c_rarg1 - int array elements n (the modulus)
10943 * c_rarg2 - int length
10944 * c_rarg3 - int inv
10945 * c_rarg4 - int array elements m (the result)
10946 *
10947 */
10948 address generate_square() {
10949 Label argh;
10950 bind(argh);
10951 stop("MontgomeryMultiply total_allocation must be <= 8192");
10952
10953 align(CodeEntryAlignment);
10954 address entry = pc();
10955
10956 enter();
10957
10958 // Make room.
10959 cmpw(Rlen, 512);
10960 br(Assembler::HI, argh);
10961 sub(Ra, sp, Rlen, ext::uxtw, exact_log2(4 * sizeof (jint)));
10962 andr(sp, Ra, -2 * wordSize);
10963
10964 lsrw(Rlen, Rlen, 1); // length in longwords = len/2
10965
10966 {
10967 // Copy input args, reversing as we go. We use Ra as a
10968 // temporary variable.
10969 reverse(Ra, Pa_base, Rlen, t0, t1);
10970 reverse(Ra, Pn_base, Rlen, t0, t1);
10971 }
10972
10973 // Push all call-saved registers and also Pm_base which we'll need
10974 // at the end.
10975 save_regs();
10976
10977 mov(Pm_base, Ra);
10978
10979 mov(t0, zr);
10980 mov(t1, zr);
10981 mov(t2, zr);
10982
10983 block_comment("for (int i = 0; i < len; i++) {");
10984 mov(Ri, zr); {
10985 Label loop, end;
10986 bind(loop);
10987 cmp(Ri, Rlen);
10988 br(Assembler::GE, end);
10989
10990 pre1(Ri);
10991
10992 block_comment("for (j = (i+1)/2; j; j--) {"); {
10993 add(Rj, Ri, 1);
10994 lsr(Rj, Rj, 1);
10995 unroll_2(Rj, &MontgomeryMultiplyGenerator::step_squaring);
10996 } block_comment(" } // j");
10997
10998 last_squaring(Ri);
10999
11000 block_comment(" for (j = i/2; j; j--) {"); {
11001 lsr(Rj, Ri, 1);
11002 unroll_2(Rj, &MontgomeryMultiplyGenerator::extra_step_squaring);
11003 } block_comment(" } // j");
11004
11005 post1_squaring();
11006 add(Ri, Ri, 1);
11007 cmp(Ri, Rlen);
11008 br(Assembler::LT, loop);
11009
11010 bind(end);
11011 block_comment("} // i");
11012 }
11013
11014 block_comment("for (int i = len; i < 2*len; i++) {");
11015 mov(Ri, Rlen); {
11016 Label loop, end;
11017 bind(loop);
11018 cmp(Ri, Rlen, Assembler::LSL, 1);
11019 br(Assembler::GE, end);
11020
11021 pre2(Ri, Rlen);
11022
11023 block_comment(" for (j = (2*len-i-1)/2; j; j--) {"); {
11024 lsl(Rj, Rlen, 1);
11025 sub(Rj, Rj, Ri);
11026 sub(Rj, Rj, 1);
11027 lsr(Rj, Rj, 1);
11028 unroll_2(Rj, &MontgomeryMultiplyGenerator::step_squaring);
11029 } block_comment(" } // j");
11030
11031 last_squaring(Ri);
11032
11033 block_comment(" for (j = (2*len-i)/2; j; j--) {"); {
11034 lsl(Rj, Rlen, 1);
11035 sub(Rj, Rj, Ri);
11036 lsr(Rj, Rj, 1);
11037 unroll_2(Rj, &MontgomeryMultiplyGenerator::extra_step_squaring);
11038 } block_comment(" } // j");
11039
11040 post2(Ri, Rlen);
11041 add(Ri, Ri, 1);
11042 cmp(Ri, Rlen, Assembler::LSL, 1);
11043
11044 br(Assembler::LT, loop);
11045 bind(end);
11046 block_comment("} // i");
11047 }
11048
11049 normalize(Rlen);
11050
11051 mov(Ra, Pm_base); // Save Pm_base in Ra
11052 restore_regs(); // Restore caller's Pm_base
11053
11054 // Copy our result into caller's Pm_base
11055 reverse(Pm_base, Ra, Rlen, t0, t1);
11056
11057 leave();
11058 ret(lr);
11059
11060 return entry;
11061 }
11062 // In C, approximately:
11063
11064 // void
11065 // montgomery_square(julong Pa_base[], julong Pn_base[],
11066 // julong Pm_base[], julong inv, int len) {
11067 // julong t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator
11068 // julong *Pa, *Pb, *Pn, *Pm;
11069 // julong Ra, Rb, Rn, Rm;
11070
11071 // int i;
11072
11073 // assert(inv * Pn_base[0] == -1UL, "broken inverse in Montgomery multiply");
11074
11075 // for (i = 0; i < len; i++) {
11076 // int j;
11077
11078 // Pa = Pa_base;
11079 // Pb = Pa_base + i;
11080 // Pm = Pm_base;
11081 // Pn = Pn_base + i;
11082
11083 // Ra = *Pa;
11084 // Rb = *Pb;
11085 // Rm = *Pm;
11086 // Rn = *Pn;
11087
11088 // int iters = (i+1)/2;
11089 // for (j = 0; iters--; j++) {
11090 // assert(Ra == Pa_base[j] && Rb == Pa_base[i-j], "must be");
11091 // MACC2(Ra, Rb, t0, t1, t2);
11092 // Ra = *++Pa;
11093 // Rb = *--Pb;
11094 // assert(Rm == Pm_base[j] && Rn == Pn_base[i-j], "must be");
11095 // MACC(Rm, Rn, t0, t1, t2);
11096 // Rm = *++Pm;
11097 // Rn = *--Pn;
11098 // }
11099 // if ((i & 1) == 0) {
11100 // assert(Ra == Pa_base[j], "must be");
11101 // MACC(Ra, Ra, t0, t1, t2);
11102 // }
11103 // iters = i/2;
11104 // assert(iters == i-j, "must be");
11105 // for (; iters--; j++) {
11106 // assert(Rm == Pm_base[j] && Rn == Pn_base[i-j], "must be");
11107 // MACC(Rm, Rn, t0, t1, t2);
11108 // Rm = *++Pm;
11109 // Rn = *--Pn;
11110 // }
11111
11112 // *Pm = Rm = t0 * inv;
11113 // assert(Rm == Pm_base[i] && Rn == Pn_base[0], "must be");
11114 // MACC(Rm, Rn, t0, t1, t2);
11115
11116 // assert(t0 == 0, "broken Montgomery multiply");
11117
11118 // t0 = t1; t1 = t2; t2 = 0;
11119 // }
11120
11121 // for (i = len; i < 2*len; i++) {
11122 // int start = i-len+1;
11123 // int end = start + (len - start)/2;
11124 // int j;
11125
11126 // Pa = Pa_base + i-len;
11127 // Pb = Pa_base + len;
11128 // Pm = Pm_base + i-len;
11129 // Pn = Pn_base + len;
11130
11131 // Ra = *++Pa;
11132 // Rb = *--Pb;
11133 // Rm = *++Pm;
11134 // Rn = *--Pn;
11135
11136 // int iters = (2*len-i-1)/2;
11137 // assert(iters == end-start, "must be");
11138 // for (j = start; iters--; j++) {
11139 // assert(Ra == Pa_base[j] && Rb == Pa_base[i-j], "must be");
11140 // MACC2(Ra, Rb, t0, t1, t2);
11141 // Ra = *++Pa;
11142 // Rb = *--Pb;
11143 // assert(Rm == Pm_base[j] && Rn == Pn_base[i-j], "must be");
11144 // MACC(Rm, Rn, t0, t1, t2);
11145 // Rm = *++Pm;
11146 // Rn = *--Pn;
11147 // }
11148 // if ((i & 1) == 0) {
11149 // assert(Ra == Pa_base[j], "must be");
11150 // MACC(Ra, Ra, t0, t1, t2);
11151 // }
11152 // iters = (2*len-i)/2;
11153 // assert(iters == len-j, "must be");
11154 // for (; iters--; j++) {
11155 // assert(Rm == Pm_base[j] && Rn == Pn_base[i-j], "must be");
11156 // MACC(Rm, Rn, t0, t1, t2);
11157 // Rm = *++Pm;
11158 // Rn = *--Pn;
11159 // }
11160 // Pm_base[i-len] = t0;
11161 // t0 = t1; t1 = t2; t2 = 0;
11162 // }
11163
11164 // while (t0)
11165 // t0 = sub(Pm_base, Pn_base, t0, len);
11166 // }
11167 };
11168
11169 void generate_vector_math_stubs() {
11170 // Get native vector math stub routine addresses
11171 void* libsleef = nullptr;
11172 char ebuf[1024];
11173 char dll_name[JVM_MAXPATHLEN];
11174 if (os::dll_locate_lib(dll_name, sizeof(dll_name), Arguments::get_dll_dir(), "sleef")) {
11175 libsleef = os::dll_load(dll_name, ebuf, sizeof ebuf);
11176 }
11177 if (libsleef == nullptr) {
11178 log_info(library)("Failed to load native vector math library, %s!", ebuf);
11179 return;
11180 }
11181 // Method naming convention
11182 // All the methods are named as <OP><T><N>_<U><suffix>
11183 // Where:
11184 // <OP> is the operation name, e.g. sin
11185 // <T> is optional to indicate float/double
11186 // "f/d" for vector float/double operation
11187 // <N> is the number of elements in the vector
11188 // "2/4" for neon, and "x" for sve
11189 // <U> is the precision level
11190 // "u10/u05" represents 1.0/0.5 ULP error bounds
11191 // We use "u10" for all operations by default
11192 // But for those functions do not have u10 support, we use "u05" instead
11193 // <suffix> indicates neon/sve
11194 // "sve/advsimd" for sve/neon implementations
11195 // e.g. sinfx_u10sve is the method for computing vector float sin using SVE instructions
11196 // cosd2_u10advsimd is the method for computing 2 elements vector double cos using NEON instructions
11197 //
11198 log_info(library)("Loaded library %s, handle " INTPTR_FORMAT, JNI_LIB_PREFIX "sleef" JNI_LIB_SUFFIX, p2i(libsleef));
11199
11200 // Math vector stubs implemented with SVE for scalable vector size.
11201 if (UseSVE > 0) {
11202 for (int op = 0; op < VectorSupport::NUM_VECTOR_OP_MATH; op++) {
11203 int vop = VectorSupport::VECTOR_OP_MATH_START + op;
11204 // Skip "tanh" because there is performance regression
11205 if (vop == VectorSupport::VECTOR_OP_TANH) {
11206 continue;
11207 }
11208
11209 // The native library does not support u10 level of "hypot".
11210 const char* ulf = (vop == VectorSupport::VECTOR_OP_HYPOT) ? "u05" : "u10";
11211
11212 snprintf(ebuf, sizeof(ebuf), "%sfx_%ssve", VectorSupport::mathname[op], ulf);
11213 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_SCALABLE][op] = (address)os::dll_lookup(libsleef, ebuf);
11214
11215 snprintf(ebuf, sizeof(ebuf), "%sdx_%ssve", VectorSupport::mathname[op], ulf);
11216 StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_SCALABLE][op] = (address)os::dll_lookup(libsleef, ebuf);
11217 }
11218 }
11219
11220 // Math vector stubs implemented with NEON for 64/128 bits vector size.
11221 for (int op = 0; op < VectorSupport::NUM_VECTOR_OP_MATH; op++) {
11222 int vop = VectorSupport::VECTOR_OP_MATH_START + op;
11223 // Skip "tanh" because there is performance regression
11224 if (vop == VectorSupport::VECTOR_OP_TANH) {
11225 continue;
11226 }
11227
11228 // The native library does not support u10 level of "hypot".
11229 const char* ulf = (vop == VectorSupport::VECTOR_OP_HYPOT) ? "u05" : "u10";
11230
11231 snprintf(ebuf, sizeof(ebuf), "%sf4_%sadvsimd", VectorSupport::mathname[op], ulf);
11232 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_64][op] = (address)os::dll_lookup(libsleef, ebuf);
11233
11234 snprintf(ebuf, sizeof(ebuf), "%sf4_%sadvsimd", VectorSupport::mathname[op], ulf);
11235 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_128][op] = (address)os::dll_lookup(libsleef, ebuf);
11236
11237 snprintf(ebuf, sizeof(ebuf), "%sd2_%sadvsimd", VectorSupport::mathname[op], ulf);
11238 StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_128][op] = (address)os::dll_lookup(libsleef, ebuf);
11239 }
11240 }
11241
11242 // Initialization
11243 void generate_initial_stubs() {
11244 // Generate initial stubs and initializes the entry points
11245
11246 // entry points that exist in all platforms Note: This is code
11247 // that could be shared among different platforms - however the
11248 // benefit seems to be smaller than the disadvantage of having a
11249 // much more complicated generator structure. See also comment in
11250 // stubRoutines.hpp.
11251
11252 StubRoutines::_forward_exception_entry = generate_forward_exception();
11253
11254 StubRoutines::_call_stub_entry =
11255 generate_call_stub(StubRoutines::_call_stub_return_address);
11256
11257 // is referenced by megamorphic call
11258 StubRoutines::_catch_exception_entry = generate_catch_exception();
11259
11260 // Initialize table for copy memory (arraycopy) check.
11261 if (UnsafeMemoryAccess::_table == nullptr) {
11262 UnsafeMemoryAccess::create_table(8 + 4); // 8 for copyMemory; 4 for setMemory
11263 }
11264
11265 if (UseCRC32Intrinsics) {
11266 // set table address before stub generation which use it
11267 StubRoutines::_crc_table_adr = (address)StubRoutines::aarch64::_crc_table;
11268 StubRoutines::_updateBytesCRC32 = generate_updateBytesCRC32();
11269 }
11270
11271 if (UseCRC32CIntrinsics) {
11272 StubRoutines::_updateBytesCRC32C = generate_updateBytesCRC32C();
11273 }
11274
11275 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dsin)) {
11276 StubRoutines::_dsin = generate_dsin_dcos(/* isCos = */ false);
11277 }
11278
11279 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dcos)) {
11280 StubRoutines::_dcos = generate_dsin_dcos(/* isCos = */ true);
11281 }
11282
11283 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_float16ToFloat) &&
11284 vmIntrinsics::is_intrinsic_available(vmIntrinsics::_floatToFloat16)) {
11285 StubRoutines::_hf2f = generate_float16ToFloat();
11286 StubRoutines::_f2hf = generate_floatToFloat16();
11287 }
11288 }
11289
11290 void generate_continuation_stubs() {
11291 // Continuation stubs:
11292 StubRoutines::_cont_thaw = generate_cont_thaw();
11293 StubRoutines::_cont_returnBarrier = generate_cont_returnBarrier();
11294 StubRoutines::_cont_returnBarrierExc = generate_cont_returnBarrier_exception();
11295 StubRoutines::_cont_preempt_stub = generate_cont_preempt_stub();
11296 }
11297
11298 void generate_final_stubs() {
11299 // support for verify_oop (must happen after universe_init)
11300 if (VerifyOops) {
11301 StubRoutines::_verify_oop_subroutine_entry = generate_verify_oop();
11302 }
11303
11304 // arraycopy stubs used by compilers
11305 generate_arraycopy_stubs();
11306
11307 StubRoutines::_method_entry_barrier = generate_method_entry_barrier();
11308
11309 StubRoutines::aarch64::_spin_wait = generate_spin_wait();
11310
11311 StubRoutines::_upcall_stub_exception_handler = generate_upcall_stub_exception_handler();
11312 StubRoutines::_upcall_stub_load_target = generate_upcall_stub_load_target();
11313
11314 #if defined (LINUX) && !defined (__ARM_FEATURE_ATOMICS)
11315
11316 generate_atomic_entry_points();
11317
11318 #endif // LINUX
11319
11320 #ifdef COMPILER2
11321 if (UseSecondarySupersTable) {
11322 StubRoutines::_lookup_secondary_supers_table_slow_path_stub = generate_lookup_secondary_supers_table_slow_path_stub();
11323 if (! InlineSecondarySupersTest) {
11324 generate_lookup_secondary_supers_table_stub();
11325 }
11326 }
11327 #endif
11328
11329 StubRoutines::aarch64::set_completed(); // Inidicate that arraycopy and zero_blocks stubs are generated
11330 }
11331
11332 void generate_compiler_stubs() {
11333 #if COMPILER2_OR_JVMCI
11334
11335 if (UseSVE == 0) {
11336 StubRoutines::aarch64::_vector_iota_indices = generate_iota_indices(StubGenStubId::vector_iota_indices_id);
11337 }
11338
11339 // array equals stub for large arrays.
11340 if (!UseSimpleArrayEquals) {
11341 StubRoutines::aarch64::_large_array_equals = generate_large_array_equals();
11342 }
11343
11344 // arrays_hascode stub for large arrays.
11345 StubRoutines::aarch64::_large_arrays_hashcode_boolean = generate_large_arrays_hashcode(T_BOOLEAN);
11346 StubRoutines::aarch64::_large_arrays_hashcode_byte = generate_large_arrays_hashcode(T_BYTE);
11347 StubRoutines::aarch64::_large_arrays_hashcode_char = generate_large_arrays_hashcode(T_CHAR);
11348 StubRoutines::aarch64::_large_arrays_hashcode_int = generate_large_arrays_hashcode(T_INT);
11349 StubRoutines::aarch64::_large_arrays_hashcode_short = generate_large_arrays_hashcode(T_SHORT);
11350
11351 // byte_array_inflate stub for large arrays.
11352 StubRoutines::aarch64::_large_byte_array_inflate = generate_large_byte_array_inflate();
11353
11354 // countPositives stub for large arrays.
11355 StubRoutines::aarch64::_count_positives = generate_count_positives(StubRoutines::aarch64::_count_positives_long);
11356
11357 generate_compare_long_strings();
11358
11359 generate_string_indexof_stubs();
11360
11361 #ifdef COMPILER2
11362 if (UseMultiplyToLenIntrinsic) {
11363 StubRoutines::_multiplyToLen = generate_multiplyToLen();
11364 }
11365
11366 if (UseSquareToLenIntrinsic) {
11367 StubRoutines::_squareToLen = generate_squareToLen();
11368 }
11369
11370 if (UseMulAddIntrinsic) {
11371 StubRoutines::_mulAdd = generate_mulAdd();
11372 }
11373
11374 if (UseSIMDForBigIntegerShiftIntrinsics) {
11375 StubRoutines::_bigIntegerRightShiftWorker = generate_bigIntegerRightShift();
11376 StubRoutines::_bigIntegerLeftShiftWorker = generate_bigIntegerLeftShift();
11377 }
11378
11379 if (UseMontgomeryMultiplyIntrinsic) {
11380 StubGenStubId stub_id = StubGenStubId::montgomeryMultiply_id;
11381 StubCodeMark mark(this, stub_id);
11382 MontgomeryMultiplyGenerator g(_masm, /*squaring*/false);
11383 StubRoutines::_montgomeryMultiply = g.generate_multiply();
11384 }
11385
11386 if (UseMontgomerySquareIntrinsic) {
11387 StubGenStubId stub_id = StubGenStubId::montgomerySquare_id;
11388 StubCodeMark mark(this, stub_id);
11389 MontgomeryMultiplyGenerator g(_masm, /*squaring*/true);
11390 // We use generate_multiply() rather than generate_square()
11391 // because it's faster for the sizes of modulus we care about.
11392 StubRoutines::_montgomerySquare = g.generate_multiply();
11393 }
11394
11395 generate_vector_math_stubs();
11396
11397 #endif // COMPILER2
11398
11399 if (UseChaCha20Intrinsics) {
11400 StubRoutines::_chacha20Block = generate_chacha20Block_blockpar();
11401 }
11402
11403 if (UseKyberIntrinsics) {
11404 StubRoutines::_kyberNtt = generate_kyberNtt();
11405 StubRoutines::_kyberInverseNtt = generate_kyberInverseNtt();
11406 StubRoutines::_kyberNttMult = generate_kyberNttMult();
11407 StubRoutines::_kyberAddPoly_2 = generate_kyberAddPoly_2();
11408 StubRoutines::_kyberAddPoly_3 = generate_kyberAddPoly_3();
11409 StubRoutines::_kyber12To16 = generate_kyber12To16();
11410 StubRoutines::_kyberBarrettReduce = generate_kyberBarrettReduce();
11411 }
11412
11413 if (UseDilithiumIntrinsics) {
11414 StubRoutines::_dilithiumAlmostNtt = generate_dilithiumAlmostNtt();
11415 StubRoutines::_dilithiumAlmostInverseNtt = generate_dilithiumAlmostInverseNtt();
11416 StubRoutines::_dilithiumNttMult = generate_dilithiumNttMult();
11417 StubRoutines::_dilithiumMontMulByConstant = generate_dilithiumMontMulByConstant();
11418 StubRoutines::_dilithiumDecomposePoly = generate_dilithiumDecomposePoly();
11419 }
11420
11421 if (UseBASE64Intrinsics) {
11422 StubRoutines::_base64_encodeBlock = generate_base64_encodeBlock();
11423 StubRoutines::_base64_decodeBlock = generate_base64_decodeBlock();
11424 }
11425
11426 // data cache line writeback
11427 StubRoutines::_data_cache_writeback = generate_data_cache_writeback();
11428 StubRoutines::_data_cache_writeback_sync = generate_data_cache_writeback_sync();
11429
11430 if (UseAESIntrinsics) {
11431 StubRoutines::_aescrypt_encryptBlock = generate_aescrypt_encryptBlock();
11432 StubRoutines::_aescrypt_decryptBlock = generate_aescrypt_decryptBlock();
11433 StubRoutines::_cipherBlockChaining_encryptAESCrypt = generate_cipherBlockChaining_encryptAESCrypt();
11434 StubRoutines::_cipherBlockChaining_decryptAESCrypt = generate_cipherBlockChaining_decryptAESCrypt();
11435 StubRoutines::_counterMode_AESCrypt = generate_counterMode_AESCrypt();
11436 }
11437 if (UseGHASHIntrinsics) {
11438 // StubRoutines::_ghash_processBlocks = generate_ghash_processBlocks();
11439 StubRoutines::_ghash_processBlocks = generate_ghash_processBlocks_wide();
11440 }
11441 if (UseAESIntrinsics && UseGHASHIntrinsics) {
11442 StubRoutines::_galoisCounterMode_AESCrypt = generate_galoisCounterMode_AESCrypt();
11443 }
11444
11445 if (UseMD5Intrinsics) {
11446 StubRoutines::_md5_implCompress = generate_md5_implCompress(StubGenStubId::md5_implCompress_id);
11447 StubRoutines::_md5_implCompressMB = generate_md5_implCompress(StubGenStubId::md5_implCompressMB_id);
11448 }
11449 if (UseSHA1Intrinsics) {
11450 StubRoutines::_sha1_implCompress = generate_sha1_implCompress(StubGenStubId::sha1_implCompress_id);
11451 StubRoutines::_sha1_implCompressMB = generate_sha1_implCompress(StubGenStubId::sha1_implCompressMB_id);
11452 }
11453 if (UseSHA256Intrinsics) {
11454 StubRoutines::_sha256_implCompress = generate_sha256_implCompress(StubGenStubId::sha256_implCompress_id);
11455 StubRoutines::_sha256_implCompressMB = generate_sha256_implCompress(StubGenStubId::sha256_implCompressMB_id);
11456 }
11457 if (UseSHA512Intrinsics) {
11458 StubRoutines::_sha512_implCompress = generate_sha512_implCompress(StubGenStubId::sha512_implCompress_id);
11459 StubRoutines::_sha512_implCompressMB = generate_sha512_implCompress(StubGenStubId::sha512_implCompressMB_id);
11460 }
11461 if (UseSHA3Intrinsics) {
11462 StubRoutines::_sha3_implCompress = generate_sha3_implCompress(StubGenStubId::sha3_implCompress_id);
11463 StubRoutines::_double_keccak = generate_double_keccak();
11464 StubRoutines::_sha3_implCompressMB = generate_sha3_implCompress(StubGenStubId::sha3_implCompressMB_id);
11465 }
11466
11467 if (UsePoly1305Intrinsics) {
11468 StubRoutines::_poly1305_processBlocks = generate_poly1305_processBlocks();
11469 }
11470
11471 // generate Adler32 intrinsics code
11472 if (UseAdler32Intrinsics) {
11473 StubRoutines::_updateBytesAdler32 = generate_updateBytesAdler32();
11474 }
11475
11476 #endif // COMPILER2_OR_JVMCI
11477 }
11478
11479 public:
11480 StubGenerator(CodeBuffer* code, StubGenBlobId blob_id) : StubCodeGenerator(code, blob_id) {
11481 switch(blob_id) {
11482 case initial_id:
11483 generate_initial_stubs();
11484 break;
11485 case continuation_id:
11486 generate_continuation_stubs();
11487 break;
11488 case compiler_id:
11489 generate_compiler_stubs();
11490 break;
11491 case final_id:
11492 generate_final_stubs();
11493 break;
11494 default:
11495 fatal("unexpected blob id: %d", blob_id);
11496 break;
11497 };
11498 }
11499 }; // end class declaration
11500
11501 void StubGenerator_generate(CodeBuffer* code, StubGenBlobId blob_id) {
11502 StubGenerator g(code, blob_id);
11503 }
11504
11505
11506 #if defined (LINUX)
11507
11508 // Define pointers to atomic stubs and initialize them to point to the
11509 // code in atomic_aarch64.S.
11510
11511 #define DEFAULT_ATOMIC_OP(OPNAME, SIZE, RELAXED) \
11512 extern "C" uint64_t aarch64_atomic_ ## OPNAME ## _ ## SIZE ## RELAXED ## _default_impl \
11513 (volatile void *ptr, uint64_t arg1, uint64_t arg2); \
11514 aarch64_atomic_stub_t aarch64_atomic_ ## OPNAME ## _ ## SIZE ## RELAXED ## _impl \
11515 = aarch64_atomic_ ## OPNAME ## _ ## SIZE ## RELAXED ## _default_impl;
11516
11517 DEFAULT_ATOMIC_OP(fetch_add, 4, )
11518 DEFAULT_ATOMIC_OP(fetch_add, 8, )
11519 DEFAULT_ATOMIC_OP(fetch_add, 4, _relaxed)
11520 DEFAULT_ATOMIC_OP(fetch_add, 8, _relaxed)
11521 DEFAULT_ATOMIC_OP(xchg, 4, )
11522 DEFAULT_ATOMIC_OP(xchg, 8, )
11523 DEFAULT_ATOMIC_OP(cmpxchg, 1, )
11524 DEFAULT_ATOMIC_OP(cmpxchg, 4, )
11525 DEFAULT_ATOMIC_OP(cmpxchg, 8, )
11526 DEFAULT_ATOMIC_OP(cmpxchg, 1, _relaxed)
11527 DEFAULT_ATOMIC_OP(cmpxchg, 4, _relaxed)
11528 DEFAULT_ATOMIC_OP(cmpxchg, 8, _relaxed)
11529 DEFAULT_ATOMIC_OP(cmpxchg, 4, _release)
11530 DEFAULT_ATOMIC_OP(cmpxchg, 8, _release)
11531 DEFAULT_ATOMIC_OP(cmpxchg, 4, _seq_cst)
11532 DEFAULT_ATOMIC_OP(cmpxchg, 8, _seq_cst)
11533
11534 #undef DEFAULT_ATOMIC_OP
11535
11536 #endif // LINUX