1 /*
2 * Copyright (c) 2003, 2019, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "ci/ciUtilities.hpp"
29 #include "gc/shared/barrierSet.hpp"
30 #include "gc/shared/barrierSetAssembler.hpp"
31 #include "gc/shared/barrierSetNMethod.hpp"
32 #include "interpreter/interpreter.hpp"
33 #include "memory/universe.hpp"
34 #include "nativeInst_x86.hpp"
35 #include "oops/instanceOop.hpp"
36 #include "oops/method.hpp"
37 #include "oops/objArrayKlass.hpp"
38 #include "oops/oop.inline.hpp"
39 #include "prims/methodHandles.hpp"
40 #include "runtime/frame.inline.hpp"
41 #include "runtime/handles.inline.hpp"
42 #include "runtime/sharedRuntime.hpp"
43 #include "runtime/stubCodeGenerator.hpp"
44 #include "runtime/stubRoutines.hpp"
45 #include "runtime/thread.inline.hpp"
46 #ifdef COMPILER2
47 #include "opto/runtime.hpp"
48 #endif
49 #if INCLUDE_ZGC
50 #include "gc/z/zThreadLocalData.hpp"
51 #endif
52
53 // Declaration and definition of StubGenerator (no .hpp file).
54 // For a more detailed description of the stub routine structure
55 // see the comment in stubRoutines.hpp
56
57 #define __ _masm->
58 #define TIMES_OOP (UseCompressedOops ? Address::times_4 : Address::times_8)
59 #define a__ ((Assembler*)_masm)->
60
61 #ifdef PRODUCT
62 #define BLOCK_COMMENT(str) /* nothing */
63 #else
64 #define BLOCK_COMMENT(str) __ block_comment(str)
65 #endif
66
67 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
68 const int MXCSR_MASK = 0xFFC0; // Mask out any pending exceptions
69
70 // Stub Code definitions
71
72 class StubGenerator: public StubCodeGenerator {
73 private:
74
75 #ifdef PRODUCT
76 #define inc_counter_np(counter) ((void)0)
77 #else
78 void inc_counter_np_(int& counter) {
79 // This can destroy rscratch1 if counter is far from the code cache
80 __ incrementl(ExternalAddress((address)&counter));
81 }
82 #define inc_counter_np(counter) \
83 BLOCK_COMMENT("inc_counter " #counter); \
84 inc_counter_np_(counter);
85 #endif
86
87 // Call stubs are used to call Java from C
88 //
89 // Linux Arguments:
90 // c_rarg0: call wrapper address address
91 // c_rarg1: result address
92 // c_rarg2: result type BasicType
93 // c_rarg3: method Method*
94 // c_rarg4: (interpreter) entry point address
95 // c_rarg5: parameters intptr_t*
96 // 16(rbp): parameter size (in words) int
97 // 24(rbp): thread Thread*
98 //
99 // [ return_from_Java ] <--- rsp
100 // [ argument word n ]
101 // ...
102 // -12 [ argument word 1 ]
103 // -11 [ saved r15 ] <--- rsp_after_call
104 // -10 [ saved r14 ]
105 // -9 [ saved r13 ]
106 // -8 [ saved r12 ]
107 // -7 [ saved rbx ]
108 // -6 [ call wrapper ]
109 // -5 [ result ]
110 // -4 [ result type ]
111 // -3 [ method ]
112 // -2 [ entry point ]
113 // -1 [ parameters ]
114 // 0 [ saved rbp ] <--- rbp
115 // 1 [ return address ]
116 // 2 [ parameter size ]
117 // 3 [ thread ]
118 //
119 // Windows Arguments:
120 // c_rarg0: call wrapper address address
121 // c_rarg1: result address
122 // c_rarg2: result type BasicType
123 // c_rarg3: method Method*
124 // 48(rbp): (interpreter) entry point address
125 // 56(rbp): parameters intptr_t*
126 // 64(rbp): parameter size (in words) int
127 // 72(rbp): thread Thread*
128 //
129 // [ return_from_Java ] <--- rsp
130 // [ argument word n ]
131 // ...
132 // -60 [ argument word 1 ]
133 // -59 [ saved xmm31 ] <--- rsp after_call
134 // [ saved xmm16-xmm30 ] (EVEX enabled, else the space is blank)
135 // -27 [ saved xmm15 ]
136 // [ saved xmm7-xmm14 ]
137 // -9 [ saved xmm6 ] (each xmm register takes 2 slots)
138 // -7 [ saved r15 ]
139 // -6 [ saved r14 ]
140 // -5 [ saved r13 ]
141 // -4 [ saved r12 ]
142 // -3 [ saved rdi ]
143 // -2 [ saved rsi ]
144 // -1 [ saved rbx ]
145 // 0 [ saved rbp ] <--- rbp
146 // 1 [ return address ]
147 // 2 [ call wrapper ]
148 // 3 [ result ]
149 // 4 [ result type ]
150 // 5 [ method ]
151 // 6 [ entry point ]
152 // 7 [ parameters ]
153 // 8 [ parameter size ]
154 // 9 [ thread ]
155 //
156 // Windows reserves the callers stack space for arguments 1-4.
157 // We spill c_rarg0-c_rarg3 to this space.
158
159 // Call stub stack layout word offsets from rbp
160 enum call_stub_layout {
161 #ifdef _WIN64
162 xmm_save_first = 6, // save from xmm6
163 xmm_save_last = 31, // to xmm31
164 xmm_save_base = -9,
165 rsp_after_call_off = xmm_save_base - 2 * (xmm_save_last - xmm_save_first), // -27
166 r15_off = -7,
167 r14_off = -6,
168 r13_off = -5,
169 r12_off = -4,
170 rdi_off = -3,
171 rsi_off = -2,
172 rbx_off = -1,
173 rbp_off = 0,
174 retaddr_off = 1,
175 call_wrapper_off = 2,
176 result_off = 3,
177 result_type_off = 4,
178 method_off = 5,
179 entry_point_off = 6,
180 parameters_off = 7,
181 parameter_size_off = 8,
182 thread_off = 9
183 #else
184 rsp_after_call_off = -12,
185 mxcsr_off = rsp_after_call_off,
186 r15_off = -11,
187 r14_off = -10,
188 r13_off = -9,
189 r12_off = -8,
190 rbx_off = -7,
191 call_wrapper_off = -6,
192 result_off = -5,
193 result_type_off = -4,
194 method_off = -3,
195 entry_point_off = -2,
196 parameters_off = -1,
197 rbp_off = 0,
198 retaddr_off = 1,
199 parameter_size_off = 2,
200 thread_off = 3
201 #endif
202 };
203
204 #ifdef _WIN64
205 Address xmm_save(int reg) {
206 assert(reg >= xmm_save_first && reg <= xmm_save_last, "XMM register number out of range");
207 return Address(rbp, (xmm_save_base - (reg - xmm_save_first) * 2) * wordSize);
208 }
209 #endif
210
211 address generate_call_stub(address& return_address) {
212 assert((int)frame::entry_frame_after_call_words == -(int)rsp_after_call_off + 1 &&
213 (int)frame::entry_frame_call_wrapper_offset == (int)call_wrapper_off,
214 "adjust this code");
215 StubCodeMark mark(this, "StubRoutines", "call_stub");
216 address start = __ pc();
217
218 // same as in generate_catch_exception()!
219 const Address rsp_after_call(rbp, rsp_after_call_off * wordSize);
220
221 const Address call_wrapper (rbp, call_wrapper_off * wordSize);
222 const Address result (rbp, result_off * wordSize);
223 const Address result_type (rbp, result_type_off * wordSize);
224 const Address method (rbp, method_off * wordSize);
225 const Address entry_point (rbp, entry_point_off * wordSize);
226 const Address parameters (rbp, parameters_off * wordSize);
227 const Address parameter_size(rbp, parameter_size_off * wordSize);
228
229 // same as in generate_catch_exception()!
230 const Address thread (rbp, thread_off * wordSize);
231
232 const Address r15_save(rbp, r15_off * wordSize);
233 const Address r14_save(rbp, r14_off * wordSize);
234 const Address r13_save(rbp, r13_off * wordSize);
235 const Address r12_save(rbp, r12_off * wordSize);
236 const Address rbx_save(rbp, rbx_off * wordSize);
237
238 // stub code
239 __ enter();
240 __ subptr(rsp, -rsp_after_call_off * wordSize);
241
242 // save register parameters
243 #ifndef _WIN64
244 __ movptr(parameters, c_rarg5); // parameters
245 __ movptr(entry_point, c_rarg4); // entry_point
246 #endif
247
248 __ movptr(method, c_rarg3); // method
249 __ movl(result_type, c_rarg2); // result type
250 __ movptr(result, c_rarg1); // result
251 __ movptr(call_wrapper, c_rarg0); // call wrapper
252
253 // save regs belonging to calling function
254 __ movptr(rbx_save, rbx);
255 __ movptr(r12_save, r12);
256 __ movptr(r13_save, r13);
257 __ movptr(r14_save, r14);
258 __ movptr(r15_save, r15);
259
260 #ifdef _WIN64
261 int last_reg = 15;
262 if (UseAVX > 2) {
263 last_reg = 31;
264 }
265 if (VM_Version::supports_evex()) {
266 for (int i = xmm_save_first; i <= last_reg; i++) {
267 __ vextractf32x4(xmm_save(i), as_XMMRegister(i), 0);
268 }
269 } else {
270 for (int i = xmm_save_first; i <= last_reg; i++) {
271 __ movdqu(xmm_save(i), as_XMMRegister(i));
272 }
273 }
274
275 const Address rdi_save(rbp, rdi_off * wordSize);
276 const Address rsi_save(rbp, rsi_off * wordSize);
277
278 __ movptr(rsi_save, rsi);
279 __ movptr(rdi_save, rdi);
280 #else
281 const Address mxcsr_save(rbp, mxcsr_off * wordSize);
282 {
283 Label skip_ldmx;
284 __ stmxcsr(mxcsr_save);
285 __ movl(rax, mxcsr_save);
286 __ andl(rax, MXCSR_MASK); // Only check control and mask bits
287 ExternalAddress mxcsr_std(StubRoutines::addr_mxcsr_std());
288 __ cmp32(rax, mxcsr_std);
289 __ jcc(Assembler::equal, skip_ldmx);
290 __ ldmxcsr(mxcsr_std);
291 __ bind(skip_ldmx);
292 }
293 #endif
294
295 // Load up thread register
296 __ movptr(r15_thread, thread);
297 __ reinit_heapbase();
298
299 #ifdef ASSERT
300 // make sure we have no pending exceptions
301 {
302 Label L;
303 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
304 __ jcc(Assembler::equal, L);
305 __ stop("StubRoutines::call_stub: entered with pending exception");
306 __ bind(L);
307 }
308 #endif
309
310 // pass parameters if any
311 BLOCK_COMMENT("pass parameters if any");
312 Label parameters_done;
313 __ movl(c_rarg3, parameter_size);
314 __ testl(c_rarg3, c_rarg3);
315 __ jcc(Assembler::zero, parameters_done);
316
317 Label loop;
318 __ movptr(c_rarg2, parameters); // parameter pointer
319 __ movl(c_rarg1, c_rarg3); // parameter counter is in c_rarg1
320 __ BIND(loop);
321 __ movptr(rax, Address(c_rarg2, 0));// get parameter
322 __ addptr(c_rarg2, wordSize); // advance to next parameter
323 __ decrementl(c_rarg1); // decrement counter
324 __ push(rax); // pass parameter
325 __ jcc(Assembler::notZero, loop);
326
327 // call Java function
328 __ BIND(parameters_done);
329 __ movptr(rbx, method); // get Method*
330 __ movptr(c_rarg1, entry_point); // get entry_point
331 __ mov(r13, rsp); // set sender sp
332 BLOCK_COMMENT("call Java function");
333 __ call(c_rarg1);
334
335 BLOCK_COMMENT("call_stub_return_address:");
336 return_address = __ pc();
337
338 // store result depending on type (everything that is not
339 // T_OBJECT, T_LONG, T_FLOAT or T_DOUBLE is treated as T_INT)
340 __ movptr(c_rarg0, result);
341 Label is_long, is_float, is_double, exit;
342 __ movl(c_rarg1, result_type);
343 __ cmpl(c_rarg1, T_OBJECT);
344 __ jcc(Assembler::equal, is_long);
345 __ cmpl(c_rarg1, T_LONG);
346 __ jcc(Assembler::equal, is_long);
347 __ cmpl(c_rarg1, T_FLOAT);
348 __ jcc(Assembler::equal, is_float);
349 __ cmpl(c_rarg1, T_DOUBLE);
350 __ jcc(Assembler::equal, is_double);
351
352 // handle T_INT case
353 __ movl(Address(c_rarg0, 0), rax);
354
355 __ BIND(exit);
356
357 // pop parameters
358 __ lea(rsp, rsp_after_call);
359
360 #ifdef ASSERT
361 // verify that threads correspond
362 {
363 Label L1, L2, L3;
364 __ cmpptr(r15_thread, thread);
365 __ jcc(Assembler::equal, L1);
366 __ stop("StubRoutines::call_stub: r15_thread is corrupted");
367 __ bind(L1);
368 __ get_thread(rbx);
369 __ cmpptr(r15_thread, thread);
370 __ jcc(Assembler::equal, L2);
371 __ stop("StubRoutines::call_stub: r15_thread is modified by call");
372 __ bind(L2);
373 __ cmpptr(r15_thread, rbx);
374 __ jcc(Assembler::equal, L3);
375 __ stop("StubRoutines::call_stub: threads must correspond");
376 __ bind(L3);
377 }
378 #endif
379
380 // restore regs belonging to calling function
381 #ifdef _WIN64
382 // emit the restores for xmm regs
383 if (VM_Version::supports_evex()) {
384 for (int i = xmm_save_first; i <= last_reg; i++) {
385 __ vinsertf32x4(as_XMMRegister(i), as_XMMRegister(i), xmm_save(i), 0);
386 }
387 } else {
388 for (int i = xmm_save_first; i <= last_reg; i++) {
389 __ movdqu(as_XMMRegister(i), xmm_save(i));
390 }
391 }
392 #endif
393 __ movptr(r15, r15_save);
394 __ movptr(r14, r14_save);
395 __ movptr(r13, r13_save);
396 __ movptr(r12, r12_save);
397 __ movptr(rbx, rbx_save);
398
399 #ifdef _WIN64
400 __ movptr(rdi, rdi_save);
401 __ movptr(rsi, rsi_save);
402 #else
403 __ ldmxcsr(mxcsr_save);
404 #endif
405
406 // restore rsp
407 __ addptr(rsp, -rsp_after_call_off * wordSize);
408
409 // return
410 __ vzeroupper();
411 __ pop(rbp);
412 __ ret(0);
413
414 // handle return types different from T_INT
415 __ BIND(is_long);
416 __ movq(Address(c_rarg0, 0), rax);
417 __ jmp(exit);
418
419 __ BIND(is_float);
420 __ movflt(Address(c_rarg0, 0), xmm0);
421 __ jmp(exit);
422
423 __ BIND(is_double);
424 __ movdbl(Address(c_rarg0, 0), xmm0);
425 __ jmp(exit);
426
427 return start;
428 }
429
430 // Return point for a Java call if there's an exception thrown in
431 // Java code. The exception is caught and transformed into a
432 // pending exception stored in JavaThread that can be tested from
433 // within the VM.
434 //
435 // Note: Usually the parameters are removed by the callee. In case
436 // of an exception crossing an activation frame boundary, that is
437 // not the case if the callee is compiled code => need to setup the
438 // rsp.
439 //
440 // rax: exception oop
441
442 address generate_catch_exception() {
443 StubCodeMark mark(this, "StubRoutines", "catch_exception");
444 address start = __ pc();
445
446 // same as in generate_call_stub():
447 const Address rsp_after_call(rbp, rsp_after_call_off * wordSize);
448 const Address thread (rbp, thread_off * wordSize);
449
450 #ifdef ASSERT
451 // verify that threads correspond
452 {
453 Label L1, L2, L3;
454 __ cmpptr(r15_thread, thread);
455 __ jcc(Assembler::equal, L1);
456 __ stop("StubRoutines::catch_exception: r15_thread is corrupted");
457 __ bind(L1);
458 __ get_thread(rbx);
459 __ cmpptr(r15_thread, thread);
460 __ jcc(Assembler::equal, L2);
461 __ stop("StubRoutines::catch_exception: r15_thread is modified by call");
462 __ bind(L2);
463 __ cmpptr(r15_thread, rbx);
464 __ jcc(Assembler::equal, L3);
465 __ stop("StubRoutines::catch_exception: threads must correspond");
466 __ bind(L3);
467 }
468 #endif
469
470 // set pending exception
471 __ verify_oop(rax);
472
473 __ movptr(Address(r15_thread, Thread::pending_exception_offset()), rax);
474 __ lea(rscratch1, ExternalAddress((address)__FILE__));
475 __ movptr(Address(r15_thread, Thread::exception_file_offset()), rscratch1);
476 __ movl(Address(r15_thread, Thread::exception_line_offset()), (int) __LINE__);
477
478 // complete return to VM
479 assert(StubRoutines::_call_stub_return_address != NULL,
480 "_call_stub_return_address must have been generated before");
481 __ jump(RuntimeAddress(StubRoutines::_call_stub_return_address));
482
483 return start;
484 }
485
486 // Continuation point for runtime calls returning with a pending
487 // exception. The pending exception check happened in the runtime
488 // or native call stub. The pending exception in Thread is
489 // converted into a Java-level exception.
490 //
491 // Contract with Java-level exception handlers:
492 // rax: exception
493 // rdx: throwing pc
494 //
495 // NOTE: At entry of this stub, exception-pc must be on stack !!
496
497 address generate_forward_exception() {
498 StubCodeMark mark(this, "StubRoutines", "forward exception");
499 address start = __ pc();
500
501 // Upon entry, the sp points to the return address returning into
502 // Java (interpreted or compiled) code; i.e., the return address
503 // becomes the throwing pc.
504 //
505 // Arguments pushed before the runtime call are still on the stack
506 // but the exception handler will reset the stack pointer ->
507 // ignore them. A potential result in registers can be ignored as
508 // well.
509
510 #ifdef ASSERT
511 // make sure this code is only executed if there is a pending exception
512 {
513 Label L;
514 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t) NULL);
515 __ jcc(Assembler::notEqual, L);
516 __ stop("StubRoutines::forward exception: no pending exception (1)");
517 __ bind(L);
518 }
519 #endif
520
521 // compute exception handler into rbx
522 __ movptr(c_rarg0, Address(rsp, 0));
523 BLOCK_COMMENT("call exception_handler_for_return_address");
524 __ call_VM_leaf(CAST_FROM_FN_PTR(address,
525 SharedRuntime::exception_handler_for_return_address),
526 r15_thread, c_rarg0);
527 __ mov(rbx, rax);
528
529 // setup rax & rdx, remove return address & clear pending exception
530 __ pop(rdx);
531 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset()));
532 __ movptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
533
534 #ifdef ASSERT
535 // make sure exception is set
536 {
537 Label L;
538 __ testptr(rax, rax);
539 __ jcc(Assembler::notEqual, L);
540 __ stop("StubRoutines::forward exception: no pending exception (2)");
541 __ bind(L);
542 }
543 #endif
544
545 // continue at exception handler (return address removed)
546 // rax: exception
547 // rbx: exception handler
548 // rdx: throwing pc
549 __ verify_oop(rax);
550 __ jmp(rbx);
551
552 return start;
553 }
554
555 // Support for intptr_t OrderAccess::fence()
556 //
557 // Arguments :
558 //
559 // Result:
560 address generate_orderaccess_fence() {
561 StubCodeMark mark(this, "StubRoutines", "orderaccess_fence");
562 address start = __ pc();
563 __ membar(Assembler::StoreLoad);
564 __ ret(0);
565
566 return start;
567 }
568
569 // Support for intptr_t get_previous_fp()
570 //
571 // This routine is used to find the previous frame pointer for the
572 // caller (current_frame_guess). This is used as part of debugging
573 // ps() is seemingly lost trying to find frames.
574 // This code assumes that caller current_frame_guess) has a frame.
575 address generate_get_previous_fp() {
576 StubCodeMark mark(this, "StubRoutines", "get_previous_fp");
577 const Address old_fp(rbp, 0);
578 const Address older_fp(rax, 0);
579 address start = __ pc();
580
581 __ enter();
582 __ movptr(rax, old_fp); // callers fp
583 __ movptr(rax, older_fp); // the frame for ps()
584 __ pop(rbp);
585 __ ret(0);
586
587 return start;
588 }
589
590 // Support for intptr_t get_previous_sp()
591 //
592 // This routine is used to find the previous stack pointer for the
593 // caller.
594 address generate_get_previous_sp() {
595 StubCodeMark mark(this, "StubRoutines", "get_previous_sp");
596 address start = __ pc();
597
598 __ movptr(rax, rsp);
599 __ addptr(rax, 8); // return address is at the top of the stack.
600 __ ret(0);
601
602 return start;
603 }
604
605 //----------------------------------------------------------------------------------------------------
606 // Support for void verify_mxcsr()
607 //
608 // This routine is used with -Xcheck:jni to verify that native
609 // JNI code does not return to Java code without restoring the
610 // MXCSR register to our expected state.
611
612 address generate_verify_mxcsr() {
613 StubCodeMark mark(this, "StubRoutines", "verify_mxcsr");
614 address start = __ pc();
615
616 const Address mxcsr_save(rsp, 0);
617
618 if (CheckJNICalls) {
619 Label ok_ret;
620 ExternalAddress mxcsr_std(StubRoutines::addr_mxcsr_std());
621 __ push(rax);
622 __ subptr(rsp, wordSize); // allocate a temp location
623 __ stmxcsr(mxcsr_save);
624 __ movl(rax, mxcsr_save);
625 __ andl(rax, MXCSR_MASK); // Only check control and mask bits
626 __ cmp32(rax, mxcsr_std);
627 __ jcc(Assembler::equal, ok_ret);
628
629 __ warn("MXCSR changed by native JNI code, use -XX:+RestoreMXCSROnJNICall");
630
631 __ ldmxcsr(mxcsr_std);
632
633 __ bind(ok_ret);
634 __ addptr(rsp, wordSize);
635 __ pop(rax);
636 }
637
638 __ ret(0);
639
640 return start;
641 }
642
643 address generate_f2i_fixup() {
644 StubCodeMark mark(this, "StubRoutines", "f2i_fixup");
645 Address inout(rsp, 5 * wordSize); // return address + 4 saves
646
647 address start = __ pc();
648
649 Label L;
650
651 __ push(rax);
652 __ push(c_rarg3);
653 __ push(c_rarg2);
654 __ push(c_rarg1);
655
656 __ movl(rax, 0x7f800000);
657 __ xorl(c_rarg3, c_rarg3);
658 __ movl(c_rarg2, inout);
659 __ movl(c_rarg1, c_rarg2);
660 __ andl(c_rarg1, 0x7fffffff);
661 __ cmpl(rax, c_rarg1); // NaN? -> 0
662 __ jcc(Assembler::negative, L);
663 __ testl(c_rarg2, c_rarg2); // signed ? min_jint : max_jint
664 __ movl(c_rarg3, 0x80000000);
665 __ movl(rax, 0x7fffffff);
666 __ cmovl(Assembler::positive, c_rarg3, rax);
667
668 __ bind(L);
669 __ movptr(inout, c_rarg3);
670
671 __ pop(c_rarg1);
672 __ pop(c_rarg2);
673 __ pop(c_rarg3);
674 __ pop(rax);
675
676 __ ret(0);
677
678 return start;
679 }
680
681 address generate_f2l_fixup() {
682 StubCodeMark mark(this, "StubRoutines", "f2l_fixup");
683 Address inout(rsp, 5 * wordSize); // return address + 4 saves
684 address start = __ pc();
685
686 Label L;
687
688 __ push(rax);
689 __ push(c_rarg3);
690 __ push(c_rarg2);
691 __ push(c_rarg1);
692
693 __ movl(rax, 0x7f800000);
694 __ xorl(c_rarg3, c_rarg3);
695 __ movl(c_rarg2, inout);
696 __ movl(c_rarg1, c_rarg2);
697 __ andl(c_rarg1, 0x7fffffff);
698 __ cmpl(rax, c_rarg1); // NaN? -> 0
699 __ jcc(Assembler::negative, L);
700 __ testl(c_rarg2, c_rarg2); // signed ? min_jlong : max_jlong
701 __ mov64(c_rarg3, 0x8000000000000000);
702 __ mov64(rax, 0x7fffffffffffffff);
703 __ cmov(Assembler::positive, c_rarg3, rax);
704
705 __ bind(L);
706 __ movptr(inout, c_rarg3);
707
708 __ pop(c_rarg1);
709 __ pop(c_rarg2);
710 __ pop(c_rarg3);
711 __ pop(rax);
712
713 __ ret(0);
714
715 return start;
716 }
717
718 address generate_d2i_fixup() {
719 StubCodeMark mark(this, "StubRoutines", "d2i_fixup");
720 Address inout(rsp, 6 * wordSize); // return address + 5 saves
721
722 address start = __ pc();
723
724 Label L;
725
726 __ push(rax);
727 __ push(c_rarg3);
728 __ push(c_rarg2);
729 __ push(c_rarg1);
730 __ push(c_rarg0);
731
732 __ movl(rax, 0x7ff00000);
733 __ movq(c_rarg2, inout);
734 __ movl(c_rarg3, c_rarg2);
735 __ mov(c_rarg1, c_rarg2);
736 __ mov(c_rarg0, c_rarg2);
737 __ negl(c_rarg3);
738 __ shrptr(c_rarg1, 0x20);
739 __ orl(c_rarg3, c_rarg2);
740 __ andl(c_rarg1, 0x7fffffff);
741 __ xorl(c_rarg2, c_rarg2);
742 __ shrl(c_rarg3, 0x1f);
743 __ orl(c_rarg1, c_rarg3);
744 __ cmpl(rax, c_rarg1);
745 __ jcc(Assembler::negative, L); // NaN -> 0
746 __ testptr(c_rarg0, c_rarg0); // signed ? min_jint : max_jint
747 __ movl(c_rarg2, 0x80000000);
748 __ movl(rax, 0x7fffffff);
749 __ cmov(Assembler::positive, c_rarg2, rax);
750
751 __ bind(L);
752 __ movptr(inout, c_rarg2);
753
754 __ pop(c_rarg0);
755 __ pop(c_rarg1);
756 __ pop(c_rarg2);
757 __ pop(c_rarg3);
758 __ pop(rax);
759
760 __ ret(0);
761
762 return start;
763 }
764
765 address generate_d2l_fixup() {
766 StubCodeMark mark(this, "StubRoutines", "d2l_fixup");
767 Address inout(rsp, 6 * wordSize); // return address + 5 saves
768
769 address start = __ pc();
770
771 Label L;
772
773 __ push(rax);
774 __ push(c_rarg3);
775 __ push(c_rarg2);
776 __ push(c_rarg1);
777 __ push(c_rarg0);
778
779 __ movl(rax, 0x7ff00000);
780 __ movq(c_rarg2, inout);
781 __ movl(c_rarg3, c_rarg2);
782 __ mov(c_rarg1, c_rarg2);
783 __ mov(c_rarg0, c_rarg2);
784 __ negl(c_rarg3);
785 __ shrptr(c_rarg1, 0x20);
786 __ orl(c_rarg3, c_rarg2);
787 __ andl(c_rarg1, 0x7fffffff);
788 __ xorl(c_rarg2, c_rarg2);
789 __ shrl(c_rarg3, 0x1f);
790 __ orl(c_rarg1, c_rarg3);
791 __ cmpl(rax, c_rarg1);
792 __ jcc(Assembler::negative, L); // NaN -> 0
793 __ testq(c_rarg0, c_rarg0); // signed ? min_jlong : max_jlong
794 __ mov64(c_rarg2, 0x8000000000000000);
795 __ mov64(rax, 0x7fffffffffffffff);
796 __ cmovq(Assembler::positive, c_rarg2, rax);
797
798 __ bind(L);
799 __ movq(inout, c_rarg2);
800
801 __ pop(c_rarg0);
802 __ pop(c_rarg1);
803 __ pop(c_rarg2);
804 __ pop(c_rarg3);
805 __ pop(rax);
806
807 __ ret(0);
808
809 return start;
810 }
811
812 address generate_fp_mask(const char *stub_name, int64_t mask) {
813 __ align(CodeEntryAlignment);
814 StubCodeMark mark(this, "StubRoutines", stub_name);
815 address start = __ pc();
816
817 __ emit_data64( mask, relocInfo::none );
818 __ emit_data64( mask, relocInfo::none );
819
820 return start;
821 }
822
823 address generate_vector_mask(const char *stub_name, int64_t mask) {
824 __ align(CodeEntryAlignment);
825 StubCodeMark mark(this, "StubRoutines", stub_name);
826 address start = __ pc();
827
828 __ emit_data64(mask, relocInfo::none);
829 __ emit_data64(mask, relocInfo::none);
830 __ emit_data64(mask, relocInfo::none);
831 __ emit_data64(mask, relocInfo::none);
832 __ emit_data64(mask, relocInfo::none);
833 __ emit_data64(mask, relocInfo::none);
834 __ emit_data64(mask, relocInfo::none);
835 __ emit_data64(mask, relocInfo::none);
836
837 return start;
838 }
839
840 address generate_vector_byte_perm_mask(const char *stub_name) {
841 __ align(CodeEntryAlignment);
842 StubCodeMark mark(this, "StubRoutines", stub_name);
843 address start = __ pc();
844
845 __ emit_data64(0x0000000000000001, relocInfo::none);
846 __ emit_data64(0x0000000000000003, relocInfo::none);
847 __ emit_data64(0x0000000000000005, relocInfo::none);
848 __ emit_data64(0x0000000000000007, relocInfo::none);
849 __ emit_data64(0x0000000000000000, relocInfo::none);
850 __ emit_data64(0x0000000000000002, relocInfo::none);
851 __ emit_data64(0x0000000000000004, relocInfo::none);
852 __ emit_data64(0x0000000000000006, relocInfo::none);
853
854 return start;
855 }
856
857 // Non-destructive plausibility checks for oops
858 //
859 // Arguments:
860 // all args on stack!
861 //
862 // Stack after saving c_rarg3:
863 // [tos + 0]: saved c_rarg3
864 // [tos + 1]: saved c_rarg2
865 // [tos + 2]: saved r12 (several TemplateTable methods use it)
866 // [tos + 3]: saved flags
867 // [tos + 4]: return address
868 // * [tos + 5]: error message (char*)
869 // * [tos + 6]: object to verify (oop)
870 // * [tos + 7]: saved rax - saved by caller and bashed
871 // * [tos + 8]: saved r10 (rscratch1) - saved by caller
872 // * = popped on exit
873 address generate_verify_oop() {
874 StubCodeMark mark(this, "StubRoutines", "verify_oop");
875 address start = __ pc();
876
877 Label exit, error;
878
879 __ pushf();
880 __ incrementl(ExternalAddress((address) StubRoutines::verify_oop_count_addr()));
881
882 __ push(r12);
883
884 // save c_rarg2 and c_rarg3
885 __ push(c_rarg2);
886 __ push(c_rarg3);
887
888 enum {
889 // After previous pushes.
890 oop_to_verify = 6 * wordSize,
891 saved_rax = 7 * wordSize,
892 saved_r10 = 8 * wordSize,
893
894 // Before the call to MacroAssembler::debug(), see below.
895 return_addr = 16 * wordSize,
896 error_msg = 17 * wordSize
897 };
898
899 // get object
900 __ movptr(rax, Address(rsp, oop_to_verify));
901
902 // make sure object is 'reasonable'
903 __ testptr(rax, rax);
904 __ jcc(Assembler::zero, exit); // if obj is NULL it is OK
905
906 #if INCLUDE_ZGC
907 if (UseZGC) {
908 // Check if metadata bits indicate a bad oop
909 __ testptr(rax, Address(r15_thread, ZThreadLocalData::address_bad_mask_offset()));
910 __ jcc(Assembler::notZero, error);
911 }
912 #endif
913
914 // Check if the oop is in the right area of memory
915 __ movptr(c_rarg2, rax);
916 __ movptr(c_rarg3, (intptr_t) Universe::verify_oop_mask());
917 __ andptr(c_rarg2, c_rarg3);
918 __ movptr(c_rarg3, (intptr_t) Universe::verify_oop_bits());
919 __ cmpptr(c_rarg2, c_rarg3);
920 __ jcc(Assembler::notZero, error);
921
922 // make sure klass is 'reasonable', which is not zero.
923 __ load_klass(rax, rax, rscratch1); // get klass
924 __ testptr(rax, rax);
925 __ jcc(Assembler::zero, error); // if klass is NULL it is broken
926
927 // return if everything seems ok
928 __ bind(exit);
929 __ movptr(rax, Address(rsp, saved_rax)); // get saved rax back
930 __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back
931 __ pop(c_rarg3); // restore c_rarg3
932 __ pop(c_rarg2); // restore c_rarg2
933 __ pop(r12); // restore r12
934 __ popf(); // restore flags
935 __ ret(4 * wordSize); // pop caller saved stuff
936
937 // handle errors
938 __ bind(error);
939 __ movptr(rax, Address(rsp, saved_rax)); // get saved rax back
940 __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back
941 __ pop(c_rarg3); // get saved c_rarg3 back
942 __ pop(c_rarg2); // get saved c_rarg2 back
943 __ pop(r12); // get saved r12 back
944 __ popf(); // get saved flags off stack --
945 // will be ignored
946
947 __ pusha(); // push registers
948 // (rip is already
949 // already pushed)
950 // debug(char* msg, int64_t pc, int64_t regs[])
951 // We've popped the registers we'd saved (c_rarg3, c_rarg2 and flags), and
952 // pushed all the registers, so now the stack looks like:
953 // [tos + 0] 16 saved registers
954 // [tos + 16] return address
955 // * [tos + 17] error message (char*)
956 // * [tos + 18] object to verify (oop)
957 // * [tos + 19] saved rax - saved by caller and bashed
958 // * [tos + 20] saved r10 (rscratch1) - saved by caller
959 // * = popped on exit
960
961 __ movptr(c_rarg0, Address(rsp, error_msg)); // pass address of error message
962 __ movptr(c_rarg1, Address(rsp, return_addr)); // pass return address
963 __ movq(c_rarg2, rsp); // pass address of regs on stack
964 __ mov(r12, rsp); // remember rsp
965 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows
966 __ andptr(rsp, -16); // align stack as required by ABI
967 BLOCK_COMMENT("call MacroAssembler::debug");
968 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64)));
969 __ hlt();
970 return start;
971 }
972
973 //
974 // Verify that a register contains clean 32-bits positive value
975 // (high 32-bits are 0) so it could be used in 64-bits shifts.
976 //
977 // Input:
978 // Rint - 32-bits value
979 // Rtmp - scratch
980 //
981 void assert_clean_int(Register Rint, Register Rtmp) {
982 #ifdef ASSERT
983 Label L;
984 assert_different_registers(Rtmp, Rint);
985 __ movslq(Rtmp, Rint);
986 __ cmpq(Rtmp, Rint);
987 __ jcc(Assembler::equal, L);
988 __ stop("high 32-bits of int value are not 0");
989 __ bind(L);
990 #endif
991 }
992
993 // Generate overlap test for array copy stubs
994 //
995 // Input:
996 // c_rarg0 - from
997 // c_rarg1 - to
998 // c_rarg2 - element count
999 //
1000 // Output:
1001 // rax - &from[element count - 1]
1002 //
1003 void array_overlap_test(address no_overlap_target, Address::ScaleFactor sf) {
1004 assert(no_overlap_target != NULL, "must be generated");
1005 array_overlap_test(no_overlap_target, NULL, sf);
1006 }
1007 void array_overlap_test(Label& L_no_overlap, Address::ScaleFactor sf) {
1008 array_overlap_test(NULL, &L_no_overlap, sf);
1009 }
1010 void array_overlap_test(address no_overlap_target, Label* NOLp, Address::ScaleFactor sf) {
1011 const Register from = c_rarg0;
1012 const Register to = c_rarg1;
1013 const Register count = c_rarg2;
1014 const Register end_from = rax;
1015
1016 __ cmpptr(to, from);
1017 __ lea(end_from, Address(from, count, sf, 0));
1018 if (NOLp == NULL) {
1019 ExternalAddress no_overlap(no_overlap_target);
1020 __ jump_cc(Assembler::belowEqual, no_overlap);
1021 __ cmpptr(to, end_from);
1022 __ jump_cc(Assembler::aboveEqual, no_overlap);
1023 } else {
1024 __ jcc(Assembler::belowEqual, (*NOLp));
1025 __ cmpptr(to, end_from);
1026 __ jcc(Assembler::aboveEqual, (*NOLp));
1027 }
1028 }
1029
1030 // Shuffle first three arg regs on Windows into Linux/Solaris locations.
1031 //
1032 // Outputs:
1033 // rdi - rcx
1034 // rsi - rdx
1035 // rdx - r8
1036 // rcx - r9
1037 //
1038 // Registers r9 and r10 are used to save rdi and rsi on Windows, which latter
1039 // are non-volatile. r9 and r10 should not be used by the caller.
1040 //
1041 DEBUG_ONLY(bool regs_in_thread;)
1042
1043 void setup_arg_regs(int nargs = 3) {
1044 const Register saved_rdi = r9;
1045 const Register saved_rsi = r10;
1046 assert(nargs == 3 || nargs == 4, "else fix");
1047 #ifdef _WIN64
1048 assert(c_rarg0 == rcx && c_rarg1 == rdx && c_rarg2 == r8 && c_rarg3 == r9,
1049 "unexpected argument registers");
1050 if (nargs >= 4)
1051 __ mov(rax, r9); // r9 is also saved_rdi
1052 __ movptr(saved_rdi, rdi);
1053 __ movptr(saved_rsi, rsi);
1054 __ mov(rdi, rcx); // c_rarg0
1055 __ mov(rsi, rdx); // c_rarg1
1056 __ mov(rdx, r8); // c_rarg2
1057 if (nargs >= 4)
1058 __ mov(rcx, rax); // c_rarg3 (via rax)
1059 #else
1060 assert(c_rarg0 == rdi && c_rarg1 == rsi && c_rarg2 == rdx && c_rarg3 == rcx,
1061 "unexpected argument registers");
1062 #endif
1063 DEBUG_ONLY(regs_in_thread = false;)
1064 }
1065
1066 void restore_arg_regs() {
1067 assert(!regs_in_thread, "wrong call to restore_arg_regs");
1068 const Register saved_rdi = r9;
1069 const Register saved_rsi = r10;
1070 #ifdef _WIN64
1071 __ movptr(rdi, saved_rdi);
1072 __ movptr(rsi, saved_rsi);
1073 #endif
1074 }
1075
1076 // This is used in places where r10 is a scratch register, and can
1077 // be adapted if r9 is needed also.
1078 void setup_arg_regs_using_thread() {
1079 const Register saved_r15 = r9;
1080 #ifdef _WIN64
1081 __ mov(saved_r15, r15); // r15 is callee saved and needs to be restored
1082 __ get_thread(r15_thread);
1083 assert(c_rarg0 == rcx && c_rarg1 == rdx && c_rarg2 == r8 && c_rarg3 == r9,
1084 "unexpected argument registers");
1085 __ movptr(Address(r15_thread, in_bytes(JavaThread::windows_saved_rdi_offset())), rdi);
1086 __ movptr(Address(r15_thread, in_bytes(JavaThread::windows_saved_rsi_offset())), rsi);
1087
1088 __ mov(rdi, rcx); // c_rarg0
1089 __ mov(rsi, rdx); // c_rarg1
1090 __ mov(rdx, r8); // c_rarg2
1091 #else
1092 assert(c_rarg0 == rdi && c_rarg1 == rsi && c_rarg2 == rdx && c_rarg3 == rcx,
1093 "unexpected argument registers");
1094 #endif
1095 DEBUG_ONLY(regs_in_thread = true;)
1096 }
1097
1098 void restore_arg_regs_using_thread() {
1099 assert(regs_in_thread, "wrong call to restore_arg_regs");
1100 const Register saved_r15 = r9;
1101 #ifdef _WIN64
1102 __ get_thread(r15_thread);
1103 __ movptr(rsi, Address(r15_thread, in_bytes(JavaThread::windows_saved_rsi_offset())));
1104 __ movptr(rdi, Address(r15_thread, in_bytes(JavaThread::windows_saved_rdi_offset())));
1105 __ mov(r15, saved_r15); // r15 is callee saved and needs to be restored
1106 #endif
1107 }
1108
1109 // Copy big chunks forward
1110 //
1111 // Inputs:
1112 // end_from - source arrays end address
1113 // end_to - destination array end address
1114 // qword_count - 64-bits element count, negative
1115 // to - scratch
1116 // L_copy_bytes - entry label
1117 // L_copy_8_bytes - exit label
1118 //
1119 void copy_bytes_forward(Register end_from, Register end_to,
1120 Register qword_count, Register to,
1121 Label& L_copy_bytes, Label& L_copy_8_bytes) {
1122 DEBUG_ONLY(__ stop("enter at entry label, not here"));
1123 Label L_loop;
1124 __ align(OptoLoopAlignment);
1125 if (UseUnalignedLoadStores) {
1126 Label L_end;
1127 // Copy 64-bytes per iteration
1128 if (UseAVX > 2) {
1129 Label L_loop_avx512, L_loop_avx2, L_32_byte_head, L_above_threshold, L_below_threshold;
1130
1131 __ BIND(L_copy_bytes);
1132 __ cmpptr(qword_count, (-1 * AVX3Threshold / 8));
1133 __ jccb(Assembler::less, L_above_threshold);
1134 __ jmpb(L_below_threshold);
1135
1136 __ bind(L_loop_avx512);
1137 __ evmovdqul(xmm0, Address(end_from, qword_count, Address::times_8, -56), Assembler::AVX_512bit);
1138 __ evmovdqul(Address(end_to, qword_count, Address::times_8, -56), xmm0, Assembler::AVX_512bit);
1139 __ bind(L_above_threshold);
1140 __ addptr(qword_count, 8);
1141 __ jcc(Assembler::lessEqual, L_loop_avx512);
1142 __ jmpb(L_32_byte_head);
1143
1144 __ bind(L_loop_avx2);
1145 __ vmovdqu(xmm0, Address(end_from, qword_count, Address::times_8, -56));
1146 __ vmovdqu(Address(end_to, qword_count, Address::times_8, -56), xmm0);
1147 __ vmovdqu(xmm1, Address(end_from, qword_count, Address::times_8, -24));
1148 __ vmovdqu(Address(end_to, qword_count, Address::times_8, -24), xmm1);
1149 __ bind(L_below_threshold);
1150 __ addptr(qword_count, 8);
1151 __ jcc(Assembler::lessEqual, L_loop_avx2);
1152
1153 __ bind(L_32_byte_head);
1154 __ subptr(qword_count, 4); // sub(8) and add(4)
1155 __ jccb(Assembler::greater, L_end);
1156 } else {
1157 __ BIND(L_loop);
1158 if (UseAVX == 2) {
1159 __ vmovdqu(xmm0, Address(end_from, qword_count, Address::times_8, -56));
1160 __ vmovdqu(Address(end_to, qword_count, Address::times_8, -56), xmm0);
1161 __ vmovdqu(xmm1, Address(end_from, qword_count, Address::times_8, -24));
1162 __ vmovdqu(Address(end_to, qword_count, Address::times_8, -24), xmm1);
1163 } else {
1164 __ movdqu(xmm0, Address(end_from, qword_count, Address::times_8, -56));
1165 __ movdqu(Address(end_to, qword_count, Address::times_8, -56), xmm0);
1166 __ movdqu(xmm1, Address(end_from, qword_count, Address::times_8, -40));
1167 __ movdqu(Address(end_to, qword_count, Address::times_8, -40), xmm1);
1168 __ movdqu(xmm2, Address(end_from, qword_count, Address::times_8, -24));
1169 __ movdqu(Address(end_to, qword_count, Address::times_8, -24), xmm2);
1170 __ movdqu(xmm3, Address(end_from, qword_count, Address::times_8, - 8));
1171 __ movdqu(Address(end_to, qword_count, Address::times_8, - 8), xmm3);
1172 }
1173
1174 __ BIND(L_copy_bytes);
1175 __ addptr(qword_count, 8);
1176 __ jcc(Assembler::lessEqual, L_loop);
1177 __ subptr(qword_count, 4); // sub(8) and add(4)
1178 __ jccb(Assembler::greater, L_end);
1179 }
1180 // Copy trailing 32 bytes
1181 if (UseAVX >= 2) {
1182 __ vmovdqu(xmm0, Address(end_from, qword_count, Address::times_8, -24));
1183 __ vmovdqu(Address(end_to, qword_count, Address::times_8, -24), xmm0);
1184 } else {
1185 __ movdqu(xmm0, Address(end_from, qword_count, Address::times_8, -24));
1186 __ movdqu(Address(end_to, qword_count, Address::times_8, -24), xmm0);
1187 __ movdqu(xmm1, Address(end_from, qword_count, Address::times_8, - 8));
1188 __ movdqu(Address(end_to, qword_count, Address::times_8, - 8), xmm1);
1189 }
1190 __ addptr(qword_count, 4);
1191 __ BIND(L_end);
1192 if (UseAVX >= 2) {
1193 // clean upper bits of YMM registers
1194 __ vpxor(xmm0, xmm0);
1195 __ vpxor(xmm1, xmm1);
1196 }
1197 } else {
1198 // Copy 32-bytes per iteration
1199 __ BIND(L_loop);
1200 __ movq(to, Address(end_from, qword_count, Address::times_8, -24));
1201 __ movq(Address(end_to, qword_count, Address::times_8, -24), to);
1202 __ movq(to, Address(end_from, qword_count, Address::times_8, -16));
1203 __ movq(Address(end_to, qword_count, Address::times_8, -16), to);
1204 __ movq(to, Address(end_from, qword_count, Address::times_8, - 8));
1205 __ movq(Address(end_to, qword_count, Address::times_8, - 8), to);
1206 __ movq(to, Address(end_from, qword_count, Address::times_8, - 0));
1207 __ movq(Address(end_to, qword_count, Address::times_8, - 0), to);
1208
1209 __ BIND(L_copy_bytes);
1210 __ addptr(qword_count, 4);
1211 __ jcc(Assembler::lessEqual, L_loop);
1212 }
1213 __ subptr(qword_count, 4);
1214 __ jcc(Assembler::less, L_copy_8_bytes); // Copy trailing qwords
1215 }
1216
1217 // Copy big chunks backward
1218 //
1219 // Inputs:
1220 // from - source arrays address
1221 // dest - destination array address
1222 // qword_count - 64-bits element count
1223 // to - scratch
1224 // L_copy_bytes - entry label
1225 // L_copy_8_bytes - exit label
1226 //
1227 void copy_bytes_backward(Register from, Register dest,
1228 Register qword_count, Register to,
1229 Label& L_copy_bytes, Label& L_copy_8_bytes) {
1230 DEBUG_ONLY(__ stop("enter at entry label, not here"));
1231 Label L_loop;
1232 __ align(OptoLoopAlignment);
1233 if (UseUnalignedLoadStores) {
1234 Label L_end;
1235 // Copy 64-bytes per iteration
1236 if (UseAVX > 2) {
1237 Label L_loop_avx512, L_loop_avx2, L_32_byte_head, L_above_threshold, L_below_threshold;
1238
1239 __ BIND(L_copy_bytes);
1240 __ cmpptr(qword_count, (AVX3Threshold / 8));
1241 __ jccb(Assembler::greater, L_above_threshold);
1242 __ jmpb(L_below_threshold);
1243
1244 __ BIND(L_loop_avx512);
1245 __ evmovdqul(xmm0, Address(from, qword_count, Address::times_8, 0), Assembler::AVX_512bit);
1246 __ evmovdqul(Address(dest, qword_count, Address::times_8, 0), xmm0, Assembler::AVX_512bit);
1247 __ bind(L_above_threshold);
1248 __ subptr(qword_count, 8);
1249 __ jcc(Assembler::greaterEqual, L_loop_avx512);
1250 __ jmpb(L_32_byte_head);
1251
1252 __ bind(L_loop_avx2);
1253 __ vmovdqu(xmm0, Address(from, qword_count, Address::times_8, 32));
1254 __ vmovdqu(Address(dest, qword_count, Address::times_8, 32), xmm0);
1255 __ vmovdqu(xmm1, Address(from, qword_count, Address::times_8, 0));
1256 __ vmovdqu(Address(dest, qword_count, Address::times_8, 0), xmm1);
1257 __ bind(L_below_threshold);
1258 __ subptr(qword_count, 8);
1259 __ jcc(Assembler::greaterEqual, L_loop_avx2);
1260
1261 __ bind(L_32_byte_head);
1262 __ addptr(qword_count, 4); // add(8) and sub(4)
1263 __ jccb(Assembler::less, L_end);
1264 } else {
1265 __ BIND(L_loop);
1266 if (UseAVX == 2) {
1267 __ vmovdqu(xmm0, Address(from, qword_count, Address::times_8, 32));
1268 __ vmovdqu(Address(dest, qword_count, Address::times_8, 32), xmm0);
1269 __ vmovdqu(xmm1, Address(from, qword_count, Address::times_8, 0));
1270 __ vmovdqu(Address(dest, qword_count, Address::times_8, 0), xmm1);
1271 } else {
1272 __ movdqu(xmm0, Address(from, qword_count, Address::times_8, 48));
1273 __ movdqu(Address(dest, qword_count, Address::times_8, 48), xmm0);
1274 __ movdqu(xmm1, Address(from, qword_count, Address::times_8, 32));
1275 __ movdqu(Address(dest, qword_count, Address::times_8, 32), xmm1);
1276 __ movdqu(xmm2, Address(from, qword_count, Address::times_8, 16));
1277 __ movdqu(Address(dest, qword_count, Address::times_8, 16), xmm2);
1278 __ movdqu(xmm3, Address(from, qword_count, Address::times_8, 0));
1279 __ movdqu(Address(dest, qword_count, Address::times_8, 0), xmm3);
1280 }
1281
1282 __ BIND(L_copy_bytes);
1283 __ subptr(qword_count, 8);
1284 __ jcc(Assembler::greaterEqual, L_loop);
1285
1286 __ addptr(qword_count, 4); // add(8) and sub(4)
1287 __ jccb(Assembler::less, L_end);
1288 }
1289 // Copy trailing 32 bytes
1290 if (UseAVX >= 2) {
1291 __ vmovdqu(xmm0, Address(from, qword_count, Address::times_8, 0));
1292 __ vmovdqu(Address(dest, qword_count, Address::times_8, 0), xmm0);
1293 } else {
1294 __ movdqu(xmm0, Address(from, qword_count, Address::times_8, 16));
1295 __ movdqu(Address(dest, qword_count, Address::times_8, 16), xmm0);
1296 __ movdqu(xmm1, Address(from, qword_count, Address::times_8, 0));
1297 __ movdqu(Address(dest, qword_count, Address::times_8, 0), xmm1);
1298 }
1299 __ subptr(qword_count, 4);
1300 __ BIND(L_end);
1301 if (UseAVX >= 2) {
1302 // clean upper bits of YMM registers
1303 __ vpxor(xmm0, xmm0);
1304 __ vpxor(xmm1, xmm1);
1305 }
1306 } else {
1307 // Copy 32-bytes per iteration
1308 __ BIND(L_loop);
1309 __ movq(to, Address(from, qword_count, Address::times_8, 24));
1310 __ movq(Address(dest, qword_count, Address::times_8, 24), to);
1311 __ movq(to, Address(from, qword_count, Address::times_8, 16));
1312 __ movq(Address(dest, qword_count, Address::times_8, 16), to);
1313 __ movq(to, Address(from, qword_count, Address::times_8, 8));
1314 __ movq(Address(dest, qword_count, Address::times_8, 8), to);
1315 __ movq(to, Address(from, qword_count, Address::times_8, 0));
1316 __ movq(Address(dest, qword_count, Address::times_8, 0), to);
1317
1318 __ BIND(L_copy_bytes);
1319 __ subptr(qword_count, 4);
1320 __ jcc(Assembler::greaterEqual, L_loop);
1321 }
1322 __ addptr(qword_count, 4);
1323 __ jcc(Assembler::greater, L_copy_8_bytes); // Copy trailing qwords
1324 }
1325
1326 // Arguments:
1327 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1328 // ignored
1329 // name - stub name string
1330 //
1331 // Inputs:
1332 // c_rarg0 - source array address
1333 // c_rarg1 - destination array address
1334 // c_rarg2 - element count, treated as ssize_t, can be zero
1335 //
1336 // If 'from' and/or 'to' are aligned on 4-, 2-, or 1-byte boundaries,
1337 // we let the hardware handle it. The one to eight bytes within words,
1338 // dwords or qwords that span cache line boundaries will still be loaded
1339 // and stored atomically.
1340 //
1341 // Side Effects:
1342 // disjoint_byte_copy_entry is set to the no-overlap entry point
1343 // used by generate_conjoint_byte_copy().
1344 //
1345 address generate_disjoint_byte_copy(bool aligned, address* entry, const char *name) {
1346 __ align(CodeEntryAlignment);
1347 StubCodeMark mark(this, "StubRoutines", name);
1348 address start = __ pc();
1349
1350 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes, L_copy_2_bytes;
1351 Label L_copy_byte, L_exit;
1352 const Register from = rdi; // source array address
1353 const Register to = rsi; // destination array address
1354 const Register count = rdx; // elements count
1355 const Register byte_count = rcx;
1356 const Register qword_count = count;
1357 const Register end_from = from; // source array end address
1358 const Register end_to = to; // destination array end address
1359 // End pointers are inclusive, and if count is not zero they point
1360 // to the last unit copied: end_to[0] := end_from[0]
1361
1362 __ enter(); // required for proper stackwalking of RuntimeStub frame
1363 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1364
1365 if (entry != NULL) {
1366 *entry = __ pc();
1367 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1368 BLOCK_COMMENT("Entry:");
1369 }
1370
1371 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1372 // r9 and r10 may be used to save non-volatile registers
1373
1374 {
1375 // UnsafeCopyMemory page error: continue after ucm
1376 UnsafeCopyMemoryMark ucmm(this, !aligned, true);
1377 // 'from', 'to' and 'count' are now valid
1378 __ movptr(byte_count, count);
1379 __ shrptr(count, 3); // count => qword_count
1380
1381 // Copy from low to high addresses. Use 'to' as scratch.
1382 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
1383 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
1384 __ negptr(qword_count); // make the count negative
1385 __ jmp(L_copy_bytes);
1386
1387 // Copy trailing qwords
1388 __ BIND(L_copy_8_bytes);
1389 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
1390 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
1391 __ increment(qword_count);
1392 __ jcc(Assembler::notZero, L_copy_8_bytes);
1393
1394 // Check for and copy trailing dword
1395 __ BIND(L_copy_4_bytes);
1396 __ testl(byte_count, 4);
1397 __ jccb(Assembler::zero, L_copy_2_bytes);
1398 __ movl(rax, Address(end_from, 8));
1399 __ movl(Address(end_to, 8), rax);
1400
1401 __ addptr(end_from, 4);
1402 __ addptr(end_to, 4);
1403
1404 // Check for and copy trailing word
1405 __ BIND(L_copy_2_bytes);
1406 __ testl(byte_count, 2);
1407 __ jccb(Assembler::zero, L_copy_byte);
1408 __ movw(rax, Address(end_from, 8));
1409 __ movw(Address(end_to, 8), rax);
1410
1411 __ addptr(end_from, 2);
1412 __ addptr(end_to, 2);
1413
1414 // Check for and copy trailing byte
1415 __ BIND(L_copy_byte);
1416 __ testl(byte_count, 1);
1417 __ jccb(Assembler::zero, L_exit);
1418 __ movb(rax, Address(end_from, 8));
1419 __ movb(Address(end_to, 8), rax);
1420 }
1421 __ BIND(L_exit);
1422 address ucme_exit_pc = __ pc();
1423 restore_arg_regs();
1424 inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); // Update counter after rscratch1 is free
1425 __ xorptr(rax, rax); // return 0
1426 __ vzeroupper();
1427 __ leave(); // required for proper stackwalking of RuntimeStub frame
1428 __ ret(0);
1429
1430 {
1431 UnsafeCopyMemoryMark ucmm(this, !aligned, false, ucme_exit_pc);
1432 // Copy in multi-bytes chunks
1433 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1434 __ jmp(L_copy_4_bytes);
1435 }
1436 return start;
1437 }
1438
1439 // Arguments:
1440 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1441 // ignored
1442 // name - stub name string
1443 //
1444 // Inputs:
1445 // c_rarg0 - source array address
1446 // c_rarg1 - destination array address
1447 // c_rarg2 - element count, treated as ssize_t, can be zero
1448 //
1449 // If 'from' and/or 'to' are aligned on 4-, 2-, or 1-byte boundaries,
1450 // we let the hardware handle it. The one to eight bytes within words,
1451 // dwords or qwords that span cache line boundaries will still be loaded
1452 // and stored atomically.
1453 //
1454 address generate_conjoint_byte_copy(bool aligned, address nooverlap_target,
1455 address* entry, const char *name) {
1456 __ align(CodeEntryAlignment);
1457 StubCodeMark mark(this, "StubRoutines", name);
1458 address start = __ pc();
1459
1460 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes, L_copy_2_bytes;
1461 const Register from = rdi; // source array address
1462 const Register to = rsi; // destination array address
1463 const Register count = rdx; // elements count
1464 const Register byte_count = rcx;
1465 const Register qword_count = count;
1466
1467 __ enter(); // required for proper stackwalking of RuntimeStub frame
1468 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1469
1470 if (entry != NULL) {
1471 *entry = __ pc();
1472 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1473 BLOCK_COMMENT("Entry:");
1474 }
1475
1476 array_overlap_test(nooverlap_target, Address::times_1);
1477 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1478 // r9 and r10 may be used to save non-volatile registers
1479
1480 {
1481 // UnsafeCopyMemory page error: continue after ucm
1482 UnsafeCopyMemoryMark ucmm(this, !aligned, true);
1483 // 'from', 'to' and 'count' are now valid
1484 __ movptr(byte_count, count);
1485 __ shrptr(count, 3); // count => qword_count
1486
1487 // Copy from high to low addresses.
1488
1489 // Check for and copy trailing byte
1490 __ testl(byte_count, 1);
1491 __ jcc(Assembler::zero, L_copy_2_bytes);
1492 __ movb(rax, Address(from, byte_count, Address::times_1, -1));
1493 __ movb(Address(to, byte_count, Address::times_1, -1), rax);
1494 __ decrement(byte_count); // Adjust for possible trailing word
1495
1496 // Check for and copy trailing word
1497 __ BIND(L_copy_2_bytes);
1498 __ testl(byte_count, 2);
1499 __ jcc(Assembler::zero, L_copy_4_bytes);
1500 __ movw(rax, Address(from, byte_count, Address::times_1, -2));
1501 __ movw(Address(to, byte_count, Address::times_1, -2), rax);
1502
1503 // Check for and copy trailing dword
1504 __ BIND(L_copy_4_bytes);
1505 __ testl(byte_count, 4);
1506 __ jcc(Assembler::zero, L_copy_bytes);
1507 __ movl(rax, Address(from, qword_count, Address::times_8));
1508 __ movl(Address(to, qword_count, Address::times_8), rax);
1509 __ jmp(L_copy_bytes);
1510
1511 // Copy trailing qwords
1512 __ BIND(L_copy_8_bytes);
1513 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
1514 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
1515 __ decrement(qword_count);
1516 __ jcc(Assembler::notZero, L_copy_8_bytes);
1517 }
1518 restore_arg_regs();
1519 inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); // Update counter after rscratch1 is free
1520 __ xorptr(rax, rax); // return 0
1521 __ vzeroupper();
1522 __ leave(); // required for proper stackwalking of RuntimeStub frame
1523 __ ret(0);
1524
1525 {
1526 // UnsafeCopyMemory page error: continue after ucm
1527 UnsafeCopyMemoryMark ucmm(this, !aligned, true);
1528 // Copy in multi-bytes chunks
1529 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1530 }
1531 restore_arg_regs();
1532 inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); // Update counter after rscratch1 is free
1533 __ xorptr(rax, rax); // return 0
1534 __ vzeroupper();
1535 __ leave(); // required for proper stackwalking of RuntimeStub frame
1536 __ ret(0);
1537
1538 return start;
1539 }
1540
1541 // Arguments:
1542 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1543 // ignored
1544 // name - stub name string
1545 //
1546 // Inputs:
1547 // c_rarg0 - source array address
1548 // c_rarg1 - destination array address
1549 // c_rarg2 - element count, treated as ssize_t, can be zero
1550 //
1551 // If 'from' and/or 'to' are aligned on 4- or 2-byte boundaries, we
1552 // let the hardware handle it. The two or four words within dwords
1553 // or qwords that span cache line boundaries will still be loaded
1554 // and stored atomically.
1555 //
1556 // Side Effects:
1557 // disjoint_short_copy_entry is set to the no-overlap entry point
1558 // used by generate_conjoint_short_copy().
1559 //
1560 address generate_disjoint_short_copy(bool aligned, address *entry, const char *name) {
1561 __ align(CodeEntryAlignment);
1562 StubCodeMark mark(this, "StubRoutines", name);
1563 address start = __ pc();
1564
1565 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes,L_copy_2_bytes,L_exit;
1566 const Register from = rdi; // source array address
1567 const Register to = rsi; // destination array address
1568 const Register count = rdx; // elements count
1569 const Register word_count = rcx;
1570 const Register qword_count = count;
1571 const Register end_from = from; // source array end address
1572 const Register end_to = to; // destination array end address
1573 // End pointers are inclusive, and if count is not zero they point
1574 // to the last unit copied: end_to[0] := end_from[0]
1575
1576 __ enter(); // required for proper stackwalking of RuntimeStub frame
1577 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1578
1579 if (entry != NULL) {
1580 *entry = __ pc();
1581 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1582 BLOCK_COMMENT("Entry:");
1583 }
1584
1585 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1586 // r9 and r10 may be used to save non-volatile registers
1587
1588 {
1589 // UnsafeCopyMemory page error: continue after ucm
1590 UnsafeCopyMemoryMark ucmm(this, !aligned, true);
1591 // 'from', 'to' and 'count' are now valid
1592 __ movptr(word_count, count);
1593 __ shrptr(count, 2); // count => qword_count
1594
1595 // Copy from low to high addresses. Use 'to' as scratch.
1596 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
1597 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
1598 __ negptr(qword_count);
1599 __ jmp(L_copy_bytes);
1600
1601 // Copy trailing qwords
1602 __ BIND(L_copy_8_bytes);
1603 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
1604 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
1605 __ increment(qword_count);
1606 __ jcc(Assembler::notZero, L_copy_8_bytes);
1607
1608 // Original 'dest' is trashed, so we can't use it as a
1609 // base register for a possible trailing word copy
1610
1611 // Check for and copy trailing dword
1612 __ BIND(L_copy_4_bytes);
1613 __ testl(word_count, 2);
1614 __ jccb(Assembler::zero, L_copy_2_bytes);
1615 __ movl(rax, Address(end_from, 8));
1616 __ movl(Address(end_to, 8), rax);
1617
1618 __ addptr(end_from, 4);
1619 __ addptr(end_to, 4);
1620
1621 // Check for and copy trailing word
1622 __ BIND(L_copy_2_bytes);
1623 __ testl(word_count, 1);
1624 __ jccb(Assembler::zero, L_exit);
1625 __ movw(rax, Address(end_from, 8));
1626 __ movw(Address(end_to, 8), rax);
1627 }
1628 __ BIND(L_exit);
1629 address ucme_exit_pc = __ pc();
1630 restore_arg_regs();
1631 inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); // Update counter after rscratch1 is free
1632 __ xorptr(rax, rax); // return 0
1633 __ vzeroupper();
1634 __ leave(); // required for proper stackwalking of RuntimeStub frame
1635 __ ret(0);
1636
1637 {
1638 UnsafeCopyMemoryMark ucmm(this, !aligned, false, ucme_exit_pc);
1639 // Copy in multi-bytes chunks
1640 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1641 __ jmp(L_copy_4_bytes);
1642 }
1643
1644 return start;
1645 }
1646
1647 address generate_fill(BasicType t, bool aligned, const char *name) {
1648 __ align(CodeEntryAlignment);
1649 StubCodeMark mark(this, "StubRoutines", name);
1650 address start = __ pc();
1651
1652 BLOCK_COMMENT("Entry:");
1653
1654 const Register to = c_rarg0; // source array address
1655 const Register value = c_rarg1; // value
1656 const Register count = c_rarg2; // elements count
1657
1658 __ enter(); // required for proper stackwalking of RuntimeStub frame
1659
1660 __ generate_fill(t, aligned, to, value, count, rax, xmm0);
1661
1662 __ vzeroupper();
1663 __ leave(); // required for proper stackwalking of RuntimeStub frame
1664 __ ret(0);
1665 return start;
1666 }
1667
1668 // Arguments:
1669 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1670 // ignored
1671 // name - stub name string
1672 //
1673 // Inputs:
1674 // c_rarg0 - source array address
1675 // c_rarg1 - destination array address
1676 // c_rarg2 - element count, treated as ssize_t, can be zero
1677 //
1678 // If 'from' and/or 'to' are aligned on 4- or 2-byte boundaries, we
1679 // let the hardware handle it. The two or four words within dwords
1680 // or qwords that span cache line boundaries will still be loaded
1681 // and stored atomically.
1682 //
1683 address generate_conjoint_short_copy(bool aligned, address nooverlap_target,
1684 address *entry, const char *name) {
1685 __ align(CodeEntryAlignment);
1686 StubCodeMark mark(this, "StubRoutines", name);
1687 address start = __ pc();
1688
1689 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes;
1690 const Register from = rdi; // source array address
1691 const Register to = rsi; // destination array address
1692 const Register count = rdx; // elements count
1693 const Register word_count = rcx;
1694 const Register qword_count = count;
1695
1696 __ enter(); // required for proper stackwalking of RuntimeStub frame
1697 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1698
1699 if (entry != NULL) {
1700 *entry = __ pc();
1701 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1702 BLOCK_COMMENT("Entry:");
1703 }
1704
1705 array_overlap_test(nooverlap_target, Address::times_2);
1706 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1707 // r9 and r10 may be used to save non-volatile registers
1708
1709 {
1710 // UnsafeCopyMemory page error: continue after ucm
1711 UnsafeCopyMemoryMark ucmm(this, !aligned, true);
1712 // 'from', 'to' and 'count' are now valid
1713 __ movptr(word_count, count);
1714 __ shrptr(count, 2); // count => qword_count
1715
1716 // Copy from high to low addresses. Use 'to' as scratch.
1717
1718 // Check for and copy trailing word
1719 __ testl(word_count, 1);
1720 __ jccb(Assembler::zero, L_copy_4_bytes);
1721 __ movw(rax, Address(from, word_count, Address::times_2, -2));
1722 __ movw(Address(to, word_count, Address::times_2, -2), rax);
1723
1724 // Check for and copy trailing dword
1725 __ BIND(L_copy_4_bytes);
1726 __ testl(word_count, 2);
1727 __ jcc(Assembler::zero, L_copy_bytes);
1728 __ movl(rax, Address(from, qword_count, Address::times_8));
1729 __ movl(Address(to, qword_count, Address::times_8), rax);
1730 __ jmp(L_copy_bytes);
1731
1732 // Copy trailing qwords
1733 __ BIND(L_copy_8_bytes);
1734 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
1735 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
1736 __ decrement(qword_count);
1737 __ jcc(Assembler::notZero, L_copy_8_bytes);
1738 }
1739 restore_arg_regs();
1740 inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); // Update counter after rscratch1 is free
1741 __ xorptr(rax, rax); // return 0
1742 __ vzeroupper();
1743 __ leave(); // required for proper stackwalking of RuntimeStub frame
1744 __ ret(0);
1745
1746 {
1747 // UnsafeCopyMemory page error: continue after ucm
1748 UnsafeCopyMemoryMark ucmm(this, !aligned, true);
1749 // Copy in multi-bytes chunks
1750 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1751 }
1752 restore_arg_regs();
1753 inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); // Update counter after rscratch1 is free
1754 __ xorptr(rax, rax); // return 0
1755 __ vzeroupper();
1756 __ leave(); // required for proper stackwalking of RuntimeStub frame
1757 __ ret(0);
1758
1759 return start;
1760 }
1761
1762 // Arguments:
1763 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1764 // ignored
1765 // is_oop - true => oop array, so generate store check code
1766 // name - stub name string
1767 //
1768 // Inputs:
1769 // c_rarg0 - source array address
1770 // c_rarg1 - destination array address
1771 // c_rarg2 - element count, treated as ssize_t, can be zero
1772 //
1773 // If 'from' and/or 'to' are aligned on 4-byte boundaries, we let
1774 // the hardware handle it. The two dwords within qwords that span
1775 // cache line boundaries will still be loaded and stored atomicly.
1776 //
1777 // Side Effects:
1778 // disjoint_int_copy_entry is set to the no-overlap entry point
1779 // used by generate_conjoint_int_oop_copy().
1780 //
1781 address generate_disjoint_int_oop_copy(bool aligned, bool is_oop, address* entry,
1782 const char *name, bool dest_uninitialized = false) {
1783 __ align(CodeEntryAlignment);
1784 StubCodeMark mark(this, "StubRoutines", name);
1785 address start = __ pc();
1786
1787 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes, L_exit;
1788 const Register from = rdi; // source array address
1789 const Register to = rsi; // destination array address
1790 const Register count = rdx; // elements count
1791 const Register dword_count = rcx;
1792 const Register qword_count = count;
1793 const Register end_from = from; // source array end address
1794 const Register end_to = to; // destination array end address
1795 // End pointers are inclusive, and if count is not zero they point
1796 // to the last unit copied: end_to[0] := end_from[0]
1797
1798 __ enter(); // required for proper stackwalking of RuntimeStub frame
1799 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1800
1801 if (entry != NULL) {
1802 *entry = __ pc();
1803 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1804 BLOCK_COMMENT("Entry:");
1805 }
1806
1807 setup_arg_regs_using_thread(); // from => rdi, to => rsi, count => rdx
1808 // r9 is used to save r15_thread
1809
1810 DecoratorSet decorators = IN_HEAP | IS_ARRAY | ARRAYCOPY_DISJOINT;
1811 if (dest_uninitialized) {
1812 decorators |= IS_DEST_UNINITIALIZED;
1813 }
1814 if (aligned) {
1815 decorators |= ARRAYCOPY_ALIGNED;
1816 }
1817
1818 BasicType type = is_oop ? T_OBJECT : T_INT;
1819 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler();
1820 bs->arraycopy_prologue(_masm, decorators, type, from, to, count);
1821
1822 {
1823 // UnsafeCopyMemory page error: continue after ucm
1824 UnsafeCopyMemoryMark ucmm(this, !is_oop && !aligned, true);
1825 // 'from', 'to' and 'count' are now valid
1826 __ movptr(dword_count, count);
1827 __ shrptr(count, 1); // count => qword_count
1828
1829 // Copy from low to high addresses. Use 'to' as scratch.
1830 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
1831 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
1832 __ negptr(qword_count);
1833 __ jmp(L_copy_bytes);
1834
1835 // Copy trailing qwords
1836 __ BIND(L_copy_8_bytes);
1837 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
1838 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
1839 __ increment(qword_count);
1840 __ jcc(Assembler::notZero, L_copy_8_bytes);
1841
1842 // Check for and copy trailing dword
1843 __ BIND(L_copy_4_bytes);
1844 __ testl(dword_count, 1); // Only byte test since the value is 0 or 1
1845 __ jccb(Assembler::zero, L_exit);
1846 __ movl(rax, Address(end_from, 8));
1847 __ movl(Address(end_to, 8), rax);
1848 }
1849 __ BIND(L_exit);
1850 address ucme_exit_pc = __ pc();
1851 bs->arraycopy_epilogue(_masm, decorators, type, from, to, dword_count);
1852 restore_arg_regs_using_thread();
1853 inc_counter_np(SharedRuntime::_jint_array_copy_ctr); // Update counter after rscratch1 is free
1854 __ vzeroupper();
1855 __ xorptr(rax, rax); // return 0
1856 __ leave(); // required for proper stackwalking of RuntimeStub frame
1857 __ ret(0);
1858
1859 {
1860 UnsafeCopyMemoryMark ucmm(this, !is_oop && !aligned, false, ucme_exit_pc);
1861 // Copy in multi-bytes chunks
1862 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1863 __ jmp(L_copy_4_bytes);
1864 }
1865
1866 return start;
1867 }
1868
1869 // Arguments:
1870 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1871 // ignored
1872 // is_oop - true => oop array, so generate store check code
1873 // name - stub name string
1874 //
1875 // Inputs:
1876 // c_rarg0 - source array address
1877 // c_rarg1 - destination array address
1878 // c_rarg2 - element count, treated as ssize_t, can be zero
1879 //
1880 // If 'from' and/or 'to' are aligned on 4-byte boundaries, we let
1881 // the hardware handle it. The two dwords within qwords that span
1882 // cache line boundaries will still be loaded and stored atomicly.
1883 //
1884 address generate_conjoint_int_oop_copy(bool aligned, bool is_oop, address nooverlap_target,
1885 address *entry, const char *name,
1886 bool dest_uninitialized = false) {
1887 __ align(CodeEntryAlignment);
1888 StubCodeMark mark(this, "StubRoutines", name);
1889 address start = __ pc();
1890
1891 Label L_copy_bytes, L_copy_8_bytes, L_exit;
1892 const Register from = rdi; // source array address
1893 const Register to = rsi; // destination array address
1894 const Register count = rdx; // elements count
1895 const Register dword_count = rcx;
1896 const Register qword_count = count;
1897
1898 __ enter(); // required for proper stackwalking of RuntimeStub frame
1899 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1900
1901 if (entry != NULL) {
1902 *entry = __ pc();
1903 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1904 BLOCK_COMMENT("Entry:");
1905 }
1906
1907 array_overlap_test(nooverlap_target, Address::times_4);
1908 setup_arg_regs_using_thread(); // from => rdi, to => rsi, count => rdx
1909 // r9 is used to save r15_thread
1910
1911 DecoratorSet decorators = IN_HEAP | IS_ARRAY;
1912 if (dest_uninitialized) {
1913 decorators |= IS_DEST_UNINITIALIZED;
1914 }
1915 if (aligned) {
1916 decorators |= ARRAYCOPY_ALIGNED;
1917 }
1918
1919 BasicType type = is_oop ? T_OBJECT : T_INT;
1920 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler();
1921 // no registers are destroyed by this call
1922 bs->arraycopy_prologue(_masm, decorators, type, from, to, count);
1923
1924 assert_clean_int(count, rax); // Make sure 'count' is clean int.
1925 {
1926 // UnsafeCopyMemory page error: continue after ucm
1927 UnsafeCopyMemoryMark ucmm(this, !is_oop && !aligned, true);
1928 // 'from', 'to' and 'count' are now valid
1929 __ movptr(dword_count, count);
1930 __ shrptr(count, 1); // count => qword_count
1931
1932 // Copy from high to low addresses. Use 'to' as scratch.
1933
1934 // Check for and copy trailing dword
1935 __ testl(dword_count, 1);
1936 __ jcc(Assembler::zero, L_copy_bytes);
1937 __ movl(rax, Address(from, dword_count, Address::times_4, -4));
1938 __ movl(Address(to, dword_count, Address::times_4, -4), rax);
1939 __ jmp(L_copy_bytes);
1940
1941 // Copy trailing qwords
1942 __ BIND(L_copy_8_bytes);
1943 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
1944 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
1945 __ decrement(qword_count);
1946 __ jcc(Assembler::notZero, L_copy_8_bytes);
1947 }
1948 if (is_oop) {
1949 __ jmp(L_exit);
1950 }
1951 restore_arg_regs_using_thread();
1952 inc_counter_np(SharedRuntime::_jint_array_copy_ctr); // Update counter after rscratch1 is free
1953 __ xorptr(rax, rax); // return 0
1954 __ vzeroupper();
1955 __ leave(); // required for proper stackwalking of RuntimeStub frame
1956 __ ret(0);
1957
1958 {
1959 // UnsafeCopyMemory page error: continue after ucm
1960 UnsafeCopyMemoryMark ucmm(this, !is_oop && !aligned, true);
1961 // Copy in multi-bytes chunks
1962 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1963 }
1964
1965 __ BIND(L_exit);
1966 bs->arraycopy_epilogue(_masm, decorators, type, from, to, dword_count);
1967 restore_arg_regs_using_thread();
1968 inc_counter_np(SharedRuntime::_jint_array_copy_ctr); // Update counter after rscratch1 is free
1969 __ xorptr(rax, rax); // return 0
1970 __ vzeroupper();
1971 __ leave(); // required for proper stackwalking of RuntimeStub frame
1972 __ ret(0);
1973
1974 return start;
1975 }
1976
1977 // Arguments:
1978 // aligned - true => Input and output aligned on a HeapWord boundary == 8 bytes
1979 // ignored
1980 // is_oop - true => oop array, so generate store check code
1981 // name - stub name string
1982 //
1983 // Inputs:
1984 // c_rarg0 - source array address
1985 // c_rarg1 - destination array address
1986 // c_rarg2 - element count, treated as ssize_t, can be zero
1987 //
1988 // Side Effects:
1989 // disjoint_oop_copy_entry or disjoint_long_copy_entry is set to the
1990 // no-overlap entry point used by generate_conjoint_long_oop_copy().
1991 //
1992 address generate_disjoint_long_oop_copy(bool aligned, bool is_oop, address *entry,
1993 const char *name, bool dest_uninitialized = false) {
1994 __ align(CodeEntryAlignment);
1995 StubCodeMark mark(this, "StubRoutines", name);
1996 address start = __ pc();
1997
1998 Label L_copy_bytes, L_copy_8_bytes, L_exit;
1999 const Register from = rdi; // source array address
2000 const Register to = rsi; // destination array address
2001 const Register qword_count = rdx; // elements count
2002 const Register end_from = from; // source array end address
2003 const Register end_to = rcx; // destination array end address
2004 const Register saved_count = r11;
2005 // End pointers are inclusive, and if count is not zero they point
2006 // to the last unit copied: end_to[0] := end_from[0]
2007
2008 __ enter(); // required for proper stackwalking of RuntimeStub frame
2009 // Save no-overlap entry point for generate_conjoint_long_oop_copy()
2010 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
2011
2012 if (entry != NULL) {
2013 *entry = __ pc();
2014 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
2015 BLOCK_COMMENT("Entry:");
2016 }
2017
2018 setup_arg_regs_using_thread(); // from => rdi, to => rsi, count => rdx
2019 // r9 is used to save r15_thread
2020 // 'from', 'to' and 'qword_count' are now valid
2021
2022 DecoratorSet decorators = IN_HEAP | IS_ARRAY | ARRAYCOPY_DISJOINT;
2023 if (dest_uninitialized) {
2024 decorators |= IS_DEST_UNINITIALIZED;
2025 }
2026 if (aligned) {
2027 decorators |= ARRAYCOPY_ALIGNED;
2028 }
2029
2030 BasicType type = is_oop ? T_OBJECT : T_LONG;
2031 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler();
2032 bs->arraycopy_prologue(_masm, decorators, type, from, to, qword_count);
2033 {
2034 // UnsafeCopyMemory page error: continue after ucm
2035 UnsafeCopyMemoryMark ucmm(this, !is_oop && !aligned, true);
2036
2037 // Copy from low to high addresses. Use 'to' as scratch.
2038 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
2039 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
2040 __ negptr(qword_count);
2041 __ jmp(L_copy_bytes);
2042
2043 // Copy trailing qwords
2044 __ BIND(L_copy_8_bytes);
2045 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
2046 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
2047 __ increment(qword_count);
2048 __ jcc(Assembler::notZero, L_copy_8_bytes);
2049 }
2050 if (is_oop) {
2051 __ jmp(L_exit);
2052 } else {
2053 restore_arg_regs_using_thread();
2054 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2055 __ xorptr(rax, rax); // return 0
2056 __ vzeroupper();
2057 __ leave(); // required for proper stackwalking of RuntimeStub frame
2058 __ ret(0);
2059 }
2060
2061 {
2062 // UnsafeCopyMemory page error: continue after ucm
2063 UnsafeCopyMemoryMark ucmm(this, !is_oop && !aligned, true);
2064 // Copy in multi-bytes chunks
2065 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
2066 }
2067
2068 __ BIND(L_exit);
2069 bs->arraycopy_epilogue(_masm, decorators, type, from, to, qword_count);
2070 restore_arg_regs_using_thread();
2071 if (is_oop) {
2072 inc_counter_np(SharedRuntime::_oop_array_copy_ctr); // Update counter after rscratch1 is free
2073 } else {
2074 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2075 }
2076 __ vzeroupper();
2077 __ xorptr(rax, rax); // return 0
2078 __ leave(); // required for proper stackwalking of RuntimeStub frame
2079 __ ret(0);
2080
2081 return start;
2082 }
2083
2084 // Arguments:
2085 // aligned - true => Input and output aligned on a HeapWord boundary == 8 bytes
2086 // ignored
2087 // is_oop - true => oop array, so generate store check code
2088 // name - stub name string
2089 //
2090 // Inputs:
2091 // c_rarg0 - source array address
2092 // c_rarg1 - destination array address
2093 // c_rarg2 - element count, treated as ssize_t, can be zero
2094 //
2095 address generate_conjoint_long_oop_copy(bool aligned, bool is_oop,
2096 address nooverlap_target, address *entry,
2097 const char *name, bool dest_uninitialized = false) {
2098 __ align(CodeEntryAlignment);
2099 StubCodeMark mark(this, "StubRoutines", name);
2100 address start = __ pc();
2101
2102 Label L_copy_bytes, L_copy_8_bytes, L_exit;
2103 const Register from = rdi; // source array address
2104 const Register to = rsi; // destination array address
2105 const Register qword_count = rdx; // elements count
2106 const Register saved_count = rcx;
2107
2108 __ enter(); // required for proper stackwalking of RuntimeStub frame
2109 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
2110
2111 if (entry != NULL) {
2112 *entry = __ pc();
2113 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
2114 BLOCK_COMMENT("Entry:");
2115 }
2116
2117 array_overlap_test(nooverlap_target, Address::times_8);
2118 setup_arg_regs_using_thread(); // from => rdi, to => rsi, count => rdx
2119 // r9 is used to save r15_thread
2120 // 'from', 'to' and 'qword_count' are now valid
2121
2122 DecoratorSet decorators = IN_HEAP | IS_ARRAY;
2123 if (dest_uninitialized) {
2124 decorators |= IS_DEST_UNINITIALIZED;
2125 }
2126 if (aligned) {
2127 decorators |= ARRAYCOPY_ALIGNED;
2128 }
2129
2130 BasicType type = is_oop ? T_OBJECT : T_LONG;
2131 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler();
2132 bs->arraycopy_prologue(_masm, decorators, type, from, to, qword_count);
2133 {
2134 // UnsafeCopyMemory page error: continue after ucm
2135 UnsafeCopyMemoryMark ucmm(this, !is_oop && !aligned, true);
2136
2137 __ jmp(L_copy_bytes);
2138
2139 // Copy trailing qwords
2140 __ BIND(L_copy_8_bytes);
2141 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
2142 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
2143 __ decrement(qword_count);
2144 __ jcc(Assembler::notZero, L_copy_8_bytes);
2145 }
2146 if (is_oop) {
2147 __ jmp(L_exit);
2148 } else {
2149 restore_arg_regs_using_thread();
2150 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2151 __ xorptr(rax, rax); // return 0
2152 __ vzeroupper();
2153 __ leave(); // required for proper stackwalking of RuntimeStub frame
2154 __ ret(0);
2155 }
2156 {
2157 // UnsafeCopyMemory page error: continue after ucm
2158 UnsafeCopyMemoryMark ucmm(this, !is_oop && !aligned, true);
2159
2160 // Copy in multi-bytes chunks
2161 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
2162 }
2163 __ BIND(L_exit);
2164 bs->arraycopy_epilogue(_masm, decorators, type, from, to, qword_count);
2165 restore_arg_regs_using_thread();
2166 if (is_oop) {
2167 inc_counter_np(SharedRuntime::_oop_array_copy_ctr); // Update counter after rscratch1 is free
2168 } else {
2169 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2170 }
2171 __ vzeroupper();
2172 __ xorptr(rax, rax); // return 0
2173 __ leave(); // required for proper stackwalking of RuntimeStub frame
2174 __ ret(0);
2175
2176 return start;
2177 }
2178
2179
2180 // Helper for generating a dynamic type check.
2181 // Smashes no registers.
2182 void generate_type_check(Register sub_klass,
2183 Register super_check_offset,
2184 Register super_klass,
2185 Label& L_success) {
2186 assert_different_registers(sub_klass, super_check_offset, super_klass);
2187
2188 BLOCK_COMMENT("type_check:");
2189
2190 Label L_miss;
2191
2192 __ check_klass_subtype_fast_path(sub_klass, super_klass, noreg, &L_success, &L_miss, NULL,
2193 super_check_offset);
2194 __ check_klass_subtype_slow_path(sub_klass, super_klass, noreg, noreg, &L_success, NULL);
2195
2196 // Fall through on failure!
2197 __ BIND(L_miss);
2198 }
2199
2200 //
2201 // Generate checkcasting array copy stub
2202 //
2203 // Input:
2204 // c_rarg0 - source array address
2205 // c_rarg1 - destination array address
2206 // c_rarg2 - element count, treated as ssize_t, can be zero
2207 // c_rarg3 - size_t ckoff (super_check_offset)
2208 // not Win64
2209 // c_rarg4 - oop ckval (super_klass)
2210 // Win64
2211 // rsp+40 - oop ckval (super_klass)
2212 //
2213 // Output:
2214 // rax == 0 - success
2215 // rax == -1^K - failure, where K is partial transfer count
2216 //
2217 address generate_checkcast_copy(const char *name, address *entry,
2218 bool dest_uninitialized = false) {
2219
2220 Label L_load_element, L_store_element, L_do_card_marks, L_done;
2221
2222 // Input registers (after setup_arg_regs)
2223 const Register from = rdi; // source array address
2224 const Register to = rsi; // destination array address
2225 const Register length = rdx; // elements count
2226 const Register ckoff = rcx; // super_check_offset
2227 const Register ckval = r8; // super_klass
2228
2229 // Registers used as temps (r13, r14 are save-on-entry)
2230 const Register end_from = from; // source array end address
2231 const Register end_to = r13; // destination array end address
2232 const Register count = rdx; // -(count_remaining)
2233 const Register r14_length = r14; // saved copy of length
2234 // End pointers are inclusive, and if length is not zero they point
2235 // to the last unit copied: end_to[0] := end_from[0]
2236
2237 const Register rax_oop = rax; // actual oop copied
2238 const Register r11_klass = r11; // oop._klass
2239
2240 //---------------------------------------------------------------
2241 // Assembler stub will be used for this call to arraycopy
2242 // if the two arrays are subtypes of Object[] but the
2243 // destination array type is not equal to or a supertype
2244 // of the source type. Each element must be separately
2245 // checked.
2246
2247 __ align(CodeEntryAlignment);
2248 StubCodeMark mark(this, "StubRoutines", name);
2249 address start = __ pc();
2250
2251 __ enter(); // required for proper stackwalking of RuntimeStub frame
2252
2253 #ifdef ASSERT
2254 // caller guarantees that the arrays really are different
2255 // otherwise, we would have to make conjoint checks
2256 { Label L;
2257 array_overlap_test(L, TIMES_OOP);
2258 __ stop("checkcast_copy within a single array");
2259 __ bind(L);
2260 }
2261 #endif //ASSERT
2262
2263 setup_arg_regs(4); // from => rdi, to => rsi, length => rdx
2264 // ckoff => rcx, ckval => r8
2265 // r9 and r10 may be used to save non-volatile registers
2266 #ifdef _WIN64
2267 // last argument (#4) is on stack on Win64
2268 __ movptr(ckval, Address(rsp, 6 * wordSize));
2269 #endif
2270
2271 // Caller of this entry point must set up the argument registers.
2272 if (entry != NULL) {
2273 *entry = __ pc();
2274 BLOCK_COMMENT("Entry:");
2275 }
2276
2277 // allocate spill slots for r13, r14
2278 enum {
2279 saved_r13_offset,
2280 saved_r14_offset,
2281 saved_r10_offset,
2282 saved_rbp_offset
2283 };
2284 __ subptr(rsp, saved_rbp_offset * wordSize);
2285 __ movptr(Address(rsp, saved_r13_offset * wordSize), r13);
2286 __ movptr(Address(rsp, saved_r14_offset * wordSize), r14);
2287 __ movptr(Address(rsp, saved_r10_offset * wordSize), r10);
2288
2289 #ifdef ASSERT
2290 Label L2;
2291 __ get_thread(r14);
2292 __ cmpptr(r15_thread, r14);
2293 __ jcc(Assembler::equal, L2);
2294 __ stop("StubRoutines::call_stub: r15_thread is modified by call");
2295 __ bind(L2);
2296 #endif // ASSERT
2297
2298 // check that int operands are properly extended to size_t
2299 assert_clean_int(length, rax);
2300 assert_clean_int(ckoff, rax);
2301
2302 #ifdef ASSERT
2303 BLOCK_COMMENT("assert consistent ckoff/ckval");
2304 // The ckoff and ckval must be mutually consistent,
2305 // even though caller generates both.
2306 { Label L;
2307 int sco_offset = in_bytes(Klass::super_check_offset_offset());
2308 __ cmpl(ckoff, Address(ckval, sco_offset));
2309 __ jcc(Assembler::equal, L);
2310 __ stop("super_check_offset inconsistent");
2311 __ bind(L);
2312 }
2313 #endif //ASSERT
2314
2315 // Loop-invariant addresses. They are exclusive end pointers.
2316 Address end_from_addr(from, length, TIMES_OOP, 0);
2317 Address end_to_addr(to, length, TIMES_OOP, 0);
2318 // Loop-variant addresses. They assume post-incremented count < 0.
2319 Address from_element_addr(end_from, count, TIMES_OOP, 0);
2320 Address to_element_addr(end_to, count, TIMES_OOP, 0);
2321
2322 DecoratorSet decorators = IN_HEAP | IS_ARRAY | ARRAYCOPY_CHECKCAST | ARRAYCOPY_DISJOINT;
2323 if (dest_uninitialized) {
2324 decorators |= IS_DEST_UNINITIALIZED;
2325 }
2326
2327 BasicType type = T_OBJECT;
2328 BarrierSetAssembler *bs = BarrierSet::barrier_set()->barrier_set_assembler();
2329 bs->arraycopy_prologue(_masm, decorators, type, from, to, count);
2330
2331 // Copy from low to high addresses, indexed from the end of each array.
2332 __ lea(end_from, end_from_addr);
2333 __ lea(end_to, end_to_addr);
2334 __ movptr(r14_length, length); // save a copy of the length
2335 assert(length == count, ""); // else fix next line:
2336 __ negptr(count); // negate and test the length
2337 __ jcc(Assembler::notZero, L_load_element);
2338
2339 // Empty array: Nothing to do.
2340 __ xorptr(rax, rax); // return 0 on (trivial) success
2341 __ jmp(L_done);
2342
2343 // ======== begin loop ========
2344 // (Loop is rotated; its entry is L_load_element.)
2345 // Loop control:
2346 // for (count = -count; count != 0; count++)
2347 // Base pointers src, dst are biased by 8*(count-1),to last element.
2348 __ align(OptoLoopAlignment);
2349
2350 __ BIND(L_store_element);
2351 __ store_heap_oop(to_element_addr, rax_oop, noreg, noreg, AS_RAW); // store the oop
2352 __ increment(count); // increment the count toward zero
2353 __ jcc(Assembler::zero, L_do_card_marks);
2354
2355 // ======== loop entry is here ========
2356 __ BIND(L_load_element);
2357 __ load_heap_oop(rax_oop, from_element_addr, noreg, noreg, AS_RAW); // load the oop
2358 __ testptr(rax_oop, rax_oop);
2359 __ jcc(Assembler::zero, L_store_element);
2360
2361 __ load_klass(r11_klass, rax_oop, rscratch1);// query the object klass
2362 generate_type_check(r11_klass, ckoff, ckval, L_store_element);
2363 // ======== end loop ========
2364
2365 // It was a real error; we must depend on the caller to finish the job.
2366 // Register rdx = -1 * number of *remaining* oops, r14 = *total* oops.
2367 // Emit GC store barriers for the oops we have copied (r14 + rdx),
2368 // and report their number to the caller.
2369 assert_different_registers(rax, r14_length, count, to, end_to, rcx, rscratch1);
2370 Label L_post_barrier;
2371 __ addptr(r14_length, count); // K = (original - remaining) oops
2372 __ movptr(rax, r14_length); // save the value
2373 __ notptr(rax); // report (-1^K) to caller (does not affect flags)
2374 __ jccb(Assembler::notZero, L_post_barrier);
2375 __ jmp(L_done); // K == 0, nothing was copied, skip post barrier
2376
2377 // Come here on success only.
2378 __ BIND(L_do_card_marks);
2379 __ xorptr(rax, rax); // return 0 on success
2380
2381 __ BIND(L_post_barrier);
2382 bs->arraycopy_epilogue(_masm, decorators, type, from, to, r14_length);
2383
2384 // Common exit point (success or failure).
2385 __ BIND(L_done);
2386 __ movptr(r13, Address(rsp, saved_r13_offset * wordSize));
2387 __ movptr(r14, Address(rsp, saved_r14_offset * wordSize));
2388 __ movptr(r10, Address(rsp, saved_r10_offset * wordSize));
2389 restore_arg_regs();
2390 inc_counter_np(SharedRuntime::_checkcast_array_copy_ctr); // Update counter after rscratch1 is free
2391 __ leave(); // required for proper stackwalking of RuntimeStub frame
2392 __ ret(0);
2393
2394 return start;
2395 }
2396
2397 //
2398 // Generate 'unsafe' array copy stub
2399 // Though just as safe as the other stubs, it takes an unscaled
2400 // size_t argument instead of an element count.
2401 //
2402 // Input:
2403 // c_rarg0 - source array address
2404 // c_rarg1 - destination array address
2405 // c_rarg2 - byte count, treated as ssize_t, can be zero
2406 //
2407 // Examines the alignment of the operands and dispatches
2408 // to a long, int, short, or byte copy loop.
2409 //
2410 address generate_unsafe_copy(const char *name,
2411 address byte_copy_entry, address short_copy_entry,
2412 address int_copy_entry, address long_copy_entry) {
2413
2414 Label L_long_aligned, L_int_aligned, L_short_aligned;
2415
2416 // Input registers (before setup_arg_regs)
2417 const Register from = c_rarg0; // source array address
2418 const Register to = c_rarg1; // destination array address
2419 const Register size = c_rarg2; // byte count (size_t)
2420
2421 // Register used as a temp
2422 const Register bits = rax; // test copy of low bits
2423
2424 __ align(CodeEntryAlignment);
2425 StubCodeMark mark(this, "StubRoutines", name);
2426 address start = __ pc();
2427
2428 __ enter(); // required for proper stackwalking of RuntimeStub frame
2429
2430 // bump this on entry, not on exit:
2431 inc_counter_np(SharedRuntime::_unsafe_array_copy_ctr);
2432
2433 __ mov(bits, from);
2434 __ orptr(bits, to);
2435 __ orptr(bits, size);
2436
2437 __ testb(bits, BytesPerLong-1);
2438 __ jccb(Assembler::zero, L_long_aligned);
2439
2440 __ testb(bits, BytesPerInt-1);
2441 __ jccb(Assembler::zero, L_int_aligned);
2442
2443 __ testb(bits, BytesPerShort-1);
2444 __ jump_cc(Assembler::notZero, RuntimeAddress(byte_copy_entry));
2445
2446 __ BIND(L_short_aligned);
2447 __ shrptr(size, LogBytesPerShort); // size => short_count
2448 __ jump(RuntimeAddress(short_copy_entry));
2449
2450 __ BIND(L_int_aligned);
2451 __ shrptr(size, LogBytesPerInt); // size => int_count
2452 __ jump(RuntimeAddress(int_copy_entry));
2453
2454 __ BIND(L_long_aligned);
2455 __ shrptr(size, LogBytesPerLong); // size => qword_count
2456 __ jump(RuntimeAddress(long_copy_entry));
2457
2458 return start;
2459 }
2460
2461 // Perform range checks on the proposed arraycopy.
2462 // Kills temp, but nothing else.
2463 // Also, clean the sign bits of src_pos and dst_pos.
2464 void arraycopy_range_checks(Register src, // source array oop (c_rarg0)
2465 Register src_pos, // source position (c_rarg1)
2466 Register dst, // destination array oo (c_rarg2)
2467 Register dst_pos, // destination position (c_rarg3)
2468 Register length,
2469 Register temp,
2470 Label& L_failed) {
2471 BLOCK_COMMENT("arraycopy_range_checks:");
2472
2473 // if (src_pos + length > arrayOop(src)->length()) FAIL;
2474 __ movl(temp, length);
2475 __ addl(temp, src_pos); // src_pos + length
2476 __ cmpl(temp, Address(src, arrayOopDesc::length_offset_in_bytes()));
2477 __ jcc(Assembler::above, L_failed);
2478
2479 // if (dst_pos + length > arrayOop(dst)->length()) FAIL;
2480 __ movl(temp, length);
2481 __ addl(temp, dst_pos); // dst_pos + length
2482 __ cmpl(temp, Address(dst, arrayOopDesc::length_offset_in_bytes()));
2483 __ jcc(Assembler::above, L_failed);
2484
2485 // Have to clean up high 32-bits of 'src_pos' and 'dst_pos'.
2486 // Move with sign extension can be used since they are positive.
2487 __ movslq(src_pos, src_pos);
2488 __ movslq(dst_pos, dst_pos);
2489
2490 BLOCK_COMMENT("arraycopy_range_checks done");
2491 }
2492
2493 //
2494 // Generate generic array copy stubs
2495 //
2496 // Input:
2497 // c_rarg0 - src oop
2498 // c_rarg1 - src_pos (32-bits)
2499 // c_rarg2 - dst oop
2500 // c_rarg3 - dst_pos (32-bits)
2501 // not Win64
2502 // c_rarg4 - element count (32-bits)
2503 // Win64
2504 // rsp+40 - element count (32-bits)
2505 //
2506 // Output:
2507 // rax == 0 - success
2508 // rax == -1^K - failure, where K is partial transfer count
2509 //
2510 address generate_generic_copy(const char *name,
2511 address byte_copy_entry, address short_copy_entry,
2512 address int_copy_entry, address oop_copy_entry,
2513 address long_copy_entry, address checkcast_copy_entry) {
2514
2515 Label L_failed, L_failed_0, L_objArray;
2516 Label L_copy_bytes, L_copy_shorts, L_copy_ints, L_copy_longs;
2517
2518 // Input registers
2519 const Register src = c_rarg0; // source array oop
2520 const Register src_pos = c_rarg1; // source position
2521 const Register dst = c_rarg2; // destination array oop
2522 const Register dst_pos = c_rarg3; // destination position
2523 #ifndef _WIN64
2524 const Register length = c_rarg4;
2525 const Register rklass_tmp = r9; // load_klass
2526 #else
2527 const Address length(rsp, 6 * wordSize); // elements count is on stack on Win64
2528 const Register rklass_tmp = rdi; // load_klass
2529 #endif
2530
2531 { int modulus = CodeEntryAlignment;
2532 int target = modulus - 5; // 5 = sizeof jmp(L_failed)
2533 int advance = target - (__ offset() % modulus);
2534 if (advance < 0) advance += modulus;
2535 if (advance > 0) __ nop(advance);
2536 }
2537 StubCodeMark mark(this, "StubRoutines", name);
2538
2539 // Short-hop target to L_failed. Makes for denser prologue code.
2540 __ BIND(L_failed_0);
2541 __ jmp(L_failed);
2542 assert(__ offset() % CodeEntryAlignment == 0, "no further alignment needed");
2543
2544 __ align(CodeEntryAlignment);
2545 address start = __ pc();
2546
2547 __ enter(); // required for proper stackwalking of RuntimeStub frame
2548
2549 // bump this on entry, not on exit:
2550 inc_counter_np(SharedRuntime::_generic_array_copy_ctr);
2551
2552 //-----------------------------------------------------------------------
2553 // Assembler stub will be used for this call to arraycopy
2554 // if the following conditions are met:
2555 //
2556 // (1) src and dst must not be null.
2557 // (2) src_pos must not be negative.
2558 // (3) dst_pos must not be negative.
2559 // (4) length must not be negative.
2560 // (5) src klass and dst klass should be the same and not NULL.
2561 // (6) src and dst should be arrays.
2562 // (7) src_pos + length must not exceed length of src.
2563 // (8) dst_pos + length must not exceed length of dst.
2564 //
2565
2566 // if (src == NULL) return -1;
2567 __ testptr(src, src); // src oop
2568 size_t j1off = __ offset();
2569 __ jccb(Assembler::zero, L_failed_0);
2570
2571 // if (src_pos < 0) return -1;
2572 __ testl(src_pos, src_pos); // src_pos (32-bits)
2573 __ jccb(Assembler::negative, L_failed_0);
2574
2575 // if (dst == NULL) return -1;
2576 __ testptr(dst, dst); // dst oop
2577 __ jccb(Assembler::zero, L_failed_0);
2578
2579 // if (dst_pos < 0) return -1;
2580 __ testl(dst_pos, dst_pos); // dst_pos (32-bits)
2581 size_t j4off = __ offset();
2582 __ jccb(Assembler::negative, L_failed_0);
2583
2584 // The first four tests are very dense code,
2585 // but not quite dense enough to put four
2586 // jumps in a 16-byte instruction fetch buffer.
2587 // That's good, because some branch predicters
2588 // do not like jumps so close together.
2589 // Make sure of this.
2590 guarantee(((j1off ^ j4off) & ~15) != 0, "I$ line of 1st & 4th jumps");
2591
2592 // registers used as temp
2593 const Register r11_length = r11; // elements count to copy
2594 const Register r10_src_klass = r10; // array klass
2595
2596 // if (length < 0) return -1;
2597 __ movl(r11_length, length); // length (elements count, 32-bits value)
2598 __ testl(r11_length, r11_length);
2599 __ jccb(Assembler::negative, L_failed_0);
2600
2601 __ load_klass(r10_src_klass, src, rklass_tmp);
2602 #ifdef ASSERT
2603 // assert(src->klass() != NULL);
2604 {
2605 BLOCK_COMMENT("assert klasses not null {");
2606 Label L1, L2;
2607 __ testptr(r10_src_klass, r10_src_klass);
2608 __ jcc(Assembler::notZero, L2); // it is broken if klass is NULL
2609 __ bind(L1);
2610 __ stop("broken null klass");
2611 __ bind(L2);
2612 __ load_klass(rax, dst, rklass_tmp);
2613 __ cmpq(rax, 0);
2614 __ jcc(Assembler::equal, L1); // this would be broken also
2615 BLOCK_COMMENT("} assert klasses not null done");
2616 }
2617 #endif
2618
2619 // Load layout helper (32-bits)
2620 //
2621 // |array_tag| | header_size | element_type | |log2_element_size|
2622 // 32 30 24 16 8 2 0
2623 //
2624 // array_tag: typeArray = 0x3, objArray = 0x2, non-array = 0x0
2625 //
2626
2627 const int lh_offset = in_bytes(Klass::layout_helper_offset());
2628
2629 // Handle objArrays completely differently...
2630 const jint objArray_lh = Klass::array_layout_helper(T_OBJECT);
2631 __ cmpl(Address(r10_src_klass, lh_offset), objArray_lh);
2632 __ jcc(Assembler::equal, L_objArray);
2633
2634 // if (src->klass() != dst->klass()) return -1;
2635 __ load_klass(rax, dst, rklass_tmp);
2636 __ cmpq(r10_src_klass, rax);
2637 __ jcc(Assembler::notEqual, L_failed);
2638
2639 const Register rax_lh = rax; // layout helper
2640 __ movl(rax_lh, Address(r10_src_klass, lh_offset));
2641
2642 // if (!src->is_Array()) return -1;
2643 __ cmpl(rax_lh, Klass::_lh_neutral_value);
2644 __ jcc(Assembler::greaterEqual, L_failed);
2645
2646 // At this point, it is known to be a typeArray (array_tag 0x3).
2647 #ifdef ASSERT
2648 {
2649 BLOCK_COMMENT("assert primitive array {");
2650 Label L;
2651 __ cmpl(rax_lh, (Klass::_lh_array_tag_type_value << Klass::_lh_array_tag_shift));
2652 __ jcc(Assembler::greaterEqual, L);
2653 __ stop("must be a primitive array");
2654 __ bind(L);
2655 BLOCK_COMMENT("} assert primitive array done");
2656 }
2657 #endif
2658
2659 arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length,
2660 r10, L_failed);
2661
2662 // TypeArrayKlass
2663 //
2664 // src_addr = (src + array_header_in_bytes()) + (src_pos << log2elemsize);
2665 // dst_addr = (dst + array_header_in_bytes()) + (dst_pos << log2elemsize);
2666 //
2667
2668 const Register r10_offset = r10; // array offset
2669 const Register rax_elsize = rax_lh; // element size
2670
2671 __ movl(r10_offset, rax_lh);
2672 __ shrl(r10_offset, Klass::_lh_header_size_shift);
2673 __ andptr(r10_offset, Klass::_lh_header_size_mask); // array_offset
2674 __ addptr(src, r10_offset); // src array offset
2675 __ addptr(dst, r10_offset); // dst array offset
2676 BLOCK_COMMENT("choose copy loop based on element size");
2677 __ andl(rax_lh, Klass::_lh_log2_element_size_mask); // rax_lh -> rax_elsize
2678
2679 // next registers should be set before the jump to corresponding stub
2680 const Register from = c_rarg0; // source array address
2681 const Register to = c_rarg1; // destination array address
2682 const Register count = c_rarg2; // elements count
2683
2684 // 'from', 'to', 'count' registers should be set in such order
2685 // since they are the same as 'src', 'src_pos', 'dst'.
2686
2687 __ BIND(L_copy_bytes);
2688 __ cmpl(rax_elsize, 0);
2689 __ jccb(Assembler::notEqual, L_copy_shorts);
2690 __ lea(from, Address(src, src_pos, Address::times_1, 0));// src_addr
2691 __ lea(to, Address(dst, dst_pos, Address::times_1, 0));// dst_addr
2692 __ movl2ptr(count, r11_length); // length
2693 __ jump(RuntimeAddress(byte_copy_entry));
2694
2695 __ BIND(L_copy_shorts);
2696 __ cmpl(rax_elsize, LogBytesPerShort);
2697 __ jccb(Assembler::notEqual, L_copy_ints);
2698 __ lea(from, Address(src, src_pos, Address::times_2, 0));// src_addr
2699 __ lea(to, Address(dst, dst_pos, Address::times_2, 0));// dst_addr
2700 __ movl2ptr(count, r11_length); // length
2701 __ jump(RuntimeAddress(short_copy_entry));
2702
2703 __ BIND(L_copy_ints);
2704 __ cmpl(rax_elsize, LogBytesPerInt);
2705 __ jccb(Assembler::notEqual, L_copy_longs);
2706 __ lea(from, Address(src, src_pos, Address::times_4, 0));// src_addr
2707 __ lea(to, Address(dst, dst_pos, Address::times_4, 0));// dst_addr
2708 __ movl2ptr(count, r11_length); // length
2709 __ jump(RuntimeAddress(int_copy_entry));
2710
2711 __ BIND(L_copy_longs);
2712 #ifdef ASSERT
2713 {
2714 BLOCK_COMMENT("assert long copy {");
2715 Label L;
2716 __ cmpl(rax_elsize, LogBytesPerLong);
2717 __ jcc(Assembler::equal, L);
2718 __ stop("must be long copy, but elsize is wrong");
2719 __ bind(L);
2720 BLOCK_COMMENT("} assert long copy done");
2721 }
2722 #endif
2723 __ lea(from, Address(src, src_pos, Address::times_8, 0));// src_addr
2724 __ lea(to, Address(dst, dst_pos, Address::times_8, 0));// dst_addr
2725 __ movl2ptr(count, r11_length); // length
2726 __ jump(RuntimeAddress(long_copy_entry));
2727
2728 // ObjArrayKlass
2729 __ BIND(L_objArray);
2730 // live at this point: r10_src_klass, r11_length, src[_pos], dst[_pos]
2731
2732 Label L_plain_copy, L_checkcast_copy;
2733 // test array classes for subtyping
2734 __ load_klass(rax, dst, rklass_tmp);
2735 __ cmpq(r10_src_klass, rax); // usual case is exact equality
2736 __ jcc(Assembler::notEqual, L_checkcast_copy);
2737
2738 // Identically typed arrays can be copied without element-wise checks.
2739 arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length,
2740 r10, L_failed);
2741
2742 __ lea(from, Address(src, src_pos, TIMES_OOP,
2743 arrayOopDesc::base_offset_in_bytes(T_OBJECT))); // src_addr
2744 __ lea(to, Address(dst, dst_pos, TIMES_OOP,
2745 arrayOopDesc::base_offset_in_bytes(T_OBJECT))); // dst_addr
2746 __ movl2ptr(count, r11_length); // length
2747 __ BIND(L_plain_copy);
2748 __ jump(RuntimeAddress(oop_copy_entry));
2749
2750 __ BIND(L_checkcast_copy);
2751 // live at this point: r10_src_klass, r11_length, rax (dst_klass)
2752 {
2753 // Before looking at dst.length, make sure dst is also an objArray.
2754 __ cmpl(Address(rax, lh_offset), objArray_lh);
2755 __ jcc(Assembler::notEqual, L_failed);
2756
2757 // It is safe to examine both src.length and dst.length.
2758 arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length,
2759 rax, L_failed);
2760
2761 const Register r11_dst_klass = r11;
2762 __ load_klass(r11_dst_klass, dst, rklass_tmp); // reload
2763
2764 // Marshal the base address arguments now, freeing registers.
2765 __ lea(from, Address(src, src_pos, TIMES_OOP,
2766 arrayOopDesc::base_offset_in_bytes(T_OBJECT)));
2767 __ lea(to, Address(dst, dst_pos, TIMES_OOP,
2768 arrayOopDesc::base_offset_in_bytes(T_OBJECT)));
2769 __ movl(count, length); // length (reloaded)
2770 Register sco_temp = c_rarg3; // this register is free now
2771 assert_different_registers(from, to, count, sco_temp,
2772 r11_dst_klass, r10_src_klass);
2773 assert_clean_int(count, sco_temp);
2774
2775 // Generate the type check.
2776 const int sco_offset = in_bytes(Klass::super_check_offset_offset());
2777 __ movl(sco_temp, Address(r11_dst_klass, sco_offset));
2778 assert_clean_int(sco_temp, rax);
2779 generate_type_check(r10_src_klass, sco_temp, r11_dst_klass, L_plain_copy);
2780
2781 // Fetch destination element klass from the ObjArrayKlass header.
2782 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
2783 __ movptr(r11_dst_klass, Address(r11_dst_klass, ek_offset));
2784 __ movl( sco_temp, Address(r11_dst_klass, sco_offset));
2785 assert_clean_int(sco_temp, rax);
2786
2787 // the checkcast_copy loop needs two extra arguments:
2788 assert(c_rarg3 == sco_temp, "#3 already in place");
2789 // Set up arguments for checkcast_copy_entry.
2790 setup_arg_regs(4);
2791 __ movptr(r8, r11_dst_klass); // dst.klass.element_klass, r8 is c_rarg4 on Linux/Solaris
2792 __ jump(RuntimeAddress(checkcast_copy_entry));
2793 }
2794
2795 __ BIND(L_failed);
2796 __ xorptr(rax, rax);
2797 __ notptr(rax); // return -1
2798 __ leave(); // required for proper stackwalking of RuntimeStub frame
2799 __ ret(0);
2800
2801 return start;
2802 }
2803
2804 address generate_data_cache_writeback() {
2805 const Register src = c_rarg0; // source address
2806
2807 __ align(CodeEntryAlignment);
2808
2809 StubCodeMark mark(this, "StubRoutines", "_data_cache_writeback");
2810
2811 address start = __ pc();
2812 __ enter();
2813 __ cache_wb(Address(src, 0));
2814 __ leave();
2815 __ ret(0);
2816
2817 return start;
2818 }
2819
2820 address generate_data_cache_writeback_sync() {
2821 const Register is_pre = c_rarg0; // pre or post sync
2822
2823 __ align(CodeEntryAlignment);
2824
2825 StubCodeMark mark(this, "StubRoutines", "_data_cache_writeback_sync");
2826
2827 // pre wbsync is a no-op
2828 // post wbsync translates to an sfence
2829
2830 Label skip;
2831 address start = __ pc();
2832 __ enter();
2833 __ cmpl(is_pre, 0);
2834 __ jcc(Assembler::notEqual, skip);
2835 __ cache_wbsync(false);
2836 __ bind(skip);
2837 __ leave();
2838 __ ret(0);
2839
2840 return start;
2841 }
2842
2843 void generate_arraycopy_stubs() {
2844 address entry;
2845 address entry_jbyte_arraycopy;
2846 address entry_jshort_arraycopy;
2847 address entry_jint_arraycopy;
2848 address entry_oop_arraycopy;
2849 address entry_jlong_arraycopy;
2850 address entry_checkcast_arraycopy;
2851
2852 StubRoutines::_jbyte_disjoint_arraycopy = generate_disjoint_byte_copy(false, &entry,
2853 "jbyte_disjoint_arraycopy");
2854 StubRoutines::_jbyte_arraycopy = generate_conjoint_byte_copy(false, entry, &entry_jbyte_arraycopy,
2855 "jbyte_arraycopy");
2856
2857 StubRoutines::_jshort_disjoint_arraycopy = generate_disjoint_short_copy(false, &entry,
2858 "jshort_disjoint_arraycopy");
2859 StubRoutines::_jshort_arraycopy = generate_conjoint_short_copy(false, entry, &entry_jshort_arraycopy,
2860 "jshort_arraycopy");
2861
2862 StubRoutines::_jint_disjoint_arraycopy = generate_disjoint_int_oop_copy(false, false, &entry,
2863 "jint_disjoint_arraycopy");
2864 StubRoutines::_jint_arraycopy = generate_conjoint_int_oop_copy(false, false, entry,
2865 &entry_jint_arraycopy, "jint_arraycopy");
2866
2867 StubRoutines::_jlong_disjoint_arraycopy = generate_disjoint_long_oop_copy(false, false, &entry,
2868 "jlong_disjoint_arraycopy");
2869 StubRoutines::_jlong_arraycopy = generate_conjoint_long_oop_copy(false, false, entry,
2870 &entry_jlong_arraycopy, "jlong_arraycopy");
2871
2872
2873 if (UseCompressedOops) {
2874 StubRoutines::_oop_disjoint_arraycopy = generate_disjoint_int_oop_copy(false, true, &entry,
2875 "oop_disjoint_arraycopy");
2876 StubRoutines::_oop_arraycopy = generate_conjoint_int_oop_copy(false, true, entry,
2877 &entry_oop_arraycopy, "oop_arraycopy");
2878 StubRoutines::_oop_disjoint_arraycopy_uninit = generate_disjoint_int_oop_copy(false, true, &entry,
2879 "oop_disjoint_arraycopy_uninit",
2880 /*dest_uninitialized*/true);
2881 StubRoutines::_oop_arraycopy_uninit = generate_conjoint_int_oop_copy(false, true, entry,
2882 NULL, "oop_arraycopy_uninit",
2883 /*dest_uninitialized*/true);
2884 } else {
2885 StubRoutines::_oop_disjoint_arraycopy = generate_disjoint_long_oop_copy(false, true, &entry,
2886 "oop_disjoint_arraycopy");
2887 StubRoutines::_oop_arraycopy = generate_conjoint_long_oop_copy(false, true, entry,
2888 &entry_oop_arraycopy, "oop_arraycopy");
2889 StubRoutines::_oop_disjoint_arraycopy_uninit = generate_disjoint_long_oop_copy(false, true, &entry,
2890 "oop_disjoint_arraycopy_uninit",
2891 /*dest_uninitialized*/true);
2892 StubRoutines::_oop_arraycopy_uninit = generate_conjoint_long_oop_copy(false, true, entry,
2893 NULL, "oop_arraycopy_uninit",
2894 /*dest_uninitialized*/true);
2895 }
2896
2897 StubRoutines::_checkcast_arraycopy = generate_checkcast_copy("checkcast_arraycopy", &entry_checkcast_arraycopy);
2898 StubRoutines::_checkcast_arraycopy_uninit = generate_checkcast_copy("checkcast_arraycopy_uninit", NULL,
2899 /*dest_uninitialized*/true);
2900
2901 StubRoutines::_unsafe_arraycopy = generate_unsafe_copy("unsafe_arraycopy",
2902 entry_jbyte_arraycopy,
2903 entry_jshort_arraycopy,
2904 entry_jint_arraycopy,
2905 entry_jlong_arraycopy);
2906 StubRoutines::_generic_arraycopy = generate_generic_copy("generic_arraycopy",
2907 entry_jbyte_arraycopy,
2908 entry_jshort_arraycopy,
2909 entry_jint_arraycopy,
2910 entry_oop_arraycopy,
2911 entry_jlong_arraycopy,
2912 entry_checkcast_arraycopy);
2913
2914 StubRoutines::_jbyte_fill = generate_fill(T_BYTE, false, "jbyte_fill");
2915 StubRoutines::_jshort_fill = generate_fill(T_SHORT, false, "jshort_fill");
2916 StubRoutines::_jint_fill = generate_fill(T_INT, false, "jint_fill");
2917 StubRoutines::_arrayof_jbyte_fill = generate_fill(T_BYTE, true, "arrayof_jbyte_fill");
2918 StubRoutines::_arrayof_jshort_fill = generate_fill(T_SHORT, true, "arrayof_jshort_fill");
2919 StubRoutines::_arrayof_jint_fill = generate_fill(T_INT, true, "arrayof_jint_fill");
2920
2921 // We don't generate specialized code for HeapWord-aligned source
2922 // arrays, so just use the code we've already generated
2923 StubRoutines::_arrayof_jbyte_disjoint_arraycopy = StubRoutines::_jbyte_disjoint_arraycopy;
2924 StubRoutines::_arrayof_jbyte_arraycopy = StubRoutines::_jbyte_arraycopy;
2925
2926 StubRoutines::_arrayof_jshort_disjoint_arraycopy = StubRoutines::_jshort_disjoint_arraycopy;
2927 StubRoutines::_arrayof_jshort_arraycopy = StubRoutines::_jshort_arraycopy;
2928
2929 StubRoutines::_arrayof_jint_disjoint_arraycopy = StubRoutines::_jint_disjoint_arraycopy;
2930 StubRoutines::_arrayof_jint_arraycopy = StubRoutines::_jint_arraycopy;
2931
2932 StubRoutines::_arrayof_jlong_disjoint_arraycopy = StubRoutines::_jlong_disjoint_arraycopy;
2933 StubRoutines::_arrayof_jlong_arraycopy = StubRoutines::_jlong_arraycopy;
2934
2935 StubRoutines::_arrayof_oop_disjoint_arraycopy = StubRoutines::_oop_disjoint_arraycopy;
2936 StubRoutines::_arrayof_oop_arraycopy = StubRoutines::_oop_arraycopy;
2937
2938 StubRoutines::_arrayof_oop_disjoint_arraycopy_uninit = StubRoutines::_oop_disjoint_arraycopy_uninit;
2939 StubRoutines::_arrayof_oop_arraycopy_uninit = StubRoutines::_oop_arraycopy_uninit;
2940 }
2941
2942 // AES intrinsic stubs
2943 enum {AESBlockSize = 16};
2944
2945 address generate_key_shuffle_mask() {
2946 __ align(16);
2947 StubCodeMark mark(this, "StubRoutines", "key_shuffle_mask");
2948 address start = __ pc();
2949 __ emit_data64( 0x0405060700010203, relocInfo::none );
2950 __ emit_data64( 0x0c0d0e0f08090a0b, relocInfo::none );
2951 return start;
2952 }
2953
2954 address generate_counter_shuffle_mask() {
2955 __ align(16);
2956 StubCodeMark mark(this, "StubRoutines", "counter_shuffle_mask");
2957 address start = __ pc();
2958 __ emit_data64(0x08090a0b0c0d0e0f, relocInfo::none);
2959 __ emit_data64(0x0001020304050607, relocInfo::none);
2960 return start;
2961 }
2962
2963 // Utility routine for loading a 128-bit key word in little endian format
2964 // can optionally specify that the shuffle mask is already in an xmmregister
2965 void load_key(XMMRegister xmmdst, Register key, int offset, XMMRegister xmm_shuf_mask=NULL) {
2966 __ movdqu(xmmdst, Address(key, offset));
2967 if (xmm_shuf_mask != NULL) {
2968 __ pshufb(xmmdst, xmm_shuf_mask);
2969 } else {
2970 __ pshufb(xmmdst, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
2971 }
2972 }
2973
2974 // Utility routine for increase 128bit counter (iv in CTR mode)
2975 void inc_counter(Register reg, XMMRegister xmmdst, int inc_delta, Label& next_block) {
2976 __ pextrq(reg, xmmdst, 0x0);
2977 __ addq(reg, inc_delta);
2978 __ pinsrq(xmmdst, reg, 0x0);
2979 __ jcc(Assembler::carryClear, next_block); // jump if no carry
2980 __ pextrq(reg, xmmdst, 0x01); // Carry
2981 __ addq(reg, 0x01);
2982 __ pinsrq(xmmdst, reg, 0x01); //Carry end
2983 __ BIND(next_block); // next instruction
2984 }
2985
2986 // Arguments:
2987 //
2988 // Inputs:
2989 // c_rarg0 - source byte array address
2990 // c_rarg1 - destination byte array address
2991 // c_rarg2 - K (key) in little endian int array
2992 //
2993 address generate_aescrypt_encryptBlock() {
2994 assert(UseAES, "need AES instructions and misaligned SSE support");
2995 __ align(CodeEntryAlignment);
2996 StubCodeMark mark(this, "StubRoutines", "aescrypt_encryptBlock");
2997 Label L_doLast;
2998 address start = __ pc();
2999
3000 const Register from = c_rarg0; // source array address
3001 const Register to = c_rarg1; // destination array address
3002 const Register key = c_rarg2; // key array address
3003 const Register keylen = rax;
3004
3005 const XMMRegister xmm_result = xmm0;
3006 const XMMRegister xmm_key_shuf_mask = xmm1;
3007 // On win64 xmm6-xmm15 must be preserved so don't use them.
3008 const XMMRegister xmm_temp1 = xmm2;
3009 const XMMRegister xmm_temp2 = xmm3;
3010 const XMMRegister xmm_temp3 = xmm4;
3011 const XMMRegister xmm_temp4 = xmm5;
3012
3013 __ enter(); // required for proper stackwalking of RuntimeStub frame
3014
3015 // keylen could be only {11, 13, 15} * 4 = {44, 52, 60}
3016 __ movl(keylen, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3017
3018 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3019 __ movdqu(xmm_result, Address(from, 0)); // get 16 bytes of input
3020
3021 // For encryption, the java expanded key ordering is just what we need
3022 // we don't know if the key is aligned, hence not using load-execute form
3023
3024 load_key(xmm_temp1, key, 0x00, xmm_key_shuf_mask);
3025 __ pxor(xmm_result, xmm_temp1);
3026
3027 load_key(xmm_temp1, key, 0x10, xmm_key_shuf_mask);
3028 load_key(xmm_temp2, key, 0x20, xmm_key_shuf_mask);
3029 load_key(xmm_temp3, key, 0x30, xmm_key_shuf_mask);
3030 load_key(xmm_temp4, key, 0x40, xmm_key_shuf_mask);
3031
3032 __ aesenc(xmm_result, xmm_temp1);
3033 __ aesenc(xmm_result, xmm_temp2);
3034 __ aesenc(xmm_result, xmm_temp3);
3035 __ aesenc(xmm_result, xmm_temp4);
3036
3037 load_key(xmm_temp1, key, 0x50, xmm_key_shuf_mask);
3038 load_key(xmm_temp2, key, 0x60, xmm_key_shuf_mask);
3039 load_key(xmm_temp3, key, 0x70, xmm_key_shuf_mask);
3040 load_key(xmm_temp4, key, 0x80, xmm_key_shuf_mask);
3041
3042 __ aesenc(xmm_result, xmm_temp1);
3043 __ aesenc(xmm_result, xmm_temp2);
3044 __ aesenc(xmm_result, xmm_temp3);
3045 __ aesenc(xmm_result, xmm_temp4);
3046
3047 load_key(xmm_temp1, key, 0x90, xmm_key_shuf_mask);
3048 load_key(xmm_temp2, key, 0xa0, xmm_key_shuf_mask);
3049
3050 __ cmpl(keylen, 44);
3051 __ jccb(Assembler::equal, L_doLast);
3052
3053 __ aesenc(xmm_result, xmm_temp1);
3054 __ aesenc(xmm_result, xmm_temp2);
3055
3056 load_key(xmm_temp1, key, 0xb0, xmm_key_shuf_mask);
3057 load_key(xmm_temp2, key, 0xc0, xmm_key_shuf_mask);
3058
3059 __ cmpl(keylen, 52);
3060 __ jccb(Assembler::equal, L_doLast);
3061
3062 __ aesenc(xmm_result, xmm_temp1);
3063 __ aesenc(xmm_result, xmm_temp2);
3064
3065 load_key(xmm_temp1, key, 0xd0, xmm_key_shuf_mask);
3066 load_key(xmm_temp2, key, 0xe0, xmm_key_shuf_mask);
3067
3068 __ BIND(L_doLast);
3069 __ aesenc(xmm_result, xmm_temp1);
3070 __ aesenclast(xmm_result, xmm_temp2);
3071 __ movdqu(Address(to, 0), xmm_result); // store the result
3072 __ xorptr(rax, rax); // return 0
3073 __ leave(); // required for proper stackwalking of RuntimeStub frame
3074 __ ret(0);
3075
3076 return start;
3077 }
3078
3079
3080 // Arguments:
3081 //
3082 // Inputs:
3083 // c_rarg0 - source byte array address
3084 // c_rarg1 - destination byte array address
3085 // c_rarg2 - K (key) in little endian int array
3086 //
3087 address generate_aescrypt_decryptBlock() {
3088 assert(UseAES, "need AES instructions and misaligned SSE support");
3089 __ align(CodeEntryAlignment);
3090 StubCodeMark mark(this, "StubRoutines", "aescrypt_decryptBlock");
3091 Label L_doLast;
3092 address start = __ pc();
3093
3094 const Register from = c_rarg0; // source array address
3095 const Register to = c_rarg1; // destination array address
3096 const Register key = c_rarg2; // key array address
3097 const Register keylen = rax;
3098
3099 const XMMRegister xmm_result = xmm0;
3100 const XMMRegister xmm_key_shuf_mask = xmm1;
3101 // On win64 xmm6-xmm15 must be preserved so don't use them.
3102 const XMMRegister xmm_temp1 = xmm2;
3103 const XMMRegister xmm_temp2 = xmm3;
3104 const XMMRegister xmm_temp3 = xmm4;
3105 const XMMRegister xmm_temp4 = xmm5;
3106
3107 __ enter(); // required for proper stackwalking of RuntimeStub frame
3108
3109 // keylen could be only {11, 13, 15} * 4 = {44, 52, 60}
3110 __ movl(keylen, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3111
3112 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3113 __ movdqu(xmm_result, Address(from, 0));
3114
3115 // for decryption java expanded key ordering is rotated one position from what we want
3116 // so we start from 0x10 here and hit 0x00 last
3117 // we don't know if the key is aligned, hence not using load-execute form
3118 load_key(xmm_temp1, key, 0x10, xmm_key_shuf_mask);
3119 load_key(xmm_temp2, key, 0x20, xmm_key_shuf_mask);
3120 load_key(xmm_temp3, key, 0x30, xmm_key_shuf_mask);
3121 load_key(xmm_temp4, key, 0x40, xmm_key_shuf_mask);
3122
3123 __ pxor (xmm_result, xmm_temp1);
3124 __ aesdec(xmm_result, xmm_temp2);
3125 __ aesdec(xmm_result, xmm_temp3);
3126 __ aesdec(xmm_result, xmm_temp4);
3127
3128 load_key(xmm_temp1, key, 0x50, xmm_key_shuf_mask);
3129 load_key(xmm_temp2, key, 0x60, xmm_key_shuf_mask);
3130 load_key(xmm_temp3, key, 0x70, xmm_key_shuf_mask);
3131 load_key(xmm_temp4, key, 0x80, xmm_key_shuf_mask);
3132
3133 __ aesdec(xmm_result, xmm_temp1);
3134 __ aesdec(xmm_result, xmm_temp2);
3135 __ aesdec(xmm_result, xmm_temp3);
3136 __ aesdec(xmm_result, xmm_temp4);
3137
3138 load_key(xmm_temp1, key, 0x90, xmm_key_shuf_mask);
3139 load_key(xmm_temp2, key, 0xa0, xmm_key_shuf_mask);
3140 load_key(xmm_temp3, key, 0x00, xmm_key_shuf_mask);
3141
3142 __ cmpl(keylen, 44);
3143 __ jccb(Assembler::equal, L_doLast);
3144
3145 __ aesdec(xmm_result, xmm_temp1);
3146 __ aesdec(xmm_result, xmm_temp2);
3147
3148 load_key(xmm_temp1, key, 0xb0, xmm_key_shuf_mask);
3149 load_key(xmm_temp2, key, 0xc0, xmm_key_shuf_mask);
3150
3151 __ cmpl(keylen, 52);
3152 __ jccb(Assembler::equal, L_doLast);
3153
3154 __ aesdec(xmm_result, xmm_temp1);
3155 __ aesdec(xmm_result, xmm_temp2);
3156
3157 load_key(xmm_temp1, key, 0xd0, xmm_key_shuf_mask);
3158 load_key(xmm_temp2, key, 0xe0, xmm_key_shuf_mask);
3159
3160 __ BIND(L_doLast);
3161 __ aesdec(xmm_result, xmm_temp1);
3162 __ aesdec(xmm_result, xmm_temp2);
3163
3164 // for decryption the aesdeclast operation is always on key+0x00
3165 __ aesdeclast(xmm_result, xmm_temp3);
3166 __ movdqu(Address(to, 0), xmm_result); // store the result
3167 __ xorptr(rax, rax); // return 0
3168 __ leave(); // required for proper stackwalking of RuntimeStub frame
3169 __ ret(0);
3170
3171 return start;
3172 }
3173
3174
3175 // Arguments:
3176 //
3177 // Inputs:
3178 // c_rarg0 - source byte array address
3179 // c_rarg1 - destination byte array address
3180 // c_rarg2 - K (key) in little endian int array
3181 // c_rarg3 - r vector byte array address
3182 // c_rarg4 - input length
3183 //
3184 // Output:
3185 // rax - input length
3186 //
3187 address generate_cipherBlockChaining_encryptAESCrypt() {
3188 assert(UseAES, "need AES instructions and misaligned SSE support");
3189 __ align(CodeEntryAlignment);
3190 StubCodeMark mark(this, "StubRoutines", "cipherBlockChaining_encryptAESCrypt");
3191 address start = __ pc();
3192
3193 Label L_exit, L_key_192_256, L_key_256, L_loopTop_128, L_loopTop_192, L_loopTop_256;
3194 const Register from = c_rarg0; // source array address
3195 const Register to = c_rarg1; // destination array address
3196 const Register key = c_rarg2; // key array address
3197 const Register rvec = c_rarg3; // r byte array initialized from initvector array address
3198 // and left with the results of the last encryption block
3199 #ifndef _WIN64
3200 const Register len_reg = c_rarg4; // src len (must be multiple of blocksize 16)
3201 #else
3202 const Address len_mem(rbp, 6 * wordSize); // length is on stack on Win64
3203 const Register len_reg = r11; // pick the volatile windows register
3204 #endif
3205 const Register pos = rax;
3206
3207 // xmm register assignments for the loops below
3208 const XMMRegister xmm_result = xmm0;
3209 const XMMRegister xmm_temp = xmm1;
3210 // keys 0-10 preloaded into xmm2-xmm12
3211 const int XMM_REG_NUM_KEY_FIRST = 2;
3212 const int XMM_REG_NUM_KEY_LAST = 15;
3213 const XMMRegister xmm_key0 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST);
3214 const XMMRegister xmm_key10 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+10);
3215 const XMMRegister xmm_key11 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+11);
3216 const XMMRegister xmm_key12 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+12);
3217 const XMMRegister xmm_key13 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+13);
3218
3219 __ enter(); // required for proper stackwalking of RuntimeStub frame
3220
3221 #ifdef _WIN64
3222 // on win64, fill len_reg from stack position
3223 __ movl(len_reg, len_mem);
3224 #else
3225 __ push(len_reg); // Save
3226 #endif
3227
3228 const XMMRegister xmm_key_shuf_mask = xmm_temp; // used temporarily to swap key bytes up front
3229 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3230 // load up xmm regs xmm2 thru xmm12 with key 0x00 - 0xa0
3231 for (int rnum = XMM_REG_NUM_KEY_FIRST, offset = 0x00; rnum <= XMM_REG_NUM_KEY_FIRST+10; rnum++) {
3232 load_key(as_XMMRegister(rnum), key, offset, xmm_key_shuf_mask);
3233 offset += 0x10;
3234 }
3235 __ movdqu(xmm_result, Address(rvec, 0x00)); // initialize xmm_result with r vec
3236
3237 // now split to different paths depending on the keylen (len in ints of AESCrypt.KLE array (52=192, or 60=256))
3238 __ movl(rax, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3239 __ cmpl(rax, 44);
3240 __ jcc(Assembler::notEqual, L_key_192_256);
3241
3242 // 128 bit code follows here
3243 __ movptr(pos, 0);
3244 __ align(OptoLoopAlignment);
3245
3246 __ BIND(L_loopTop_128);
3247 __ movdqu(xmm_temp, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of input
3248 __ pxor (xmm_result, xmm_temp); // xor with the current r vector
3249 __ pxor (xmm_result, xmm_key0); // do the aes rounds
3250 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_FIRST + 9; rnum++) {
3251 __ aesenc(xmm_result, as_XMMRegister(rnum));
3252 }
3253 __ aesenclast(xmm_result, xmm_key10);
3254 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3255 // no need to store r to memory until we exit
3256 __ addptr(pos, AESBlockSize);
3257 __ subptr(len_reg, AESBlockSize);
3258 __ jcc(Assembler::notEqual, L_loopTop_128);
3259
3260 __ BIND(L_exit);
3261 __ movdqu(Address(rvec, 0), xmm_result); // final value of r stored in rvec of CipherBlockChaining object
3262
3263 #ifdef _WIN64
3264 __ movl(rax, len_mem);
3265 #else
3266 __ pop(rax); // return length
3267 #endif
3268 __ leave(); // required for proper stackwalking of RuntimeStub frame
3269 __ ret(0);
3270
3271 __ BIND(L_key_192_256);
3272 // here rax = len in ints of AESCrypt.KLE array (52=192, or 60=256)
3273 load_key(xmm_key11, key, 0xb0, xmm_key_shuf_mask);
3274 load_key(xmm_key12, key, 0xc0, xmm_key_shuf_mask);
3275 __ cmpl(rax, 52);
3276 __ jcc(Assembler::notEqual, L_key_256);
3277
3278 // 192-bit code follows here (could be changed to use more xmm registers)
3279 __ movptr(pos, 0);
3280 __ align(OptoLoopAlignment);
3281
3282 __ BIND(L_loopTop_192);
3283 __ movdqu(xmm_temp, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of input
3284 __ pxor (xmm_result, xmm_temp); // xor with the current r vector
3285 __ pxor (xmm_result, xmm_key0); // do the aes rounds
3286 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_FIRST + 11; rnum++) {
3287 __ aesenc(xmm_result, as_XMMRegister(rnum));
3288 }
3289 __ aesenclast(xmm_result, xmm_key12);
3290 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3291 // no need to store r to memory until we exit
3292 __ addptr(pos, AESBlockSize);
3293 __ subptr(len_reg, AESBlockSize);
3294 __ jcc(Assembler::notEqual, L_loopTop_192);
3295 __ jmp(L_exit);
3296
3297 __ BIND(L_key_256);
3298 // 256-bit code follows here (could be changed to use more xmm registers)
3299 load_key(xmm_key13, key, 0xd0, xmm_key_shuf_mask);
3300 __ movptr(pos, 0);
3301 __ align(OptoLoopAlignment);
3302
3303 __ BIND(L_loopTop_256);
3304 __ movdqu(xmm_temp, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of input
3305 __ pxor (xmm_result, xmm_temp); // xor with the current r vector
3306 __ pxor (xmm_result, xmm_key0); // do the aes rounds
3307 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_FIRST + 13; rnum++) {
3308 __ aesenc(xmm_result, as_XMMRegister(rnum));
3309 }
3310 load_key(xmm_temp, key, 0xe0);
3311 __ aesenclast(xmm_result, xmm_temp);
3312 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3313 // no need to store r to memory until we exit
3314 __ addptr(pos, AESBlockSize);
3315 __ subptr(len_reg, AESBlockSize);
3316 __ jcc(Assembler::notEqual, L_loopTop_256);
3317 __ jmp(L_exit);
3318
3319 return start;
3320 }
3321
3322 // Safefetch stubs.
3323 void generate_safefetch(const char* name, int size, address* entry,
3324 address* fault_pc, address* continuation_pc) {
3325 // safefetch signatures:
3326 // int SafeFetch32(int* adr, int errValue);
3327 // intptr_t SafeFetchN (intptr_t* adr, intptr_t errValue);
3328 //
3329 // arguments:
3330 // c_rarg0 = adr
3331 // c_rarg1 = errValue
3332 //
3333 // result:
3334 // PPC_RET = *adr or errValue
3335
3336 StubCodeMark mark(this, "StubRoutines", name);
3337
3338 // Entry point, pc or function descriptor.
3339 *entry = __ pc();
3340
3341 // Load *adr into c_rarg1, may fault.
3342 *fault_pc = __ pc();
3343 switch (size) {
3344 case 4:
3345 // int32_t
3346 __ movl(c_rarg1, Address(c_rarg0, 0));
3347 break;
3348 case 8:
3349 // int64_t
3350 __ movq(c_rarg1, Address(c_rarg0, 0));
3351 break;
3352 default:
3353 ShouldNotReachHere();
3354 }
3355
3356 // return errValue or *adr
3357 *continuation_pc = __ pc();
3358 __ movq(rax, c_rarg1);
3359 __ ret(0);
3360 }
3361
3362 // This is a version of CBC/AES Decrypt which does 4 blocks in a loop at a time
3363 // to hide instruction latency
3364 //
3365 // Arguments:
3366 //
3367 // Inputs:
3368 // c_rarg0 - source byte array address
3369 // c_rarg1 - destination byte array address
3370 // c_rarg2 - K (key) in little endian int array
3371 // c_rarg3 - r vector byte array address
3372 // c_rarg4 - input length
3373 //
3374 // Output:
3375 // rax - input length
3376 //
3377 address generate_cipherBlockChaining_decryptAESCrypt_Parallel() {
3378 assert(UseAES, "need AES instructions and misaligned SSE support");
3379 __ align(CodeEntryAlignment);
3380 StubCodeMark mark(this, "StubRoutines", "cipherBlockChaining_decryptAESCrypt");
3381 address start = __ pc();
3382
3383 const Register from = c_rarg0; // source array address
3384 const Register to = c_rarg1; // destination array address
3385 const Register key = c_rarg2; // key array address
3386 const Register rvec = c_rarg3; // r byte array initialized from initvector array address
3387 // and left with the results of the last encryption block
3388 #ifndef _WIN64
3389 const Register len_reg = c_rarg4; // src len (must be multiple of blocksize 16)
3390 #else
3391 const Address len_mem(rbp, 6 * wordSize); // length is on stack on Win64
3392 const Register len_reg = r11; // pick the volatile windows register
3393 #endif
3394 const Register pos = rax;
3395
3396 const int PARALLEL_FACTOR = 4;
3397 const int ROUNDS[3] = { 10, 12, 14 }; // aes rounds for key128, key192, key256
3398
3399 Label L_exit;
3400 Label L_singleBlock_loopTopHead[3]; // 128, 192, 256
3401 Label L_singleBlock_loopTopHead2[3]; // 128, 192, 256
3402 Label L_singleBlock_loopTop[3]; // 128, 192, 256
3403 Label L_multiBlock_loopTopHead[3]; // 128, 192, 256
3404 Label L_multiBlock_loopTop[3]; // 128, 192, 256
3405
3406 // keys 0-10 preloaded into xmm5-xmm15
3407 const int XMM_REG_NUM_KEY_FIRST = 5;
3408 const int XMM_REG_NUM_KEY_LAST = 15;
3409 const XMMRegister xmm_key_first = as_XMMRegister(XMM_REG_NUM_KEY_FIRST);
3410 const XMMRegister xmm_key_last = as_XMMRegister(XMM_REG_NUM_KEY_LAST);
3411
3412 __ enter(); // required for proper stackwalking of RuntimeStub frame
3413
3414 #ifdef _WIN64
3415 // on win64, fill len_reg from stack position
3416 __ movl(len_reg, len_mem);
3417 #else
3418 __ push(len_reg); // Save
3419 #endif
3420 __ push(rbx);
3421 // the java expanded key ordering is rotated one position from what we want
3422 // so we start from 0x10 here and hit 0x00 last
3423 const XMMRegister xmm_key_shuf_mask = xmm1; // used temporarily to swap key bytes up front
3424 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3425 // load up xmm regs 5 thru 15 with key 0x10 - 0xa0 - 0x00
3426 for (int rnum = XMM_REG_NUM_KEY_FIRST, offset = 0x10; rnum < XMM_REG_NUM_KEY_LAST; rnum++) {
3427 load_key(as_XMMRegister(rnum), key, offset, xmm_key_shuf_mask);
3428 offset += 0x10;
3429 }
3430 load_key(xmm_key_last, key, 0x00, xmm_key_shuf_mask);
3431
3432 const XMMRegister xmm_prev_block_cipher = xmm1; // holds cipher of previous block
3433
3434 // registers holding the four results in the parallelized loop
3435 const XMMRegister xmm_result0 = xmm0;
3436 const XMMRegister xmm_result1 = xmm2;
3437 const XMMRegister xmm_result2 = xmm3;
3438 const XMMRegister xmm_result3 = xmm4;
3439
3440 __ movdqu(xmm_prev_block_cipher, Address(rvec, 0x00)); // initialize with initial rvec
3441
3442 __ xorptr(pos, pos);
3443
3444 // now split to different paths depending on the keylen (len in ints of AESCrypt.KLE array (52=192, or 60=256))
3445 __ movl(rbx, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3446 __ cmpl(rbx, 52);
3447 __ jcc(Assembler::equal, L_multiBlock_loopTopHead[1]);
3448 __ cmpl(rbx, 60);
3449 __ jcc(Assembler::equal, L_multiBlock_loopTopHead[2]);
3450
3451 #define DoFour(opc, src_reg) \
3452 __ opc(xmm_result0, src_reg); \
3453 __ opc(xmm_result1, src_reg); \
3454 __ opc(xmm_result2, src_reg); \
3455 __ opc(xmm_result3, src_reg); \
3456
3457 for (int k = 0; k < 3; ++k) {
3458 __ BIND(L_multiBlock_loopTopHead[k]);
3459 if (k != 0) {
3460 __ cmpptr(len_reg, PARALLEL_FACTOR * AESBlockSize); // see if at least 4 blocks left
3461 __ jcc(Assembler::less, L_singleBlock_loopTopHead2[k]);
3462 }
3463 if (k == 1) {
3464 __ subptr(rsp, 6 * wordSize);
3465 __ movdqu(Address(rsp, 0), xmm15); //save last_key from xmm15
3466 load_key(xmm15, key, 0xb0); // 0xb0; 192-bit key goes up to 0xc0
3467 __ movdqu(Address(rsp, 2 * wordSize), xmm15);
3468 load_key(xmm1, key, 0xc0); // 0xc0;
3469 __ movdqu(Address(rsp, 4 * wordSize), xmm1);
3470 } else if (k == 2) {
3471 __ subptr(rsp, 10 * wordSize);
3472 __ movdqu(Address(rsp, 0), xmm15); //save last_key from xmm15
3473 load_key(xmm15, key, 0xd0); // 0xd0; 256-bit key goes upto 0xe0
3474 __ movdqu(Address(rsp, 6 * wordSize), xmm15);
3475 load_key(xmm1, key, 0xe0); // 0xe0;
3476 __ movdqu(Address(rsp, 8 * wordSize), xmm1);
3477 load_key(xmm15, key, 0xb0); // 0xb0;
3478 __ movdqu(Address(rsp, 2 * wordSize), xmm15);
3479 load_key(xmm1, key, 0xc0); // 0xc0;
3480 __ movdqu(Address(rsp, 4 * wordSize), xmm1);
3481 }
3482 __ align(OptoLoopAlignment);
3483 __ BIND(L_multiBlock_loopTop[k]);
3484 __ cmpptr(len_reg, PARALLEL_FACTOR * AESBlockSize); // see if at least 4 blocks left
3485 __ jcc(Assembler::less, L_singleBlock_loopTopHead[k]);
3486
3487 if (k != 0) {
3488 __ movdqu(xmm15, Address(rsp, 2 * wordSize));
3489 __ movdqu(xmm1, Address(rsp, 4 * wordSize));
3490 }
3491
3492 __ movdqu(xmm_result0, Address(from, pos, Address::times_1, 0 * AESBlockSize)); // get next 4 blocks into xmmresult registers
3493 __ movdqu(xmm_result1, Address(from, pos, Address::times_1, 1 * AESBlockSize));
3494 __ movdqu(xmm_result2, Address(from, pos, Address::times_1, 2 * AESBlockSize));
3495 __ movdqu(xmm_result3, Address(from, pos, Address::times_1, 3 * AESBlockSize));
3496
3497 DoFour(pxor, xmm_key_first);
3498 if (k == 0) {
3499 for (int rnum = 1; rnum < ROUNDS[k]; rnum++) {
3500 DoFour(aesdec, as_XMMRegister(rnum + XMM_REG_NUM_KEY_FIRST));
3501 }
3502 DoFour(aesdeclast, xmm_key_last);
3503 } else if (k == 1) {
3504 for (int rnum = 1; rnum <= ROUNDS[k]-2; rnum++) {
3505 DoFour(aesdec, as_XMMRegister(rnum + XMM_REG_NUM_KEY_FIRST));
3506 }
3507 __ movdqu(xmm_key_last, Address(rsp, 0)); // xmm15 needs to be loaded again.
3508 DoFour(aesdec, xmm1); // key : 0xc0
3509 __ movdqu(xmm_prev_block_cipher, Address(rvec, 0x00)); // xmm1 needs to be loaded again
3510 DoFour(aesdeclast, xmm_key_last);
3511 } else if (k == 2) {
3512 for (int rnum = 1; rnum <= ROUNDS[k] - 4; rnum++) {
3513 DoFour(aesdec, as_XMMRegister(rnum + XMM_REG_NUM_KEY_FIRST));
3514 }
3515 DoFour(aesdec, xmm1); // key : 0xc0
3516 __ movdqu(xmm15, Address(rsp, 6 * wordSize));
3517 __ movdqu(xmm1, Address(rsp, 8 * wordSize));
3518 DoFour(aesdec, xmm15); // key : 0xd0
3519 __ movdqu(xmm_key_last, Address(rsp, 0)); // xmm15 needs to be loaded again.
3520 DoFour(aesdec, xmm1); // key : 0xe0
3521 __ movdqu(xmm_prev_block_cipher, Address(rvec, 0x00)); // xmm1 needs to be loaded again
3522 DoFour(aesdeclast, xmm_key_last);
3523 }
3524
3525 // for each result, xor with the r vector of previous cipher block
3526 __ pxor(xmm_result0, xmm_prev_block_cipher);
3527 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 0 * AESBlockSize));
3528 __ pxor(xmm_result1, xmm_prev_block_cipher);
3529 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 1 * AESBlockSize));
3530 __ pxor(xmm_result2, xmm_prev_block_cipher);
3531 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 2 * AESBlockSize));
3532 __ pxor(xmm_result3, xmm_prev_block_cipher);
3533 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 3 * AESBlockSize)); // this will carry over to next set of blocks
3534 if (k != 0) {
3535 __ movdqu(Address(rvec, 0x00), xmm_prev_block_cipher);
3536 }
3537
3538 __ movdqu(Address(to, pos, Address::times_1, 0 * AESBlockSize), xmm_result0); // store 4 results into the next 64 bytes of output
3539 __ movdqu(Address(to, pos, Address::times_1, 1 * AESBlockSize), xmm_result1);
3540 __ movdqu(Address(to, pos, Address::times_1, 2 * AESBlockSize), xmm_result2);
3541 __ movdqu(Address(to, pos, Address::times_1, 3 * AESBlockSize), xmm_result3);
3542
3543 __ addptr(pos, PARALLEL_FACTOR * AESBlockSize);
3544 __ subptr(len_reg, PARALLEL_FACTOR * AESBlockSize);
3545 __ jmp(L_multiBlock_loopTop[k]);
3546
3547 // registers used in the non-parallelized loops
3548 // xmm register assignments for the loops below
3549 const XMMRegister xmm_result = xmm0;
3550 const XMMRegister xmm_prev_block_cipher_save = xmm2;
3551 const XMMRegister xmm_key11 = xmm3;
3552 const XMMRegister xmm_key12 = xmm4;
3553 const XMMRegister key_tmp = xmm4;
3554
3555 __ BIND(L_singleBlock_loopTopHead[k]);
3556 if (k == 1) {
3557 __ addptr(rsp, 6 * wordSize);
3558 } else if (k == 2) {
3559 __ addptr(rsp, 10 * wordSize);
3560 }
3561 __ cmpptr(len_reg, 0); // any blocks left??
3562 __ jcc(Assembler::equal, L_exit);
3563 __ BIND(L_singleBlock_loopTopHead2[k]);
3564 if (k == 1) {
3565 load_key(xmm_key11, key, 0xb0); // 0xb0; 192-bit key goes upto 0xc0
3566 load_key(xmm_key12, key, 0xc0); // 0xc0; 192-bit key goes upto 0xc0
3567 }
3568 if (k == 2) {
3569 load_key(xmm_key11, key, 0xb0); // 0xb0; 256-bit key goes upto 0xe0
3570 }
3571 __ align(OptoLoopAlignment);
3572 __ BIND(L_singleBlock_loopTop[k]);
3573 __ movdqu(xmm_result, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of cipher input
3574 __ movdqa(xmm_prev_block_cipher_save, xmm_result); // save for next r vector
3575 __ pxor(xmm_result, xmm_key_first); // do the aes dec rounds
3576 for (int rnum = 1; rnum <= 9 ; rnum++) {
3577 __ aesdec(xmm_result, as_XMMRegister(rnum + XMM_REG_NUM_KEY_FIRST));
3578 }
3579 if (k == 1) {
3580 __ aesdec(xmm_result, xmm_key11);
3581 __ aesdec(xmm_result, xmm_key12);
3582 }
3583 if (k == 2) {
3584 __ aesdec(xmm_result, xmm_key11);
3585 load_key(key_tmp, key, 0xc0);
3586 __ aesdec(xmm_result, key_tmp);
3587 load_key(key_tmp, key, 0xd0);
3588 __ aesdec(xmm_result, key_tmp);
3589 load_key(key_tmp, key, 0xe0);
3590 __ aesdec(xmm_result, key_tmp);
3591 }
3592
3593 __ aesdeclast(xmm_result, xmm_key_last); // xmm15 always came from key+0
3594 __ pxor(xmm_result, xmm_prev_block_cipher); // xor with the current r vector
3595 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3596 // no need to store r to memory until we exit
3597 __ movdqa(xmm_prev_block_cipher, xmm_prev_block_cipher_save); // set up next r vector with cipher input from this block
3598 __ addptr(pos, AESBlockSize);
3599 __ subptr(len_reg, AESBlockSize);
3600 __ jcc(Assembler::notEqual, L_singleBlock_loopTop[k]);
3601 if (k != 2) {
3602 __ jmp(L_exit);
3603 }
3604 } //for 128/192/256
3605
3606 __ BIND(L_exit);
3607 __ movdqu(Address(rvec, 0), xmm_prev_block_cipher); // final value of r stored in rvec of CipherBlockChaining object
3608 __ pop(rbx);
3609 #ifdef _WIN64
3610 __ movl(rax, len_mem);
3611 #else
3612 __ pop(rax); // return length
3613 #endif
3614 __ leave(); // required for proper stackwalking of RuntimeStub frame
3615 __ ret(0);
3616 return start;
3617 }
3618
3619 address generate_electronicCodeBook_encryptAESCrypt() {
3620 __ align(CodeEntryAlignment);
3621 StubCodeMark mark(this, "StubRoutines", "electronicCodeBook_encryptAESCrypt");
3622 address start = __ pc();
3623 const Register from = c_rarg0; // source array address
3624 const Register to = c_rarg1; // destination array address
3625 const Register key = c_rarg2; // key array address
3626 const Register len = c_rarg3; // src len (must be multiple of blocksize 16)
3627 __ enter(); // required for proper stackwalking of RuntimeStub frame
3628 __ aesecb_encrypt(from, to, key, len);
3629 __ leave(); // required for proper stackwalking of RuntimeStub frame
3630 __ ret(0);
3631 return start;
3632 }
3633
3634 address generate_electronicCodeBook_decryptAESCrypt() {
3635 __ align(CodeEntryAlignment);
3636 StubCodeMark mark(this, "StubRoutines", "electronicCodeBook_decryptAESCrypt");
3637 address start = __ pc();
3638 const Register from = c_rarg0; // source array address
3639 const Register to = c_rarg1; // destination array address
3640 const Register key = c_rarg2; // key array address
3641 const Register len = c_rarg3; // src len (must be multiple of blocksize 16)
3642 __ enter(); // required for proper stackwalking of RuntimeStub frame
3643 __ aesecb_decrypt(from, to, key, len);
3644 __ leave(); // required for proper stackwalking of RuntimeStub frame
3645 __ ret(0);
3646 return start;
3647 }
3648
3649 address generate_upper_word_mask() {
3650 __ align(64);
3651 StubCodeMark mark(this, "StubRoutines", "upper_word_mask");
3652 address start = __ pc();
3653 __ emit_data64(0x0000000000000000, relocInfo::none);
3654 __ emit_data64(0xFFFFFFFF00000000, relocInfo::none);
3655 return start;
3656 }
3657
3658 address generate_shuffle_byte_flip_mask() {
3659 __ align(64);
3660 StubCodeMark mark(this, "StubRoutines", "shuffle_byte_flip_mask");
3661 address start = __ pc();
3662 __ emit_data64(0x08090a0b0c0d0e0f, relocInfo::none);
3663 __ emit_data64(0x0001020304050607, relocInfo::none);
3664 return start;
3665 }
3666
3667 // ofs and limit are use for multi-block byte array.
3668 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
3669 address generate_sha1_implCompress(bool multi_block, const char *name) {
3670 __ align(CodeEntryAlignment);
3671 StubCodeMark mark(this, "StubRoutines", name);
3672 address start = __ pc();
3673
3674 Register buf = c_rarg0;
3675 Register state = c_rarg1;
3676 Register ofs = c_rarg2;
3677 Register limit = c_rarg3;
3678
3679 const XMMRegister abcd = xmm0;
3680 const XMMRegister e0 = xmm1;
3681 const XMMRegister e1 = xmm2;
3682 const XMMRegister msg0 = xmm3;
3683
3684 const XMMRegister msg1 = xmm4;
3685 const XMMRegister msg2 = xmm5;
3686 const XMMRegister msg3 = xmm6;
3687 const XMMRegister shuf_mask = xmm7;
3688
3689 __ enter();
3690
3691 __ subptr(rsp, 4 * wordSize);
3692
3693 __ fast_sha1(abcd, e0, e1, msg0, msg1, msg2, msg3, shuf_mask,
3694 buf, state, ofs, limit, rsp, multi_block);
3695
3696 __ addptr(rsp, 4 * wordSize);
3697
3698 __ leave();
3699 __ ret(0);
3700 return start;
3701 }
3702
3703 address generate_pshuffle_byte_flip_mask() {
3704 __ align(64);
3705 StubCodeMark mark(this, "StubRoutines", "pshuffle_byte_flip_mask");
3706 address start = __ pc();
3707 __ emit_data64(0x0405060700010203, relocInfo::none);
3708 __ emit_data64(0x0c0d0e0f08090a0b, relocInfo::none);
3709
3710 if (VM_Version::supports_avx2()) {
3711 __ emit_data64(0x0405060700010203, relocInfo::none); // second copy
3712 __ emit_data64(0x0c0d0e0f08090a0b, relocInfo::none);
3713 // _SHUF_00BA
3714 __ emit_data64(0x0b0a090803020100, relocInfo::none);
3715 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
3716 __ emit_data64(0x0b0a090803020100, relocInfo::none);
3717 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
3718 // _SHUF_DC00
3719 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
3720 __ emit_data64(0x0b0a090803020100, relocInfo::none);
3721 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
3722 __ emit_data64(0x0b0a090803020100, relocInfo::none);
3723 }
3724
3725 return start;
3726 }
3727
3728 //Mask for byte-swapping a couple of qwords in an XMM register using (v)pshufb.
3729 address generate_pshuffle_byte_flip_mask_sha512() {
3730 __ align(32);
3731 StubCodeMark mark(this, "StubRoutines", "pshuffle_byte_flip_mask_sha512");
3732 address start = __ pc();
3733 if (VM_Version::supports_avx2()) {
3734 __ emit_data64(0x0001020304050607, relocInfo::none); // PSHUFFLE_BYTE_FLIP_MASK
3735 __ emit_data64(0x08090a0b0c0d0e0f, relocInfo::none);
3736 __ emit_data64(0x1011121314151617, relocInfo::none);
3737 __ emit_data64(0x18191a1b1c1d1e1f, relocInfo::none);
3738 __ emit_data64(0x0000000000000000, relocInfo::none); //MASK_YMM_LO
3739 __ emit_data64(0x0000000000000000, relocInfo::none);
3740 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
3741 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
3742 }
3743
3744 return start;
3745 }
3746
3747 // ofs and limit are use for multi-block byte array.
3748 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
3749 address generate_sha256_implCompress(bool multi_block, const char *name) {
3750 assert(VM_Version::supports_sha() || VM_Version::supports_avx2(), "");
3751 __ align(CodeEntryAlignment);
3752 StubCodeMark mark(this, "StubRoutines", name);
3753 address start = __ pc();
3754
3755 Register buf = c_rarg0;
3756 Register state = c_rarg1;
3757 Register ofs = c_rarg2;
3758 Register limit = c_rarg3;
3759
3760 const XMMRegister msg = xmm0;
3761 const XMMRegister state0 = xmm1;
3762 const XMMRegister state1 = xmm2;
3763 const XMMRegister msgtmp0 = xmm3;
3764
3765 const XMMRegister msgtmp1 = xmm4;
3766 const XMMRegister msgtmp2 = xmm5;
3767 const XMMRegister msgtmp3 = xmm6;
3768 const XMMRegister msgtmp4 = xmm7;
3769
3770 const XMMRegister shuf_mask = xmm8;
3771
3772 __ enter();
3773
3774 __ subptr(rsp, 4 * wordSize);
3775
3776 if (VM_Version::supports_sha()) {
3777 __ fast_sha256(msg, state0, state1, msgtmp0, msgtmp1, msgtmp2, msgtmp3, msgtmp4,
3778 buf, state, ofs, limit, rsp, multi_block, shuf_mask);
3779 } else if (VM_Version::supports_avx2()) {
3780 __ sha256_AVX2(msg, state0, state1, msgtmp0, msgtmp1, msgtmp2, msgtmp3, msgtmp4,
3781 buf, state, ofs, limit, rsp, multi_block, shuf_mask);
3782 }
3783 __ addptr(rsp, 4 * wordSize);
3784 __ vzeroupper();
3785 __ leave();
3786 __ ret(0);
3787 return start;
3788 }
3789
3790 address generate_sha512_implCompress(bool multi_block, const char *name) {
3791 assert(VM_Version::supports_avx2(), "");
3792 assert(VM_Version::supports_bmi2(), "");
3793 __ align(CodeEntryAlignment);
3794 StubCodeMark mark(this, "StubRoutines", name);
3795 address start = __ pc();
3796
3797 Register buf = c_rarg0;
3798 Register state = c_rarg1;
3799 Register ofs = c_rarg2;
3800 Register limit = c_rarg3;
3801
3802 const XMMRegister msg = xmm0;
3803 const XMMRegister state0 = xmm1;
3804 const XMMRegister state1 = xmm2;
3805 const XMMRegister msgtmp0 = xmm3;
3806 const XMMRegister msgtmp1 = xmm4;
3807 const XMMRegister msgtmp2 = xmm5;
3808 const XMMRegister msgtmp3 = xmm6;
3809 const XMMRegister msgtmp4 = xmm7;
3810
3811 const XMMRegister shuf_mask = xmm8;
3812
3813 __ enter();
3814
3815 __ sha512_AVX2(msg, state0, state1, msgtmp0, msgtmp1, msgtmp2, msgtmp3, msgtmp4,
3816 buf, state, ofs, limit, rsp, multi_block, shuf_mask);
3817
3818 __ vzeroupper();
3819 __ leave();
3820 __ ret(0);
3821 return start;
3822 }
3823
3824 // This mask is used for incrementing counter value(linc0, linc4, etc.)
3825 address counter_mask_addr() {
3826 __ align(64);
3827 StubCodeMark mark(this, "StubRoutines", "counter_mask_addr");
3828 address start = __ pc();
3829 __ emit_data64(0x08090a0b0c0d0e0f, relocInfo::none);//lbswapmask
3830 __ emit_data64(0x0001020304050607, relocInfo::none);
3831 __ emit_data64(0x08090a0b0c0d0e0f, relocInfo::none);
3832 __ emit_data64(0x0001020304050607, relocInfo::none);
3833 __ emit_data64(0x08090a0b0c0d0e0f, relocInfo::none);
3834 __ emit_data64(0x0001020304050607, relocInfo::none);
3835 __ emit_data64(0x08090a0b0c0d0e0f, relocInfo::none);
3836 __ emit_data64(0x0001020304050607, relocInfo::none);
3837 __ emit_data64(0x0000000000000000, relocInfo::none);//linc0 = counter_mask_addr+64
3838 __ emit_data64(0x0000000000000000, relocInfo::none);
3839 __ emit_data64(0x0000000000000001, relocInfo::none);//counter_mask_addr() + 80
3840 __ emit_data64(0x0000000000000000, relocInfo::none);
3841 __ emit_data64(0x0000000000000002, relocInfo::none);
3842 __ emit_data64(0x0000000000000000, relocInfo::none);
3843 __ emit_data64(0x0000000000000003, relocInfo::none);
3844 __ emit_data64(0x0000000000000000, relocInfo::none);
3845 __ emit_data64(0x0000000000000004, relocInfo::none);//linc4 = counter_mask_addr() + 128
3846 __ emit_data64(0x0000000000000000, relocInfo::none);
3847 __ emit_data64(0x0000000000000004, relocInfo::none);
3848 __ emit_data64(0x0000000000000000, relocInfo::none);
3849 __ emit_data64(0x0000000000000004, relocInfo::none);
3850 __ emit_data64(0x0000000000000000, relocInfo::none);
3851 __ emit_data64(0x0000000000000004, relocInfo::none);
3852 __ emit_data64(0x0000000000000000, relocInfo::none);
3853 __ emit_data64(0x0000000000000008, relocInfo::none);//linc8 = counter_mask_addr() + 192
3854 __ emit_data64(0x0000000000000000, relocInfo::none);
3855 __ emit_data64(0x0000000000000008, relocInfo::none);
3856 __ emit_data64(0x0000000000000000, relocInfo::none);
3857 __ emit_data64(0x0000000000000008, relocInfo::none);
3858 __ emit_data64(0x0000000000000000, relocInfo::none);
3859 __ emit_data64(0x0000000000000008, relocInfo::none);
3860 __ emit_data64(0x0000000000000000, relocInfo::none);
3861 __ emit_data64(0x0000000000000020, relocInfo::none);//linc32 = counter_mask_addr() + 256
3862 __ emit_data64(0x0000000000000000, relocInfo::none);
3863 __ emit_data64(0x0000000000000020, relocInfo::none);
3864 __ emit_data64(0x0000000000000000, relocInfo::none);
3865 __ emit_data64(0x0000000000000020, relocInfo::none);
3866 __ emit_data64(0x0000000000000000, relocInfo::none);
3867 __ emit_data64(0x0000000000000020, relocInfo::none);
3868 __ emit_data64(0x0000000000000000, relocInfo::none);
3869 __ emit_data64(0x0000000000000010, relocInfo::none);//linc16 = counter_mask_addr() + 320
3870 __ emit_data64(0x0000000000000000, relocInfo::none);
3871 __ emit_data64(0x0000000000000010, relocInfo::none);
3872 __ emit_data64(0x0000000000000000, relocInfo::none);
3873 __ emit_data64(0x0000000000000010, relocInfo::none);
3874 __ emit_data64(0x0000000000000000, relocInfo::none);
3875 __ emit_data64(0x0000000000000010, relocInfo::none);
3876 __ emit_data64(0x0000000000000000, relocInfo::none);
3877 return start;
3878 }
3879
3880 // Vector AES Counter implementation
3881 address generate_counterMode_VectorAESCrypt() {
3882 __ align(CodeEntryAlignment);
3883 StubCodeMark mark(this, "StubRoutines", "counterMode_AESCrypt");
3884 address start = __ pc();
3885 const Register from = c_rarg0; // source array address
3886 const Register to = c_rarg1; // destination array address
3887 const Register key = c_rarg2; // key array address r8
3888 const Register counter = c_rarg3; // counter byte array initialized from counter array address
3889 // and updated with the incremented counter in the end
3890 #ifndef _WIN64
3891 const Register len_reg = c_rarg4;
3892 const Register saved_encCounter_start = c_rarg5;
3893 const Register used_addr = r10;
3894 const Address used_mem(rbp, 2 * wordSize);
3895 const Register used = r11;
3896 #else
3897 const Address len_mem(rbp, 6 * wordSize); // length is on stack on Win64
3898 const Address saved_encCounter_mem(rbp, 7 * wordSize); // saved encrypted counter is on stack on Win64
3899 const Address used_mem(rbp, 8 * wordSize); // used length is on stack on Win64
3900 const Register len_reg = r10; // pick the first volatile windows register
3901 const Register saved_encCounter_start = r11;
3902 const Register used_addr = r13;
3903 const Register used = r14;
3904 #endif
3905 __ enter();
3906 // Save state before entering routine
3907 __ push(r12);
3908 __ push(r13);
3909 __ push(r14);
3910 __ push(r15);
3911 #ifdef _WIN64
3912 // on win64, fill len_reg from stack position
3913 __ movl(len_reg, len_mem);
3914 __ movptr(saved_encCounter_start, saved_encCounter_mem);
3915 __ movptr(used_addr, used_mem);
3916 __ movl(used, Address(used_addr, 0));
3917 #else
3918 __ push(len_reg); // Save
3919 __ movptr(used_addr, used_mem);
3920 __ movl(used, Address(used_addr, 0));
3921 #endif
3922 __ push(rbx);
3923 __ aesctr_encrypt(from, to, key, counter, len_reg, used, used_addr, saved_encCounter_start);
3924 // Restore state before leaving routine
3925 __ pop(rbx);
3926 #ifdef _WIN64
3927 __ movl(rax, len_mem); // return length
3928 #else
3929 __ pop(rax); // return length
3930 #endif
3931 __ pop(r15);
3932 __ pop(r14);
3933 __ pop(r13);
3934 __ pop(r12);
3935
3936 __ leave(); // required for proper stackwalking of RuntimeStub frame
3937 __ ret(0);
3938 return start;
3939 }
3940
3941 // This is a version of CTR/AES crypt which does 6 blocks in a loop at a time
3942 // to hide instruction latency
3943 //
3944 // Arguments:
3945 //
3946 // Inputs:
3947 // c_rarg0 - source byte array address
3948 // c_rarg1 - destination byte array address
3949 // c_rarg2 - K (key) in little endian int array
3950 // c_rarg3 - counter vector byte array address
3951 // Linux
3952 // c_rarg4 - input length
3953 // c_rarg5 - saved encryptedCounter start
3954 // rbp + 6 * wordSize - saved used length
3955 // Windows
3956 // rbp + 6 * wordSize - input length
3957 // rbp + 7 * wordSize - saved encryptedCounter start
3958 // rbp + 8 * wordSize - saved used length
3959 //
3960 // Output:
3961 // rax - input length
3962 //
3963 address generate_counterMode_AESCrypt_Parallel() {
3964 assert(UseAES, "need AES instructions and misaligned SSE support");
3965 __ align(CodeEntryAlignment);
3966 StubCodeMark mark(this, "StubRoutines", "counterMode_AESCrypt");
3967 address start = __ pc();
3968 const Register from = c_rarg0; // source array address
3969 const Register to = c_rarg1; // destination array address
3970 const Register key = c_rarg2; // key array address
3971 const Register counter = c_rarg3; // counter byte array initialized from counter array address
3972 // and updated with the incremented counter in the end
3973 #ifndef _WIN64
3974 const Register len_reg = c_rarg4;
3975 const Register saved_encCounter_start = c_rarg5;
3976 const Register used_addr = r10;
3977 const Address used_mem(rbp, 2 * wordSize);
3978 const Register used = r11;
3979 #else
3980 const Address len_mem(rbp, 6 * wordSize); // length is on stack on Win64
3981 const Address saved_encCounter_mem(rbp, 7 * wordSize); // length is on stack on Win64
3982 const Address used_mem(rbp, 8 * wordSize); // length is on stack on Win64
3983 const Register len_reg = r10; // pick the first volatile windows register
3984 const Register saved_encCounter_start = r11;
3985 const Register used_addr = r13;
3986 const Register used = r14;
3987 #endif
3988 const Register pos = rax;
3989
3990 const int PARALLEL_FACTOR = 6;
3991 const XMMRegister xmm_counter_shuf_mask = xmm0;
3992 const XMMRegister xmm_key_shuf_mask = xmm1; // used temporarily to swap key bytes up front
3993 const XMMRegister xmm_curr_counter = xmm2;
3994
3995 const XMMRegister xmm_key_tmp0 = xmm3;
3996 const XMMRegister xmm_key_tmp1 = xmm4;
3997
3998 // registers holding the four results in the parallelized loop
3999 const XMMRegister xmm_result0 = xmm5;
4000 const XMMRegister xmm_result1 = xmm6;
4001 const XMMRegister xmm_result2 = xmm7;
4002 const XMMRegister xmm_result3 = xmm8;
4003 const XMMRegister xmm_result4 = xmm9;
4004 const XMMRegister xmm_result5 = xmm10;
4005
4006 const XMMRegister xmm_from0 = xmm11;
4007 const XMMRegister xmm_from1 = xmm12;
4008 const XMMRegister xmm_from2 = xmm13;
4009 const XMMRegister xmm_from3 = xmm14; //the last one is xmm14. we have to preserve it on WIN64.
4010 const XMMRegister xmm_from4 = xmm3; //reuse xmm3~4. Because xmm_key_tmp0~1 are useless when loading input text
4011 const XMMRegister xmm_from5 = xmm4;
4012
4013 //for key_128, key_192, key_256
4014 const int rounds[3] = {10, 12, 14};
4015 Label L_exit_preLoop, L_preLoop_start;
4016 Label L_multiBlock_loopTop[3];
4017 Label L_singleBlockLoopTop[3];
4018 Label L__incCounter[3][6]; //for 6 blocks
4019 Label L__incCounter_single[3]; //for single block, key128, key192, key256
4020 Label L_processTail_insr[3], L_processTail_4_insr[3], L_processTail_2_insr[3], L_processTail_1_insr[3], L_processTail_exit_insr[3];
4021 Label L_processTail_4_extr[3], L_processTail_2_extr[3], L_processTail_1_extr[3], L_processTail_exit_extr[3];
4022
4023 Label L_exit;
4024
4025 __ enter(); // required for proper stackwalking of RuntimeStub frame
4026
4027 #ifdef _WIN64
4028 // allocate spill slots for r13, r14
4029 enum {
4030 saved_r13_offset,
4031 saved_r14_offset
4032 };
4033 __ subptr(rsp, 2 * wordSize);
4034 __ movptr(Address(rsp, saved_r13_offset * wordSize), r13);
4035 __ movptr(Address(rsp, saved_r14_offset * wordSize), r14);
4036
4037 // on win64, fill len_reg from stack position
4038 __ movl(len_reg, len_mem);
4039 __ movptr(saved_encCounter_start, saved_encCounter_mem);
4040 __ movptr(used_addr, used_mem);
4041 __ movl(used, Address(used_addr, 0));
4042 #else
4043 __ push(len_reg); // Save
4044 __ movptr(used_addr, used_mem);
4045 __ movl(used, Address(used_addr, 0));
4046 #endif
4047
4048 __ push(rbx); // Save RBX
4049 __ movdqu(xmm_curr_counter, Address(counter, 0x00)); // initialize counter with initial counter
4050 __ movdqu(xmm_counter_shuf_mask, ExternalAddress(StubRoutines::x86::counter_shuffle_mask_addr()), pos); // pos as scratch
4051 __ pshufb(xmm_curr_counter, xmm_counter_shuf_mask); //counter is shuffled
4052 __ movptr(pos, 0);
4053
4054 // Use the partially used encrpyted counter from last invocation
4055 __ BIND(L_preLoop_start);
4056 __ cmpptr(used, 16);
4057 __ jcc(Assembler::aboveEqual, L_exit_preLoop);
4058 __ cmpptr(len_reg, 0);
4059 __ jcc(Assembler::lessEqual, L_exit_preLoop);
4060 __ movb(rbx, Address(saved_encCounter_start, used));
4061 __ xorb(rbx, Address(from, pos));
4062 __ movb(Address(to, pos), rbx);
4063 __ addptr(pos, 1);
4064 __ addptr(used, 1);
4065 __ subptr(len_reg, 1);
4066
4067 __ jmp(L_preLoop_start);
4068
4069 __ BIND(L_exit_preLoop);
4070 __ movl(Address(used_addr, 0), used);
4071
4072 // key length could be only {11, 13, 15} * 4 = {44, 52, 60}
4073 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()), rbx); // rbx as scratch
4074 __ movl(rbx, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
4075 __ cmpl(rbx, 52);
4076 __ jcc(Assembler::equal, L_multiBlock_loopTop[1]);
4077 __ cmpl(rbx, 60);
4078 __ jcc(Assembler::equal, L_multiBlock_loopTop[2]);
4079
4080 #define CTR_DoSix(opc, src_reg) \
4081 __ opc(xmm_result0, src_reg); \
4082 __ opc(xmm_result1, src_reg); \
4083 __ opc(xmm_result2, src_reg); \
4084 __ opc(xmm_result3, src_reg); \
4085 __ opc(xmm_result4, src_reg); \
4086 __ opc(xmm_result5, src_reg);
4087
4088 // k == 0 : generate code for key_128
4089 // k == 1 : generate code for key_192
4090 // k == 2 : generate code for key_256
4091 for (int k = 0; k < 3; ++k) {
4092 //multi blocks starts here
4093 __ align(OptoLoopAlignment);
4094 __ BIND(L_multiBlock_loopTop[k]);
4095 __ cmpptr(len_reg, PARALLEL_FACTOR * AESBlockSize); // see if at least PARALLEL_FACTOR blocks left
4096 __ jcc(Assembler::less, L_singleBlockLoopTop[k]);
4097 load_key(xmm_key_tmp0, key, 0x00, xmm_key_shuf_mask);
4098
4099 //load, then increase counters
4100 CTR_DoSix(movdqa, xmm_curr_counter);
4101 inc_counter(rbx, xmm_result1, 0x01, L__incCounter[k][0]);
4102 inc_counter(rbx, xmm_result2, 0x02, L__incCounter[k][1]);
4103 inc_counter(rbx, xmm_result3, 0x03, L__incCounter[k][2]);
4104 inc_counter(rbx, xmm_result4, 0x04, L__incCounter[k][3]);
4105 inc_counter(rbx, xmm_result5, 0x05, L__incCounter[k][4]);
4106 inc_counter(rbx, xmm_curr_counter, 0x06, L__incCounter[k][5]);
4107 CTR_DoSix(pshufb, xmm_counter_shuf_mask); // after increased, shuffled counters back for PXOR
4108 CTR_DoSix(pxor, xmm_key_tmp0); //PXOR with Round 0 key
4109
4110 //load two ROUND_KEYs at a time
4111 for (int i = 1; i < rounds[k]; ) {
4112 load_key(xmm_key_tmp1, key, (0x10 * i), xmm_key_shuf_mask);
4113 load_key(xmm_key_tmp0, key, (0x10 * (i+1)), xmm_key_shuf_mask);
4114 CTR_DoSix(aesenc, xmm_key_tmp1);
4115 i++;
4116 if (i != rounds[k]) {
4117 CTR_DoSix(aesenc, xmm_key_tmp0);
4118 } else {
4119 CTR_DoSix(aesenclast, xmm_key_tmp0);
4120 }
4121 i++;
4122 }
4123
4124 // get next PARALLEL_FACTOR blocks into xmm_result registers
4125 __ movdqu(xmm_from0, Address(from, pos, Address::times_1, 0 * AESBlockSize));
4126 __ movdqu(xmm_from1, Address(from, pos, Address::times_1, 1 * AESBlockSize));
4127 __ movdqu(xmm_from2, Address(from, pos, Address::times_1, 2 * AESBlockSize));
4128 __ movdqu(xmm_from3, Address(from, pos, Address::times_1, 3 * AESBlockSize));
4129 __ movdqu(xmm_from4, Address(from, pos, Address::times_1, 4 * AESBlockSize));
4130 __ movdqu(xmm_from5, Address(from, pos, Address::times_1, 5 * AESBlockSize));
4131
4132 __ pxor(xmm_result0, xmm_from0);
4133 __ pxor(xmm_result1, xmm_from1);
4134 __ pxor(xmm_result2, xmm_from2);
4135 __ pxor(xmm_result3, xmm_from3);
4136 __ pxor(xmm_result4, xmm_from4);
4137 __ pxor(xmm_result5, xmm_from5);
4138
4139 // store 6 results into the next 64 bytes of output
4140 __ movdqu(Address(to, pos, Address::times_1, 0 * AESBlockSize), xmm_result0);
4141 __ movdqu(Address(to, pos, Address::times_1, 1 * AESBlockSize), xmm_result1);
4142 __ movdqu(Address(to, pos, Address::times_1, 2 * AESBlockSize), xmm_result2);
4143 __ movdqu(Address(to, pos, Address::times_1, 3 * AESBlockSize), xmm_result3);
4144 __ movdqu(Address(to, pos, Address::times_1, 4 * AESBlockSize), xmm_result4);
4145 __ movdqu(Address(to, pos, Address::times_1, 5 * AESBlockSize), xmm_result5);
4146
4147 __ addptr(pos, PARALLEL_FACTOR * AESBlockSize); // increase the length of crypt text
4148 __ subptr(len_reg, PARALLEL_FACTOR * AESBlockSize); // decrease the remaining length
4149 __ jmp(L_multiBlock_loopTop[k]);
4150
4151 // singleBlock starts here
4152 __ align(OptoLoopAlignment);
4153 __ BIND(L_singleBlockLoopTop[k]);
4154 __ cmpptr(len_reg, 0);
4155 __ jcc(Assembler::lessEqual, L_exit);
4156 load_key(xmm_key_tmp0, key, 0x00, xmm_key_shuf_mask);
4157 __ movdqa(xmm_result0, xmm_curr_counter);
4158 inc_counter(rbx, xmm_curr_counter, 0x01, L__incCounter_single[k]);
4159 __ pshufb(xmm_result0, xmm_counter_shuf_mask);
4160 __ pxor(xmm_result0, xmm_key_tmp0);
4161 for (int i = 1; i < rounds[k]; i++) {
4162 load_key(xmm_key_tmp0, key, (0x10 * i), xmm_key_shuf_mask);
4163 __ aesenc(xmm_result0, xmm_key_tmp0);
4164 }
4165 load_key(xmm_key_tmp0, key, (rounds[k] * 0x10), xmm_key_shuf_mask);
4166 __ aesenclast(xmm_result0, xmm_key_tmp0);
4167 __ cmpptr(len_reg, AESBlockSize);
4168 __ jcc(Assembler::less, L_processTail_insr[k]);
4169 __ movdqu(xmm_from0, Address(from, pos, Address::times_1, 0 * AESBlockSize));
4170 __ pxor(xmm_result0, xmm_from0);
4171 __ movdqu(Address(to, pos, Address::times_1, 0 * AESBlockSize), xmm_result0);
4172 __ addptr(pos, AESBlockSize);
4173 __ subptr(len_reg, AESBlockSize);
4174 __ jmp(L_singleBlockLoopTop[k]);
4175 __ BIND(L_processTail_insr[k]); // Process the tail part of the input array
4176 __ addptr(pos, len_reg); // 1. Insert bytes from src array into xmm_from0 register
4177 __ testptr(len_reg, 8);
4178 __ jcc(Assembler::zero, L_processTail_4_insr[k]);
4179 __ subptr(pos,8);
4180 __ pinsrq(xmm_from0, Address(from, pos), 0);
4181 __ BIND(L_processTail_4_insr[k]);
4182 __ testptr(len_reg, 4);
4183 __ jcc(Assembler::zero, L_processTail_2_insr[k]);
4184 __ subptr(pos,4);
4185 __ pslldq(xmm_from0, 4);
4186 __ pinsrd(xmm_from0, Address(from, pos), 0);
4187 __ BIND(L_processTail_2_insr[k]);
4188 __ testptr(len_reg, 2);
4189 __ jcc(Assembler::zero, L_processTail_1_insr[k]);
4190 __ subptr(pos, 2);
4191 __ pslldq(xmm_from0, 2);
4192 __ pinsrw(xmm_from0, Address(from, pos), 0);
4193 __ BIND(L_processTail_1_insr[k]);
4194 __ testptr(len_reg, 1);
4195 __ jcc(Assembler::zero, L_processTail_exit_insr[k]);
4196 __ subptr(pos, 1);
4197 __ pslldq(xmm_from0, 1);
4198 __ pinsrb(xmm_from0, Address(from, pos), 0);
4199 __ BIND(L_processTail_exit_insr[k]);
4200
4201 __ movdqu(Address(saved_encCounter_start, 0), xmm_result0); // 2. Perform pxor of the encrypted counter and plaintext Bytes.
4202 __ pxor(xmm_result0, xmm_from0); // Also the encrypted counter is saved for next invocation.
4203
4204 __ testptr(len_reg, 8);
4205 __ jcc(Assembler::zero, L_processTail_4_extr[k]); // 3. Extract bytes from xmm_result0 into the dest. array
4206 __ pextrq(Address(to, pos), xmm_result0, 0);
4207 __ psrldq(xmm_result0, 8);
4208 __ addptr(pos, 8);
4209 __ BIND(L_processTail_4_extr[k]);
4210 __ testptr(len_reg, 4);
4211 __ jcc(Assembler::zero, L_processTail_2_extr[k]);
4212 __ pextrd(Address(to, pos), xmm_result0, 0);
4213 __ psrldq(xmm_result0, 4);
4214 __ addptr(pos, 4);
4215 __ BIND(L_processTail_2_extr[k]);
4216 __ testptr(len_reg, 2);
4217 __ jcc(Assembler::zero, L_processTail_1_extr[k]);
4218 __ pextrw(Address(to, pos), xmm_result0, 0);
4219 __ psrldq(xmm_result0, 2);
4220 __ addptr(pos, 2);
4221 __ BIND(L_processTail_1_extr[k]);
4222 __ testptr(len_reg, 1);
4223 __ jcc(Assembler::zero, L_processTail_exit_extr[k]);
4224 __ pextrb(Address(to, pos), xmm_result0, 0);
4225
4226 __ BIND(L_processTail_exit_extr[k]);
4227 __ movl(Address(used_addr, 0), len_reg);
4228 __ jmp(L_exit);
4229
4230 }
4231
4232 __ BIND(L_exit);
4233 __ pshufb(xmm_curr_counter, xmm_counter_shuf_mask); //counter is shuffled back.
4234 __ movdqu(Address(counter, 0), xmm_curr_counter); //save counter back
4235 __ pop(rbx); // pop the saved RBX.
4236 #ifdef _WIN64
4237 __ movl(rax, len_mem);
4238 __ movptr(r13, Address(rsp, saved_r13_offset * wordSize));
4239 __ movptr(r14, Address(rsp, saved_r14_offset * wordSize));
4240 __ addptr(rsp, 2 * wordSize);
4241 #else
4242 __ pop(rax); // return 'len'
4243 #endif
4244 __ leave(); // required for proper stackwalking of RuntimeStub frame
4245 __ ret(0);
4246 return start;
4247 }
4248
4249 void roundDec(XMMRegister xmm_reg) {
4250 __ vaesdec(xmm1, xmm1, xmm_reg, Assembler::AVX_512bit);
4251 __ vaesdec(xmm2, xmm2, xmm_reg, Assembler::AVX_512bit);
4252 __ vaesdec(xmm3, xmm3, xmm_reg, Assembler::AVX_512bit);
4253 __ vaesdec(xmm4, xmm4, xmm_reg, Assembler::AVX_512bit);
4254 __ vaesdec(xmm5, xmm5, xmm_reg, Assembler::AVX_512bit);
4255 __ vaesdec(xmm6, xmm6, xmm_reg, Assembler::AVX_512bit);
4256 __ vaesdec(xmm7, xmm7, xmm_reg, Assembler::AVX_512bit);
4257 __ vaesdec(xmm8, xmm8, xmm_reg, Assembler::AVX_512bit);
4258 }
4259
4260 void roundDeclast(XMMRegister xmm_reg) {
4261 __ vaesdeclast(xmm1, xmm1, xmm_reg, Assembler::AVX_512bit);
4262 __ vaesdeclast(xmm2, xmm2, xmm_reg, Assembler::AVX_512bit);
4263 __ vaesdeclast(xmm3, xmm3, xmm_reg, Assembler::AVX_512bit);
4264 __ vaesdeclast(xmm4, xmm4, xmm_reg, Assembler::AVX_512bit);
4265 __ vaesdeclast(xmm5, xmm5, xmm_reg, Assembler::AVX_512bit);
4266 __ vaesdeclast(xmm6, xmm6, xmm_reg, Assembler::AVX_512bit);
4267 __ vaesdeclast(xmm7, xmm7, xmm_reg, Assembler::AVX_512bit);
4268 __ vaesdeclast(xmm8, xmm8, xmm_reg, Assembler::AVX_512bit);
4269 }
4270
4271 void ev_load_key(XMMRegister xmmdst, Register key, int offset, XMMRegister xmm_shuf_mask = NULL) {
4272 __ movdqu(xmmdst, Address(key, offset));
4273 if (xmm_shuf_mask != NULL) {
4274 __ pshufb(xmmdst, xmm_shuf_mask);
4275 } else {
4276 __ pshufb(xmmdst, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
4277 }
4278 __ evshufi64x2(xmmdst, xmmdst, xmmdst, 0x0, Assembler::AVX_512bit);
4279
4280 }
4281
4282 address generate_cipherBlockChaining_decryptVectorAESCrypt() {
4283 assert(VM_Version::supports_avx512_vaes(), "need AES instructions and misaligned SSE support");
4284 __ align(CodeEntryAlignment);
4285 StubCodeMark mark(this, "StubRoutines", "cipherBlockChaining_decryptAESCrypt");
4286 address start = __ pc();
4287
4288 const Register from = c_rarg0; // source array address
4289 const Register to = c_rarg1; // destination array address
4290 const Register key = c_rarg2; // key array address
4291 const Register rvec = c_rarg3; // r byte array initialized from initvector array address
4292 // and left with the results of the last encryption block
4293 #ifndef _WIN64
4294 const Register len_reg = c_rarg4; // src len (must be multiple of blocksize 16)
4295 #else
4296 const Address len_mem(rbp, 6 * wordSize); // length is on stack on Win64
4297 const Register len_reg = r11; // pick the volatile windows register
4298 #endif
4299
4300 Label Loop, Loop1, L_128, L_256, L_192, KEY_192, KEY_256, Loop2, Lcbc_dec_rem_loop,
4301 Lcbc_dec_rem_last, Lcbc_dec_ret, Lcbc_dec_rem, Lcbc_exit;
4302
4303 __ enter();
4304
4305 #ifdef _WIN64
4306 // on win64, fill len_reg from stack position
4307 __ movl(len_reg, len_mem);
4308 #else
4309 __ push(len_reg); // Save
4310 #endif
4311 __ push(rbx);
4312 __ vzeroupper();
4313
4314 // Temporary variable declaration for swapping key bytes
4315 const XMMRegister xmm_key_shuf_mask = xmm1;
4316 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
4317
4318 // Calculate number of rounds from key size: 44 for 10-rounds, 52 for 12-rounds, 60 for 14-rounds
4319 const Register rounds = rbx;
4320 __ movl(rounds, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
4321
4322 const XMMRegister IV = xmm0;
4323 // Load IV and broadcast value to 512-bits
4324 __ evbroadcasti64x2(IV, Address(rvec, 0), Assembler::AVX_512bit);
4325
4326 // Temporary variables for storing round keys
4327 const XMMRegister RK0 = xmm30;
4328 const XMMRegister RK1 = xmm9;
4329 const XMMRegister RK2 = xmm18;
4330 const XMMRegister RK3 = xmm19;
4331 const XMMRegister RK4 = xmm20;
4332 const XMMRegister RK5 = xmm21;
4333 const XMMRegister RK6 = xmm22;
4334 const XMMRegister RK7 = xmm23;
4335 const XMMRegister RK8 = xmm24;
4336 const XMMRegister RK9 = xmm25;
4337 const XMMRegister RK10 = xmm26;
4338
4339 // Load and shuffle key
4340 // the java expanded key ordering is rotated one position from what we want
4341 // so we start from 1*16 here and hit 0*16 last
4342 ev_load_key(RK1, key, 1 * 16, xmm_key_shuf_mask);
4343 ev_load_key(RK2, key, 2 * 16, xmm_key_shuf_mask);
4344 ev_load_key(RK3, key, 3 * 16, xmm_key_shuf_mask);
4345 ev_load_key(RK4, key, 4 * 16, xmm_key_shuf_mask);
4346 ev_load_key(RK5, key, 5 * 16, xmm_key_shuf_mask);
4347 ev_load_key(RK6, key, 6 * 16, xmm_key_shuf_mask);
4348 ev_load_key(RK7, key, 7 * 16, xmm_key_shuf_mask);
4349 ev_load_key(RK8, key, 8 * 16, xmm_key_shuf_mask);
4350 ev_load_key(RK9, key, 9 * 16, xmm_key_shuf_mask);
4351 ev_load_key(RK10, key, 10 * 16, xmm_key_shuf_mask);
4352 ev_load_key(RK0, key, 0*16, xmm_key_shuf_mask);
4353
4354 // Variables for storing source cipher text
4355 const XMMRegister S0 = xmm10;
4356 const XMMRegister S1 = xmm11;
4357 const XMMRegister S2 = xmm12;
4358 const XMMRegister S3 = xmm13;
4359 const XMMRegister S4 = xmm14;
4360 const XMMRegister S5 = xmm15;
4361 const XMMRegister S6 = xmm16;
4362 const XMMRegister S7 = xmm17;
4363
4364 // Variables for storing decrypted text
4365 const XMMRegister B0 = xmm1;
4366 const XMMRegister B1 = xmm2;
4367 const XMMRegister B2 = xmm3;
4368 const XMMRegister B3 = xmm4;
4369 const XMMRegister B4 = xmm5;
4370 const XMMRegister B5 = xmm6;
4371 const XMMRegister B6 = xmm7;
4372 const XMMRegister B7 = xmm8;
4373
4374 __ cmpl(rounds, 44);
4375 __ jcc(Assembler::greater, KEY_192);
4376 __ jmp(Loop);
4377
4378 __ BIND(KEY_192);
4379 const XMMRegister RK11 = xmm27;
4380 const XMMRegister RK12 = xmm28;
4381 ev_load_key(RK11, key, 11*16, xmm_key_shuf_mask);
4382 ev_load_key(RK12, key, 12*16, xmm_key_shuf_mask);
4383
4384 __ cmpl(rounds, 52);
4385 __ jcc(Assembler::greater, KEY_256);
4386 __ jmp(Loop);
4387
4388 __ BIND(KEY_256);
4389 const XMMRegister RK13 = xmm29;
4390 const XMMRegister RK14 = xmm31;
4391 ev_load_key(RK13, key, 13*16, xmm_key_shuf_mask);
4392 ev_load_key(RK14, key, 14*16, xmm_key_shuf_mask);
4393
4394 __ BIND(Loop);
4395 __ cmpl(len_reg, 512);
4396 __ jcc(Assembler::below, Lcbc_dec_rem);
4397 __ BIND(Loop1);
4398 __ subl(len_reg, 512);
4399 __ evmovdquq(S0, Address(from, 0 * 64), Assembler::AVX_512bit);
4400 __ evmovdquq(S1, Address(from, 1 * 64), Assembler::AVX_512bit);
4401 __ evmovdquq(S2, Address(from, 2 * 64), Assembler::AVX_512bit);
4402 __ evmovdquq(S3, Address(from, 3 * 64), Assembler::AVX_512bit);
4403 __ evmovdquq(S4, Address(from, 4 * 64), Assembler::AVX_512bit);
4404 __ evmovdquq(S5, Address(from, 5 * 64), Assembler::AVX_512bit);
4405 __ evmovdquq(S6, Address(from, 6 * 64), Assembler::AVX_512bit);
4406 __ evmovdquq(S7, Address(from, 7 * 64), Assembler::AVX_512bit);
4407 __ leaq(from, Address(from, 8 * 64));
4408
4409 __ evpxorq(B0, S0, RK1, Assembler::AVX_512bit);
4410 __ evpxorq(B1, S1, RK1, Assembler::AVX_512bit);
4411 __ evpxorq(B2, S2, RK1, Assembler::AVX_512bit);
4412 __ evpxorq(B3, S3, RK1, Assembler::AVX_512bit);
4413 __ evpxorq(B4, S4, RK1, Assembler::AVX_512bit);
4414 __ evpxorq(B5, S5, RK1, Assembler::AVX_512bit);
4415 __ evpxorq(B6, S6, RK1, Assembler::AVX_512bit);
4416 __ evpxorq(B7, S7, RK1, Assembler::AVX_512bit);
4417
4418 __ evalignq(IV, S0, IV, 0x06);
4419 __ evalignq(S0, S1, S0, 0x06);
4420 __ evalignq(S1, S2, S1, 0x06);
4421 __ evalignq(S2, S3, S2, 0x06);
4422 __ evalignq(S3, S4, S3, 0x06);
4423 __ evalignq(S4, S5, S4, 0x06);
4424 __ evalignq(S5, S6, S5, 0x06);
4425 __ evalignq(S6, S7, S6, 0x06);
4426
4427 roundDec(RK2);
4428 roundDec(RK3);
4429 roundDec(RK4);
4430 roundDec(RK5);
4431 roundDec(RK6);
4432 roundDec(RK7);
4433 roundDec(RK8);
4434 roundDec(RK9);
4435 roundDec(RK10);
4436
4437 __ cmpl(rounds, 44);
4438 __ jcc(Assembler::belowEqual, L_128);
4439 roundDec(RK11);
4440 roundDec(RK12);
4441
4442 __ cmpl(rounds, 52);
4443 __ jcc(Assembler::belowEqual, L_192);
4444 roundDec(RK13);
4445 roundDec(RK14);
4446
4447 __ BIND(L_256);
4448 roundDeclast(RK0);
4449 __ jmp(Loop2);
4450
4451 __ BIND(L_128);
4452 roundDeclast(RK0);
4453 __ jmp(Loop2);
4454
4455 __ BIND(L_192);
4456 roundDeclast(RK0);
4457
4458 __ BIND(Loop2);
4459 __ evpxorq(B0, B0, IV, Assembler::AVX_512bit);
4460 __ evpxorq(B1, B1, S0, Assembler::AVX_512bit);
4461 __ evpxorq(B2, B2, S1, Assembler::AVX_512bit);
4462 __ evpxorq(B3, B3, S2, Assembler::AVX_512bit);
4463 __ evpxorq(B4, B4, S3, Assembler::AVX_512bit);
4464 __ evpxorq(B5, B5, S4, Assembler::AVX_512bit);
4465 __ evpxorq(B6, B6, S5, Assembler::AVX_512bit);
4466 __ evpxorq(B7, B7, S6, Assembler::AVX_512bit);
4467 __ evmovdquq(IV, S7, Assembler::AVX_512bit);
4468
4469 __ evmovdquq(Address(to, 0 * 64), B0, Assembler::AVX_512bit);
4470 __ evmovdquq(Address(to, 1 * 64), B1, Assembler::AVX_512bit);
4471 __ evmovdquq(Address(to, 2 * 64), B2, Assembler::AVX_512bit);
4472 __ evmovdquq(Address(to, 3 * 64), B3, Assembler::AVX_512bit);
4473 __ evmovdquq(Address(to, 4 * 64), B4, Assembler::AVX_512bit);
4474 __ evmovdquq(Address(to, 5 * 64), B5, Assembler::AVX_512bit);
4475 __ evmovdquq(Address(to, 6 * 64), B6, Assembler::AVX_512bit);
4476 __ evmovdquq(Address(to, 7 * 64), B7, Assembler::AVX_512bit);
4477 __ leaq(to, Address(to, 8 * 64));
4478 __ jmp(Loop);
4479
4480 __ BIND(Lcbc_dec_rem);
4481 __ evshufi64x2(IV, IV, IV, 0x03, Assembler::AVX_512bit);
4482
4483 __ BIND(Lcbc_dec_rem_loop);
4484 __ subl(len_reg, 16);
4485 __ jcc(Assembler::carrySet, Lcbc_dec_ret);
4486
4487 __ movdqu(S0, Address(from, 0));
4488 __ evpxorq(B0, S0, RK1, Assembler::AVX_512bit);
4489 __ vaesdec(B0, B0, RK2, Assembler::AVX_512bit);
4490 __ vaesdec(B0, B0, RK3, Assembler::AVX_512bit);
4491 __ vaesdec(B0, B0, RK4, Assembler::AVX_512bit);
4492 __ vaesdec(B0, B0, RK5, Assembler::AVX_512bit);
4493 __ vaesdec(B0, B0, RK6, Assembler::AVX_512bit);
4494 __ vaesdec(B0, B0, RK7, Assembler::AVX_512bit);
4495 __ vaesdec(B0, B0, RK8, Assembler::AVX_512bit);
4496 __ vaesdec(B0, B0, RK9, Assembler::AVX_512bit);
4497 __ vaesdec(B0, B0, RK10, Assembler::AVX_512bit);
4498 __ cmpl(rounds, 44);
4499 __ jcc(Assembler::belowEqual, Lcbc_dec_rem_last);
4500
4501 __ vaesdec(B0, B0, RK11, Assembler::AVX_512bit);
4502 __ vaesdec(B0, B0, RK12, Assembler::AVX_512bit);
4503 __ cmpl(rounds, 52);
4504 __ jcc(Assembler::belowEqual, Lcbc_dec_rem_last);
4505
4506 __ vaesdec(B0, B0, RK13, Assembler::AVX_512bit);
4507 __ vaesdec(B0, B0, RK14, Assembler::AVX_512bit);
4508
4509 __ BIND(Lcbc_dec_rem_last);
4510 __ vaesdeclast(B0, B0, RK0, Assembler::AVX_512bit);
4511
4512 __ evpxorq(B0, B0, IV, Assembler::AVX_512bit);
4513 __ evmovdquq(IV, S0, Assembler::AVX_512bit);
4514 __ movdqu(Address(to, 0), B0);
4515 __ leaq(from, Address(from, 16));
4516 __ leaq(to, Address(to, 16));
4517 __ jmp(Lcbc_dec_rem_loop);
4518
4519 __ BIND(Lcbc_dec_ret);
4520 __ movdqu(Address(rvec, 0), IV);
4521
4522 // Zero out the round keys
4523 __ evpxorq(RK0, RK0, RK0, Assembler::AVX_512bit);
4524 __ evpxorq(RK1, RK1, RK1, Assembler::AVX_512bit);
4525 __ evpxorq(RK2, RK2, RK2, Assembler::AVX_512bit);
4526 __ evpxorq(RK3, RK3, RK3, Assembler::AVX_512bit);
4527 __ evpxorq(RK4, RK4, RK4, Assembler::AVX_512bit);
4528 __ evpxorq(RK5, RK5, RK5, Assembler::AVX_512bit);
4529 __ evpxorq(RK6, RK6, RK6, Assembler::AVX_512bit);
4530 __ evpxorq(RK7, RK7, RK7, Assembler::AVX_512bit);
4531 __ evpxorq(RK8, RK8, RK8, Assembler::AVX_512bit);
4532 __ evpxorq(RK9, RK9, RK9, Assembler::AVX_512bit);
4533 __ evpxorq(RK10, RK10, RK10, Assembler::AVX_512bit);
4534 __ cmpl(rounds, 44);
4535 __ jcc(Assembler::belowEqual, Lcbc_exit);
4536 __ evpxorq(RK11, RK11, RK11, Assembler::AVX_512bit);
4537 __ evpxorq(RK12, RK12, RK12, Assembler::AVX_512bit);
4538 __ cmpl(rounds, 52);
4539 __ jcc(Assembler::belowEqual, Lcbc_exit);
4540 __ evpxorq(RK13, RK13, RK13, Assembler::AVX_512bit);
4541 __ evpxorq(RK14, RK14, RK14, Assembler::AVX_512bit);
4542
4543 __ BIND(Lcbc_exit);
4544 __ pop(rbx);
4545 #ifdef _WIN64
4546 __ movl(rax, len_mem);
4547 #else
4548 __ pop(rax); // return length
4549 #endif
4550 __ leave(); // required for proper stackwalking of RuntimeStub frame
4551 __ ret(0);
4552 return start;
4553 }
4554
4555 // Polynomial x^128+x^127+x^126+x^121+1
4556 address ghash_polynomial_addr() {
4557 __ align(CodeEntryAlignment);
4558 StubCodeMark mark(this, "StubRoutines", "_ghash_poly_addr");
4559 address start = __ pc();
4560 __ emit_data64(0x0000000000000001, relocInfo::none);
4561 __ emit_data64(0xc200000000000000, relocInfo::none);
4562 return start;
4563 }
4564
4565 address ghash_shufflemask_addr() {
4566 __ align(CodeEntryAlignment);
4567 StubCodeMark mark(this, "StubRoutines", "_ghash_shuffmask_addr");
4568 address start = __ pc();
4569 __ emit_data64(0x0f0f0f0f0f0f0f0f, relocInfo::none);
4570 __ emit_data64(0x0f0f0f0f0f0f0f0f, relocInfo::none);
4571 return start;
4572 }
4573
4574 // Ghash single and multi block operations using AVX instructions
4575 address generate_avx_ghash_processBlocks() {
4576 __ align(CodeEntryAlignment);
4577
4578 StubCodeMark mark(this, "StubRoutines", "ghash_processBlocks");
4579 address start = __ pc();
4580
4581 // arguments
4582 const Register state = c_rarg0;
4583 const Register htbl = c_rarg1;
4584 const Register data = c_rarg2;
4585 const Register blocks = c_rarg3;
4586 __ enter();
4587 // Save state before entering routine
4588 __ avx_ghash(state, htbl, data, blocks);
4589 __ leave(); // required for proper stackwalking of RuntimeStub frame
4590 __ ret(0);
4591 return start;
4592 }
4593
4594 // byte swap x86 long
4595 address generate_ghash_long_swap_mask() {
4596 __ align(CodeEntryAlignment);
4597 StubCodeMark mark(this, "StubRoutines", "ghash_long_swap_mask");
4598 address start = __ pc();
4599 __ emit_data64(0x0f0e0d0c0b0a0908, relocInfo::none );
4600 __ emit_data64(0x0706050403020100, relocInfo::none );
4601 return start;
4602 }
4603
4604 // byte swap x86 byte array
4605 address generate_ghash_byte_swap_mask() {
4606 __ align(CodeEntryAlignment);
4607 StubCodeMark mark(this, "StubRoutines", "ghash_byte_swap_mask");
4608 address start = __ pc();
4609 __ emit_data64(0x08090a0b0c0d0e0f, relocInfo::none );
4610 __ emit_data64(0x0001020304050607, relocInfo::none );
4611 return start;
4612 }
4613
4614 /* Single and multi-block ghash operations */
4615 address generate_ghash_processBlocks() {
4616 __ align(CodeEntryAlignment);
4617 Label L_ghash_loop, L_exit;
4618 StubCodeMark mark(this, "StubRoutines", "ghash_processBlocks");
4619 address start = __ pc();
4620
4621 const Register state = c_rarg0;
4622 const Register subkeyH = c_rarg1;
4623 const Register data = c_rarg2;
4624 const Register blocks = c_rarg3;
4625
4626 const XMMRegister xmm_temp0 = xmm0;
4627 const XMMRegister xmm_temp1 = xmm1;
4628 const XMMRegister xmm_temp2 = xmm2;
4629 const XMMRegister xmm_temp3 = xmm3;
4630 const XMMRegister xmm_temp4 = xmm4;
4631 const XMMRegister xmm_temp5 = xmm5;
4632 const XMMRegister xmm_temp6 = xmm6;
4633 const XMMRegister xmm_temp7 = xmm7;
4634 const XMMRegister xmm_temp8 = xmm8;
4635 const XMMRegister xmm_temp9 = xmm9;
4636 const XMMRegister xmm_temp10 = xmm10;
4637
4638 __ enter();
4639
4640 __ movdqu(xmm_temp10, ExternalAddress(StubRoutines::x86::ghash_long_swap_mask_addr()));
4641
4642 __ movdqu(xmm_temp0, Address(state, 0));
4643 __ pshufb(xmm_temp0, xmm_temp10);
4644
4645
4646 __ BIND(L_ghash_loop);
4647 __ movdqu(xmm_temp2, Address(data, 0));
4648 __ pshufb(xmm_temp2, ExternalAddress(StubRoutines::x86::ghash_byte_swap_mask_addr()));
4649
4650 __ movdqu(xmm_temp1, Address(subkeyH, 0));
4651 __ pshufb(xmm_temp1, xmm_temp10);
4652
4653 __ pxor(xmm_temp0, xmm_temp2);
4654
4655 //
4656 // Multiply with the hash key
4657 //
4658 __ movdqu(xmm_temp3, xmm_temp0);
4659 __ pclmulqdq(xmm_temp3, xmm_temp1, 0); // xmm3 holds a0*b0
4660 __ movdqu(xmm_temp4, xmm_temp0);
4661 __ pclmulqdq(xmm_temp4, xmm_temp1, 16); // xmm4 holds a0*b1
4662
4663 __ movdqu(xmm_temp5, xmm_temp0);
4664 __ pclmulqdq(xmm_temp5, xmm_temp1, 1); // xmm5 holds a1*b0
4665 __ movdqu(xmm_temp6, xmm_temp0);
4666 __ pclmulqdq(xmm_temp6, xmm_temp1, 17); // xmm6 holds a1*b1
4667
4668 __ pxor(xmm_temp4, xmm_temp5); // xmm4 holds a0*b1 + a1*b0
4669
4670 __ movdqu(xmm_temp5, xmm_temp4); // move the contents of xmm4 to xmm5
4671 __ psrldq(xmm_temp4, 8); // shift by xmm4 64 bits to the right
4672 __ pslldq(xmm_temp5, 8); // shift by xmm5 64 bits to the left
4673 __ pxor(xmm_temp3, xmm_temp5);
4674 __ pxor(xmm_temp6, xmm_temp4); // Register pair <xmm6:xmm3> holds the result
4675 // of the carry-less multiplication of
4676 // xmm0 by xmm1.
4677
4678 // We shift the result of the multiplication by one bit position
4679 // to the left to cope for the fact that the bits are reversed.
4680 __ movdqu(xmm_temp7, xmm_temp3);
4681 __ movdqu(xmm_temp8, xmm_temp6);
4682 __ pslld(xmm_temp3, 1);
4683 __ pslld(xmm_temp6, 1);
4684 __ psrld(xmm_temp7, 31);
4685 __ psrld(xmm_temp8, 31);
4686 __ movdqu(xmm_temp9, xmm_temp7);
4687 __ pslldq(xmm_temp8, 4);
4688 __ pslldq(xmm_temp7, 4);
4689 __ psrldq(xmm_temp9, 12);
4690 __ por(xmm_temp3, xmm_temp7);
4691 __ por(xmm_temp6, xmm_temp8);
4692 __ por(xmm_temp6, xmm_temp9);
4693
4694 //
4695 // First phase of the reduction
4696 //
4697 // Move xmm3 into xmm7, xmm8, xmm9 in order to perform the shifts
4698 // independently.
4699 __ movdqu(xmm_temp7, xmm_temp3);
4700 __ movdqu(xmm_temp8, xmm_temp3);
4701 __ movdqu(xmm_temp9, xmm_temp3);
4702 __ pslld(xmm_temp7, 31); // packed right shift shifting << 31
4703 __ pslld(xmm_temp8, 30); // packed right shift shifting << 30
4704 __ pslld(xmm_temp9, 25); // packed right shift shifting << 25
4705 __ pxor(xmm_temp7, xmm_temp8); // xor the shifted versions
4706 __ pxor(xmm_temp7, xmm_temp9);
4707 __ movdqu(xmm_temp8, xmm_temp7);
4708 __ pslldq(xmm_temp7, 12);
4709 __ psrldq(xmm_temp8, 4);
4710 __ pxor(xmm_temp3, xmm_temp7); // first phase of the reduction complete
4711
4712 //
4713 // Second phase of the reduction
4714 //
4715 // Make 3 copies of xmm3 in xmm2, xmm4, xmm5 for doing these
4716 // shift operations.
4717 __ movdqu(xmm_temp2, xmm_temp3);
4718 __ movdqu(xmm_temp4, xmm_temp3);
4719 __ movdqu(xmm_temp5, xmm_temp3);
4720 __ psrld(xmm_temp2, 1); // packed left shifting >> 1
4721 __ psrld(xmm_temp4, 2); // packed left shifting >> 2
4722 __ psrld(xmm_temp5, 7); // packed left shifting >> 7
4723 __ pxor(xmm_temp2, xmm_temp4); // xor the shifted versions
4724 __ pxor(xmm_temp2, xmm_temp5);
4725 __ pxor(xmm_temp2, xmm_temp8);
4726 __ pxor(xmm_temp3, xmm_temp2);
4727 __ pxor(xmm_temp6, xmm_temp3); // the result is in xmm6
4728
4729 __ decrement(blocks);
4730 __ jcc(Assembler::zero, L_exit);
4731 __ movdqu(xmm_temp0, xmm_temp6);
4732 __ addptr(data, 16);
4733 __ jmp(L_ghash_loop);
4734
4735 __ BIND(L_exit);
4736 __ pshufb(xmm_temp6, xmm_temp10); // Byte swap 16-byte result
4737 __ movdqu(Address(state, 0), xmm_temp6); // store the result
4738 __ leave();
4739 __ ret(0);
4740 return start;
4741 }
4742
4743 //base64 character set
4744 address base64_charset_addr() {
4745 __ align(CodeEntryAlignment);
4746 StubCodeMark mark(this, "StubRoutines", "base64_charset");
4747 address start = __ pc();
4748 __ emit_data64(0x0000004200000041, relocInfo::none);
4749 __ emit_data64(0x0000004400000043, relocInfo::none);
4750 __ emit_data64(0x0000004600000045, relocInfo::none);
4751 __ emit_data64(0x0000004800000047, relocInfo::none);
4752 __ emit_data64(0x0000004a00000049, relocInfo::none);
4753 __ emit_data64(0x0000004c0000004b, relocInfo::none);
4754 __ emit_data64(0x0000004e0000004d, relocInfo::none);
4755 __ emit_data64(0x000000500000004f, relocInfo::none);
4756 __ emit_data64(0x0000005200000051, relocInfo::none);
4757 __ emit_data64(0x0000005400000053, relocInfo::none);
4758 __ emit_data64(0x0000005600000055, relocInfo::none);
4759 __ emit_data64(0x0000005800000057, relocInfo::none);
4760 __ emit_data64(0x0000005a00000059, relocInfo::none);
4761 __ emit_data64(0x0000006200000061, relocInfo::none);
4762 __ emit_data64(0x0000006400000063, relocInfo::none);
4763 __ emit_data64(0x0000006600000065, relocInfo::none);
4764 __ emit_data64(0x0000006800000067, relocInfo::none);
4765 __ emit_data64(0x0000006a00000069, relocInfo::none);
4766 __ emit_data64(0x0000006c0000006b, relocInfo::none);
4767 __ emit_data64(0x0000006e0000006d, relocInfo::none);
4768 __ emit_data64(0x000000700000006f, relocInfo::none);
4769 __ emit_data64(0x0000007200000071, relocInfo::none);
4770 __ emit_data64(0x0000007400000073, relocInfo::none);
4771 __ emit_data64(0x0000007600000075, relocInfo::none);
4772 __ emit_data64(0x0000007800000077, relocInfo::none);
4773 __ emit_data64(0x0000007a00000079, relocInfo::none);
4774 __ emit_data64(0x0000003100000030, relocInfo::none);
4775 __ emit_data64(0x0000003300000032, relocInfo::none);
4776 __ emit_data64(0x0000003500000034, relocInfo::none);
4777 __ emit_data64(0x0000003700000036, relocInfo::none);
4778 __ emit_data64(0x0000003900000038, relocInfo::none);
4779 __ emit_data64(0x0000002f0000002b, relocInfo::none);
4780 return start;
4781 }
4782
4783 //base64 url character set
4784 address base64url_charset_addr() {
4785 __ align(CodeEntryAlignment);
4786 StubCodeMark mark(this, "StubRoutines", "base64url_charset");
4787 address start = __ pc();
4788 __ emit_data64(0x0000004200000041, relocInfo::none);
4789 __ emit_data64(0x0000004400000043, relocInfo::none);
4790 __ emit_data64(0x0000004600000045, relocInfo::none);
4791 __ emit_data64(0x0000004800000047, relocInfo::none);
4792 __ emit_data64(0x0000004a00000049, relocInfo::none);
4793 __ emit_data64(0x0000004c0000004b, relocInfo::none);
4794 __ emit_data64(0x0000004e0000004d, relocInfo::none);
4795 __ emit_data64(0x000000500000004f, relocInfo::none);
4796 __ emit_data64(0x0000005200000051, relocInfo::none);
4797 __ emit_data64(0x0000005400000053, relocInfo::none);
4798 __ emit_data64(0x0000005600000055, relocInfo::none);
4799 __ emit_data64(0x0000005800000057, relocInfo::none);
4800 __ emit_data64(0x0000005a00000059, relocInfo::none);
4801 __ emit_data64(0x0000006200000061, relocInfo::none);
4802 __ emit_data64(0x0000006400000063, relocInfo::none);
4803 __ emit_data64(0x0000006600000065, relocInfo::none);
4804 __ emit_data64(0x0000006800000067, relocInfo::none);
4805 __ emit_data64(0x0000006a00000069, relocInfo::none);
4806 __ emit_data64(0x0000006c0000006b, relocInfo::none);
4807 __ emit_data64(0x0000006e0000006d, relocInfo::none);
4808 __ emit_data64(0x000000700000006f, relocInfo::none);
4809 __ emit_data64(0x0000007200000071, relocInfo::none);
4810 __ emit_data64(0x0000007400000073, relocInfo::none);
4811 __ emit_data64(0x0000007600000075, relocInfo::none);
4812 __ emit_data64(0x0000007800000077, relocInfo::none);
4813 __ emit_data64(0x0000007a00000079, relocInfo::none);
4814 __ emit_data64(0x0000003100000030, relocInfo::none);
4815 __ emit_data64(0x0000003300000032, relocInfo::none);
4816 __ emit_data64(0x0000003500000034, relocInfo::none);
4817 __ emit_data64(0x0000003700000036, relocInfo::none);
4818 __ emit_data64(0x0000003900000038, relocInfo::none);
4819 __ emit_data64(0x0000005f0000002d, relocInfo::none);
4820
4821 return start;
4822 }
4823
4824 address base64_bswap_mask_addr() {
4825 __ align(CodeEntryAlignment);
4826 StubCodeMark mark(this, "StubRoutines", "bswap_mask_base64");
4827 address start = __ pc();
4828 __ emit_data64(0x0504038002010080, relocInfo::none);
4829 __ emit_data64(0x0b0a098008070680, relocInfo::none);
4830 __ emit_data64(0x0908078006050480, relocInfo::none);
4831 __ emit_data64(0x0f0e0d800c0b0a80, relocInfo::none);
4832 __ emit_data64(0x0605048003020180, relocInfo::none);
4833 __ emit_data64(0x0c0b0a8009080780, relocInfo::none);
4834 __ emit_data64(0x0504038002010080, relocInfo::none);
4835 __ emit_data64(0x0b0a098008070680, relocInfo::none);
4836
4837 return start;
4838 }
4839
4840 address base64_right_shift_mask_addr() {
4841 __ align(CodeEntryAlignment);
4842 StubCodeMark mark(this, "StubRoutines", "right_shift_mask");
4843 address start = __ pc();
4844 __ emit_data64(0x0006000400020000, relocInfo::none);
4845 __ emit_data64(0x0006000400020000, relocInfo::none);
4846 __ emit_data64(0x0006000400020000, relocInfo::none);
4847 __ emit_data64(0x0006000400020000, relocInfo::none);
4848 __ emit_data64(0x0006000400020000, relocInfo::none);
4849 __ emit_data64(0x0006000400020000, relocInfo::none);
4850 __ emit_data64(0x0006000400020000, relocInfo::none);
4851 __ emit_data64(0x0006000400020000, relocInfo::none);
4852
4853 return start;
4854 }
4855
4856 address base64_left_shift_mask_addr() {
4857 __ align(CodeEntryAlignment);
4858 StubCodeMark mark(this, "StubRoutines", "left_shift_mask");
4859 address start = __ pc();
4860 __ emit_data64(0x0000000200040000, relocInfo::none);
4861 __ emit_data64(0x0000000200040000, relocInfo::none);
4862 __ emit_data64(0x0000000200040000, relocInfo::none);
4863 __ emit_data64(0x0000000200040000, relocInfo::none);
4864 __ emit_data64(0x0000000200040000, relocInfo::none);
4865 __ emit_data64(0x0000000200040000, relocInfo::none);
4866 __ emit_data64(0x0000000200040000, relocInfo::none);
4867 __ emit_data64(0x0000000200040000, relocInfo::none);
4868
4869 return start;
4870 }
4871
4872 address base64_and_mask_addr() {
4873 __ align(CodeEntryAlignment);
4874 StubCodeMark mark(this, "StubRoutines", "and_mask");
4875 address start = __ pc();
4876 __ emit_data64(0x3f003f003f000000, relocInfo::none);
4877 __ emit_data64(0x3f003f003f000000, relocInfo::none);
4878 __ emit_data64(0x3f003f003f000000, relocInfo::none);
4879 __ emit_data64(0x3f003f003f000000, relocInfo::none);
4880 __ emit_data64(0x3f003f003f000000, relocInfo::none);
4881 __ emit_data64(0x3f003f003f000000, relocInfo::none);
4882 __ emit_data64(0x3f003f003f000000, relocInfo::none);
4883 __ emit_data64(0x3f003f003f000000, relocInfo::none);
4884 return start;
4885 }
4886
4887 address base64_gather_mask_addr() {
4888 __ align(CodeEntryAlignment);
4889 StubCodeMark mark(this, "StubRoutines", "gather_mask");
4890 address start = __ pc();
4891 __ emit_data64(0xffffffffffffffff, relocInfo::none);
4892 return start;
4893 }
4894
4895 // Code for generating Base64 encoding.
4896 // Intrinsic function prototype in Base64.java:
4897 // private void encodeBlock(byte[] src, int sp, int sl, byte[] dst, int dp, boolean isURL) {
4898 address generate_base64_encodeBlock() {
4899 __ align(CodeEntryAlignment);
4900 StubCodeMark mark(this, "StubRoutines", "implEncode");
4901 address start = __ pc();
4902 __ enter();
4903
4904 // Save callee-saved registers before using them
4905 __ push(r12);
4906 __ push(r13);
4907 __ push(r14);
4908 __ push(r15);
4909
4910 // arguments
4911 const Register source = c_rarg0; // Source Array
4912 const Register start_offset = c_rarg1; // start offset
4913 const Register end_offset = c_rarg2; // end offset
4914 const Register dest = c_rarg3; // destination array
4915
4916 #ifndef _WIN64
4917 const Register dp = c_rarg4; // Position for writing to dest array
4918 const Register isURL = c_rarg5;// Base64 or URL character set
4919 #else
4920 const Address dp_mem(rbp, 6 * wordSize); // length is on stack on Win64
4921 const Address isURL_mem(rbp, 7 * wordSize);
4922 const Register isURL = r10; // pick the volatile windows register
4923 const Register dp = r12;
4924 __ movl(dp, dp_mem);
4925 __ movl(isURL, isURL_mem);
4926 #endif
4927
4928 const Register length = r14;
4929 Label L_process80, L_process32, L_process3, L_exit, L_processdata;
4930
4931 // calculate length from offsets
4932 __ movl(length, end_offset);
4933 __ subl(length, start_offset);
4934 __ cmpl(length, 0);
4935 __ jcc(Assembler::lessEqual, L_exit);
4936
4937 __ lea(r11, ExternalAddress(StubRoutines::x86::base64_charset_addr()));
4938 // check if base64 charset(isURL=0) or base64 url charset(isURL=1) needs to be loaded
4939 __ cmpl(isURL, 0);
4940 __ jcc(Assembler::equal, L_processdata);
4941 __ lea(r11, ExternalAddress(StubRoutines::x86::base64url_charset_addr()));
4942
4943 // load masks required for encoding data
4944 __ BIND(L_processdata);
4945 __ movdqu(xmm16, ExternalAddress(StubRoutines::x86::base64_gather_mask_addr()));
4946 // Set 64 bits of K register.
4947 __ evpcmpeqb(k3, xmm16, xmm16, Assembler::AVX_512bit);
4948 __ evmovdquq(xmm12, ExternalAddress(StubRoutines::x86::base64_bswap_mask_addr()), Assembler::AVX_256bit, r13);
4949 __ evmovdquq(xmm13, ExternalAddress(StubRoutines::x86::base64_right_shift_mask_addr()), Assembler::AVX_512bit, r13);
4950 __ evmovdquq(xmm14, ExternalAddress(StubRoutines::x86::base64_left_shift_mask_addr()), Assembler::AVX_512bit, r13);
4951 __ evmovdquq(xmm15, ExternalAddress(StubRoutines::x86::base64_and_mask_addr()), Assembler::AVX_512bit, r13);
4952
4953 // Vector Base64 implementation, producing 96 bytes of encoded data
4954 __ BIND(L_process80);
4955 __ cmpl(length, 80);
4956 __ jcc(Assembler::below, L_process32);
4957 __ evmovdquq(xmm0, Address(source, start_offset, Address::times_1, 0), Assembler::AVX_256bit);
4958 __ evmovdquq(xmm1, Address(source, start_offset, Address::times_1, 24), Assembler::AVX_256bit);
4959 __ evmovdquq(xmm2, Address(source, start_offset, Address::times_1, 48), Assembler::AVX_256bit);
4960
4961 //permute the input data in such a manner that we have continuity of the source
4962 __ vpermq(xmm3, xmm0, 148, Assembler::AVX_256bit);
4963 __ vpermq(xmm4, xmm1, 148, Assembler::AVX_256bit);
4964 __ vpermq(xmm5, xmm2, 148, Assembler::AVX_256bit);
4965
4966 //shuffle input and group 3 bytes of data and to it add 0 as the 4th byte.
4967 //we can deal with 12 bytes at a time in a 128 bit register
4968 __ vpshufb(xmm3, xmm3, xmm12, Assembler::AVX_256bit);
4969 __ vpshufb(xmm4, xmm4, xmm12, Assembler::AVX_256bit);
4970 __ vpshufb(xmm5, xmm5, xmm12, Assembler::AVX_256bit);
4971
4972 //convert byte to word. Each 128 bit register will have 6 bytes for processing
4973 __ vpmovzxbw(xmm3, xmm3, Assembler::AVX_512bit);
4974 __ vpmovzxbw(xmm4, xmm4, Assembler::AVX_512bit);
4975 __ vpmovzxbw(xmm5, xmm5, Assembler::AVX_512bit);
4976
4977 // Extract bits in the following pattern 6, 4+2, 2+4, 6 to convert 3, 8 bit numbers to 4, 6 bit numbers
4978 __ evpsrlvw(xmm0, xmm3, xmm13, Assembler::AVX_512bit);
4979 __ evpsrlvw(xmm1, xmm4, xmm13, Assembler::AVX_512bit);
4980 __ evpsrlvw(xmm2, xmm5, xmm13, Assembler::AVX_512bit);
4981
4982 __ evpsllvw(xmm3, xmm3, xmm14, Assembler::AVX_512bit);
4983 __ evpsllvw(xmm4, xmm4, xmm14, Assembler::AVX_512bit);
4984 __ evpsllvw(xmm5, xmm5, xmm14, Assembler::AVX_512bit);
4985
4986 __ vpsrlq(xmm0, xmm0, 8, Assembler::AVX_512bit);
4987 __ vpsrlq(xmm1, xmm1, 8, Assembler::AVX_512bit);
4988 __ vpsrlq(xmm2, xmm2, 8, Assembler::AVX_512bit);
4989
4990 __ vpsllq(xmm3, xmm3, 8, Assembler::AVX_512bit);
4991 __ vpsllq(xmm4, xmm4, 8, Assembler::AVX_512bit);
4992 __ vpsllq(xmm5, xmm5, 8, Assembler::AVX_512bit);
4993
4994 __ vpandq(xmm3, xmm3, xmm15, Assembler::AVX_512bit);
4995 __ vpandq(xmm4, xmm4, xmm15, Assembler::AVX_512bit);
4996 __ vpandq(xmm5, xmm5, xmm15, Assembler::AVX_512bit);
4997
4998 // Get the final 4*6 bits base64 encoding
4999 __ vporq(xmm3, xmm3, xmm0, Assembler::AVX_512bit);
5000 __ vporq(xmm4, xmm4, xmm1, Assembler::AVX_512bit);
5001 __ vporq(xmm5, xmm5, xmm2, Assembler::AVX_512bit);
5002
5003 // Shift
5004 __ vpsrlq(xmm3, xmm3, 8, Assembler::AVX_512bit);
5005 __ vpsrlq(xmm4, xmm4, 8, Assembler::AVX_512bit);
5006 __ vpsrlq(xmm5, xmm5, 8, Assembler::AVX_512bit);
5007
5008 // look up 6 bits in the base64 character set to fetch the encoding
5009 // we are converting word to dword as gather instructions need dword indices for looking up encoding
5010 __ vextracti64x4(xmm6, xmm3, 0);
5011 __ vpmovzxwd(xmm0, xmm6, Assembler::AVX_512bit);
5012 __ vextracti64x4(xmm6, xmm3, 1);
5013 __ vpmovzxwd(xmm1, xmm6, Assembler::AVX_512bit);
5014
5015 __ vextracti64x4(xmm6, xmm4, 0);
5016 __ vpmovzxwd(xmm2, xmm6, Assembler::AVX_512bit);
5017 __ vextracti64x4(xmm6, xmm4, 1);
5018 __ vpmovzxwd(xmm3, xmm6, Assembler::AVX_512bit);
5019
5020 __ vextracti64x4(xmm4, xmm5, 0);
5021 __ vpmovzxwd(xmm6, xmm4, Assembler::AVX_512bit);
5022
5023 __ vextracti64x4(xmm4, xmm5, 1);
5024 __ vpmovzxwd(xmm7, xmm4, Assembler::AVX_512bit);
5025
5026 __ kmovql(k2, k3);
5027 __ evpgatherdd(xmm4, k2, Address(r11, xmm0, Address::times_4, 0), Assembler::AVX_512bit);
5028 __ kmovql(k2, k3);
5029 __ evpgatherdd(xmm5, k2, Address(r11, xmm1, Address::times_4, 0), Assembler::AVX_512bit);
5030 __ kmovql(k2, k3);
5031 __ evpgatherdd(xmm8, k2, Address(r11, xmm2, Address::times_4, 0), Assembler::AVX_512bit);
5032 __ kmovql(k2, k3);
5033 __ evpgatherdd(xmm9, k2, Address(r11, xmm3, Address::times_4, 0), Assembler::AVX_512bit);
5034 __ kmovql(k2, k3);
5035 __ evpgatherdd(xmm10, k2, Address(r11, xmm6, Address::times_4, 0), Assembler::AVX_512bit);
5036 __ kmovql(k2, k3);
5037 __ evpgatherdd(xmm11, k2, Address(r11, xmm7, Address::times_4, 0), Assembler::AVX_512bit);
5038
5039 //Down convert dword to byte. Final output is 16*6 = 96 bytes long
5040 __ evpmovdb(Address(dest, dp, Address::times_1, 0), xmm4, Assembler::AVX_512bit);
5041 __ evpmovdb(Address(dest, dp, Address::times_1, 16), xmm5, Assembler::AVX_512bit);
5042 __ evpmovdb(Address(dest, dp, Address::times_1, 32), xmm8, Assembler::AVX_512bit);
5043 __ evpmovdb(Address(dest, dp, Address::times_1, 48), xmm9, Assembler::AVX_512bit);
5044 __ evpmovdb(Address(dest, dp, Address::times_1, 64), xmm10, Assembler::AVX_512bit);
5045 __ evpmovdb(Address(dest, dp, Address::times_1, 80), xmm11, Assembler::AVX_512bit);
5046
5047 __ addq(dest, 96);
5048 __ addq(source, 72);
5049 __ subq(length, 72);
5050 __ jmp(L_process80);
5051
5052 // Vector Base64 implementation generating 32 bytes of encoded data
5053 __ BIND(L_process32);
5054 __ cmpl(length, 32);
5055 __ jcc(Assembler::below, L_process3);
5056 __ evmovdquq(xmm0, Address(source, start_offset), Assembler::AVX_256bit);
5057 __ vpermq(xmm0, xmm0, 148, Assembler::AVX_256bit);
5058 __ vpshufb(xmm6, xmm0, xmm12, Assembler::AVX_256bit);
5059 __ vpmovzxbw(xmm6, xmm6, Assembler::AVX_512bit);
5060 __ evpsrlvw(xmm2, xmm6, xmm13, Assembler::AVX_512bit);
5061 __ evpsllvw(xmm3, xmm6, xmm14, Assembler::AVX_512bit);
5062
5063 __ vpsrlq(xmm2, xmm2, 8, Assembler::AVX_512bit);
5064 __ vpsllq(xmm3, xmm3, 8, Assembler::AVX_512bit);
5065 __ vpandq(xmm3, xmm3, xmm15, Assembler::AVX_512bit);
5066 __ vporq(xmm1, xmm2, xmm3, Assembler::AVX_512bit);
5067 __ vpsrlq(xmm1, xmm1, 8, Assembler::AVX_512bit);
5068 __ vextracti64x4(xmm9, xmm1, 0);
5069 __ vpmovzxwd(xmm6, xmm9, Assembler::AVX_512bit);
5070 __ vextracti64x4(xmm9, xmm1, 1);
5071 __ vpmovzxwd(xmm5, xmm9, Assembler::AVX_512bit);
5072 __ kmovql(k2, k3);
5073 __ evpgatherdd(xmm8, k2, Address(r11, xmm6, Address::times_4, 0), Assembler::AVX_512bit);
5074 __ kmovql(k2, k3);
5075 __ evpgatherdd(xmm10, k2, Address(r11, xmm5, Address::times_4, 0), Assembler::AVX_512bit);
5076 __ evpmovdb(Address(dest, dp, Address::times_1, 0), xmm8, Assembler::AVX_512bit);
5077 __ evpmovdb(Address(dest, dp, Address::times_1, 16), xmm10, Assembler::AVX_512bit);
5078 __ subq(length, 24);
5079 __ addq(dest, 32);
5080 __ addq(source, 24);
5081 __ jmp(L_process32);
5082
5083 // Scalar data processing takes 3 bytes at a time and produces 4 bytes of encoded data
5084 /* This code corresponds to the scalar version of the following snippet in Base64.java
5085 ** int bits = (src[sp0++] & 0xff) << 16 |(src[sp0++] & 0xff) << 8 |(src[sp0++] & 0xff);
5086 ** dst[dp0++] = (byte)base64[(bits >> > 18) & 0x3f];
5087 ** dst[dp0++] = (byte)base64[(bits >> > 12) & 0x3f];
5088 ** dst[dp0++] = (byte)base64[(bits >> > 6) & 0x3f];
5089 ** dst[dp0++] = (byte)base64[bits & 0x3f];*/
5090 __ BIND(L_process3);
5091 __ cmpl(length, 3);
5092 __ jcc(Assembler::below, L_exit);
5093 // Read 1 byte at a time
5094 __ movzbl(rax, Address(source, start_offset));
5095 __ shll(rax, 0x10);
5096 __ movl(r15, rax);
5097 __ movzbl(rax, Address(source, start_offset, Address::times_1, 1));
5098 __ shll(rax, 0x8);
5099 __ movzwl(rax, rax);
5100 __ orl(r15, rax);
5101 __ movzbl(rax, Address(source, start_offset, Address::times_1, 2));
5102 __ orl(rax, r15);
5103 // Save 3 bytes read in r15
5104 __ movl(r15, rax);
5105 __ shrl(rax, 0x12);
5106 __ andl(rax, 0x3f);
5107 // rax contains the index, r11 contains base64 lookup table
5108 __ movb(rax, Address(r11, rax, Address::times_4));
5109 // Write the encoded byte to destination
5110 __ movb(Address(dest, dp, Address::times_1, 0), rax);
5111 __ movl(rax, r15);
5112 __ shrl(rax, 0xc);
5113 __ andl(rax, 0x3f);
5114 __ movb(rax, Address(r11, rax, Address::times_4));
5115 __ movb(Address(dest, dp, Address::times_1, 1), rax);
5116 __ movl(rax, r15);
5117 __ shrl(rax, 0x6);
5118 __ andl(rax, 0x3f);
5119 __ movb(rax, Address(r11, rax, Address::times_4));
5120 __ movb(Address(dest, dp, Address::times_1, 2), rax);
5121 __ movl(rax, r15);
5122 __ andl(rax, 0x3f);
5123 __ movb(rax, Address(r11, rax, Address::times_4));
5124 __ movb(Address(dest, dp, Address::times_1, 3), rax);
5125 __ subl(length, 3);
5126 __ addq(dest, 4);
5127 __ addq(source, 3);
5128 __ jmp(L_process3);
5129 __ BIND(L_exit);
5130 __ pop(r15);
5131 __ pop(r14);
5132 __ pop(r13);
5133 __ pop(r12);
5134 __ leave();
5135 __ ret(0);
5136 return start;
5137 }
5138
5139 /**
5140 * Arguments:
5141 *
5142 * Inputs:
5143 * c_rarg0 - int crc
5144 * c_rarg1 - byte* buf
5145 * c_rarg2 - int length
5146 *
5147 * Ouput:
5148 * rax - int crc result
5149 */
5150 address generate_updateBytesCRC32() {
5151 assert(UseCRC32Intrinsics, "need AVX and CLMUL instructions");
5152
5153 __ align(CodeEntryAlignment);
5154 StubCodeMark mark(this, "StubRoutines", "updateBytesCRC32");
5155
5156 address start = __ pc();
5157 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
5158 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
5159 // rscratch1: r10
5160 const Register crc = c_rarg0; // crc
5161 const Register buf = c_rarg1; // source java byte array address
5162 const Register len = c_rarg2; // length
5163 const Register table = c_rarg3; // crc_table address (reuse register)
5164 const Register tmp1 = r11;
5165 const Register tmp2 = r10;
5166 assert_different_registers(crc, buf, len, table, tmp1, tmp2, rax);
5167
5168 BLOCK_COMMENT("Entry:");
5169 __ enter(); // required for proper stackwalking of RuntimeStub frame
5170
5171 if (VM_Version::supports_sse4_1() && VM_Version::supports_avx512_vpclmulqdq() &&
5172 VM_Version::supports_avx512bw() &&
5173 VM_Version::supports_avx512vl()) {
5174 __ kernel_crc32_avx512(crc, buf, len, table, tmp1, tmp2);
5175 } else {
5176 __ kernel_crc32(crc, buf, len, table, tmp1);
5177 }
5178
5179 __ movl(rax, crc);
5180 __ vzeroupper();
5181 __ leave(); // required for proper stackwalking of RuntimeStub frame
5182 __ ret(0);
5183
5184 return start;
5185 }
5186
5187 /**
5188 * Arguments:
5189 *
5190 * Inputs:
5191 * c_rarg0 - int crc
5192 * c_rarg1 - byte* buf
5193 * c_rarg2 - long length
5194 * c_rarg3 - table_start - optional (present only when doing a library_call,
5195 * not used by x86 algorithm)
5196 *
5197 * Ouput:
5198 * rax - int crc result
5199 */
5200 address generate_updateBytesCRC32C(bool is_pclmulqdq_supported) {
5201 assert(UseCRC32CIntrinsics, "need SSE4_2");
5202 __ align(CodeEntryAlignment);
5203 StubCodeMark mark(this, "StubRoutines", "updateBytesCRC32C");
5204 address start = __ pc();
5205 //reg.arg int#0 int#1 int#2 int#3 int#4 int#5 float regs
5206 //Windows RCX RDX R8 R9 none none XMM0..XMM3
5207 //Lin / Sol RDI RSI RDX RCX R8 R9 XMM0..XMM7
5208 const Register crc = c_rarg0; // crc
5209 const Register buf = c_rarg1; // source java byte array address
5210 const Register len = c_rarg2; // length
5211 const Register a = rax;
5212 const Register j = r9;
5213 const Register k = r10;
5214 const Register l = r11;
5215 #ifdef _WIN64
5216 const Register y = rdi;
5217 const Register z = rsi;
5218 #else
5219 const Register y = rcx;
5220 const Register z = r8;
5221 #endif
5222 assert_different_registers(crc, buf, len, a, j, k, l, y, z);
5223
5224 BLOCK_COMMENT("Entry:");
5225 __ enter(); // required for proper stackwalking of RuntimeStub frame
5226 #ifdef _WIN64
5227 __ push(y);
5228 __ push(z);
5229 #endif
5230 __ crc32c_ipl_alg2_alt2(crc, buf, len,
5231 a, j, k,
5232 l, y, z,
5233 c_farg0, c_farg1, c_farg2,
5234 is_pclmulqdq_supported);
5235 __ movl(rax, crc);
5236 #ifdef _WIN64
5237 __ pop(z);
5238 __ pop(y);
5239 #endif
5240 __ vzeroupper();
5241 __ leave(); // required for proper stackwalking of RuntimeStub frame
5242 __ ret(0);
5243
5244 return start;
5245 }
5246
5247 /**
5248 * Arguments:
5249 *
5250 * Input:
5251 * c_rarg0 - x address
5252 * c_rarg1 - x length
5253 * c_rarg2 - y address
5254 * c_rarg3 - y length
5255 * not Win64
5256 * c_rarg4 - z address
5257 * c_rarg5 - z length
5258 * Win64
5259 * rsp+40 - z address
5260 * rsp+48 - z length
5261 */
5262 address generate_multiplyToLen() {
5263 __ align(CodeEntryAlignment);
5264 StubCodeMark mark(this, "StubRoutines", "multiplyToLen");
5265
5266 address start = __ pc();
5267 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
5268 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
5269 const Register x = rdi;
5270 const Register xlen = rax;
5271 const Register y = rsi;
5272 const Register ylen = rcx;
5273 const Register z = r8;
5274 const Register zlen = r11;
5275
5276 // Next registers will be saved on stack in multiply_to_len().
5277 const Register tmp1 = r12;
5278 const Register tmp2 = r13;
5279 const Register tmp3 = r14;
5280 const Register tmp4 = r15;
5281 const Register tmp5 = rbx;
5282
5283 BLOCK_COMMENT("Entry:");
5284 __ enter(); // required for proper stackwalking of RuntimeStub frame
5285
5286 #ifndef _WIN64
5287 __ movptr(zlen, r9); // Save r9 in r11 - zlen
5288 #endif
5289 setup_arg_regs(4); // x => rdi, xlen => rsi, y => rdx
5290 // ylen => rcx, z => r8, zlen => r11
5291 // r9 and r10 may be used to save non-volatile registers
5292 #ifdef _WIN64
5293 // last 2 arguments (#4, #5) are on stack on Win64
5294 __ movptr(z, Address(rsp, 6 * wordSize));
5295 __ movptr(zlen, Address(rsp, 7 * wordSize));
5296 #endif
5297
5298 __ movptr(xlen, rsi);
5299 __ movptr(y, rdx);
5300 __ multiply_to_len(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5);
5301
5302 restore_arg_regs();
5303
5304 __ leave(); // required for proper stackwalking of RuntimeStub frame
5305 __ ret(0);
5306
5307 return start;
5308 }
5309
5310 /**
5311 * Arguments:
5312 *
5313 * Input:
5314 * c_rarg0 - obja address
5315 * c_rarg1 - objb address
5316 * c_rarg3 - length length
5317 * c_rarg4 - scale log2_array_indxscale
5318 *
5319 * Output:
5320 * rax - int >= mismatched index, < 0 bitwise complement of tail
5321 */
5322 address generate_vectorizedMismatch() {
5323 __ align(CodeEntryAlignment);
5324 StubCodeMark mark(this, "StubRoutines", "vectorizedMismatch");
5325 address start = __ pc();
5326
5327 BLOCK_COMMENT("Entry:");
5328 __ enter();
5329
5330 #ifdef _WIN64 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
5331 const Register scale = c_rarg0; //rcx, will exchange with r9
5332 const Register objb = c_rarg1; //rdx
5333 const Register length = c_rarg2; //r8
5334 const Register obja = c_rarg3; //r9
5335 __ xchgq(obja, scale); //now obja and scale contains the correct contents
5336
5337 const Register tmp1 = r10;
5338 const Register tmp2 = r11;
5339 #endif
5340 #ifndef _WIN64 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
5341 const Register obja = c_rarg0; //U:rdi
5342 const Register objb = c_rarg1; //U:rsi
5343 const Register length = c_rarg2; //U:rdx
5344 const Register scale = c_rarg3; //U:rcx
5345 const Register tmp1 = r8;
5346 const Register tmp2 = r9;
5347 #endif
5348 const Register result = rax; //return value
5349 const XMMRegister vec0 = xmm0;
5350 const XMMRegister vec1 = xmm1;
5351 const XMMRegister vec2 = xmm2;
5352
5353 __ vectorized_mismatch(obja, objb, length, scale, result, tmp1, tmp2, vec0, vec1, vec2);
5354
5355 __ vzeroupper();
5356 __ leave();
5357 __ ret(0);
5358
5359 return start;
5360 }
5361
5362 /**
5363 * Arguments:
5364 *
5365 // Input:
5366 // c_rarg0 - x address
5367 // c_rarg1 - x length
5368 // c_rarg2 - z address
5369 // c_rarg3 - z lenth
5370 *
5371 */
5372 address generate_squareToLen() {
5373
5374 __ align(CodeEntryAlignment);
5375 StubCodeMark mark(this, "StubRoutines", "squareToLen");
5376
5377 address start = __ pc();
5378 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
5379 // Unix: rdi, rsi, rdx, rcx (c_rarg0, c_rarg1, ...)
5380 const Register x = rdi;
5381 const Register len = rsi;
5382 const Register z = r8;
5383 const Register zlen = rcx;
5384
5385 const Register tmp1 = r12;
5386 const Register tmp2 = r13;
5387 const Register tmp3 = r14;
5388 const Register tmp4 = r15;
5389 const Register tmp5 = rbx;
5390
5391 BLOCK_COMMENT("Entry:");
5392 __ enter(); // required for proper stackwalking of RuntimeStub frame
5393
5394 setup_arg_regs(4); // x => rdi, len => rsi, z => rdx
5395 // zlen => rcx
5396 // r9 and r10 may be used to save non-volatile registers
5397 __ movptr(r8, rdx);
5398 __ square_to_len(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx, rax);
5399
5400 restore_arg_regs();
5401
5402 __ leave(); // required for proper stackwalking of RuntimeStub frame
5403 __ ret(0);
5404
5405 return start;
5406 }
5407
5408 address generate_method_entry_barrier() {
5409 __ align(CodeEntryAlignment);
5410 StubCodeMark mark(this, "StubRoutines", "nmethod_entry_barrier");
5411
5412 Label deoptimize_label;
5413
5414 address start = __ pc();
5415
5416 __ push(-1); // cookie, this is used for writing the new rsp when deoptimizing
5417
5418 BLOCK_COMMENT("Entry:");
5419 __ enter(); // save rbp
5420
5421 // save c_rarg0, because we want to use that value.
5422 // We could do without it but then we depend on the number of slots used by pusha
5423 __ push(c_rarg0);
5424
5425 __ lea(c_rarg0, Address(rsp, wordSize * 3)); // 1 for cookie, 1 for rbp, 1 for c_rarg0 - this should be the return address
5426
5427 __ pusha();
5428
5429 // The method may have floats as arguments, and we must spill them before calling
5430 // the VM runtime.
5431 assert(Argument::n_float_register_parameters_j == 8, "Assumption");
5432 const int xmm_size = wordSize * 2;
5433 const int xmm_spill_size = xmm_size * Argument::n_float_register_parameters_j;
5434 __ subptr(rsp, xmm_spill_size);
5435 __ movdqu(Address(rsp, xmm_size * 7), xmm7);
5436 __ movdqu(Address(rsp, xmm_size * 6), xmm6);
5437 __ movdqu(Address(rsp, xmm_size * 5), xmm5);
5438 __ movdqu(Address(rsp, xmm_size * 4), xmm4);
5439 __ movdqu(Address(rsp, xmm_size * 3), xmm3);
5440 __ movdqu(Address(rsp, xmm_size * 2), xmm2);
5441 __ movdqu(Address(rsp, xmm_size * 1), xmm1);
5442 __ movdqu(Address(rsp, xmm_size * 0), xmm0);
5443
5444 __ call_VM_leaf(CAST_FROM_FN_PTR(address, static_cast<int (*)(address*)>(BarrierSetNMethod::nmethod_stub_entry_barrier)), 1);
5445
5446 __ movdqu(xmm0, Address(rsp, xmm_size * 0));
5447 __ movdqu(xmm1, Address(rsp, xmm_size * 1));
5448 __ movdqu(xmm2, Address(rsp, xmm_size * 2));
5449 __ movdqu(xmm3, Address(rsp, xmm_size * 3));
5450 __ movdqu(xmm4, Address(rsp, xmm_size * 4));
5451 __ movdqu(xmm5, Address(rsp, xmm_size * 5));
5452 __ movdqu(xmm6, Address(rsp, xmm_size * 6));
5453 __ movdqu(xmm7, Address(rsp, xmm_size * 7));
5454 __ addptr(rsp, xmm_spill_size);
5455
5456 __ cmpl(rax, 1); // 1 means deoptimize
5457 __ jcc(Assembler::equal, deoptimize_label);
5458
5459 __ popa();
5460 __ pop(c_rarg0);
5461
5462 __ leave();
5463
5464 __ addptr(rsp, 1 * wordSize); // cookie
5465 __ ret(0);
5466
5467
5468 __ BIND(deoptimize_label);
5469
5470 __ popa();
5471 __ pop(c_rarg0);
5472
5473 __ leave();
5474
5475 // this can be taken out, but is good for verification purposes. getting a SIGSEGV
5476 // here while still having a correct stack is valuable
5477 __ testptr(rsp, Address(rsp, 0));
5478
5479 __ movptr(rsp, Address(rsp, 0)); // new rsp was written in the barrier
5480 __ jmp(Address(rsp, -1 * wordSize)); // jmp target should be callers verified_entry_point
5481
5482 return start;
5483 }
5484
5485 /**
5486 * Arguments:
5487 *
5488 * Input:
5489 * c_rarg0 - out address
5490 * c_rarg1 - in address
5491 * c_rarg2 - offset
5492 * c_rarg3 - len
5493 * not Win64
5494 * c_rarg4 - k
5495 * Win64
5496 * rsp+40 - k
5497 */
5498 address generate_mulAdd() {
5499 __ align(CodeEntryAlignment);
5500 StubCodeMark mark(this, "StubRoutines", "mulAdd");
5501
5502 address start = __ pc();
5503 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
5504 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
5505 const Register out = rdi;
5506 const Register in = rsi;
5507 const Register offset = r11;
5508 const Register len = rcx;
5509 const Register k = r8;
5510
5511 // Next registers will be saved on stack in mul_add().
5512 const Register tmp1 = r12;
5513 const Register tmp2 = r13;
5514 const Register tmp3 = r14;
5515 const Register tmp4 = r15;
5516 const Register tmp5 = rbx;
5517
5518 BLOCK_COMMENT("Entry:");
5519 __ enter(); // required for proper stackwalking of RuntimeStub frame
5520
5521 setup_arg_regs(4); // out => rdi, in => rsi, offset => rdx
5522 // len => rcx, k => r8
5523 // r9 and r10 may be used to save non-volatile registers
5524 #ifdef _WIN64
5525 // last argument is on stack on Win64
5526 __ movl(k, Address(rsp, 6 * wordSize));
5527 #endif
5528 __ movptr(r11, rdx); // move offset in rdx to offset(r11)
5529 __ mul_add(out, in, offset, len, k, tmp1, tmp2, tmp3, tmp4, tmp5, rdx, rax);
5530
5531 restore_arg_regs();
5532
5533 __ leave(); // required for proper stackwalking of RuntimeStub frame
5534 __ ret(0);
5535
5536 return start;
5537 }
5538
5539 address generate_bigIntegerRightShift() {
5540 __ align(CodeEntryAlignment);
5541 StubCodeMark mark(this, "StubRoutines", "bigIntegerRightShiftWorker");
5542
5543 address start = __ pc();
5544 Label Shift512Loop, ShiftTwo, ShiftTwoLoop, ShiftOne, Exit;
5545 // For Unix, the arguments are as follows: rdi, rsi, rdx, rcx, r8.
5546 const Register newArr = rdi;
5547 const Register oldArr = rsi;
5548 const Register newIdx = rdx;
5549 const Register shiftCount = rcx; // It was intentional to have shiftCount in rcx since it is used implicitly for shift.
5550 const Register totalNumIter = r8;
5551
5552 // For windows, we use r9 and r10 as temps to save rdi and rsi. Thus we cannot allocate them for our temps.
5553 // For everything else, we prefer using r9 and r10 since we do not have to save them before use.
5554 const Register tmp1 = r11; // Caller save.
5555 const Register tmp2 = rax; // Caller save.
5556 const Register tmp3 = WINDOWS_ONLY(r12) NOT_WINDOWS(r9); // Windows: Callee save. Linux: Caller save.
5557 const Register tmp4 = WINDOWS_ONLY(r13) NOT_WINDOWS(r10); // Windows: Callee save. Linux: Caller save.
5558 const Register tmp5 = r14; // Callee save.
5559 const Register tmp6 = r15;
5560
5561 const XMMRegister x0 = xmm0;
5562 const XMMRegister x1 = xmm1;
5563 const XMMRegister x2 = xmm2;
5564
5565 BLOCK_COMMENT("Entry:");
5566 __ enter(); // required for proper stackwalking of RuntimeStub frame
5567
5568 #ifdef _WINDOWS
5569 setup_arg_regs(4);
5570 // For windows, since last argument is on stack, we need to move it to the appropriate register.
5571 __ movl(totalNumIter, Address(rsp, 6 * wordSize));
5572 // Save callee save registers.
5573 __ push(tmp3);
5574 __ push(tmp4);
5575 #endif
5576 __ push(tmp5);
5577
5578 // Rename temps used throughout the code.
5579 const Register idx = tmp1;
5580 const Register nIdx = tmp2;
5581
5582 __ xorl(idx, idx);
5583
5584 // Start right shift from end of the array.
5585 // For example, if #iteration = 4 and newIdx = 1
5586 // then dest[4] = src[4] >> shiftCount | src[3] <<< (shiftCount - 32)
5587 // if #iteration = 4 and newIdx = 0
5588 // then dest[3] = src[4] >> shiftCount | src[3] <<< (shiftCount - 32)
5589 __ movl(idx, totalNumIter);
5590 __ movl(nIdx, idx);
5591 __ addl(nIdx, newIdx);
5592
5593 // If vectorization is enabled, check if the number of iterations is at least 64
5594 // If not, then go to ShifTwo processing 2 iterations
5595 if (VM_Version::supports_avx512_vbmi2()) {
5596 __ cmpptr(totalNumIter, (AVX3Threshold/64));
5597 __ jcc(Assembler::less, ShiftTwo);
5598
5599 if (AVX3Threshold < 16 * 64) {
5600 __ cmpl(totalNumIter, 16);
5601 __ jcc(Assembler::less, ShiftTwo);
5602 }
5603 __ evpbroadcastd(x0, shiftCount, Assembler::AVX_512bit);
5604 __ subl(idx, 16);
5605 __ subl(nIdx, 16);
5606 __ BIND(Shift512Loop);
5607 __ evmovdqul(x2, Address(oldArr, idx, Address::times_4, 4), Assembler::AVX_512bit);
5608 __ evmovdqul(x1, Address(oldArr, idx, Address::times_4), Assembler::AVX_512bit);
5609 __ vpshrdvd(x2, x1, x0, Assembler::AVX_512bit);
5610 __ evmovdqul(Address(newArr, nIdx, Address::times_4), x2, Assembler::AVX_512bit);
5611 __ subl(nIdx, 16);
5612 __ subl(idx, 16);
5613 __ jcc(Assembler::greaterEqual, Shift512Loop);
5614 __ addl(idx, 16);
5615 __ addl(nIdx, 16);
5616 }
5617 __ BIND(ShiftTwo);
5618 __ cmpl(idx, 2);
5619 __ jcc(Assembler::less, ShiftOne);
5620 __ subl(idx, 2);
5621 __ subl(nIdx, 2);
5622 __ BIND(ShiftTwoLoop);
5623 __ movl(tmp5, Address(oldArr, idx, Address::times_4, 8));
5624 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 4));
5625 __ movl(tmp3, Address(oldArr, idx, Address::times_4));
5626 __ shrdl(tmp5, tmp4);
5627 __ shrdl(tmp4, tmp3);
5628 __ movl(Address(newArr, nIdx, Address::times_4, 4), tmp5);
5629 __ movl(Address(newArr, nIdx, Address::times_4), tmp4);
5630 __ subl(nIdx, 2);
5631 __ subl(idx, 2);
5632 __ jcc(Assembler::greaterEqual, ShiftTwoLoop);
5633 __ addl(idx, 2);
5634 __ addl(nIdx, 2);
5635
5636 // Do the last iteration
5637 __ BIND(ShiftOne);
5638 __ cmpl(idx, 1);
5639 __ jcc(Assembler::less, Exit);
5640 __ subl(idx, 1);
5641 __ subl(nIdx, 1);
5642 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 4));
5643 __ movl(tmp3, Address(oldArr, idx, Address::times_4));
5644 __ shrdl(tmp4, tmp3);
5645 __ movl(Address(newArr, nIdx, Address::times_4), tmp4);
5646 __ BIND(Exit);
5647 // Restore callee save registers.
5648 __ pop(tmp5);
5649 #ifdef _WINDOWS
5650 __ pop(tmp4);
5651 __ pop(tmp3);
5652 restore_arg_regs();
5653 #endif
5654 __ leave(); // required for proper stackwalking of RuntimeStub frame
5655 __ ret(0);
5656 return start;
5657 }
5658
5659 /**
5660 * Arguments:
5661 *
5662 * Input:
5663 * c_rarg0 - newArr address
5664 * c_rarg1 - oldArr address
5665 * c_rarg2 - newIdx
5666 * c_rarg3 - shiftCount
5667 * not Win64
5668 * c_rarg4 - numIter
5669 * Win64
5670 * rsp40 - numIter
5671 */
5672 address generate_bigIntegerLeftShift() {
5673 __ align(CodeEntryAlignment);
5674 StubCodeMark mark(this, "StubRoutines", "bigIntegerLeftShiftWorker");
5675 address start = __ pc();
5676 Label Shift512Loop, ShiftTwo, ShiftTwoLoop, ShiftOne, Exit;
5677 // For Unix, the arguments are as follows: rdi, rsi, rdx, rcx, r8.
5678 const Register newArr = rdi;
5679 const Register oldArr = rsi;
5680 const Register newIdx = rdx;
5681 const Register shiftCount = rcx; // It was intentional to have shiftCount in rcx since it is used implicitly for shift.
5682 const Register totalNumIter = r8;
5683 // For windows, we use r9 and r10 as temps to save rdi and rsi. Thus we cannot allocate them for our temps.
5684 // For everything else, we prefer using r9 and r10 since we do not have to save them before use.
5685 const Register tmp1 = r11; // Caller save.
5686 const Register tmp2 = rax; // Caller save.
5687 const Register tmp3 = WINDOWS_ONLY(r12) NOT_WINDOWS(r9); // Windows: Callee save. Linux: Caller save.
5688 const Register tmp4 = WINDOWS_ONLY(r13) NOT_WINDOWS(r10); // Windows: Callee save. Linux: Caller save.
5689 const Register tmp5 = r14; // Callee save.
5690
5691 const XMMRegister x0 = xmm0;
5692 const XMMRegister x1 = xmm1;
5693 const XMMRegister x2 = xmm2;
5694 BLOCK_COMMENT("Entry:");
5695 __ enter(); // required for proper stackwalking of RuntimeStub frame
5696
5697 #ifdef _WINDOWS
5698 setup_arg_regs(4);
5699 // For windows, since last argument is on stack, we need to move it to the appropriate register.
5700 __ movl(totalNumIter, Address(rsp, 6 * wordSize));
5701 // Save callee save registers.
5702 __ push(tmp3);
5703 __ push(tmp4);
5704 #endif
5705 __ push(tmp5);
5706
5707 // Rename temps used throughout the code
5708 const Register idx = tmp1;
5709 const Register numIterTmp = tmp2;
5710
5711 // Start idx from zero.
5712 __ xorl(idx, idx);
5713 // Compute interior pointer for new array. We do this so that we can use same index for both old and new arrays.
5714 __ lea(newArr, Address(newArr, newIdx, Address::times_4));
5715 __ movl(numIterTmp, totalNumIter);
5716
5717 // If vectorization is enabled, check if the number of iterations is at least 64
5718 // If not, then go to ShiftTwo shifting two numbers at a time
5719 if (VM_Version::supports_avx512_vbmi2()) {
5720 __ cmpl(totalNumIter, (AVX3Threshold/64));
5721 __ jcc(Assembler::less, ShiftTwo);
5722
5723 if (AVX3Threshold < 16 * 64) {
5724 __ cmpl(totalNumIter, 16);
5725 __ jcc(Assembler::less, ShiftTwo);
5726 }
5727 __ evpbroadcastd(x0, shiftCount, Assembler::AVX_512bit);
5728 __ subl(numIterTmp, 16);
5729 __ BIND(Shift512Loop);
5730 __ evmovdqul(x1, Address(oldArr, idx, Address::times_4), Assembler::AVX_512bit);
5731 __ evmovdqul(x2, Address(oldArr, idx, Address::times_4, 0x4), Assembler::AVX_512bit);
5732 __ vpshldvd(x1, x2, x0, Assembler::AVX_512bit);
5733 __ evmovdqul(Address(newArr, idx, Address::times_4), x1, Assembler::AVX_512bit);
5734 __ addl(idx, 16);
5735 __ subl(numIterTmp, 16);
5736 __ jcc(Assembler::greaterEqual, Shift512Loop);
5737 __ addl(numIterTmp, 16);
5738 }
5739 __ BIND(ShiftTwo);
5740 __ cmpl(totalNumIter, 1);
5741 __ jcc(Assembler::less, Exit);
5742 __ movl(tmp3, Address(oldArr, idx, Address::times_4));
5743 __ subl(numIterTmp, 2);
5744 __ jcc(Assembler::less, ShiftOne);
5745
5746 __ BIND(ShiftTwoLoop);
5747 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 0x4));
5748 __ movl(tmp5, Address(oldArr, idx, Address::times_4, 0x8));
5749 __ shldl(tmp3, tmp4);
5750 __ shldl(tmp4, tmp5);
5751 __ movl(Address(newArr, idx, Address::times_4), tmp3);
5752 __ movl(Address(newArr, idx, Address::times_4, 0x4), tmp4);
5753 __ movl(tmp3, tmp5);
5754 __ addl(idx, 2);
5755 __ subl(numIterTmp, 2);
5756 __ jcc(Assembler::greaterEqual, ShiftTwoLoop);
5757
5758 // Do the last iteration
5759 __ BIND(ShiftOne);
5760 __ addl(numIterTmp, 2);
5761 __ cmpl(numIterTmp, 1);
5762 __ jcc(Assembler::less, Exit);
5763 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 0x4));
5764 __ shldl(tmp3, tmp4);
5765 __ movl(Address(newArr, idx, Address::times_4), tmp3);
5766
5767 __ BIND(Exit);
5768 // Restore callee save registers.
5769 __ pop(tmp5);
5770 #ifdef _WINDOWS
5771 __ pop(tmp4);
5772 __ pop(tmp3);
5773 restore_arg_regs();
5774 #endif
5775 __ leave(); // required for proper stackwalking of RuntimeStub frame
5776 __ ret(0);
5777 return start;
5778 }
5779
5780 address generate_libmExp() {
5781 StubCodeMark mark(this, "StubRoutines", "libmExp");
5782
5783 address start = __ pc();
5784
5785 const XMMRegister x0 = xmm0;
5786 const XMMRegister x1 = xmm1;
5787 const XMMRegister x2 = xmm2;
5788 const XMMRegister x3 = xmm3;
5789
5790 const XMMRegister x4 = xmm4;
5791 const XMMRegister x5 = xmm5;
5792 const XMMRegister x6 = xmm6;
5793 const XMMRegister x7 = xmm7;
5794
5795 const Register tmp = r11;
5796
5797 BLOCK_COMMENT("Entry:");
5798 __ enter(); // required for proper stackwalking of RuntimeStub frame
5799
5800 __ fast_exp(x0, x1, x2, x3, x4, x5, x6, x7, rax, rcx, rdx, tmp);
5801
5802 __ leave(); // required for proper stackwalking of RuntimeStub frame
5803 __ ret(0);
5804
5805 return start;
5806
5807 }
5808
5809 address generate_libmLog() {
5810 StubCodeMark mark(this, "StubRoutines", "libmLog");
5811
5812 address start = __ pc();
5813
5814 const XMMRegister x0 = xmm0;
5815 const XMMRegister x1 = xmm1;
5816 const XMMRegister x2 = xmm2;
5817 const XMMRegister x3 = xmm3;
5818
5819 const XMMRegister x4 = xmm4;
5820 const XMMRegister x5 = xmm5;
5821 const XMMRegister x6 = xmm6;
5822 const XMMRegister x7 = xmm7;
5823
5824 const Register tmp1 = r11;
5825 const Register tmp2 = r8;
5826
5827 BLOCK_COMMENT("Entry:");
5828 __ enter(); // required for proper stackwalking of RuntimeStub frame
5829
5830 __ fast_log(x0, x1, x2, x3, x4, x5, x6, x7, rax, rcx, rdx, tmp1, tmp2);
5831
5832 __ leave(); // required for proper stackwalking of RuntimeStub frame
5833 __ ret(0);
5834
5835 return start;
5836
5837 }
5838
5839 address generate_libmLog10() {
5840 StubCodeMark mark(this, "StubRoutines", "libmLog10");
5841
5842 address start = __ pc();
5843
5844 const XMMRegister x0 = xmm0;
5845 const XMMRegister x1 = xmm1;
5846 const XMMRegister x2 = xmm2;
5847 const XMMRegister x3 = xmm3;
5848
5849 const XMMRegister x4 = xmm4;
5850 const XMMRegister x5 = xmm5;
5851 const XMMRegister x6 = xmm6;
5852 const XMMRegister x7 = xmm7;
5853
5854 const Register tmp = r11;
5855
5856 BLOCK_COMMENT("Entry:");
5857 __ enter(); // required for proper stackwalking of RuntimeStub frame
5858
5859 __ fast_log10(x0, x1, x2, x3, x4, x5, x6, x7, rax, rcx, rdx, tmp);
5860
5861 __ leave(); // required for proper stackwalking of RuntimeStub frame
5862 __ ret(0);
5863
5864 return start;
5865
5866 }
5867
5868 address generate_libmPow() {
5869 StubCodeMark mark(this, "StubRoutines", "libmPow");
5870
5871 address start = __ pc();
5872
5873 const XMMRegister x0 = xmm0;
5874 const XMMRegister x1 = xmm1;
5875 const XMMRegister x2 = xmm2;
5876 const XMMRegister x3 = xmm3;
5877
5878 const XMMRegister x4 = xmm4;
5879 const XMMRegister x5 = xmm5;
5880 const XMMRegister x6 = xmm6;
5881 const XMMRegister x7 = xmm7;
5882
5883 const Register tmp1 = r8;
5884 const Register tmp2 = r9;
5885 const Register tmp3 = r10;
5886 const Register tmp4 = r11;
5887
5888 BLOCK_COMMENT("Entry:");
5889 __ enter(); // required for proper stackwalking of RuntimeStub frame
5890
5891 __ fast_pow(x0, x1, x2, x3, x4, x5, x6, x7, rax, rcx, rdx, tmp1, tmp2, tmp3, tmp4);
5892
5893 __ leave(); // required for proper stackwalking of RuntimeStub frame
5894 __ ret(0);
5895
5896 return start;
5897
5898 }
5899
5900 address generate_libmSin() {
5901 StubCodeMark mark(this, "StubRoutines", "libmSin");
5902
5903 address start = __ pc();
5904
5905 const XMMRegister x0 = xmm0;
5906 const XMMRegister x1 = xmm1;
5907 const XMMRegister x2 = xmm2;
5908 const XMMRegister x3 = xmm3;
5909
5910 const XMMRegister x4 = xmm4;
5911 const XMMRegister x5 = xmm5;
5912 const XMMRegister x6 = xmm6;
5913 const XMMRegister x7 = xmm7;
5914
5915 const Register tmp1 = r8;
5916 const Register tmp2 = r9;
5917 const Register tmp3 = r10;
5918 const Register tmp4 = r11;
5919
5920 BLOCK_COMMENT("Entry:");
5921 __ enter(); // required for proper stackwalking of RuntimeStub frame
5922
5923 #ifdef _WIN64
5924 __ push(rsi);
5925 __ push(rdi);
5926 #endif
5927 __ fast_sin(x0, x1, x2, x3, x4, x5, x6, x7, rax, rbx, rcx, rdx, tmp1, tmp2, tmp3, tmp4);
5928
5929 #ifdef _WIN64
5930 __ pop(rdi);
5931 __ pop(rsi);
5932 #endif
5933
5934 __ leave(); // required for proper stackwalking of RuntimeStub frame
5935 __ ret(0);
5936
5937 return start;
5938
5939 }
5940
5941 address generate_libmCos() {
5942 StubCodeMark mark(this, "StubRoutines", "libmCos");
5943
5944 address start = __ pc();
5945
5946 const XMMRegister x0 = xmm0;
5947 const XMMRegister x1 = xmm1;
5948 const XMMRegister x2 = xmm2;
5949 const XMMRegister x3 = xmm3;
5950
5951 const XMMRegister x4 = xmm4;
5952 const XMMRegister x5 = xmm5;
5953 const XMMRegister x6 = xmm6;
5954 const XMMRegister x7 = xmm7;
5955
5956 const Register tmp1 = r8;
5957 const Register tmp2 = r9;
5958 const Register tmp3 = r10;
5959 const Register tmp4 = r11;
5960
5961 BLOCK_COMMENT("Entry:");
5962 __ enter(); // required for proper stackwalking of RuntimeStub frame
5963
5964 #ifdef _WIN64
5965 __ push(rsi);
5966 __ push(rdi);
5967 #endif
5968 __ fast_cos(x0, x1, x2, x3, x4, x5, x6, x7, rax, rcx, rdx, tmp1, tmp2, tmp3, tmp4);
5969
5970 #ifdef _WIN64
5971 __ pop(rdi);
5972 __ pop(rsi);
5973 #endif
5974
5975 __ leave(); // required for proper stackwalking of RuntimeStub frame
5976 __ ret(0);
5977
5978 return start;
5979
5980 }
5981
5982 address generate_libmTan() {
5983 StubCodeMark mark(this, "StubRoutines", "libmTan");
5984
5985 address start = __ pc();
5986
5987 const XMMRegister x0 = xmm0;
5988 const XMMRegister x1 = xmm1;
5989 const XMMRegister x2 = xmm2;
5990 const XMMRegister x3 = xmm3;
5991
5992 const XMMRegister x4 = xmm4;
5993 const XMMRegister x5 = xmm5;
5994 const XMMRegister x6 = xmm6;
5995 const XMMRegister x7 = xmm7;
5996
5997 const Register tmp1 = r8;
5998 const Register tmp2 = r9;
5999 const Register tmp3 = r10;
6000 const Register tmp4 = r11;
6001
6002 BLOCK_COMMENT("Entry:");
6003 __ enter(); // required for proper stackwalking of RuntimeStub frame
6004
6005 #ifdef _WIN64
6006 __ push(rsi);
6007 __ push(rdi);
6008 #endif
6009 __ fast_tan(x0, x1, x2, x3, x4, x5, x6, x7, rax, rcx, rdx, tmp1, tmp2, tmp3, tmp4);
6010
6011 #ifdef _WIN64
6012 __ pop(rdi);
6013 __ pop(rsi);
6014 #endif
6015
6016 __ leave(); // required for proper stackwalking of RuntimeStub frame
6017 __ ret(0);
6018
6019 return start;
6020
6021 }
6022
6023 #undef __
6024 #define __ masm->
6025
6026 // Continuation point for throwing of implicit exceptions that are
6027 // not handled in the current activation. Fabricates an exception
6028 // oop and initiates normal exception dispatching in this
6029 // frame. Since we need to preserve callee-saved values (currently
6030 // only for C2, but done for C1 as well) we need a callee-saved oop
6031 // map and therefore have to make these stubs into RuntimeStubs
6032 // rather than BufferBlobs. If the compiler needs all registers to
6033 // be preserved between the fault point and the exception handler
6034 // then it must assume responsibility for that in
6035 // AbstractCompiler::continuation_for_implicit_null_exception or
6036 // continuation_for_implicit_division_by_zero_exception. All other
6037 // implicit exceptions (e.g., NullPointerException or
6038 // AbstractMethodError on entry) are either at call sites or
6039 // otherwise assume that stack unwinding will be initiated, so
6040 // caller saved registers were assumed volatile in the compiler.
6041 address generate_throw_exception(const char* name,
6042 address runtime_entry,
6043 Register arg1 = noreg,
6044 Register arg2 = noreg) {
6045 // Information about frame layout at time of blocking runtime call.
6046 // Note that we only have to preserve callee-saved registers since
6047 // the compilers are responsible for supplying a continuation point
6048 // if they expect all registers to be preserved.
6049 enum layout {
6050 rbp_off = frame::arg_reg_save_area_bytes/BytesPerInt,
6051 rbp_off2,
6052 return_off,
6053 return_off2,
6054 framesize // inclusive of return address
6055 };
6056
6057 int insts_size = 512;
6058 int locs_size = 64;
6059
6060 CodeBuffer code(name, insts_size, locs_size);
6061 OopMapSet* oop_maps = new OopMapSet();
6062 MacroAssembler* masm = new MacroAssembler(&code);
6063
6064 address start = __ pc();
6065
6066 // This is an inlined and slightly modified version of call_VM
6067 // which has the ability to fetch the return PC out of
6068 // thread-local storage and also sets up last_Java_sp slightly
6069 // differently than the real call_VM
6070
6071 __ enter(); // required for proper stackwalking of RuntimeStub frame
6072
6073 assert(is_even(framesize/2), "sp not 16-byte aligned");
6074
6075 // return address and rbp are already in place
6076 __ subptr(rsp, (framesize-4) << LogBytesPerInt); // prolog
6077
6078 int frame_complete = __ pc() - start;
6079
6080 // Set up last_Java_sp and last_Java_fp
6081 address the_pc = __ pc();
6082 __ set_last_Java_frame(rsp, rbp, the_pc);
6083 __ andptr(rsp, -(StackAlignmentInBytes)); // Align stack
6084
6085 // Call runtime
6086 if (arg1 != noreg) {
6087 assert(arg2 != c_rarg1, "clobbered");
6088 __ movptr(c_rarg1, arg1);
6089 }
6090 if (arg2 != noreg) {
6091 __ movptr(c_rarg2, arg2);
6092 }
6093 __ movptr(c_rarg0, r15_thread);
6094 BLOCK_COMMENT("call runtime_entry");
6095 __ call(RuntimeAddress(runtime_entry));
6096
6097 // Generate oop map
6098 OopMap* map = new OopMap(framesize, 0);
6099
6100 oop_maps->add_gc_map(the_pc - start, map);
6101
6102 __ reset_last_Java_frame(true);
6103
6104 __ leave(); // required for proper stackwalking of RuntimeStub frame
6105
6106 // check for pending exceptions
6107 #ifdef ASSERT
6108 Label L;
6109 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()),
6110 (int32_t) NULL_WORD);
6111 __ jcc(Assembler::notEqual, L);
6112 __ should_not_reach_here();
6113 __ bind(L);
6114 #endif // ASSERT
6115 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
6116
6117
6118 // codeBlob framesize is in words (not VMRegImpl::slot_size)
6119 RuntimeStub* stub =
6120 RuntimeStub::new_runtime_stub(name,
6121 &code,
6122 frame_complete,
6123 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
6124 oop_maps, false);
6125 return stub->entry_point();
6126 }
6127
6128 void create_control_words() {
6129 // Round to nearest, 53-bit mode, exceptions masked
6130 StubRoutines::_fpu_cntrl_wrd_std = 0x027F;
6131 // Round to zero, 53-bit mode, exception mased
6132 StubRoutines::_fpu_cntrl_wrd_trunc = 0x0D7F;
6133 // Round to nearest, 24-bit mode, exceptions masked
6134 StubRoutines::_fpu_cntrl_wrd_24 = 0x007F;
6135 // Round to nearest, 64-bit mode, exceptions masked
6136 StubRoutines::_mxcsr_std = 0x1F80;
6137 // Note: the following two constants are 80-bit values
6138 // layout is critical for correct loading by FPU.
6139 // Bias for strict fp multiply/divide
6140 StubRoutines::_fpu_subnormal_bias1[0]= 0x00000000; // 2^(-15360) == 0x03ff 8000 0000 0000 0000
6141 StubRoutines::_fpu_subnormal_bias1[1]= 0x80000000;
6142 StubRoutines::_fpu_subnormal_bias1[2]= 0x03ff;
6143 // Un-Bias for strict fp multiply/divide
6144 StubRoutines::_fpu_subnormal_bias2[0]= 0x00000000; // 2^(+15360) == 0x7bff 8000 0000 0000 0000
6145 StubRoutines::_fpu_subnormal_bias2[1]= 0x80000000;
6146 StubRoutines::_fpu_subnormal_bias2[2]= 0x7bff;
6147 }
6148
6149 // Initialization
6150 void generate_initial() {
6151 // Generates all stubs and initializes the entry points
6152
6153 // This platform-specific settings are needed by generate_call_stub()
6154 create_control_words();
6155
6156 // entry points that exist in all platforms Note: This is code
6157 // that could be shared among different platforms - however the
6158 // benefit seems to be smaller than the disadvantage of having a
6159 // much more complicated generator structure. See also comment in
6160 // stubRoutines.hpp.
6161
6162 StubRoutines::_forward_exception_entry = generate_forward_exception();
6163
6164 StubRoutines::_call_stub_entry =
6165 generate_call_stub(StubRoutines::_call_stub_return_address);
6166
6167 // is referenced by megamorphic call
6168 StubRoutines::_catch_exception_entry = generate_catch_exception();
6169
6170 // atomic calls
6171 StubRoutines::_fence_entry = generate_orderaccess_fence();
6172
6173 // platform dependent
6174 StubRoutines::x86::_get_previous_fp_entry = generate_get_previous_fp();
6175 StubRoutines::x86::_get_previous_sp_entry = generate_get_previous_sp();
6176
6177 StubRoutines::x86::_verify_mxcsr_entry = generate_verify_mxcsr();
6178
6179 StubRoutines::x86::_f2i_fixup = generate_f2i_fixup();
6180 StubRoutines::x86::_f2l_fixup = generate_f2l_fixup();
6181 StubRoutines::x86::_d2i_fixup = generate_d2i_fixup();
6182 StubRoutines::x86::_d2l_fixup = generate_d2l_fixup();
6183
6184 StubRoutines::x86::_float_sign_mask = generate_fp_mask("float_sign_mask", 0x7FFFFFFF7FFFFFFF);
6185 StubRoutines::x86::_float_sign_flip = generate_fp_mask("float_sign_flip", 0x8000000080000000);
6186 StubRoutines::x86::_double_sign_mask = generate_fp_mask("double_sign_mask", 0x7FFFFFFFFFFFFFFF);
6187 StubRoutines::x86::_double_sign_flip = generate_fp_mask("double_sign_flip", 0x8000000000000000);
6188
6189 // Build this early so it's available for the interpreter.
6190 StubRoutines::_throw_StackOverflowError_entry =
6191 generate_throw_exception("StackOverflowError throw_exception",
6192 CAST_FROM_FN_PTR(address,
6193 SharedRuntime::
6194 throw_StackOverflowError));
6195 StubRoutines::_throw_delayed_StackOverflowError_entry =
6196 generate_throw_exception("delayed StackOverflowError throw_exception",
6197 CAST_FROM_FN_PTR(address,
6198 SharedRuntime::
6199 throw_delayed_StackOverflowError));
6200 if (UseCRC32Intrinsics) {
6201 // set table address before stub generation which use it
6202 StubRoutines::_crc_table_adr = (address)StubRoutines::x86::_crc_table;
6203 StubRoutines::_updateBytesCRC32 = generate_updateBytesCRC32();
6204 }
6205
6206 if (UseCRC32CIntrinsics) {
6207 bool supports_clmul = VM_Version::supports_clmul();
6208 StubRoutines::x86::generate_CRC32C_table(supports_clmul);
6209 StubRoutines::_crc32c_table_addr = (address)StubRoutines::x86::_crc32c_table;
6210 StubRoutines::_updateBytesCRC32C = generate_updateBytesCRC32C(supports_clmul);
6211 }
6212 if (UseLibmIntrinsic && InlineIntrinsics) {
6213 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dsin) ||
6214 vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dcos) ||
6215 vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dtan)) {
6216 StubRoutines::x86::_ONEHALF_adr = (address)StubRoutines::x86::_ONEHALF;
6217 StubRoutines::x86::_P_2_adr = (address)StubRoutines::x86::_P_2;
6218 StubRoutines::x86::_SC_4_adr = (address)StubRoutines::x86::_SC_4;
6219 StubRoutines::x86::_Ctable_adr = (address)StubRoutines::x86::_Ctable;
6220 StubRoutines::x86::_SC_2_adr = (address)StubRoutines::x86::_SC_2;
6221 StubRoutines::x86::_SC_3_adr = (address)StubRoutines::x86::_SC_3;
6222 StubRoutines::x86::_SC_1_adr = (address)StubRoutines::x86::_SC_1;
6223 StubRoutines::x86::_PI_INV_TABLE_adr = (address)StubRoutines::x86::_PI_INV_TABLE;
6224 StubRoutines::x86::_PI_4_adr = (address)StubRoutines::x86::_PI_4;
6225 StubRoutines::x86::_PI32INV_adr = (address)StubRoutines::x86::_PI32INV;
6226 StubRoutines::x86::_SIGN_MASK_adr = (address)StubRoutines::x86::_SIGN_MASK;
6227 StubRoutines::x86::_P_1_adr = (address)StubRoutines::x86::_P_1;
6228 StubRoutines::x86::_P_3_adr = (address)StubRoutines::x86::_P_3;
6229 StubRoutines::x86::_NEG_ZERO_adr = (address)StubRoutines::x86::_NEG_ZERO;
6230 }
6231 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dexp)) {
6232 StubRoutines::_dexp = generate_libmExp();
6233 }
6234 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dlog)) {
6235 StubRoutines::_dlog = generate_libmLog();
6236 }
6237 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dlog10)) {
6238 StubRoutines::_dlog10 = generate_libmLog10();
6239 }
6240 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dpow)) {
6241 StubRoutines::_dpow = generate_libmPow();
6242 }
6243 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dsin)) {
6244 StubRoutines::_dsin = generate_libmSin();
6245 }
6246 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dcos)) {
6247 StubRoutines::_dcos = generate_libmCos();
6248 }
6249 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dtan)) {
6250 StubRoutines::_dtan = generate_libmTan();
6251 }
6252 }
6253
6254 // Safefetch stubs.
6255 generate_safefetch("SafeFetch32", sizeof(int), &StubRoutines::_safefetch32_entry,
6256 &StubRoutines::_safefetch32_fault_pc,
6257 &StubRoutines::_safefetch32_continuation_pc);
6258 generate_safefetch("SafeFetchN", sizeof(intptr_t), &StubRoutines::_safefetchN_entry,
6259 &StubRoutines::_safefetchN_fault_pc,
6260 &StubRoutines::_safefetchN_continuation_pc);
6261 }
6262
6263 void generate_all() {
6264 // Generates all stubs and initializes the entry points
6265
6266 // These entry points require SharedInfo::stack0 to be set up in
6267 // non-core builds and need to be relocatable, so they each
6268 // fabricate a RuntimeStub internally.
6269 StubRoutines::_throw_AbstractMethodError_entry =
6270 generate_throw_exception("AbstractMethodError throw_exception",
6271 CAST_FROM_FN_PTR(address,
6272 SharedRuntime::
6273 throw_AbstractMethodError));
6274
6275 StubRoutines::_throw_IncompatibleClassChangeError_entry =
6276 generate_throw_exception("IncompatibleClassChangeError throw_exception",
6277 CAST_FROM_FN_PTR(address,
6278 SharedRuntime::
6279 throw_IncompatibleClassChangeError));
6280
6281 StubRoutines::_throw_NullPointerException_at_call_entry =
6282 generate_throw_exception("NullPointerException at call throw_exception",
6283 CAST_FROM_FN_PTR(address,
6284 SharedRuntime::
6285 throw_NullPointerException_at_call));
6286
6287 // entry points that are platform specific
6288 StubRoutines::x86::_vector_float_sign_mask = generate_vector_mask("vector_float_sign_mask", 0x7FFFFFFF7FFFFFFF);
6289 StubRoutines::x86::_vector_float_sign_flip = generate_vector_mask("vector_float_sign_flip", 0x8000000080000000);
6290 StubRoutines::x86::_vector_double_sign_mask = generate_vector_mask("vector_double_sign_mask", 0x7FFFFFFFFFFFFFFF);
6291 StubRoutines::x86::_vector_double_sign_flip = generate_vector_mask("vector_double_sign_flip", 0x8000000000000000);
6292 StubRoutines::x86::_vector_short_to_byte_mask = generate_vector_mask("vector_short_to_byte_mask", 0x00ff00ff00ff00ff);
6293 StubRoutines::x86::_vector_byte_perm_mask = generate_vector_byte_perm_mask("vector_byte_perm_mask");
6294 StubRoutines::x86::_vector_long_sign_mask = generate_vector_mask("vector_long_sign_mask", 0x8000000000000000);
6295
6296 // support for verify_oop (must happen after universe_init)
6297 StubRoutines::_verify_oop_subroutine_entry = generate_verify_oop();
6298
6299 // data cache line writeback
6300 StubRoutines::_data_cache_writeback = generate_data_cache_writeback();
6301 StubRoutines::_data_cache_writeback_sync = generate_data_cache_writeback_sync();
6302
6303 // arraycopy stubs used by compilers
6304 generate_arraycopy_stubs();
6305
6306 // don't bother generating these AES intrinsic stubs unless global flag is set
6307 if (UseAESIntrinsics) {
6308 StubRoutines::x86::_key_shuffle_mask_addr = generate_key_shuffle_mask(); // needed by the others
6309 StubRoutines::_aescrypt_encryptBlock = generate_aescrypt_encryptBlock();
6310 StubRoutines::_aescrypt_decryptBlock = generate_aescrypt_decryptBlock();
6311 StubRoutines::_cipherBlockChaining_encryptAESCrypt = generate_cipherBlockChaining_encryptAESCrypt();
6312 if (VM_Version::supports_avx512_vaes() && VM_Version::supports_avx512vl() && VM_Version::supports_avx512dq() ) {
6313 StubRoutines::_cipherBlockChaining_decryptAESCrypt = generate_cipherBlockChaining_decryptVectorAESCrypt();
6314 StubRoutines::_electronicCodeBook_encryptAESCrypt = generate_electronicCodeBook_encryptAESCrypt();
6315 StubRoutines::_electronicCodeBook_decryptAESCrypt = generate_electronicCodeBook_decryptAESCrypt();
6316 } else {
6317 StubRoutines::_cipherBlockChaining_decryptAESCrypt = generate_cipherBlockChaining_decryptAESCrypt_Parallel();
6318 }
6319 }
6320 if (UseAESCTRIntrinsics) {
6321 if (VM_Version::supports_avx512_vaes() && VM_Version::supports_avx512bw() && VM_Version::supports_avx512vl()) {
6322 StubRoutines::x86::_counter_mask_addr = counter_mask_addr();
6323 StubRoutines::_counterMode_AESCrypt = generate_counterMode_VectorAESCrypt();
6324 } else {
6325 StubRoutines::x86::_counter_shuffle_mask_addr = generate_counter_shuffle_mask();
6326 StubRoutines::_counterMode_AESCrypt = generate_counterMode_AESCrypt_Parallel();
6327 }
6328 }
6329
6330 if (UseSHA1Intrinsics) {
6331 StubRoutines::x86::_upper_word_mask_addr = generate_upper_word_mask();
6332 StubRoutines::x86::_shuffle_byte_flip_mask_addr = generate_shuffle_byte_flip_mask();
6333 StubRoutines::_sha1_implCompress = generate_sha1_implCompress(false, "sha1_implCompress");
6334 StubRoutines::_sha1_implCompressMB = generate_sha1_implCompress(true, "sha1_implCompressMB");
6335 }
6336 if (UseSHA256Intrinsics) {
6337 StubRoutines::x86::_k256_adr = (address)StubRoutines::x86::_k256;
6338 char* dst = (char*)StubRoutines::x86::_k256_W;
6339 char* src = (char*)StubRoutines::x86::_k256;
6340 for (int ii = 0; ii < 16; ++ii) {
6341 memcpy(dst + 32 * ii, src + 16 * ii, 16);
6342 memcpy(dst + 32 * ii + 16, src + 16 * ii, 16);
6343 }
6344 StubRoutines::x86::_k256_W_adr = (address)StubRoutines::x86::_k256_W;
6345 StubRoutines::x86::_pshuffle_byte_flip_mask_addr = generate_pshuffle_byte_flip_mask();
6346 StubRoutines::_sha256_implCompress = generate_sha256_implCompress(false, "sha256_implCompress");
6347 StubRoutines::_sha256_implCompressMB = generate_sha256_implCompress(true, "sha256_implCompressMB");
6348 }
6349 if (UseSHA512Intrinsics) {
6350 StubRoutines::x86::_k512_W_addr = (address)StubRoutines::x86::_k512_W;
6351 StubRoutines::x86::_pshuffle_byte_flip_mask_addr_sha512 = generate_pshuffle_byte_flip_mask_sha512();
6352 StubRoutines::_sha512_implCompress = generate_sha512_implCompress(false, "sha512_implCompress");
6353 StubRoutines::_sha512_implCompressMB = generate_sha512_implCompress(true, "sha512_implCompressMB");
6354 }
6355
6356 // Generate GHASH intrinsics code
6357 if (UseGHASHIntrinsics) {
6358 StubRoutines::x86::_ghash_long_swap_mask_addr = generate_ghash_long_swap_mask();
6359 StubRoutines::x86::_ghash_byte_swap_mask_addr = generate_ghash_byte_swap_mask();
6360 if (VM_Version::supports_avx()) {
6361 StubRoutines::x86::_ghash_shuffmask_addr = ghash_shufflemask_addr();
6362 StubRoutines::x86::_ghash_poly_addr = ghash_polynomial_addr();
6363 StubRoutines::_ghash_processBlocks = generate_avx_ghash_processBlocks();
6364 } else {
6365 StubRoutines::_ghash_processBlocks = generate_ghash_processBlocks();
6366 }
6367 }
6368
6369 if (UseBASE64Intrinsics) {
6370 StubRoutines::x86::_and_mask = base64_and_mask_addr();
6371 StubRoutines::x86::_bswap_mask = base64_bswap_mask_addr();
6372 StubRoutines::x86::_base64_charset = base64_charset_addr();
6373 StubRoutines::x86::_url_charset = base64url_charset_addr();
6374 StubRoutines::x86::_gather_mask = base64_gather_mask_addr();
6375 StubRoutines::x86::_left_shift_mask = base64_left_shift_mask_addr();
6376 StubRoutines::x86::_right_shift_mask = base64_right_shift_mask_addr();
6377 StubRoutines::_base64_encodeBlock = generate_base64_encodeBlock();
6378 }
6379
6380 BarrierSetNMethod* bs_nm = BarrierSet::barrier_set()->barrier_set_nmethod();
6381 if (bs_nm != NULL) {
6382 StubRoutines::x86::_method_entry_barrier = generate_method_entry_barrier();
6383 }
6384 #ifdef COMPILER2
6385 if (UseMultiplyToLenIntrinsic) {
6386 StubRoutines::_multiplyToLen = generate_multiplyToLen();
6387 }
6388 if (UseSquareToLenIntrinsic) {
6389 StubRoutines::_squareToLen = generate_squareToLen();
6390 }
6391 if (UseMulAddIntrinsic) {
6392 StubRoutines::_mulAdd = generate_mulAdd();
6393 }
6394 if (VM_Version::supports_avx512_vbmi2()) {
6395 StubRoutines::_bigIntegerRightShiftWorker = generate_bigIntegerRightShift();
6396 StubRoutines::_bigIntegerLeftShiftWorker = generate_bigIntegerLeftShift();
6397 }
6398 if (UseMontgomeryMultiplyIntrinsic) {
6399 StubRoutines::_montgomeryMultiply
6400 = CAST_FROM_FN_PTR(address, SharedRuntime::montgomery_multiply);
6401 }
6402 if (UseMontgomerySquareIntrinsic) {
6403 StubRoutines::_montgomerySquare
6404 = CAST_FROM_FN_PTR(address, SharedRuntime::montgomery_square);
6405 }
6406 #endif // COMPILER2
6407
6408 if (UseVectorizedMismatchIntrinsic) {
6409 StubRoutines::_vectorizedMismatch = generate_vectorizedMismatch();
6410 }
6411 }
6412
6413 public:
6414 StubGenerator(CodeBuffer* code, bool all) : StubCodeGenerator(code) {
6415 if (all) {
6416 generate_all();
6417 } else {
6418 generate_initial();
6419 }
6420 }
6421 }; // end class declaration
6422
6423 #define UCM_TABLE_MAX_ENTRIES 16
6424 void StubGenerator_generate(CodeBuffer* code, bool all) {
6425 if (UnsafeCopyMemory::_table == NULL) {
6426 UnsafeCopyMemory::create_table(UCM_TABLE_MAX_ENTRIES);
6427 }
6428 StubGenerator g(code, all);
6429 }