1 /*
2 * Copyright (c) 2003, 2011, 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/assembler.hpp"
27 #include "assembler_x86.inline.hpp"
28 #include "interpreter/interpreter.hpp"
29 #include "nativeInst_x86.hpp"
30 #include "oops/instanceOop.hpp"
31 #include "oops/methodOop.hpp"
32 #include "oops/objArrayKlass.hpp"
33 #include "oops/oop.inline.hpp"
34 #include "prims/methodHandles.hpp"
35 #include "runtime/frame.inline.hpp"
36 #include "runtime/handles.inline.hpp"
37 #include "runtime/sharedRuntime.hpp"
38 #include "runtime/stubCodeGenerator.hpp"
39 #include "runtime/stubRoutines.hpp"
40 #include "utilities/top.hpp"
41 #ifdef TARGET_OS_FAMILY_linux
42 # include "thread_linux.inline.hpp"
43 #endif
44 #ifdef TARGET_OS_FAMILY_solaris
45 # include "thread_solaris.inline.hpp"
46 #endif
47 #ifdef TARGET_OS_FAMILY_windows
48 # include "thread_windows.inline.hpp"
49 #endif
50 #ifdef TARGET_OS_FAMILY_bsd
51 # include "thread_bsd.inline.hpp"
52 #endif
53 #ifdef COMPILER2
54 #include "opto/runtime.hpp"
55 #endif
56
57 // Declaration and definition of StubGenerator (no .hpp file).
58 // For a more detailed description of the stub routine structure
59 // see the comment in stubRoutines.hpp
60
61 #define __ _masm->
62 #define TIMES_OOP (UseCompressedOops ? Address::times_4 : Address::times_8)
63 #define a__ ((Assembler*)_masm)->
64
65 #ifdef PRODUCT
66 #define BLOCK_COMMENT(str) /* nothing */
67 #else
68 #define BLOCK_COMMENT(str) __ block_comment(str)
69 #endif
70
71 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
72 const int MXCSR_MASK = 0xFFC0; // Mask out any pending exceptions
73
74 // Stub Code definitions
75
76 static address handle_unsafe_access() {
77 JavaThread* thread = JavaThread::current();
78 address pc = thread->saved_exception_pc();
79 // pc is the instruction which we must emulate
80 // doing a no-op is fine: return garbage from the load
81 // therefore, compute npc
82 address npc = Assembler::locate_next_instruction(pc);
83
84 // request an async exception
85 thread->set_pending_unsafe_access_error();
86
87 // return address of next instruction to execute
88 return npc;
89 }
90
91 class StubGenerator: public StubCodeGenerator {
92 private:
93
94 #ifdef PRODUCT
95 #define inc_counter_np(counter) (0)
96 #else
97 void inc_counter_np_(int& counter) {
98 // This can destroy rscratch1 if counter is far from the code cache
99 __ incrementl(ExternalAddress((address)&counter));
100 }
101 #define inc_counter_np(counter) \
102 BLOCK_COMMENT("inc_counter " #counter); \
103 inc_counter_np_(counter);
104 #endif
105
106 // Call stubs are used to call Java from C
107 //
108 // Linux Arguments:
109 // c_rarg0: call wrapper address address
110 // c_rarg1: result address
111 // c_rarg2: result type BasicType
112 // c_rarg3: method methodOop
113 // c_rarg4: (interpreter) entry point address
114 // c_rarg5: parameters intptr_t*
115 // 16(rbp): parameter size (in words) int
116 // 24(rbp): thread Thread*
117 //
118 // [ return_from_Java ] <--- rsp
119 // [ argument word n ]
120 // ...
121 // -12 [ argument word 1 ]
122 // -11 [ saved r15 ] <--- rsp_after_call
123 // -10 [ saved r14 ]
124 // -9 [ saved r13 ]
125 // -8 [ saved r12 ]
126 // -7 [ saved rbx ]
127 // -6 [ call wrapper ]
128 // -5 [ result ]
129 // -4 [ result type ]
130 // -3 [ method ]
131 // -2 [ entry point ]
132 // -1 [ parameters ]
133 // 0 [ saved rbp ] <--- rbp
134 // 1 [ return address ]
135 // 2 [ parameter size ]
136 // 3 [ thread ]
137 //
138 // Windows Arguments:
139 // c_rarg0: call wrapper address address
140 // c_rarg1: result address
141 // c_rarg2: result type BasicType
142 // c_rarg3: method methodOop
143 // 48(rbp): (interpreter) entry point address
144 // 56(rbp): parameters intptr_t*
145 // 64(rbp): parameter size (in words) int
146 // 72(rbp): thread Thread*
147 //
148 // [ return_from_Java ] <--- rsp
149 // [ argument word n ]
150 // ...
151 // -28 [ argument word 1 ]
152 // -27 [ saved xmm15 ] <--- rsp_after_call
153 // [ saved xmm7-xmm14 ]
154 // -9 [ saved xmm6 ] (each xmm register takes 2 slots)
155 // -7 [ saved r15 ]
156 // -6 [ saved r14 ]
157 // -5 [ saved r13 ]
158 // -4 [ saved r12 ]
159 // -3 [ saved rdi ]
160 // -2 [ saved rsi ]
161 // -1 [ saved rbx ]
162 // 0 [ saved rbp ] <--- rbp
163 // 1 [ return address ]
164 // 2 [ call wrapper ]
165 // 3 [ result ]
166 // 4 [ result type ]
167 // 5 [ method ]
168 // 6 [ entry point ]
169 // 7 [ parameters ]
170 // 8 [ parameter size ]
171 // 9 [ thread ]
172 //
173 // Windows reserves the callers stack space for arguments 1-4.
174 // We spill c_rarg0-c_rarg3 to this space.
175
176 // Call stub stack layout word offsets from rbp
177 enum call_stub_layout {
178 #ifdef _WIN64
179 xmm_save_first = 6, // save from xmm6
180 xmm_save_last = 15, // to xmm15
181 xmm_save_base = -9,
182 rsp_after_call_off = xmm_save_base - 2 * (xmm_save_last - xmm_save_first), // -27
183 r15_off = -7,
184 r14_off = -6,
185 r13_off = -5,
186 r12_off = -4,
187 rdi_off = -3,
188 rsi_off = -2,
189 rbx_off = -1,
190 rbp_off = 0,
191 retaddr_off = 1,
192 call_wrapper_off = 2,
193 result_off = 3,
194 result_type_off = 4,
195 method_off = 5,
196 entry_point_off = 6,
197 parameters_off = 7,
198 parameter_size_off = 8,
199 thread_off = 9
200 #else
201 rsp_after_call_off = -12,
202 mxcsr_off = rsp_after_call_off,
203 r15_off = -11,
204 r14_off = -10,
205 r13_off = -9,
206 r12_off = -8,
207 rbx_off = -7,
208 call_wrapper_off = -6,
209 result_off = -5,
210 result_type_off = -4,
211 method_off = -3,
212 entry_point_off = -2,
213 parameters_off = -1,
214 rbp_off = 0,
215 retaddr_off = 1,
216 parameter_size_off = 2,
217 thread_off = 3
218 #endif
219 };
220
221 #ifdef _WIN64
222 Address xmm_save(int reg) {
223 assert(reg >= xmm_save_first && reg <= xmm_save_last, "XMM register number out of range");
224 return Address(rbp, (xmm_save_base - (reg - xmm_save_first) * 2) * wordSize);
225 }
226 #endif
227
228 address generate_call_stub(address& return_address) {
229 assert((int)frame::entry_frame_after_call_words == -(int)rsp_after_call_off + 1 &&
230 (int)frame::entry_frame_call_wrapper_offset == (int)call_wrapper_off,
231 "adjust this code");
232 StubCodeMark mark(this, "StubRoutines", "call_stub");
233 address start = __ pc();
234
235 // same as in generate_catch_exception()!
236 const Address rsp_after_call(rbp, rsp_after_call_off * wordSize);
237
238 const Address call_wrapper (rbp, call_wrapper_off * wordSize);
239 const Address result (rbp, result_off * wordSize);
240 const Address result_type (rbp, result_type_off * wordSize);
241 const Address method (rbp, method_off * wordSize);
242 const Address entry_point (rbp, entry_point_off * wordSize);
243 const Address parameters (rbp, parameters_off * wordSize);
244 const Address parameter_size(rbp, parameter_size_off * wordSize);
245
246 // same as in generate_catch_exception()!
247 const Address thread (rbp, thread_off * wordSize);
248
249 const Address r15_save(rbp, r15_off * wordSize);
250 const Address r14_save(rbp, r14_off * wordSize);
251 const Address r13_save(rbp, r13_off * wordSize);
252 const Address r12_save(rbp, r12_off * wordSize);
253 const Address rbx_save(rbp, rbx_off * wordSize);
254
255 // stub code
256 __ enter();
257 __ subptr(rsp, -rsp_after_call_off * wordSize);
258
259 // save register parameters
260 #ifndef _WIN64
261 __ movptr(parameters, c_rarg5); // parameters
262 __ movptr(entry_point, c_rarg4); // entry_point
263 #endif
264
265 __ movptr(method, c_rarg3); // method
266 __ movl(result_type, c_rarg2); // result type
267 __ movptr(result, c_rarg1); // result
268 __ movptr(call_wrapper, c_rarg0); // call wrapper
269
270 // save regs belonging to calling function
271 __ movptr(rbx_save, rbx);
272 __ movptr(r12_save, r12);
273 __ movptr(r13_save, r13);
274 __ movptr(r14_save, r14);
275 __ movptr(r15_save, r15);
276 #ifdef _WIN64
277 for (int i = 6; i <= 15; i++) {
278 __ movdqu(xmm_save(i), as_XMMRegister(i));
279 }
280
281 const Address rdi_save(rbp, rdi_off * wordSize);
282 const Address rsi_save(rbp, rsi_off * wordSize);
283
284 __ movptr(rsi_save, rsi);
285 __ movptr(rdi_save, rdi);
286 #else
287 const Address mxcsr_save(rbp, mxcsr_off * wordSize);
288 {
289 Label skip_ldmx;
290 __ stmxcsr(mxcsr_save);
291 __ movl(rax, mxcsr_save);
292 __ andl(rax, MXCSR_MASK); // Only check control and mask bits
293 ExternalAddress mxcsr_std(StubRoutines::x86::mxcsr_std());
294 __ cmp32(rax, mxcsr_std);
295 __ jcc(Assembler::equal, skip_ldmx);
296 __ ldmxcsr(mxcsr_std);
297 __ bind(skip_ldmx);
298 }
299 #endif
300
301 // Load up thread register
302 __ movptr(r15_thread, thread);
303 __ reinit_heapbase();
304
305 #ifdef ASSERT
306 // make sure we have no pending exceptions
307 {
308 Label L;
309 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
310 __ jcc(Assembler::equal, L);
311 __ stop("StubRoutines::call_stub: entered with pending exception");
312 __ bind(L);
313 }
314 #endif
315
316 // pass parameters if any
317 BLOCK_COMMENT("pass parameters if any");
318 Label parameters_done;
319 __ movl(c_rarg3, parameter_size);
320 __ testl(c_rarg3, c_rarg3);
321 __ jcc(Assembler::zero, parameters_done);
322
323 Label loop;
324 __ movptr(c_rarg2, parameters); // parameter pointer
325 __ movl(c_rarg1, c_rarg3); // parameter counter is in c_rarg1
326 __ BIND(loop);
327 __ movptr(rax, Address(c_rarg2, 0));// get parameter
328 __ addptr(c_rarg2, wordSize); // advance to next parameter
329 __ decrementl(c_rarg1); // decrement counter
330 __ push(rax); // pass parameter
331 __ jcc(Assembler::notZero, loop);
332
333 // call Java function
334 __ BIND(parameters_done);
335 __ movptr(rbx, method); // get methodOop
336 __ movptr(c_rarg1, entry_point); // get entry_point
337 __ mov(r13, rsp); // set sender sp
338 BLOCK_COMMENT("call Java function");
339 __ call(c_rarg1);
340
341 BLOCK_COMMENT("call_stub_return_address:");
342 return_address = __ pc();
343
344 // store result depending on type (everything that is not
345 // T_OBJECT, T_LONG, T_FLOAT or T_DOUBLE is treated as T_INT)
346 __ movptr(c_rarg0, result);
347 Label is_long, is_float, is_double, exit;
348 __ movl(c_rarg1, result_type);
349 __ cmpl(c_rarg1, T_OBJECT);
350 __ jcc(Assembler::equal, is_long);
351 __ cmpl(c_rarg1, T_LONG);
352 __ jcc(Assembler::equal, is_long);
353 __ cmpl(c_rarg1, T_FLOAT);
354 __ jcc(Assembler::equal, is_float);
355 __ cmpl(c_rarg1, T_DOUBLE);
356 __ jcc(Assembler::equal, is_double);
357
358 // handle T_INT case
359 __ movl(Address(c_rarg0, 0), rax);
360
361 __ BIND(exit);
362
363 // pop parameters
364 __ lea(rsp, rsp_after_call);
365
366 #ifdef ASSERT
367 // verify that threads correspond
368 {
369 Label L, S;
370 __ cmpptr(r15_thread, thread);
371 __ jcc(Assembler::notEqual, S);
372 __ get_thread(rbx);
373 __ cmpptr(r15_thread, rbx);
374 __ jcc(Assembler::equal, L);
375 __ bind(S);
376 __ jcc(Assembler::equal, L);
377 __ stop("StubRoutines::call_stub: threads must correspond");
378 __ bind(L);
379 }
380 #endif
381
382 // restore regs belonging to calling function
383 #ifdef _WIN64
384 for (int i = 15; i >= 6; i--) {
385 __ movdqu(as_XMMRegister(i), xmm_save(i));
386 }
387 #endif
388 __ movptr(r15, r15_save);
389 __ movptr(r14, r14_save);
390 __ movptr(r13, r13_save);
391 __ movptr(r12, r12_save);
392 __ movptr(rbx, rbx_save);
393
394 #ifdef _WIN64
395 __ movptr(rdi, rdi_save);
396 __ movptr(rsi, rsi_save);
397 #else
398 __ ldmxcsr(mxcsr_save);
399 #endif
400
401 // restore rsp
402 __ addptr(rsp, -rsp_after_call_off * wordSize);
403
404 // return
405 __ pop(rbp);
406 __ ret(0);
407
408 // handle return types different from T_INT
409 __ BIND(is_long);
410 __ movq(Address(c_rarg0, 0), rax);
411 __ jmp(exit);
412
413 __ BIND(is_float);
414 __ movflt(Address(c_rarg0, 0), xmm0);
415 __ jmp(exit);
416
417 __ BIND(is_double);
418 __ movdbl(Address(c_rarg0, 0), xmm0);
419 __ jmp(exit);
420
421 return start;
422 }
423
424 // Return point for a Java call if there's an exception thrown in
425 // Java code. The exception is caught and transformed into a
426 // pending exception stored in JavaThread that can be tested from
427 // within the VM.
428 //
429 // Note: Usually the parameters are removed by the callee. In case
430 // of an exception crossing an activation frame boundary, that is
431 // not the case if the callee is compiled code => need to setup the
432 // rsp.
433 //
434 // rax: exception oop
435
436 address generate_catch_exception() {
437 StubCodeMark mark(this, "StubRoutines", "catch_exception");
438 address start = __ pc();
439
440 // same as in generate_call_stub():
441 const Address rsp_after_call(rbp, rsp_after_call_off * wordSize);
442 const Address thread (rbp, thread_off * wordSize);
443
444 #ifdef ASSERT
445 // verify that threads correspond
446 {
447 Label L, S;
448 __ cmpptr(r15_thread, thread);
449 __ jcc(Assembler::notEqual, S);
450 __ get_thread(rbx);
451 __ cmpptr(r15_thread, rbx);
452 __ jcc(Assembler::equal, L);
453 __ bind(S);
454 __ stop("StubRoutines::catch_exception: threads must correspond");
455 __ bind(L);
456 }
457 #endif
458
459 // set pending exception
460 __ verify_oop(rax);
461
462 __ movptr(Address(r15_thread, Thread::pending_exception_offset()), rax);
463 __ lea(rscratch1, ExternalAddress((address)__FILE__));
464 __ movptr(Address(r15_thread, Thread::exception_file_offset()), rscratch1);
465 __ movl(Address(r15_thread, Thread::exception_line_offset()), (int) __LINE__);
466
467 // complete return to VM
468 assert(StubRoutines::_call_stub_return_address != NULL,
469 "_call_stub_return_address must have been generated before");
470 __ jump(RuntimeAddress(StubRoutines::_call_stub_return_address));
471
472 return start;
473 }
474
475 // Continuation point for runtime calls returning with a pending
476 // exception. The pending exception check happened in the runtime
477 // or native call stub. The pending exception in Thread is
478 // converted into a Java-level exception.
479 //
480 // Contract with Java-level exception handlers:
481 // rax: exception
482 // rdx: throwing pc
483 //
484 // NOTE: At entry of this stub, exception-pc must be on stack !!
485
486 address generate_forward_exception() {
487 StubCodeMark mark(this, "StubRoutines", "forward exception");
488 address start = __ pc();
489
490 // Upon entry, the sp points to the return address returning into
491 // Java (interpreted or compiled) code; i.e., the return address
492 // becomes the throwing pc.
493 //
494 // Arguments pushed before the runtime call are still on the stack
495 // but the exception handler will reset the stack pointer ->
496 // ignore them. A potential result in registers can be ignored as
497 // well.
498
499 #ifdef ASSERT
500 // make sure this code is only executed if there is a pending exception
501 {
502 Label L;
503 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t) NULL);
504 __ jcc(Assembler::notEqual, L);
505 __ stop("StubRoutines::forward exception: no pending exception (1)");
506 __ bind(L);
507 }
508 #endif
509
510 // compute exception handler into rbx
511 __ movptr(c_rarg0, Address(rsp, 0));
512 BLOCK_COMMENT("call exception_handler_for_return_address");
513 __ call_VM_leaf(CAST_FROM_FN_PTR(address,
514 SharedRuntime::exception_handler_for_return_address),
515 r15_thread, c_rarg0);
516 __ mov(rbx, rax);
517
518 // setup rax & rdx, remove return address & clear pending exception
519 __ pop(rdx);
520 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset()));
521 __ movptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
522
523 #ifdef ASSERT
524 // make sure exception is set
525 {
526 Label L;
527 __ testptr(rax, rax);
528 __ jcc(Assembler::notEqual, L);
529 __ stop("StubRoutines::forward exception: no pending exception (2)");
530 __ bind(L);
531 }
532 #endif
533
534 // continue at exception handler (return address removed)
535 // rax: exception
536 // rbx: exception handler
537 // rdx: throwing pc
538 __ verify_oop(rax);
539 __ jmp(rbx);
540
541 return start;
542 }
543
544 // Support for jint atomic::xchg(jint exchange_value, volatile jint* dest)
545 //
546 // Arguments :
547 // c_rarg0: exchange_value
548 // c_rarg0: dest
549 //
550 // Result:
551 // *dest <- ex, return (orig *dest)
552 address generate_atomic_xchg() {
553 StubCodeMark mark(this, "StubRoutines", "atomic_xchg");
554 address start = __ pc();
555
556 __ movl(rax, c_rarg0); // Copy to eax we need a return value anyhow
557 __ xchgl(rax, Address(c_rarg1, 0)); // automatic LOCK
558 __ ret(0);
559
560 return start;
561 }
562
563 // Support for intptr_t atomic::xchg_ptr(intptr_t exchange_value, volatile intptr_t* dest)
564 //
565 // Arguments :
566 // c_rarg0: exchange_value
567 // c_rarg1: dest
568 //
569 // Result:
570 // *dest <- ex, return (orig *dest)
571 address generate_atomic_xchg_ptr() {
572 StubCodeMark mark(this, "StubRoutines", "atomic_xchg_ptr");
573 address start = __ pc();
574
575 __ movptr(rax, c_rarg0); // Copy to eax we need a return value anyhow
576 __ xchgptr(rax, Address(c_rarg1, 0)); // automatic LOCK
577 __ ret(0);
578
579 return start;
580 }
581
582 // Support for jint atomic::atomic_cmpxchg(jint exchange_value, volatile jint* dest,
583 // jint compare_value)
584 //
585 // Arguments :
586 // c_rarg0: exchange_value
587 // c_rarg1: dest
588 // c_rarg2: compare_value
589 //
590 // Result:
591 // if ( compare_value == *dest ) {
592 // *dest = exchange_value
593 // return compare_value;
594 // else
595 // return *dest;
596 address generate_atomic_cmpxchg() {
597 StubCodeMark mark(this, "StubRoutines", "atomic_cmpxchg");
598 address start = __ pc();
599
600 __ movl(rax, c_rarg2);
601 if ( os::is_MP() ) __ lock();
602 __ cmpxchgl(c_rarg0, Address(c_rarg1, 0));
603 __ ret(0);
604
605 return start;
606 }
607
608 // Support for jint atomic::atomic_cmpxchg_long(jlong exchange_value,
609 // volatile jlong* dest,
610 // jlong compare_value)
611 // Arguments :
612 // c_rarg0: exchange_value
613 // c_rarg1: dest
614 // c_rarg2: compare_value
615 //
616 // Result:
617 // if ( compare_value == *dest ) {
618 // *dest = exchange_value
619 // return compare_value;
620 // else
621 // return *dest;
622 address generate_atomic_cmpxchg_long() {
623 StubCodeMark mark(this, "StubRoutines", "atomic_cmpxchg_long");
624 address start = __ pc();
625
626 __ movq(rax, c_rarg2);
627 if ( os::is_MP() ) __ lock();
628 __ cmpxchgq(c_rarg0, Address(c_rarg1, 0));
629 __ ret(0);
630
631 return start;
632 }
633
634 // Support for jint atomic::add(jint add_value, volatile jint* dest)
635 //
636 // Arguments :
637 // c_rarg0: add_value
638 // c_rarg1: dest
639 //
640 // Result:
641 // *dest += add_value
642 // return *dest;
643 address generate_atomic_add() {
644 StubCodeMark mark(this, "StubRoutines", "atomic_add");
645 address start = __ pc();
646
647 __ movl(rax, c_rarg0);
648 if ( os::is_MP() ) __ lock();
649 __ xaddl(Address(c_rarg1, 0), c_rarg0);
650 __ addl(rax, c_rarg0);
651 __ ret(0);
652
653 return start;
654 }
655
656 // Support for intptr_t atomic::add_ptr(intptr_t add_value, volatile intptr_t* dest)
657 //
658 // Arguments :
659 // c_rarg0: add_value
660 // c_rarg1: dest
661 //
662 // Result:
663 // *dest += add_value
664 // return *dest;
665 address generate_atomic_add_ptr() {
666 StubCodeMark mark(this, "StubRoutines", "atomic_add_ptr");
667 address start = __ pc();
668
669 __ movptr(rax, c_rarg0); // Copy to eax we need a return value anyhow
670 if ( os::is_MP() ) __ lock();
671 __ xaddptr(Address(c_rarg1, 0), c_rarg0);
672 __ addptr(rax, c_rarg0);
673 __ ret(0);
674
675 return start;
676 }
677
678 // Support for intptr_t OrderAccess::fence()
679 //
680 // Arguments :
681 //
682 // Result:
683 address generate_orderaccess_fence() {
684 StubCodeMark mark(this, "StubRoutines", "orderaccess_fence");
685 address start = __ pc();
686 __ membar(Assembler::StoreLoad);
687 __ ret(0);
688
689 return start;
690 }
691
692 // Support for intptr_t get_previous_fp()
693 //
694 // This routine is used to find the previous frame pointer for the
695 // caller (current_frame_guess). This is used as part of debugging
696 // ps() is seemingly lost trying to find frames.
697 // This code assumes that caller current_frame_guess) has a frame.
698 address generate_get_previous_fp() {
699 StubCodeMark mark(this, "StubRoutines", "get_previous_fp");
700 const Address old_fp(rbp, 0);
701 const Address older_fp(rax, 0);
702 address start = __ pc();
703
704 __ enter();
705 __ movptr(rax, old_fp); // callers fp
706 __ movptr(rax, older_fp); // the frame for ps()
707 __ pop(rbp);
708 __ ret(0);
709
710 return start;
711 }
712
713 // Support for intptr_t get_previous_sp()
714 //
715 // This routine is used to find the previous stack pointer for the
716 // caller.
717 address generate_get_previous_sp() {
718 StubCodeMark mark(this, "StubRoutines", "get_previous_sp");
719 address start = __ pc();
720
721 __ movptr(rax, rsp);
722 __ addptr(rax, 8); // return address is at the top of the stack.
723 __ ret(0);
724
725 return start;
726 }
727
728 //----------------------------------------------------------------------------------------------------
729 // Support for void verify_mxcsr()
730 //
731 // This routine is used with -Xcheck:jni to verify that native
732 // JNI code does not return to Java code without restoring the
733 // MXCSR register to our expected state.
734
735 address generate_verify_mxcsr() {
736 StubCodeMark mark(this, "StubRoutines", "verify_mxcsr");
737 address start = __ pc();
738
739 const Address mxcsr_save(rsp, 0);
740
741 if (CheckJNICalls) {
742 Label ok_ret;
743 __ push(rax);
744 __ subptr(rsp, wordSize); // allocate a temp location
745 __ stmxcsr(mxcsr_save);
746 __ movl(rax, mxcsr_save);
747 __ andl(rax, MXCSR_MASK); // Only check control and mask bits
748 __ cmpl(rax, *(int *)(StubRoutines::x86::mxcsr_std()));
749 __ jcc(Assembler::equal, ok_ret);
750
751 __ warn("MXCSR changed by native JNI code, use -XX:+RestoreMXCSROnJNICall");
752
753 __ ldmxcsr(ExternalAddress(StubRoutines::x86::mxcsr_std()));
754
755 __ bind(ok_ret);
756 __ addptr(rsp, wordSize);
757 __ pop(rax);
758 }
759
760 __ ret(0);
761
762 return start;
763 }
764
765 address generate_f2i_fixup() {
766 StubCodeMark mark(this, "StubRoutines", "f2i_fixup");
767 Address inout(rsp, 5 * wordSize); // return address + 4 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
778 __ movl(rax, 0x7f800000);
779 __ xorl(c_rarg3, c_rarg3);
780 __ movl(c_rarg2, inout);
781 __ movl(c_rarg1, c_rarg2);
782 __ andl(c_rarg1, 0x7fffffff);
783 __ cmpl(rax, c_rarg1); // NaN? -> 0
784 __ jcc(Assembler::negative, L);
785 __ testl(c_rarg2, c_rarg2); // signed ? min_jint : max_jint
786 __ movl(c_rarg3, 0x80000000);
787 __ movl(rax, 0x7fffffff);
788 __ cmovl(Assembler::positive, c_rarg3, rax);
789
790 __ bind(L);
791 __ movptr(inout, c_rarg3);
792
793 __ pop(c_rarg1);
794 __ pop(c_rarg2);
795 __ pop(c_rarg3);
796 __ pop(rax);
797
798 __ ret(0);
799
800 return start;
801 }
802
803 address generate_f2l_fixup() {
804 StubCodeMark mark(this, "StubRoutines", "f2l_fixup");
805 Address inout(rsp, 5 * wordSize); // return address + 4 saves
806 address start = __ pc();
807
808 Label L;
809
810 __ push(rax);
811 __ push(c_rarg3);
812 __ push(c_rarg2);
813 __ push(c_rarg1);
814
815 __ movl(rax, 0x7f800000);
816 __ xorl(c_rarg3, c_rarg3);
817 __ movl(c_rarg2, inout);
818 __ movl(c_rarg1, c_rarg2);
819 __ andl(c_rarg1, 0x7fffffff);
820 __ cmpl(rax, c_rarg1); // NaN? -> 0
821 __ jcc(Assembler::negative, L);
822 __ testl(c_rarg2, c_rarg2); // signed ? min_jlong : max_jlong
823 __ mov64(c_rarg3, 0x8000000000000000);
824 __ mov64(rax, 0x7fffffffffffffff);
825 __ cmov(Assembler::positive, c_rarg3, rax);
826
827 __ bind(L);
828 __ movptr(inout, c_rarg3);
829
830 __ pop(c_rarg1);
831 __ pop(c_rarg2);
832 __ pop(c_rarg3);
833 __ pop(rax);
834
835 __ ret(0);
836
837 return start;
838 }
839
840 address generate_d2i_fixup() {
841 StubCodeMark mark(this, "StubRoutines", "d2i_fixup");
842 Address inout(rsp, 6 * wordSize); // return address + 5 saves
843
844 address start = __ pc();
845
846 Label L;
847
848 __ push(rax);
849 __ push(c_rarg3);
850 __ push(c_rarg2);
851 __ push(c_rarg1);
852 __ push(c_rarg0);
853
854 __ movl(rax, 0x7ff00000);
855 __ movq(c_rarg2, inout);
856 __ movl(c_rarg3, c_rarg2);
857 __ mov(c_rarg1, c_rarg2);
858 __ mov(c_rarg0, c_rarg2);
859 __ negl(c_rarg3);
860 __ shrptr(c_rarg1, 0x20);
861 __ orl(c_rarg3, c_rarg2);
862 __ andl(c_rarg1, 0x7fffffff);
863 __ xorl(c_rarg2, c_rarg2);
864 __ shrl(c_rarg3, 0x1f);
865 __ orl(c_rarg1, c_rarg3);
866 __ cmpl(rax, c_rarg1);
867 __ jcc(Assembler::negative, L); // NaN -> 0
868 __ testptr(c_rarg0, c_rarg0); // signed ? min_jint : max_jint
869 __ movl(c_rarg2, 0x80000000);
870 __ movl(rax, 0x7fffffff);
871 __ cmov(Assembler::positive, c_rarg2, rax);
872
873 __ bind(L);
874 __ movptr(inout, c_rarg2);
875
876 __ pop(c_rarg0);
877 __ pop(c_rarg1);
878 __ pop(c_rarg2);
879 __ pop(c_rarg3);
880 __ pop(rax);
881
882 __ ret(0);
883
884 return start;
885 }
886
887 address generate_d2l_fixup() {
888 StubCodeMark mark(this, "StubRoutines", "d2l_fixup");
889 Address inout(rsp, 6 * wordSize); // return address + 5 saves
890
891 address start = __ pc();
892
893 Label L;
894
895 __ push(rax);
896 __ push(c_rarg3);
897 __ push(c_rarg2);
898 __ push(c_rarg1);
899 __ push(c_rarg0);
900
901 __ movl(rax, 0x7ff00000);
902 __ movq(c_rarg2, inout);
903 __ movl(c_rarg3, c_rarg2);
904 __ mov(c_rarg1, c_rarg2);
905 __ mov(c_rarg0, c_rarg2);
906 __ negl(c_rarg3);
907 __ shrptr(c_rarg1, 0x20);
908 __ orl(c_rarg3, c_rarg2);
909 __ andl(c_rarg1, 0x7fffffff);
910 __ xorl(c_rarg2, c_rarg2);
911 __ shrl(c_rarg3, 0x1f);
912 __ orl(c_rarg1, c_rarg3);
913 __ cmpl(rax, c_rarg1);
914 __ jcc(Assembler::negative, L); // NaN -> 0
915 __ testq(c_rarg0, c_rarg0); // signed ? min_jlong : max_jlong
916 __ mov64(c_rarg2, 0x8000000000000000);
917 __ mov64(rax, 0x7fffffffffffffff);
918 __ cmovq(Assembler::positive, c_rarg2, rax);
919
920 __ bind(L);
921 __ movq(inout, c_rarg2);
922
923 __ pop(c_rarg0);
924 __ pop(c_rarg1);
925 __ pop(c_rarg2);
926 __ pop(c_rarg3);
927 __ pop(rax);
928
929 __ ret(0);
930
931 return start;
932 }
933
934 address generate_fp_mask(const char *stub_name, int64_t mask) {
935 __ align(CodeEntryAlignment);
936 StubCodeMark mark(this, "StubRoutines", stub_name);
937 address start = __ pc();
938
939 __ emit_data64( mask, relocInfo::none );
940 __ emit_data64( mask, relocInfo::none );
941
942 return start;
943 }
944
945 // The following routine generates a subroutine to throw an
946 // asynchronous UnknownError when an unsafe access gets a fault that
947 // could not be reasonably prevented by the programmer. (Example:
948 // SIGBUS/OBJERR.)
949 address generate_handler_for_unsafe_access() {
950 StubCodeMark mark(this, "StubRoutines", "handler_for_unsafe_access");
951 address start = __ pc();
952
953 __ push(0); // hole for return address-to-be
954 __ pusha(); // push registers
955 Address next_pc(rsp, RegisterImpl::number_of_registers * BytesPerWord);
956
957 // FIXME: this probably needs alignment logic
958
959 __ subptr(rsp, frame::arg_reg_save_area_bytes);
960 BLOCK_COMMENT("call handle_unsafe_access");
961 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, handle_unsafe_access)));
962 __ addptr(rsp, frame::arg_reg_save_area_bytes);
963
964 __ movptr(next_pc, rax); // stuff next address
965 __ popa();
966 __ ret(0); // jump to next address
967
968 return start;
969 }
970
971 // Non-destructive plausibility checks for oops
972 //
973 // Arguments:
974 // all args on stack!
975 //
976 // Stack after saving c_rarg3:
977 // [tos + 0]: saved c_rarg3
978 // [tos + 1]: saved c_rarg2
979 // [tos + 2]: saved r12 (several TemplateTable methods use it)
980 // [tos + 3]: saved flags
981 // [tos + 4]: return address
982 // * [tos + 5]: error message (char*)
983 // * [tos + 6]: object to verify (oop)
984 // * [tos + 7]: saved rax - saved by caller and bashed
985 // * [tos + 8]: saved r10 (rscratch1) - saved by caller
986 // * = popped on exit
987 address generate_verify_oop() {
988 StubCodeMark mark(this, "StubRoutines", "verify_oop");
989 address start = __ pc();
990
991 Label exit, error;
992
993 __ pushf();
994 __ incrementl(ExternalAddress((address) StubRoutines::verify_oop_count_addr()));
995
996 __ push(r12);
997
998 // save c_rarg2 and c_rarg3
999 __ push(c_rarg2);
1000 __ push(c_rarg3);
1001
1002 enum {
1003 // After previous pushes.
1004 oop_to_verify = 6 * wordSize,
1005 saved_rax = 7 * wordSize,
1006 saved_r10 = 8 * wordSize,
1007
1008 // Before the call to MacroAssembler::debug(), see below.
1009 return_addr = 16 * wordSize,
1010 error_msg = 17 * wordSize
1011 };
1012
1013 // get object
1014 __ movptr(rax, Address(rsp, oop_to_verify));
1015
1016 // make sure object is 'reasonable'
1017 __ testptr(rax, rax);
1018 __ jcc(Assembler::zero, exit); // if obj is NULL it is OK
1019 // Check if the oop is in the right area of memory
1020 __ movptr(c_rarg2, rax);
1021 __ movptr(c_rarg3, (intptr_t) Universe::verify_oop_mask());
1022 __ andptr(c_rarg2, c_rarg3);
1023 __ movptr(c_rarg3, (intptr_t) Universe::verify_oop_bits());
1024 __ cmpptr(c_rarg2, c_rarg3);
1025 __ jcc(Assembler::notZero, error);
1026
1027 // set r12 to heapbase for load_klass()
1028 __ reinit_heapbase();
1029
1030 // make sure klass is 'reasonable'
1031 __ load_klass(rax, rax); // get klass
1032 __ testptr(rax, rax);
1033 __ jcc(Assembler::zero, error); // if klass is NULL it is broken
1034 // Check if the klass is in the right area of memory
1035 __ mov(c_rarg2, rax);
1036 __ movptr(c_rarg3, (intptr_t) Universe::verify_klass_mask());
1037 __ andptr(c_rarg2, c_rarg3);
1038 __ movptr(c_rarg3, (intptr_t) Universe::verify_klass_bits());
1039 __ cmpptr(c_rarg2, c_rarg3);
1040 __ jcc(Assembler::notZero, error);
1041
1042 // make sure klass' klass is 'reasonable'
1043 __ load_klass(rax, rax);
1044 __ testptr(rax, rax);
1045 __ jcc(Assembler::zero, error); // if klass' klass is NULL it is broken
1046 // Check if the klass' klass is in the right area of memory
1047 __ movptr(c_rarg3, (intptr_t) Universe::verify_klass_mask());
1048 __ andptr(rax, c_rarg3);
1049 __ movptr(c_rarg3, (intptr_t) Universe::verify_klass_bits());
1050 __ cmpptr(rax, c_rarg3);
1051 __ jcc(Assembler::notZero, error);
1052
1053 // return if everything seems ok
1054 __ bind(exit);
1055 __ movptr(rax, Address(rsp, saved_rax)); // get saved rax back
1056 __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back
1057 __ pop(c_rarg3); // restore c_rarg3
1058 __ pop(c_rarg2); // restore c_rarg2
1059 __ pop(r12); // restore r12
1060 __ popf(); // restore flags
1061 __ ret(4 * wordSize); // pop caller saved stuff
1062
1063 // handle errors
1064 __ bind(error);
1065 __ movptr(rax, Address(rsp, saved_rax)); // get saved rax back
1066 __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back
1067 __ pop(c_rarg3); // get saved c_rarg3 back
1068 __ pop(c_rarg2); // get saved c_rarg2 back
1069 __ pop(r12); // get saved r12 back
1070 __ popf(); // get saved flags off stack --
1071 // will be ignored
1072
1073 __ pusha(); // push registers
1074 // (rip is already
1075 // already pushed)
1076 // debug(char* msg, int64_t pc, int64_t regs[])
1077 // We've popped the registers we'd saved (c_rarg3, c_rarg2 and flags), and
1078 // pushed all the registers, so now the stack looks like:
1079 // [tos + 0] 16 saved registers
1080 // [tos + 16] return address
1081 // * [tos + 17] error message (char*)
1082 // * [tos + 18] object to verify (oop)
1083 // * [tos + 19] saved rax - saved by caller and bashed
1084 // * [tos + 20] saved r10 (rscratch1) - saved by caller
1085 // * = popped on exit
1086
1087 __ movptr(c_rarg0, Address(rsp, error_msg)); // pass address of error message
1088 __ movptr(c_rarg1, Address(rsp, return_addr)); // pass return address
1089 __ movq(c_rarg2, rsp); // pass address of regs on stack
1090 __ mov(r12, rsp); // remember rsp
1091 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows
1092 __ andptr(rsp, -16); // align stack as required by ABI
1093 BLOCK_COMMENT("call MacroAssembler::debug");
1094 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64)));
1095 __ mov(rsp, r12); // restore rsp
1096 __ popa(); // pop registers (includes r12)
1097 __ ret(4 * wordSize); // pop caller saved stuff
1098
1099 return start;
1100 }
1101
1102 //
1103 // Verify that a register contains clean 32-bits positive value
1104 // (high 32-bits are 0) so it could be used in 64-bits shifts.
1105 //
1106 // Input:
1107 // Rint - 32-bits value
1108 // Rtmp - scratch
1109 //
1110 void assert_clean_int(Register Rint, Register Rtmp) {
1111 #ifdef ASSERT
1112 Label L;
1113 assert_different_registers(Rtmp, Rint);
1114 __ movslq(Rtmp, Rint);
1115 __ cmpq(Rtmp, Rint);
1116 __ jcc(Assembler::equal, L);
1117 __ stop("high 32-bits of int value are not 0");
1118 __ bind(L);
1119 #endif
1120 }
1121
1122 // Generate overlap test for array copy stubs
1123 //
1124 // Input:
1125 // c_rarg0 - from
1126 // c_rarg1 - to
1127 // c_rarg2 - element count
1128 //
1129 // Output:
1130 // rax - &from[element count - 1]
1131 //
1132 void array_overlap_test(address no_overlap_target, Address::ScaleFactor sf) {
1133 assert(no_overlap_target != NULL, "must be generated");
1134 array_overlap_test(no_overlap_target, NULL, sf);
1135 }
1136 void array_overlap_test(Label& L_no_overlap, Address::ScaleFactor sf) {
1137 array_overlap_test(NULL, &L_no_overlap, sf);
1138 }
1139 void array_overlap_test(address no_overlap_target, Label* NOLp, Address::ScaleFactor sf) {
1140 const Register from = c_rarg0;
1141 const Register to = c_rarg1;
1142 const Register count = c_rarg2;
1143 const Register end_from = rax;
1144
1145 __ cmpptr(to, from);
1146 __ lea(end_from, Address(from, count, sf, 0));
1147 if (NOLp == NULL) {
1148 ExternalAddress no_overlap(no_overlap_target);
1149 __ jump_cc(Assembler::belowEqual, no_overlap);
1150 __ cmpptr(to, end_from);
1151 __ jump_cc(Assembler::aboveEqual, no_overlap);
1152 } else {
1153 __ jcc(Assembler::belowEqual, (*NOLp));
1154 __ cmpptr(to, end_from);
1155 __ jcc(Assembler::aboveEqual, (*NOLp));
1156 }
1157 }
1158
1159 // Shuffle first three arg regs on Windows into Linux/Solaris locations.
1160 //
1161 // Outputs:
1162 // rdi - rcx
1163 // rsi - rdx
1164 // rdx - r8
1165 // rcx - r9
1166 //
1167 // Registers r9 and r10 are used to save rdi and rsi on Windows, which latter
1168 // are non-volatile. r9 and r10 should not be used by the caller.
1169 //
1170 void setup_arg_regs(int nargs = 3) {
1171 const Register saved_rdi = r9;
1172 const Register saved_rsi = r10;
1173 assert(nargs == 3 || nargs == 4, "else fix");
1174 #ifdef _WIN64
1175 assert(c_rarg0 == rcx && c_rarg1 == rdx && c_rarg2 == r8 && c_rarg3 == r9,
1176 "unexpected argument registers");
1177 if (nargs >= 4)
1178 __ mov(rax, r9); // r9 is also saved_rdi
1179 __ movptr(saved_rdi, rdi);
1180 __ movptr(saved_rsi, rsi);
1181 __ mov(rdi, rcx); // c_rarg0
1182 __ mov(rsi, rdx); // c_rarg1
1183 __ mov(rdx, r8); // c_rarg2
1184 if (nargs >= 4)
1185 __ mov(rcx, rax); // c_rarg3 (via rax)
1186 #else
1187 assert(c_rarg0 == rdi && c_rarg1 == rsi && c_rarg2 == rdx && c_rarg3 == rcx,
1188 "unexpected argument registers");
1189 #endif
1190 }
1191
1192 void restore_arg_regs() {
1193 const Register saved_rdi = r9;
1194 const Register saved_rsi = r10;
1195 #ifdef _WIN64
1196 __ movptr(rdi, saved_rdi);
1197 __ movptr(rsi, saved_rsi);
1198 #endif
1199 }
1200
1201 // Generate code for an array write pre barrier
1202 //
1203 // addr - starting address
1204 // count - element count
1205 // tmp - scratch register
1206 //
1207 // Destroy no registers!
1208 //
1209 void gen_write_ref_array_pre_barrier(Register addr, Register count, bool dest_uninitialized) {
1210 BarrierSet* bs = Universe::heap()->barrier_set();
1211 switch (bs->kind()) {
1212 case BarrierSet::G1SATBCT:
1213 case BarrierSet::G1SATBCTLogging:
1214 // With G1, don't generate the call if we statically know that the target in uninitialized
1215 if (!dest_uninitialized) {
1216 __ pusha(); // push registers
1217 if (count == c_rarg0) {
1218 if (addr == c_rarg1) {
1219 // exactly backwards!!
1220 __ xchgptr(c_rarg1, c_rarg0);
1221 } else {
1222 __ movptr(c_rarg1, count);
1223 __ movptr(c_rarg0, addr);
1224 }
1225 } else {
1226 __ movptr(c_rarg0, addr);
1227 __ movptr(c_rarg1, count);
1228 }
1229 __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_pre), 2);
1230 __ popa();
1231 }
1232 break;
1233 case BarrierSet::CardTableModRef:
1234 case BarrierSet::CardTableExtension:
1235 case BarrierSet::ModRef:
1236 break;
1237 default:
1238 ShouldNotReachHere();
1239
1240 }
1241 }
1242
1243 //
1244 // Generate code for an array write post barrier
1245 //
1246 // Input:
1247 // start - register containing starting address of destination array
1248 // end - register containing ending address of destination array
1249 // scratch - scratch register
1250 //
1251 // The input registers are overwritten.
1252 // The ending address is inclusive.
1253 void gen_write_ref_array_post_barrier(Register start, Register end, Register scratch) {
1254 assert_different_registers(start, end, scratch);
1255 BarrierSet* bs = Universe::heap()->barrier_set();
1256 switch (bs->kind()) {
1257 case BarrierSet::G1SATBCT:
1258 case BarrierSet::G1SATBCTLogging:
1259
1260 {
1261 __ pusha(); // push registers (overkill)
1262 // must compute element count unless barrier set interface is changed (other platforms supply count)
1263 assert_different_registers(start, end, scratch);
1264 __ lea(scratch, Address(end, BytesPerHeapOop));
1265 __ subptr(scratch, start); // subtract start to get #bytes
1266 __ shrptr(scratch, LogBytesPerHeapOop); // convert to element count
1267 __ mov(c_rarg0, start);
1268 __ mov(c_rarg1, scratch);
1269 __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_post), 2);
1270 __ popa();
1271 }
1272 break;
1273 case BarrierSet::CardTableModRef:
1274 case BarrierSet::CardTableExtension:
1275 {
1276 CardTableModRefBS* ct = (CardTableModRefBS*)bs;
1277 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
1278
1279 Label L_loop;
1280
1281 __ shrptr(start, CardTableModRefBS::card_shift);
1282 __ addptr(end, BytesPerHeapOop);
1283 __ shrptr(end, CardTableModRefBS::card_shift);
1284 __ subptr(end, start); // number of bytes to copy
1285
1286 intptr_t disp = (intptr_t) ct->byte_map_base;
1287 if (Assembler::is_simm32(disp)) {
1288 Address cardtable(noreg, noreg, Address::no_scale, disp);
1289 __ lea(scratch, cardtable);
1290 } else {
1291 ExternalAddress cardtable((address)disp);
1292 __ lea(scratch, cardtable);
1293 }
1294
1295 const Register count = end; // 'end' register contains bytes count now
1296 __ addptr(start, scratch);
1297 __ BIND(L_loop);
1298 __ movb(Address(start, count, Address::times_1), 0);
1299 __ decrement(count);
1300 __ jcc(Assembler::greaterEqual, L_loop);
1301 }
1302 break;
1303 default:
1304 ShouldNotReachHere();
1305
1306 }
1307 }
1308
1309
1310 // Copy big chunks forward
1311 //
1312 // Inputs:
1313 // end_from - source arrays end address
1314 // end_to - destination array end address
1315 // qword_count - 64-bits element count, negative
1316 // to - scratch
1317 // L_copy_bytes - entry label
1318 // L_copy_8_bytes - exit label
1319 //
1320 void copy_bytes_forward(Register end_from, Register end_to,
1321 Register qword_count, Register to,
1322 Label& L_copy_bytes, Label& L_copy_8_bytes) {
1323 DEBUG_ONLY(__ stop("enter at entry label, not here"));
1324 Label L_loop;
1325 __ align(OptoLoopAlignment);
1326 if (UseUnalignedLoadStores) {
1327 Label L_end;
1328 // Copy 64-bytes per iteration
1329 __ BIND(L_loop);
1330 if (UseAVX >= 2) {
1331 __ vmovdqu(xmm0, Address(end_from, qword_count, Address::times_8, -56));
1332 __ vmovdqu(Address(end_to, qword_count, Address::times_8, -56), xmm0);
1333 __ vmovdqu(xmm1, Address(end_from, qword_count, Address::times_8, -24));
1334 __ vmovdqu(Address(end_to, qword_count, Address::times_8, -24), xmm1);
1335 } else {
1336 __ movdqu(xmm0, Address(end_from, qword_count, Address::times_8, -56));
1337 __ movdqu(Address(end_to, qword_count, Address::times_8, -56), xmm0);
1338 __ movdqu(xmm1, Address(end_from, qword_count, Address::times_8, -40));
1339 __ movdqu(Address(end_to, qword_count, Address::times_8, -40), xmm1);
1340 __ movdqu(xmm2, Address(end_from, qword_count, Address::times_8, -24));
1341 __ movdqu(Address(end_to, qword_count, Address::times_8, -24), xmm2);
1342 __ movdqu(xmm3, Address(end_from, qword_count, Address::times_8, - 8));
1343 __ movdqu(Address(end_to, qword_count, Address::times_8, - 8), xmm3);
1344 }
1345 __ BIND(L_copy_bytes);
1346 __ addptr(qword_count, 8);
1347 __ jcc(Assembler::lessEqual, L_loop);
1348 __ subptr(qword_count, 4); // sub(8) and add(4)
1349 __ jccb(Assembler::greater, L_end);
1350 // Copy trailing 32 bytes
1351 if (UseAVX >= 2) {
1352 __ vmovdqu(xmm0, Address(end_from, qword_count, Address::times_8, -24));
1353 __ vmovdqu(Address(end_to, qword_count, Address::times_8, -24), xmm0);
1354 } else {
1355 __ movdqu(xmm0, Address(end_from, qword_count, Address::times_8, -24));
1356 __ movdqu(Address(end_to, qword_count, Address::times_8, -24), xmm0);
1357 __ movdqu(xmm1, Address(end_from, qword_count, Address::times_8, - 8));
1358 __ movdqu(Address(end_to, qword_count, Address::times_8, - 8), xmm1);
1359 }
1360 __ addptr(qword_count, 4);
1361 __ BIND(L_end);
1362 if (UseAVX >= 2) {
1363 // clean upper bits of YMM registers
1364 __ vzeroupper();
1365 }
1366 } else {
1367 // Copy 32-bytes per iteration
1368 __ BIND(L_loop);
1369 __ movq(to, Address(end_from, qword_count, Address::times_8, -24));
1370 __ movq(Address(end_to, qword_count, Address::times_8, -24), to);
1371 __ movq(to, Address(end_from, qword_count, Address::times_8, -16));
1372 __ movq(Address(end_to, qword_count, Address::times_8, -16), to);
1373 __ movq(to, Address(end_from, qword_count, Address::times_8, - 8));
1374 __ movq(Address(end_to, qword_count, Address::times_8, - 8), to);
1375 __ movq(to, Address(end_from, qword_count, Address::times_8, - 0));
1376 __ movq(Address(end_to, qword_count, Address::times_8, - 0), to);
1377
1378 __ BIND(L_copy_bytes);
1379 __ addptr(qword_count, 4);
1380 __ jcc(Assembler::lessEqual, L_loop);
1381 }
1382 __ subptr(qword_count, 4);
1383 __ jcc(Assembler::less, L_copy_8_bytes); // Copy trailing qwords
1384 }
1385
1386 // Copy big chunks backward
1387 //
1388 // Inputs:
1389 // from - source arrays address
1390 // dest - destination array address
1391 // qword_count - 64-bits element count
1392 // to - scratch
1393 // L_copy_bytes - entry label
1394 // L_copy_8_bytes - exit label
1395 //
1396 void copy_bytes_backward(Register from, Register dest,
1397 Register qword_count, Register to,
1398 Label& L_copy_bytes, Label& L_copy_8_bytes) {
1399 DEBUG_ONLY(__ stop("enter at entry label, not here"));
1400 Label L_loop;
1401 __ align(OptoLoopAlignment);
1402 if (UseUnalignedLoadStores) {
1403 Label L_end;
1404 // Copy 64-bytes per iteration
1405 __ BIND(L_loop);
1406 if (UseAVX >= 2) {
1407 __ vmovdqu(xmm0, Address(from, qword_count, Address::times_8, 32));
1408 __ vmovdqu(Address(dest, qword_count, Address::times_8, 32), xmm0);
1409 __ vmovdqu(xmm1, Address(from, qword_count, Address::times_8, 0));
1410 __ vmovdqu(Address(dest, qword_count, Address::times_8, 0), xmm1);
1411 } else {
1412 __ movdqu(xmm0, Address(from, qword_count, Address::times_8, 48));
1413 __ movdqu(Address(dest, qword_count, Address::times_8, 48), xmm0);
1414 __ movdqu(xmm1, Address(from, qword_count, Address::times_8, 32));
1415 __ movdqu(Address(dest, qword_count, Address::times_8, 32), xmm1);
1416 __ movdqu(xmm2, Address(from, qword_count, Address::times_8, 16));
1417 __ movdqu(Address(dest, qword_count, Address::times_8, 16), xmm2);
1418 __ movdqu(xmm3, Address(from, qword_count, Address::times_8, 0));
1419 __ movdqu(Address(dest, qword_count, Address::times_8, 0), xmm3);
1420 }
1421 __ BIND(L_copy_bytes);
1422 __ subptr(qword_count, 8);
1423 __ jcc(Assembler::greaterEqual, L_loop);
1424
1425 __ addptr(qword_count, 4); // add(8) and sub(4)
1426 __ jccb(Assembler::less, L_end);
1427 // Copy trailing 32 bytes
1428 if (UseAVX >= 2) {
1429 __ vmovdqu(xmm0, Address(from, qword_count, Address::times_8, 0));
1430 __ vmovdqu(Address(dest, qword_count, Address::times_8, 0), xmm0);
1431 } else {
1432 __ movdqu(xmm0, Address(from, qword_count, Address::times_8, 16));
1433 __ movdqu(Address(dest, qword_count, Address::times_8, 16), xmm0);
1434 __ movdqu(xmm1, Address(from, qword_count, Address::times_8, 0));
1435 __ movdqu(Address(dest, qword_count, Address::times_8, 0), xmm1);
1436 }
1437 __ subptr(qword_count, 4);
1438 __ BIND(L_end);
1439 if (UseAVX >= 2) {
1440 // clean upper bits of YMM registers
1441 __ vzeroupper();
1442 }
1443 } else {
1444 // Copy 32-bytes per iteration
1445 __ BIND(L_loop);
1446 __ movq(to, Address(from, qword_count, Address::times_8, 24));
1447 __ movq(Address(dest, qword_count, Address::times_8, 24), to);
1448 __ movq(to, Address(from, qword_count, Address::times_8, 16));
1449 __ movq(Address(dest, qword_count, Address::times_8, 16), to);
1450 __ movq(to, Address(from, qword_count, Address::times_8, 8));
1451 __ movq(Address(dest, qword_count, Address::times_8, 8), to);
1452 __ movq(to, Address(from, qword_count, Address::times_8, 0));
1453 __ movq(Address(dest, qword_count, Address::times_8, 0), to);
1454
1455 __ BIND(L_copy_bytes);
1456 __ subptr(qword_count, 4);
1457 __ jcc(Assembler::greaterEqual, L_loop);
1458 }
1459 __ addptr(qword_count, 4);
1460 __ jcc(Assembler::greater, L_copy_8_bytes); // Copy trailing qwords
1461 }
1462
1463
1464 // Arguments:
1465 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1466 // ignored
1467 // name - stub name string
1468 //
1469 // Inputs:
1470 // c_rarg0 - source array address
1471 // c_rarg1 - destination array address
1472 // c_rarg2 - element count, treated as ssize_t, can be zero
1473 //
1474 // If 'from' and/or 'to' are aligned on 4-, 2-, or 1-byte boundaries,
1475 // we let the hardware handle it. The one to eight bytes within words,
1476 // dwords or qwords that span cache line boundaries will still be loaded
1477 // and stored atomically.
1478 //
1479 // Side Effects:
1480 // disjoint_byte_copy_entry is set to the no-overlap entry point
1481 // used by generate_conjoint_byte_copy().
1482 //
1483 address generate_disjoint_byte_copy(bool aligned, address* entry, const char *name) {
1484 __ align(CodeEntryAlignment);
1485 StubCodeMark mark(this, "StubRoutines", name);
1486 address start = __ pc();
1487
1488 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes, L_copy_2_bytes;
1489 Label L_copy_byte, L_exit;
1490 const Register from = rdi; // source array address
1491 const Register to = rsi; // destination array address
1492 const Register count = rdx; // elements count
1493 const Register byte_count = rcx;
1494 const Register qword_count = count;
1495 const Register end_from = from; // source array end address
1496 const Register end_to = to; // destination array end address
1497 // End pointers are inclusive, and if count is not zero they point
1498 // to the last unit copied: end_to[0] := end_from[0]
1499
1500 __ enter(); // required for proper stackwalking of RuntimeStub frame
1501 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1502
1503 if (entry != NULL) {
1504 *entry = __ pc();
1505 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1506 BLOCK_COMMENT("Entry:");
1507 }
1508
1509 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1510 // r9 and r10 may be used to save non-volatile registers
1511
1512 // 'from', 'to' and 'count' are now valid
1513 __ movptr(byte_count, count);
1514 __ shrptr(count, 3); // count => qword_count
1515
1516 // Copy from low to high addresses. Use 'to' as scratch.
1517 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
1518 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
1519 __ negptr(qword_count); // make the count negative
1520 __ jmp(L_copy_bytes);
1521
1522 // Copy trailing qwords
1523 __ BIND(L_copy_8_bytes);
1524 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
1525 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
1526 __ increment(qword_count);
1527 __ jcc(Assembler::notZero, L_copy_8_bytes);
1528
1529 // Check for and copy trailing dword
1530 __ BIND(L_copy_4_bytes);
1531 __ testl(byte_count, 4);
1532 __ jccb(Assembler::zero, L_copy_2_bytes);
1533 __ movl(rax, Address(end_from, 8));
1534 __ movl(Address(end_to, 8), rax);
1535
1536 __ addptr(end_from, 4);
1537 __ addptr(end_to, 4);
1538
1539 // Check for and copy trailing word
1540 __ BIND(L_copy_2_bytes);
1541 __ testl(byte_count, 2);
1542 __ jccb(Assembler::zero, L_copy_byte);
1543 __ movw(rax, Address(end_from, 8));
1544 __ movw(Address(end_to, 8), rax);
1545
1546 __ addptr(end_from, 2);
1547 __ addptr(end_to, 2);
1548
1549 // Check for and copy trailing byte
1550 __ BIND(L_copy_byte);
1551 __ testl(byte_count, 1);
1552 __ jccb(Assembler::zero, L_exit);
1553 __ movb(rax, Address(end_from, 8));
1554 __ movb(Address(end_to, 8), rax);
1555
1556 __ BIND(L_exit);
1557 restore_arg_regs();
1558 inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); // Update counter after rscratch1 is free
1559 __ xorptr(rax, rax); // return 0
1560 __ leave(); // required for proper stackwalking of RuntimeStub frame
1561 __ ret(0);
1562
1563 // Copy in multi-bytes chunks
1564 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1565 __ jmp(L_copy_4_bytes);
1566
1567 return start;
1568 }
1569
1570 // Arguments:
1571 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1572 // ignored
1573 // name - stub name string
1574 //
1575 // Inputs:
1576 // c_rarg0 - source array address
1577 // c_rarg1 - destination array address
1578 // c_rarg2 - element count, treated as ssize_t, can be zero
1579 //
1580 // If 'from' and/or 'to' are aligned on 4-, 2-, or 1-byte boundaries,
1581 // we let the hardware handle it. The one to eight bytes within words,
1582 // dwords or qwords that span cache line boundaries will still be loaded
1583 // and stored atomically.
1584 //
1585 address generate_conjoint_byte_copy(bool aligned, address nooverlap_target,
1586 address* entry, const char *name) {
1587 __ align(CodeEntryAlignment);
1588 StubCodeMark mark(this, "StubRoutines", name);
1589 address start = __ pc();
1590
1591 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes, L_copy_2_bytes;
1592 const Register from = rdi; // source array address
1593 const Register to = rsi; // destination array address
1594 const Register count = rdx; // elements count
1595 const Register byte_count = rcx;
1596 const Register qword_count = count;
1597
1598 __ enter(); // required for proper stackwalking of RuntimeStub frame
1599 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1600
1601 if (entry != NULL) {
1602 *entry = __ pc();
1603 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1604 BLOCK_COMMENT("Entry:");
1605 }
1606
1607 array_overlap_test(nooverlap_target, Address::times_1);
1608 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1609 // r9 and r10 may be used to save non-volatile registers
1610
1611 // 'from', 'to' and 'count' are now valid
1612 __ movptr(byte_count, count);
1613 __ shrptr(count, 3); // count => qword_count
1614
1615 // Copy from high to low addresses.
1616
1617 // Check for and copy trailing byte
1618 __ testl(byte_count, 1);
1619 __ jcc(Assembler::zero, L_copy_2_bytes);
1620 __ movb(rax, Address(from, byte_count, Address::times_1, -1));
1621 __ movb(Address(to, byte_count, Address::times_1, -1), rax);
1622 __ decrement(byte_count); // Adjust for possible trailing word
1623
1624 // Check for and copy trailing word
1625 __ BIND(L_copy_2_bytes);
1626 __ testl(byte_count, 2);
1627 __ jcc(Assembler::zero, L_copy_4_bytes);
1628 __ movw(rax, Address(from, byte_count, Address::times_1, -2));
1629 __ movw(Address(to, byte_count, Address::times_1, -2), rax);
1630
1631 // Check for and copy trailing dword
1632 __ BIND(L_copy_4_bytes);
1633 __ testl(byte_count, 4);
1634 __ jcc(Assembler::zero, L_copy_bytes);
1635 __ movl(rax, Address(from, qword_count, Address::times_8));
1636 __ movl(Address(to, qword_count, Address::times_8), rax);
1637 __ jmp(L_copy_bytes);
1638
1639 // Copy trailing qwords
1640 __ BIND(L_copy_8_bytes);
1641 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
1642 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
1643 __ decrement(qword_count);
1644 __ jcc(Assembler::notZero, L_copy_8_bytes);
1645
1646 restore_arg_regs();
1647 inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); // Update counter after rscratch1 is free
1648 __ xorptr(rax, rax); // return 0
1649 __ leave(); // required for proper stackwalking of RuntimeStub frame
1650 __ ret(0);
1651
1652 // Copy in multi-bytes chunks
1653 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1654
1655 restore_arg_regs();
1656 inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); // Update counter after rscratch1 is free
1657 __ xorptr(rax, rax); // return 0
1658 __ leave(); // required for proper stackwalking of RuntimeStub frame
1659 __ ret(0);
1660
1661 return start;
1662 }
1663
1664 // Arguments:
1665 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1666 // ignored
1667 // name - stub name string
1668 //
1669 // Inputs:
1670 // c_rarg0 - source array address
1671 // c_rarg1 - destination array address
1672 // c_rarg2 - element count, treated as ssize_t, can be zero
1673 //
1674 // If 'from' and/or 'to' are aligned on 4- or 2-byte boundaries, we
1675 // let the hardware handle it. The two or four words within dwords
1676 // or qwords that span cache line boundaries will still be loaded
1677 // and stored atomically.
1678 //
1679 // Side Effects:
1680 // disjoint_short_copy_entry is set to the no-overlap entry point
1681 // used by generate_conjoint_short_copy().
1682 //
1683 address generate_disjoint_short_copy(bool aligned, address *entry, const char *name) {
1684 __ align(CodeEntryAlignment);
1685 StubCodeMark mark(this, "StubRoutines", name);
1686 address start = __ pc();
1687
1688 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes,L_copy_2_bytes,L_exit;
1689 const Register from = rdi; // source array address
1690 const Register to = rsi; // destination array address
1691 const Register count = rdx; // elements count
1692 const Register word_count = rcx;
1693 const Register qword_count = count;
1694 const Register end_from = from; // source array end address
1695 const Register end_to = to; // destination array end address
1696 // End pointers are inclusive, and if count is not zero they point
1697 // to the last unit copied: end_to[0] := end_from[0]
1698
1699 __ enter(); // required for proper stackwalking of RuntimeStub frame
1700 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1701
1702 if (entry != NULL) {
1703 *entry = __ pc();
1704 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1705 BLOCK_COMMENT("Entry:");
1706 }
1707
1708 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1709 // r9 and r10 may be used to save non-volatile registers
1710
1711 // 'from', 'to' and 'count' are now valid
1712 __ movptr(word_count, count);
1713 __ shrptr(count, 2); // count => qword_count
1714
1715 // Copy from low to high addresses. Use 'to' as scratch.
1716 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
1717 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
1718 __ negptr(qword_count);
1719 __ jmp(L_copy_bytes);
1720
1721 // Copy trailing qwords
1722 __ BIND(L_copy_8_bytes);
1723 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
1724 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
1725 __ increment(qword_count);
1726 __ jcc(Assembler::notZero, L_copy_8_bytes);
1727
1728 // Original 'dest' is trashed, so we can't use it as a
1729 // base register for a possible trailing word copy
1730
1731 // Check for and copy trailing dword
1732 __ BIND(L_copy_4_bytes);
1733 __ testl(word_count, 2);
1734 __ jccb(Assembler::zero, L_copy_2_bytes);
1735 __ movl(rax, Address(end_from, 8));
1736 __ movl(Address(end_to, 8), rax);
1737
1738 __ addptr(end_from, 4);
1739 __ addptr(end_to, 4);
1740
1741 // Check for and copy trailing word
1742 __ BIND(L_copy_2_bytes);
1743 __ testl(word_count, 1);
1744 __ jccb(Assembler::zero, L_exit);
1745 __ movw(rax, Address(end_from, 8));
1746 __ movw(Address(end_to, 8), rax);
1747
1748 __ BIND(L_exit);
1749 restore_arg_regs();
1750 inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); // Update counter after rscratch1 is free
1751 __ xorptr(rax, rax); // return 0
1752 __ leave(); // required for proper stackwalking of RuntimeStub frame
1753 __ ret(0);
1754
1755 // Copy in multi-bytes chunks
1756 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1757 __ jmp(L_copy_4_bytes);
1758
1759 return start;
1760 }
1761
1762 address generate_fill(BasicType t, bool aligned, const char *name) {
1763 __ align(CodeEntryAlignment);
1764 StubCodeMark mark(this, "StubRoutines", name);
1765 address start = __ pc();
1766
1767 BLOCK_COMMENT("Entry:");
1768
1769 const Register to = c_rarg0; // source array address
1770 const Register value = c_rarg1; // value
1771 const Register count = c_rarg2; // elements count
1772
1773 __ enter(); // required for proper stackwalking of RuntimeStub frame
1774
1775 __ generate_fill(t, aligned, to, value, count, rax, xmm0);
1776
1777 __ leave(); // required for proper stackwalking of RuntimeStub frame
1778 __ ret(0);
1779 return start;
1780 }
1781
1782 // Arguments:
1783 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1784 // ignored
1785 // name - stub name string
1786 //
1787 // Inputs:
1788 // c_rarg0 - source array address
1789 // c_rarg1 - destination array address
1790 // c_rarg2 - element count, treated as ssize_t, can be zero
1791 //
1792 // If 'from' and/or 'to' are aligned on 4- or 2-byte boundaries, we
1793 // let the hardware handle it. The two or four words within dwords
1794 // or qwords that span cache line boundaries will still be loaded
1795 // and stored atomically.
1796 //
1797 address generate_conjoint_short_copy(bool aligned, address nooverlap_target,
1798 address *entry, const char *name) {
1799 __ align(CodeEntryAlignment);
1800 StubCodeMark mark(this, "StubRoutines", name);
1801 address start = __ pc();
1802
1803 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes;
1804 const Register from = rdi; // source array address
1805 const Register to = rsi; // destination array address
1806 const Register count = rdx; // elements count
1807 const Register word_count = rcx;
1808 const Register qword_count = count;
1809
1810 __ enter(); // required for proper stackwalking of RuntimeStub frame
1811 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1812
1813 if (entry != NULL) {
1814 *entry = __ pc();
1815 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1816 BLOCK_COMMENT("Entry:");
1817 }
1818
1819 array_overlap_test(nooverlap_target, Address::times_2);
1820 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1821 // r9 and r10 may be used to save non-volatile registers
1822
1823 // 'from', 'to' and 'count' are now valid
1824 __ movptr(word_count, count);
1825 __ shrptr(count, 2); // count => qword_count
1826
1827 // Copy from high to low addresses. Use 'to' as scratch.
1828
1829 // Check for and copy trailing word
1830 __ testl(word_count, 1);
1831 __ jccb(Assembler::zero, L_copy_4_bytes);
1832 __ movw(rax, Address(from, word_count, Address::times_2, -2));
1833 __ movw(Address(to, word_count, Address::times_2, -2), rax);
1834
1835 // Check for and copy trailing dword
1836 __ BIND(L_copy_4_bytes);
1837 __ testl(word_count, 2);
1838 __ jcc(Assembler::zero, L_copy_bytes);
1839 __ movl(rax, Address(from, qword_count, Address::times_8));
1840 __ movl(Address(to, qword_count, Address::times_8), rax);
1841 __ jmp(L_copy_bytes);
1842
1843 // Copy trailing qwords
1844 __ BIND(L_copy_8_bytes);
1845 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
1846 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
1847 __ decrement(qword_count);
1848 __ jcc(Assembler::notZero, L_copy_8_bytes);
1849
1850 restore_arg_regs();
1851 inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); // Update counter after rscratch1 is free
1852 __ xorptr(rax, rax); // return 0
1853 __ leave(); // required for proper stackwalking of RuntimeStub frame
1854 __ ret(0);
1855
1856 // Copy in multi-bytes chunks
1857 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1858
1859 restore_arg_regs();
1860 inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); // Update counter after rscratch1 is free
1861 __ xorptr(rax, rax); // return 0
1862 __ leave(); // required for proper stackwalking of RuntimeStub frame
1863 __ ret(0);
1864
1865 return start;
1866 }
1867
1868 // Arguments:
1869 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1870 // ignored
1871 // is_oop - true => oop array, so generate store check code
1872 // name - stub name string
1873 //
1874 // Inputs:
1875 // c_rarg0 - source array address
1876 // c_rarg1 - destination array address
1877 // c_rarg2 - element count, treated as ssize_t, can be zero
1878 //
1879 // If 'from' and/or 'to' are aligned on 4-byte boundaries, we let
1880 // the hardware handle it. The two dwords within qwords that span
1881 // cache line boundaries will still be loaded and stored atomicly.
1882 //
1883 // Side Effects:
1884 // disjoint_int_copy_entry is set to the no-overlap entry point
1885 // used by generate_conjoint_int_oop_copy().
1886 //
1887 address generate_disjoint_int_oop_copy(bool aligned, bool is_oop, address* entry,
1888 const char *name, bool dest_uninitialized = false) {
1889 __ align(CodeEntryAlignment);
1890 StubCodeMark mark(this, "StubRoutines", name);
1891 address start = __ pc();
1892
1893 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes, L_exit;
1894 const Register from = rdi; // source array address
1895 const Register to = rsi; // destination array address
1896 const Register count = rdx; // elements count
1897 const Register dword_count = rcx;
1898 const Register qword_count = count;
1899 const Register end_from = from; // source array end address
1900 const Register end_to = to; // destination array end address
1901 const Register saved_to = r11; // saved destination array address
1902 // End pointers are inclusive, and if count is not zero they point
1903 // to the last unit copied: end_to[0] := end_from[0]
1904
1905 __ enter(); // required for proper stackwalking of RuntimeStub frame
1906 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1907
1908 if (entry != NULL) {
1909 *entry = __ pc();
1910 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1911 BLOCK_COMMENT("Entry:");
1912 }
1913
1914 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1915 // r9 and r10 may be used to save non-volatile registers
1916 if (is_oop) {
1917 __ movq(saved_to, to);
1918 gen_write_ref_array_pre_barrier(to, count, dest_uninitialized);
1919 }
1920
1921 // 'from', 'to' and 'count' are now valid
1922 __ movptr(dword_count, count);
1923 __ shrptr(count, 1); // count => qword_count
1924
1925 // Copy from low to high addresses. Use 'to' as scratch.
1926 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
1927 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
1928 __ negptr(qword_count);
1929 __ jmp(L_copy_bytes);
1930
1931 // Copy trailing qwords
1932 __ BIND(L_copy_8_bytes);
1933 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
1934 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
1935 __ increment(qword_count);
1936 __ jcc(Assembler::notZero, L_copy_8_bytes);
1937
1938 // Check for and copy trailing dword
1939 __ BIND(L_copy_4_bytes);
1940 __ testl(dword_count, 1); // Only byte test since the value is 0 or 1
1941 __ jccb(Assembler::zero, L_exit);
1942 __ movl(rax, Address(end_from, 8));
1943 __ movl(Address(end_to, 8), rax);
1944
1945 __ BIND(L_exit);
1946 if (is_oop) {
1947 __ leaq(end_to, Address(saved_to, dword_count, Address::times_4, -4));
1948 gen_write_ref_array_post_barrier(saved_to, end_to, rax);
1949 }
1950 restore_arg_regs();
1951 inc_counter_np(SharedRuntime::_jint_array_copy_ctr); // Update counter after rscratch1 is free
1952 __ xorptr(rax, rax); // return 0
1953 __ leave(); // required for proper stackwalking of RuntimeStub frame
1954 __ ret(0);
1955
1956 // Copy in multi-bytes chunks
1957 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1958 __ jmp(L_copy_4_bytes);
1959
1960 return start;
1961 }
1962
1963 // Arguments:
1964 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1965 // ignored
1966 // is_oop - true => oop array, so generate store check code
1967 // name - stub name string
1968 //
1969 // Inputs:
1970 // c_rarg0 - source array address
1971 // c_rarg1 - destination array address
1972 // c_rarg2 - element count, treated as ssize_t, can be zero
1973 //
1974 // If 'from' and/or 'to' are aligned on 4-byte boundaries, we let
1975 // the hardware handle it. The two dwords within qwords that span
1976 // cache line boundaries will still be loaded and stored atomicly.
1977 //
1978 address generate_conjoint_int_oop_copy(bool aligned, bool is_oop, address nooverlap_target,
1979 address *entry, const char *name,
1980 bool dest_uninitialized = false) {
1981 __ align(CodeEntryAlignment);
1982 StubCodeMark mark(this, "StubRoutines", name);
1983 address start = __ pc();
1984
1985 Label L_copy_bytes, L_copy_8_bytes, L_copy_2_bytes, L_exit;
1986 const Register from = rdi; // source array address
1987 const Register to = rsi; // destination array address
1988 const Register count = rdx; // elements count
1989 const Register dword_count = rcx;
1990 const Register qword_count = count;
1991
1992 __ enter(); // required for proper stackwalking of RuntimeStub frame
1993 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1994
1995 if (entry != NULL) {
1996 *entry = __ pc();
1997 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1998 BLOCK_COMMENT("Entry:");
1999 }
2000
2001 array_overlap_test(nooverlap_target, Address::times_4);
2002 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
2003 // r9 and r10 may be used to save non-volatile registers
2004
2005 if (is_oop) {
2006 // no registers are destroyed by this call
2007 gen_write_ref_array_pre_barrier(to, count, dest_uninitialized);
2008 }
2009
2010 assert_clean_int(count, rax); // Make sure 'count' is clean int.
2011 // 'from', 'to' and 'count' are now valid
2012 __ movptr(dword_count, count);
2013 __ shrptr(count, 1); // count => qword_count
2014
2015 // Copy from high to low addresses. Use 'to' as scratch.
2016
2017 // Check for and copy trailing dword
2018 __ testl(dword_count, 1);
2019 __ jcc(Assembler::zero, L_copy_bytes);
2020 __ movl(rax, Address(from, dword_count, Address::times_4, -4));
2021 __ movl(Address(to, dword_count, Address::times_4, -4), rax);
2022 __ jmp(L_copy_bytes);
2023
2024 // Copy trailing qwords
2025 __ BIND(L_copy_8_bytes);
2026 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
2027 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
2028 __ decrement(qword_count);
2029 __ jcc(Assembler::notZero, L_copy_8_bytes);
2030
2031 if (is_oop) {
2032 __ jmp(L_exit);
2033 }
2034 restore_arg_regs();
2035 inc_counter_np(SharedRuntime::_jint_array_copy_ctr); // Update counter after rscratch1 is free
2036 __ xorptr(rax, rax); // return 0
2037 __ leave(); // required for proper stackwalking of RuntimeStub frame
2038 __ ret(0);
2039
2040 // Copy in multi-bytes chunks
2041 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
2042
2043 __ bind(L_exit);
2044 if (is_oop) {
2045 Register end_to = rdx;
2046 __ leaq(end_to, Address(to, dword_count, Address::times_4, -4));
2047 gen_write_ref_array_post_barrier(to, end_to, rax);
2048 }
2049 restore_arg_regs();
2050 inc_counter_np(SharedRuntime::_jint_array_copy_ctr); // Update counter after rscratch1 is free
2051 __ xorptr(rax, rax); // return 0
2052 __ leave(); // required for proper stackwalking of RuntimeStub frame
2053 __ ret(0);
2054
2055 return start;
2056 }
2057
2058 // Arguments:
2059 // aligned - true => Input and output aligned on a HeapWord boundary == 8 bytes
2060 // ignored
2061 // is_oop - true => oop array, so generate store check code
2062 // name - stub name string
2063 //
2064 // Inputs:
2065 // c_rarg0 - source array address
2066 // c_rarg1 - destination array address
2067 // c_rarg2 - element count, treated as ssize_t, can be zero
2068 //
2069 // Side Effects:
2070 // disjoint_oop_copy_entry or disjoint_long_copy_entry is set to the
2071 // no-overlap entry point used by generate_conjoint_long_oop_copy().
2072 //
2073 address generate_disjoint_long_oop_copy(bool aligned, bool is_oop, address *entry,
2074 const char *name, bool dest_uninitialized = false) {
2075 __ align(CodeEntryAlignment);
2076 StubCodeMark mark(this, "StubRoutines", name);
2077 address start = __ pc();
2078
2079 Label L_copy_bytes, L_copy_8_bytes, L_exit;
2080 const Register from = rdi; // source array address
2081 const Register to = rsi; // destination array address
2082 const Register qword_count = rdx; // elements count
2083 const Register end_from = from; // source array end address
2084 const Register end_to = rcx; // destination array end address
2085 const Register saved_to = to;
2086 // End pointers are inclusive, and if count is not zero they point
2087 // to the last unit copied: end_to[0] := end_from[0]
2088
2089 __ enter(); // required for proper stackwalking of RuntimeStub frame
2090 // Save no-overlap entry point for generate_conjoint_long_oop_copy()
2091 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
2092
2093 if (entry != NULL) {
2094 *entry = __ pc();
2095 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
2096 BLOCK_COMMENT("Entry:");
2097 }
2098
2099 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
2100 // r9 and r10 may be used to save non-volatile registers
2101 // 'from', 'to' and 'qword_count' are now valid
2102 if (is_oop) {
2103 // no registers are destroyed by this call
2104 gen_write_ref_array_pre_barrier(to, qword_count, dest_uninitialized);
2105 }
2106
2107 // Copy from low to high addresses. Use 'to' as scratch.
2108 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
2109 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
2110 __ negptr(qword_count);
2111 __ jmp(L_copy_bytes);
2112
2113 // Copy trailing qwords
2114 __ BIND(L_copy_8_bytes);
2115 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
2116 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
2117 __ increment(qword_count);
2118 __ jcc(Assembler::notZero, L_copy_8_bytes);
2119
2120 if (is_oop) {
2121 __ jmp(L_exit);
2122 } else {
2123 restore_arg_regs();
2124 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2125 __ xorptr(rax, rax); // return 0
2126 __ leave(); // required for proper stackwalking of RuntimeStub frame
2127 __ ret(0);
2128 }
2129
2130 // Copy in multi-bytes chunks
2131 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
2132
2133 if (is_oop) {
2134 __ BIND(L_exit);
2135 gen_write_ref_array_post_barrier(saved_to, end_to, rax);
2136 }
2137 restore_arg_regs();
2138 if (is_oop) {
2139 inc_counter_np(SharedRuntime::_oop_array_copy_ctr); // Update counter after rscratch1 is free
2140 } else {
2141 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2142 }
2143 __ xorptr(rax, rax); // return 0
2144 __ leave(); // required for proper stackwalking of RuntimeStub frame
2145 __ ret(0);
2146
2147 return start;
2148 }
2149
2150 // Arguments:
2151 // aligned - true => Input and output aligned on a HeapWord boundary == 8 bytes
2152 // ignored
2153 // is_oop - true => oop array, so generate store check code
2154 // name - stub name string
2155 //
2156 // Inputs:
2157 // c_rarg0 - source array address
2158 // c_rarg1 - destination array address
2159 // c_rarg2 - element count, treated as ssize_t, can be zero
2160 //
2161 address generate_conjoint_long_oop_copy(bool aligned, bool is_oop,
2162 address nooverlap_target, address *entry,
2163 const char *name, bool dest_uninitialized = false) {
2164 __ align(CodeEntryAlignment);
2165 StubCodeMark mark(this, "StubRoutines", name);
2166 address start = __ pc();
2167
2168 Label L_copy_bytes, L_copy_8_bytes, L_exit;
2169 const Register from = rdi; // source array address
2170 const Register to = rsi; // destination array address
2171 const Register qword_count = rdx; // elements count
2172 const Register saved_count = rcx;
2173
2174 __ enter(); // required for proper stackwalking of RuntimeStub frame
2175 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
2176
2177 if (entry != NULL) {
2178 *entry = __ pc();
2179 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
2180 BLOCK_COMMENT("Entry:");
2181 }
2182
2183 array_overlap_test(nooverlap_target, Address::times_8);
2184 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
2185 // r9 and r10 may be used to save non-volatile registers
2186 // 'from', 'to' and 'qword_count' are now valid
2187 if (is_oop) {
2188 // Save to and count for store barrier
2189 __ movptr(saved_count, qword_count);
2190 // No registers are destroyed by this call
2191 gen_write_ref_array_pre_barrier(to, saved_count, dest_uninitialized);
2192 }
2193
2194 __ jmp(L_copy_bytes);
2195
2196 // Copy trailing qwords
2197 __ BIND(L_copy_8_bytes);
2198 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
2199 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
2200 __ decrement(qword_count);
2201 __ jcc(Assembler::notZero, L_copy_8_bytes);
2202
2203 if (is_oop) {
2204 __ jmp(L_exit);
2205 } else {
2206 restore_arg_regs();
2207 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2208 __ xorptr(rax, rax); // return 0
2209 __ leave(); // required for proper stackwalking of RuntimeStub frame
2210 __ ret(0);
2211 }
2212
2213 // Copy in multi-bytes chunks
2214 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
2215
2216 if (is_oop) {
2217 __ BIND(L_exit);
2218 __ lea(rcx, Address(to, saved_count, Address::times_8, -8));
2219 gen_write_ref_array_post_barrier(to, rcx, rax);
2220 }
2221 restore_arg_regs();
2222 if (is_oop) {
2223 inc_counter_np(SharedRuntime::_oop_array_copy_ctr); // Update counter after rscratch1 is free
2224 } else {
2225 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2226 }
2227 __ xorptr(rax, rax); // return 0
2228 __ leave(); // required for proper stackwalking of RuntimeStub frame
2229 __ ret(0);
2230
2231 return start;
2232 }
2233
2234
2235 // Helper for generating a dynamic type check.
2236 // Smashes no registers.
2237 void generate_type_check(Register sub_klass,
2238 Register super_check_offset,
2239 Register super_klass,
2240 Label& L_success) {
2241 assert_different_registers(sub_klass, super_check_offset, super_klass);
2242
2243 BLOCK_COMMENT("type_check:");
2244
2245 Label L_miss;
2246
2247 __ check_klass_subtype_fast_path(sub_klass, super_klass, noreg, &L_success, &L_miss, NULL,
2248 super_check_offset);
2249 __ check_klass_subtype_slow_path(sub_klass, super_klass, noreg, noreg, &L_success, NULL);
2250
2251 // Fall through on failure!
2252 __ BIND(L_miss);
2253 }
2254
2255 //
2256 // Generate checkcasting array copy stub
2257 //
2258 // Input:
2259 // c_rarg0 - source array address
2260 // c_rarg1 - destination array address
2261 // c_rarg2 - element count, treated as ssize_t, can be zero
2262 // c_rarg3 - size_t ckoff (super_check_offset)
2263 // not Win64
2264 // c_rarg4 - oop ckval (super_klass)
2265 // Win64
2266 // rsp+40 - oop ckval (super_klass)
2267 //
2268 // Output:
2269 // rax == 0 - success
2270 // rax == -1^K - failure, where K is partial transfer count
2271 //
2272 address generate_checkcast_copy(const char *name, address *entry,
2273 bool dest_uninitialized = false) {
2274
2275 Label L_load_element, L_store_element, L_do_card_marks, L_done;
2276
2277 // Input registers (after setup_arg_regs)
2278 const Register from = rdi; // source array address
2279 const Register to = rsi; // destination array address
2280 const Register length = rdx; // elements count
2281 const Register ckoff = rcx; // super_check_offset
2282 const Register ckval = r8; // super_klass
2283
2284 // Registers used as temps (r13, r14 are save-on-entry)
2285 const Register end_from = from; // source array end address
2286 const Register end_to = r13; // destination array end address
2287 const Register count = rdx; // -(count_remaining)
2288 const Register r14_length = r14; // saved copy of length
2289 // End pointers are inclusive, and if length is not zero they point
2290 // to the last unit copied: end_to[0] := end_from[0]
2291
2292 const Register rax_oop = rax; // actual oop copied
2293 const Register r11_klass = r11; // oop._klass
2294
2295 //---------------------------------------------------------------
2296 // Assembler stub will be used for this call to arraycopy
2297 // if the two arrays are subtypes of Object[] but the
2298 // destination array type is not equal to or a supertype
2299 // of the source type. Each element must be separately
2300 // checked.
2301
2302 __ align(CodeEntryAlignment);
2303 StubCodeMark mark(this, "StubRoutines", name);
2304 address start = __ pc();
2305
2306 __ enter(); // required for proper stackwalking of RuntimeStub frame
2307
2308 #ifdef ASSERT
2309 // caller guarantees that the arrays really are different
2310 // otherwise, we would have to make conjoint checks
2311 { Label L;
2312 array_overlap_test(L, TIMES_OOP);
2313 __ stop("checkcast_copy within a single array");
2314 __ bind(L);
2315 }
2316 #endif //ASSERT
2317
2318 setup_arg_regs(4); // from => rdi, to => rsi, length => rdx
2319 // ckoff => rcx, ckval => r8
2320 // r9 and r10 may be used to save non-volatile registers
2321 #ifdef _WIN64
2322 // last argument (#4) is on stack on Win64
2323 __ movptr(ckval, Address(rsp, 6 * wordSize));
2324 #endif
2325
2326 // Caller of this entry point must set up the argument registers.
2327 if (entry != NULL) {
2328 *entry = __ pc();
2329 BLOCK_COMMENT("Entry:");
2330 }
2331
2332 // allocate spill slots for r13, r14
2333 enum {
2334 saved_r13_offset,
2335 saved_r14_offset,
2336 saved_rbp_offset
2337 };
2338 __ subptr(rsp, saved_rbp_offset * wordSize);
2339 __ movptr(Address(rsp, saved_r13_offset * wordSize), r13);
2340 __ movptr(Address(rsp, saved_r14_offset * wordSize), r14);
2341
2342 // check that int operands are properly extended to size_t
2343 assert_clean_int(length, rax);
2344 assert_clean_int(ckoff, rax);
2345
2346 #ifdef ASSERT
2347 BLOCK_COMMENT("assert consistent ckoff/ckval");
2348 // The ckoff and ckval must be mutually consistent,
2349 // even though caller generates both.
2350 { Label L;
2351 int sco_offset = in_bytes(Klass::super_check_offset_offset());
2352 __ cmpl(ckoff, Address(ckval, sco_offset));
2353 __ jcc(Assembler::equal, L);
2354 __ stop("super_check_offset inconsistent");
2355 __ bind(L);
2356 }
2357 #endif //ASSERT
2358
2359 // Loop-invariant addresses. They are exclusive end pointers.
2360 Address end_from_addr(from, length, TIMES_OOP, 0);
2361 Address end_to_addr(to, length, TIMES_OOP, 0);
2362 // Loop-variant addresses. They assume post-incremented count < 0.
2363 Address from_element_addr(end_from, count, TIMES_OOP, 0);
2364 Address to_element_addr(end_to, count, TIMES_OOP, 0);
2365
2366 gen_write_ref_array_pre_barrier(to, count, dest_uninitialized);
2367
2368 // Copy from low to high addresses, indexed from the end of each array.
2369 __ lea(end_from, end_from_addr);
2370 __ lea(end_to, end_to_addr);
2371 __ movptr(r14_length, length); // save a copy of the length
2372 assert(length == count, ""); // else fix next line:
2373 __ negptr(count); // negate and test the length
2374 __ jcc(Assembler::notZero, L_load_element);
2375
2376 // Empty array: Nothing to do.
2377 __ xorptr(rax, rax); // return 0 on (trivial) success
2378 __ jmp(L_done);
2379
2380 // ======== begin loop ========
2381 // (Loop is rotated; its entry is L_load_element.)
2382 // Loop control:
2383 // for (count = -count; count != 0; count++)
2384 // Base pointers src, dst are biased by 8*(count-1),to last element.
2385 __ align(OptoLoopAlignment);
2386
2387 __ BIND(L_store_element);
2388 __ store_heap_oop(to_element_addr, rax_oop); // store the oop
2389 __ increment(count); // increment the count toward zero
2390 __ jcc(Assembler::zero, L_do_card_marks);
2391
2392 // ======== loop entry is here ========
2393 __ BIND(L_load_element);
2394 __ load_heap_oop(rax_oop, from_element_addr); // load the oop
2395 __ testptr(rax_oop, rax_oop);
2396 __ jcc(Assembler::zero, L_store_element);
2397
2398 __ load_klass(r11_klass, rax_oop);// query the object klass
2399 generate_type_check(r11_klass, ckoff, ckval, L_store_element);
2400 // ======== end loop ========
2401
2402 // It was a real error; we must depend on the caller to finish the job.
2403 // Register rdx = -1 * number of *remaining* oops, r14 = *total* oops.
2404 // Emit GC store barriers for the oops we have copied (r14 + rdx),
2405 // and report their number to the caller.
2406 assert_different_registers(rax, r14_length, count, to, end_to, rcx);
2407 __ lea(end_to, to_element_addr);
2408 __ addptr(end_to, -heapOopSize); // make an inclusive end pointer
2409 gen_write_ref_array_post_barrier(to, end_to, rscratch1);
2410 __ movptr(rax, r14_length); // original oops
2411 __ addptr(rax, count); // K = (original - remaining) oops
2412 __ notptr(rax); // report (-1^K) to caller
2413 __ jmp(L_done);
2414
2415 // Come here on success only.
2416 __ BIND(L_do_card_marks);
2417 __ addptr(end_to, -heapOopSize); // make an inclusive end pointer
2418 gen_write_ref_array_post_barrier(to, end_to, rscratch1);
2419 __ xorptr(rax, rax); // return 0 on success
2420
2421 // Common exit point (success or failure).
2422 __ BIND(L_done);
2423 __ movptr(r13, Address(rsp, saved_r13_offset * wordSize));
2424 __ movptr(r14, Address(rsp, saved_r14_offset * wordSize));
2425 restore_arg_regs();
2426 inc_counter_np(SharedRuntime::_checkcast_array_copy_ctr); // Update counter after rscratch1 is free
2427 __ leave(); // required for proper stackwalking of RuntimeStub frame
2428 __ ret(0);
2429
2430 return start;
2431 }
2432
2433 //
2434 // Generate 'unsafe' array copy stub
2435 // Though just as safe as the other stubs, it takes an unscaled
2436 // size_t argument instead of an element count.
2437 //
2438 // Input:
2439 // c_rarg0 - source array address
2440 // c_rarg1 - destination array address
2441 // c_rarg2 - byte count, treated as ssize_t, can be zero
2442 //
2443 // Examines the alignment of the operands and dispatches
2444 // to a long, int, short, or byte copy loop.
2445 //
2446 address generate_unsafe_copy(const char *name,
2447 address byte_copy_entry, address short_copy_entry,
2448 address int_copy_entry, address long_copy_entry) {
2449
2450 Label L_long_aligned, L_int_aligned, L_short_aligned;
2451
2452 // Input registers (before setup_arg_regs)
2453 const Register from = c_rarg0; // source array address
2454 const Register to = c_rarg1; // destination array address
2455 const Register size = c_rarg2; // byte count (size_t)
2456
2457 // Register used as a temp
2458 const Register bits = rax; // test copy of low bits
2459
2460 __ align(CodeEntryAlignment);
2461 StubCodeMark mark(this, "StubRoutines", name);
2462 address start = __ pc();
2463
2464 __ enter(); // required for proper stackwalking of RuntimeStub frame
2465
2466 // bump this on entry, not on exit:
2467 inc_counter_np(SharedRuntime::_unsafe_array_copy_ctr);
2468
2469 __ mov(bits, from);
2470 __ orptr(bits, to);
2471 __ orptr(bits, size);
2472
2473 __ testb(bits, BytesPerLong-1);
2474 __ jccb(Assembler::zero, L_long_aligned);
2475
2476 __ testb(bits, BytesPerInt-1);
2477 __ jccb(Assembler::zero, L_int_aligned);
2478
2479 __ testb(bits, BytesPerShort-1);
2480 __ jump_cc(Assembler::notZero, RuntimeAddress(byte_copy_entry));
2481
2482 __ BIND(L_short_aligned);
2483 __ shrptr(size, LogBytesPerShort); // size => short_count
2484 __ jump(RuntimeAddress(short_copy_entry));
2485
2486 __ BIND(L_int_aligned);
2487 __ shrptr(size, LogBytesPerInt); // size => int_count
2488 __ jump(RuntimeAddress(int_copy_entry));
2489
2490 __ BIND(L_long_aligned);
2491 __ shrptr(size, LogBytesPerLong); // size => qword_count
2492 __ jump(RuntimeAddress(long_copy_entry));
2493
2494 return start;
2495 }
2496
2497 // Perform range checks on the proposed arraycopy.
2498 // Kills temp, but nothing else.
2499 // Also, clean the sign bits of src_pos and dst_pos.
2500 void arraycopy_range_checks(Register src, // source array oop (c_rarg0)
2501 Register src_pos, // source position (c_rarg1)
2502 Register dst, // destination array oo (c_rarg2)
2503 Register dst_pos, // destination position (c_rarg3)
2504 Register length,
2505 Register temp,
2506 Label& L_failed) {
2507 BLOCK_COMMENT("arraycopy_range_checks:");
2508
2509 // if (src_pos + length > arrayOop(src)->length()) FAIL;
2510 __ movl(temp, length);
2511 __ addl(temp, src_pos); // src_pos + length
2512 __ cmpl(temp, Address(src, arrayOopDesc::length_offset_in_bytes()));
2513 __ jcc(Assembler::above, L_failed);
2514
2515 // if (dst_pos + length > arrayOop(dst)->length()) FAIL;
2516 __ movl(temp, length);
2517 __ addl(temp, dst_pos); // dst_pos + length
2518 __ cmpl(temp, Address(dst, arrayOopDesc::length_offset_in_bytes()));
2519 __ jcc(Assembler::above, L_failed);
2520
2521 // Have to clean up high 32-bits of 'src_pos' and 'dst_pos'.
2522 // Move with sign extension can be used since they are positive.
2523 __ movslq(src_pos, src_pos);
2524 __ movslq(dst_pos, dst_pos);
2525
2526 BLOCK_COMMENT("arraycopy_range_checks done");
2527 }
2528
2529 //
2530 // Generate generic array copy stubs
2531 //
2532 // Input:
2533 // c_rarg0 - src oop
2534 // c_rarg1 - src_pos (32-bits)
2535 // c_rarg2 - dst oop
2536 // c_rarg3 - dst_pos (32-bits)
2537 // not Win64
2538 // c_rarg4 - element count (32-bits)
2539 // Win64
2540 // rsp+40 - element count (32-bits)
2541 //
2542 // Output:
2543 // rax == 0 - success
2544 // rax == -1^K - failure, where K is partial transfer count
2545 //
2546 address generate_generic_copy(const char *name,
2547 address byte_copy_entry, address short_copy_entry,
2548 address int_copy_entry, address oop_copy_entry,
2549 address long_copy_entry, address checkcast_copy_entry) {
2550
2551 Label L_failed, L_failed_0, L_objArray;
2552 Label L_copy_bytes, L_copy_shorts, L_copy_ints, L_copy_longs;
2553
2554 // Input registers
2555 const Register src = c_rarg0; // source array oop
2556 const Register src_pos = c_rarg1; // source position
2557 const Register dst = c_rarg2; // destination array oop
2558 const Register dst_pos = c_rarg3; // destination position
2559 #ifndef _WIN64
2560 const Register length = c_rarg4;
2561 #else
2562 const Address length(rsp, 6 * wordSize); // elements count is on stack on Win64
2563 #endif
2564
2565 { int modulus = CodeEntryAlignment;
2566 int target = modulus - 5; // 5 = sizeof jmp(L_failed)
2567 int advance = target - (__ offset() % modulus);
2568 if (advance < 0) advance += modulus;
2569 if (advance > 0) __ nop(advance);
2570 }
2571 StubCodeMark mark(this, "StubRoutines", name);
2572
2573 // Short-hop target to L_failed. Makes for denser prologue code.
2574 __ BIND(L_failed_0);
2575 __ jmp(L_failed);
2576 assert(__ offset() % CodeEntryAlignment == 0, "no further alignment needed");
2577
2578 __ align(CodeEntryAlignment);
2579 address start = __ pc();
2580
2581 __ enter(); // required for proper stackwalking of RuntimeStub frame
2582
2583 // bump this on entry, not on exit:
2584 inc_counter_np(SharedRuntime::_generic_array_copy_ctr);
2585
2586 //-----------------------------------------------------------------------
2587 // Assembler stub will be used for this call to arraycopy
2588 // if the following conditions are met:
2589 //
2590 // (1) src and dst must not be null.
2591 // (2) src_pos must not be negative.
2592 // (3) dst_pos must not be negative.
2593 // (4) length must not be negative.
2594 // (5) src klass and dst klass should be the same and not NULL.
2595 // (6) src and dst should be arrays.
2596 // (7) src_pos + length must not exceed length of src.
2597 // (8) dst_pos + length must not exceed length of dst.
2598 //
2599
2600 // if (src == NULL) return -1;
2601 __ testptr(src, src); // src oop
2602 size_t j1off = __ offset();
2603 __ jccb(Assembler::zero, L_failed_0);
2604
2605 // if (src_pos < 0) return -1;
2606 __ testl(src_pos, src_pos); // src_pos (32-bits)
2607 __ jccb(Assembler::negative, L_failed_0);
2608
2609 // if (dst == NULL) return -1;
2610 __ testptr(dst, dst); // dst oop
2611 __ jccb(Assembler::zero, L_failed_0);
2612
2613 // if (dst_pos < 0) return -1;
2614 __ testl(dst_pos, dst_pos); // dst_pos (32-bits)
2615 size_t j4off = __ offset();
2616 __ jccb(Assembler::negative, L_failed_0);
2617
2618 // The first four tests are very dense code,
2619 // but not quite dense enough to put four
2620 // jumps in a 16-byte instruction fetch buffer.
2621 // That's good, because some branch predicters
2622 // do not like jumps so close together.
2623 // Make sure of this.
2624 guarantee(((j1off ^ j4off) & ~15) != 0, "I$ line of 1st & 4th jumps");
2625
2626 // registers used as temp
2627 const Register r11_length = r11; // elements count to copy
2628 const Register r10_src_klass = r10; // array klass
2629
2630 // if (length < 0) return -1;
2631 __ movl(r11_length, length); // length (elements count, 32-bits value)
2632 __ testl(r11_length, r11_length);
2633 __ jccb(Assembler::negative, L_failed_0);
2634
2635 __ load_klass(r10_src_klass, src);
2636 #ifdef ASSERT
2637 // assert(src->klass() != NULL);
2638 {
2639 BLOCK_COMMENT("assert klasses not null {");
2640 Label L1, L2;
2641 __ testptr(r10_src_klass, r10_src_klass);
2642 __ jcc(Assembler::notZero, L2); // it is broken if klass is NULL
2643 __ bind(L1);
2644 __ stop("broken null klass");
2645 __ bind(L2);
2646 __ load_klass(rax, dst);
2647 __ cmpq(rax, 0);
2648 __ jcc(Assembler::equal, L1); // this would be broken also
2649 BLOCK_COMMENT("} assert klasses not null done");
2650 }
2651 #endif
2652
2653 // Load layout helper (32-bits)
2654 //
2655 // |array_tag| | header_size | element_type | |log2_element_size|
2656 // 32 30 24 16 8 2 0
2657 //
2658 // array_tag: typeArray = 0x3, objArray = 0x2, non-array = 0x0
2659 //
2660
2661 const int lh_offset = in_bytes(Klass::layout_helper_offset());
2662
2663 // Handle objArrays completely differently...
2664 const jint objArray_lh = Klass::array_layout_helper(T_OBJECT);
2665 __ cmpl(Address(r10_src_klass, lh_offset), objArray_lh);
2666 __ jcc(Assembler::equal, L_objArray);
2667
2668 // if (src->klass() != dst->klass()) return -1;
2669 __ load_klass(rax, dst);
2670 __ cmpq(r10_src_klass, rax);
2671 __ jcc(Assembler::notEqual, L_failed);
2672
2673 const Register rax_lh = rax; // layout helper
2674 __ movl(rax_lh, Address(r10_src_klass, lh_offset));
2675
2676 // if (!src->is_Array()) return -1;
2677 __ cmpl(rax_lh, Klass::_lh_neutral_value);
2678 __ jcc(Assembler::greaterEqual, L_failed);
2679
2680 // At this point, it is known to be a typeArray (array_tag 0x3).
2681 #ifdef ASSERT
2682 {
2683 BLOCK_COMMENT("assert primitive array {");
2684 Label L;
2685 __ cmpl(rax_lh, (Klass::_lh_array_tag_type_value << Klass::_lh_array_tag_shift));
2686 __ jcc(Assembler::greaterEqual, L);
2687 __ stop("must be a primitive array");
2688 __ bind(L);
2689 BLOCK_COMMENT("} assert primitive array done");
2690 }
2691 #endif
2692
2693 arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length,
2694 r10, L_failed);
2695
2696 // typeArrayKlass
2697 //
2698 // src_addr = (src + array_header_in_bytes()) + (src_pos << log2elemsize);
2699 // dst_addr = (dst + array_header_in_bytes()) + (dst_pos << log2elemsize);
2700 //
2701
2702 const Register r10_offset = r10; // array offset
2703 const Register rax_elsize = rax_lh; // element size
2704
2705 __ movl(r10_offset, rax_lh);
2706 __ shrl(r10_offset, Klass::_lh_header_size_shift);
2707 __ andptr(r10_offset, Klass::_lh_header_size_mask); // array_offset
2708 __ addptr(src, r10_offset); // src array offset
2709 __ addptr(dst, r10_offset); // dst array offset
2710 BLOCK_COMMENT("choose copy loop based on element size");
2711 __ andl(rax_lh, Klass::_lh_log2_element_size_mask); // rax_lh -> rax_elsize
2712
2713 // next registers should be set before the jump to corresponding stub
2714 const Register from = c_rarg0; // source array address
2715 const Register to = c_rarg1; // destination array address
2716 const Register count = c_rarg2; // elements count
2717
2718 // 'from', 'to', 'count' registers should be set in such order
2719 // since they are the same as 'src', 'src_pos', 'dst'.
2720
2721 __ BIND(L_copy_bytes);
2722 __ cmpl(rax_elsize, 0);
2723 __ jccb(Assembler::notEqual, L_copy_shorts);
2724 __ lea(from, Address(src, src_pos, Address::times_1, 0));// src_addr
2725 __ lea(to, Address(dst, dst_pos, Address::times_1, 0));// dst_addr
2726 __ movl2ptr(count, r11_length); // length
2727 __ jump(RuntimeAddress(byte_copy_entry));
2728
2729 __ BIND(L_copy_shorts);
2730 __ cmpl(rax_elsize, LogBytesPerShort);
2731 __ jccb(Assembler::notEqual, L_copy_ints);
2732 __ lea(from, Address(src, src_pos, Address::times_2, 0));// src_addr
2733 __ lea(to, Address(dst, dst_pos, Address::times_2, 0));// dst_addr
2734 __ movl2ptr(count, r11_length); // length
2735 __ jump(RuntimeAddress(short_copy_entry));
2736
2737 __ BIND(L_copy_ints);
2738 __ cmpl(rax_elsize, LogBytesPerInt);
2739 __ jccb(Assembler::notEqual, L_copy_longs);
2740 __ lea(from, Address(src, src_pos, Address::times_4, 0));// src_addr
2741 __ lea(to, Address(dst, dst_pos, Address::times_4, 0));// dst_addr
2742 __ movl2ptr(count, r11_length); // length
2743 __ jump(RuntimeAddress(int_copy_entry));
2744
2745 __ BIND(L_copy_longs);
2746 #ifdef ASSERT
2747 {
2748 BLOCK_COMMENT("assert long copy {");
2749 Label L;
2750 __ cmpl(rax_elsize, LogBytesPerLong);
2751 __ jcc(Assembler::equal, L);
2752 __ stop("must be long copy, but elsize is wrong");
2753 __ bind(L);
2754 BLOCK_COMMENT("} assert long copy done");
2755 }
2756 #endif
2757 __ lea(from, Address(src, src_pos, Address::times_8, 0));// src_addr
2758 __ lea(to, Address(dst, dst_pos, Address::times_8, 0));// dst_addr
2759 __ movl2ptr(count, r11_length); // length
2760 __ jump(RuntimeAddress(long_copy_entry));
2761
2762 // objArrayKlass
2763 __ BIND(L_objArray);
2764 // live at this point: r10_src_klass, r11_length, src[_pos], dst[_pos]
2765
2766 Label L_plain_copy, L_checkcast_copy;
2767 // test array classes for subtyping
2768 __ load_klass(rax, dst);
2769 __ cmpq(r10_src_klass, rax); // usual case is exact equality
2770 __ jcc(Assembler::notEqual, L_checkcast_copy);
2771
2772 // Identically typed arrays can be copied without element-wise checks.
2773 arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length,
2774 r10, L_failed);
2775
2776 __ lea(from, Address(src, src_pos, TIMES_OOP,
2777 arrayOopDesc::base_offset_in_bytes(T_OBJECT))); // src_addr
2778 __ lea(to, Address(dst, dst_pos, TIMES_OOP,
2779 arrayOopDesc::base_offset_in_bytes(T_OBJECT))); // dst_addr
2780 __ movl2ptr(count, r11_length); // length
2781 __ BIND(L_plain_copy);
2782 __ jump(RuntimeAddress(oop_copy_entry));
2783
2784 __ BIND(L_checkcast_copy);
2785 // live at this point: r10_src_klass, r11_length, rax (dst_klass)
2786 {
2787 // Before looking at dst.length, make sure dst is also an objArray.
2788 __ cmpl(Address(rax, lh_offset), objArray_lh);
2789 __ jcc(Assembler::notEqual, L_failed);
2790
2791 // It is safe to examine both src.length and dst.length.
2792 arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length,
2793 rax, L_failed);
2794
2795 const Register r11_dst_klass = r11;
2796 __ load_klass(r11_dst_klass, dst); // reload
2797
2798 // Marshal the base address arguments now, freeing registers.
2799 __ lea(from, Address(src, src_pos, TIMES_OOP,
2800 arrayOopDesc::base_offset_in_bytes(T_OBJECT)));
2801 __ lea(to, Address(dst, dst_pos, TIMES_OOP,
2802 arrayOopDesc::base_offset_in_bytes(T_OBJECT)));
2803 __ movl(count, length); // length (reloaded)
2804 Register sco_temp = c_rarg3; // this register is free now
2805 assert_different_registers(from, to, count, sco_temp,
2806 r11_dst_klass, r10_src_klass);
2807 assert_clean_int(count, sco_temp);
2808
2809 // Generate the type check.
2810 const int sco_offset = in_bytes(Klass::super_check_offset_offset());
2811 __ movl(sco_temp, Address(r11_dst_klass, sco_offset));
2812 assert_clean_int(sco_temp, rax);
2813 generate_type_check(r10_src_klass, sco_temp, r11_dst_klass, L_plain_copy);
2814
2815 // Fetch destination element klass from the objArrayKlass header.
2816 int ek_offset = in_bytes(objArrayKlass::element_klass_offset());
2817 __ movptr(r11_dst_klass, Address(r11_dst_klass, ek_offset));
2818 __ movl( sco_temp, Address(r11_dst_klass, sco_offset));
2819 assert_clean_int(sco_temp, rax);
2820
2821 // the checkcast_copy loop needs two extra arguments:
2822 assert(c_rarg3 == sco_temp, "#3 already in place");
2823 // Set up arguments for checkcast_copy_entry.
2824 setup_arg_regs(4);
2825 __ movptr(r8, r11_dst_klass); // dst.klass.element_klass, r8 is c_rarg4 on Linux/Solaris
2826 __ jump(RuntimeAddress(checkcast_copy_entry));
2827 }
2828
2829 __ BIND(L_failed);
2830 __ xorptr(rax, rax);
2831 __ notptr(rax); // return -1
2832 __ leave(); // required for proper stackwalking of RuntimeStub frame
2833 __ ret(0);
2834
2835 return start;
2836 }
2837
2838 void generate_arraycopy_stubs() {
2839 address entry;
2840 address entry_jbyte_arraycopy;
2841 address entry_jshort_arraycopy;
2842 address entry_jint_arraycopy;
2843 address entry_oop_arraycopy;
2844 address entry_jlong_arraycopy;
2845 address entry_checkcast_arraycopy;
2846
2847 StubRoutines::_jbyte_disjoint_arraycopy = generate_disjoint_byte_copy(false, &entry,
2848 "jbyte_disjoint_arraycopy");
2849 StubRoutines::_jbyte_arraycopy = generate_conjoint_byte_copy(false, entry, &entry_jbyte_arraycopy,
2850 "jbyte_arraycopy");
2851
2852 StubRoutines::_jshort_disjoint_arraycopy = generate_disjoint_short_copy(false, &entry,
2853 "jshort_disjoint_arraycopy");
2854 StubRoutines::_jshort_arraycopy = generate_conjoint_short_copy(false, entry, &entry_jshort_arraycopy,
2855 "jshort_arraycopy");
2856
2857 StubRoutines::_jint_disjoint_arraycopy = generate_disjoint_int_oop_copy(false, false, &entry,
2858 "jint_disjoint_arraycopy");
2859 StubRoutines::_jint_arraycopy = generate_conjoint_int_oop_copy(false, false, entry,
2860 &entry_jint_arraycopy, "jint_arraycopy");
2861
2862 StubRoutines::_jlong_disjoint_arraycopy = generate_disjoint_long_oop_copy(false, false, &entry,
2863 "jlong_disjoint_arraycopy");
2864 StubRoutines::_jlong_arraycopy = generate_conjoint_long_oop_copy(false, false, entry,
2865 &entry_jlong_arraycopy, "jlong_arraycopy");
2866
2867
2868 if (UseCompressedOops) {
2869 StubRoutines::_oop_disjoint_arraycopy = generate_disjoint_int_oop_copy(false, true, &entry,
2870 "oop_disjoint_arraycopy");
2871 StubRoutines::_oop_arraycopy = generate_conjoint_int_oop_copy(false, true, entry,
2872 &entry_oop_arraycopy, "oop_arraycopy");
2873 StubRoutines::_oop_disjoint_arraycopy_uninit = generate_disjoint_int_oop_copy(false, true, &entry,
2874 "oop_disjoint_arraycopy_uninit",
2875 /*dest_uninitialized*/true);
2876 StubRoutines::_oop_arraycopy_uninit = generate_conjoint_int_oop_copy(false, true, entry,
2877 NULL, "oop_arraycopy_uninit",
2878 /*dest_uninitialized*/true);
2879 } else {
2880 StubRoutines::_oop_disjoint_arraycopy = generate_disjoint_long_oop_copy(false, true, &entry,
2881 "oop_disjoint_arraycopy");
2882 StubRoutines::_oop_arraycopy = generate_conjoint_long_oop_copy(false, true, entry,
2883 &entry_oop_arraycopy, "oop_arraycopy");
2884 StubRoutines::_oop_disjoint_arraycopy_uninit = generate_disjoint_long_oop_copy(false, true, &entry,
2885 "oop_disjoint_arraycopy_uninit",
2886 /*dest_uninitialized*/true);
2887 StubRoutines::_oop_arraycopy_uninit = generate_conjoint_long_oop_copy(false, true, entry,
2888 NULL, "oop_arraycopy_uninit",
2889 /*dest_uninitialized*/true);
2890 }
2891
2892 StubRoutines::_checkcast_arraycopy = generate_checkcast_copy("checkcast_arraycopy", &entry_checkcast_arraycopy);
2893 StubRoutines::_checkcast_arraycopy_uninit = generate_checkcast_copy("checkcast_arraycopy_uninit", NULL,
2894 /*dest_uninitialized*/true);
2895
2896 StubRoutines::_unsafe_arraycopy = generate_unsafe_copy("unsafe_arraycopy",
2897 entry_jbyte_arraycopy,
2898 entry_jshort_arraycopy,
2899 entry_jint_arraycopy,
2900 entry_jlong_arraycopy);
2901 StubRoutines::_generic_arraycopy = generate_generic_copy("generic_arraycopy",
2902 entry_jbyte_arraycopy,
2903 entry_jshort_arraycopy,
2904 entry_jint_arraycopy,
2905 entry_oop_arraycopy,
2906 entry_jlong_arraycopy,
2907 entry_checkcast_arraycopy);
2908
2909 StubRoutines::_jbyte_fill = generate_fill(T_BYTE, false, "jbyte_fill");
2910 StubRoutines::_jshort_fill = generate_fill(T_SHORT, false, "jshort_fill");
2911 StubRoutines::_jint_fill = generate_fill(T_INT, false, "jint_fill");
2912 StubRoutines::_arrayof_jbyte_fill = generate_fill(T_BYTE, true, "arrayof_jbyte_fill");
2913 StubRoutines::_arrayof_jshort_fill = generate_fill(T_SHORT, true, "arrayof_jshort_fill");
2914 StubRoutines::_arrayof_jint_fill = generate_fill(T_INT, true, "arrayof_jint_fill");
2915
2916 // We don't generate specialized code for HeapWord-aligned source
2917 // arrays, so just use the code we've already generated
2918 StubRoutines::_arrayof_jbyte_disjoint_arraycopy = StubRoutines::_jbyte_disjoint_arraycopy;
2919 StubRoutines::_arrayof_jbyte_arraycopy = StubRoutines::_jbyte_arraycopy;
2920
2921 StubRoutines::_arrayof_jshort_disjoint_arraycopy = StubRoutines::_jshort_disjoint_arraycopy;
2922 StubRoutines::_arrayof_jshort_arraycopy = StubRoutines::_jshort_arraycopy;
2923
2924 StubRoutines::_arrayof_jint_disjoint_arraycopy = StubRoutines::_jint_disjoint_arraycopy;
2925 StubRoutines::_arrayof_jint_arraycopy = StubRoutines::_jint_arraycopy;
2926
2927 StubRoutines::_arrayof_jlong_disjoint_arraycopy = StubRoutines::_jlong_disjoint_arraycopy;
2928 StubRoutines::_arrayof_jlong_arraycopy = StubRoutines::_jlong_arraycopy;
2929
2930 StubRoutines::_arrayof_oop_disjoint_arraycopy = StubRoutines::_oop_disjoint_arraycopy;
2931 StubRoutines::_arrayof_oop_arraycopy = StubRoutines::_oop_arraycopy;
2932
2933 StubRoutines::_arrayof_oop_disjoint_arraycopy_uninit = StubRoutines::_oop_disjoint_arraycopy_uninit;
2934 StubRoutines::_arrayof_oop_arraycopy_uninit = StubRoutines::_oop_arraycopy_uninit;
2935 }
2936
2937 void generate_math_stubs() {
2938 {
2939 StubCodeMark mark(this, "StubRoutines", "log");
2940 StubRoutines::_intrinsic_log = (double (*)(double)) __ pc();
2941
2942 __ subq(rsp, 8);
2943 __ movdbl(Address(rsp, 0), xmm0);
2944 __ fld_d(Address(rsp, 0));
2945 __ flog();
2946 __ fstp_d(Address(rsp, 0));
2947 __ movdbl(xmm0, Address(rsp, 0));
2948 __ addq(rsp, 8);
2949 __ ret(0);
2950 }
2951 {
2952 StubCodeMark mark(this, "StubRoutines", "log10");
2953 StubRoutines::_intrinsic_log10 = (double (*)(double)) __ pc();
2954
2955 __ subq(rsp, 8);
2956 __ movdbl(Address(rsp, 0), xmm0);
2957 __ fld_d(Address(rsp, 0));
2958 __ flog10();
2959 __ fstp_d(Address(rsp, 0));
2960 __ movdbl(xmm0, Address(rsp, 0));
2961 __ addq(rsp, 8);
2962 __ ret(0);
2963 }
2964 {
2965 StubCodeMark mark(this, "StubRoutines", "sin");
2966 StubRoutines::_intrinsic_sin = (double (*)(double)) __ pc();
2967
2968 __ subq(rsp, 8);
2969 __ movdbl(Address(rsp, 0), xmm0);
2970 __ fld_d(Address(rsp, 0));
2971 __ trigfunc('s');
2972 __ fstp_d(Address(rsp, 0));
2973 __ movdbl(xmm0, Address(rsp, 0));
2974 __ addq(rsp, 8);
2975 __ ret(0);
2976 }
2977 {
2978 StubCodeMark mark(this, "StubRoutines", "cos");
2979 StubRoutines::_intrinsic_cos = (double (*)(double)) __ pc();
2980
2981 __ subq(rsp, 8);
2982 __ movdbl(Address(rsp, 0), xmm0);
2983 __ fld_d(Address(rsp, 0));
2984 __ trigfunc('c');
2985 __ fstp_d(Address(rsp, 0));
2986 __ movdbl(xmm0, Address(rsp, 0));
2987 __ addq(rsp, 8);
2988 __ ret(0);
2989 }
2990 {
2991 StubCodeMark mark(this, "StubRoutines", "tan");
2992 StubRoutines::_intrinsic_tan = (double (*)(double)) __ pc();
2993
2994 __ subq(rsp, 8);
2995 __ movdbl(Address(rsp, 0), xmm0);
2996 __ fld_d(Address(rsp, 0));
2997 __ trigfunc('t');
2998 __ fstp_d(Address(rsp, 0));
2999 __ movdbl(xmm0, Address(rsp, 0));
3000 __ addq(rsp, 8);
3001 __ ret(0);
3002 }
3003 {
3004 StubCodeMark mark(this, "StubRoutines", "exp");
3005 StubRoutines::_intrinsic_exp = (double (*)(double)) __ pc();
3006
3007 __ subq(rsp, 8);
3008 __ movdbl(Address(rsp, 0), xmm0);
3009 __ fld_d(Address(rsp, 0));
3010 __ exp_with_fallback(0);
3011 __ fstp_d(Address(rsp, 0));
3012 __ movdbl(xmm0, Address(rsp, 0));
3013 __ addq(rsp, 8);
3014 __ ret(0);
3015 }
3016 {
3017 StubCodeMark mark(this, "StubRoutines", "pow");
3018 StubRoutines::_intrinsic_pow = (double (*)(double,double)) __ pc();
3019
3020 __ subq(rsp, 8);
3021 __ movdbl(Address(rsp, 0), xmm1);
3022 __ fld_d(Address(rsp, 0));
3023 __ movdbl(Address(rsp, 0), xmm0);
3024 __ fld_d(Address(rsp, 0));
3025 __ pow_with_fallback(0);
3026 __ fstp_d(Address(rsp, 0));
3027 __ movdbl(xmm0, Address(rsp, 0));
3028 __ addq(rsp, 8);
3029 __ ret(0);
3030 }
3031 }
3032
3033 // AES intrinsic stubs
3034 enum {AESBlockSize = 16};
3035
3036 address generate_key_shuffle_mask() {
3037 __ align(16);
3038 StubCodeMark mark(this, "StubRoutines", "key_shuffle_mask");
3039 address start = __ pc();
3040 __ emit_data64( 0x0405060700010203, relocInfo::none );
3041 __ emit_data64( 0x0c0d0e0f08090a0b, relocInfo::none );
3042 return start;
3043 }
3044
3045 // Utility routine for loading a 128-bit key word in little endian format
3046 // can optionally specify that the shuffle mask is already in an xmmregister
3047 void load_key(XMMRegister xmmdst, Register key, int offset, XMMRegister xmm_shuf_mask=NULL) {
3048 __ movdqu(xmmdst, Address(key, offset));
3049 if (xmm_shuf_mask != NULL) {
3050 __ pshufb(xmmdst, xmm_shuf_mask);
3051 } else {
3052 __ pshufb(xmmdst, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3053 }
3054 }
3055
3056 // Arguments:
3057 //
3058 // Inputs:
3059 // c_rarg0 - source byte array address
3060 // c_rarg1 - destination byte array address
3061 // c_rarg2 - K (key) in little endian int array
3062 //
3063 address generate_aescrypt_encryptBlock() {
3064 assert(UseAES, "need AES instructions and misaligned SSE support");
3065 __ align(CodeEntryAlignment);
3066 StubCodeMark mark(this, "StubRoutines", "aescrypt_encryptBlock");
3067 Label L_doLast;
3068 address start = __ pc();
3069
3070 const Register from = c_rarg0; // source array address
3071 const Register to = c_rarg1; // destination array address
3072 const Register key = c_rarg2; // key array address
3073 const Register keylen = rax;
3074
3075 const XMMRegister xmm_result = xmm0;
3076 const XMMRegister xmm_key_shuf_mask = xmm1;
3077 // On win64 xmm6-xmm15 must be preserved so don't use them.
3078 const XMMRegister xmm_temp1 = xmm2;
3079 const XMMRegister xmm_temp2 = xmm3;
3080 const XMMRegister xmm_temp3 = xmm4;
3081 const XMMRegister xmm_temp4 = xmm5;
3082
3083 __ enter(); // required for proper stackwalking of RuntimeStub frame
3084
3085 // keylen could be only {11, 13, 15} * 4 = {44, 52, 60}
3086 __ movl(keylen, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3087
3088 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3089 __ movdqu(xmm_result, Address(from, 0)); // get 16 bytes of input
3090
3091 // For encryption, the java expanded key ordering is just what we need
3092 // we don't know if the key is aligned, hence not using load-execute form
3093
3094 load_key(xmm_temp1, key, 0x00, xmm_key_shuf_mask);
3095 __ pxor(xmm_result, xmm_temp1);
3096
3097 load_key(xmm_temp1, key, 0x10, xmm_key_shuf_mask);
3098 load_key(xmm_temp2, key, 0x20, xmm_key_shuf_mask);
3099 load_key(xmm_temp3, key, 0x30, xmm_key_shuf_mask);
3100 load_key(xmm_temp4, key, 0x40, xmm_key_shuf_mask);
3101
3102 __ aesenc(xmm_result, xmm_temp1);
3103 __ aesenc(xmm_result, xmm_temp2);
3104 __ aesenc(xmm_result, xmm_temp3);
3105 __ aesenc(xmm_result, xmm_temp4);
3106
3107 load_key(xmm_temp1, key, 0x50, xmm_key_shuf_mask);
3108 load_key(xmm_temp2, key, 0x60, xmm_key_shuf_mask);
3109 load_key(xmm_temp3, key, 0x70, xmm_key_shuf_mask);
3110 load_key(xmm_temp4, key, 0x80, xmm_key_shuf_mask);
3111
3112 __ aesenc(xmm_result, xmm_temp1);
3113 __ aesenc(xmm_result, xmm_temp2);
3114 __ aesenc(xmm_result, xmm_temp3);
3115 __ aesenc(xmm_result, xmm_temp4);
3116
3117 load_key(xmm_temp1, key, 0x90, xmm_key_shuf_mask);
3118 load_key(xmm_temp2, key, 0xa0, xmm_key_shuf_mask);
3119
3120 __ cmpl(keylen, 44);
3121 __ jccb(Assembler::equal, L_doLast);
3122
3123 __ aesenc(xmm_result, xmm_temp1);
3124 __ aesenc(xmm_result, xmm_temp2);
3125
3126 load_key(xmm_temp1, key, 0xb0, xmm_key_shuf_mask);
3127 load_key(xmm_temp2, key, 0xc0, xmm_key_shuf_mask);
3128
3129 __ cmpl(keylen, 52);
3130 __ jccb(Assembler::equal, L_doLast);
3131
3132 __ aesenc(xmm_result, xmm_temp1);
3133 __ aesenc(xmm_result, xmm_temp2);
3134
3135 load_key(xmm_temp1, key, 0xd0, xmm_key_shuf_mask);
3136 load_key(xmm_temp2, key, 0xe0, xmm_key_shuf_mask);
3137
3138 __ BIND(L_doLast);
3139 __ aesenc(xmm_result, xmm_temp1);
3140 __ aesenclast(xmm_result, xmm_temp2);
3141 __ movdqu(Address(to, 0), xmm_result); // store the result
3142 __ xorptr(rax, rax); // return 0
3143 __ leave(); // required for proper stackwalking of RuntimeStub frame
3144 __ ret(0);
3145
3146 return start;
3147 }
3148
3149
3150 // Arguments:
3151 //
3152 // Inputs:
3153 // c_rarg0 - source byte array address
3154 // c_rarg1 - destination byte array address
3155 // c_rarg2 - K (key) in little endian int array
3156 //
3157 address generate_aescrypt_decryptBlock() {
3158 assert(UseAES, "need AES instructions and misaligned SSE support");
3159 __ align(CodeEntryAlignment);
3160 StubCodeMark mark(this, "StubRoutines", "aescrypt_decryptBlock");
3161 Label L_doLast;
3162 address start = __ pc();
3163
3164 const Register from = c_rarg0; // source array address
3165 const Register to = c_rarg1; // destination array address
3166 const Register key = c_rarg2; // key array address
3167 const Register keylen = rax;
3168
3169 const XMMRegister xmm_result = xmm0;
3170 const XMMRegister xmm_key_shuf_mask = xmm1;
3171 // On win64 xmm6-xmm15 must be preserved so don't use them.
3172 const XMMRegister xmm_temp1 = xmm2;
3173 const XMMRegister xmm_temp2 = xmm3;
3174 const XMMRegister xmm_temp3 = xmm4;
3175 const XMMRegister xmm_temp4 = xmm5;
3176
3177 __ enter(); // required for proper stackwalking of RuntimeStub frame
3178
3179 // keylen could be only {11, 13, 15} * 4 = {44, 52, 60}
3180 __ movl(keylen, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3181
3182 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3183 __ movdqu(xmm_result, Address(from, 0));
3184
3185 // for decryption java expanded key ordering is rotated one position from what we want
3186 // so we start from 0x10 here and hit 0x00 last
3187 // we don't know if the key is aligned, hence not using load-execute form
3188 load_key(xmm_temp1, key, 0x10, xmm_key_shuf_mask);
3189 load_key(xmm_temp2, key, 0x20, xmm_key_shuf_mask);
3190 load_key(xmm_temp3, key, 0x30, xmm_key_shuf_mask);
3191 load_key(xmm_temp4, key, 0x40, xmm_key_shuf_mask);
3192
3193 __ pxor (xmm_result, xmm_temp1);
3194 __ aesdec(xmm_result, xmm_temp2);
3195 __ aesdec(xmm_result, xmm_temp3);
3196 __ aesdec(xmm_result, xmm_temp4);
3197
3198 load_key(xmm_temp1, key, 0x50, xmm_key_shuf_mask);
3199 load_key(xmm_temp2, key, 0x60, xmm_key_shuf_mask);
3200 load_key(xmm_temp3, key, 0x70, xmm_key_shuf_mask);
3201 load_key(xmm_temp4, key, 0x80, xmm_key_shuf_mask);
3202
3203 __ aesdec(xmm_result, xmm_temp1);
3204 __ aesdec(xmm_result, xmm_temp2);
3205 __ aesdec(xmm_result, xmm_temp3);
3206 __ aesdec(xmm_result, xmm_temp4);
3207
3208 load_key(xmm_temp1, key, 0x90, xmm_key_shuf_mask);
3209 load_key(xmm_temp2, key, 0xa0, xmm_key_shuf_mask);
3210 load_key(xmm_temp3, key, 0x00, xmm_key_shuf_mask);
3211
3212 __ cmpl(keylen, 44);
3213 __ jccb(Assembler::equal, L_doLast);
3214
3215 __ aesdec(xmm_result, xmm_temp1);
3216 __ aesdec(xmm_result, xmm_temp2);
3217
3218 load_key(xmm_temp1, key, 0xb0, xmm_key_shuf_mask);
3219 load_key(xmm_temp2, key, 0xc0, xmm_key_shuf_mask);
3220
3221 __ cmpl(keylen, 52);
3222 __ jccb(Assembler::equal, L_doLast);
3223
3224 __ aesdec(xmm_result, xmm_temp1);
3225 __ aesdec(xmm_result, xmm_temp2);
3226
3227 load_key(xmm_temp1, key, 0xd0, xmm_key_shuf_mask);
3228 load_key(xmm_temp2, key, 0xe0, xmm_key_shuf_mask);
3229
3230 __ BIND(L_doLast);
3231 __ aesdec(xmm_result, xmm_temp1);
3232 __ aesdec(xmm_result, xmm_temp2);
3233
3234 // for decryption the aesdeclast operation is always on key+0x00
3235 __ aesdeclast(xmm_result, xmm_temp3);
3236 __ movdqu(Address(to, 0), xmm_result); // store the result
3237 __ xorptr(rax, rax); // return 0
3238 __ leave(); // required for proper stackwalking of RuntimeStub frame
3239 __ ret(0);
3240
3241 return start;
3242 }
3243
3244
3245 // Arguments:
3246 //
3247 // Inputs:
3248 // c_rarg0 - source byte array address
3249 // c_rarg1 - destination byte array address
3250 // c_rarg2 - K (key) in little endian int array
3251 // c_rarg3 - r vector byte array address
3252 // c_rarg4 - input length
3253 //
3254 address generate_cipherBlockChaining_encryptAESCrypt() {
3255 assert(UseAES, "need AES instructions and misaligned SSE support");
3256 __ align(CodeEntryAlignment);
3257 StubCodeMark mark(this, "StubRoutines", "cipherBlockChaining_encryptAESCrypt");
3258 address start = __ pc();
3259
3260 Label L_exit, L_key_192_256, L_key_256, L_loopTop_128, L_loopTop_192, L_loopTop_256;
3261 const Register from = c_rarg0; // source array address
3262 const Register to = c_rarg1; // destination array address
3263 const Register key = c_rarg2; // key array address
3264 const Register rvec = c_rarg3; // r byte array initialized from initvector array address
3265 // and left with the results of the last encryption block
3266 #ifndef _WIN64
3267 const Register len_reg = c_rarg4; // src len (must be multiple of blocksize 16)
3268 #else
3269 const Address len_mem(rsp, 6 * wordSize); // length is on stack on Win64
3270 const Register len_reg = r10; // pick the first volatile windows register
3271 #endif
3272 const Register pos = rax;
3273
3274 // xmm register assignments for the loops below
3275 const XMMRegister xmm_result = xmm0;
3276 const XMMRegister xmm_temp = xmm1;
3277 // keys 0-10 preloaded into xmm2-xmm12
3278 const int XMM_REG_NUM_KEY_FIRST = 2;
3279 const int XMM_REG_NUM_KEY_LAST = 15;
3280 const XMMRegister xmm_key0 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST);
3281 const XMMRegister xmm_key10 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+10);
3282 const XMMRegister xmm_key11 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+11);
3283 const XMMRegister xmm_key12 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+12);
3284 const XMMRegister xmm_key13 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+13);
3285
3286 __ enter(); // required for proper stackwalking of RuntimeStub frame
3287
3288 #ifdef _WIN64
3289 // on win64, fill len_reg from stack position
3290 __ movl(len_reg, len_mem);
3291 // save the xmm registers which must be preserved 6-15
3292 __ subptr(rsp, -rsp_after_call_off * wordSize);
3293 for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3294 __ movdqu(xmm_save(i), as_XMMRegister(i));
3295 }
3296 #endif
3297
3298 const XMMRegister xmm_key_shuf_mask = xmm_temp; // used temporarily to swap key bytes up front
3299 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3300 // load up xmm regs xmm2 thru xmm12 with key 0x00 - 0xa0
3301 for (int rnum = XMM_REG_NUM_KEY_FIRST, offset = 0x00; rnum <= XMM_REG_NUM_KEY_FIRST+10; rnum++) {
3302 load_key(as_XMMRegister(rnum), key, offset, xmm_key_shuf_mask);
3303 offset += 0x10;
3304 }
3305 __ movdqu(xmm_result, Address(rvec, 0x00)); // initialize xmm_result with r vec
3306
3307 // now split to different paths depending on the keylen (len in ints of AESCrypt.KLE array (52=192, or 60=256))
3308 __ movl(rax, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3309 __ cmpl(rax, 44);
3310 __ jcc(Assembler::notEqual, L_key_192_256);
3311
3312 // 128 bit code follows here
3313 __ movptr(pos, 0);
3314 __ align(OptoLoopAlignment);
3315
3316 __ BIND(L_loopTop_128);
3317 __ movdqu(xmm_temp, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of input
3318 __ pxor (xmm_result, xmm_temp); // xor with the current r vector
3319 __ pxor (xmm_result, xmm_key0); // do the aes rounds
3320 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_FIRST + 9; rnum++) {
3321 __ aesenc(xmm_result, as_XMMRegister(rnum));
3322 }
3323 __ aesenclast(xmm_result, xmm_key10);
3324 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3325 // no need to store r to memory until we exit
3326 __ addptr(pos, AESBlockSize);
3327 __ subptr(len_reg, AESBlockSize);
3328 __ jcc(Assembler::notEqual, L_loopTop_128);
3329
3330 __ BIND(L_exit);
3331 __ movdqu(Address(rvec, 0), xmm_result); // final value of r stored in rvec of CipherBlockChaining object
3332
3333 #ifdef _WIN64
3334 // restore xmm regs belonging to calling function
3335 for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3336 __ movdqu(as_XMMRegister(i), xmm_save(i));
3337 }
3338 #endif
3339 __ movl(rax, 0); // return 0 (why?)
3340 __ leave(); // required for proper stackwalking of RuntimeStub frame
3341 __ ret(0);
3342
3343 __ BIND(L_key_192_256);
3344 // here rax = len in ints of AESCrypt.KLE array (52=192, or 60=256)
3345 load_key(xmm_key11, key, 0xb0, xmm_key_shuf_mask);
3346 load_key(xmm_key12, key, 0xc0, xmm_key_shuf_mask);
3347 __ cmpl(rax, 52);
3348 __ jcc(Assembler::notEqual, L_key_256);
3349
3350 // 192-bit code follows here (could be changed to use more xmm registers)
3351 __ movptr(pos, 0);
3352 __ align(OptoLoopAlignment);
3353
3354 __ BIND(L_loopTop_192);
3355 __ movdqu(xmm_temp, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of input
3356 __ pxor (xmm_result, xmm_temp); // xor with the current r vector
3357 __ pxor (xmm_result, xmm_key0); // do the aes rounds
3358 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_FIRST + 11; rnum++) {
3359 __ aesenc(xmm_result, as_XMMRegister(rnum));
3360 }
3361 __ aesenclast(xmm_result, xmm_key12);
3362 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3363 // no need to store r to memory until we exit
3364 __ addptr(pos, AESBlockSize);
3365 __ subptr(len_reg, AESBlockSize);
3366 __ jcc(Assembler::notEqual, L_loopTop_192);
3367 __ jmp(L_exit);
3368
3369 __ BIND(L_key_256);
3370 // 256-bit code follows here (could be changed to use more xmm registers)
3371 load_key(xmm_key13, key, 0xd0, xmm_key_shuf_mask);
3372 __ movptr(pos, 0);
3373 __ align(OptoLoopAlignment);
3374
3375 __ BIND(L_loopTop_256);
3376 __ movdqu(xmm_temp, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of input
3377 __ pxor (xmm_result, xmm_temp); // xor with the current r vector
3378 __ pxor (xmm_result, xmm_key0); // do the aes rounds
3379 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_FIRST + 13; rnum++) {
3380 __ aesenc(xmm_result, as_XMMRegister(rnum));
3381 }
3382 load_key(xmm_temp, key, 0xe0);
3383 __ aesenclast(xmm_result, xmm_temp);
3384 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3385 // no need to store r to memory until we exit
3386 __ addptr(pos, AESBlockSize);
3387 __ subptr(len_reg, AESBlockSize);
3388 __ jcc(Assembler::notEqual, L_loopTop_256);
3389 __ jmp(L_exit);
3390
3391 return start;
3392 }
3393
3394
3395
3396 // This is a version of CBC/AES Decrypt which does 4 blocks in a loop at a time
3397 // to hide instruction latency
3398 //
3399 // Arguments:
3400 //
3401 // Inputs:
3402 // c_rarg0 - source byte array address
3403 // c_rarg1 - destination byte array address
3404 // c_rarg2 - K (key) in little endian int array
3405 // c_rarg3 - r vector byte array address
3406 // c_rarg4 - input length
3407 //
3408
3409 address generate_cipherBlockChaining_decryptAESCrypt_Parallel() {
3410 assert(UseAES, "need AES instructions and misaligned SSE support");
3411 __ align(CodeEntryAlignment);
3412 StubCodeMark mark(this, "StubRoutines", "cipherBlockChaining_decryptAESCrypt");
3413 address start = __ pc();
3414
3415 Label L_exit, L_key_192_256, L_key_256;
3416 Label L_singleBlock_loopTop_128, L_multiBlock_loopTop_128;
3417 Label L_singleBlock_loopTop_192, L_singleBlock_loopTop_256;
3418 const Register from = c_rarg0; // source array address
3419 const Register to = c_rarg1; // destination array address
3420 const Register key = c_rarg2; // key array address
3421 const Register rvec = c_rarg3; // r byte array initialized from initvector array address
3422 // and left with the results of the last encryption block
3423 #ifndef _WIN64
3424 const Register len_reg = c_rarg4; // src len (must be multiple of blocksize 16)
3425 #else
3426 const Address len_mem(rsp, 6 * wordSize); // length is on stack on Win64
3427 const Register len_reg = r10; // pick the first volatile windows register
3428 #endif
3429 const Register pos = rax;
3430
3431 // keys 0-10 preloaded into xmm2-xmm12
3432 const int XMM_REG_NUM_KEY_FIRST = 5;
3433 const int XMM_REG_NUM_KEY_LAST = 15;
3434 const XMMRegister xmm_key_first = as_XMMRegister(XMM_REG_NUM_KEY_FIRST);
3435 const XMMRegister xmm_key_last = as_XMMRegister(XMM_REG_NUM_KEY_LAST);
3436
3437 __ enter(); // required for proper stackwalking of RuntimeStub frame
3438
3439 #ifdef _WIN64
3440 // on win64, fill len_reg from stack position
3441 __ movl(len_reg, len_mem);
3442 // save the xmm registers which must be preserved 6-15
3443 __ subptr(rsp, -rsp_after_call_off * wordSize);
3444 for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3445 __ movdqu(xmm_save(i), as_XMMRegister(i));
3446 }
3447 #endif
3448 // the java expanded key ordering is rotated one position from what we want
3449 // so we start from 0x10 here and hit 0x00 last
3450 const XMMRegister xmm_key_shuf_mask = xmm1; // used temporarily to swap key bytes up front
3451 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3452 // load up xmm regs 5 thru 15 with key 0x10 - 0xa0 - 0x00
3453 for (int rnum = XMM_REG_NUM_KEY_FIRST, offset = 0x10; rnum < XMM_REG_NUM_KEY_LAST; rnum++) {
3454 load_key(as_XMMRegister(rnum), key, offset, xmm_key_shuf_mask);
3455 offset += 0x10;
3456 }
3457 load_key(xmm_key_last, key, 0x00, xmm_key_shuf_mask);
3458
3459 const XMMRegister xmm_prev_block_cipher = xmm1; // holds cipher of previous block
3460
3461 // registers holding the four results in the parallelized loop
3462 const XMMRegister xmm_result0 = xmm0;
3463 const XMMRegister xmm_result1 = xmm2;
3464 const XMMRegister xmm_result2 = xmm3;
3465 const XMMRegister xmm_result3 = xmm4;
3466
3467 __ movdqu(xmm_prev_block_cipher, Address(rvec, 0x00)); // initialize with initial rvec
3468
3469 // now split to different paths depending on the keylen (len in ints of AESCrypt.KLE array (52=192, or 60=256))
3470 __ movl(rax, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3471 __ cmpl(rax, 44);
3472 __ jcc(Assembler::notEqual, L_key_192_256);
3473
3474
3475 // 128-bit code follows here, parallelized
3476 __ movptr(pos, 0);
3477 __ align(OptoLoopAlignment);
3478 __ BIND(L_multiBlock_loopTop_128);
3479 __ cmpptr(len_reg, 4*AESBlockSize); // see if at least 4 blocks left
3480 __ jcc(Assembler::less, L_singleBlock_loopTop_128);
3481
3482 __ movdqu(xmm_result0, Address(from, pos, Address::times_1, 0*AESBlockSize)); // get next 4 blocks into xmmresult registers
3483 __ movdqu(xmm_result1, Address(from, pos, Address::times_1, 1*AESBlockSize));
3484 __ movdqu(xmm_result2, Address(from, pos, Address::times_1, 2*AESBlockSize));
3485 __ movdqu(xmm_result3, Address(from, pos, Address::times_1, 3*AESBlockSize));
3486
3487 #define DoFour(opc, src_reg) \
3488 __ opc(xmm_result0, src_reg); \
3489 __ opc(xmm_result1, src_reg); \
3490 __ opc(xmm_result2, src_reg); \
3491 __ opc(xmm_result3, src_reg);
3492
3493 DoFour(pxor, xmm_key_first);
3494 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3495 DoFour(aesdec, as_XMMRegister(rnum));
3496 }
3497 DoFour(aesdeclast, xmm_key_last);
3498 // for each result, xor with the r vector of previous cipher block
3499 __ pxor(xmm_result0, xmm_prev_block_cipher);
3500 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 0*AESBlockSize));
3501 __ pxor(xmm_result1, xmm_prev_block_cipher);
3502 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 1*AESBlockSize));
3503 __ pxor(xmm_result2, xmm_prev_block_cipher);
3504 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 2*AESBlockSize));
3505 __ pxor(xmm_result3, xmm_prev_block_cipher);
3506 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 3*AESBlockSize)); // this will carry over to next set of blocks
3507
3508 __ movdqu(Address(to, pos, Address::times_1, 0*AESBlockSize), xmm_result0); // store 4 results into the next 64 bytes of output
3509 __ movdqu(Address(to, pos, Address::times_1, 1*AESBlockSize), xmm_result1);
3510 __ movdqu(Address(to, pos, Address::times_1, 2*AESBlockSize), xmm_result2);
3511 __ movdqu(Address(to, pos, Address::times_1, 3*AESBlockSize), xmm_result3);
3512
3513 __ addptr(pos, 4*AESBlockSize);
3514 __ subptr(len_reg, 4*AESBlockSize);
3515 __ jmp(L_multiBlock_loopTop_128);
3516
3517 // registers used in the non-parallelized loops
3518 // xmm register assignments for the loops below
3519 const XMMRegister xmm_result = xmm0;
3520 const XMMRegister xmm_prev_block_cipher_save = xmm2;
3521 const XMMRegister xmm_key11 = xmm3;
3522 const XMMRegister xmm_key12 = xmm4;
3523 const XMMRegister xmm_temp = xmm4;
3524
3525 __ align(OptoLoopAlignment);
3526 __ BIND(L_singleBlock_loopTop_128);
3527 __ cmpptr(len_reg, 0); // any blocks left??
3528 __ jcc(Assembler::equal, L_exit);
3529 __ movdqu(xmm_result, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of cipher input
3530 __ movdqa(xmm_prev_block_cipher_save, xmm_result); // save for next r vector
3531 __ pxor (xmm_result, xmm_key_first); // do the aes dec rounds
3532 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3533 __ aesdec(xmm_result, as_XMMRegister(rnum));
3534 }
3535 __ aesdeclast(xmm_result, xmm_key_last);
3536 __ pxor (xmm_result, xmm_prev_block_cipher); // xor with the current r vector
3537 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3538 // no need to store r to memory until we exit
3539 __ movdqa(xmm_prev_block_cipher, xmm_prev_block_cipher_save); // set up next r vector with cipher input from this block
3540
3541 __ addptr(pos, AESBlockSize);
3542 __ subptr(len_reg, AESBlockSize);
3543 __ jmp(L_singleBlock_loopTop_128);
3544
3545
3546 __ BIND(L_exit);
3547 __ movdqu(Address(rvec, 0), xmm_prev_block_cipher); // final value of r stored in rvec of CipherBlockChaining object
3548 #ifdef _WIN64
3549 // restore regs belonging to calling function
3550 for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3551 __ movdqu(as_XMMRegister(i), xmm_save(i));
3552 }
3553 #endif
3554 __ movl(rax, 0); // return 0 (why?)
3555 __ leave(); // required for proper stackwalking of RuntimeStub frame
3556 __ ret(0);
3557
3558
3559 __ BIND(L_key_192_256);
3560 // here rax = len in ints of AESCrypt.KLE array (52=192, or 60=256)
3561 load_key(xmm_key11, key, 0xb0);
3562 __ cmpl(rax, 52);
3563 __ jcc(Assembler::notEqual, L_key_256);
3564
3565 // 192-bit code follows here (could be optimized to use parallelism)
3566 load_key(xmm_key12, key, 0xc0); // 192-bit key goes up to c0
3567 __ movptr(pos, 0);
3568 __ align(OptoLoopAlignment);
3569
3570 __ BIND(L_singleBlock_loopTop_192);
3571 __ movdqu(xmm_result, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of cipher input
3572 __ movdqa(xmm_prev_block_cipher_save, xmm_result); // save for next r vector
3573 __ pxor (xmm_result, xmm_key_first); // do the aes dec rounds
3574 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3575 __ aesdec(xmm_result, as_XMMRegister(rnum));
3576 }
3577 __ aesdec(xmm_result, xmm_key11);
3578 __ aesdec(xmm_result, xmm_key12);
3579 __ aesdeclast(xmm_result, xmm_key_last); // xmm15 always came from key+0
3580 __ pxor (xmm_result, xmm_prev_block_cipher); // xor with the current r vector
3581 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3582 // no need to store r to memory until we exit
3583 __ movdqa(xmm_prev_block_cipher, xmm_prev_block_cipher_save); // set up next r vector with cipher input from this block
3584 __ addptr(pos, AESBlockSize);
3585 __ subptr(len_reg, AESBlockSize);
3586 __ jcc(Assembler::notEqual,L_singleBlock_loopTop_192);
3587 __ jmp(L_exit);
3588
3589 __ BIND(L_key_256);
3590 // 256-bit code follows here (could be optimized to use parallelism)
3591 __ movptr(pos, 0);
3592 __ align(OptoLoopAlignment);
3593
3594 __ BIND(L_singleBlock_loopTop_256);
3595 __ movdqu(xmm_result, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of cipher input
3596 __ movdqa(xmm_prev_block_cipher_save, xmm_result); // save for next r vector
3597 __ pxor (xmm_result, xmm_key_first); // do the aes dec rounds
3598 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3599 __ aesdec(xmm_result, as_XMMRegister(rnum));
3600 }
3601 __ aesdec(xmm_result, xmm_key11);
3602 load_key(xmm_temp, key, 0xc0);
3603 __ aesdec(xmm_result, xmm_temp);
3604 load_key(xmm_temp, key, 0xd0);
3605 __ aesdec(xmm_result, xmm_temp);
3606 load_key(xmm_temp, key, 0xe0); // 256-bit key goes up to e0
3607 __ aesdec(xmm_result, xmm_temp);
3608 __ aesdeclast(xmm_result, xmm_key_last); // xmm15 came from key+0
3609 __ pxor (xmm_result, xmm_prev_block_cipher); // xor with the current r vector
3610 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3611 // no need to store r to memory until we exit
3612 __ movdqa(xmm_prev_block_cipher, xmm_prev_block_cipher_save); // set up next r vector with cipher input from this block
3613 __ addptr(pos, AESBlockSize);
3614 __ subptr(len_reg, AESBlockSize);
3615 __ jcc(Assembler::notEqual,L_singleBlock_loopTop_256);
3616 __ jmp(L_exit);
3617
3618 return start;
3619 }
3620
3621
3622
3623 #undef __
3624 #define __ masm->
3625
3626 // Continuation point for throwing of implicit exceptions that are
3627 // not handled in the current activation. Fabricates an exception
3628 // oop and initiates normal exception dispatching in this
3629 // frame. Since we need to preserve callee-saved values (currently
3630 // only for C2, but done for C1 as well) we need a callee-saved oop
3631 // map and therefore have to make these stubs into RuntimeStubs
3632 // rather than BufferBlobs. If the compiler needs all registers to
3633 // be preserved between the fault point and the exception handler
3634 // then it must assume responsibility for that in
3635 // AbstractCompiler::continuation_for_implicit_null_exception or
3636 // continuation_for_implicit_division_by_zero_exception. All other
3637 // implicit exceptions (e.g., NullPointerException or
3638 // AbstractMethodError on entry) are either at call sites or
3639 // otherwise assume that stack unwinding will be initiated, so
3640 // caller saved registers were assumed volatile in the compiler.
3641 address generate_throw_exception(const char* name,
3642 address runtime_entry,
3643 Register arg1 = noreg,
3644 Register arg2 = noreg) {
3645 // Information about frame layout at time of blocking runtime call.
3646 // Note that we only have to preserve callee-saved registers since
3647 // the compilers are responsible for supplying a continuation point
3648 // if they expect all registers to be preserved.
3649 enum layout {
3650 rbp_off = frame::arg_reg_save_area_bytes/BytesPerInt,
3651 rbp_off2,
3652 return_off,
3653 return_off2,
3654 framesize // inclusive of return address
3655 };
3656
3657 int insts_size = 512;
3658 int locs_size = 64;
3659
3660 CodeBuffer code(name, insts_size, locs_size);
3661 OopMapSet* oop_maps = new OopMapSet();
3662 MacroAssembler* masm = new MacroAssembler(&code);
3663
3664 address start = __ pc();
3665
3666 // This is an inlined and slightly modified version of call_VM
3667 // which has the ability to fetch the return PC out of
3668 // thread-local storage and also sets up last_Java_sp slightly
3669 // differently than the real call_VM
3670
3671 __ enter(); // required for proper stackwalking of RuntimeStub frame
3672
3673 assert(is_even(framesize/2), "sp not 16-byte aligned");
3674
3675 // return address and rbp are already in place
3676 __ subptr(rsp, (framesize-4) << LogBytesPerInt); // prolog
3677
3678 int frame_complete = __ pc() - start;
3679
3680 // Set up last_Java_sp and last_Java_fp
3681 address the_pc = __ pc();
3682 __ set_last_Java_frame(rsp, rbp, the_pc);
3683 __ andptr(rsp, -(StackAlignmentInBytes)); // Align stack
3684
3685 // Call runtime
3686 if (arg1 != noreg) {
3687 assert(arg2 != c_rarg1, "clobbered");
3688 __ movptr(c_rarg1, arg1);
3689 }
3690 if (arg2 != noreg) {
3691 __ movptr(c_rarg2, arg2);
3692 }
3693 __ movptr(c_rarg0, r15_thread);
3694 BLOCK_COMMENT("call runtime_entry");
3695 __ call(RuntimeAddress(runtime_entry));
3696
3697 // Generate oop map
3698 OopMap* map = new OopMap(framesize, 0);
3699
3700 oop_maps->add_gc_map(the_pc - start, map);
3701
3702 __ reset_last_Java_frame(true, true);
3703
3704 __ leave(); // required for proper stackwalking of RuntimeStub frame
3705
3706 // check for pending exceptions
3707 #ifdef ASSERT
3708 Label L;
3709 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()),
3710 (int32_t) NULL_WORD);
3711 __ jcc(Assembler::notEqual, L);
3712 __ should_not_reach_here();
3713 __ bind(L);
3714 #endif // ASSERT
3715 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3716
3717
3718 // codeBlob framesize is in words (not VMRegImpl::slot_size)
3719 RuntimeStub* stub =
3720 RuntimeStub::new_runtime_stub(name,
3721 &code,
3722 frame_complete,
3723 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3724 oop_maps, false);
3725 return stub->entry_point();
3726 }
3727
3728 // Initialization
3729 void generate_initial() {
3730 // Generates all stubs and initializes the entry points
3731
3732 // This platform-specific stub is needed by generate_call_stub()
3733 StubRoutines::x86::_mxcsr_std = generate_fp_mask("mxcsr_std", 0x0000000000001F80);
3734
3735 // entry points that exist in all platforms Note: This is code
3736 // that could be shared among different platforms - however the
3737 // benefit seems to be smaller than the disadvantage of having a
3738 // much more complicated generator structure. See also comment in
3739 // stubRoutines.hpp.
3740
3741 StubRoutines::_forward_exception_entry = generate_forward_exception();
3742
3743 StubRoutines::_call_stub_entry =
3744 generate_call_stub(StubRoutines::_call_stub_return_address);
3745
3746 // is referenced by megamorphic call
3747 StubRoutines::_catch_exception_entry = generate_catch_exception();
3748
3749 // atomic calls
3750 StubRoutines::_atomic_xchg_entry = generate_atomic_xchg();
3751 StubRoutines::_atomic_xchg_ptr_entry = generate_atomic_xchg_ptr();
3752 StubRoutines::_atomic_cmpxchg_entry = generate_atomic_cmpxchg();
3753 StubRoutines::_atomic_cmpxchg_long_entry = generate_atomic_cmpxchg_long();
3754 StubRoutines::_atomic_add_entry = generate_atomic_add();
3755 StubRoutines::_atomic_add_ptr_entry = generate_atomic_add_ptr();
3756 StubRoutines::_fence_entry = generate_orderaccess_fence();
3757
3758 StubRoutines::_handler_for_unsafe_access_entry =
3759 generate_handler_for_unsafe_access();
3760
3761 // platform dependent
3762 StubRoutines::x86::_get_previous_fp_entry = generate_get_previous_fp();
3763 StubRoutines::x86::_get_previous_sp_entry = generate_get_previous_sp();
3764
3765 StubRoutines::x86::_verify_mxcsr_entry = generate_verify_mxcsr();
3766
3767 // Build this early so it's available for the interpreter.
3768 StubRoutines::_throw_StackOverflowError_entry =
3769 generate_throw_exception("StackOverflowError throw_exception",
3770 CAST_FROM_FN_PTR(address,
3771 SharedRuntime::
3772 throw_StackOverflowError));
3773 }
3774
3775 void generate_all() {
3776 // Generates all stubs and initializes the entry points
3777
3778 // These entry points require SharedInfo::stack0 to be set up in
3779 // non-core builds and need to be relocatable, so they each
3780 // fabricate a RuntimeStub internally.
3781 StubRoutines::_throw_AbstractMethodError_entry =
3782 generate_throw_exception("AbstractMethodError throw_exception",
3783 CAST_FROM_FN_PTR(address,
3784 SharedRuntime::
3785 throw_AbstractMethodError));
3786
3787 StubRoutines::_throw_IncompatibleClassChangeError_entry =
3788 generate_throw_exception("IncompatibleClassChangeError throw_exception",
3789 CAST_FROM_FN_PTR(address,
3790 SharedRuntime::
3791 throw_IncompatibleClassChangeError));
3792
3793 StubRoutines::_throw_NullPointerException_at_call_entry =
3794 generate_throw_exception("NullPointerException at call throw_exception",
3795 CAST_FROM_FN_PTR(address,
3796 SharedRuntime::
3797 throw_NullPointerException_at_call));
3798
3799 // entry points that are platform specific
3800 StubRoutines::x86::_f2i_fixup = generate_f2i_fixup();
3801 StubRoutines::x86::_f2l_fixup = generate_f2l_fixup();
3802 StubRoutines::x86::_d2i_fixup = generate_d2i_fixup();
3803 StubRoutines::x86::_d2l_fixup = generate_d2l_fixup();
3804
3805 StubRoutines::x86::_float_sign_mask = generate_fp_mask("float_sign_mask", 0x7FFFFFFF7FFFFFFF);
3806 StubRoutines::x86::_float_sign_flip = generate_fp_mask("float_sign_flip", 0x8000000080000000);
3807 StubRoutines::x86::_double_sign_mask = generate_fp_mask("double_sign_mask", 0x7FFFFFFFFFFFFFFF);
3808 StubRoutines::x86::_double_sign_flip = generate_fp_mask("double_sign_flip", 0x8000000000000000);
3809
3810 // support for verify_oop (must happen after universe_init)
3811 StubRoutines::_verify_oop_subroutine_entry = generate_verify_oop();
3812
3813 // arraycopy stubs used by compilers
3814 generate_arraycopy_stubs();
3815
3816 generate_math_stubs();
3817
3818 // don't bother generating these AES intrinsic stubs unless global flag is set
3819 if (UseAESIntrinsics) {
3820 StubRoutines::x86::_key_shuffle_mask_addr = generate_key_shuffle_mask(); // needed by the others
3821
3822 StubRoutines::_aescrypt_encryptBlock = generate_aescrypt_encryptBlock();
3823 StubRoutines::_aescrypt_decryptBlock = generate_aescrypt_decryptBlock();
3824 StubRoutines::_cipherBlockChaining_encryptAESCrypt = generate_cipherBlockChaining_encryptAESCrypt();
3825 StubRoutines::_cipherBlockChaining_decryptAESCrypt = generate_cipherBlockChaining_decryptAESCrypt_Parallel();
3826 }
3827 }
3828
3829 public:
3830 StubGenerator(CodeBuffer* code, bool all) : StubCodeGenerator(code) {
3831 if (all) {
3832 generate_all();
3833 } else {
3834 generate_initial();
3835 }
3836 }
3837 }; // end class declaration
3838
3839 void StubGenerator_generate(CodeBuffer* code, bool all) {
3840 StubGenerator g(code, all);
3841 }