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 // count - elements count
1249 // scratch - scratch register
1250 //
1251 // The input registers are overwritten.
1252 //
1253 void gen_write_ref_array_post_barrier(Register start, Register count, Register scratch) {
1254 assert_different_registers(start, count, scratch);
1255 BarrierSet* bs = Universe::heap()->barrier_set();
1256 switch (bs->kind()) {
1257 case BarrierSet::G1SATBCT:
1258 case BarrierSet::G1SATBCTLogging:
1259 {
1260 __ pusha(); // push registers (overkill)
1261 if (c_rarg0 == count) { // On win64 c_rarg0 == rcx
1262 assert_different_registers(c_rarg1, start);
1263 __ mov(c_rarg1, count);
1264 __ mov(c_rarg0, start);
1265 } else {
1266 assert_different_registers(c_rarg0, count);
1267 __ mov(c_rarg0, start);
1268 __ mov(c_rarg1, count);
1269 }
1270 __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_post), 2);
1271 __ popa();
1272 }
1273 break;
1274 case BarrierSet::CardTableModRef:
1275 case BarrierSet::CardTableExtension:
1276 {
1277 CardTableModRefBS* ct = (CardTableModRefBS*)bs;
1278 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
1279
1280 Label L_loop;
1281 const Register end = count;
1282
1283 __ leaq(end, Address(start, count, TIMES_OOP, 0)); // end == start+count*oop_size
1284 __ subptr(end, BytesPerHeapOop); // end - 1 to make inclusive
1285 __ shrptr(start, CardTableModRefBS::card_shift);
1286 __ shrptr(end, CardTableModRefBS::card_shift);
1287 __ subptr(end, start); // end --> cards count
1288
1289 int64_t disp = (int64_t) ct->byte_map_base;
1290 __ mov64(scratch, disp);
1291 __ addptr(start, scratch);
1292 __ BIND(L_loop);
1293 __ movb(Address(start, count, Address::times_1), 0);
1294 __ decrement(count);
1295 __ jcc(Assembler::greaterEqual, L_loop);
1296 }
1297 break;
1298 default:
1299 ShouldNotReachHere();
1300
1301 }
1302 }
1303
1304
1305 // Copy big chunks forward
1306 //
1307 // Inputs:
1308 // end_from - source arrays end address
1309 // end_to - destination array end address
1310 // qword_count - 64-bits element count, negative
1311 // to - scratch
1312 // L_copy_bytes - entry label
1313 // L_copy_8_bytes - exit label
1314 //
1315 void copy_bytes_forward(Register end_from, Register end_to,
1316 Register qword_count, Register to,
1317 Label& L_copy_bytes, Label& L_copy_8_bytes) {
1318 DEBUG_ONLY(__ stop("enter at entry label, not here"));
1319 Label L_loop;
1320 __ align(OptoLoopAlignment);
1321 if (UseUnalignedLoadStores) {
1322 Label L_end;
1323 // Copy 64-bytes per iteration
1324 __ BIND(L_loop);
1325 if (UseAVX >= 2) {
1326 __ vmovdqu(xmm0, Address(end_from, qword_count, Address::times_8, -56));
1327 __ vmovdqu(Address(end_to, qword_count, Address::times_8, -56), xmm0);
1328 __ vmovdqu(xmm1, Address(end_from, qword_count, Address::times_8, -24));
1329 __ vmovdqu(Address(end_to, qword_count, Address::times_8, -24), xmm1);
1330 } else {
1331 __ movdqu(xmm0, Address(end_from, qword_count, Address::times_8, -56));
1332 __ movdqu(Address(end_to, qword_count, Address::times_8, -56), xmm0);
1333 __ movdqu(xmm1, Address(end_from, qword_count, Address::times_8, -40));
1334 __ movdqu(Address(end_to, qword_count, Address::times_8, -40), xmm1);
1335 __ movdqu(xmm2, Address(end_from, qword_count, Address::times_8, -24));
1336 __ movdqu(Address(end_to, qword_count, Address::times_8, -24), xmm2);
1337 __ movdqu(xmm3, Address(end_from, qword_count, Address::times_8, - 8));
1338 __ movdqu(Address(end_to, qword_count, Address::times_8, - 8), xmm3);
1339 }
1340 __ BIND(L_copy_bytes);
1341 __ addptr(qword_count, 8);
1342 __ jcc(Assembler::lessEqual, L_loop);
1343 __ subptr(qword_count, 4); // sub(8) and add(4)
1344 __ jccb(Assembler::greater, L_end);
1345 // Copy trailing 32 bytes
1346 if (UseAVX >= 2) {
1347 __ vmovdqu(xmm0, Address(end_from, qword_count, Address::times_8, -24));
1348 __ vmovdqu(Address(end_to, qword_count, Address::times_8, -24), xmm0);
1349 } else {
1350 __ movdqu(xmm0, Address(end_from, qword_count, Address::times_8, -24));
1351 __ movdqu(Address(end_to, qword_count, Address::times_8, -24), xmm0);
1352 __ movdqu(xmm1, Address(end_from, qword_count, Address::times_8, - 8));
1353 __ movdqu(Address(end_to, qword_count, Address::times_8, - 8), xmm1);
1354 }
1355 __ addptr(qword_count, 4);
1356 __ BIND(L_end);
1357 if (UseAVX >= 2) {
1358 // clean upper bits of YMM registers
1359 __ vzeroupper();
1360 }
1361 } else {
1362 // Copy 32-bytes per iteration
1363 __ BIND(L_loop);
1364 __ movq(to, Address(end_from, qword_count, Address::times_8, -24));
1365 __ movq(Address(end_to, qword_count, Address::times_8, -24), to);
1366 __ movq(to, Address(end_from, qword_count, Address::times_8, -16));
1367 __ movq(Address(end_to, qword_count, Address::times_8, -16), to);
1368 __ movq(to, Address(end_from, qword_count, Address::times_8, - 8));
1369 __ movq(Address(end_to, qword_count, Address::times_8, - 8), to);
1370 __ movq(to, Address(end_from, qword_count, Address::times_8, - 0));
1371 __ movq(Address(end_to, qword_count, Address::times_8, - 0), to);
1372
1373 __ BIND(L_copy_bytes);
1374 __ addptr(qword_count, 4);
1375 __ jcc(Assembler::lessEqual, L_loop);
1376 }
1377 __ subptr(qword_count, 4);
1378 __ jcc(Assembler::less, L_copy_8_bytes); // Copy trailing qwords
1379 }
1380
1381 // Copy big chunks backward
1382 //
1383 // Inputs:
1384 // from - source arrays address
1385 // dest - destination array address
1386 // qword_count - 64-bits element count
1387 // to - scratch
1388 // L_copy_bytes - entry label
1389 // L_copy_8_bytes - exit label
1390 //
1391 void copy_bytes_backward(Register from, Register dest,
1392 Register qword_count, Register to,
1393 Label& L_copy_bytes, Label& L_copy_8_bytes) {
1394 DEBUG_ONLY(__ stop("enter at entry label, not here"));
1395 Label L_loop;
1396 __ align(OptoLoopAlignment);
1397 if (UseUnalignedLoadStores) {
1398 Label L_end;
1399 // Copy 64-bytes per iteration
1400 __ BIND(L_loop);
1401 if (UseAVX >= 2) {
1402 __ vmovdqu(xmm0, Address(from, qword_count, Address::times_8, 32));
1403 __ vmovdqu(Address(dest, qword_count, Address::times_8, 32), xmm0);
1404 __ vmovdqu(xmm1, Address(from, qword_count, Address::times_8, 0));
1405 __ vmovdqu(Address(dest, qword_count, Address::times_8, 0), xmm1);
1406 } else {
1407 __ movdqu(xmm0, Address(from, qword_count, Address::times_8, 48));
1408 __ movdqu(Address(dest, qword_count, Address::times_8, 48), xmm0);
1409 __ movdqu(xmm1, Address(from, qword_count, Address::times_8, 32));
1410 __ movdqu(Address(dest, qword_count, Address::times_8, 32), xmm1);
1411 __ movdqu(xmm2, Address(from, qword_count, Address::times_8, 16));
1412 __ movdqu(Address(dest, qword_count, Address::times_8, 16), xmm2);
1413 __ movdqu(xmm3, Address(from, qword_count, Address::times_8, 0));
1414 __ movdqu(Address(dest, qword_count, Address::times_8, 0), xmm3);
1415 }
1416 __ BIND(L_copy_bytes);
1417 __ subptr(qword_count, 8);
1418 __ jcc(Assembler::greaterEqual, L_loop);
1419
1420 __ addptr(qword_count, 4); // add(8) and sub(4)
1421 __ jccb(Assembler::less, L_end);
1422 // Copy trailing 32 bytes
1423 if (UseAVX >= 2) {
1424 __ vmovdqu(xmm0, Address(from, qword_count, Address::times_8, 0));
1425 __ vmovdqu(Address(dest, qword_count, Address::times_8, 0), xmm0);
1426 } else {
1427 __ movdqu(xmm0, Address(from, qword_count, Address::times_8, 16));
1428 __ movdqu(Address(dest, qword_count, Address::times_8, 16), xmm0);
1429 __ movdqu(xmm1, Address(from, qword_count, Address::times_8, 0));
1430 __ movdqu(Address(dest, qword_count, Address::times_8, 0), xmm1);
1431 }
1432 __ subptr(qword_count, 4);
1433 __ BIND(L_end);
1434 if (UseAVX >= 2) {
1435 // clean upper bits of YMM registers
1436 __ vzeroupper();
1437 }
1438 } else {
1439 // Copy 32-bytes per iteration
1440 __ BIND(L_loop);
1441 __ movq(to, Address(from, qword_count, Address::times_8, 24));
1442 __ movq(Address(dest, qword_count, Address::times_8, 24), to);
1443 __ movq(to, Address(from, qword_count, Address::times_8, 16));
1444 __ movq(Address(dest, qword_count, Address::times_8, 16), to);
1445 __ movq(to, Address(from, qword_count, Address::times_8, 8));
1446 __ movq(Address(dest, qword_count, Address::times_8, 8), to);
1447 __ movq(to, Address(from, qword_count, Address::times_8, 0));
1448 __ movq(Address(dest, qword_count, Address::times_8, 0), to);
1449
1450 __ BIND(L_copy_bytes);
1451 __ subptr(qword_count, 4);
1452 __ jcc(Assembler::greaterEqual, L_loop);
1453 }
1454 __ addptr(qword_count, 4);
1455 __ jcc(Assembler::greater, L_copy_8_bytes); // Copy trailing qwords
1456 }
1457
1458
1459 // Arguments:
1460 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1461 // ignored
1462 // name - stub name string
1463 //
1464 // Inputs:
1465 // c_rarg0 - source array address
1466 // c_rarg1 - destination array address
1467 // c_rarg2 - element count, treated as ssize_t, can be zero
1468 //
1469 // If 'from' and/or 'to' are aligned on 4-, 2-, or 1-byte boundaries,
1470 // we let the hardware handle it. The one to eight bytes within words,
1471 // dwords or qwords that span cache line boundaries will still be loaded
1472 // and stored atomically.
1473 //
1474 // Side Effects:
1475 // disjoint_byte_copy_entry is set to the no-overlap entry point
1476 // used by generate_conjoint_byte_copy().
1477 //
1478 address generate_disjoint_byte_copy(bool aligned, address* entry, const char *name) {
1479 __ align(CodeEntryAlignment);
1480 StubCodeMark mark(this, "StubRoutines", name);
1481 address start = __ pc();
1482
1483 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes, L_copy_2_bytes;
1484 Label L_copy_byte, L_exit;
1485 const Register from = rdi; // source array address
1486 const Register to = rsi; // destination array address
1487 const Register count = rdx; // elements count
1488 const Register byte_count = rcx;
1489 const Register qword_count = count;
1490 const Register end_from = from; // source array end address
1491 const Register end_to = to; // destination array end address
1492 // End pointers are inclusive, and if count is not zero they point
1493 // to the last unit copied: end_to[0] := end_from[0]
1494
1495 __ enter(); // required for proper stackwalking of RuntimeStub frame
1496 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1497
1498 if (entry != NULL) {
1499 *entry = __ pc();
1500 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1501 BLOCK_COMMENT("Entry:");
1502 }
1503
1504 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1505 // r9 and r10 may be used to save non-volatile registers
1506
1507 // 'from', 'to' and 'count' are now valid
1508 __ movptr(byte_count, count);
1509 __ shrptr(count, 3); // count => qword_count
1510
1511 // Copy from low to high addresses. Use 'to' as scratch.
1512 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
1513 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
1514 __ negptr(qword_count); // make the count negative
1515 __ jmp(L_copy_bytes);
1516
1517 // Copy trailing qwords
1518 __ BIND(L_copy_8_bytes);
1519 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
1520 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
1521 __ increment(qword_count);
1522 __ jcc(Assembler::notZero, L_copy_8_bytes);
1523
1524 // Check for and copy trailing dword
1525 __ BIND(L_copy_4_bytes);
1526 __ testl(byte_count, 4);
1527 __ jccb(Assembler::zero, L_copy_2_bytes);
1528 __ movl(rax, Address(end_from, 8));
1529 __ movl(Address(end_to, 8), rax);
1530
1531 __ addptr(end_from, 4);
1532 __ addptr(end_to, 4);
1533
1534 // Check for and copy trailing word
1535 __ BIND(L_copy_2_bytes);
1536 __ testl(byte_count, 2);
1537 __ jccb(Assembler::zero, L_copy_byte);
1538 __ movw(rax, Address(end_from, 8));
1539 __ movw(Address(end_to, 8), rax);
1540
1541 __ addptr(end_from, 2);
1542 __ addptr(end_to, 2);
1543
1544 // Check for and copy trailing byte
1545 __ BIND(L_copy_byte);
1546 __ testl(byte_count, 1);
1547 __ jccb(Assembler::zero, L_exit);
1548 __ movb(rax, Address(end_from, 8));
1549 __ movb(Address(end_to, 8), rax);
1550
1551 __ BIND(L_exit);
1552 restore_arg_regs();
1553 inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); // Update counter after rscratch1 is free
1554 __ xorptr(rax, rax); // return 0
1555 __ leave(); // required for proper stackwalking of RuntimeStub frame
1556 __ ret(0);
1557
1558 // Copy in multi-bytes chunks
1559 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1560 __ jmp(L_copy_4_bytes);
1561
1562 return start;
1563 }
1564
1565 // Arguments:
1566 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1567 // ignored
1568 // name - stub name string
1569 //
1570 // Inputs:
1571 // c_rarg0 - source array address
1572 // c_rarg1 - destination array address
1573 // c_rarg2 - element count, treated as ssize_t, can be zero
1574 //
1575 // If 'from' and/or 'to' are aligned on 4-, 2-, or 1-byte boundaries,
1576 // we let the hardware handle it. The one to eight bytes within words,
1577 // dwords or qwords that span cache line boundaries will still be loaded
1578 // and stored atomically.
1579 //
1580 address generate_conjoint_byte_copy(bool aligned, address nooverlap_target,
1581 address* entry, const char *name) {
1582 __ align(CodeEntryAlignment);
1583 StubCodeMark mark(this, "StubRoutines", name);
1584 address start = __ pc();
1585
1586 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes, L_copy_2_bytes;
1587 const Register from = rdi; // source array address
1588 const Register to = rsi; // destination array address
1589 const Register count = rdx; // elements count
1590 const Register byte_count = rcx;
1591 const Register qword_count = count;
1592
1593 __ enter(); // required for proper stackwalking of RuntimeStub frame
1594 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1595
1596 if (entry != NULL) {
1597 *entry = __ pc();
1598 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1599 BLOCK_COMMENT("Entry:");
1600 }
1601
1602 array_overlap_test(nooverlap_target, Address::times_1);
1603 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1604 // r9 and r10 may be used to save non-volatile registers
1605
1606 // 'from', 'to' and 'count' are now valid
1607 __ movptr(byte_count, count);
1608 __ shrptr(count, 3); // count => qword_count
1609
1610 // Copy from high to low addresses.
1611
1612 // Check for and copy trailing byte
1613 __ testl(byte_count, 1);
1614 __ jcc(Assembler::zero, L_copy_2_bytes);
1615 __ movb(rax, Address(from, byte_count, Address::times_1, -1));
1616 __ movb(Address(to, byte_count, Address::times_1, -1), rax);
1617 __ decrement(byte_count); // Adjust for possible trailing word
1618
1619 // Check for and copy trailing word
1620 __ BIND(L_copy_2_bytes);
1621 __ testl(byte_count, 2);
1622 __ jcc(Assembler::zero, L_copy_4_bytes);
1623 __ movw(rax, Address(from, byte_count, Address::times_1, -2));
1624 __ movw(Address(to, byte_count, Address::times_1, -2), rax);
1625
1626 // Check for and copy trailing dword
1627 __ BIND(L_copy_4_bytes);
1628 __ testl(byte_count, 4);
1629 __ jcc(Assembler::zero, L_copy_bytes);
1630 __ movl(rax, Address(from, qword_count, Address::times_8));
1631 __ movl(Address(to, qword_count, Address::times_8), rax);
1632 __ jmp(L_copy_bytes);
1633
1634 // Copy trailing qwords
1635 __ BIND(L_copy_8_bytes);
1636 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
1637 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
1638 __ decrement(qword_count);
1639 __ jcc(Assembler::notZero, L_copy_8_bytes);
1640
1641 restore_arg_regs();
1642 inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); // Update counter after rscratch1 is free
1643 __ xorptr(rax, rax); // return 0
1644 __ leave(); // required for proper stackwalking of RuntimeStub frame
1645 __ ret(0);
1646
1647 // Copy in multi-bytes chunks
1648 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1649
1650 restore_arg_regs();
1651 inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); // Update counter after rscratch1 is free
1652 __ xorptr(rax, rax); // return 0
1653 __ leave(); // required for proper stackwalking of RuntimeStub frame
1654 __ ret(0);
1655
1656 return start;
1657 }
1658
1659 // Arguments:
1660 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1661 // ignored
1662 // name - stub name string
1663 //
1664 // Inputs:
1665 // c_rarg0 - source array address
1666 // c_rarg1 - destination array address
1667 // c_rarg2 - element count, treated as ssize_t, can be zero
1668 //
1669 // If 'from' and/or 'to' are aligned on 4- or 2-byte boundaries, we
1670 // let the hardware handle it. The two or four words within dwords
1671 // or qwords that span cache line boundaries will still be loaded
1672 // and stored atomically.
1673 //
1674 // Side Effects:
1675 // disjoint_short_copy_entry is set to the no-overlap entry point
1676 // used by generate_conjoint_short_copy().
1677 //
1678 address generate_disjoint_short_copy(bool aligned, address *entry, const char *name) {
1679 __ align(CodeEntryAlignment);
1680 StubCodeMark mark(this, "StubRoutines", name);
1681 address start = __ pc();
1682
1683 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes,L_copy_2_bytes,L_exit;
1684 const Register from = rdi; // source array address
1685 const Register to = rsi; // destination array address
1686 const Register count = rdx; // elements count
1687 const Register word_count = rcx;
1688 const Register qword_count = count;
1689 const Register end_from = from; // source array end address
1690 const Register end_to = to; // destination array end address
1691 // End pointers are inclusive, and if count is not zero they point
1692 // to the last unit copied: end_to[0] := end_from[0]
1693
1694 __ enter(); // required for proper stackwalking of RuntimeStub frame
1695 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1696
1697 if (entry != NULL) {
1698 *entry = __ pc();
1699 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1700 BLOCK_COMMENT("Entry:");
1701 }
1702
1703 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1704 // r9 and r10 may be used to save non-volatile registers
1705
1706 // 'from', 'to' and 'count' are now valid
1707 __ movptr(word_count, count);
1708 __ shrptr(count, 2); // count => qword_count
1709
1710 // Copy from low to high addresses. Use 'to' as scratch.
1711 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
1712 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
1713 __ negptr(qword_count);
1714 __ jmp(L_copy_bytes);
1715
1716 // Copy trailing qwords
1717 __ BIND(L_copy_8_bytes);
1718 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
1719 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
1720 __ increment(qword_count);
1721 __ jcc(Assembler::notZero, L_copy_8_bytes);
1722
1723 // Original 'dest' is trashed, so we can't use it as a
1724 // base register for a possible trailing word copy
1725
1726 // Check for and copy trailing dword
1727 __ BIND(L_copy_4_bytes);
1728 __ testl(word_count, 2);
1729 __ jccb(Assembler::zero, L_copy_2_bytes);
1730 __ movl(rax, Address(end_from, 8));
1731 __ movl(Address(end_to, 8), rax);
1732
1733 __ addptr(end_from, 4);
1734 __ addptr(end_to, 4);
1735
1736 // Check for and copy trailing word
1737 __ BIND(L_copy_2_bytes);
1738 __ testl(word_count, 1);
1739 __ jccb(Assembler::zero, L_exit);
1740 __ movw(rax, Address(end_from, 8));
1741 __ movw(Address(end_to, 8), rax);
1742
1743 __ BIND(L_exit);
1744 restore_arg_regs();
1745 inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); // Update counter after rscratch1 is free
1746 __ xorptr(rax, rax); // return 0
1747 __ leave(); // required for proper stackwalking of RuntimeStub frame
1748 __ ret(0);
1749
1750 // Copy in multi-bytes chunks
1751 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1752 __ jmp(L_copy_4_bytes);
1753
1754 return start;
1755 }
1756
1757 address generate_fill(BasicType t, bool aligned, const char *name) {
1758 __ align(CodeEntryAlignment);
1759 StubCodeMark mark(this, "StubRoutines", name);
1760 address start = __ pc();
1761
1762 BLOCK_COMMENT("Entry:");
1763
1764 const Register to = c_rarg0; // source array address
1765 const Register value = c_rarg1; // value
1766 const Register count = c_rarg2; // elements count
1767
1768 __ enter(); // required for proper stackwalking of RuntimeStub frame
1769
1770 __ generate_fill(t, aligned, to, value, count, rax, xmm0);
1771
1772 __ leave(); // required for proper stackwalking of RuntimeStub frame
1773 __ ret(0);
1774 return start;
1775 }
1776
1777 // Arguments:
1778 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1779 // ignored
1780 // name - stub name string
1781 //
1782 // Inputs:
1783 // c_rarg0 - source array address
1784 // c_rarg1 - destination array address
1785 // c_rarg2 - element count, treated as ssize_t, can be zero
1786 //
1787 // If 'from' and/or 'to' are aligned on 4- or 2-byte boundaries, we
1788 // let the hardware handle it. The two or four words within dwords
1789 // or qwords that span cache line boundaries will still be loaded
1790 // and stored atomically.
1791 //
1792 address generate_conjoint_short_copy(bool aligned, address nooverlap_target,
1793 address *entry, const char *name) {
1794 __ align(CodeEntryAlignment);
1795 StubCodeMark mark(this, "StubRoutines", name);
1796 address start = __ pc();
1797
1798 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes;
1799 const Register from = rdi; // source array address
1800 const Register to = rsi; // destination array address
1801 const Register count = rdx; // elements count
1802 const Register word_count = rcx;
1803 const Register qword_count = count;
1804
1805 __ enter(); // required for proper stackwalking of RuntimeStub frame
1806 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1807
1808 if (entry != NULL) {
1809 *entry = __ pc();
1810 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1811 BLOCK_COMMENT("Entry:");
1812 }
1813
1814 array_overlap_test(nooverlap_target, Address::times_2);
1815 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1816 // r9 and r10 may be used to save non-volatile registers
1817
1818 // 'from', 'to' and 'count' are now valid
1819 __ movptr(word_count, count);
1820 __ shrptr(count, 2); // count => qword_count
1821
1822 // Copy from high to low addresses. Use 'to' as scratch.
1823
1824 // Check for and copy trailing word
1825 __ testl(word_count, 1);
1826 __ jccb(Assembler::zero, L_copy_4_bytes);
1827 __ movw(rax, Address(from, word_count, Address::times_2, -2));
1828 __ movw(Address(to, word_count, Address::times_2, -2), rax);
1829
1830 // Check for and copy trailing dword
1831 __ BIND(L_copy_4_bytes);
1832 __ testl(word_count, 2);
1833 __ jcc(Assembler::zero, L_copy_bytes);
1834 __ movl(rax, Address(from, qword_count, Address::times_8));
1835 __ movl(Address(to, qword_count, Address::times_8), rax);
1836 __ jmp(L_copy_bytes);
1837
1838 // Copy trailing qwords
1839 __ BIND(L_copy_8_bytes);
1840 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
1841 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
1842 __ decrement(qword_count);
1843 __ jcc(Assembler::notZero, L_copy_8_bytes);
1844
1845 restore_arg_regs();
1846 inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); // Update counter after rscratch1 is free
1847 __ xorptr(rax, rax); // return 0
1848 __ leave(); // required for proper stackwalking of RuntimeStub frame
1849 __ ret(0);
1850
1851 // Copy in multi-bytes chunks
1852 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1853
1854 restore_arg_regs();
1855 inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); // Update counter after rscratch1 is free
1856 __ xorptr(rax, rax); // return 0
1857 __ leave(); // required for proper stackwalking of RuntimeStub frame
1858 __ ret(0);
1859
1860 return start;
1861 }
1862
1863 // Arguments:
1864 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1865 // ignored
1866 // is_oop - true => oop array, so generate store check code
1867 // name - stub name string
1868 //
1869 // Inputs:
1870 // c_rarg0 - source array address
1871 // c_rarg1 - destination array address
1872 // c_rarg2 - element count, treated as ssize_t, can be zero
1873 //
1874 // If 'from' and/or 'to' are aligned on 4-byte boundaries, we let
1875 // the hardware handle it. The two dwords within qwords that span
1876 // cache line boundaries will still be loaded and stored atomicly.
1877 //
1878 // Side Effects:
1879 // disjoint_int_copy_entry is set to the no-overlap entry point
1880 // used by generate_conjoint_int_oop_copy().
1881 //
1882 address generate_disjoint_int_oop_copy(bool aligned, bool is_oop, address* entry,
1883 const char *name, bool dest_uninitialized = false) {
1884 __ align(CodeEntryAlignment);
1885 StubCodeMark mark(this, "StubRoutines", name);
1886 address start = __ pc();
1887
1888 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes, L_exit;
1889 const Register from = rdi; // source array address
1890 const Register to = rsi; // destination array address
1891 const Register count = rdx; // elements count
1892 const Register dword_count = rcx;
1893 const Register qword_count = count;
1894 const Register end_from = from; // source array end address
1895 const Register end_to = to; // destination array end address
1896 const Register saved_to = r11; // saved destination array address
1897 // End pointers are inclusive, and if count is not zero they point
1898 // to the last unit copied: end_to[0] := end_from[0]
1899
1900 __ enter(); // required for proper stackwalking of RuntimeStub frame
1901 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1902
1903 if (entry != NULL) {
1904 *entry = __ pc();
1905 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1906 BLOCK_COMMENT("Entry:");
1907 }
1908
1909 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1910 // r9 and r10 may be used to save non-volatile registers
1911 if (is_oop) {
1912 __ movq(saved_to, to);
1913 gen_write_ref_array_pre_barrier(to, count, dest_uninitialized);
1914 }
1915
1916 // 'from', 'to' and 'count' are now valid
1917 __ movptr(dword_count, count);
1918 __ shrptr(count, 1); // count => qword_count
1919
1920 // Copy from low to high addresses. Use 'to' as scratch.
1921 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
1922 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
1923 __ negptr(qword_count);
1924 __ jmp(L_copy_bytes);
1925
1926 // Copy trailing qwords
1927 __ BIND(L_copy_8_bytes);
1928 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
1929 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
1930 __ increment(qword_count);
1931 __ jcc(Assembler::notZero, L_copy_8_bytes);
1932
1933 // Check for and copy trailing dword
1934 __ BIND(L_copy_4_bytes);
1935 __ testl(dword_count, 1); // Only byte test since the value is 0 or 1
1936 __ jccb(Assembler::zero, L_exit);
1937 __ movl(rax, Address(end_from, 8));
1938 __ movl(Address(end_to, 8), rax);
1939
1940 __ BIND(L_exit);
1941 if (is_oop) {
1942 gen_write_ref_array_post_barrier(saved_to, dword_count, rax);
1943 }
1944 restore_arg_regs();
1945 inc_counter_np(SharedRuntime::_jint_array_copy_ctr); // Update counter after rscratch1 is free
1946 __ xorptr(rax, rax); // return 0
1947 __ leave(); // required for proper stackwalking of RuntimeStub frame
1948 __ ret(0);
1949
1950 // Copy in multi-bytes chunks
1951 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1952 __ jmp(L_copy_4_bytes);
1953
1954 return start;
1955 }
1956
1957 // Arguments:
1958 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1959 // ignored
1960 // is_oop - true => oop array, so generate store check code
1961 // name - stub name string
1962 //
1963 // Inputs:
1964 // c_rarg0 - source array address
1965 // c_rarg1 - destination array address
1966 // c_rarg2 - element count, treated as ssize_t, can be zero
1967 //
1968 // If 'from' and/or 'to' are aligned on 4-byte boundaries, we let
1969 // the hardware handle it. The two dwords within qwords that span
1970 // cache line boundaries will still be loaded and stored atomicly.
1971 //
1972 address generate_conjoint_int_oop_copy(bool aligned, bool is_oop, address nooverlap_target,
1973 address *entry, const char *name,
1974 bool dest_uninitialized = false) {
1975 __ align(CodeEntryAlignment);
1976 StubCodeMark mark(this, "StubRoutines", name);
1977 address start = __ pc();
1978
1979 Label L_copy_bytes, L_copy_8_bytes, L_copy_2_bytes, L_exit;
1980 const Register from = rdi; // source array address
1981 const Register to = rsi; // destination array address
1982 const Register count = rdx; // elements count
1983 const Register dword_count = rcx;
1984 const Register qword_count = count;
1985
1986 __ enter(); // required for proper stackwalking of RuntimeStub frame
1987 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1988
1989 if (entry != NULL) {
1990 *entry = __ pc();
1991 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1992 BLOCK_COMMENT("Entry:");
1993 }
1994
1995 array_overlap_test(nooverlap_target, Address::times_4);
1996 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1997 // r9 and r10 may be used to save non-volatile registers
1998
1999 if (is_oop) {
2000 // no registers are destroyed by this call
2001 gen_write_ref_array_pre_barrier(to, count, dest_uninitialized);
2002 }
2003
2004 assert_clean_int(count, rax); // Make sure 'count' is clean int.
2005 // 'from', 'to' and 'count' are now valid
2006 __ movptr(dword_count, count);
2007 __ shrptr(count, 1); // count => qword_count
2008
2009 // Copy from high to low addresses. Use 'to' as scratch.
2010
2011 // Check for and copy trailing dword
2012 __ testl(dword_count, 1);
2013 __ jcc(Assembler::zero, L_copy_bytes);
2014 __ movl(rax, Address(from, dword_count, Address::times_4, -4));
2015 __ movl(Address(to, dword_count, Address::times_4, -4), rax);
2016 __ jmp(L_copy_bytes);
2017
2018 // Copy trailing qwords
2019 __ BIND(L_copy_8_bytes);
2020 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
2021 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
2022 __ decrement(qword_count);
2023 __ jcc(Assembler::notZero, L_copy_8_bytes);
2024
2025 if (is_oop) {
2026 __ jmp(L_exit);
2027 }
2028 restore_arg_regs();
2029 inc_counter_np(SharedRuntime::_jint_array_copy_ctr); // Update counter after rscratch1 is free
2030 __ xorptr(rax, rax); // return 0
2031 __ leave(); // required for proper stackwalking of RuntimeStub frame
2032 __ ret(0);
2033
2034 // Copy in multi-bytes chunks
2035 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
2036
2037 __ BIND(L_exit);
2038 if (is_oop) {
2039 gen_write_ref_array_post_barrier(to, dword_count, rax);
2040 }
2041 restore_arg_regs();
2042 inc_counter_np(SharedRuntime::_jint_array_copy_ctr); // Update counter after rscratch1 is free
2043 __ xorptr(rax, rax); // return 0
2044 __ leave(); // required for proper stackwalking of RuntimeStub frame
2045 __ ret(0);
2046
2047 return start;
2048 }
2049
2050 // Arguments:
2051 // aligned - true => Input and output aligned on a HeapWord boundary == 8 bytes
2052 // ignored
2053 // is_oop - true => oop array, so generate store check code
2054 // name - stub name string
2055 //
2056 // Inputs:
2057 // c_rarg0 - source array address
2058 // c_rarg1 - destination array address
2059 // c_rarg2 - element count, treated as ssize_t, can be zero
2060 //
2061 // Side Effects:
2062 // disjoint_oop_copy_entry or disjoint_long_copy_entry is set to the
2063 // no-overlap entry point used by generate_conjoint_long_oop_copy().
2064 //
2065 address generate_disjoint_long_oop_copy(bool aligned, bool is_oop, address *entry,
2066 const char *name, bool dest_uninitialized = false) {
2067 __ align(CodeEntryAlignment);
2068 StubCodeMark mark(this, "StubRoutines", name);
2069 address start = __ pc();
2070
2071 Label L_copy_bytes, L_copy_8_bytes, L_exit;
2072 const Register from = rdi; // source array address
2073 const Register to = rsi; // destination array address
2074 const Register qword_count = rdx; // elements count
2075 const Register end_from = from; // source array end address
2076 const Register end_to = rcx; // destination array end address
2077 const Register saved_to = to;
2078 const Register saved_count = r11;
2079 // End pointers are inclusive, and if count is not zero they point
2080 // to the last unit copied: end_to[0] := end_from[0]
2081
2082 __ enter(); // required for proper stackwalking of RuntimeStub frame
2083 // Save no-overlap entry point for generate_conjoint_long_oop_copy()
2084 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
2085
2086 if (entry != NULL) {
2087 *entry = __ pc();
2088 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
2089 BLOCK_COMMENT("Entry:");
2090 }
2091
2092 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
2093 // r9 and r10 may be used to save non-volatile registers
2094 // 'from', 'to' and 'qword_count' are now valid
2095 if (is_oop) {
2096 // Save to and count for store barrier
2097 __ movptr(saved_count, qword_count);
2098 // no registers are destroyed by this call
2099 gen_write_ref_array_pre_barrier(to, qword_count, dest_uninitialized);
2100 }
2101
2102 // Copy from low to high addresses. Use 'to' as scratch.
2103 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
2104 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
2105 __ negptr(qword_count);
2106 __ jmp(L_copy_bytes);
2107
2108 // Copy trailing qwords
2109 __ BIND(L_copy_8_bytes);
2110 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
2111 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
2112 __ increment(qword_count);
2113 __ jcc(Assembler::notZero, L_copy_8_bytes);
2114
2115 if (is_oop) {
2116 __ jmp(L_exit);
2117 } else {
2118 restore_arg_regs();
2119 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2120 __ xorptr(rax, rax); // return 0
2121 __ leave(); // required for proper stackwalking of RuntimeStub frame
2122 __ ret(0);
2123 }
2124
2125 // Copy in multi-bytes chunks
2126 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
2127
2128 if (is_oop) {
2129 __ BIND(L_exit);
2130 gen_write_ref_array_post_barrier(saved_to, saved_count, rax);
2131 }
2132 restore_arg_regs();
2133 if (is_oop) {
2134 inc_counter_np(SharedRuntime::_oop_array_copy_ctr); // Update counter after rscratch1 is free
2135 } else {
2136 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2137 }
2138 __ xorptr(rax, rax); // return 0
2139 __ leave(); // required for proper stackwalking of RuntimeStub frame
2140 __ ret(0);
2141
2142 return start;
2143 }
2144
2145 // Arguments:
2146 // aligned - true => Input and output aligned on a HeapWord boundary == 8 bytes
2147 // ignored
2148 // is_oop - true => oop array, so generate store check code
2149 // name - stub name string
2150 //
2151 // Inputs:
2152 // c_rarg0 - source array address
2153 // c_rarg1 - destination array address
2154 // c_rarg2 - element count, treated as ssize_t, can be zero
2155 //
2156 address generate_conjoint_long_oop_copy(bool aligned, bool is_oop,
2157 address nooverlap_target, address *entry,
2158 const char *name, bool dest_uninitialized = false) {
2159 __ align(CodeEntryAlignment);
2160 StubCodeMark mark(this, "StubRoutines", name);
2161 address start = __ pc();
2162
2163 Label L_copy_bytes, L_copy_8_bytes, L_exit;
2164 const Register from = rdi; // source array address
2165 const Register to = rsi; // destination array address
2166 const Register qword_count = rdx; // elements count
2167 const Register saved_count = rcx;
2168
2169 __ enter(); // required for proper stackwalking of RuntimeStub frame
2170 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
2171
2172 if (entry != NULL) {
2173 *entry = __ pc();
2174 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
2175 BLOCK_COMMENT("Entry:");
2176 }
2177
2178 array_overlap_test(nooverlap_target, Address::times_8);
2179 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
2180 // r9 and r10 may be used to save non-volatile registers
2181 // 'from', 'to' and 'qword_count' are now valid
2182 if (is_oop) {
2183 // Save to and count for store barrier
2184 __ movptr(saved_count, qword_count);
2185 // No registers are destroyed by this call
2186 gen_write_ref_array_pre_barrier(to, saved_count, dest_uninitialized);
2187 }
2188
2189 __ jmp(L_copy_bytes);
2190
2191 // Copy trailing qwords
2192 __ BIND(L_copy_8_bytes);
2193 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
2194 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
2195 __ decrement(qword_count);
2196 __ jcc(Assembler::notZero, L_copy_8_bytes);
2197
2198 if (is_oop) {
2199 __ jmp(L_exit);
2200 } else {
2201 restore_arg_regs();
2202 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2203 __ xorptr(rax, rax); // return 0
2204 __ leave(); // required for proper stackwalking of RuntimeStub frame
2205 __ ret(0);
2206 }
2207
2208 // Copy in multi-bytes chunks
2209 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
2210
2211 if (is_oop) {
2212 __ BIND(L_exit);
2213 gen_write_ref_array_post_barrier(to, saved_count, rax);
2214 }
2215 restore_arg_regs();
2216 if (is_oop) {
2217 inc_counter_np(SharedRuntime::_oop_array_copy_ctr); // Update counter after rscratch1 is free
2218 } else {
2219 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2220 }
2221 __ xorptr(rax, rax); // return 0
2222 __ leave(); // required for proper stackwalking of RuntimeStub frame
2223 __ ret(0);
2224
2225 return start;
2226 }
2227
2228
2229 // Helper for generating a dynamic type check.
2230 // Smashes no registers.
2231 void generate_type_check(Register sub_klass,
2232 Register super_check_offset,
2233 Register super_klass,
2234 Label& L_success) {
2235 assert_different_registers(sub_klass, super_check_offset, super_klass);
2236
2237 BLOCK_COMMENT("type_check:");
2238
2239 Label L_miss;
2240
2241 __ check_klass_subtype_fast_path(sub_klass, super_klass, noreg, &L_success, &L_miss, NULL,
2242 super_check_offset);
2243 __ check_klass_subtype_slow_path(sub_klass, super_klass, noreg, noreg, &L_success, NULL);
2244
2245 // Fall through on failure!
2246 __ BIND(L_miss);
2247 }
2248
2249 //
2250 // Generate checkcasting array copy stub
2251 //
2252 // Input:
2253 // c_rarg0 - source array address
2254 // c_rarg1 - destination array address
2255 // c_rarg2 - element count, treated as ssize_t, can be zero
2256 // c_rarg3 - size_t ckoff (super_check_offset)
2257 // not Win64
2258 // c_rarg4 - oop ckval (super_klass)
2259 // Win64
2260 // rsp+40 - oop ckval (super_klass)
2261 //
2262 // Output:
2263 // rax == 0 - success
2264 // rax == -1^K - failure, where K is partial transfer count
2265 //
2266 address generate_checkcast_copy(const char *name, address *entry,
2267 bool dest_uninitialized = false) {
2268
2269 Label L_load_element, L_store_element, L_do_card_marks, L_done;
2270
2271 // Input registers (after setup_arg_regs)
2272 const Register from = rdi; // source array address
2273 const Register to = rsi; // destination array address
2274 const Register length = rdx; // elements count
2275 const Register ckoff = rcx; // super_check_offset
2276 const Register ckval = r8; // super_klass
2277
2278 // Registers used as temps (r13, r14 are save-on-entry)
2279 const Register end_from = from; // source array end address
2280 const Register end_to = r13; // destination array end address
2281 const Register count = rdx; // -(count_remaining)
2282 const Register r14_length = r14; // saved copy of length
2283 // End pointers are inclusive, and if length is not zero they point
2284 // to the last unit copied: end_to[0] := end_from[0]
2285
2286 const Register rax_oop = rax; // actual oop copied
2287 const Register r11_klass = r11; // oop._klass
2288
2289 //---------------------------------------------------------------
2290 // Assembler stub will be used for this call to arraycopy
2291 // if the two arrays are subtypes of Object[] but the
2292 // destination array type is not equal to or a supertype
2293 // of the source type. Each element must be separately
2294 // checked.
2295
2296 __ align(CodeEntryAlignment);
2297 StubCodeMark mark(this, "StubRoutines", name);
2298 address start = __ pc();
2299
2300 __ enter(); // required for proper stackwalking of RuntimeStub frame
2301
2302 #ifdef ASSERT
2303 // caller guarantees that the arrays really are different
2304 // otherwise, we would have to make conjoint checks
2305 { Label L;
2306 array_overlap_test(L, TIMES_OOP);
2307 __ stop("checkcast_copy within a single array");
2308 __ bind(L);
2309 }
2310 #endif //ASSERT
2311
2312 setup_arg_regs(4); // from => rdi, to => rsi, length => rdx
2313 // ckoff => rcx, ckval => r8
2314 // r9 and r10 may be used to save non-volatile registers
2315 #ifdef _WIN64
2316 // last argument (#4) is on stack on Win64
2317 __ movptr(ckval, Address(rsp, 6 * wordSize));
2318 #endif
2319
2320 // Caller of this entry point must set up the argument registers.
2321 if (entry != NULL) {
2322 *entry = __ pc();
2323 BLOCK_COMMENT("Entry:");
2324 }
2325
2326 // allocate spill slots for r13, r14
2327 enum {
2328 saved_r13_offset,
2329 saved_r14_offset,
2330 saved_rbp_offset
2331 };
2332 __ subptr(rsp, saved_rbp_offset * wordSize);
2333 __ movptr(Address(rsp, saved_r13_offset * wordSize), r13);
2334 __ movptr(Address(rsp, saved_r14_offset * wordSize), r14);
2335
2336 // check that int operands are properly extended to size_t
2337 assert_clean_int(length, rax);
2338 assert_clean_int(ckoff, rax);
2339
2340 #ifdef ASSERT
2341 BLOCK_COMMENT("assert consistent ckoff/ckval");
2342 // The ckoff and ckval must be mutually consistent,
2343 // even though caller generates both.
2344 { Label L;
2345 int sco_offset = in_bytes(Klass::super_check_offset_offset());
2346 __ cmpl(ckoff, Address(ckval, sco_offset));
2347 __ jcc(Assembler::equal, L);
2348 __ stop("super_check_offset inconsistent");
2349 __ bind(L);
2350 }
2351 #endif //ASSERT
2352
2353 // Loop-invariant addresses. They are exclusive end pointers.
2354 Address end_from_addr(from, length, TIMES_OOP, 0);
2355 Address end_to_addr(to, length, TIMES_OOP, 0);
2356 // Loop-variant addresses. They assume post-incremented count < 0.
2357 Address from_element_addr(end_from, count, TIMES_OOP, 0);
2358 Address to_element_addr(end_to, count, TIMES_OOP, 0);
2359
2360 gen_write_ref_array_pre_barrier(to, count, dest_uninitialized);
2361
2362 // Copy from low to high addresses, indexed from the end of each array.
2363 __ lea(end_from, end_from_addr);
2364 __ lea(end_to, end_to_addr);
2365 __ movptr(r14_length, length); // save a copy of the length
2366 assert(length == count, ""); // else fix next line:
2367 __ negptr(count); // negate and test the length
2368 __ jcc(Assembler::notZero, L_load_element);
2369
2370 // Empty array: Nothing to do.
2371 __ xorptr(rax, rax); // return 0 on (trivial) success
2372 __ jmp(L_done);
2373
2374 // ======== begin loop ========
2375 // (Loop is rotated; its entry is L_load_element.)
2376 // Loop control:
2377 // for (count = -count; count != 0; count++)
2378 // Base pointers src, dst are biased by 8*(count-1),to last element.
2379 __ align(OptoLoopAlignment);
2380
2381 __ BIND(L_store_element);
2382 __ store_heap_oop(to_element_addr, rax_oop); // store the oop
2383 __ increment(count); // increment the count toward zero
2384 __ jcc(Assembler::zero, L_do_card_marks);
2385
2386 // ======== loop entry is here ========
2387 __ BIND(L_load_element);
2388 __ load_heap_oop(rax_oop, from_element_addr); // load the oop
2389 __ testptr(rax_oop, rax_oop);
2390 __ jcc(Assembler::zero, L_store_element);
2391
2392 __ load_klass(r11_klass, rax_oop);// query the object klass
2393 generate_type_check(r11_klass, ckoff, ckval, L_store_element);
2394 // ======== end loop ========
2395
2396 // It was a real error; we must depend on the caller to finish the job.
2397 // Register rdx = -1 * number of *remaining* oops, r14 = *total* oops.
2398 // Emit GC store barriers for the oops we have copied (r14 + rdx),
2399 // and report their number to the caller.
2400 assert_different_registers(rax, r14_length, count, to, end_to, rcx, rscratch1);
2401 Label L_post_barrier;
2402 __ addptr(r14_length, count); // K = (original - remaining) oops
2403 __ movptr(rax, r14_length); // save the value
2404 __ notptr(rax); // report (-1^K) to caller (does not affect flags)
2405 __ jccb(Assembler::notZero, L_post_barrier);
2406 __ jmp(L_done); // K == 0, nothing was copied, skip post barrier
2407
2408 // Come here on success only.
2409 __ BIND(L_do_card_marks);
2410 __ xorptr(rax, rax); // return 0 on success
2411
2412 __ BIND(L_post_barrier);
2413 gen_write_ref_array_post_barrier(to, r14_length, rscratch1);
2414
2415 // Common exit point (success or failure).
2416 __ BIND(L_done);
2417 __ movptr(r13, Address(rsp, saved_r13_offset * wordSize));
2418 __ movptr(r14, Address(rsp, saved_r14_offset * wordSize));
2419 restore_arg_regs();
2420 inc_counter_np(SharedRuntime::_checkcast_array_copy_ctr); // Update counter after rscratch1 is free
2421 __ leave(); // required for proper stackwalking of RuntimeStub frame
2422 __ ret(0);
2423
2424 return start;
2425 }
2426
2427 //
2428 // Generate 'unsafe' array copy stub
2429 // Though just as safe as the other stubs, it takes an unscaled
2430 // size_t argument instead of an element count.
2431 //
2432 // Input:
2433 // c_rarg0 - source array address
2434 // c_rarg1 - destination array address
2435 // c_rarg2 - byte count, treated as ssize_t, can be zero
2436 //
2437 // Examines the alignment of the operands and dispatches
2438 // to a long, int, short, or byte copy loop.
2439 //
2440 address generate_unsafe_copy(const char *name,
2441 address byte_copy_entry, address short_copy_entry,
2442 address int_copy_entry, address long_copy_entry) {
2443
2444 Label L_long_aligned, L_int_aligned, L_short_aligned;
2445
2446 // Input registers (before setup_arg_regs)
2447 const Register from = c_rarg0; // source array address
2448 const Register to = c_rarg1; // destination array address
2449 const Register size = c_rarg2; // byte count (size_t)
2450
2451 // Register used as a temp
2452 const Register bits = rax; // test copy of low bits
2453
2454 __ align(CodeEntryAlignment);
2455 StubCodeMark mark(this, "StubRoutines", name);
2456 address start = __ pc();
2457
2458 __ enter(); // required for proper stackwalking of RuntimeStub frame
2459
2460 // bump this on entry, not on exit:
2461 inc_counter_np(SharedRuntime::_unsafe_array_copy_ctr);
2462
2463 __ mov(bits, from);
2464 __ orptr(bits, to);
2465 __ orptr(bits, size);
2466
2467 __ testb(bits, BytesPerLong-1);
2468 __ jccb(Assembler::zero, L_long_aligned);
2469
2470 __ testb(bits, BytesPerInt-1);
2471 __ jccb(Assembler::zero, L_int_aligned);
2472
2473 __ testb(bits, BytesPerShort-1);
2474 __ jump_cc(Assembler::notZero, RuntimeAddress(byte_copy_entry));
2475
2476 __ BIND(L_short_aligned);
2477 __ shrptr(size, LogBytesPerShort); // size => short_count
2478 __ jump(RuntimeAddress(short_copy_entry));
2479
2480 __ BIND(L_int_aligned);
2481 __ shrptr(size, LogBytesPerInt); // size => int_count
2482 __ jump(RuntimeAddress(int_copy_entry));
2483
2484 __ BIND(L_long_aligned);
2485 __ shrptr(size, LogBytesPerLong); // size => qword_count
2486 __ jump(RuntimeAddress(long_copy_entry));
2487
2488 return start;
2489 }
2490
2491 // Perform range checks on the proposed arraycopy.
2492 // Kills temp, but nothing else.
2493 // Also, clean the sign bits of src_pos and dst_pos.
2494 void arraycopy_range_checks(Register src, // source array oop (c_rarg0)
2495 Register src_pos, // source position (c_rarg1)
2496 Register dst, // destination array oo (c_rarg2)
2497 Register dst_pos, // destination position (c_rarg3)
2498 Register length,
2499 Register temp,
2500 Label& L_failed) {
2501 BLOCK_COMMENT("arraycopy_range_checks:");
2502
2503 // if (src_pos + length > arrayOop(src)->length()) FAIL;
2504 __ movl(temp, length);
2505 __ addl(temp, src_pos); // src_pos + length
2506 __ cmpl(temp, Address(src, arrayOopDesc::length_offset_in_bytes()));
2507 __ jcc(Assembler::above, L_failed);
2508
2509 // if (dst_pos + length > arrayOop(dst)->length()) FAIL;
2510 __ movl(temp, length);
2511 __ addl(temp, dst_pos); // dst_pos + length
2512 __ cmpl(temp, Address(dst, arrayOopDesc::length_offset_in_bytes()));
2513 __ jcc(Assembler::above, L_failed);
2514
2515 // Have to clean up high 32-bits of 'src_pos' and 'dst_pos'.
2516 // Move with sign extension can be used since they are positive.
2517 __ movslq(src_pos, src_pos);
2518 __ movslq(dst_pos, dst_pos);
2519
2520 BLOCK_COMMENT("arraycopy_range_checks done");
2521 }
2522
2523 //
2524 // Generate generic array copy stubs
2525 //
2526 // Input:
2527 // c_rarg0 - src oop
2528 // c_rarg1 - src_pos (32-bits)
2529 // c_rarg2 - dst oop
2530 // c_rarg3 - dst_pos (32-bits)
2531 // not Win64
2532 // c_rarg4 - element count (32-bits)
2533 // Win64
2534 // rsp+40 - element count (32-bits)
2535 //
2536 // Output:
2537 // rax == 0 - success
2538 // rax == -1^K - failure, where K is partial transfer count
2539 //
2540 address generate_generic_copy(const char *name,
2541 address byte_copy_entry, address short_copy_entry,
2542 address int_copy_entry, address oop_copy_entry,
2543 address long_copy_entry, address checkcast_copy_entry) {
2544
2545 Label L_failed, L_failed_0, L_objArray;
2546 Label L_copy_bytes, L_copy_shorts, L_copy_ints, L_copy_longs;
2547
2548 // Input registers
2549 const Register src = c_rarg0; // source array oop
2550 const Register src_pos = c_rarg1; // source position
2551 const Register dst = c_rarg2; // destination array oop
2552 const Register dst_pos = c_rarg3; // destination position
2553 #ifndef _WIN64
2554 const Register length = c_rarg4;
2555 #else
2556 const Address length(rsp, 6 * wordSize); // elements count is on stack on Win64
2557 #endif
2558
2559 { int modulus = CodeEntryAlignment;
2560 int target = modulus - 5; // 5 = sizeof jmp(L_failed)
2561 int advance = target - (__ offset() % modulus);
2562 if (advance < 0) advance += modulus;
2563 if (advance > 0) __ nop(advance);
2564 }
2565 StubCodeMark mark(this, "StubRoutines", name);
2566
2567 // Short-hop target to L_failed. Makes for denser prologue code.
2568 __ BIND(L_failed_0);
2569 __ jmp(L_failed);
2570 assert(__ offset() % CodeEntryAlignment == 0, "no further alignment needed");
2571
2572 __ align(CodeEntryAlignment);
2573 address start = __ pc();
2574
2575 __ enter(); // required for proper stackwalking of RuntimeStub frame
2576
2577 // bump this on entry, not on exit:
2578 inc_counter_np(SharedRuntime::_generic_array_copy_ctr);
2579
2580 //-----------------------------------------------------------------------
2581 // Assembler stub will be used for this call to arraycopy
2582 // if the following conditions are met:
2583 //
2584 // (1) src and dst must not be null.
2585 // (2) src_pos must not be negative.
2586 // (3) dst_pos must not be negative.
2587 // (4) length must not be negative.
2588 // (5) src klass and dst klass should be the same and not NULL.
2589 // (6) src and dst should be arrays.
2590 // (7) src_pos + length must not exceed length of src.
2591 // (8) dst_pos + length must not exceed length of dst.
2592 //
2593
2594 // if (src == NULL) return -1;
2595 __ testptr(src, src); // src oop
2596 size_t j1off = __ offset();
2597 __ jccb(Assembler::zero, L_failed_0);
2598
2599 // if (src_pos < 0) return -1;
2600 __ testl(src_pos, src_pos); // src_pos (32-bits)
2601 __ jccb(Assembler::negative, L_failed_0);
2602
2603 // if (dst == NULL) return -1;
2604 __ testptr(dst, dst); // dst oop
2605 __ jccb(Assembler::zero, L_failed_0);
2606
2607 // if (dst_pos < 0) return -1;
2608 __ testl(dst_pos, dst_pos); // dst_pos (32-bits)
2609 size_t j4off = __ offset();
2610 __ jccb(Assembler::negative, L_failed_0);
2611
2612 // The first four tests are very dense code,
2613 // but not quite dense enough to put four
2614 // jumps in a 16-byte instruction fetch buffer.
2615 // That's good, because some branch predicters
2616 // do not like jumps so close together.
2617 // Make sure of this.
2618 guarantee(((j1off ^ j4off) & ~15) != 0, "I$ line of 1st & 4th jumps");
2619
2620 // registers used as temp
2621 const Register r11_length = r11; // elements count to copy
2622 const Register r10_src_klass = r10; // array klass
2623
2624 // if (length < 0) return -1;
2625 __ movl(r11_length, length); // length (elements count, 32-bits value)
2626 __ testl(r11_length, r11_length);
2627 __ jccb(Assembler::negative, L_failed_0);
2628
2629 __ load_klass(r10_src_klass, src);
2630 #ifdef ASSERT
2631 // assert(src->klass() != NULL);
2632 {
2633 BLOCK_COMMENT("assert klasses not null {");
2634 Label L1, L2;
2635 __ testptr(r10_src_klass, r10_src_klass);
2636 __ jcc(Assembler::notZero, L2); // it is broken if klass is NULL
2637 __ bind(L1);
2638 __ stop("broken null klass");
2639 __ bind(L2);
2640 __ load_klass(rax, dst);
2641 __ cmpq(rax, 0);
2642 __ jcc(Assembler::equal, L1); // this would be broken also
2643 BLOCK_COMMENT("} assert klasses not null done");
2644 }
2645 #endif
2646
2647 // Load layout helper (32-bits)
2648 //
2649 // |array_tag| | header_size | element_type | |log2_element_size|
2650 // 32 30 24 16 8 2 0
2651 //
2652 // array_tag: typeArray = 0x3, objArray = 0x2, non-array = 0x0
2653 //
2654
2655 const int lh_offset = in_bytes(Klass::layout_helper_offset());
2656
2657 // Handle objArrays completely differently...
2658 const jint objArray_lh = Klass::array_layout_helper(T_OBJECT);
2659 __ cmpl(Address(r10_src_klass, lh_offset), objArray_lh);
2660 __ jcc(Assembler::equal, L_objArray);
2661
2662 // if (src->klass() != dst->klass()) return -1;
2663 __ load_klass(rax, dst);
2664 __ cmpq(r10_src_klass, rax);
2665 __ jcc(Assembler::notEqual, L_failed);
2666
2667 const Register rax_lh = rax; // layout helper
2668 __ movl(rax_lh, Address(r10_src_klass, lh_offset));
2669
2670 // if (!src->is_Array()) return -1;
2671 __ cmpl(rax_lh, Klass::_lh_neutral_value);
2672 __ jcc(Assembler::greaterEqual, L_failed);
2673
2674 // At this point, it is known to be a typeArray (array_tag 0x3).
2675 #ifdef ASSERT
2676 {
2677 BLOCK_COMMENT("assert primitive array {");
2678 Label L;
2679 __ cmpl(rax_lh, (Klass::_lh_array_tag_type_value << Klass::_lh_array_tag_shift));
2680 __ jcc(Assembler::greaterEqual, L);
2681 __ stop("must be a primitive array");
2682 __ bind(L);
2683 BLOCK_COMMENT("} assert primitive array done");
2684 }
2685 #endif
2686
2687 arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length,
2688 r10, L_failed);
2689
2690 // typeArrayKlass
2691 //
2692 // src_addr = (src + array_header_in_bytes()) + (src_pos << log2elemsize);
2693 // dst_addr = (dst + array_header_in_bytes()) + (dst_pos << log2elemsize);
2694 //
2695
2696 const Register r10_offset = r10; // array offset
2697 const Register rax_elsize = rax_lh; // element size
2698
2699 __ movl(r10_offset, rax_lh);
2700 __ shrl(r10_offset, Klass::_lh_header_size_shift);
2701 __ andptr(r10_offset, Klass::_lh_header_size_mask); // array_offset
2702 __ addptr(src, r10_offset); // src array offset
2703 __ addptr(dst, r10_offset); // dst array offset
2704 BLOCK_COMMENT("choose copy loop based on element size");
2705 __ andl(rax_lh, Klass::_lh_log2_element_size_mask); // rax_lh -> rax_elsize
2706
2707 // next registers should be set before the jump to corresponding stub
2708 const Register from = c_rarg0; // source array address
2709 const Register to = c_rarg1; // destination array address
2710 const Register count = c_rarg2; // elements count
2711
2712 // 'from', 'to', 'count' registers should be set in such order
2713 // since they are the same as 'src', 'src_pos', 'dst'.
2714
2715 __ BIND(L_copy_bytes);
2716 __ cmpl(rax_elsize, 0);
2717 __ jccb(Assembler::notEqual, L_copy_shorts);
2718 __ lea(from, Address(src, src_pos, Address::times_1, 0));// src_addr
2719 __ lea(to, Address(dst, dst_pos, Address::times_1, 0));// dst_addr
2720 __ movl2ptr(count, r11_length); // length
2721 __ jump(RuntimeAddress(byte_copy_entry));
2722
2723 __ BIND(L_copy_shorts);
2724 __ cmpl(rax_elsize, LogBytesPerShort);
2725 __ jccb(Assembler::notEqual, L_copy_ints);
2726 __ lea(from, Address(src, src_pos, Address::times_2, 0));// src_addr
2727 __ lea(to, Address(dst, dst_pos, Address::times_2, 0));// dst_addr
2728 __ movl2ptr(count, r11_length); // length
2729 __ jump(RuntimeAddress(short_copy_entry));
2730
2731 __ BIND(L_copy_ints);
2732 __ cmpl(rax_elsize, LogBytesPerInt);
2733 __ jccb(Assembler::notEqual, L_copy_longs);
2734 __ lea(from, Address(src, src_pos, Address::times_4, 0));// src_addr
2735 __ lea(to, Address(dst, dst_pos, Address::times_4, 0));// dst_addr
2736 __ movl2ptr(count, r11_length); // length
2737 __ jump(RuntimeAddress(int_copy_entry));
2738
2739 __ BIND(L_copy_longs);
2740 #ifdef ASSERT
2741 {
2742 BLOCK_COMMENT("assert long copy {");
2743 Label L;
2744 __ cmpl(rax_elsize, LogBytesPerLong);
2745 __ jcc(Assembler::equal, L);
2746 __ stop("must be long copy, but elsize is wrong");
2747 __ bind(L);
2748 BLOCK_COMMENT("} assert long copy done");
2749 }
2750 #endif
2751 __ lea(from, Address(src, src_pos, Address::times_8, 0));// src_addr
2752 __ lea(to, Address(dst, dst_pos, Address::times_8, 0));// dst_addr
2753 __ movl2ptr(count, r11_length); // length
2754 __ jump(RuntimeAddress(long_copy_entry));
2755
2756 // objArrayKlass
2757 __ BIND(L_objArray);
2758 // live at this point: r10_src_klass, r11_length, src[_pos], dst[_pos]
2759
2760 Label L_plain_copy, L_checkcast_copy;
2761 // test array classes for subtyping
2762 __ load_klass(rax, dst);
2763 __ cmpq(r10_src_klass, rax); // usual case is exact equality
2764 __ jcc(Assembler::notEqual, L_checkcast_copy);
2765
2766 // Identically typed arrays can be copied without element-wise checks.
2767 arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length,
2768 r10, L_failed);
2769
2770 __ lea(from, Address(src, src_pos, TIMES_OOP,
2771 arrayOopDesc::base_offset_in_bytes(T_OBJECT))); // src_addr
2772 __ lea(to, Address(dst, dst_pos, TIMES_OOP,
2773 arrayOopDesc::base_offset_in_bytes(T_OBJECT))); // dst_addr
2774 __ movl2ptr(count, r11_length); // length
2775 __ BIND(L_plain_copy);
2776 __ jump(RuntimeAddress(oop_copy_entry));
2777
2778 __ BIND(L_checkcast_copy);
2779 // live at this point: r10_src_klass, r11_length, rax (dst_klass)
2780 {
2781 // Before looking at dst.length, make sure dst is also an objArray.
2782 __ cmpl(Address(rax, lh_offset), objArray_lh);
2783 __ jcc(Assembler::notEqual, L_failed);
2784
2785 // It is safe to examine both src.length and dst.length.
2786 arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length,
2787 rax, L_failed);
2788
2789 const Register r11_dst_klass = r11;
2790 __ load_klass(r11_dst_klass, dst); // reload
2791
2792 // Marshal the base address arguments now, freeing registers.
2793 __ lea(from, Address(src, src_pos, TIMES_OOP,
2794 arrayOopDesc::base_offset_in_bytes(T_OBJECT)));
2795 __ lea(to, Address(dst, dst_pos, TIMES_OOP,
2796 arrayOopDesc::base_offset_in_bytes(T_OBJECT)));
2797 __ movl(count, length); // length (reloaded)
2798 Register sco_temp = c_rarg3; // this register is free now
2799 assert_different_registers(from, to, count, sco_temp,
2800 r11_dst_klass, r10_src_klass);
2801 assert_clean_int(count, sco_temp);
2802
2803 // Generate the type check.
2804 const int sco_offset = in_bytes(Klass::super_check_offset_offset());
2805 __ movl(sco_temp, Address(r11_dst_klass, sco_offset));
2806 assert_clean_int(sco_temp, rax);
2807 generate_type_check(r10_src_klass, sco_temp, r11_dst_klass, L_plain_copy);
2808
2809 // Fetch destination element klass from the objArrayKlass header.
2810 int ek_offset = in_bytes(objArrayKlass::element_klass_offset());
2811 __ movptr(r11_dst_klass, Address(r11_dst_klass, ek_offset));
2812 __ movl( sco_temp, Address(r11_dst_klass, sco_offset));
2813 assert_clean_int(sco_temp, rax);
2814
2815 // the checkcast_copy loop needs two extra arguments:
2816 assert(c_rarg3 == sco_temp, "#3 already in place");
2817 // Set up arguments for checkcast_copy_entry.
2818 setup_arg_regs(4);
2819 __ movptr(r8, r11_dst_klass); // dst.klass.element_klass, r8 is c_rarg4 on Linux/Solaris
2820 __ jump(RuntimeAddress(checkcast_copy_entry));
2821 }
2822
2823 __ BIND(L_failed);
2824 __ xorptr(rax, rax);
2825 __ notptr(rax); // return -1
2826 __ leave(); // required for proper stackwalking of RuntimeStub frame
2827 __ ret(0);
2828
2829 return start;
2830 }
2831
2832 void generate_arraycopy_stubs() {
2833 address entry;
2834 address entry_jbyte_arraycopy;
2835 address entry_jshort_arraycopy;
2836 address entry_jint_arraycopy;
2837 address entry_oop_arraycopy;
2838 address entry_jlong_arraycopy;
2839 address entry_checkcast_arraycopy;
2840
2841 StubRoutines::_jbyte_disjoint_arraycopy = generate_disjoint_byte_copy(false, &entry,
2842 "jbyte_disjoint_arraycopy");
2843 StubRoutines::_jbyte_arraycopy = generate_conjoint_byte_copy(false, entry, &entry_jbyte_arraycopy,
2844 "jbyte_arraycopy");
2845
2846 StubRoutines::_jshort_disjoint_arraycopy = generate_disjoint_short_copy(false, &entry,
2847 "jshort_disjoint_arraycopy");
2848 StubRoutines::_jshort_arraycopy = generate_conjoint_short_copy(false, entry, &entry_jshort_arraycopy,
2849 "jshort_arraycopy");
2850
2851 StubRoutines::_jint_disjoint_arraycopy = generate_disjoint_int_oop_copy(false, false, &entry,
2852 "jint_disjoint_arraycopy");
2853 StubRoutines::_jint_arraycopy = generate_conjoint_int_oop_copy(false, false, entry,
2854 &entry_jint_arraycopy, "jint_arraycopy");
2855
2856 StubRoutines::_jlong_disjoint_arraycopy = generate_disjoint_long_oop_copy(false, false, &entry,
2857 "jlong_disjoint_arraycopy");
2858 StubRoutines::_jlong_arraycopy = generate_conjoint_long_oop_copy(false, false, entry,
2859 &entry_jlong_arraycopy, "jlong_arraycopy");
2860
2861
2862 if (UseCompressedOops) {
2863 StubRoutines::_oop_disjoint_arraycopy = generate_disjoint_int_oop_copy(false, true, &entry,
2864 "oop_disjoint_arraycopy");
2865 StubRoutines::_oop_arraycopy = generate_conjoint_int_oop_copy(false, true, entry,
2866 &entry_oop_arraycopy, "oop_arraycopy");
2867 StubRoutines::_oop_disjoint_arraycopy_uninit = generate_disjoint_int_oop_copy(false, true, &entry,
2868 "oop_disjoint_arraycopy_uninit",
2869 /*dest_uninitialized*/true);
2870 StubRoutines::_oop_arraycopy_uninit = generate_conjoint_int_oop_copy(false, true, entry,
2871 NULL, "oop_arraycopy_uninit",
2872 /*dest_uninitialized*/true);
2873 } else {
2874 StubRoutines::_oop_disjoint_arraycopy = generate_disjoint_long_oop_copy(false, true, &entry,
2875 "oop_disjoint_arraycopy");
2876 StubRoutines::_oop_arraycopy = generate_conjoint_long_oop_copy(false, true, entry,
2877 &entry_oop_arraycopy, "oop_arraycopy");
2878 StubRoutines::_oop_disjoint_arraycopy_uninit = generate_disjoint_long_oop_copy(false, true, &entry,
2879 "oop_disjoint_arraycopy_uninit",
2880 /*dest_uninitialized*/true);
2881 StubRoutines::_oop_arraycopy_uninit = generate_conjoint_long_oop_copy(false, true, entry,
2882 NULL, "oop_arraycopy_uninit",
2883 /*dest_uninitialized*/true);
2884 }
2885
2886 StubRoutines::_checkcast_arraycopy = generate_checkcast_copy("checkcast_arraycopy", &entry_checkcast_arraycopy);
2887 StubRoutines::_checkcast_arraycopy_uninit = generate_checkcast_copy("checkcast_arraycopy_uninit", NULL,
2888 /*dest_uninitialized*/true);
2889
2890 StubRoutines::_unsafe_arraycopy = generate_unsafe_copy("unsafe_arraycopy",
2891 entry_jbyte_arraycopy,
2892 entry_jshort_arraycopy,
2893 entry_jint_arraycopy,
2894 entry_jlong_arraycopy);
2895 StubRoutines::_generic_arraycopy = generate_generic_copy("generic_arraycopy",
2896 entry_jbyte_arraycopy,
2897 entry_jshort_arraycopy,
2898 entry_jint_arraycopy,
2899 entry_oop_arraycopy,
2900 entry_jlong_arraycopy,
2901 entry_checkcast_arraycopy);
2902
2903 StubRoutines::_jbyte_fill = generate_fill(T_BYTE, false, "jbyte_fill");
2904 StubRoutines::_jshort_fill = generate_fill(T_SHORT, false, "jshort_fill");
2905 StubRoutines::_jint_fill = generate_fill(T_INT, false, "jint_fill");
2906 StubRoutines::_arrayof_jbyte_fill = generate_fill(T_BYTE, true, "arrayof_jbyte_fill");
2907 StubRoutines::_arrayof_jshort_fill = generate_fill(T_SHORT, true, "arrayof_jshort_fill");
2908 StubRoutines::_arrayof_jint_fill = generate_fill(T_INT, true, "arrayof_jint_fill");
2909
2910 // We don't generate specialized code for HeapWord-aligned source
2911 // arrays, so just use the code we've already generated
2912 StubRoutines::_arrayof_jbyte_disjoint_arraycopy = StubRoutines::_jbyte_disjoint_arraycopy;
2913 StubRoutines::_arrayof_jbyte_arraycopy = StubRoutines::_jbyte_arraycopy;
2914
2915 StubRoutines::_arrayof_jshort_disjoint_arraycopy = StubRoutines::_jshort_disjoint_arraycopy;
2916 StubRoutines::_arrayof_jshort_arraycopy = StubRoutines::_jshort_arraycopy;
2917
2918 StubRoutines::_arrayof_jint_disjoint_arraycopy = StubRoutines::_jint_disjoint_arraycopy;
2919 StubRoutines::_arrayof_jint_arraycopy = StubRoutines::_jint_arraycopy;
2920
2921 StubRoutines::_arrayof_jlong_disjoint_arraycopy = StubRoutines::_jlong_disjoint_arraycopy;
2922 StubRoutines::_arrayof_jlong_arraycopy = StubRoutines::_jlong_arraycopy;
2923
2924 StubRoutines::_arrayof_oop_disjoint_arraycopy = StubRoutines::_oop_disjoint_arraycopy;
2925 StubRoutines::_arrayof_oop_arraycopy = StubRoutines::_oop_arraycopy;
2926
2927 StubRoutines::_arrayof_oop_disjoint_arraycopy_uninit = StubRoutines::_oop_disjoint_arraycopy_uninit;
2928 StubRoutines::_arrayof_oop_arraycopy_uninit = StubRoutines::_oop_arraycopy_uninit;
2929 }
2930
2931 void generate_math_stubs() {
2932 {
2933 StubCodeMark mark(this, "StubRoutines", "log");
2934 StubRoutines::_intrinsic_log = (double (*)(double)) __ pc();
2935
2936 __ subq(rsp, 8);
2937 __ movdbl(Address(rsp, 0), xmm0);
2938 __ fld_d(Address(rsp, 0));
2939 __ flog();
2940 __ fstp_d(Address(rsp, 0));
2941 __ movdbl(xmm0, Address(rsp, 0));
2942 __ addq(rsp, 8);
2943 __ ret(0);
2944 }
2945 {
2946 StubCodeMark mark(this, "StubRoutines", "log10");
2947 StubRoutines::_intrinsic_log10 = (double (*)(double)) __ pc();
2948
2949 __ subq(rsp, 8);
2950 __ movdbl(Address(rsp, 0), xmm0);
2951 __ fld_d(Address(rsp, 0));
2952 __ flog10();
2953 __ fstp_d(Address(rsp, 0));
2954 __ movdbl(xmm0, Address(rsp, 0));
2955 __ addq(rsp, 8);
2956 __ ret(0);
2957 }
2958 {
2959 StubCodeMark mark(this, "StubRoutines", "sin");
2960 StubRoutines::_intrinsic_sin = (double (*)(double)) __ pc();
2961
2962 __ subq(rsp, 8);
2963 __ movdbl(Address(rsp, 0), xmm0);
2964 __ fld_d(Address(rsp, 0));
2965 __ trigfunc('s');
2966 __ fstp_d(Address(rsp, 0));
2967 __ movdbl(xmm0, Address(rsp, 0));
2968 __ addq(rsp, 8);
2969 __ ret(0);
2970 }
2971 {
2972 StubCodeMark mark(this, "StubRoutines", "cos");
2973 StubRoutines::_intrinsic_cos = (double (*)(double)) __ pc();
2974
2975 __ subq(rsp, 8);
2976 __ movdbl(Address(rsp, 0), xmm0);
2977 __ fld_d(Address(rsp, 0));
2978 __ trigfunc('c');
2979 __ fstp_d(Address(rsp, 0));
2980 __ movdbl(xmm0, Address(rsp, 0));
2981 __ addq(rsp, 8);
2982 __ ret(0);
2983 }
2984 {
2985 StubCodeMark mark(this, "StubRoutines", "tan");
2986 StubRoutines::_intrinsic_tan = (double (*)(double)) __ pc();
2987
2988 __ subq(rsp, 8);
2989 __ movdbl(Address(rsp, 0), xmm0);
2990 __ fld_d(Address(rsp, 0));
2991 __ trigfunc('t');
2992 __ fstp_d(Address(rsp, 0));
2993 __ movdbl(xmm0, Address(rsp, 0));
2994 __ addq(rsp, 8);
2995 __ ret(0);
2996 }
2997 {
2998 StubCodeMark mark(this, "StubRoutines", "exp");
2999 StubRoutines::_intrinsic_exp = (double (*)(double)) __ pc();
3000
3001 __ subq(rsp, 8);
3002 __ movdbl(Address(rsp, 0), xmm0);
3003 __ fld_d(Address(rsp, 0));
3004 __ exp_with_fallback(0);
3005 __ fstp_d(Address(rsp, 0));
3006 __ movdbl(xmm0, Address(rsp, 0));
3007 __ addq(rsp, 8);
3008 __ ret(0);
3009 }
3010 {
3011 StubCodeMark mark(this, "StubRoutines", "pow");
3012 StubRoutines::_intrinsic_pow = (double (*)(double,double)) __ pc();
3013
3014 __ subq(rsp, 8);
3015 __ movdbl(Address(rsp, 0), xmm1);
3016 __ fld_d(Address(rsp, 0));
3017 __ movdbl(Address(rsp, 0), xmm0);
3018 __ fld_d(Address(rsp, 0));
3019 __ pow_with_fallback(0);
3020 __ fstp_d(Address(rsp, 0));
3021 __ movdbl(xmm0, Address(rsp, 0));
3022 __ addq(rsp, 8);
3023 __ ret(0);
3024 }
3025 }
3026
3027 // AES intrinsic stubs
3028 enum {AESBlockSize = 16};
3029
3030 address generate_key_shuffle_mask() {
3031 __ align(16);
3032 StubCodeMark mark(this, "StubRoutines", "key_shuffle_mask");
3033 address start = __ pc();
3034 __ emit_data64( 0x0405060700010203, relocInfo::none );
3035 __ emit_data64( 0x0c0d0e0f08090a0b, relocInfo::none );
3036 return start;
3037 }
3038
3039 // Utility routine for loading a 128-bit key word in little endian format
3040 // can optionally specify that the shuffle mask is already in an xmmregister
3041 void load_key(XMMRegister xmmdst, Register key, int offset, XMMRegister xmm_shuf_mask=NULL) {
3042 __ movdqu(xmmdst, Address(key, offset));
3043 if (xmm_shuf_mask != NULL) {
3044 __ pshufb(xmmdst, xmm_shuf_mask);
3045 } else {
3046 __ pshufb(xmmdst, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3047 }
3048 }
3049
3050 // Arguments:
3051 //
3052 // Inputs:
3053 // c_rarg0 - source byte array address
3054 // c_rarg1 - destination byte array address
3055 // c_rarg2 - K (key) in little endian int array
3056 //
3057 address generate_aescrypt_encryptBlock() {
3058 assert(UseAES, "need AES instructions and misaligned SSE support");
3059 __ align(CodeEntryAlignment);
3060 StubCodeMark mark(this, "StubRoutines", "aescrypt_encryptBlock");
3061 Label L_doLast;
3062 address start = __ pc();
3063
3064 const Register from = c_rarg0; // source array address
3065 const Register to = c_rarg1; // destination array address
3066 const Register key = c_rarg2; // key array address
3067 const Register keylen = rax;
3068
3069 const XMMRegister xmm_result = xmm0;
3070 const XMMRegister xmm_key_shuf_mask = xmm1;
3071 // On win64 xmm6-xmm15 must be preserved so don't use them.
3072 const XMMRegister xmm_temp1 = xmm2;
3073 const XMMRegister xmm_temp2 = xmm3;
3074 const XMMRegister xmm_temp3 = xmm4;
3075 const XMMRegister xmm_temp4 = xmm5;
3076
3077 __ enter(); // required for proper stackwalking of RuntimeStub frame
3078
3079 // keylen could be only {11, 13, 15} * 4 = {44, 52, 60}
3080 __ movl(keylen, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3081
3082 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3083 __ movdqu(xmm_result, Address(from, 0)); // get 16 bytes of input
3084
3085 // For encryption, the java expanded key ordering is just what we need
3086 // we don't know if the key is aligned, hence not using load-execute form
3087
3088 load_key(xmm_temp1, key, 0x00, xmm_key_shuf_mask);
3089 __ pxor(xmm_result, xmm_temp1);
3090
3091 load_key(xmm_temp1, key, 0x10, xmm_key_shuf_mask);
3092 load_key(xmm_temp2, key, 0x20, xmm_key_shuf_mask);
3093 load_key(xmm_temp3, key, 0x30, xmm_key_shuf_mask);
3094 load_key(xmm_temp4, key, 0x40, xmm_key_shuf_mask);
3095
3096 __ aesenc(xmm_result, xmm_temp1);
3097 __ aesenc(xmm_result, xmm_temp2);
3098 __ aesenc(xmm_result, xmm_temp3);
3099 __ aesenc(xmm_result, xmm_temp4);
3100
3101 load_key(xmm_temp1, key, 0x50, xmm_key_shuf_mask);
3102 load_key(xmm_temp2, key, 0x60, xmm_key_shuf_mask);
3103 load_key(xmm_temp3, key, 0x70, xmm_key_shuf_mask);
3104 load_key(xmm_temp4, key, 0x80, xmm_key_shuf_mask);
3105
3106 __ aesenc(xmm_result, xmm_temp1);
3107 __ aesenc(xmm_result, xmm_temp2);
3108 __ aesenc(xmm_result, xmm_temp3);
3109 __ aesenc(xmm_result, xmm_temp4);
3110
3111 load_key(xmm_temp1, key, 0x90, xmm_key_shuf_mask);
3112 load_key(xmm_temp2, key, 0xa0, xmm_key_shuf_mask);
3113
3114 __ cmpl(keylen, 44);
3115 __ jccb(Assembler::equal, L_doLast);
3116
3117 __ aesenc(xmm_result, xmm_temp1);
3118 __ aesenc(xmm_result, xmm_temp2);
3119
3120 load_key(xmm_temp1, key, 0xb0, xmm_key_shuf_mask);
3121 load_key(xmm_temp2, key, 0xc0, xmm_key_shuf_mask);
3122
3123 __ cmpl(keylen, 52);
3124 __ jccb(Assembler::equal, L_doLast);
3125
3126 __ aesenc(xmm_result, xmm_temp1);
3127 __ aesenc(xmm_result, xmm_temp2);
3128
3129 load_key(xmm_temp1, key, 0xd0, xmm_key_shuf_mask);
3130 load_key(xmm_temp2, key, 0xe0, xmm_key_shuf_mask);
3131
3132 __ BIND(L_doLast);
3133 __ aesenc(xmm_result, xmm_temp1);
3134 __ aesenclast(xmm_result, xmm_temp2);
3135 __ movdqu(Address(to, 0), xmm_result); // store the result
3136 __ xorptr(rax, rax); // return 0
3137 __ leave(); // required for proper stackwalking of RuntimeStub frame
3138 __ ret(0);
3139
3140 return start;
3141 }
3142
3143
3144 // Arguments:
3145 //
3146 // Inputs:
3147 // c_rarg0 - source byte array address
3148 // c_rarg1 - destination byte array address
3149 // c_rarg2 - K (key) in little endian int array
3150 //
3151 address generate_aescrypt_decryptBlock() {
3152 assert(UseAES, "need AES instructions and misaligned SSE support");
3153 __ align(CodeEntryAlignment);
3154 StubCodeMark mark(this, "StubRoutines", "aescrypt_decryptBlock");
3155 Label L_doLast;
3156 address start = __ pc();
3157
3158 const Register from = c_rarg0; // source array address
3159 const Register to = c_rarg1; // destination array address
3160 const Register key = c_rarg2; // key array address
3161 const Register keylen = rax;
3162
3163 const XMMRegister xmm_result = xmm0;
3164 const XMMRegister xmm_key_shuf_mask = xmm1;
3165 // On win64 xmm6-xmm15 must be preserved so don't use them.
3166 const XMMRegister xmm_temp1 = xmm2;
3167 const XMMRegister xmm_temp2 = xmm3;
3168 const XMMRegister xmm_temp3 = xmm4;
3169 const XMMRegister xmm_temp4 = xmm5;
3170
3171 __ enter(); // required for proper stackwalking of RuntimeStub frame
3172
3173 // keylen could be only {11, 13, 15} * 4 = {44, 52, 60}
3174 __ movl(keylen, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3175
3176 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3177 __ movdqu(xmm_result, Address(from, 0));
3178
3179 // for decryption java expanded key ordering is rotated one position from what we want
3180 // so we start from 0x10 here and hit 0x00 last
3181 // we don't know if the key is aligned, hence not using load-execute form
3182 load_key(xmm_temp1, key, 0x10, xmm_key_shuf_mask);
3183 load_key(xmm_temp2, key, 0x20, xmm_key_shuf_mask);
3184 load_key(xmm_temp3, key, 0x30, xmm_key_shuf_mask);
3185 load_key(xmm_temp4, key, 0x40, xmm_key_shuf_mask);
3186
3187 __ pxor (xmm_result, xmm_temp1);
3188 __ aesdec(xmm_result, xmm_temp2);
3189 __ aesdec(xmm_result, xmm_temp3);
3190 __ aesdec(xmm_result, xmm_temp4);
3191
3192 load_key(xmm_temp1, key, 0x50, xmm_key_shuf_mask);
3193 load_key(xmm_temp2, key, 0x60, xmm_key_shuf_mask);
3194 load_key(xmm_temp3, key, 0x70, xmm_key_shuf_mask);
3195 load_key(xmm_temp4, key, 0x80, xmm_key_shuf_mask);
3196
3197 __ aesdec(xmm_result, xmm_temp1);
3198 __ aesdec(xmm_result, xmm_temp2);
3199 __ aesdec(xmm_result, xmm_temp3);
3200 __ aesdec(xmm_result, xmm_temp4);
3201
3202 load_key(xmm_temp1, key, 0x90, xmm_key_shuf_mask);
3203 load_key(xmm_temp2, key, 0xa0, xmm_key_shuf_mask);
3204 load_key(xmm_temp3, key, 0x00, xmm_key_shuf_mask);
3205
3206 __ cmpl(keylen, 44);
3207 __ jccb(Assembler::equal, L_doLast);
3208
3209 __ aesdec(xmm_result, xmm_temp1);
3210 __ aesdec(xmm_result, xmm_temp2);
3211
3212 load_key(xmm_temp1, key, 0xb0, xmm_key_shuf_mask);
3213 load_key(xmm_temp2, key, 0xc0, xmm_key_shuf_mask);
3214
3215 __ cmpl(keylen, 52);
3216 __ jccb(Assembler::equal, L_doLast);
3217
3218 __ aesdec(xmm_result, xmm_temp1);
3219 __ aesdec(xmm_result, xmm_temp2);
3220
3221 load_key(xmm_temp1, key, 0xd0, xmm_key_shuf_mask);
3222 load_key(xmm_temp2, key, 0xe0, xmm_key_shuf_mask);
3223
3224 __ BIND(L_doLast);
3225 __ aesdec(xmm_result, xmm_temp1);
3226 __ aesdec(xmm_result, xmm_temp2);
3227
3228 // for decryption the aesdeclast operation is always on key+0x00
3229 __ aesdeclast(xmm_result, xmm_temp3);
3230 __ movdqu(Address(to, 0), xmm_result); // store the result
3231 __ xorptr(rax, rax); // return 0
3232 __ leave(); // required for proper stackwalking of RuntimeStub frame
3233 __ ret(0);
3234
3235 return start;
3236 }
3237
3238
3239 // Arguments:
3240 //
3241 // Inputs:
3242 // c_rarg0 - source byte array address
3243 // c_rarg1 - destination byte array address
3244 // c_rarg2 - K (key) in little endian int array
3245 // c_rarg3 - r vector byte array address
3246 // c_rarg4 - input length
3247 //
3248 address generate_cipherBlockChaining_encryptAESCrypt() {
3249 assert(UseAES, "need AES instructions and misaligned SSE support");
3250 __ align(CodeEntryAlignment);
3251 StubCodeMark mark(this, "StubRoutines", "cipherBlockChaining_encryptAESCrypt");
3252 address start = __ pc();
3253
3254 Label L_exit, L_key_192_256, L_key_256, L_loopTop_128, L_loopTop_192, L_loopTop_256;
3255 const Register from = c_rarg0; // source array address
3256 const Register to = c_rarg1; // destination array address
3257 const Register key = c_rarg2; // key array address
3258 const Register rvec = c_rarg3; // r byte array initialized from initvector array address
3259 // and left with the results of the last encryption block
3260 #ifndef _WIN64
3261 const Register len_reg = c_rarg4; // src len (must be multiple of blocksize 16)
3262 #else
3263 const Address len_mem(rsp, 6 * wordSize); // length is on stack on Win64
3264 const Register len_reg = r10; // pick the first volatile windows register
3265 #endif
3266 const Register pos = rax;
3267
3268 // xmm register assignments for the loops below
3269 const XMMRegister xmm_result = xmm0;
3270 const XMMRegister xmm_temp = xmm1;
3271 // keys 0-10 preloaded into xmm2-xmm12
3272 const int XMM_REG_NUM_KEY_FIRST = 2;
3273 const int XMM_REG_NUM_KEY_LAST = 15;
3274 const XMMRegister xmm_key0 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST);
3275 const XMMRegister xmm_key10 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+10);
3276 const XMMRegister xmm_key11 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+11);
3277 const XMMRegister xmm_key12 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+12);
3278 const XMMRegister xmm_key13 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+13);
3279
3280 __ enter(); // required for proper stackwalking of RuntimeStub frame
3281
3282 #ifdef _WIN64
3283 // on win64, fill len_reg from stack position
3284 __ movl(len_reg, len_mem);
3285 // save the xmm registers which must be preserved 6-15
3286 __ subptr(rsp, -rsp_after_call_off * wordSize);
3287 for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3288 __ movdqu(xmm_save(i), as_XMMRegister(i));
3289 }
3290 #endif
3291
3292 const XMMRegister xmm_key_shuf_mask = xmm_temp; // used temporarily to swap key bytes up front
3293 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3294 // load up xmm regs xmm2 thru xmm12 with key 0x00 - 0xa0
3295 for (int rnum = XMM_REG_NUM_KEY_FIRST, offset = 0x00; rnum <= XMM_REG_NUM_KEY_FIRST+10; rnum++) {
3296 load_key(as_XMMRegister(rnum), key, offset, xmm_key_shuf_mask);
3297 offset += 0x10;
3298 }
3299 __ movdqu(xmm_result, Address(rvec, 0x00)); // initialize xmm_result with r vec
3300
3301 // now split to different paths depending on the keylen (len in ints of AESCrypt.KLE array (52=192, or 60=256))
3302 __ movl(rax, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3303 __ cmpl(rax, 44);
3304 __ jcc(Assembler::notEqual, L_key_192_256);
3305
3306 // 128 bit code follows here
3307 __ movptr(pos, 0);
3308 __ align(OptoLoopAlignment);
3309
3310 __ BIND(L_loopTop_128);
3311 __ movdqu(xmm_temp, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of input
3312 __ pxor (xmm_result, xmm_temp); // xor with the current r vector
3313 __ pxor (xmm_result, xmm_key0); // do the aes rounds
3314 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_FIRST + 9; rnum++) {
3315 __ aesenc(xmm_result, as_XMMRegister(rnum));
3316 }
3317 __ aesenclast(xmm_result, xmm_key10);
3318 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3319 // no need to store r to memory until we exit
3320 __ addptr(pos, AESBlockSize);
3321 __ subptr(len_reg, AESBlockSize);
3322 __ jcc(Assembler::notEqual, L_loopTop_128);
3323
3324 __ BIND(L_exit);
3325 __ movdqu(Address(rvec, 0), xmm_result); // final value of r stored in rvec of CipherBlockChaining object
3326
3327 #ifdef _WIN64
3328 // restore xmm regs belonging to calling function
3329 for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3330 __ movdqu(as_XMMRegister(i), xmm_save(i));
3331 }
3332 #endif
3333 __ movl(rax, 0); // return 0 (why?)
3334 __ leave(); // required for proper stackwalking of RuntimeStub frame
3335 __ ret(0);
3336
3337 __ BIND(L_key_192_256);
3338 // here rax = len in ints of AESCrypt.KLE array (52=192, or 60=256)
3339 load_key(xmm_key11, key, 0xb0, xmm_key_shuf_mask);
3340 load_key(xmm_key12, key, 0xc0, xmm_key_shuf_mask);
3341 __ cmpl(rax, 52);
3342 __ jcc(Assembler::notEqual, L_key_256);
3343
3344 // 192-bit code follows here (could be changed to use more xmm registers)
3345 __ movptr(pos, 0);
3346 __ align(OptoLoopAlignment);
3347
3348 __ BIND(L_loopTop_192);
3349 __ movdqu(xmm_temp, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of input
3350 __ pxor (xmm_result, xmm_temp); // xor with the current r vector
3351 __ pxor (xmm_result, xmm_key0); // do the aes rounds
3352 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_FIRST + 11; rnum++) {
3353 __ aesenc(xmm_result, as_XMMRegister(rnum));
3354 }
3355 __ aesenclast(xmm_result, xmm_key12);
3356 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3357 // no need to store r to memory until we exit
3358 __ addptr(pos, AESBlockSize);
3359 __ subptr(len_reg, AESBlockSize);
3360 __ jcc(Assembler::notEqual, L_loopTop_192);
3361 __ jmp(L_exit);
3362
3363 __ BIND(L_key_256);
3364 // 256-bit code follows here (could be changed to use more xmm registers)
3365 load_key(xmm_key13, key, 0xd0, xmm_key_shuf_mask);
3366 __ movptr(pos, 0);
3367 __ align(OptoLoopAlignment);
3368
3369 __ BIND(L_loopTop_256);
3370 __ movdqu(xmm_temp, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of input
3371 __ pxor (xmm_result, xmm_temp); // xor with the current r vector
3372 __ pxor (xmm_result, xmm_key0); // do the aes rounds
3373 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_FIRST + 13; rnum++) {
3374 __ aesenc(xmm_result, as_XMMRegister(rnum));
3375 }
3376 load_key(xmm_temp, key, 0xe0);
3377 __ aesenclast(xmm_result, xmm_temp);
3378 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3379 // no need to store r to memory until we exit
3380 __ addptr(pos, AESBlockSize);
3381 __ subptr(len_reg, AESBlockSize);
3382 __ jcc(Assembler::notEqual, L_loopTop_256);
3383 __ jmp(L_exit);
3384
3385 return start;
3386 }
3387
3388
3389
3390 // This is a version of CBC/AES Decrypt which does 4 blocks in a loop at a time
3391 // to hide instruction latency
3392 //
3393 // Arguments:
3394 //
3395 // Inputs:
3396 // c_rarg0 - source byte array address
3397 // c_rarg1 - destination byte array address
3398 // c_rarg2 - K (key) in little endian int array
3399 // c_rarg3 - r vector byte array address
3400 // c_rarg4 - input length
3401 //
3402
3403 address generate_cipherBlockChaining_decryptAESCrypt_Parallel() {
3404 assert(UseAES, "need AES instructions and misaligned SSE support");
3405 __ align(CodeEntryAlignment);
3406 StubCodeMark mark(this, "StubRoutines", "cipherBlockChaining_decryptAESCrypt");
3407 address start = __ pc();
3408
3409 Label L_exit, L_key_192_256, L_key_256;
3410 Label L_singleBlock_loopTop_128, L_multiBlock_loopTop_128;
3411 Label L_singleBlock_loopTop_192, L_singleBlock_loopTop_256;
3412 const Register from = c_rarg0; // source array address
3413 const Register to = c_rarg1; // destination array address
3414 const Register key = c_rarg2; // key array address
3415 const Register rvec = c_rarg3; // r byte array initialized from initvector array address
3416 // and left with the results of the last encryption block
3417 #ifndef _WIN64
3418 const Register len_reg = c_rarg4; // src len (must be multiple of blocksize 16)
3419 #else
3420 const Address len_mem(rsp, 6 * wordSize); // length is on stack on Win64
3421 const Register len_reg = r10; // pick the first volatile windows register
3422 #endif
3423 const Register pos = rax;
3424
3425 // keys 0-10 preloaded into xmm2-xmm12
3426 const int XMM_REG_NUM_KEY_FIRST = 5;
3427 const int XMM_REG_NUM_KEY_LAST = 15;
3428 const XMMRegister xmm_key_first = as_XMMRegister(XMM_REG_NUM_KEY_FIRST);
3429 const XMMRegister xmm_key_last = as_XMMRegister(XMM_REG_NUM_KEY_LAST);
3430
3431 __ enter(); // required for proper stackwalking of RuntimeStub frame
3432
3433 #ifdef _WIN64
3434 // on win64, fill len_reg from stack position
3435 __ movl(len_reg, len_mem);
3436 // save the xmm registers which must be preserved 6-15
3437 __ subptr(rsp, -rsp_after_call_off * wordSize);
3438 for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3439 __ movdqu(xmm_save(i), as_XMMRegister(i));
3440 }
3441 #endif
3442 // the java expanded key ordering is rotated one position from what we want
3443 // so we start from 0x10 here and hit 0x00 last
3444 const XMMRegister xmm_key_shuf_mask = xmm1; // used temporarily to swap key bytes up front
3445 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3446 // load up xmm regs 5 thru 15 with key 0x10 - 0xa0 - 0x00
3447 for (int rnum = XMM_REG_NUM_KEY_FIRST, offset = 0x10; rnum < XMM_REG_NUM_KEY_LAST; rnum++) {
3448 load_key(as_XMMRegister(rnum), key, offset, xmm_key_shuf_mask);
3449 offset += 0x10;
3450 }
3451 load_key(xmm_key_last, key, 0x00, xmm_key_shuf_mask);
3452
3453 const XMMRegister xmm_prev_block_cipher = xmm1; // holds cipher of previous block
3454
3455 // registers holding the four results in the parallelized loop
3456 const XMMRegister xmm_result0 = xmm0;
3457 const XMMRegister xmm_result1 = xmm2;
3458 const XMMRegister xmm_result2 = xmm3;
3459 const XMMRegister xmm_result3 = xmm4;
3460
3461 __ movdqu(xmm_prev_block_cipher, Address(rvec, 0x00)); // initialize with initial rvec
3462
3463 // now split to different paths depending on the keylen (len in ints of AESCrypt.KLE array (52=192, or 60=256))
3464 __ movl(rax, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3465 __ cmpl(rax, 44);
3466 __ jcc(Assembler::notEqual, L_key_192_256);
3467
3468
3469 // 128-bit code follows here, parallelized
3470 __ movptr(pos, 0);
3471 __ align(OptoLoopAlignment);
3472 __ BIND(L_multiBlock_loopTop_128);
3473 __ cmpptr(len_reg, 4*AESBlockSize); // see if at least 4 blocks left
3474 __ jcc(Assembler::less, L_singleBlock_loopTop_128);
3475
3476 __ movdqu(xmm_result0, Address(from, pos, Address::times_1, 0*AESBlockSize)); // get next 4 blocks into xmmresult registers
3477 __ movdqu(xmm_result1, Address(from, pos, Address::times_1, 1*AESBlockSize));
3478 __ movdqu(xmm_result2, Address(from, pos, Address::times_1, 2*AESBlockSize));
3479 __ movdqu(xmm_result3, Address(from, pos, Address::times_1, 3*AESBlockSize));
3480
3481 #define DoFour(opc, src_reg) \
3482 __ opc(xmm_result0, src_reg); \
3483 __ opc(xmm_result1, src_reg); \
3484 __ opc(xmm_result2, src_reg); \
3485 __ opc(xmm_result3, src_reg);
3486
3487 DoFour(pxor, xmm_key_first);
3488 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3489 DoFour(aesdec, as_XMMRegister(rnum));
3490 }
3491 DoFour(aesdeclast, xmm_key_last);
3492 // for each result, xor with the r vector of previous cipher block
3493 __ pxor(xmm_result0, xmm_prev_block_cipher);
3494 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 0*AESBlockSize));
3495 __ pxor(xmm_result1, xmm_prev_block_cipher);
3496 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 1*AESBlockSize));
3497 __ pxor(xmm_result2, xmm_prev_block_cipher);
3498 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 2*AESBlockSize));
3499 __ pxor(xmm_result3, xmm_prev_block_cipher);
3500 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 3*AESBlockSize)); // this will carry over to next set of blocks
3501
3502 __ movdqu(Address(to, pos, Address::times_1, 0*AESBlockSize), xmm_result0); // store 4 results into the next 64 bytes of output
3503 __ movdqu(Address(to, pos, Address::times_1, 1*AESBlockSize), xmm_result1);
3504 __ movdqu(Address(to, pos, Address::times_1, 2*AESBlockSize), xmm_result2);
3505 __ movdqu(Address(to, pos, Address::times_1, 3*AESBlockSize), xmm_result3);
3506
3507 __ addptr(pos, 4*AESBlockSize);
3508 __ subptr(len_reg, 4*AESBlockSize);
3509 __ jmp(L_multiBlock_loopTop_128);
3510
3511 // registers used in the non-parallelized loops
3512 // xmm register assignments for the loops below
3513 const XMMRegister xmm_result = xmm0;
3514 const XMMRegister xmm_prev_block_cipher_save = xmm2;
3515 const XMMRegister xmm_key11 = xmm3;
3516 const XMMRegister xmm_key12 = xmm4;
3517 const XMMRegister xmm_temp = xmm4;
3518
3519 __ align(OptoLoopAlignment);
3520 __ BIND(L_singleBlock_loopTop_128);
3521 __ cmpptr(len_reg, 0); // any blocks left??
3522 __ jcc(Assembler::equal, L_exit);
3523 __ movdqu(xmm_result, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of cipher input
3524 __ movdqa(xmm_prev_block_cipher_save, xmm_result); // save for next r vector
3525 __ pxor (xmm_result, xmm_key_first); // do the aes dec rounds
3526 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3527 __ aesdec(xmm_result, as_XMMRegister(rnum));
3528 }
3529 __ aesdeclast(xmm_result, xmm_key_last);
3530 __ pxor (xmm_result, xmm_prev_block_cipher); // xor with the current r vector
3531 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3532 // no need to store r to memory until we exit
3533 __ movdqa(xmm_prev_block_cipher, xmm_prev_block_cipher_save); // set up next r vector with cipher input from this block
3534
3535 __ addptr(pos, AESBlockSize);
3536 __ subptr(len_reg, AESBlockSize);
3537 __ jmp(L_singleBlock_loopTop_128);
3538
3539
3540 __ BIND(L_exit);
3541 __ movdqu(Address(rvec, 0), xmm_prev_block_cipher); // final value of r stored in rvec of CipherBlockChaining object
3542 #ifdef _WIN64
3543 // restore regs belonging to calling function
3544 for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3545 __ movdqu(as_XMMRegister(i), xmm_save(i));
3546 }
3547 #endif
3548 __ movl(rax, 0); // return 0 (why?)
3549 __ leave(); // required for proper stackwalking of RuntimeStub frame
3550 __ ret(0);
3551
3552
3553 __ BIND(L_key_192_256);
3554 // here rax = len in ints of AESCrypt.KLE array (52=192, or 60=256)
3555 load_key(xmm_key11, key, 0xb0);
3556 __ cmpl(rax, 52);
3557 __ jcc(Assembler::notEqual, L_key_256);
3558
3559 // 192-bit code follows here (could be optimized to use parallelism)
3560 load_key(xmm_key12, key, 0xc0); // 192-bit key goes up to c0
3561 __ movptr(pos, 0);
3562 __ align(OptoLoopAlignment);
3563
3564 __ BIND(L_singleBlock_loopTop_192);
3565 __ movdqu(xmm_result, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of cipher input
3566 __ movdqa(xmm_prev_block_cipher_save, xmm_result); // save for next r vector
3567 __ pxor (xmm_result, xmm_key_first); // do the aes dec rounds
3568 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3569 __ aesdec(xmm_result, as_XMMRegister(rnum));
3570 }
3571 __ aesdec(xmm_result, xmm_key11);
3572 __ aesdec(xmm_result, xmm_key12);
3573 __ aesdeclast(xmm_result, xmm_key_last); // xmm15 always came from key+0
3574 __ pxor (xmm_result, xmm_prev_block_cipher); // xor with the current r vector
3575 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3576 // no need to store r to memory until we exit
3577 __ movdqa(xmm_prev_block_cipher, xmm_prev_block_cipher_save); // set up next r vector with cipher input from this block
3578 __ addptr(pos, AESBlockSize);
3579 __ subptr(len_reg, AESBlockSize);
3580 __ jcc(Assembler::notEqual,L_singleBlock_loopTop_192);
3581 __ jmp(L_exit);
3582
3583 __ BIND(L_key_256);
3584 // 256-bit code follows here (could be optimized to use parallelism)
3585 __ movptr(pos, 0);
3586 __ align(OptoLoopAlignment);
3587
3588 __ BIND(L_singleBlock_loopTop_256);
3589 __ movdqu(xmm_result, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of cipher input
3590 __ movdqa(xmm_prev_block_cipher_save, xmm_result); // save for next r vector
3591 __ pxor (xmm_result, xmm_key_first); // do the aes dec rounds
3592 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3593 __ aesdec(xmm_result, as_XMMRegister(rnum));
3594 }
3595 __ aesdec(xmm_result, xmm_key11);
3596 load_key(xmm_temp, key, 0xc0);
3597 __ aesdec(xmm_result, xmm_temp);
3598 load_key(xmm_temp, key, 0xd0);
3599 __ aesdec(xmm_result, xmm_temp);
3600 load_key(xmm_temp, key, 0xe0); // 256-bit key goes up to e0
3601 __ aesdec(xmm_result, xmm_temp);
3602 __ aesdeclast(xmm_result, xmm_key_last); // xmm15 came from key+0
3603 __ pxor (xmm_result, xmm_prev_block_cipher); // xor with the current r vector
3604 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3605 // no need to store r to memory until we exit
3606 __ movdqa(xmm_prev_block_cipher, xmm_prev_block_cipher_save); // set up next r vector with cipher input from this block
3607 __ addptr(pos, AESBlockSize);
3608 __ subptr(len_reg, AESBlockSize);
3609 __ jcc(Assembler::notEqual,L_singleBlock_loopTop_256);
3610 __ jmp(L_exit);
3611
3612 return start;
3613 }
3614
3615
3616
3617 #undef __
3618 #define __ masm->
3619
3620 // Continuation point for throwing of implicit exceptions that are
3621 // not handled in the current activation. Fabricates an exception
3622 // oop and initiates normal exception dispatching in this
3623 // frame. Since we need to preserve callee-saved values (currently
3624 // only for C2, but done for C1 as well) we need a callee-saved oop
3625 // map and therefore have to make these stubs into RuntimeStubs
3626 // rather than BufferBlobs. If the compiler needs all registers to
3627 // be preserved between the fault point and the exception handler
3628 // then it must assume responsibility for that in
3629 // AbstractCompiler::continuation_for_implicit_null_exception or
3630 // continuation_for_implicit_division_by_zero_exception. All other
3631 // implicit exceptions (e.g., NullPointerException or
3632 // AbstractMethodError on entry) are either at call sites or
3633 // otherwise assume that stack unwinding will be initiated, so
3634 // caller saved registers were assumed volatile in the compiler.
3635 address generate_throw_exception(const char* name,
3636 address runtime_entry,
3637 Register arg1 = noreg,
3638 Register arg2 = noreg) {
3639 // Information about frame layout at time of blocking runtime call.
3640 // Note that we only have to preserve callee-saved registers since
3641 // the compilers are responsible for supplying a continuation point
3642 // if they expect all registers to be preserved.
3643 enum layout {
3644 rbp_off = frame::arg_reg_save_area_bytes/BytesPerInt,
3645 rbp_off2,
3646 return_off,
3647 return_off2,
3648 framesize // inclusive of return address
3649 };
3650
3651 int insts_size = 512;
3652 int locs_size = 64;
3653
3654 CodeBuffer code(name, insts_size, locs_size);
3655 OopMapSet* oop_maps = new OopMapSet();
3656 MacroAssembler* masm = new MacroAssembler(&code);
3657
3658 address start = __ pc();
3659
3660 // This is an inlined and slightly modified version of call_VM
3661 // which has the ability to fetch the return PC out of
3662 // thread-local storage and also sets up last_Java_sp slightly
3663 // differently than the real call_VM
3664
3665 __ enter(); // required for proper stackwalking of RuntimeStub frame
3666
3667 assert(is_even(framesize/2), "sp not 16-byte aligned");
3668
3669 // return address and rbp are already in place
3670 __ subptr(rsp, (framesize-4) << LogBytesPerInt); // prolog
3671
3672 int frame_complete = __ pc() - start;
3673
3674 // Set up last_Java_sp and last_Java_fp
3675 address the_pc = __ pc();
3676 __ set_last_Java_frame(rsp, rbp, the_pc);
3677 __ andptr(rsp, -(StackAlignmentInBytes)); // Align stack
3678
3679 // Call runtime
3680 if (arg1 != noreg) {
3681 assert(arg2 != c_rarg1, "clobbered");
3682 __ movptr(c_rarg1, arg1);
3683 }
3684 if (arg2 != noreg) {
3685 __ movptr(c_rarg2, arg2);
3686 }
3687 __ movptr(c_rarg0, r15_thread);
3688 BLOCK_COMMENT("call runtime_entry");
3689 __ call(RuntimeAddress(runtime_entry));
3690
3691 // Generate oop map
3692 OopMap* map = new OopMap(framesize, 0);
3693
3694 oop_maps->add_gc_map(the_pc - start, map);
3695
3696 __ reset_last_Java_frame(true, true);
3697
3698 __ leave(); // required for proper stackwalking of RuntimeStub frame
3699
3700 // check for pending exceptions
3701 #ifdef ASSERT
3702 Label L;
3703 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()),
3704 (int32_t) NULL_WORD);
3705 __ jcc(Assembler::notEqual, L);
3706 __ should_not_reach_here();
3707 __ bind(L);
3708 #endif // ASSERT
3709 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3710
3711
3712 // codeBlob framesize is in words (not VMRegImpl::slot_size)
3713 RuntimeStub* stub =
3714 RuntimeStub::new_runtime_stub(name,
3715 &code,
3716 frame_complete,
3717 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3718 oop_maps, false);
3719 return stub->entry_point();
3720 }
3721
3722 // Initialization
3723 void generate_initial() {
3724 // Generates all stubs and initializes the entry points
3725
3726 // This platform-specific stub is needed by generate_call_stub()
3727 StubRoutines::x86::_mxcsr_std = generate_fp_mask("mxcsr_std", 0x0000000000001F80);
3728
3729 // entry points that exist in all platforms Note: This is code
3730 // that could be shared among different platforms - however the
3731 // benefit seems to be smaller than the disadvantage of having a
3732 // much more complicated generator structure. See also comment in
3733 // stubRoutines.hpp.
3734
3735 StubRoutines::_forward_exception_entry = generate_forward_exception();
3736
3737 StubRoutines::_call_stub_entry =
3738 generate_call_stub(StubRoutines::_call_stub_return_address);
3739
3740 // is referenced by megamorphic call
3741 StubRoutines::_catch_exception_entry = generate_catch_exception();
3742
3743 // atomic calls
3744 StubRoutines::_atomic_xchg_entry = generate_atomic_xchg();
3745 StubRoutines::_atomic_xchg_ptr_entry = generate_atomic_xchg_ptr();
3746 StubRoutines::_atomic_cmpxchg_entry = generate_atomic_cmpxchg();
3747 StubRoutines::_atomic_cmpxchg_long_entry = generate_atomic_cmpxchg_long();
3748 StubRoutines::_atomic_add_entry = generate_atomic_add();
3749 StubRoutines::_atomic_add_ptr_entry = generate_atomic_add_ptr();
3750 StubRoutines::_fence_entry = generate_orderaccess_fence();
3751
3752 StubRoutines::_handler_for_unsafe_access_entry =
3753 generate_handler_for_unsafe_access();
3754
3755 // platform dependent
3756 StubRoutines::x86::_get_previous_fp_entry = generate_get_previous_fp();
3757 StubRoutines::x86::_get_previous_sp_entry = generate_get_previous_sp();
3758
3759 StubRoutines::x86::_verify_mxcsr_entry = generate_verify_mxcsr();
3760
3761 // Build this early so it's available for the interpreter.
3762 StubRoutines::_throw_StackOverflowError_entry =
3763 generate_throw_exception("StackOverflowError throw_exception",
3764 CAST_FROM_FN_PTR(address,
3765 SharedRuntime::
3766 throw_StackOverflowError));
3767 }
3768
3769 void generate_all() {
3770 // Generates all stubs and initializes the entry points
3771
3772 // These entry points require SharedInfo::stack0 to be set up in
3773 // non-core builds and need to be relocatable, so they each
3774 // fabricate a RuntimeStub internally.
3775 StubRoutines::_throw_AbstractMethodError_entry =
3776 generate_throw_exception("AbstractMethodError throw_exception",
3777 CAST_FROM_FN_PTR(address,
3778 SharedRuntime::
3779 throw_AbstractMethodError));
3780
3781 StubRoutines::_throw_IncompatibleClassChangeError_entry =
3782 generate_throw_exception("IncompatibleClassChangeError throw_exception",
3783 CAST_FROM_FN_PTR(address,
3784 SharedRuntime::
3785 throw_IncompatibleClassChangeError));
3786
3787 StubRoutines::_throw_NullPointerException_at_call_entry =
3788 generate_throw_exception("NullPointerException at call throw_exception",
3789 CAST_FROM_FN_PTR(address,
3790 SharedRuntime::
3791 throw_NullPointerException_at_call));
3792
3793 // entry points that are platform specific
3794 StubRoutines::x86::_f2i_fixup = generate_f2i_fixup();
3795 StubRoutines::x86::_f2l_fixup = generate_f2l_fixup();
3796 StubRoutines::x86::_d2i_fixup = generate_d2i_fixup();
3797 StubRoutines::x86::_d2l_fixup = generate_d2l_fixup();
3798
3799 StubRoutines::x86::_float_sign_mask = generate_fp_mask("float_sign_mask", 0x7FFFFFFF7FFFFFFF);
3800 StubRoutines::x86::_float_sign_flip = generate_fp_mask("float_sign_flip", 0x8000000080000000);
3801 StubRoutines::x86::_double_sign_mask = generate_fp_mask("double_sign_mask", 0x7FFFFFFFFFFFFFFF);
3802 StubRoutines::x86::_double_sign_flip = generate_fp_mask("double_sign_flip", 0x8000000000000000);
3803
3804 // support for verify_oop (must happen after universe_init)
3805 StubRoutines::_verify_oop_subroutine_entry = generate_verify_oop();
3806
3807 // arraycopy stubs used by compilers
3808 generate_arraycopy_stubs();
3809
3810 generate_math_stubs();
3811
3812 // don't bother generating these AES intrinsic stubs unless global flag is set
3813 if (UseAESIntrinsics) {
3814 StubRoutines::x86::_key_shuffle_mask_addr = generate_key_shuffle_mask(); // needed by the others
3815
3816 StubRoutines::_aescrypt_encryptBlock = generate_aescrypt_encryptBlock();
3817 StubRoutines::_aescrypt_decryptBlock = generate_aescrypt_decryptBlock();
3818 StubRoutines::_cipherBlockChaining_encryptAESCrypt = generate_cipherBlockChaining_encryptAESCrypt();
3819 StubRoutines::_cipherBlockChaining_decryptAESCrypt = generate_cipherBlockChaining_decryptAESCrypt_Parallel();
3820 }
3821 }
3822
3823 public:
3824 StubGenerator(CodeBuffer* code, bool all) : StubCodeGenerator(code) {
3825 if (all) {
3826 generate_all();
3827 } else {
3828 generate_initial();
3829 }
3830 }
3831 }; // end class declaration
3832
3833 void StubGenerator_generate(CodeBuffer* code, bool all) {
3834 StubGenerator g(code, all);
3835 }