1 /*
2 * Copyright (c) 2003, 2015, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "interpreter/interpreter.hpp"
29 #include "nativeInst_x86.hpp"
30 #include "oops/instanceOop.hpp"
31 #include "oops/method.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 "runtime/thread.inline.hpp"
41 #include "utilities/top.hpp"
42 #ifdef COMPILER2
43 #include "opto/runtime.hpp"
44 #endif
45
46 // Declaration and definition of StubGenerator (no .hpp file).
47 // For a more detailed description of the stub routine structure
48 // see the comment in stubRoutines.hpp
49
50 #define __ _masm->
51 #define TIMES_OOP (UseCompressedOops ? Address::times_4 : Address::times_8)
52 #define a__ ((Assembler*)_masm)->
53
54 #ifdef PRODUCT
55 #define BLOCK_COMMENT(str) /* nothing */
56 #else
57 #define BLOCK_COMMENT(str) __ block_comment(str)
58 #endif
59
60 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
61 const int MXCSR_MASK = 0xFFC0; // Mask out any pending exceptions
62
63 // Stub Code definitions
64
65 static address handle_unsafe_access() {
66 JavaThread* thread = JavaThread::current();
67 address pc = thread->saved_exception_pc();
68 // pc is the instruction which we must emulate
69 // doing a no-op is fine: return garbage from the load
70 // therefore, compute npc
71 address npc = Assembler::locate_next_instruction(pc);
72
73 // request an async exception
74 thread->set_pending_unsafe_access_error();
75
76 // return address of next instruction to execute
77 return npc;
78 }
79
80 class StubGenerator: public StubCodeGenerator {
81 private:
82
83 #ifdef PRODUCT
84 #define inc_counter_np(counter) ((void)0)
85 #else
86 void inc_counter_np_(int& counter) {
87 // This can destroy rscratch1 if counter is far from the code cache
88 __ incrementl(ExternalAddress((address)&counter));
89 }
90 #define inc_counter_np(counter) \
91 BLOCK_COMMENT("inc_counter " #counter); \
92 inc_counter_np_(counter);
93 #endif
94
95 // Call stubs are used to call Java from C
96 //
97 // Linux Arguments:
98 // c_rarg0: call wrapper address address
99 // c_rarg1: result address
100 // c_rarg2: result type BasicType
101 // c_rarg3: method Method*
102 // c_rarg4: (interpreter) entry point address
103 // c_rarg5: parameters intptr_t*
104 // 16(rbp): parameter size (in words) int
105 // 24(rbp): thread Thread*
106 //
107 // [ return_from_Java ] <--- rsp
108 // [ argument word n ]
109 // ...
110 // -12 [ argument word 1 ]
111 // -11 [ saved r15 ] <--- rsp_after_call
112 // -10 [ saved r14 ]
113 // -9 [ saved r13 ]
114 // -8 [ saved r12 ]
115 // -7 [ saved rbx ]
116 // -6 [ call wrapper ]
117 // -5 [ result ]
118 // -4 [ result type ]
119 // -3 [ method ]
120 // -2 [ entry point ]
121 // -1 [ parameters ]
122 // 0 [ saved rbp ] <--- rbp
123 // 1 [ return address ]
124 // 2 [ parameter size ]
125 // 3 [ thread ]
126 //
127 // Windows Arguments:
128 // c_rarg0: call wrapper address address
129 // c_rarg1: result address
130 // c_rarg2: result type BasicType
131 // c_rarg3: method Method*
132 // 48(rbp): (interpreter) entry point address
133 // 56(rbp): parameters intptr_t*
134 // 64(rbp): parameter size (in words) int
135 // 72(rbp): thread Thread*
136 //
137 // [ return_from_Java ] <--- rsp
138 // [ argument word n ]
139 // ...
140 // -60 [ argument word 1 ]
141 // -59 [ saved xmm31 ] <--- rsp after_call
142 // [ saved xmm16-xmm30 ] (EVEX enabled, else the space is blank)
143 // -27 [ saved xmm15 ]
144 // [ saved xmm7-xmm14 ]
145 // -9 [ saved xmm6 ] (each xmm register takes 2 slots)
146 // -7 [ saved r15 ]
147 // -6 [ saved r14 ]
148 // -5 [ saved r13 ]
149 // -4 [ saved r12 ]
150 // -3 [ saved rdi ]
151 // -2 [ saved rsi ]
152 // -1 [ saved rbx ]
153 // 0 [ saved rbp ] <--- rbp
154 // 1 [ return address ]
155 // 2 [ call wrapper ]
156 // 3 [ result ]
157 // 4 [ result type ]
158 // 5 [ method ]
159 // 6 [ entry point ]
160 // 7 [ parameters ]
161 // 8 [ parameter size ]
162 // 9 [ thread ]
163 //
164 // Windows reserves the callers stack space for arguments 1-4.
165 // We spill c_rarg0-c_rarg3 to this space.
166
167 // Call stub stack layout word offsets from rbp
168 enum call_stub_layout {
169 #ifdef _WIN64
170 xmm_save_first = 6, // save from xmm6
171 xmm_save_last = 31, // to xmm31
172 xmm_save_base = -9,
173 rsp_after_call_off = xmm_save_base - 2 * (xmm_save_last - xmm_save_first), // -27
174 r15_off = -7,
175 r14_off = -6,
176 r13_off = -5,
177 r12_off = -4,
178 rdi_off = -3,
179 rsi_off = -2,
180 rbx_off = -1,
181 rbp_off = 0,
182 retaddr_off = 1,
183 call_wrapper_off = 2,
184 result_off = 3,
185 result_type_off = 4,
186 method_off = 5,
187 entry_point_off = 6,
188 parameters_off = 7,
189 parameter_size_off = 8,
190 thread_off = 9
191 #else
192 rsp_after_call_off = -12,
193 mxcsr_off = rsp_after_call_off,
194 r15_off = -11,
195 r14_off = -10,
196 r13_off = -9,
197 r12_off = -8,
198 rbx_off = -7,
199 call_wrapper_off = -6,
200 result_off = -5,
201 result_type_off = -4,
202 method_off = -3,
203 entry_point_off = -2,
204 parameters_off = -1,
205 rbp_off = 0,
206 retaddr_off = 1,
207 parameter_size_off = 2,
208 thread_off = 3
209 #endif
210 };
211
212 #ifdef _WIN64
213 Address xmm_save(int reg) {
214 assert(reg >= xmm_save_first && reg <= xmm_save_last, "XMM register number out of range");
215 return Address(rbp, (xmm_save_base - (reg - xmm_save_first) * 2) * wordSize);
216 }
217 #endif
218
219 address generate_call_stub(address& return_address) {
220 assert((int)frame::entry_frame_after_call_words == -(int)rsp_after_call_off + 1 &&
221 (int)frame::entry_frame_call_wrapper_offset == (int)call_wrapper_off,
222 "adjust this code");
223 StubCodeMark mark(this, "StubRoutines", "call_stub");
224 address start = __ pc();
225
226 // same as in generate_catch_exception()!
227 const Address rsp_after_call(rbp, rsp_after_call_off * wordSize);
228
229 const Address call_wrapper (rbp, call_wrapper_off * wordSize);
230 const Address result (rbp, result_off * wordSize);
231 const Address result_type (rbp, result_type_off * wordSize);
232 const Address method (rbp, method_off * wordSize);
233 const Address entry_point (rbp, entry_point_off * wordSize);
234 const Address parameters (rbp, parameters_off * wordSize);
235 const Address parameter_size(rbp, parameter_size_off * wordSize);
236
237 // same as in generate_catch_exception()!
238 const Address thread (rbp, thread_off * wordSize);
239
240 const Address r15_save(rbp, r15_off * wordSize);
241 const Address r14_save(rbp, r14_off * wordSize);
242 const Address r13_save(rbp, r13_off * wordSize);
243 const Address r12_save(rbp, r12_off * wordSize);
244 const Address rbx_save(rbp, rbx_off * wordSize);
245
246 // stub code
247 __ enter();
248 __ subptr(rsp, -rsp_after_call_off * wordSize);
249
250 // save register parameters
251 #ifndef _WIN64
252 __ movptr(parameters, c_rarg5); // parameters
253 __ movptr(entry_point, c_rarg4); // entry_point
254 #endif
255
256 __ movptr(method, c_rarg3); // method
257 __ movl(result_type, c_rarg2); // result type
258 __ movptr(result, c_rarg1); // result
259 __ movptr(call_wrapper, c_rarg0); // call wrapper
260
261 // save regs belonging to calling function
262 __ movptr(rbx_save, rbx);
263 __ movptr(r12_save, r12);
264 __ movptr(r13_save, r13);
265 __ movptr(r14_save, r14);
266 __ movptr(r15_save, r15);
267 if (UseAVX > 2) {
268 __ movl(rbx, 0xffff);
269 __ kmovql(k1, rbx);
270 }
271 #ifdef _WIN64
272 if (UseAVX > 2) {
273 for (int i = 6; i <= 31; i++) {
274 __ movdqu(xmm_save(i), as_XMMRegister(i));
275 }
276 } else {
277 for (int i = 6; i <= 15; i++) {
278 __ movdqu(xmm_save(i), as_XMMRegister(i));
279 }
280 }
281
282 const Address rdi_save(rbp, rdi_off * wordSize);
283 const Address rsi_save(rbp, rsi_off * wordSize);
284
285 __ movptr(rsi_save, rsi);
286 __ movptr(rdi_save, rdi);
287 #else
288 const Address mxcsr_save(rbp, mxcsr_off * wordSize);
289 {
290 Label skip_ldmx;
291 __ stmxcsr(mxcsr_save);
292 __ movl(rax, mxcsr_save);
293 __ andl(rax, MXCSR_MASK); // Only check control and mask bits
294 ExternalAddress mxcsr_std(StubRoutines::addr_mxcsr_std());
295 __ cmp32(rax, mxcsr_std);
296 __ jcc(Assembler::equal, skip_ldmx);
297 __ ldmxcsr(mxcsr_std);
298 __ bind(skip_ldmx);
299 }
300 #endif
301
302 // Load up thread register
303 __ movptr(r15_thread, thread);
304 __ reinit_heapbase();
305
306 #ifdef ASSERT
307 // make sure we have no pending exceptions
308 {
309 Label L;
310 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
311 __ jcc(Assembler::equal, L);
312 __ stop("StubRoutines::call_stub: entered with pending exception");
313 __ bind(L);
314 }
315 #endif
316
317 // pass parameters if any
318 BLOCK_COMMENT("pass parameters if any");
319 Label parameters_done;
320 __ movl(c_rarg3, parameter_size);
321 __ testl(c_rarg3, c_rarg3);
322 __ jcc(Assembler::zero, parameters_done);
323
324 Label loop;
325 __ movptr(c_rarg2, parameters); // parameter pointer
326 __ movl(c_rarg1, c_rarg3); // parameter counter is in c_rarg1
327 __ BIND(loop);
328 __ movptr(rax, Address(c_rarg2, 0));// get parameter
329 __ addptr(c_rarg2, wordSize); // advance to next parameter
330 __ decrementl(c_rarg1); // decrement counter
331 __ push(rax); // pass parameter
332 __ jcc(Assembler::notZero, loop);
333
334 // call Java function
335 __ BIND(parameters_done);
336 __ movptr(rbx, method); // get Method*
337 __ movptr(c_rarg1, entry_point); // get entry_point
338 __ mov(r13, rsp); // set sender sp
339 BLOCK_COMMENT("call Java function");
340 __ call(c_rarg1);
341
342 BLOCK_COMMENT("call_stub_return_address:");
343 return_address = __ pc();
344
345 // store result depending on type (everything that is not
346 // T_OBJECT, T_LONG, T_FLOAT or T_DOUBLE is treated as T_INT)
347 __ movptr(c_rarg0, result);
348 Label is_long, is_float, is_double, exit;
349 __ movl(c_rarg1, result_type);
350 __ cmpl(c_rarg1, T_OBJECT);
351 __ jcc(Assembler::equal, is_long);
352 __ cmpl(c_rarg1, T_LONG);
353 __ jcc(Assembler::equal, is_long);
354 __ cmpl(c_rarg1, T_FLOAT);
355 __ jcc(Assembler::equal, is_float);
356 __ cmpl(c_rarg1, T_DOUBLE);
357 __ jcc(Assembler::equal, is_double);
358
359 // handle T_INT case
360 __ movl(Address(c_rarg0, 0), rax);
361
362 __ BIND(exit);
363
364 // pop parameters
365 __ lea(rsp, rsp_after_call);
366
367 #ifdef ASSERT
368 // verify that threads correspond
369 {
370 Label L1, L2, L3;
371 __ cmpptr(r15_thread, thread);
372 __ jcc(Assembler::equal, L1);
373 __ stop("StubRoutines::call_stub: r15_thread is corrupted");
374 __ bind(L1);
375 __ get_thread(rbx);
376 __ cmpptr(r15_thread, thread);
377 __ jcc(Assembler::equal, L2);
378 __ stop("StubRoutines::call_stub: r15_thread is modified by call");
379 __ bind(L2);
380 __ cmpptr(r15_thread, rbx);
381 __ jcc(Assembler::equal, L3);
382 __ stop("StubRoutines::call_stub: threads must correspond");
383 __ bind(L3);
384 }
385 #endif
386
387 // restore regs belonging to calling function
388 #ifdef _WIN64
389 int xmm_ub = 15;
390 if (UseAVX > 2) {
391 xmm_ub = 31;
392 }
393 // emit the restores for xmm regs
394 for (int i = 6; i <= xmm_ub; i++) {
395 __ movdqu(as_XMMRegister(i), xmm_save(i));
396 }
397 #endif
398 __ movptr(r15, r15_save);
399 __ movptr(r14, r14_save);
400 __ movptr(r13, r13_save);
401 __ movptr(r12, r12_save);
402 __ movptr(rbx, rbx_save);
403
404 #ifdef _WIN64
405 __ movptr(rdi, rdi_save);
406 __ movptr(rsi, rsi_save);
407 #else
408 __ ldmxcsr(mxcsr_save);
409 #endif
410
411 // restore rsp
412 __ addptr(rsp, -rsp_after_call_off * wordSize);
413
414 // return
415 __ pop(rbp);
416 __ ret(0);
417
418 // handle return types different from T_INT
419 __ BIND(is_long);
420 __ movq(Address(c_rarg0, 0), rax);
421 __ jmp(exit);
422
423 __ BIND(is_float);
424 __ movflt(Address(c_rarg0, 0), xmm0);
425 __ jmp(exit);
426
427 __ BIND(is_double);
428 __ movdbl(Address(c_rarg0, 0), xmm0);
429 __ jmp(exit);
430
431 return start;
432 }
433
434 // Return point for a Java call if there's an exception thrown in
435 // Java code. The exception is caught and transformed into a
436 // pending exception stored in JavaThread that can be tested from
437 // within the VM.
438 //
439 // Note: Usually the parameters are removed by the callee. In case
440 // of an exception crossing an activation frame boundary, that is
441 // not the case if the callee is compiled code => need to setup the
442 // rsp.
443 //
444 // rax: exception oop
445
446 address generate_catch_exception() {
447 StubCodeMark mark(this, "StubRoutines", "catch_exception");
448 address start = __ pc();
449
450 // same as in generate_call_stub():
451 const Address rsp_after_call(rbp, rsp_after_call_off * wordSize);
452 const Address thread (rbp, thread_off * wordSize);
453
454 #ifdef ASSERT
455 // verify that threads correspond
456 {
457 Label L1, L2, L3;
458 __ cmpptr(r15_thread, thread);
459 __ jcc(Assembler::equal, L1);
460 __ stop("StubRoutines::catch_exception: r15_thread is corrupted");
461 __ bind(L1);
462 __ get_thread(rbx);
463 __ cmpptr(r15_thread, thread);
464 __ jcc(Assembler::equal, L2);
465 __ stop("StubRoutines::catch_exception: r15_thread is modified by call");
466 __ bind(L2);
467 __ cmpptr(r15_thread, rbx);
468 __ jcc(Assembler::equal, L3);
469 __ stop("StubRoutines::catch_exception: threads must correspond");
470 __ bind(L3);
471 }
472 #endif
473
474 // set pending exception
475 __ verify_oop(rax);
476
477 __ movptr(Address(r15_thread, Thread::pending_exception_offset()), rax);
478 __ lea(rscratch1, ExternalAddress((address)__FILE__));
479 __ movptr(Address(r15_thread, Thread::exception_file_offset()), rscratch1);
480 __ movl(Address(r15_thread, Thread::exception_line_offset()), (int) __LINE__);
481
482 // complete return to VM
483 assert(StubRoutines::_call_stub_return_address != NULL,
484 "_call_stub_return_address must have been generated before");
485 __ jump(RuntimeAddress(StubRoutines::_call_stub_return_address));
486
487 return start;
488 }
489
490 // Continuation point for runtime calls returning with a pending
491 // exception. The pending exception check happened in the runtime
492 // or native call stub. The pending exception in Thread is
493 // converted into a Java-level exception.
494 //
495 // Contract with Java-level exception handlers:
496 // rax: exception
497 // rdx: throwing pc
498 //
499 // NOTE: At entry of this stub, exception-pc must be on stack !!
500
501 address generate_forward_exception() {
502 StubCodeMark mark(this, "StubRoutines", "forward exception");
503 address start = __ pc();
504
505 // Upon entry, the sp points to the return address returning into
506 // Java (interpreted or compiled) code; i.e., the return address
507 // becomes the throwing pc.
508 //
509 // Arguments pushed before the runtime call are still on the stack
510 // but the exception handler will reset the stack pointer ->
511 // ignore them. A potential result in registers can be ignored as
512 // well.
513
514 #ifdef ASSERT
515 // make sure this code is only executed if there is a pending exception
516 {
517 Label L;
518 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t) NULL);
519 __ jcc(Assembler::notEqual, L);
520 __ stop("StubRoutines::forward exception: no pending exception (1)");
521 __ bind(L);
522 }
523 #endif
524
525 // compute exception handler into rbx
526 __ movptr(c_rarg0, Address(rsp, 0));
527 BLOCK_COMMENT("call exception_handler_for_return_address");
528 __ call_VM_leaf(CAST_FROM_FN_PTR(address,
529 SharedRuntime::exception_handler_for_return_address),
530 r15_thread, c_rarg0);
531 __ mov(rbx, rax);
532
533 // setup rax & rdx, remove return address & clear pending exception
534 __ pop(rdx);
535 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset()));
536 __ movptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
537
538 #ifdef ASSERT
539 // make sure exception is set
540 {
541 Label L;
542 __ testptr(rax, rax);
543 __ jcc(Assembler::notEqual, L);
544 __ stop("StubRoutines::forward exception: no pending exception (2)");
545 __ bind(L);
546 }
547 #endif
548
549 // continue at exception handler (return address removed)
550 // rax: exception
551 // rbx: exception handler
552 // rdx: throwing pc
553 __ verify_oop(rax);
554 __ jmp(rbx);
555
556 return start;
557 }
558
559 // Support for jint atomic::xchg(jint exchange_value, volatile jint* dest)
560 //
561 // Arguments :
562 // c_rarg0: exchange_value
563 // c_rarg0: dest
564 //
565 // Result:
566 // *dest <- ex, return (orig *dest)
567 address generate_atomic_xchg() {
568 StubCodeMark mark(this, "StubRoutines", "atomic_xchg");
569 address start = __ pc();
570
571 __ movl(rax, c_rarg0); // Copy to eax we need a return value anyhow
572 __ xchgl(rax, Address(c_rarg1, 0)); // automatic LOCK
573 __ ret(0);
574
575 return start;
576 }
577
578 // Support for intptr_t atomic::xchg_ptr(intptr_t exchange_value, volatile intptr_t* dest)
579 //
580 // Arguments :
581 // c_rarg0: exchange_value
582 // c_rarg1: dest
583 //
584 // Result:
585 // *dest <- ex, return (orig *dest)
586 address generate_atomic_xchg_ptr() {
587 StubCodeMark mark(this, "StubRoutines", "atomic_xchg_ptr");
588 address start = __ pc();
589
590 __ movptr(rax, c_rarg0); // Copy to eax we need a return value anyhow
591 __ xchgptr(rax, Address(c_rarg1, 0)); // automatic LOCK
592 __ ret(0);
593
594 return start;
595 }
596
597 // Support for jint atomic::atomic_cmpxchg(jint exchange_value, volatile jint* dest,
598 // jint compare_value)
599 //
600 // Arguments :
601 // c_rarg0: exchange_value
602 // c_rarg1: dest
603 // c_rarg2: compare_value
604 //
605 // Result:
606 // if ( compare_value == *dest ) {
607 // *dest = exchange_value
608 // return compare_value;
609 // else
610 // return *dest;
611 address generate_atomic_cmpxchg() {
612 StubCodeMark mark(this, "StubRoutines", "atomic_cmpxchg");
613 address start = __ pc();
614
615 __ movl(rax, c_rarg2);
616 if ( os::is_MP() ) __ lock();
617 __ cmpxchgl(c_rarg0, Address(c_rarg1, 0));
618 __ ret(0);
619
620 return start;
621 }
622
623 // Support for jbyte atomic::atomic_cmpxchg(jbyte exchange_value, volatile jbyte* dest,
624 // jbyte compare_value)
625 //
626 // Arguments :
627 // c_rarg0: exchange_value
628 // c_rarg1: dest
629 // c_rarg2: compare_value
630 //
631 // Result:
632 // if ( compare_value == *dest ) {
633 // *dest = exchange_value
634 // return compare_value;
635 // else
636 // return *dest;
637 address generate_atomic_cmpxchg_byte() {
638 StubCodeMark mark(this, "StubRoutines", "atomic_cmpxchg_byte");
639 address start = __ pc();
640
641 __ movsbq(rax, c_rarg2);
642 if ( os::is_MP() ) __ lock();
643 __ cmpxchgb(c_rarg0, Address(c_rarg1, 0));
644 __ ret(0);
645
646 return start;
647 }
648
649 // Support for jlong atomic::atomic_cmpxchg(jlong exchange_value,
650 // volatile jlong* dest,
651 // jlong compare_value)
652 // Arguments :
653 // c_rarg0: exchange_value
654 // c_rarg1: dest
655 // c_rarg2: compare_value
656 //
657 // Result:
658 // if ( compare_value == *dest ) {
659 // *dest = exchange_value
660 // return compare_value;
661 // else
662 // return *dest;
663 address generate_atomic_cmpxchg_long() {
664 StubCodeMark mark(this, "StubRoutines", "atomic_cmpxchg_long");
665 address start = __ pc();
666
667 __ movq(rax, c_rarg2);
668 if ( os::is_MP() ) __ lock();
669 __ cmpxchgq(c_rarg0, Address(c_rarg1, 0));
670 __ ret(0);
671
672 return start;
673 }
674
675 // Support for jint atomic::add(jint add_value, volatile jint* dest)
676 //
677 // Arguments :
678 // c_rarg0: add_value
679 // c_rarg1: dest
680 //
681 // Result:
682 // *dest += add_value
683 // return *dest;
684 address generate_atomic_add() {
685 StubCodeMark mark(this, "StubRoutines", "atomic_add");
686 address start = __ pc();
687
688 __ movl(rax, c_rarg0);
689 if ( os::is_MP() ) __ lock();
690 __ xaddl(Address(c_rarg1, 0), c_rarg0);
691 __ addl(rax, c_rarg0);
692 __ ret(0);
693
694 return start;
695 }
696
697 // Support for intptr_t atomic::add_ptr(intptr_t add_value, volatile intptr_t* dest)
698 //
699 // Arguments :
700 // c_rarg0: add_value
701 // c_rarg1: dest
702 //
703 // Result:
704 // *dest += add_value
705 // return *dest;
706 address generate_atomic_add_ptr() {
707 StubCodeMark mark(this, "StubRoutines", "atomic_add_ptr");
708 address start = __ pc();
709
710 __ movptr(rax, c_rarg0); // Copy to eax we need a return value anyhow
711 if ( os::is_MP() ) __ lock();
712 __ xaddptr(Address(c_rarg1, 0), c_rarg0);
713 __ addptr(rax, c_rarg0);
714 __ ret(0);
715
716 return start;
717 }
718
719 // Support for intptr_t OrderAccess::fence()
720 //
721 // Arguments :
722 //
723 // Result:
724 address generate_orderaccess_fence() {
725 StubCodeMark mark(this, "StubRoutines", "orderaccess_fence");
726 address start = __ pc();
727 __ membar(Assembler::StoreLoad);
728 __ ret(0);
729
730 return start;
731 }
732
733 // Support for intptr_t get_previous_fp()
734 //
735 // This routine is used to find the previous frame pointer for the
736 // caller (current_frame_guess). This is used as part of debugging
737 // ps() is seemingly lost trying to find frames.
738 // This code assumes that caller current_frame_guess) has a frame.
739 address generate_get_previous_fp() {
740 StubCodeMark mark(this, "StubRoutines", "get_previous_fp");
741 const Address old_fp(rbp, 0);
742 const Address older_fp(rax, 0);
743 address start = __ pc();
744
745 __ enter();
746 __ movptr(rax, old_fp); // callers fp
747 __ movptr(rax, older_fp); // the frame for ps()
748 __ pop(rbp);
749 __ ret(0);
750
751 return start;
752 }
753
754 // Support for intptr_t get_previous_sp()
755 //
756 // This routine is used to find the previous stack pointer for the
757 // caller.
758 address generate_get_previous_sp() {
759 StubCodeMark mark(this, "StubRoutines", "get_previous_sp");
760 address start = __ pc();
761
762 __ movptr(rax, rsp);
763 __ addptr(rax, 8); // return address is at the top of the stack.
764 __ ret(0);
765
766 return start;
767 }
768
769 //----------------------------------------------------------------------------------------------------
770 // Support for void verify_mxcsr()
771 //
772 // This routine is used with -Xcheck:jni to verify that native
773 // JNI code does not return to Java code without restoring the
774 // MXCSR register to our expected state.
775
776 address generate_verify_mxcsr() {
777 StubCodeMark mark(this, "StubRoutines", "verify_mxcsr");
778 address start = __ pc();
779
780 const Address mxcsr_save(rsp, 0);
781
782 if (CheckJNICalls) {
783 Label ok_ret;
784 ExternalAddress mxcsr_std(StubRoutines::addr_mxcsr_std());
785 __ push(rax);
786 __ subptr(rsp, wordSize); // allocate a temp location
787 __ stmxcsr(mxcsr_save);
788 __ movl(rax, mxcsr_save);
789 __ andl(rax, MXCSR_MASK); // Only check control and mask bits
790 __ cmp32(rax, mxcsr_std);
791 __ jcc(Assembler::equal, ok_ret);
792
793 __ warn("MXCSR changed by native JNI code, use -XX:+RestoreMXCSROnJNICall");
794
795 __ ldmxcsr(mxcsr_std);
796
797 __ bind(ok_ret);
798 __ addptr(rsp, wordSize);
799 __ pop(rax);
800 }
801
802 __ ret(0);
803
804 return start;
805 }
806
807 address generate_shenandoah_wb() {
808 StubCodeMark mark(this, "StubRoutines", "shenandoah_wb");
809 address start = __ pc();
810
811 Label done;
812
813 __ push(rbx);
814 // Check for object beeing in the collection set.
815 // TODO: Can we use only 1 register here?
816 __ movptr(rdi, rax);
817 __ shrptr(rdi, ShenandoahHeapRegion::RegionSizeShift);
818 __ movptr(rbx, (intptr_t) ShenandoahHeap::in_cset_fast_test_addr());
819 __ movbool(rbx, Address(rbx, rdi, Address::times_1));
820 __ testbool(rbx);
821 __ jcc(Assembler::zero, done);
822
823 __ push(rcx);
824 __ push(rdx);
825 __ push(rdi);
826 __ push(rsi);
827 __ push(r8);
828 __ push(r9);
829 __ push(r10);
830 __ push(r11);
831 __ push(r12);
832 __ push(r13);
833 __ push(r14);
834 __ push(r15);
835 __ subptr(rsp, 128);
836 __ movdbl(Address(rsp, 0), xmm0);
837 __ movdbl(Address(rsp, 8), xmm1);
838 __ movdbl(Address(rsp, 16), xmm2);
839 __ movdbl(Address(rsp, 24), xmm3);
840 __ movdbl(Address(rsp, 32), xmm4);
841 __ movdbl(Address(rsp, 40), xmm5);
842 __ movdbl(Address(rsp, 48), xmm6);
843 __ movdbl(Address(rsp, 56), xmm7);
844 __ movdbl(Address(rsp, 64), xmm8);
845 __ movdbl(Address(rsp, 72), xmm9);
846 __ movdbl(Address(rsp, 80), xmm10);
847 __ movdbl(Address(rsp, 88), xmm11);
848 __ movdbl(Address(rsp, 96), xmm12);
849 __ movdbl(Address(rsp, 104), xmm13);
850 __ movdbl(Address(rsp, 112), xmm14);
851 __ movdbl(Address(rsp, 120), xmm15);
852 __ movptr(rdi, rax);
853 __ call_VM_leaf(CAST_FROM_FN_PTR(address, ShenandoahBarrierSet::resolve_and_maybe_copy_oop_c2), rdi);
854 __ movdbl(xmm0, Address(rsp, 0));
855 __ movdbl(xmm1, Address(rsp, 8));
856 __ movdbl(xmm2, Address(rsp, 16));
857 __ movdbl(xmm3, Address(rsp, 24));
858 __ movdbl(xmm4, Address(rsp, 32));
859 __ movdbl(xmm5, Address(rsp, 40));
860 __ movdbl(xmm6, Address(rsp, 48));
861 __ movdbl(xmm7, Address(rsp, 56));
862 __ movdbl(xmm8, Address(rsp, 64));
863 __ movdbl(xmm9, Address(rsp, 72));
864 __ movdbl(xmm10, Address(rsp, 80));
865 __ movdbl(xmm11, Address(rsp, 88));
866 __ movdbl(xmm12, Address(rsp, 96));
867 __ movdbl(xmm13, Address(rsp, 104));
868 __ movdbl(xmm14, Address(rsp, 112));
869 __ movdbl(xmm15, Address(rsp, 120));
870 __ addptr(rsp, 128);
871 __ pop(r15);
872 __ pop(r14);
873 __ pop(r13);
874 __ pop(r12);
875 __ pop(r11);
876 __ pop(r10);
877 __ pop(r9);
878 __ pop(r8);
879 __ pop(rsi);
880 __ pop(rdi);
881 __ pop(rdx);
882 __ pop(rcx);
883
884 __ bind(done);
885
886 __ pop(rbx);
887
888 __ ret(0);
889
890 return start;
891 }
892
893 address generate_f2i_fixup() {
894 StubCodeMark mark(this, "StubRoutines", "f2i_fixup");
895 Address inout(rsp, 5 * wordSize); // return address + 4 saves
896
897 address start = __ pc();
898
899 Label L;
900
901 __ push(rax);
902 __ push(c_rarg3);
903 __ push(c_rarg2);
904 __ push(c_rarg1);
905
906 __ movl(rax, 0x7f800000);
907 __ xorl(c_rarg3, c_rarg3);
908 __ movl(c_rarg2, inout);
909 __ movl(c_rarg1, c_rarg2);
910 __ andl(c_rarg1, 0x7fffffff);
911 __ cmpl(rax, c_rarg1); // NaN? -> 0
912 __ jcc(Assembler::negative, L);
913 __ testl(c_rarg2, c_rarg2); // signed ? min_jint : max_jint
914 __ movl(c_rarg3, 0x80000000);
915 __ movl(rax, 0x7fffffff);
916 __ cmovl(Assembler::positive, c_rarg3, rax);
917
918 __ bind(L);
919 __ movptr(inout, c_rarg3);
920
921 __ pop(c_rarg1);
922 __ pop(c_rarg2);
923 __ pop(c_rarg3);
924 __ pop(rax);
925
926 __ ret(0);
927
928 return start;
929 }
930
931 address generate_f2l_fixup() {
932 StubCodeMark mark(this, "StubRoutines", "f2l_fixup");
933 Address inout(rsp, 5 * wordSize); // return address + 4 saves
934 address start = __ pc();
935
936 Label L;
937
938 __ push(rax);
939 __ push(c_rarg3);
940 __ push(c_rarg2);
941 __ push(c_rarg1);
942
943 __ movl(rax, 0x7f800000);
944 __ xorl(c_rarg3, c_rarg3);
945 __ movl(c_rarg2, inout);
946 __ movl(c_rarg1, c_rarg2);
947 __ andl(c_rarg1, 0x7fffffff);
948 __ cmpl(rax, c_rarg1); // NaN? -> 0
949 __ jcc(Assembler::negative, L);
950 __ testl(c_rarg2, c_rarg2); // signed ? min_jlong : max_jlong
951 __ mov64(c_rarg3, 0x8000000000000000);
952 __ mov64(rax, 0x7fffffffffffffff);
953 __ cmov(Assembler::positive, c_rarg3, rax);
954
955 __ bind(L);
956 __ movptr(inout, c_rarg3);
957
958 __ pop(c_rarg1);
959 __ pop(c_rarg2);
960 __ pop(c_rarg3);
961 __ pop(rax);
962
963 __ ret(0);
964
965 return start;
966 }
967
968 address generate_d2i_fixup() {
969 StubCodeMark mark(this, "StubRoutines", "d2i_fixup");
970 Address inout(rsp, 6 * wordSize); // return address + 5 saves
971
972 address start = __ pc();
973
974 Label L;
975
976 __ push(rax);
977 __ push(c_rarg3);
978 __ push(c_rarg2);
979 __ push(c_rarg1);
980 __ push(c_rarg0);
981
982 __ movl(rax, 0x7ff00000);
983 __ movq(c_rarg2, inout);
984 __ movl(c_rarg3, c_rarg2);
985 __ mov(c_rarg1, c_rarg2);
986 __ mov(c_rarg0, c_rarg2);
987 __ negl(c_rarg3);
988 __ shrptr(c_rarg1, 0x20);
989 __ orl(c_rarg3, c_rarg2);
990 __ andl(c_rarg1, 0x7fffffff);
991 __ xorl(c_rarg2, c_rarg2);
992 __ shrl(c_rarg3, 0x1f);
993 __ orl(c_rarg1, c_rarg3);
994 __ cmpl(rax, c_rarg1);
995 __ jcc(Assembler::negative, L); // NaN -> 0
996 __ testptr(c_rarg0, c_rarg0); // signed ? min_jint : max_jint
997 __ movl(c_rarg2, 0x80000000);
998 __ movl(rax, 0x7fffffff);
999 __ cmov(Assembler::positive, c_rarg2, rax);
1000
1001 __ bind(L);
1002 __ movptr(inout, c_rarg2);
1003
1004 __ pop(c_rarg0);
1005 __ pop(c_rarg1);
1006 __ pop(c_rarg2);
1007 __ pop(c_rarg3);
1008 __ pop(rax);
1009
1010 __ ret(0);
1011
1012 return start;
1013 }
1014
1015 address generate_d2l_fixup() {
1016 StubCodeMark mark(this, "StubRoutines", "d2l_fixup");
1017 Address inout(rsp, 6 * wordSize); // return address + 5 saves
1018
1019 address start = __ pc();
1020
1021 Label L;
1022
1023 __ push(rax);
1024 __ push(c_rarg3);
1025 __ push(c_rarg2);
1026 __ push(c_rarg1);
1027 __ push(c_rarg0);
1028
1029 __ movl(rax, 0x7ff00000);
1030 __ movq(c_rarg2, inout);
1031 __ movl(c_rarg3, c_rarg2);
1032 __ mov(c_rarg1, c_rarg2);
1033 __ mov(c_rarg0, c_rarg2);
1034 __ negl(c_rarg3);
1035 __ shrptr(c_rarg1, 0x20);
1036 __ orl(c_rarg3, c_rarg2);
1037 __ andl(c_rarg1, 0x7fffffff);
1038 __ xorl(c_rarg2, c_rarg2);
1039 __ shrl(c_rarg3, 0x1f);
1040 __ orl(c_rarg1, c_rarg3);
1041 __ cmpl(rax, c_rarg1);
1042 __ jcc(Assembler::negative, L); // NaN -> 0
1043 __ testq(c_rarg0, c_rarg0); // signed ? min_jlong : max_jlong
1044 __ mov64(c_rarg2, 0x8000000000000000);
1045 __ mov64(rax, 0x7fffffffffffffff);
1046 __ cmovq(Assembler::positive, c_rarg2, rax);
1047
1048 __ bind(L);
1049 __ movq(inout, c_rarg2);
1050
1051 __ pop(c_rarg0);
1052 __ pop(c_rarg1);
1053 __ pop(c_rarg2);
1054 __ pop(c_rarg3);
1055 __ pop(rax);
1056
1057 __ ret(0);
1058
1059 return start;
1060 }
1061
1062 address generate_fp_mask(const char *stub_name, int64_t mask) {
1063 __ align(CodeEntryAlignment);
1064 StubCodeMark mark(this, "StubRoutines", stub_name);
1065 address start = __ pc();
1066
1067 __ emit_data64( mask, relocInfo::none );
1068 __ emit_data64( mask, relocInfo::none );
1069
1070 return start;
1071 }
1072
1073 // The following routine generates a subroutine to throw an
1074 // asynchronous UnknownError when an unsafe access gets a fault that
1075 // could not be reasonably prevented by the programmer. (Example:
1076 // SIGBUS/OBJERR.)
1077 address generate_handler_for_unsafe_access() {
1078 StubCodeMark mark(this, "StubRoutines", "handler_for_unsafe_access");
1079 address start = __ pc();
1080
1081 __ push(0); // hole for return address-to-be
1082 __ pusha(); // push registers
1083 Address next_pc(rsp, RegisterImpl::number_of_registers * BytesPerWord);
1084
1085 // FIXME: this probably needs alignment logic
1086
1087 __ subptr(rsp, frame::arg_reg_save_area_bytes);
1088 BLOCK_COMMENT("call handle_unsafe_access");
1089 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, handle_unsafe_access)));
1090 __ addptr(rsp, frame::arg_reg_save_area_bytes);
1091
1092 __ movptr(next_pc, rax); // stuff next address
1093 __ popa();
1094 __ ret(0); // jump to next address
1095
1096 return start;
1097 }
1098
1099 // Non-destructive plausibility checks for oops
1100 //
1101 // Arguments:
1102 // all args on stack!
1103 //
1104 // Stack after saving c_rarg3:
1105 // [tos + 0]: saved c_rarg3
1106 // [tos + 1]: saved c_rarg2
1107 // [tos + 2]: saved r12 (several TemplateTable methods use it)
1108 // [tos + 3]: saved flags
1109 // [tos + 4]: return address
1110 // * [tos + 5]: error message (char*)
1111 // * [tos + 6]: object to verify (oop)
1112 // * [tos + 7]: saved rax - saved by caller and bashed
1113 // * [tos + 8]: saved r10 (rscratch1) - saved by caller
1114 // * = popped on exit
1115 address generate_verify_oop() {
1116 StubCodeMark mark(this, "StubRoutines", "verify_oop");
1117 address start = __ pc();
1118
1119 Label exit, error;
1120
1121 __ pushf();
1122 __ incrementl(ExternalAddress((address) StubRoutines::verify_oop_count_addr()));
1123
1124 __ push(r12);
1125
1126 // save c_rarg2 and c_rarg3
1127 __ push(c_rarg2);
1128 __ push(c_rarg3);
1129
1130 enum {
1131 // After previous pushes.
1132 oop_to_verify = 6 * wordSize,
1133 saved_rax = 7 * wordSize,
1134 saved_r10 = 8 * wordSize,
1135
1136 // Before the call to MacroAssembler::debug(), see below.
1137 return_addr = 16 * wordSize,
1138 error_msg = 17 * wordSize
1139 };
1140
1141 // get object
1142 __ movptr(rax, Address(rsp, oop_to_verify));
1143
1144 // make sure object is 'reasonable'
1145 __ testptr(rax, rax);
1146 __ jcc(Assembler::zero, exit); // if obj is NULL it is OK
1147 // Check if the oop is in the right area of memory
1148 __ movptr(c_rarg2, rax);
1149 __ movptr(c_rarg3, (intptr_t) Universe::verify_oop_mask());
1150 __ andptr(c_rarg2, c_rarg3);
1151 __ movptr(c_rarg3, (intptr_t) Universe::verify_oop_bits());
1152 __ cmpptr(c_rarg2, c_rarg3);
1153 __ jcc(Assembler::notZero, error);
1154
1155 // set r12 to heapbase for load_klass()
1156 __ reinit_heapbase();
1157
1158 // make sure klass is 'reasonable', which is not zero.
1159 __ load_klass(rax, rax); // get klass
1160 __ testptr(rax, rax);
1161 __ jcc(Assembler::zero, error); // if klass is NULL it is broken
1162
1163 // return if everything seems ok
1164 __ bind(exit);
1165 __ movptr(rax, Address(rsp, saved_rax)); // get saved rax back
1166 __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back
1167 __ pop(c_rarg3); // restore c_rarg3
1168 __ pop(c_rarg2); // restore c_rarg2
1169 __ pop(r12); // restore r12
1170 __ popf(); // restore flags
1171 __ ret(4 * wordSize); // pop caller saved stuff
1172
1173 // handle errors
1174 __ bind(error);
1175 __ movptr(rax, Address(rsp, saved_rax)); // get saved rax back
1176 __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back
1177 __ pop(c_rarg3); // get saved c_rarg3 back
1178 __ pop(c_rarg2); // get saved c_rarg2 back
1179 __ pop(r12); // get saved r12 back
1180 __ popf(); // get saved flags off stack --
1181 // will be ignored
1182
1183 __ pusha(); // push registers
1184 // (rip is already
1185 // already pushed)
1186 // debug(char* msg, int64_t pc, int64_t regs[])
1187 // We've popped the registers we'd saved (c_rarg3, c_rarg2 and flags), and
1188 // pushed all the registers, so now the stack looks like:
1189 // [tos + 0] 16 saved registers
1190 // [tos + 16] return address
1191 // * [tos + 17] error message (char*)
1192 // * [tos + 18] object to verify (oop)
1193 // * [tos + 19] saved rax - saved by caller and bashed
1194 // * [tos + 20] saved r10 (rscratch1) - saved by caller
1195 // * = popped on exit
1196
1197 __ movptr(c_rarg0, Address(rsp, error_msg)); // pass address of error message
1198 __ movptr(c_rarg1, Address(rsp, return_addr)); // pass return address
1199 __ movq(c_rarg2, rsp); // pass address of regs on stack
1200 __ mov(r12, rsp); // remember rsp
1201 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows
1202 __ andptr(rsp, -16); // align stack as required by ABI
1203 BLOCK_COMMENT("call MacroAssembler::debug");
1204 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64)));
1205 __ mov(rsp, r12); // restore rsp
1206 __ popa(); // pop registers (includes r12)
1207 __ ret(4 * wordSize); // pop caller saved stuff
1208
1209 return start;
1210 }
1211
1212 //
1213 // Verify that a register contains clean 32-bits positive value
1214 // (high 32-bits are 0) so it could be used in 64-bits shifts.
1215 //
1216 // Input:
1217 // Rint - 32-bits value
1218 // Rtmp - scratch
1219 //
1220 void assert_clean_int(Register Rint, Register Rtmp) {
1221 #ifdef ASSERT
1222 Label L;
1223 assert_different_registers(Rtmp, Rint);
1224 __ movslq(Rtmp, Rint);
1225 __ cmpq(Rtmp, Rint);
1226 __ jcc(Assembler::equal, L);
1227 __ stop("high 32-bits of int value are not 0");
1228 __ bind(L);
1229 #endif
1230 }
1231
1232 // Generate overlap test for array copy stubs
1233 //
1234 // Input:
1235 // c_rarg0 - from
1236 // c_rarg1 - to
1237 // c_rarg2 - element count
1238 //
1239 // Output:
1240 // rax - &from[element count - 1]
1241 //
1242 void array_overlap_test(address no_overlap_target, Address::ScaleFactor sf) {
1243 assert(no_overlap_target != NULL, "must be generated");
1244 array_overlap_test(no_overlap_target, NULL, sf);
1245 }
1246 void array_overlap_test(Label& L_no_overlap, Address::ScaleFactor sf) {
1247 array_overlap_test(NULL, &L_no_overlap, sf);
1248 }
1249 void array_overlap_test(address no_overlap_target, Label* NOLp, Address::ScaleFactor sf) {
1250 const Register from = c_rarg0;
1251 const Register to = c_rarg1;
1252 const Register count = c_rarg2;
1253 const Register end_from = rax;
1254
1255 __ cmpptr(to, from);
1256 __ lea(end_from, Address(from, count, sf, 0));
1257 if (NOLp == NULL) {
1258 ExternalAddress no_overlap(no_overlap_target);
1259 __ jump_cc(Assembler::belowEqual, no_overlap);
1260 __ cmpptr(to, end_from);
1261 __ jump_cc(Assembler::aboveEqual, no_overlap);
1262 } else {
1263 __ jcc(Assembler::belowEqual, (*NOLp));
1264 __ cmpptr(to, end_from);
1265 __ jcc(Assembler::aboveEqual, (*NOLp));
1266 }
1267 }
1268
1269 // Shuffle first three arg regs on Windows into Linux/Solaris locations.
1270 //
1271 // Outputs:
1272 // rdi - rcx
1273 // rsi - rdx
1274 // rdx - r8
1275 // rcx - r9
1276 //
1277 // Registers r9 and r10 are used to save rdi and rsi on Windows, which latter
1278 // are non-volatile. r9 and r10 should not be used by the caller.
1279 //
1280 void setup_arg_regs(int nargs = 3) {
1281 const Register saved_rdi = r9;
1282 const Register saved_rsi = r10;
1283 assert(nargs == 3 || nargs == 4, "else fix");
1284 #ifdef _WIN64
1285 assert(c_rarg0 == rcx && c_rarg1 == rdx && c_rarg2 == r8 && c_rarg3 == r9,
1286 "unexpected argument registers");
1287 if (nargs >= 4)
1288 __ mov(rax, r9); // r9 is also saved_rdi
1289 __ movptr(saved_rdi, rdi);
1290 __ movptr(saved_rsi, rsi);
1291 __ mov(rdi, rcx); // c_rarg0
1292 __ mov(rsi, rdx); // c_rarg1
1293 __ mov(rdx, r8); // c_rarg2
1294 if (nargs >= 4)
1295 __ mov(rcx, rax); // c_rarg3 (via rax)
1296 #else
1297 assert(c_rarg0 == rdi && c_rarg1 == rsi && c_rarg2 == rdx && c_rarg3 == rcx,
1298 "unexpected argument registers");
1299 #endif
1300 }
1301
1302 void restore_arg_regs() {
1303 const Register saved_rdi = r9;
1304 const Register saved_rsi = r10;
1305 #ifdef _WIN64
1306 __ movptr(rdi, saved_rdi);
1307 __ movptr(rsi, saved_rsi);
1308 #endif
1309 }
1310
1311 // Generate code for an array write pre barrier
1312 //
1313 // addr - starting address
1314 // count - element count
1315 // tmp - scratch register
1316 //
1317 // Destroy no registers!
1318 //
1319 void gen_write_ref_array_pre_barrier(Register addr, Register count, bool dest_uninitialized) {
1320 BarrierSet* bs = Universe::heap()->barrier_set();
1321 switch (bs->kind()) {
1322 case BarrierSet::G1SATBCTLogging:
1323 case BarrierSet::ShenandoahBarrierSet:
1324 // With G1, don't generate the call if we statically know that the target in uninitialized
1325 if (!dest_uninitialized) {
1326 __ pusha(); // push registers
1327 if (count == c_rarg0) {
1328 if (addr == c_rarg1) {
1329 // exactly backwards!!
1330 __ xchgptr(c_rarg1, c_rarg0);
1331 } else {
1332 __ movptr(c_rarg1, count);
1333 __ movptr(c_rarg0, addr);
1334 }
1335 } else {
1336 __ movptr(c_rarg0, addr);
1337 __ movptr(c_rarg1, count);
1338 }
1339 __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_pre), 2);
1340 __ popa();
1341 }
1342 break;
1343 case BarrierSet::CardTableForRS:
1344 case BarrierSet::CardTableExtension:
1345 case BarrierSet::ModRef:
1346 break;
1347 default:
1348 ShouldNotReachHere();
1349
1350 }
1351 }
1352
1353 //
1354 // Generate code for an array write post barrier
1355 //
1356 // Input:
1357 // start - register containing starting address of destination array
1358 // count - elements count
1359 // scratch - scratch register
1360 //
1361 // The input registers are overwritten.
1362 //
1363 void gen_write_ref_array_post_barrier(Register start, Register count, Register scratch) {
1364 assert_different_registers(start, count, scratch);
1365 BarrierSet* bs = Universe::heap()->barrier_set();
1366 switch (bs->kind()) {
1367 case BarrierSet::G1SATBCTLogging:
1368 case BarrierSet::ShenandoahBarrierSet:
1369 {
1370 __ pusha(); // push registers (overkill)
1371 if (c_rarg0 == count) { // On win64 c_rarg0 == rcx
1372 assert_different_registers(c_rarg1, start);
1373 __ mov(c_rarg1, count);
1374 __ mov(c_rarg0, start);
1375 } else {
1376 assert_different_registers(c_rarg0, count);
1377 __ mov(c_rarg0, start);
1378 __ mov(c_rarg1, count);
1379 }
1380 __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_post), 2);
1381 __ popa();
1382 }
1383 break;
1384 case BarrierSet::CardTableForRS:
1385 case BarrierSet::CardTableExtension:
1386 {
1387 CardTableModRefBS* ct = barrier_set_cast<CardTableModRefBS>(bs);
1388 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
1389
1390 Label L_loop;
1391 const Register end = count;
1392
1393 __ leaq(end, Address(start, count, TIMES_OOP, 0)); // end == start+count*oop_size
1394 __ subptr(end, BytesPerHeapOop); // end - 1 to make inclusive
1395 __ shrptr(start, CardTableModRefBS::card_shift);
1396 __ shrptr(end, CardTableModRefBS::card_shift);
1397 __ subptr(end, start); // end --> cards count
1398
1399 int64_t disp = (int64_t) ct->byte_map_base;
1400 __ mov64(scratch, disp);
1401 __ addptr(start, scratch);
1402 __ BIND(L_loop);
1403 __ movb(Address(start, count, Address::times_1), 0);
1404 __ decrement(count);
1405 __ jcc(Assembler::greaterEqual, L_loop);
1406 }
1407 break;
1408 default:
1409 ShouldNotReachHere();
1410
1411 }
1412 }
1413
1414
1415 // Copy big chunks forward
1416 //
1417 // Inputs:
1418 // end_from - source arrays end address
1419 // end_to - destination array end address
1420 // qword_count - 64-bits element count, negative
1421 // to - scratch
1422 // L_copy_bytes - entry label
1423 // L_copy_8_bytes - exit label
1424 //
1425 void copy_bytes_forward(Register end_from, Register end_to,
1426 Register qword_count, Register to,
1427 Label& L_copy_bytes, Label& L_copy_8_bytes) {
1428 DEBUG_ONLY(__ stop("enter at entry label, not here"));
1429 Label L_loop;
1430 __ align(OptoLoopAlignment);
1431 if (UseUnalignedLoadStores) {
1432 Label L_end;
1433 // Copy 64-bytes per iteration
1434 __ BIND(L_loop);
1435 if (UseAVX > 2) {
1436 __ evmovdqu(xmm0, Address(end_from, qword_count, Address::times_8, -56), Assembler::AVX_512bit);
1437 __ evmovdqu(Address(end_to, qword_count, Address::times_8, -56), xmm0, Assembler::AVX_512bit);
1438 } else if (UseAVX == 2) {
1439 __ vmovdqu(xmm0, Address(end_from, qword_count, Address::times_8, -56));
1440 __ vmovdqu(Address(end_to, qword_count, Address::times_8, -56), xmm0);
1441 __ vmovdqu(xmm1, Address(end_from, qword_count, Address::times_8, -24));
1442 __ vmovdqu(Address(end_to, qword_count, Address::times_8, -24), xmm1);
1443 } else {
1444 __ movdqu(xmm0, Address(end_from, qword_count, Address::times_8, -56));
1445 __ movdqu(Address(end_to, qword_count, Address::times_8, -56), xmm0);
1446 __ movdqu(xmm1, Address(end_from, qword_count, Address::times_8, -40));
1447 __ movdqu(Address(end_to, qword_count, Address::times_8, -40), xmm1);
1448 __ movdqu(xmm2, Address(end_from, qword_count, Address::times_8, -24));
1449 __ movdqu(Address(end_to, qword_count, Address::times_8, -24), xmm2);
1450 __ movdqu(xmm3, Address(end_from, qword_count, Address::times_8, - 8));
1451 __ movdqu(Address(end_to, qword_count, Address::times_8, - 8), xmm3);
1452 }
1453 __ BIND(L_copy_bytes);
1454 __ addptr(qword_count, 8);
1455 __ jcc(Assembler::lessEqual, L_loop);
1456 __ subptr(qword_count, 4); // sub(8) and add(4)
1457 __ jccb(Assembler::greater, L_end);
1458 // Copy trailing 32 bytes
1459 if (UseAVX >= 2) {
1460 __ vmovdqu(xmm0, Address(end_from, qword_count, Address::times_8, -24));
1461 __ vmovdqu(Address(end_to, qword_count, Address::times_8, -24), xmm0);
1462 } else {
1463 __ movdqu(xmm0, Address(end_from, qword_count, Address::times_8, -24));
1464 __ movdqu(Address(end_to, qword_count, Address::times_8, -24), xmm0);
1465 __ movdqu(xmm1, Address(end_from, qword_count, Address::times_8, - 8));
1466 __ movdqu(Address(end_to, qword_count, Address::times_8, - 8), xmm1);
1467 }
1468 __ addptr(qword_count, 4);
1469 __ BIND(L_end);
1470 if (UseAVX >= 2) {
1471 // clean upper bits of YMM registers
1472 __ vpxor(xmm0, xmm0);
1473 __ vpxor(xmm1, xmm1);
1474 }
1475 } else {
1476 // Copy 32-bytes per iteration
1477 __ BIND(L_loop);
1478 __ movq(to, Address(end_from, qword_count, Address::times_8, -24));
1479 __ movq(Address(end_to, qword_count, Address::times_8, -24), to);
1480 __ movq(to, Address(end_from, qword_count, Address::times_8, -16));
1481 __ movq(Address(end_to, qword_count, Address::times_8, -16), to);
1482 __ movq(to, Address(end_from, qword_count, Address::times_8, - 8));
1483 __ movq(Address(end_to, qword_count, Address::times_8, - 8), to);
1484 __ movq(to, Address(end_from, qword_count, Address::times_8, - 0));
1485 __ movq(Address(end_to, qword_count, Address::times_8, - 0), to);
1486
1487 __ BIND(L_copy_bytes);
1488 __ addptr(qword_count, 4);
1489 __ jcc(Assembler::lessEqual, L_loop);
1490 }
1491 __ subptr(qword_count, 4);
1492 __ jcc(Assembler::less, L_copy_8_bytes); // Copy trailing qwords
1493 }
1494
1495 // Copy big chunks backward
1496 //
1497 // Inputs:
1498 // from - source arrays address
1499 // dest - destination array address
1500 // qword_count - 64-bits element count
1501 // to - scratch
1502 // L_copy_bytes - entry label
1503 // L_copy_8_bytes - exit label
1504 //
1505 void copy_bytes_backward(Register from, Register dest,
1506 Register qword_count, Register to,
1507 Label& L_copy_bytes, Label& L_copy_8_bytes) {
1508 DEBUG_ONLY(__ stop("enter at entry label, not here"));
1509 Label L_loop;
1510 __ align(OptoLoopAlignment);
1511 if (UseUnalignedLoadStores) {
1512 Label L_end;
1513 // Copy 64-bytes per iteration
1514 __ BIND(L_loop);
1515 if (UseAVX > 2) {
1516 __ evmovdqu(xmm0, Address(from, qword_count, Address::times_8, 32), Assembler::AVX_512bit);
1517 __ evmovdqu(Address(dest, qword_count, Address::times_8, 32), xmm0, Assembler::AVX_512bit);
1518 } else if (UseAVX == 2) {
1519 __ vmovdqu(xmm0, Address(from, qword_count, Address::times_8, 32));
1520 __ vmovdqu(Address(dest, qword_count, Address::times_8, 32), xmm0);
1521 __ vmovdqu(xmm1, Address(from, qword_count, Address::times_8, 0));
1522 __ vmovdqu(Address(dest, qword_count, Address::times_8, 0), xmm1);
1523 } else {
1524 __ movdqu(xmm0, Address(from, qword_count, Address::times_8, 48));
1525 __ movdqu(Address(dest, qword_count, Address::times_8, 48), xmm0);
1526 __ movdqu(xmm1, Address(from, qword_count, Address::times_8, 32));
1527 __ movdqu(Address(dest, qword_count, Address::times_8, 32), xmm1);
1528 __ movdqu(xmm2, Address(from, qword_count, Address::times_8, 16));
1529 __ movdqu(Address(dest, qword_count, Address::times_8, 16), xmm2);
1530 __ movdqu(xmm3, Address(from, qword_count, Address::times_8, 0));
1531 __ movdqu(Address(dest, qword_count, Address::times_8, 0), xmm3);
1532 }
1533 __ BIND(L_copy_bytes);
1534 __ subptr(qword_count, 8);
1535 __ jcc(Assembler::greaterEqual, L_loop);
1536
1537 __ addptr(qword_count, 4); // add(8) and sub(4)
1538 __ jccb(Assembler::less, L_end);
1539 // Copy trailing 32 bytes
1540 if (UseAVX >= 2) {
1541 __ vmovdqu(xmm0, Address(from, qword_count, Address::times_8, 0));
1542 __ vmovdqu(Address(dest, qword_count, Address::times_8, 0), xmm0);
1543 } else {
1544 __ movdqu(xmm0, Address(from, qword_count, Address::times_8, 16));
1545 __ movdqu(Address(dest, qword_count, Address::times_8, 16), xmm0);
1546 __ movdqu(xmm1, Address(from, qword_count, Address::times_8, 0));
1547 __ movdqu(Address(dest, qword_count, Address::times_8, 0), xmm1);
1548 }
1549 __ subptr(qword_count, 4);
1550 __ BIND(L_end);
1551 if (UseAVX >= 2) {
1552 // clean upper bits of YMM registers
1553 __ vpxor(xmm0, xmm0);
1554 __ vpxor(xmm1, xmm1);
1555 }
1556 } else {
1557 // Copy 32-bytes per iteration
1558 __ BIND(L_loop);
1559 __ movq(to, Address(from, qword_count, Address::times_8, 24));
1560 __ movq(Address(dest, qword_count, Address::times_8, 24), to);
1561 __ movq(to, Address(from, qword_count, Address::times_8, 16));
1562 __ movq(Address(dest, qword_count, Address::times_8, 16), to);
1563 __ movq(to, Address(from, qword_count, Address::times_8, 8));
1564 __ movq(Address(dest, qword_count, Address::times_8, 8), to);
1565 __ movq(to, Address(from, qword_count, Address::times_8, 0));
1566 __ movq(Address(dest, qword_count, Address::times_8, 0), to);
1567
1568 __ BIND(L_copy_bytes);
1569 __ subptr(qword_count, 4);
1570 __ jcc(Assembler::greaterEqual, L_loop);
1571 }
1572 __ addptr(qword_count, 4);
1573 __ jcc(Assembler::greater, L_copy_8_bytes); // Copy trailing qwords
1574 }
1575
1576
1577 // Arguments:
1578 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1579 // ignored
1580 // name - stub name string
1581 //
1582 // Inputs:
1583 // c_rarg0 - source array address
1584 // c_rarg1 - destination array address
1585 // c_rarg2 - element count, treated as ssize_t, can be zero
1586 //
1587 // If 'from' and/or 'to' are aligned on 4-, 2-, or 1-byte boundaries,
1588 // we let the hardware handle it. The one to eight bytes within words,
1589 // dwords or qwords that span cache line boundaries will still be loaded
1590 // and stored atomically.
1591 //
1592 // Side Effects:
1593 // disjoint_byte_copy_entry is set to the no-overlap entry point
1594 // used by generate_conjoint_byte_copy().
1595 //
1596 address generate_disjoint_byte_copy(bool aligned, address* entry, const char *name) {
1597 __ align(CodeEntryAlignment);
1598 StubCodeMark mark(this, "StubRoutines", name);
1599 address start = __ pc();
1600
1601 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes, L_copy_2_bytes;
1602 Label L_copy_byte, L_exit;
1603 const Register from = rdi; // source array address
1604 const Register to = rsi; // destination array address
1605 const Register count = rdx; // elements count
1606 const Register byte_count = rcx;
1607 const Register qword_count = count;
1608 const Register end_from = from; // source array end address
1609 const Register end_to = to; // destination array end address
1610 // End pointers are inclusive, and if count is not zero they point
1611 // to the last unit copied: end_to[0] := end_from[0]
1612
1613 __ enter(); // required for proper stackwalking of RuntimeStub frame
1614 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1615
1616 if (entry != NULL) {
1617 *entry = __ pc();
1618 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1619 BLOCK_COMMENT("Entry:");
1620 }
1621
1622 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1623 // r9 and r10 may be used to save non-volatile registers
1624
1625 // 'from', 'to' and 'count' are now valid
1626 __ movptr(byte_count, count);
1627 __ shrptr(count, 3); // count => qword_count
1628
1629 // Copy from low to high addresses. Use 'to' as scratch.
1630 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
1631 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
1632 __ negptr(qword_count); // make the count negative
1633 __ jmp(L_copy_bytes);
1634
1635 // Copy trailing qwords
1636 __ BIND(L_copy_8_bytes);
1637 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
1638 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
1639 __ increment(qword_count);
1640 __ jcc(Assembler::notZero, L_copy_8_bytes);
1641
1642 // Check for and copy trailing dword
1643 __ BIND(L_copy_4_bytes);
1644 __ testl(byte_count, 4);
1645 __ jccb(Assembler::zero, L_copy_2_bytes);
1646 __ movl(rax, Address(end_from, 8));
1647 __ movl(Address(end_to, 8), rax);
1648
1649 __ addptr(end_from, 4);
1650 __ addptr(end_to, 4);
1651
1652 // Check for and copy trailing word
1653 __ BIND(L_copy_2_bytes);
1654 __ testl(byte_count, 2);
1655 __ jccb(Assembler::zero, L_copy_byte);
1656 __ movw(rax, Address(end_from, 8));
1657 __ movw(Address(end_to, 8), rax);
1658
1659 __ addptr(end_from, 2);
1660 __ addptr(end_to, 2);
1661
1662 // Check for and copy trailing byte
1663 __ BIND(L_copy_byte);
1664 __ testl(byte_count, 1);
1665 __ jccb(Assembler::zero, L_exit);
1666 __ movb(rax, Address(end_from, 8));
1667 __ movb(Address(end_to, 8), rax);
1668
1669 __ BIND(L_exit);
1670 restore_arg_regs();
1671 inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); // Update counter after rscratch1 is free
1672 __ xorptr(rax, rax); // return 0
1673 __ leave(); // required for proper stackwalking of RuntimeStub frame
1674 __ ret(0);
1675
1676 // Copy in multi-bytes chunks
1677 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1678 __ jmp(L_copy_4_bytes);
1679
1680 return start;
1681 }
1682
1683 // Arguments:
1684 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1685 // ignored
1686 // name - stub name string
1687 //
1688 // Inputs:
1689 // c_rarg0 - source array address
1690 // c_rarg1 - destination array address
1691 // c_rarg2 - element count, treated as ssize_t, can be zero
1692 //
1693 // If 'from' and/or 'to' are aligned on 4-, 2-, or 1-byte boundaries,
1694 // we let the hardware handle it. The one to eight bytes within words,
1695 // dwords or qwords that span cache line boundaries will still be loaded
1696 // and stored atomically.
1697 //
1698 address generate_conjoint_byte_copy(bool aligned, address nooverlap_target,
1699 address* entry, const char *name) {
1700 __ align(CodeEntryAlignment);
1701 StubCodeMark mark(this, "StubRoutines", name);
1702 address start = __ pc();
1703
1704 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes, L_copy_2_bytes;
1705 const Register from = rdi; // source array address
1706 const Register to = rsi; // destination array address
1707 const Register count = rdx; // elements count
1708 const Register byte_count = rcx;
1709 const Register qword_count = count;
1710
1711 __ enter(); // required for proper stackwalking of RuntimeStub frame
1712 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1713
1714 if (entry != NULL) {
1715 *entry = __ pc();
1716 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1717 BLOCK_COMMENT("Entry:");
1718 }
1719
1720 array_overlap_test(nooverlap_target, Address::times_1);
1721 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1722 // r9 and r10 may be used to save non-volatile registers
1723
1724 // 'from', 'to' and 'count' are now valid
1725 __ movptr(byte_count, count);
1726 __ shrptr(count, 3); // count => qword_count
1727
1728 // Copy from high to low addresses.
1729
1730 // Check for and copy trailing byte
1731 __ testl(byte_count, 1);
1732 __ jcc(Assembler::zero, L_copy_2_bytes);
1733 __ movb(rax, Address(from, byte_count, Address::times_1, -1));
1734 __ movb(Address(to, byte_count, Address::times_1, -1), rax);
1735 __ decrement(byte_count); // Adjust for possible trailing word
1736
1737 // Check for and copy trailing word
1738 __ BIND(L_copy_2_bytes);
1739 __ testl(byte_count, 2);
1740 __ jcc(Assembler::zero, L_copy_4_bytes);
1741 __ movw(rax, Address(from, byte_count, Address::times_1, -2));
1742 __ movw(Address(to, byte_count, Address::times_1, -2), rax);
1743
1744 // Check for and copy trailing dword
1745 __ BIND(L_copy_4_bytes);
1746 __ testl(byte_count, 4);
1747 __ jcc(Assembler::zero, L_copy_bytes);
1748 __ movl(rax, Address(from, qword_count, Address::times_8));
1749 __ movl(Address(to, qword_count, Address::times_8), rax);
1750 __ jmp(L_copy_bytes);
1751
1752 // Copy trailing qwords
1753 __ BIND(L_copy_8_bytes);
1754 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
1755 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
1756 __ decrement(qword_count);
1757 __ jcc(Assembler::notZero, L_copy_8_bytes);
1758
1759 restore_arg_regs();
1760 inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); // Update counter after rscratch1 is free
1761 __ xorptr(rax, rax); // return 0
1762 __ leave(); // required for proper stackwalking of RuntimeStub frame
1763 __ ret(0);
1764
1765 // Copy in multi-bytes chunks
1766 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1767
1768 restore_arg_regs();
1769 inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); // Update counter after rscratch1 is free
1770 __ xorptr(rax, rax); // return 0
1771 __ leave(); // required for proper stackwalking of RuntimeStub frame
1772 __ ret(0);
1773
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 // Side Effects:
1793 // disjoint_short_copy_entry is set to the no-overlap entry point
1794 // used by generate_conjoint_short_copy().
1795 //
1796 address generate_disjoint_short_copy(bool aligned, address *entry, const char *name) {
1797 __ align(CodeEntryAlignment);
1798 StubCodeMark mark(this, "StubRoutines", name);
1799 address start = __ pc();
1800
1801 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes,L_copy_2_bytes,L_exit;
1802 const Register from = rdi; // source array address
1803 const Register to = rsi; // destination array address
1804 const Register count = rdx; // elements count
1805 const Register word_count = rcx;
1806 const Register qword_count = count;
1807 const Register end_from = from; // source array end address
1808 const Register end_to = to; // destination array end address
1809 // End pointers are inclusive, and if count is not zero they point
1810 // to the last unit copied: end_to[0] := end_from[0]
1811
1812 __ enter(); // required for proper stackwalking of RuntimeStub frame
1813 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1814
1815 if (entry != NULL) {
1816 *entry = __ pc();
1817 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1818 BLOCK_COMMENT("Entry:");
1819 }
1820
1821 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1822 // r9 and r10 may be used to save non-volatile registers
1823
1824 // 'from', 'to' and 'count' are now valid
1825 __ movptr(word_count, count);
1826 __ shrptr(count, 2); // count => qword_count
1827
1828 // Copy from low to high addresses. Use 'to' as scratch.
1829 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
1830 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
1831 __ negptr(qword_count);
1832 __ jmp(L_copy_bytes);
1833
1834 // Copy trailing qwords
1835 __ BIND(L_copy_8_bytes);
1836 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
1837 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
1838 __ increment(qword_count);
1839 __ jcc(Assembler::notZero, L_copy_8_bytes);
1840
1841 // Original 'dest' is trashed, so we can't use it as a
1842 // base register for a possible trailing word copy
1843
1844 // Check for and copy trailing dword
1845 __ BIND(L_copy_4_bytes);
1846 __ testl(word_count, 2);
1847 __ jccb(Assembler::zero, L_copy_2_bytes);
1848 __ movl(rax, Address(end_from, 8));
1849 __ movl(Address(end_to, 8), rax);
1850
1851 __ addptr(end_from, 4);
1852 __ addptr(end_to, 4);
1853
1854 // Check for and copy trailing word
1855 __ BIND(L_copy_2_bytes);
1856 __ testl(word_count, 1);
1857 __ jccb(Assembler::zero, L_exit);
1858 __ movw(rax, Address(end_from, 8));
1859 __ movw(Address(end_to, 8), rax);
1860
1861 __ BIND(L_exit);
1862 restore_arg_regs();
1863 inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); // Update counter after rscratch1 is free
1864 __ xorptr(rax, rax); // return 0
1865 __ leave(); // required for proper stackwalking of RuntimeStub frame
1866 __ ret(0);
1867
1868 // Copy in multi-bytes chunks
1869 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1870 __ jmp(L_copy_4_bytes);
1871
1872 return start;
1873 }
1874
1875 address generate_fill(BasicType t, bool aligned, const char *name) {
1876 __ align(CodeEntryAlignment);
1877 StubCodeMark mark(this, "StubRoutines", name);
1878 address start = __ pc();
1879
1880 BLOCK_COMMENT("Entry:");
1881
1882 const Register to = c_rarg0; // source array address
1883 const Register value = c_rarg1; // value
1884 const Register count = c_rarg2; // elements count
1885
1886 __ enter(); // required for proper stackwalking of RuntimeStub frame
1887
1888 __ generate_fill(t, aligned, to, value, count, rax, xmm0);
1889
1890 __ leave(); // required for proper stackwalking of RuntimeStub frame
1891 __ ret(0);
1892 return start;
1893 }
1894
1895 // Arguments:
1896 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1897 // ignored
1898 // name - stub name string
1899 //
1900 // Inputs:
1901 // c_rarg0 - source array address
1902 // c_rarg1 - destination array address
1903 // c_rarg2 - element count, treated as ssize_t, can be zero
1904 //
1905 // If 'from' and/or 'to' are aligned on 4- or 2-byte boundaries, we
1906 // let the hardware handle it. The two or four words within dwords
1907 // or qwords that span cache line boundaries will still be loaded
1908 // and stored atomically.
1909 //
1910 address generate_conjoint_short_copy(bool aligned, address nooverlap_target,
1911 address *entry, const char *name) {
1912 __ align(CodeEntryAlignment);
1913 StubCodeMark mark(this, "StubRoutines", name);
1914 address start = __ pc();
1915
1916 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes;
1917 const Register from = rdi; // source array address
1918 const Register to = rsi; // destination array address
1919 const Register count = rdx; // elements count
1920 const Register word_count = rcx;
1921 const Register qword_count = count;
1922
1923 __ enter(); // required for proper stackwalking of RuntimeStub frame
1924 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1925
1926 if (entry != NULL) {
1927 *entry = __ pc();
1928 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1929 BLOCK_COMMENT("Entry:");
1930 }
1931
1932 array_overlap_test(nooverlap_target, Address::times_2);
1933 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1934 // r9 and r10 may be used to save non-volatile registers
1935
1936 // 'from', 'to' and 'count' are now valid
1937 __ movptr(word_count, count);
1938 __ shrptr(count, 2); // count => qword_count
1939
1940 // Copy from high to low addresses. Use 'to' as scratch.
1941
1942 // Check for and copy trailing word
1943 __ testl(word_count, 1);
1944 __ jccb(Assembler::zero, L_copy_4_bytes);
1945 __ movw(rax, Address(from, word_count, Address::times_2, -2));
1946 __ movw(Address(to, word_count, Address::times_2, -2), rax);
1947
1948 // Check for and copy trailing dword
1949 __ BIND(L_copy_4_bytes);
1950 __ testl(word_count, 2);
1951 __ jcc(Assembler::zero, L_copy_bytes);
1952 __ movl(rax, Address(from, qword_count, Address::times_8));
1953 __ movl(Address(to, qword_count, Address::times_8), rax);
1954 __ jmp(L_copy_bytes);
1955
1956 // Copy trailing qwords
1957 __ BIND(L_copy_8_bytes);
1958 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
1959 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
1960 __ decrement(qword_count);
1961 __ jcc(Assembler::notZero, L_copy_8_bytes);
1962
1963 restore_arg_regs();
1964 inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); // Update counter after rscratch1 is free
1965 __ xorptr(rax, rax); // return 0
1966 __ leave(); // required for proper stackwalking of RuntimeStub frame
1967 __ ret(0);
1968
1969 // Copy in multi-bytes chunks
1970 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1971
1972 restore_arg_regs();
1973 inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); // Update counter after rscratch1 is free
1974 __ xorptr(rax, rax); // return 0
1975 __ leave(); // required for proper stackwalking of RuntimeStub frame
1976 __ ret(0);
1977
1978 return start;
1979 }
1980
1981 // Arguments:
1982 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1983 // ignored
1984 // is_oop - true => oop array, so generate store check code
1985 // name - stub name string
1986 //
1987 // Inputs:
1988 // c_rarg0 - source array address
1989 // c_rarg1 - destination array address
1990 // c_rarg2 - element count, treated as ssize_t, can be zero
1991 //
1992 // If 'from' and/or 'to' are aligned on 4-byte boundaries, we let
1993 // the hardware handle it. The two dwords within qwords that span
1994 // cache line boundaries will still be loaded and stored atomicly.
1995 //
1996 // Side Effects:
1997 // disjoint_int_copy_entry is set to the no-overlap entry point
1998 // used by generate_conjoint_int_oop_copy().
1999 //
2000 address generate_disjoint_int_oop_copy(bool aligned, bool is_oop, address* entry,
2001 const char *name, bool dest_uninitialized = false) {
2002 __ align(CodeEntryAlignment);
2003 StubCodeMark mark(this, "StubRoutines", name);
2004 address start = __ pc();
2005
2006 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes, L_exit;
2007 const Register from = rdi; // source array address
2008 const Register to = rsi; // destination array address
2009 const Register count = rdx; // elements count
2010 const Register dword_count = rcx;
2011 const Register qword_count = count;
2012 const Register end_from = from; // source array end address
2013 const Register end_to = to; // destination array end address
2014 const Register saved_to = r11; // saved destination array address
2015 // End pointers are inclusive, and if count is not zero they point
2016 // to the last unit copied: end_to[0] := end_from[0]
2017
2018 __ enter(); // required for proper stackwalking of RuntimeStub frame
2019 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
2020
2021 if (entry != NULL) {
2022 *entry = __ pc();
2023 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
2024 BLOCK_COMMENT("Entry:");
2025 }
2026
2027 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
2028 // r9 and r10 may be used to save non-volatile registers
2029 if (is_oop) {
2030 __ movq(saved_to, to);
2031 gen_write_ref_array_pre_barrier(to, count, dest_uninitialized);
2032 }
2033
2034 // 'from', 'to' and 'count' are now valid
2035 __ movptr(dword_count, count);
2036 __ shrptr(count, 1); // count => qword_count
2037
2038 // Copy from low to high addresses. Use 'to' as scratch.
2039 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
2040 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
2041 __ negptr(qword_count);
2042 __ jmp(L_copy_bytes);
2043
2044 // Copy trailing qwords
2045 __ BIND(L_copy_8_bytes);
2046 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
2047 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
2048 __ increment(qword_count);
2049 __ jcc(Assembler::notZero, L_copy_8_bytes);
2050
2051 // Check for and copy trailing dword
2052 __ BIND(L_copy_4_bytes);
2053 __ testl(dword_count, 1); // Only byte test since the value is 0 or 1
2054 __ jccb(Assembler::zero, L_exit);
2055 __ movl(rax, Address(end_from, 8));
2056 __ movl(Address(end_to, 8), rax);
2057
2058 __ BIND(L_exit);
2059 if (is_oop) {
2060 gen_write_ref_array_post_barrier(saved_to, dword_count, rax);
2061 }
2062 restore_arg_regs();
2063 inc_counter_np(SharedRuntime::_jint_array_copy_ctr); // Update counter after rscratch1 is free
2064 __ xorptr(rax, rax); // return 0
2065 __ leave(); // required for proper stackwalking of RuntimeStub frame
2066 __ ret(0);
2067
2068 // Copy in multi-bytes chunks
2069 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
2070 __ jmp(L_copy_4_bytes);
2071
2072 return start;
2073 }
2074
2075 // Arguments:
2076 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
2077 // ignored
2078 // is_oop - true => oop array, so generate store check code
2079 // name - stub name string
2080 //
2081 // Inputs:
2082 // c_rarg0 - source array address
2083 // c_rarg1 - destination array address
2084 // c_rarg2 - element count, treated as ssize_t, can be zero
2085 //
2086 // If 'from' and/or 'to' are aligned on 4-byte boundaries, we let
2087 // the hardware handle it. The two dwords within qwords that span
2088 // cache line boundaries will still be loaded and stored atomicly.
2089 //
2090 address generate_conjoint_int_oop_copy(bool aligned, bool is_oop, address nooverlap_target,
2091 address *entry, const char *name,
2092 bool dest_uninitialized = false) {
2093 __ align(CodeEntryAlignment);
2094 StubCodeMark mark(this, "StubRoutines", name);
2095 address start = __ pc();
2096
2097 Label L_copy_bytes, L_copy_8_bytes, L_copy_2_bytes, L_exit;
2098 const Register from = rdi; // source array address
2099 const Register to = rsi; // destination array address
2100 const Register count = rdx; // elements count
2101 const Register dword_count = rcx;
2102 const Register qword_count = count;
2103
2104 __ enter(); // required for proper stackwalking of RuntimeStub frame
2105 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
2106
2107 if (entry != NULL) {
2108 *entry = __ pc();
2109 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
2110 BLOCK_COMMENT("Entry:");
2111 }
2112
2113 array_overlap_test(nooverlap_target, Address::times_4);
2114 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
2115 // r9 and r10 may be used to save non-volatile registers
2116
2117 if (is_oop) {
2118 // no registers are destroyed by this call
2119 gen_write_ref_array_pre_barrier(to, count, dest_uninitialized);
2120 }
2121
2122 assert_clean_int(count, rax); // Make sure 'count' is clean int.
2123 // 'from', 'to' and 'count' are now valid
2124 __ movptr(dword_count, count);
2125 __ shrptr(count, 1); // count => qword_count
2126
2127 // Copy from high to low addresses. Use 'to' as scratch.
2128
2129 // Check for and copy trailing dword
2130 __ testl(dword_count, 1);
2131 __ jcc(Assembler::zero, L_copy_bytes);
2132 __ movl(rax, Address(from, dword_count, Address::times_4, -4));
2133 __ movl(Address(to, dword_count, Address::times_4, -4), rax);
2134 __ jmp(L_copy_bytes);
2135
2136 // Copy trailing qwords
2137 __ BIND(L_copy_8_bytes);
2138 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
2139 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
2140 __ decrement(qword_count);
2141 __ jcc(Assembler::notZero, L_copy_8_bytes);
2142
2143 if (is_oop) {
2144 __ jmp(L_exit);
2145 }
2146 restore_arg_regs();
2147 inc_counter_np(SharedRuntime::_jint_array_copy_ctr); // Update counter after rscratch1 is free
2148 __ xorptr(rax, rax); // return 0
2149 __ leave(); // required for proper stackwalking of RuntimeStub frame
2150 __ ret(0);
2151
2152 // Copy in multi-bytes chunks
2153 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
2154
2155 __ BIND(L_exit);
2156 if (is_oop) {
2157 gen_write_ref_array_post_barrier(to, dword_count, rax);
2158 }
2159 restore_arg_regs();
2160 inc_counter_np(SharedRuntime::_jint_array_copy_ctr); // Update counter after rscratch1 is free
2161 __ xorptr(rax, rax); // return 0
2162 __ leave(); // required for proper stackwalking of RuntimeStub frame
2163 __ ret(0);
2164
2165 return start;
2166 }
2167
2168 // Arguments:
2169 // aligned - true => Input and output aligned on a HeapWord boundary == 8 bytes
2170 // ignored
2171 // is_oop - true => oop array, so generate store check code
2172 // name - stub name string
2173 //
2174 // Inputs:
2175 // c_rarg0 - source array address
2176 // c_rarg1 - destination array address
2177 // c_rarg2 - element count, treated as ssize_t, can be zero
2178 //
2179 // Side Effects:
2180 // disjoint_oop_copy_entry or disjoint_long_copy_entry is set to the
2181 // no-overlap entry point used by generate_conjoint_long_oop_copy().
2182 //
2183 address generate_disjoint_long_oop_copy(bool aligned, bool is_oop, address *entry,
2184 const char *name, bool dest_uninitialized = false) {
2185 __ align(CodeEntryAlignment);
2186 StubCodeMark mark(this, "StubRoutines", name);
2187 address start = __ pc();
2188
2189 Label L_copy_bytes, L_copy_8_bytes, L_exit;
2190 const Register from = rdi; // source array address
2191 const Register to = rsi; // destination array address
2192 const Register qword_count = rdx; // elements count
2193 const Register end_from = from; // source array end address
2194 const Register end_to = rcx; // destination array end address
2195 const Register saved_to = to;
2196 const Register saved_count = r11;
2197 // End pointers are inclusive, and if count is not zero they point
2198 // to the last unit copied: end_to[0] := end_from[0]
2199
2200 __ enter(); // required for proper stackwalking of RuntimeStub frame
2201 // Save no-overlap entry point for generate_conjoint_long_oop_copy()
2202 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
2203
2204 if (entry != NULL) {
2205 *entry = __ pc();
2206 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
2207 BLOCK_COMMENT("Entry:");
2208 }
2209
2210 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
2211 // r9 and r10 may be used to save non-volatile registers
2212 // 'from', 'to' and 'qword_count' are now valid
2213 if (is_oop) {
2214 // Save to and count for store barrier
2215 __ movptr(saved_count, qword_count);
2216 // no registers are destroyed by this call
2217 gen_write_ref_array_pre_barrier(to, qword_count, dest_uninitialized);
2218 }
2219
2220 // Copy from low to high addresses. Use 'to' as scratch.
2221 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
2222 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
2223 __ negptr(qword_count);
2224 __ jmp(L_copy_bytes);
2225
2226 // Copy trailing qwords
2227 __ BIND(L_copy_8_bytes);
2228 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
2229 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
2230 __ increment(qword_count);
2231 __ jcc(Assembler::notZero, L_copy_8_bytes);
2232
2233 if (is_oop) {
2234 __ jmp(L_exit);
2235 } else {
2236 restore_arg_regs();
2237 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2238 __ xorptr(rax, rax); // return 0
2239 __ leave(); // required for proper stackwalking of RuntimeStub frame
2240 __ ret(0);
2241 }
2242
2243 // Copy in multi-bytes chunks
2244 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
2245
2246 if (is_oop) {
2247 __ BIND(L_exit);
2248 gen_write_ref_array_post_barrier(saved_to, saved_count, rax);
2249 }
2250 restore_arg_regs();
2251 if (is_oop) {
2252 inc_counter_np(SharedRuntime::_oop_array_copy_ctr); // Update counter after rscratch1 is free
2253 } else {
2254 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2255 }
2256 __ xorptr(rax, rax); // return 0
2257 __ leave(); // required for proper stackwalking of RuntimeStub frame
2258 __ ret(0);
2259
2260 return start;
2261 }
2262
2263 // Arguments:
2264 // aligned - true => Input and output aligned on a HeapWord boundary == 8 bytes
2265 // ignored
2266 // is_oop - true => oop array, so generate store check code
2267 // name - stub name string
2268 //
2269 // Inputs:
2270 // c_rarg0 - source array address
2271 // c_rarg1 - destination array address
2272 // c_rarg2 - element count, treated as ssize_t, can be zero
2273 //
2274 address generate_conjoint_long_oop_copy(bool aligned, bool is_oop,
2275 address nooverlap_target, address *entry,
2276 const char *name, bool dest_uninitialized = false) {
2277 __ align(CodeEntryAlignment);
2278 StubCodeMark mark(this, "StubRoutines", name);
2279 address start = __ pc();
2280
2281 Label L_copy_bytes, L_copy_8_bytes, L_exit;
2282 const Register from = rdi; // source array address
2283 const Register to = rsi; // destination array address
2284 const Register qword_count = rdx; // elements count
2285 const Register saved_count = rcx;
2286
2287 __ enter(); // required for proper stackwalking of RuntimeStub frame
2288 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
2289
2290 if (entry != NULL) {
2291 *entry = __ pc();
2292 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
2293 BLOCK_COMMENT("Entry:");
2294 }
2295
2296 array_overlap_test(nooverlap_target, Address::times_8);
2297 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
2298 // r9 and r10 may be used to save non-volatile registers
2299 // 'from', 'to' and 'qword_count' are now valid
2300 if (is_oop) {
2301 // Save to and count for store barrier
2302 __ movptr(saved_count, qword_count);
2303 // No registers are destroyed by this call
2304 gen_write_ref_array_pre_barrier(to, saved_count, dest_uninitialized);
2305 }
2306
2307 __ jmp(L_copy_bytes);
2308
2309 // Copy trailing qwords
2310 __ BIND(L_copy_8_bytes);
2311 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
2312 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
2313 __ decrement(qword_count);
2314 __ jcc(Assembler::notZero, L_copy_8_bytes);
2315
2316 if (is_oop) {
2317 __ jmp(L_exit);
2318 } else {
2319 restore_arg_regs();
2320 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2321 __ xorptr(rax, rax); // return 0
2322 __ leave(); // required for proper stackwalking of RuntimeStub frame
2323 __ ret(0);
2324 }
2325
2326 // Copy in multi-bytes chunks
2327 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
2328
2329 if (is_oop) {
2330 __ BIND(L_exit);
2331 gen_write_ref_array_post_barrier(to, saved_count, rax);
2332 }
2333 restore_arg_regs();
2334 if (is_oop) {
2335 inc_counter_np(SharedRuntime::_oop_array_copy_ctr); // Update counter after rscratch1 is free
2336 } else {
2337 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2338 }
2339 __ xorptr(rax, rax); // return 0
2340 __ leave(); // required for proper stackwalking of RuntimeStub frame
2341 __ ret(0);
2342
2343 return start;
2344 }
2345
2346
2347 // Helper for generating a dynamic type check.
2348 // Smashes no registers.
2349 void generate_type_check(Register sub_klass,
2350 Register super_check_offset,
2351 Register super_klass,
2352 Label& L_success) {
2353 assert_different_registers(sub_klass, super_check_offset, super_klass);
2354
2355 BLOCK_COMMENT("type_check:");
2356
2357 Label L_miss;
2358
2359 __ check_klass_subtype_fast_path(sub_klass, super_klass, noreg, &L_success, &L_miss, NULL,
2360 super_check_offset);
2361 __ check_klass_subtype_slow_path(sub_klass, super_klass, noreg, noreg, &L_success, NULL);
2362
2363 // Fall through on failure!
2364 __ BIND(L_miss);
2365 }
2366
2367 //
2368 // Generate checkcasting array copy stub
2369 //
2370 // Input:
2371 // c_rarg0 - source array address
2372 // c_rarg1 - destination array address
2373 // c_rarg2 - element count, treated as ssize_t, can be zero
2374 // c_rarg3 - size_t ckoff (super_check_offset)
2375 // not Win64
2376 // c_rarg4 - oop ckval (super_klass)
2377 // Win64
2378 // rsp+40 - oop ckval (super_klass)
2379 //
2380 // Output:
2381 // rax == 0 - success
2382 // rax == -1^K - failure, where K is partial transfer count
2383 //
2384 address generate_checkcast_copy(const char *name, address *entry,
2385 bool dest_uninitialized = false) {
2386
2387 Label L_load_element, L_store_element, L_do_card_marks, L_done;
2388
2389 // Input registers (after setup_arg_regs)
2390 const Register from = rdi; // source array address
2391 const Register to = rsi; // destination array address
2392 const Register length = rdx; // elements count
2393 const Register ckoff = rcx; // super_check_offset
2394 const Register ckval = r8; // super_klass
2395
2396 // Registers used as temps (r13, r14 are save-on-entry)
2397 const Register end_from = from; // source array end address
2398 const Register end_to = r13; // destination array end address
2399 const Register count = rdx; // -(count_remaining)
2400 const Register r14_length = r14; // saved copy of length
2401 // End pointers are inclusive, and if length is not zero they point
2402 // to the last unit copied: end_to[0] := end_from[0]
2403
2404 const Register rax_oop = rax; // actual oop copied
2405 const Register r11_klass = r11; // oop._klass
2406
2407 //---------------------------------------------------------------
2408 // Assembler stub will be used for this call to arraycopy
2409 // if the two arrays are subtypes of Object[] but the
2410 // destination array type is not equal to or a supertype
2411 // of the source type. Each element must be separately
2412 // checked.
2413
2414 __ align(CodeEntryAlignment);
2415 StubCodeMark mark(this, "StubRoutines", name);
2416 address start = __ pc();
2417
2418 __ enter(); // required for proper stackwalking of RuntimeStub frame
2419
2420 #ifdef ASSERT
2421 // caller guarantees that the arrays really are different
2422 // otherwise, we would have to make conjoint checks
2423 { Label L;
2424 array_overlap_test(L, TIMES_OOP);
2425 __ stop("checkcast_copy within a single array");
2426 __ bind(L);
2427 }
2428 #endif //ASSERT
2429
2430 setup_arg_regs(4); // from => rdi, to => rsi, length => rdx
2431 // ckoff => rcx, ckval => r8
2432 // r9 and r10 may be used to save non-volatile registers
2433 #ifdef _WIN64
2434 // last argument (#4) is on stack on Win64
2435 __ movptr(ckval, Address(rsp, 6 * wordSize));
2436 #endif
2437
2438 // Caller of this entry point must set up the argument registers.
2439 if (entry != NULL) {
2440 *entry = __ pc();
2441 BLOCK_COMMENT("Entry:");
2442 }
2443
2444 // allocate spill slots for r13, r14
2445 enum {
2446 saved_r13_offset,
2447 saved_r14_offset,
2448 saved_rbp_offset
2449 };
2450 __ subptr(rsp, saved_rbp_offset * wordSize);
2451 __ movptr(Address(rsp, saved_r13_offset * wordSize), r13);
2452 __ movptr(Address(rsp, saved_r14_offset * wordSize), r14);
2453
2454 // check that int operands are properly extended to size_t
2455 assert_clean_int(length, rax);
2456 assert_clean_int(ckoff, rax);
2457
2458 #ifdef ASSERT
2459 BLOCK_COMMENT("assert consistent ckoff/ckval");
2460 // The ckoff and ckval must be mutually consistent,
2461 // even though caller generates both.
2462 { Label L;
2463 int sco_offset = in_bytes(Klass::super_check_offset_offset());
2464 __ cmpl(ckoff, Address(ckval, sco_offset));
2465 __ jcc(Assembler::equal, L);
2466 __ stop("super_check_offset inconsistent");
2467 __ bind(L);
2468 }
2469 #endif //ASSERT
2470
2471 // Loop-invariant addresses. They are exclusive end pointers.
2472 Address end_from_addr(from, length, TIMES_OOP, 0);
2473 Address end_to_addr(to, length, TIMES_OOP, 0);
2474 // Loop-variant addresses. They assume post-incremented count < 0.
2475 Address from_element_addr(end_from, count, TIMES_OOP, 0);
2476 Address to_element_addr(end_to, count, TIMES_OOP, 0);
2477
2478 gen_write_ref_array_pre_barrier(to, count, dest_uninitialized);
2479
2480 // Copy from low to high addresses, indexed from the end of each array.
2481 __ lea(end_from, end_from_addr);
2482 __ lea(end_to, end_to_addr);
2483 __ movptr(r14_length, length); // save a copy of the length
2484 assert(length == count, ""); // else fix next line:
2485 __ negptr(count); // negate and test the length
2486 __ jcc(Assembler::notZero, L_load_element);
2487
2488 // Empty array: Nothing to do.
2489 __ xorptr(rax, rax); // return 0 on (trivial) success
2490 __ jmp(L_done);
2491
2492 // ======== begin loop ========
2493 // (Loop is rotated; its entry is L_load_element.)
2494 // Loop control:
2495 // for (count = -count; count != 0; count++)
2496 // Base pointers src, dst are biased by 8*(count-1),to last element.
2497 __ align(OptoLoopAlignment);
2498
2499 __ BIND(L_store_element);
2500 __ store_heap_oop(to_element_addr, rax_oop); // store the oop
2501 __ increment(count); // increment the count toward zero
2502 __ jcc(Assembler::zero, L_do_card_marks);
2503
2504 // ======== loop entry is here ========
2505 __ BIND(L_load_element);
2506 __ load_heap_oop(rax_oop, from_element_addr); // load the oop
2507 __ testptr(rax_oop, rax_oop);
2508 __ jcc(Assembler::zero, L_store_element);
2509
2510 __ load_klass(r11_klass, rax_oop);// query the object klass
2511 generate_type_check(r11_klass, ckoff, ckval, L_store_element);
2512 // ======== end loop ========
2513
2514 // It was a real error; we must depend on the caller to finish the job.
2515 // Register rdx = -1 * number of *remaining* oops, r14 = *total* oops.
2516 // Emit GC store barriers for the oops we have copied (r14 + rdx),
2517 // and report their number to the caller.
2518 assert_different_registers(rax, r14_length, count, to, end_to, rcx, rscratch1);
2519 Label L_post_barrier;
2520 __ addptr(r14_length, count); // K = (original - remaining) oops
2521 __ movptr(rax, r14_length); // save the value
2522 __ notptr(rax); // report (-1^K) to caller (does not affect flags)
2523 __ jccb(Assembler::notZero, L_post_barrier);
2524 __ jmp(L_done); // K == 0, nothing was copied, skip post barrier
2525
2526 // Come here on success only.
2527 __ BIND(L_do_card_marks);
2528 __ xorptr(rax, rax); // return 0 on success
2529
2530 __ BIND(L_post_barrier);
2531 gen_write_ref_array_post_barrier(to, r14_length, rscratch1);
2532
2533 // Common exit point (success or failure).
2534 __ BIND(L_done);
2535 __ movptr(r13, Address(rsp, saved_r13_offset * wordSize));
2536 __ movptr(r14, Address(rsp, saved_r14_offset * wordSize));
2537 restore_arg_regs();
2538 inc_counter_np(SharedRuntime::_checkcast_array_copy_ctr); // Update counter after rscratch1 is free
2539 __ leave(); // required for proper stackwalking of RuntimeStub frame
2540 __ ret(0);
2541
2542 return start;
2543 }
2544
2545 //
2546 // Generate 'unsafe' array copy stub
2547 // Though just as safe as the other stubs, it takes an unscaled
2548 // size_t argument instead of an element count.
2549 //
2550 // Input:
2551 // c_rarg0 - source array address
2552 // c_rarg1 - destination array address
2553 // c_rarg2 - byte count, treated as ssize_t, can be zero
2554 //
2555 // Examines the alignment of the operands and dispatches
2556 // to a long, int, short, or byte copy loop.
2557 //
2558 address generate_unsafe_copy(const char *name,
2559 address byte_copy_entry, address short_copy_entry,
2560 address int_copy_entry, address long_copy_entry) {
2561
2562 Label L_long_aligned, L_int_aligned, L_short_aligned;
2563
2564 // Input registers (before setup_arg_regs)
2565 const Register from = c_rarg0; // source array address
2566 const Register to = c_rarg1; // destination array address
2567 const Register size = c_rarg2; // byte count (size_t)
2568
2569 // Register used as a temp
2570 const Register bits = rax; // test copy of low bits
2571
2572 __ align(CodeEntryAlignment);
2573 StubCodeMark mark(this, "StubRoutines", name);
2574 address start = __ pc();
2575
2576 __ enter(); // required for proper stackwalking of RuntimeStub frame
2577
2578 // bump this on entry, not on exit:
2579 inc_counter_np(SharedRuntime::_unsafe_array_copy_ctr);
2580
2581 __ mov(bits, from);
2582 __ orptr(bits, to);
2583 __ orptr(bits, size);
2584
2585 __ testb(bits, BytesPerLong-1);
2586 __ jccb(Assembler::zero, L_long_aligned);
2587
2588 __ testb(bits, BytesPerInt-1);
2589 __ jccb(Assembler::zero, L_int_aligned);
2590
2591 __ testb(bits, BytesPerShort-1);
2592 __ jump_cc(Assembler::notZero, RuntimeAddress(byte_copy_entry));
2593
2594 __ BIND(L_short_aligned);
2595 __ shrptr(size, LogBytesPerShort); // size => short_count
2596 __ jump(RuntimeAddress(short_copy_entry));
2597
2598 __ BIND(L_int_aligned);
2599 __ shrptr(size, LogBytesPerInt); // size => int_count
2600 __ jump(RuntimeAddress(int_copy_entry));
2601
2602 __ BIND(L_long_aligned);
2603 __ shrptr(size, LogBytesPerLong); // size => qword_count
2604 __ jump(RuntimeAddress(long_copy_entry));
2605
2606 return start;
2607 }
2608
2609 // Perform range checks on the proposed arraycopy.
2610 // Kills temp, but nothing else.
2611 // Also, clean the sign bits of src_pos and dst_pos.
2612 void arraycopy_range_checks(Register src, // source array oop (c_rarg0)
2613 Register src_pos, // source position (c_rarg1)
2614 Register dst, // destination array oo (c_rarg2)
2615 Register dst_pos, // destination position (c_rarg3)
2616 Register length,
2617 Register temp,
2618 Label& L_failed) {
2619 BLOCK_COMMENT("arraycopy_range_checks:");
2620
2621 // if (src_pos + length > arrayOop(src)->length()) FAIL;
2622 __ movl(temp, length);
2623 __ addl(temp, src_pos); // src_pos + length
2624 __ cmpl(temp, Address(src, arrayOopDesc::length_offset_in_bytes()));
2625 __ jcc(Assembler::above, L_failed);
2626
2627 // if (dst_pos + length > arrayOop(dst)->length()) FAIL;
2628 __ movl(temp, length);
2629 __ addl(temp, dst_pos); // dst_pos + length
2630 __ cmpl(temp, Address(dst, arrayOopDesc::length_offset_in_bytes()));
2631 __ jcc(Assembler::above, L_failed);
2632
2633 // Have to clean up high 32-bits of 'src_pos' and 'dst_pos'.
2634 // Move with sign extension can be used since they are positive.
2635 __ movslq(src_pos, src_pos);
2636 __ movslq(dst_pos, dst_pos);
2637
2638 BLOCK_COMMENT("arraycopy_range_checks done");
2639 }
2640
2641 //
2642 // Generate generic array copy stubs
2643 //
2644 // Input:
2645 // c_rarg0 - src oop
2646 // c_rarg1 - src_pos (32-bits)
2647 // c_rarg2 - dst oop
2648 // c_rarg3 - dst_pos (32-bits)
2649 // not Win64
2650 // c_rarg4 - element count (32-bits)
2651 // Win64
2652 // rsp+40 - element count (32-bits)
2653 //
2654 // Output:
2655 // rax == 0 - success
2656 // rax == -1^K - failure, where K is partial transfer count
2657 //
2658 address generate_generic_copy(const char *name,
2659 address byte_copy_entry, address short_copy_entry,
2660 address int_copy_entry, address oop_copy_entry,
2661 address long_copy_entry, address checkcast_copy_entry) {
2662
2663 Label L_failed, L_failed_0, L_objArray;
2664 Label L_copy_bytes, L_copy_shorts, L_copy_ints, L_copy_longs;
2665
2666 // Input registers
2667 const Register src = c_rarg0; // source array oop
2668 const Register src_pos = c_rarg1; // source position
2669 const Register dst = c_rarg2; // destination array oop
2670 const Register dst_pos = c_rarg3; // destination position
2671 #ifndef _WIN64
2672 const Register length = c_rarg4;
2673 #else
2674 const Address length(rsp, 6 * wordSize); // elements count is on stack on Win64
2675 #endif
2676
2677 { int modulus = CodeEntryAlignment;
2678 int target = modulus - 5; // 5 = sizeof jmp(L_failed)
2679 int advance = target - (__ offset() % modulus);
2680 if (advance < 0) advance += modulus;
2681 if (advance > 0) __ nop(advance);
2682 }
2683 StubCodeMark mark(this, "StubRoutines", name);
2684
2685 // Short-hop target to L_failed. Makes for denser prologue code.
2686 __ BIND(L_failed_0);
2687 __ jmp(L_failed);
2688 assert(__ offset() % CodeEntryAlignment == 0, "no further alignment needed");
2689
2690 __ align(CodeEntryAlignment);
2691 address start = __ pc();
2692
2693 __ enter(); // required for proper stackwalking of RuntimeStub frame
2694
2695 // bump this on entry, not on exit:
2696 inc_counter_np(SharedRuntime::_generic_array_copy_ctr);
2697
2698 //-----------------------------------------------------------------------
2699 // Assembler stub will be used for this call to arraycopy
2700 // if the following conditions are met:
2701 //
2702 // (1) src and dst must not be null.
2703 // (2) src_pos must not be negative.
2704 // (3) dst_pos must not be negative.
2705 // (4) length must not be negative.
2706 // (5) src klass and dst klass should be the same and not NULL.
2707 // (6) src and dst should be arrays.
2708 // (7) src_pos + length must not exceed length of src.
2709 // (8) dst_pos + length must not exceed length of dst.
2710 //
2711
2712 // if (src == NULL) return -1;
2713 __ testptr(src, src); // src oop
2714 size_t j1off = __ offset();
2715 __ jccb(Assembler::zero, L_failed_0);
2716
2717 // if (src_pos < 0) return -1;
2718 __ testl(src_pos, src_pos); // src_pos (32-bits)
2719 __ jccb(Assembler::negative, L_failed_0);
2720
2721 // if (dst == NULL) return -1;
2722 __ testptr(dst, dst); // dst oop
2723 __ jccb(Assembler::zero, L_failed_0);
2724
2725 // if (dst_pos < 0) return -1;
2726 __ testl(dst_pos, dst_pos); // dst_pos (32-bits)
2727 size_t j4off = __ offset();
2728 __ jccb(Assembler::negative, L_failed_0);
2729
2730 // The first four tests are very dense code,
2731 // but not quite dense enough to put four
2732 // jumps in a 16-byte instruction fetch buffer.
2733 // That's good, because some branch predicters
2734 // do not like jumps so close together.
2735 // Make sure of this.
2736 guarantee(((j1off ^ j4off) & ~15) != 0, "I$ line of 1st & 4th jumps");
2737
2738 // registers used as temp
2739 const Register r11_length = r11; // elements count to copy
2740 const Register r10_src_klass = r10; // array klass
2741
2742 // if (length < 0) return -1;
2743 __ movl(r11_length, length); // length (elements count, 32-bits value)
2744 __ testl(r11_length, r11_length);
2745 __ jccb(Assembler::negative, L_failed_0);
2746
2747 __ load_klass(r10_src_klass, src);
2748 #ifdef ASSERT
2749 // assert(src->klass() != NULL);
2750 {
2751 BLOCK_COMMENT("assert klasses not null {");
2752 Label L1, L2;
2753 __ testptr(r10_src_klass, r10_src_klass);
2754 __ jcc(Assembler::notZero, L2); // it is broken if klass is NULL
2755 __ bind(L1);
2756 __ stop("broken null klass");
2757 __ bind(L2);
2758 __ load_klass(rax, dst);
2759 __ cmpq(rax, 0);
2760 __ jcc(Assembler::equal, L1); // this would be broken also
2761 BLOCK_COMMENT("} assert klasses not null done");
2762 }
2763 #endif
2764
2765 // Load layout helper (32-bits)
2766 //
2767 // |array_tag| | header_size | element_type | |log2_element_size|
2768 // 32 30 24 16 8 2 0
2769 //
2770 // array_tag: typeArray = 0x3, objArray = 0x2, non-array = 0x0
2771 //
2772
2773 const int lh_offset = in_bytes(Klass::layout_helper_offset());
2774
2775 // Handle objArrays completely differently...
2776 const jint objArray_lh = Klass::array_layout_helper(T_OBJECT);
2777 __ cmpl(Address(r10_src_klass, lh_offset), objArray_lh);
2778 __ jcc(Assembler::equal, L_objArray);
2779
2780 // if (src->klass() != dst->klass()) return -1;
2781 __ load_klass(rax, dst);
2782 __ cmpq(r10_src_klass, rax);
2783 __ jcc(Assembler::notEqual, L_failed);
2784
2785 const Register rax_lh = rax; // layout helper
2786 __ movl(rax_lh, Address(r10_src_klass, lh_offset));
2787
2788 // if (!src->is_Array()) return -1;
2789 __ cmpl(rax_lh, Klass::_lh_neutral_value);
2790 __ jcc(Assembler::greaterEqual, L_failed);
2791
2792 // At this point, it is known to be a typeArray (array_tag 0x3).
2793 #ifdef ASSERT
2794 {
2795 BLOCK_COMMENT("assert primitive array {");
2796 Label L;
2797 __ cmpl(rax_lh, (Klass::_lh_array_tag_type_value << Klass::_lh_array_tag_shift));
2798 __ jcc(Assembler::greaterEqual, L);
2799 __ stop("must be a primitive array");
2800 __ bind(L);
2801 BLOCK_COMMENT("} assert primitive array done");
2802 }
2803 #endif
2804
2805 arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length,
2806 r10, L_failed);
2807
2808 // TypeArrayKlass
2809 //
2810 // src_addr = (src + array_header_in_bytes()) + (src_pos << log2elemsize);
2811 // dst_addr = (dst + array_header_in_bytes()) + (dst_pos << log2elemsize);
2812 //
2813
2814 const Register r10_offset = r10; // array offset
2815 const Register rax_elsize = rax_lh; // element size
2816
2817 __ movl(r10_offset, rax_lh);
2818 __ shrl(r10_offset, Klass::_lh_header_size_shift);
2819 __ andptr(r10_offset, Klass::_lh_header_size_mask); // array_offset
2820 __ addptr(src, r10_offset); // src array offset
2821 __ addptr(dst, r10_offset); // dst array offset
2822 BLOCK_COMMENT("choose copy loop based on element size");
2823 __ andl(rax_lh, Klass::_lh_log2_element_size_mask); // rax_lh -> rax_elsize
2824
2825 // next registers should be set before the jump to corresponding stub
2826 const Register from = c_rarg0; // source array address
2827 const Register to = c_rarg1; // destination array address
2828 const Register count = c_rarg2; // elements count
2829
2830 // 'from', 'to', 'count' registers should be set in such order
2831 // since they are the same as 'src', 'src_pos', 'dst'.
2832
2833 __ BIND(L_copy_bytes);
2834 __ cmpl(rax_elsize, 0);
2835 __ jccb(Assembler::notEqual, L_copy_shorts);
2836 __ lea(from, Address(src, src_pos, Address::times_1, 0));// src_addr
2837 __ lea(to, Address(dst, dst_pos, Address::times_1, 0));// dst_addr
2838 __ movl2ptr(count, r11_length); // length
2839 __ jump(RuntimeAddress(byte_copy_entry));
2840
2841 __ BIND(L_copy_shorts);
2842 __ cmpl(rax_elsize, LogBytesPerShort);
2843 __ jccb(Assembler::notEqual, L_copy_ints);
2844 __ lea(from, Address(src, src_pos, Address::times_2, 0));// src_addr
2845 __ lea(to, Address(dst, dst_pos, Address::times_2, 0));// dst_addr
2846 __ movl2ptr(count, r11_length); // length
2847 __ jump(RuntimeAddress(short_copy_entry));
2848
2849 __ BIND(L_copy_ints);
2850 __ cmpl(rax_elsize, LogBytesPerInt);
2851 __ jccb(Assembler::notEqual, L_copy_longs);
2852 __ lea(from, Address(src, src_pos, Address::times_4, 0));// src_addr
2853 __ lea(to, Address(dst, dst_pos, Address::times_4, 0));// dst_addr
2854 __ movl2ptr(count, r11_length); // length
2855 __ jump(RuntimeAddress(int_copy_entry));
2856
2857 __ BIND(L_copy_longs);
2858 #ifdef ASSERT
2859 {
2860 BLOCK_COMMENT("assert long copy {");
2861 Label L;
2862 __ cmpl(rax_elsize, LogBytesPerLong);
2863 __ jcc(Assembler::equal, L);
2864 __ stop("must be long copy, but elsize is wrong");
2865 __ bind(L);
2866 BLOCK_COMMENT("} assert long copy done");
2867 }
2868 #endif
2869 __ lea(from, Address(src, src_pos, Address::times_8, 0));// src_addr
2870 __ lea(to, Address(dst, dst_pos, Address::times_8, 0));// dst_addr
2871 __ movl2ptr(count, r11_length); // length
2872 __ jump(RuntimeAddress(long_copy_entry));
2873
2874 // ObjArrayKlass
2875 __ BIND(L_objArray);
2876 // live at this point: r10_src_klass, r11_length, src[_pos], dst[_pos]
2877
2878 Label L_plain_copy, L_checkcast_copy;
2879 // test array classes for subtyping
2880 __ load_klass(rax, dst);
2881 __ cmpq(r10_src_klass, rax); // usual case is exact equality
2882 __ jcc(Assembler::notEqual, L_checkcast_copy);
2883
2884 // Identically typed arrays can be copied without element-wise checks.
2885 arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length,
2886 r10, L_failed);
2887
2888 __ lea(from, Address(src, src_pos, TIMES_OOP,
2889 arrayOopDesc::base_offset_in_bytes(T_OBJECT))); // src_addr
2890 __ lea(to, Address(dst, dst_pos, TIMES_OOP,
2891 arrayOopDesc::base_offset_in_bytes(T_OBJECT))); // dst_addr
2892 __ movl2ptr(count, r11_length); // length
2893 __ BIND(L_plain_copy);
2894 __ jump(RuntimeAddress(oop_copy_entry));
2895
2896 __ BIND(L_checkcast_copy);
2897 // live at this point: r10_src_klass, r11_length, rax (dst_klass)
2898 {
2899 // Before looking at dst.length, make sure dst is also an objArray.
2900 __ cmpl(Address(rax, lh_offset), objArray_lh);
2901 __ jcc(Assembler::notEqual, L_failed);
2902
2903 // It is safe to examine both src.length and dst.length.
2904 arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length,
2905 rax, L_failed);
2906
2907 const Register r11_dst_klass = r11;
2908 __ load_klass(r11_dst_klass, dst); // reload
2909
2910 // Marshal the base address arguments now, freeing registers.
2911 __ lea(from, Address(src, src_pos, TIMES_OOP,
2912 arrayOopDesc::base_offset_in_bytes(T_OBJECT)));
2913 __ lea(to, Address(dst, dst_pos, TIMES_OOP,
2914 arrayOopDesc::base_offset_in_bytes(T_OBJECT)));
2915 __ movl(count, length); // length (reloaded)
2916 Register sco_temp = c_rarg3; // this register is free now
2917 assert_different_registers(from, to, count, sco_temp,
2918 r11_dst_klass, r10_src_klass);
2919 assert_clean_int(count, sco_temp);
2920
2921 // Generate the type check.
2922 const int sco_offset = in_bytes(Klass::super_check_offset_offset());
2923 __ movl(sco_temp, Address(r11_dst_klass, sco_offset));
2924 assert_clean_int(sco_temp, rax);
2925 generate_type_check(r10_src_klass, sco_temp, r11_dst_klass, L_plain_copy);
2926
2927 // Fetch destination element klass from the ObjArrayKlass header.
2928 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
2929 __ movptr(r11_dst_klass, Address(r11_dst_klass, ek_offset));
2930 __ movl( sco_temp, Address(r11_dst_klass, sco_offset));
2931 assert_clean_int(sco_temp, rax);
2932
2933 // the checkcast_copy loop needs two extra arguments:
2934 assert(c_rarg3 == sco_temp, "#3 already in place");
2935 // Set up arguments for checkcast_copy_entry.
2936 setup_arg_regs(4);
2937 __ movptr(r8, r11_dst_klass); // dst.klass.element_klass, r8 is c_rarg4 on Linux/Solaris
2938 __ jump(RuntimeAddress(checkcast_copy_entry));
2939 }
2940
2941 __ BIND(L_failed);
2942 __ xorptr(rax, rax);
2943 __ notptr(rax); // return -1
2944 __ leave(); // required for proper stackwalking of RuntimeStub frame
2945 __ ret(0);
2946
2947 return start;
2948 }
2949
2950 void generate_arraycopy_stubs() {
2951 address entry;
2952 address entry_jbyte_arraycopy;
2953 address entry_jshort_arraycopy;
2954 address entry_jint_arraycopy;
2955 address entry_oop_arraycopy;
2956 address entry_jlong_arraycopy;
2957 address entry_checkcast_arraycopy;
2958
2959 StubRoutines::_jbyte_disjoint_arraycopy = generate_disjoint_byte_copy(false, &entry,
2960 "jbyte_disjoint_arraycopy");
2961 StubRoutines::_jbyte_arraycopy = generate_conjoint_byte_copy(false, entry, &entry_jbyte_arraycopy,
2962 "jbyte_arraycopy");
2963
2964 StubRoutines::_jshort_disjoint_arraycopy = generate_disjoint_short_copy(false, &entry,
2965 "jshort_disjoint_arraycopy");
2966 StubRoutines::_jshort_arraycopy = generate_conjoint_short_copy(false, entry, &entry_jshort_arraycopy,
2967 "jshort_arraycopy");
2968
2969 StubRoutines::_jint_disjoint_arraycopy = generate_disjoint_int_oop_copy(false, false, &entry,
2970 "jint_disjoint_arraycopy");
2971 StubRoutines::_jint_arraycopy = generate_conjoint_int_oop_copy(false, false, entry,
2972 &entry_jint_arraycopy, "jint_arraycopy");
2973
2974 StubRoutines::_jlong_disjoint_arraycopy = generate_disjoint_long_oop_copy(false, false, &entry,
2975 "jlong_disjoint_arraycopy");
2976 StubRoutines::_jlong_arraycopy = generate_conjoint_long_oop_copy(false, false, entry,
2977 &entry_jlong_arraycopy, "jlong_arraycopy");
2978
2979
2980 if (UseCompressedOops) {
2981 StubRoutines::_oop_disjoint_arraycopy = generate_disjoint_int_oop_copy(false, true, &entry,
2982 "oop_disjoint_arraycopy");
2983 StubRoutines::_oop_arraycopy = generate_conjoint_int_oop_copy(false, true, entry,
2984 &entry_oop_arraycopy, "oop_arraycopy");
2985 StubRoutines::_oop_disjoint_arraycopy_uninit = generate_disjoint_int_oop_copy(false, true, &entry,
2986 "oop_disjoint_arraycopy_uninit",
2987 /*dest_uninitialized*/true);
2988 StubRoutines::_oop_arraycopy_uninit = generate_conjoint_int_oop_copy(false, true, entry,
2989 NULL, "oop_arraycopy_uninit",
2990 /*dest_uninitialized*/true);
2991 } else {
2992 StubRoutines::_oop_disjoint_arraycopy = generate_disjoint_long_oop_copy(false, true, &entry,
2993 "oop_disjoint_arraycopy");
2994 StubRoutines::_oop_arraycopy = generate_conjoint_long_oop_copy(false, true, entry,
2995 &entry_oop_arraycopy, "oop_arraycopy");
2996 StubRoutines::_oop_disjoint_arraycopy_uninit = generate_disjoint_long_oop_copy(false, true, &entry,
2997 "oop_disjoint_arraycopy_uninit",
2998 /*dest_uninitialized*/true);
2999 StubRoutines::_oop_arraycopy_uninit = generate_conjoint_long_oop_copy(false, true, entry,
3000 NULL, "oop_arraycopy_uninit",
3001 /*dest_uninitialized*/true);
3002 }
3003
3004 StubRoutines::_checkcast_arraycopy = generate_checkcast_copy("checkcast_arraycopy", &entry_checkcast_arraycopy);
3005 StubRoutines::_checkcast_arraycopy_uninit = generate_checkcast_copy("checkcast_arraycopy_uninit", NULL,
3006 /*dest_uninitialized*/true);
3007
3008 StubRoutines::_unsafe_arraycopy = generate_unsafe_copy("unsafe_arraycopy",
3009 entry_jbyte_arraycopy,
3010 entry_jshort_arraycopy,
3011 entry_jint_arraycopy,
3012 entry_jlong_arraycopy);
3013 StubRoutines::_generic_arraycopy = generate_generic_copy("generic_arraycopy",
3014 entry_jbyte_arraycopy,
3015 entry_jshort_arraycopy,
3016 entry_jint_arraycopy,
3017 entry_oop_arraycopy,
3018 entry_jlong_arraycopy,
3019 entry_checkcast_arraycopy);
3020
3021 StubRoutines::_jbyte_fill = generate_fill(T_BYTE, false, "jbyte_fill");
3022 StubRoutines::_jshort_fill = generate_fill(T_SHORT, false, "jshort_fill");
3023 StubRoutines::_jint_fill = generate_fill(T_INT, false, "jint_fill");
3024 StubRoutines::_arrayof_jbyte_fill = generate_fill(T_BYTE, true, "arrayof_jbyte_fill");
3025 StubRoutines::_arrayof_jshort_fill = generate_fill(T_SHORT, true, "arrayof_jshort_fill");
3026 StubRoutines::_arrayof_jint_fill = generate_fill(T_INT, true, "arrayof_jint_fill");
3027
3028 // We don't generate specialized code for HeapWord-aligned source
3029 // arrays, so just use the code we've already generated
3030 StubRoutines::_arrayof_jbyte_disjoint_arraycopy = StubRoutines::_jbyte_disjoint_arraycopy;
3031 StubRoutines::_arrayof_jbyte_arraycopy = StubRoutines::_jbyte_arraycopy;
3032
3033 StubRoutines::_arrayof_jshort_disjoint_arraycopy = StubRoutines::_jshort_disjoint_arraycopy;
3034 StubRoutines::_arrayof_jshort_arraycopy = StubRoutines::_jshort_arraycopy;
3035
3036 StubRoutines::_arrayof_jint_disjoint_arraycopy = StubRoutines::_jint_disjoint_arraycopy;
3037 StubRoutines::_arrayof_jint_arraycopy = StubRoutines::_jint_arraycopy;
3038
3039 StubRoutines::_arrayof_jlong_disjoint_arraycopy = StubRoutines::_jlong_disjoint_arraycopy;
3040 StubRoutines::_arrayof_jlong_arraycopy = StubRoutines::_jlong_arraycopy;
3041
3042 StubRoutines::_arrayof_oop_disjoint_arraycopy = StubRoutines::_oop_disjoint_arraycopy;
3043 StubRoutines::_arrayof_oop_arraycopy = StubRoutines::_oop_arraycopy;
3044
3045 StubRoutines::_arrayof_oop_disjoint_arraycopy_uninit = StubRoutines::_oop_disjoint_arraycopy_uninit;
3046 StubRoutines::_arrayof_oop_arraycopy_uninit = StubRoutines::_oop_arraycopy_uninit;
3047 }
3048
3049 void generate_math_stubs() {
3050 {
3051 StubCodeMark mark(this, "StubRoutines", "log");
3052 StubRoutines::_intrinsic_log = (double (*)(double)) __ pc();
3053
3054 __ subq(rsp, 8);
3055 __ movdbl(Address(rsp, 0), xmm0);
3056 __ fld_d(Address(rsp, 0));
3057 __ flog();
3058 __ fstp_d(Address(rsp, 0));
3059 __ movdbl(xmm0, Address(rsp, 0));
3060 __ addq(rsp, 8);
3061 __ ret(0);
3062 }
3063 {
3064 StubCodeMark mark(this, "StubRoutines", "log10");
3065 StubRoutines::_intrinsic_log10 = (double (*)(double)) __ pc();
3066
3067 __ subq(rsp, 8);
3068 __ movdbl(Address(rsp, 0), xmm0);
3069 __ fld_d(Address(rsp, 0));
3070 __ flog10();
3071 __ fstp_d(Address(rsp, 0));
3072 __ movdbl(xmm0, Address(rsp, 0));
3073 __ addq(rsp, 8);
3074 __ ret(0);
3075 }
3076 {
3077 StubCodeMark mark(this, "StubRoutines", "sin");
3078 StubRoutines::_intrinsic_sin = (double (*)(double)) __ pc();
3079
3080 __ subq(rsp, 8);
3081 __ movdbl(Address(rsp, 0), xmm0);
3082 __ fld_d(Address(rsp, 0));
3083 __ trigfunc('s');
3084 __ fstp_d(Address(rsp, 0));
3085 __ movdbl(xmm0, Address(rsp, 0));
3086 __ addq(rsp, 8);
3087 __ ret(0);
3088 }
3089 {
3090 StubCodeMark mark(this, "StubRoutines", "cos");
3091 StubRoutines::_intrinsic_cos = (double (*)(double)) __ pc();
3092
3093 __ subq(rsp, 8);
3094 __ movdbl(Address(rsp, 0), xmm0);
3095 __ fld_d(Address(rsp, 0));
3096 __ trigfunc('c');
3097 __ fstp_d(Address(rsp, 0));
3098 __ movdbl(xmm0, Address(rsp, 0));
3099 __ addq(rsp, 8);
3100 __ ret(0);
3101 }
3102 {
3103 StubCodeMark mark(this, "StubRoutines", "tan");
3104 StubRoutines::_intrinsic_tan = (double (*)(double)) __ pc();
3105
3106 __ subq(rsp, 8);
3107 __ movdbl(Address(rsp, 0), xmm0);
3108 __ fld_d(Address(rsp, 0));
3109 __ trigfunc('t');
3110 __ fstp_d(Address(rsp, 0));
3111 __ movdbl(xmm0, Address(rsp, 0));
3112 __ addq(rsp, 8);
3113 __ ret(0);
3114 }
3115 {
3116 StubCodeMark mark(this, "StubRoutines", "exp");
3117 StubRoutines::_intrinsic_exp = (double (*)(double)) __ pc();
3118
3119 __ subq(rsp, 8);
3120 __ movdbl(Address(rsp, 0), xmm0);
3121 __ fld_d(Address(rsp, 0));
3122 __ exp_with_fallback(0);
3123 __ fstp_d(Address(rsp, 0));
3124 __ movdbl(xmm0, Address(rsp, 0));
3125 __ addq(rsp, 8);
3126 __ ret(0);
3127 }
3128 {
3129 StubCodeMark mark(this, "StubRoutines", "pow");
3130 StubRoutines::_intrinsic_pow = (double (*)(double,double)) __ pc();
3131
3132 __ subq(rsp, 8);
3133 __ movdbl(Address(rsp, 0), xmm1);
3134 __ fld_d(Address(rsp, 0));
3135 __ movdbl(Address(rsp, 0), xmm0);
3136 __ fld_d(Address(rsp, 0));
3137 __ pow_with_fallback(0);
3138 __ fstp_d(Address(rsp, 0));
3139 __ movdbl(xmm0, Address(rsp, 0));
3140 __ addq(rsp, 8);
3141 __ ret(0);
3142 }
3143 }
3144
3145 // AES intrinsic stubs
3146 enum {AESBlockSize = 16};
3147
3148 address generate_key_shuffle_mask() {
3149 __ align(16);
3150 StubCodeMark mark(this, "StubRoutines", "key_shuffle_mask");
3151 address start = __ pc();
3152 __ emit_data64( 0x0405060700010203, relocInfo::none );
3153 __ emit_data64( 0x0c0d0e0f08090a0b, relocInfo::none );
3154 return start;
3155 }
3156
3157 // Utility routine for loading a 128-bit key word in little endian format
3158 // can optionally specify that the shuffle mask is already in an xmmregister
3159 void load_key(XMMRegister xmmdst, Register key, int offset, XMMRegister xmm_shuf_mask=NULL) {
3160 __ movdqu(xmmdst, Address(key, offset));
3161 if (xmm_shuf_mask != NULL) {
3162 __ pshufb(xmmdst, xmm_shuf_mask);
3163 } else {
3164 __ pshufb(xmmdst, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3165 }
3166 }
3167
3168 // Arguments:
3169 //
3170 // Inputs:
3171 // c_rarg0 - source byte array address
3172 // c_rarg1 - destination byte array address
3173 // c_rarg2 - K (key) in little endian int array
3174 //
3175 address generate_aescrypt_encryptBlock() {
3176 assert(UseAES, "need AES instructions and misaligned SSE support");
3177 __ align(CodeEntryAlignment);
3178 StubCodeMark mark(this, "StubRoutines", "aescrypt_encryptBlock");
3179 Label L_doLast;
3180 address start = __ pc();
3181
3182 const Register from = c_rarg0; // source array address
3183 const Register to = c_rarg1; // destination array address
3184 const Register key = c_rarg2; // key array address
3185 const Register keylen = rax;
3186
3187 const XMMRegister xmm_result = xmm0;
3188 const XMMRegister xmm_key_shuf_mask = xmm1;
3189 // On win64 xmm6-xmm15 must be preserved so don't use them.
3190 const XMMRegister xmm_temp1 = xmm2;
3191 const XMMRegister xmm_temp2 = xmm3;
3192 const XMMRegister xmm_temp3 = xmm4;
3193 const XMMRegister xmm_temp4 = xmm5;
3194
3195 __ enter(); // required for proper stackwalking of RuntimeStub frame
3196
3197 // keylen could be only {11, 13, 15} * 4 = {44, 52, 60}
3198 __ movl(keylen, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3199
3200 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3201 __ movdqu(xmm_result, Address(from, 0)); // get 16 bytes of input
3202
3203 // For encryption, the java expanded key ordering is just what we need
3204 // we don't know if the key is aligned, hence not using load-execute form
3205
3206 load_key(xmm_temp1, key, 0x00, xmm_key_shuf_mask);
3207 __ pxor(xmm_result, xmm_temp1);
3208
3209 load_key(xmm_temp1, key, 0x10, xmm_key_shuf_mask);
3210 load_key(xmm_temp2, key, 0x20, xmm_key_shuf_mask);
3211 load_key(xmm_temp3, key, 0x30, xmm_key_shuf_mask);
3212 load_key(xmm_temp4, key, 0x40, xmm_key_shuf_mask);
3213
3214 __ aesenc(xmm_result, xmm_temp1);
3215 __ aesenc(xmm_result, xmm_temp2);
3216 __ aesenc(xmm_result, xmm_temp3);
3217 __ aesenc(xmm_result, xmm_temp4);
3218
3219 load_key(xmm_temp1, key, 0x50, xmm_key_shuf_mask);
3220 load_key(xmm_temp2, key, 0x60, xmm_key_shuf_mask);
3221 load_key(xmm_temp3, key, 0x70, xmm_key_shuf_mask);
3222 load_key(xmm_temp4, key, 0x80, xmm_key_shuf_mask);
3223
3224 __ aesenc(xmm_result, xmm_temp1);
3225 __ aesenc(xmm_result, xmm_temp2);
3226 __ aesenc(xmm_result, xmm_temp3);
3227 __ aesenc(xmm_result, xmm_temp4);
3228
3229 load_key(xmm_temp1, key, 0x90, xmm_key_shuf_mask);
3230 load_key(xmm_temp2, key, 0xa0, xmm_key_shuf_mask);
3231
3232 __ cmpl(keylen, 44);
3233 __ jccb(Assembler::equal, L_doLast);
3234
3235 __ aesenc(xmm_result, xmm_temp1);
3236 __ aesenc(xmm_result, xmm_temp2);
3237
3238 load_key(xmm_temp1, key, 0xb0, xmm_key_shuf_mask);
3239 load_key(xmm_temp2, key, 0xc0, xmm_key_shuf_mask);
3240
3241 __ cmpl(keylen, 52);
3242 __ jccb(Assembler::equal, L_doLast);
3243
3244 __ aesenc(xmm_result, xmm_temp1);
3245 __ aesenc(xmm_result, xmm_temp2);
3246
3247 load_key(xmm_temp1, key, 0xd0, xmm_key_shuf_mask);
3248 load_key(xmm_temp2, key, 0xe0, xmm_key_shuf_mask);
3249
3250 __ BIND(L_doLast);
3251 __ aesenc(xmm_result, xmm_temp1);
3252 __ aesenclast(xmm_result, xmm_temp2);
3253 __ movdqu(Address(to, 0), xmm_result); // store the result
3254 __ xorptr(rax, rax); // return 0
3255 __ leave(); // required for proper stackwalking of RuntimeStub frame
3256 __ ret(0);
3257
3258 return start;
3259 }
3260
3261
3262 // Arguments:
3263 //
3264 // Inputs:
3265 // c_rarg0 - source byte array address
3266 // c_rarg1 - destination byte array address
3267 // c_rarg2 - K (key) in little endian int array
3268 //
3269 address generate_aescrypt_decryptBlock() {
3270 assert(UseAES, "need AES instructions and misaligned SSE support");
3271 __ align(CodeEntryAlignment);
3272 StubCodeMark mark(this, "StubRoutines", "aescrypt_decryptBlock");
3273 Label L_doLast;
3274 address start = __ pc();
3275
3276 const Register from = c_rarg0; // source array address
3277 const Register to = c_rarg1; // destination array address
3278 const Register key = c_rarg2; // key array address
3279 const Register keylen = rax;
3280
3281 const XMMRegister xmm_result = xmm0;
3282 const XMMRegister xmm_key_shuf_mask = xmm1;
3283 // On win64 xmm6-xmm15 must be preserved so don't use them.
3284 const XMMRegister xmm_temp1 = xmm2;
3285 const XMMRegister xmm_temp2 = xmm3;
3286 const XMMRegister xmm_temp3 = xmm4;
3287 const XMMRegister xmm_temp4 = xmm5;
3288
3289 __ enter(); // required for proper stackwalking of RuntimeStub frame
3290
3291 // keylen could be only {11, 13, 15} * 4 = {44, 52, 60}
3292 __ movl(keylen, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3293
3294 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3295 __ movdqu(xmm_result, Address(from, 0));
3296
3297 // for decryption java expanded key ordering is rotated one position from what we want
3298 // so we start from 0x10 here and hit 0x00 last
3299 // we don't know if the key is aligned, hence not using load-execute form
3300 load_key(xmm_temp1, key, 0x10, xmm_key_shuf_mask);
3301 load_key(xmm_temp2, key, 0x20, xmm_key_shuf_mask);
3302 load_key(xmm_temp3, key, 0x30, xmm_key_shuf_mask);
3303 load_key(xmm_temp4, key, 0x40, xmm_key_shuf_mask);
3304
3305 __ pxor (xmm_result, xmm_temp1);
3306 __ aesdec(xmm_result, xmm_temp2);
3307 __ aesdec(xmm_result, xmm_temp3);
3308 __ aesdec(xmm_result, xmm_temp4);
3309
3310 load_key(xmm_temp1, key, 0x50, xmm_key_shuf_mask);
3311 load_key(xmm_temp2, key, 0x60, xmm_key_shuf_mask);
3312 load_key(xmm_temp3, key, 0x70, xmm_key_shuf_mask);
3313 load_key(xmm_temp4, key, 0x80, xmm_key_shuf_mask);
3314
3315 __ aesdec(xmm_result, xmm_temp1);
3316 __ aesdec(xmm_result, xmm_temp2);
3317 __ aesdec(xmm_result, xmm_temp3);
3318 __ aesdec(xmm_result, xmm_temp4);
3319
3320 load_key(xmm_temp1, key, 0x90, xmm_key_shuf_mask);
3321 load_key(xmm_temp2, key, 0xa0, xmm_key_shuf_mask);
3322 load_key(xmm_temp3, key, 0x00, xmm_key_shuf_mask);
3323
3324 __ cmpl(keylen, 44);
3325 __ jccb(Assembler::equal, L_doLast);
3326
3327 __ aesdec(xmm_result, xmm_temp1);
3328 __ aesdec(xmm_result, xmm_temp2);
3329
3330 load_key(xmm_temp1, key, 0xb0, xmm_key_shuf_mask);
3331 load_key(xmm_temp2, key, 0xc0, xmm_key_shuf_mask);
3332
3333 __ cmpl(keylen, 52);
3334 __ jccb(Assembler::equal, L_doLast);
3335
3336 __ aesdec(xmm_result, xmm_temp1);
3337 __ aesdec(xmm_result, xmm_temp2);
3338
3339 load_key(xmm_temp1, key, 0xd0, xmm_key_shuf_mask);
3340 load_key(xmm_temp2, key, 0xe0, xmm_key_shuf_mask);
3341
3342 __ BIND(L_doLast);
3343 __ aesdec(xmm_result, xmm_temp1);
3344 __ aesdec(xmm_result, xmm_temp2);
3345
3346 // for decryption the aesdeclast operation is always on key+0x00
3347 __ aesdeclast(xmm_result, xmm_temp3);
3348 __ movdqu(Address(to, 0), xmm_result); // store the result
3349 __ xorptr(rax, rax); // return 0
3350 __ leave(); // required for proper stackwalking of RuntimeStub frame
3351 __ ret(0);
3352
3353 return start;
3354 }
3355
3356
3357 // Arguments:
3358 //
3359 // Inputs:
3360 // c_rarg0 - source byte array address
3361 // c_rarg1 - destination byte array address
3362 // c_rarg2 - K (key) in little endian int array
3363 // c_rarg3 - r vector byte array address
3364 // c_rarg4 - input length
3365 //
3366 // Output:
3367 // rax - input length
3368 //
3369 address generate_cipherBlockChaining_encryptAESCrypt() {
3370 assert(UseAES, "need AES instructions and misaligned SSE support");
3371 __ align(CodeEntryAlignment);
3372 StubCodeMark mark(this, "StubRoutines", "cipherBlockChaining_encryptAESCrypt");
3373 address start = __ pc();
3374
3375 Label L_exit, L_key_192_256, L_key_256, L_loopTop_128, L_loopTop_192, L_loopTop_256;
3376 const Register from = c_rarg0; // source array address
3377 const Register to = c_rarg1; // destination array address
3378 const Register key = c_rarg2; // key array address
3379 const Register rvec = c_rarg3; // r byte array initialized from initvector array address
3380 // and left with the results of the last encryption block
3381 #ifndef _WIN64
3382 const Register len_reg = c_rarg4; // src len (must be multiple of blocksize 16)
3383 #else
3384 const Address len_mem(rbp, 6 * wordSize); // length is on stack on Win64
3385 const Register len_reg = r10; // pick the first volatile windows register
3386 #endif
3387 const Register pos = rax;
3388
3389 // xmm register assignments for the loops below
3390 const XMMRegister xmm_result = xmm0;
3391 const XMMRegister xmm_temp = xmm1;
3392 // keys 0-10 preloaded into xmm2-xmm12
3393 const int XMM_REG_NUM_KEY_FIRST = 2;
3394 const int XMM_REG_NUM_KEY_LAST = 15;
3395 const XMMRegister xmm_key0 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST);
3396 const XMMRegister xmm_key10 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+10);
3397 const XMMRegister xmm_key11 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+11);
3398 const XMMRegister xmm_key12 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+12);
3399 const XMMRegister xmm_key13 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+13);
3400
3401 __ enter(); // required for proper stackwalking of RuntimeStub frame
3402
3403 #ifdef _WIN64
3404 // on win64, fill len_reg from stack position
3405 __ movl(len_reg, len_mem);
3406 // save the xmm registers which must be preserved 6-15
3407 __ subptr(rsp, -rsp_after_call_off * wordSize);
3408 for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3409 __ movdqu(xmm_save(i), as_XMMRegister(i));
3410 }
3411 #else
3412 __ push(len_reg); // Save
3413 #endif
3414
3415 const XMMRegister xmm_key_shuf_mask = xmm_temp; // used temporarily to swap key bytes up front
3416 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3417 // load up xmm regs xmm2 thru xmm12 with key 0x00 - 0xa0
3418 for (int rnum = XMM_REG_NUM_KEY_FIRST, offset = 0x00; rnum <= XMM_REG_NUM_KEY_FIRST+10; rnum++) {
3419 load_key(as_XMMRegister(rnum), key, offset, xmm_key_shuf_mask);
3420 offset += 0x10;
3421 }
3422 __ movdqu(xmm_result, Address(rvec, 0x00)); // initialize xmm_result with r vec
3423
3424 // now split to different paths depending on the keylen (len in ints of AESCrypt.KLE array (52=192, or 60=256))
3425 __ movl(rax, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3426 __ cmpl(rax, 44);
3427 __ jcc(Assembler::notEqual, L_key_192_256);
3428
3429 // 128 bit code follows here
3430 __ movptr(pos, 0);
3431 __ align(OptoLoopAlignment);
3432
3433 __ BIND(L_loopTop_128);
3434 __ movdqu(xmm_temp, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of input
3435 __ pxor (xmm_result, xmm_temp); // xor with the current r vector
3436 __ pxor (xmm_result, xmm_key0); // do the aes rounds
3437 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_FIRST + 9; rnum++) {
3438 __ aesenc(xmm_result, as_XMMRegister(rnum));
3439 }
3440 __ aesenclast(xmm_result, xmm_key10);
3441 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3442 // no need to store r to memory until we exit
3443 __ addptr(pos, AESBlockSize);
3444 __ subptr(len_reg, AESBlockSize);
3445 __ jcc(Assembler::notEqual, L_loopTop_128);
3446
3447 __ BIND(L_exit);
3448 __ movdqu(Address(rvec, 0), xmm_result); // final value of r stored in rvec of CipherBlockChaining object
3449
3450 #ifdef _WIN64
3451 // restore xmm regs belonging to calling function
3452 for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3453 __ movdqu(as_XMMRegister(i), xmm_save(i));
3454 }
3455 __ movl(rax, len_mem);
3456 #else
3457 __ pop(rax); // return length
3458 #endif
3459 __ leave(); // required for proper stackwalking of RuntimeStub frame
3460 __ ret(0);
3461
3462 __ BIND(L_key_192_256);
3463 // here rax = len in ints of AESCrypt.KLE array (52=192, or 60=256)
3464 load_key(xmm_key11, key, 0xb0, xmm_key_shuf_mask);
3465 load_key(xmm_key12, key, 0xc0, xmm_key_shuf_mask);
3466 __ cmpl(rax, 52);
3467 __ jcc(Assembler::notEqual, L_key_256);
3468
3469 // 192-bit code follows here (could be changed to use more xmm registers)
3470 __ movptr(pos, 0);
3471 __ align(OptoLoopAlignment);
3472
3473 __ BIND(L_loopTop_192);
3474 __ movdqu(xmm_temp, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of input
3475 __ pxor (xmm_result, xmm_temp); // xor with the current r vector
3476 __ pxor (xmm_result, xmm_key0); // do the aes rounds
3477 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_FIRST + 11; rnum++) {
3478 __ aesenc(xmm_result, as_XMMRegister(rnum));
3479 }
3480 __ aesenclast(xmm_result, xmm_key12);
3481 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3482 // no need to store r to memory until we exit
3483 __ addptr(pos, AESBlockSize);
3484 __ subptr(len_reg, AESBlockSize);
3485 __ jcc(Assembler::notEqual, L_loopTop_192);
3486 __ jmp(L_exit);
3487
3488 __ BIND(L_key_256);
3489 // 256-bit code follows here (could be changed to use more xmm registers)
3490 load_key(xmm_key13, key, 0xd0, xmm_key_shuf_mask);
3491 __ movptr(pos, 0);
3492 __ align(OptoLoopAlignment);
3493
3494 __ BIND(L_loopTop_256);
3495 __ movdqu(xmm_temp, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of input
3496 __ pxor (xmm_result, xmm_temp); // xor with the current r vector
3497 __ pxor (xmm_result, xmm_key0); // do the aes rounds
3498 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_FIRST + 13; rnum++) {
3499 __ aesenc(xmm_result, as_XMMRegister(rnum));
3500 }
3501 load_key(xmm_temp, key, 0xe0);
3502 __ aesenclast(xmm_result, xmm_temp);
3503 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3504 // no need to store r to memory until we exit
3505 __ addptr(pos, AESBlockSize);
3506 __ subptr(len_reg, AESBlockSize);
3507 __ jcc(Assembler::notEqual, L_loopTop_256);
3508 __ jmp(L_exit);
3509
3510 return start;
3511 }
3512
3513 // Safefetch stubs.
3514 void generate_safefetch(const char* name, int size, address* entry,
3515 address* fault_pc, address* continuation_pc) {
3516 // safefetch signatures:
3517 // int SafeFetch32(int* adr, int errValue);
3518 // intptr_t SafeFetchN (intptr_t* adr, intptr_t errValue);
3519 //
3520 // arguments:
3521 // c_rarg0 = adr
3522 // c_rarg1 = errValue
3523 //
3524 // result:
3525 // PPC_RET = *adr or errValue
3526
3527 StubCodeMark mark(this, "StubRoutines", name);
3528
3529 // Entry point, pc or function descriptor.
3530 *entry = __ pc();
3531
3532 // Load *adr into c_rarg1, may fault.
3533 *fault_pc = __ pc();
3534 switch (size) {
3535 case 4:
3536 // int32_t
3537 __ movl(c_rarg1, Address(c_rarg0, 0));
3538 break;
3539 case 8:
3540 // int64_t
3541 __ movq(c_rarg1, Address(c_rarg0, 0));
3542 break;
3543 default:
3544 ShouldNotReachHere();
3545 }
3546
3547 // return errValue or *adr
3548 *continuation_pc = __ pc();
3549 __ movq(rax, c_rarg1);
3550 __ ret(0);
3551 }
3552
3553 // This is a version of CBC/AES Decrypt which does 4 blocks in a loop at a time
3554 // to hide instruction latency
3555 //
3556 // Arguments:
3557 //
3558 // Inputs:
3559 // c_rarg0 - source byte array address
3560 // c_rarg1 - destination byte array address
3561 // c_rarg2 - K (key) in little endian int array
3562 // c_rarg3 - r vector byte array address
3563 // c_rarg4 - input length
3564 //
3565 // Output:
3566 // rax - input length
3567 //
3568
3569 address generate_cipherBlockChaining_decryptAESCrypt_Parallel() {
3570 assert(UseAES, "need AES instructions and misaligned SSE support");
3571 __ align(CodeEntryAlignment);
3572 StubCodeMark mark(this, "StubRoutines", "cipherBlockChaining_decryptAESCrypt");
3573 address start = __ pc();
3574
3575 Label L_exit, L_key_192_256, L_key_256;
3576 Label L_singleBlock_loopTop_128, L_multiBlock_loopTop_128;
3577 Label L_singleBlock_loopTop_192, L_singleBlock_loopTop_256;
3578 const Register from = c_rarg0; // source array address
3579 const Register to = c_rarg1; // destination array address
3580 const Register key = c_rarg2; // key array address
3581 const Register rvec = c_rarg3; // r byte array initialized from initvector array address
3582 // and left with the results of the last encryption block
3583 #ifndef _WIN64
3584 const Register len_reg = c_rarg4; // src len (must be multiple of blocksize 16)
3585 #else
3586 const Address len_mem(rbp, 6 * wordSize); // length is on stack on Win64
3587 const Register len_reg = r10; // pick the first volatile windows register
3588 #endif
3589 const Register pos = rax;
3590
3591 // keys 0-10 preloaded into xmm2-xmm12
3592 const int XMM_REG_NUM_KEY_FIRST = 5;
3593 const int XMM_REG_NUM_KEY_LAST = 15;
3594 const XMMRegister xmm_key_first = as_XMMRegister(XMM_REG_NUM_KEY_FIRST);
3595 const XMMRegister xmm_key_last = as_XMMRegister(XMM_REG_NUM_KEY_LAST);
3596
3597 __ enter(); // required for proper stackwalking of RuntimeStub frame
3598
3599 #ifdef _WIN64
3600 // on win64, fill len_reg from stack position
3601 __ movl(len_reg, len_mem);
3602 // save the xmm registers which must be preserved 6-15
3603 __ subptr(rsp, -rsp_after_call_off * wordSize);
3604 for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3605 __ movdqu(xmm_save(i), as_XMMRegister(i));
3606 }
3607 #else
3608 __ push(len_reg); // Save
3609 #endif
3610
3611 // the java expanded key ordering is rotated one position from what we want
3612 // so we start from 0x10 here and hit 0x00 last
3613 const XMMRegister xmm_key_shuf_mask = xmm1; // used temporarily to swap key bytes up front
3614 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3615 // load up xmm regs 5 thru 15 with key 0x10 - 0xa0 - 0x00
3616 for (int rnum = XMM_REG_NUM_KEY_FIRST, offset = 0x10; rnum < XMM_REG_NUM_KEY_LAST; rnum++) {
3617 load_key(as_XMMRegister(rnum), key, offset, xmm_key_shuf_mask);
3618 offset += 0x10;
3619 }
3620 load_key(xmm_key_last, key, 0x00, xmm_key_shuf_mask);
3621
3622 const XMMRegister xmm_prev_block_cipher = xmm1; // holds cipher of previous block
3623
3624 // registers holding the four results in the parallelized loop
3625 const XMMRegister xmm_result0 = xmm0;
3626 const XMMRegister xmm_result1 = xmm2;
3627 const XMMRegister xmm_result2 = xmm3;
3628 const XMMRegister xmm_result3 = xmm4;
3629
3630 __ movdqu(xmm_prev_block_cipher, Address(rvec, 0x00)); // initialize with initial rvec
3631
3632 // now split to different paths depending on the keylen (len in ints of AESCrypt.KLE array (52=192, or 60=256))
3633 __ movl(rax, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3634 __ cmpl(rax, 44);
3635 __ jcc(Assembler::notEqual, L_key_192_256);
3636
3637
3638 // 128-bit code follows here, parallelized
3639 __ movptr(pos, 0);
3640 __ align(OptoLoopAlignment);
3641 __ BIND(L_multiBlock_loopTop_128);
3642 __ cmpptr(len_reg, 4*AESBlockSize); // see if at least 4 blocks left
3643 __ jcc(Assembler::less, L_singleBlock_loopTop_128);
3644
3645 __ movdqu(xmm_result0, Address(from, pos, Address::times_1, 0*AESBlockSize)); // get next 4 blocks into xmmresult registers
3646 __ movdqu(xmm_result1, Address(from, pos, Address::times_1, 1*AESBlockSize));
3647 __ movdqu(xmm_result2, Address(from, pos, Address::times_1, 2*AESBlockSize));
3648 __ movdqu(xmm_result3, Address(from, pos, Address::times_1, 3*AESBlockSize));
3649
3650 #define DoFour(opc, src_reg) \
3651 __ opc(xmm_result0, src_reg); \
3652 __ opc(xmm_result1, src_reg); \
3653 __ opc(xmm_result2, src_reg); \
3654 __ opc(xmm_result3, src_reg);
3655
3656 DoFour(pxor, xmm_key_first);
3657 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3658 DoFour(aesdec, as_XMMRegister(rnum));
3659 }
3660 DoFour(aesdeclast, xmm_key_last);
3661 // for each result, xor with the r vector of previous cipher block
3662 __ pxor(xmm_result0, xmm_prev_block_cipher);
3663 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 0*AESBlockSize));
3664 __ pxor(xmm_result1, xmm_prev_block_cipher);
3665 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 1*AESBlockSize));
3666 __ pxor(xmm_result2, xmm_prev_block_cipher);
3667 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 2*AESBlockSize));
3668 __ pxor(xmm_result3, xmm_prev_block_cipher);
3669 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 3*AESBlockSize)); // this will carry over to next set of blocks
3670
3671 __ movdqu(Address(to, pos, Address::times_1, 0*AESBlockSize), xmm_result0); // store 4 results into the next 64 bytes of output
3672 __ movdqu(Address(to, pos, Address::times_1, 1*AESBlockSize), xmm_result1);
3673 __ movdqu(Address(to, pos, Address::times_1, 2*AESBlockSize), xmm_result2);
3674 __ movdqu(Address(to, pos, Address::times_1, 3*AESBlockSize), xmm_result3);
3675
3676 __ addptr(pos, 4*AESBlockSize);
3677 __ subptr(len_reg, 4*AESBlockSize);
3678 __ jmp(L_multiBlock_loopTop_128);
3679
3680 // registers used in the non-parallelized loops
3681 // xmm register assignments for the loops below
3682 const XMMRegister xmm_result = xmm0;
3683 const XMMRegister xmm_prev_block_cipher_save = xmm2;
3684 const XMMRegister xmm_key11 = xmm3;
3685 const XMMRegister xmm_key12 = xmm4;
3686 const XMMRegister xmm_temp = xmm4;
3687
3688 __ align(OptoLoopAlignment);
3689 __ BIND(L_singleBlock_loopTop_128);
3690 __ cmpptr(len_reg, 0); // any blocks left??
3691 __ jcc(Assembler::equal, L_exit);
3692 __ movdqu(xmm_result, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of cipher input
3693 __ movdqa(xmm_prev_block_cipher_save, xmm_result); // save for next r vector
3694 __ pxor (xmm_result, xmm_key_first); // do the aes dec rounds
3695 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3696 __ aesdec(xmm_result, as_XMMRegister(rnum));
3697 }
3698 __ aesdeclast(xmm_result, xmm_key_last);
3699 __ pxor (xmm_result, xmm_prev_block_cipher); // xor with the current r vector
3700 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3701 // no need to store r to memory until we exit
3702 __ movdqa(xmm_prev_block_cipher, xmm_prev_block_cipher_save); // set up next r vector with cipher input from this block
3703
3704 __ addptr(pos, AESBlockSize);
3705 __ subptr(len_reg, AESBlockSize);
3706 __ jmp(L_singleBlock_loopTop_128);
3707
3708
3709 __ BIND(L_exit);
3710 __ movdqu(Address(rvec, 0), xmm_prev_block_cipher); // final value of r stored in rvec of CipherBlockChaining object
3711 #ifdef _WIN64
3712 // restore regs belonging to calling function
3713 for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3714 __ movdqu(as_XMMRegister(i), xmm_save(i));
3715 }
3716 __ movl(rax, len_mem);
3717 #else
3718 __ pop(rax); // return length
3719 #endif
3720 __ leave(); // required for proper stackwalking of RuntimeStub frame
3721 __ ret(0);
3722
3723
3724 __ BIND(L_key_192_256);
3725 // here rax = len in ints of AESCrypt.KLE array (52=192, or 60=256)
3726 load_key(xmm_key11, key, 0xb0);
3727 __ cmpl(rax, 52);
3728 __ jcc(Assembler::notEqual, L_key_256);
3729
3730 // 192-bit code follows here (could be optimized to use parallelism)
3731 load_key(xmm_key12, key, 0xc0); // 192-bit key goes up to c0
3732 __ movptr(pos, 0);
3733 __ align(OptoLoopAlignment);
3734
3735 __ BIND(L_singleBlock_loopTop_192);
3736 __ movdqu(xmm_result, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of cipher input
3737 __ movdqa(xmm_prev_block_cipher_save, xmm_result); // save for next r vector
3738 __ pxor (xmm_result, xmm_key_first); // do the aes dec rounds
3739 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3740 __ aesdec(xmm_result, as_XMMRegister(rnum));
3741 }
3742 __ aesdec(xmm_result, xmm_key11);
3743 __ aesdec(xmm_result, xmm_key12);
3744 __ aesdeclast(xmm_result, xmm_key_last); // xmm15 always came from key+0
3745 __ pxor (xmm_result, xmm_prev_block_cipher); // xor with the current r vector
3746 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3747 // no need to store r to memory until we exit
3748 __ movdqa(xmm_prev_block_cipher, xmm_prev_block_cipher_save); // set up next r vector with cipher input from this block
3749 __ addptr(pos, AESBlockSize);
3750 __ subptr(len_reg, AESBlockSize);
3751 __ jcc(Assembler::notEqual,L_singleBlock_loopTop_192);
3752 __ jmp(L_exit);
3753
3754 __ BIND(L_key_256);
3755 // 256-bit code follows here (could be optimized to use parallelism)
3756 __ movptr(pos, 0);
3757 __ align(OptoLoopAlignment);
3758
3759 __ BIND(L_singleBlock_loopTop_256);
3760 __ movdqu(xmm_result, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of cipher input
3761 __ movdqa(xmm_prev_block_cipher_save, xmm_result); // save for next r vector
3762 __ pxor (xmm_result, xmm_key_first); // do the aes dec rounds
3763 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3764 __ aesdec(xmm_result, as_XMMRegister(rnum));
3765 }
3766 __ aesdec(xmm_result, xmm_key11);
3767 load_key(xmm_temp, key, 0xc0);
3768 __ aesdec(xmm_result, xmm_temp);
3769 load_key(xmm_temp, key, 0xd0);
3770 __ aesdec(xmm_result, xmm_temp);
3771 load_key(xmm_temp, key, 0xe0); // 256-bit key goes up to e0
3772 __ aesdec(xmm_result, xmm_temp);
3773 __ aesdeclast(xmm_result, xmm_key_last); // xmm15 came from key+0
3774 __ pxor (xmm_result, xmm_prev_block_cipher); // xor with the current r vector
3775 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3776 // no need to store r to memory until we exit
3777 __ movdqa(xmm_prev_block_cipher, xmm_prev_block_cipher_save); // set up next r vector with cipher input from this block
3778 __ addptr(pos, AESBlockSize);
3779 __ subptr(len_reg, AESBlockSize);
3780 __ jcc(Assembler::notEqual,L_singleBlock_loopTop_256);
3781 __ jmp(L_exit);
3782
3783 return start;
3784 }
3785
3786
3787 // byte swap x86 long
3788 address generate_ghash_long_swap_mask() {
3789 __ align(CodeEntryAlignment);
3790 StubCodeMark mark(this, "StubRoutines", "ghash_long_swap_mask");
3791 address start = __ pc();
3792 __ emit_data64(0x0f0e0d0c0b0a0908, relocInfo::none );
3793 __ emit_data64(0x0706050403020100, relocInfo::none );
3794 return start;
3795 }
3796
3797 // byte swap x86 byte array
3798 address generate_ghash_byte_swap_mask() {
3799 __ align(CodeEntryAlignment);
3800 StubCodeMark mark(this, "StubRoutines", "ghash_byte_swap_mask");
3801 address start = __ pc();
3802 __ emit_data64(0x08090a0b0c0d0e0f, relocInfo::none );
3803 __ emit_data64(0x0001020304050607, relocInfo::none );
3804 return start;
3805 }
3806
3807 /* Single and multi-block ghash operations */
3808 address generate_ghash_processBlocks() {
3809 __ align(CodeEntryAlignment);
3810 Label L_ghash_loop, L_exit;
3811 StubCodeMark mark(this, "StubRoutines", "ghash_processBlocks");
3812 address start = __ pc();
3813
3814 const Register state = c_rarg0;
3815 const Register subkeyH = c_rarg1;
3816 const Register data = c_rarg2;
3817 const Register blocks = c_rarg3;
3818
3819 #ifdef _WIN64
3820 const int XMM_REG_LAST = 10;
3821 #endif
3822
3823 const XMMRegister xmm_temp0 = xmm0;
3824 const XMMRegister xmm_temp1 = xmm1;
3825 const XMMRegister xmm_temp2 = xmm2;
3826 const XMMRegister xmm_temp3 = xmm3;
3827 const XMMRegister xmm_temp4 = xmm4;
3828 const XMMRegister xmm_temp5 = xmm5;
3829 const XMMRegister xmm_temp6 = xmm6;
3830 const XMMRegister xmm_temp7 = xmm7;
3831 const XMMRegister xmm_temp8 = xmm8;
3832 const XMMRegister xmm_temp9 = xmm9;
3833 const XMMRegister xmm_temp10 = xmm10;
3834
3835 __ enter();
3836
3837 #ifdef _WIN64
3838 // save the xmm registers which must be preserved 6-10
3839 __ subptr(rsp, -rsp_after_call_off * wordSize);
3840 for (int i = 6; i <= XMM_REG_LAST; i++) {
3841 __ movdqu(xmm_save(i), as_XMMRegister(i));
3842 }
3843 #endif
3844
3845 __ movdqu(xmm_temp10, ExternalAddress(StubRoutines::x86::ghash_long_swap_mask_addr()));
3846
3847 __ movdqu(xmm_temp0, Address(state, 0));
3848 __ pshufb(xmm_temp0, xmm_temp10);
3849
3850
3851 __ BIND(L_ghash_loop);
3852 __ movdqu(xmm_temp2, Address(data, 0));
3853 __ pshufb(xmm_temp2, ExternalAddress(StubRoutines::x86::ghash_byte_swap_mask_addr()));
3854
3855 __ movdqu(xmm_temp1, Address(subkeyH, 0));
3856 __ pshufb(xmm_temp1, xmm_temp10);
3857
3858 __ pxor(xmm_temp0, xmm_temp2);
3859
3860 //
3861 // Multiply with the hash key
3862 //
3863 __ movdqu(xmm_temp3, xmm_temp0);
3864 __ pclmulqdq(xmm_temp3, xmm_temp1, 0); // xmm3 holds a0*b0
3865 __ movdqu(xmm_temp4, xmm_temp0);
3866 __ pclmulqdq(xmm_temp4, xmm_temp1, 16); // xmm4 holds a0*b1
3867
3868 __ movdqu(xmm_temp5, xmm_temp0);
3869 __ pclmulqdq(xmm_temp5, xmm_temp1, 1); // xmm5 holds a1*b0
3870 __ movdqu(xmm_temp6, xmm_temp0);
3871 __ pclmulqdq(xmm_temp6, xmm_temp1, 17); // xmm6 holds a1*b1
3872
3873 __ pxor(xmm_temp4, xmm_temp5); // xmm4 holds a0*b1 + a1*b0
3874
3875 __ movdqu(xmm_temp5, xmm_temp4); // move the contents of xmm4 to xmm5
3876 __ psrldq(xmm_temp4, 8); // shift by xmm4 64 bits to the right
3877 __ pslldq(xmm_temp5, 8); // shift by xmm5 64 bits to the left
3878 __ pxor(xmm_temp3, xmm_temp5);
3879 __ pxor(xmm_temp6, xmm_temp4); // Register pair <xmm6:xmm3> holds the result
3880 // of the carry-less multiplication of
3881 // xmm0 by xmm1.
3882
3883 // We shift the result of the multiplication by one bit position
3884 // to the left to cope for the fact that the bits are reversed.
3885 __ movdqu(xmm_temp7, xmm_temp3);
3886 __ movdqu(xmm_temp8, xmm_temp6);
3887 __ pslld(xmm_temp3, 1);
3888 __ pslld(xmm_temp6, 1);
3889 __ psrld(xmm_temp7, 31);
3890 __ psrld(xmm_temp8, 31);
3891 __ movdqu(xmm_temp9, xmm_temp7);
3892 __ pslldq(xmm_temp8, 4);
3893 __ pslldq(xmm_temp7, 4);
3894 __ psrldq(xmm_temp9, 12);
3895 __ por(xmm_temp3, xmm_temp7);
3896 __ por(xmm_temp6, xmm_temp8);
3897 __ por(xmm_temp6, xmm_temp9);
3898
3899 //
3900 // First phase of the reduction
3901 //
3902 // Move xmm3 into xmm7, xmm8, xmm9 in order to perform the shifts
3903 // independently.
3904 __ movdqu(xmm_temp7, xmm_temp3);
3905 __ movdqu(xmm_temp8, xmm_temp3);
3906 __ movdqu(xmm_temp9, xmm_temp3);
3907 __ pslld(xmm_temp7, 31); // packed right shift shifting << 31
3908 __ pslld(xmm_temp8, 30); // packed right shift shifting << 30
3909 __ pslld(xmm_temp9, 25); // packed right shift shifting << 25
3910 __ pxor(xmm_temp7, xmm_temp8); // xor the shifted versions
3911 __ pxor(xmm_temp7, xmm_temp9);
3912 __ movdqu(xmm_temp8, xmm_temp7);
3913 __ pslldq(xmm_temp7, 12);
3914 __ psrldq(xmm_temp8, 4);
3915 __ pxor(xmm_temp3, xmm_temp7); // first phase of the reduction complete
3916
3917 //
3918 // Second phase of the reduction
3919 //
3920 // Make 3 copies of xmm3 in xmm2, xmm4, xmm5 for doing these
3921 // shift operations.
3922 __ movdqu(xmm_temp2, xmm_temp3);
3923 __ movdqu(xmm_temp4, xmm_temp3);
3924 __ movdqu(xmm_temp5, xmm_temp3);
3925 __ psrld(xmm_temp2, 1); // packed left shifting >> 1
3926 __ psrld(xmm_temp4, 2); // packed left shifting >> 2
3927 __ psrld(xmm_temp5, 7); // packed left shifting >> 7
3928 __ pxor(xmm_temp2, xmm_temp4); // xor the shifted versions
3929 __ pxor(xmm_temp2, xmm_temp5);
3930 __ pxor(xmm_temp2, xmm_temp8);
3931 __ pxor(xmm_temp3, xmm_temp2);
3932 __ pxor(xmm_temp6, xmm_temp3); // the result is in xmm6
3933
3934 __ decrement(blocks);
3935 __ jcc(Assembler::zero, L_exit);
3936 __ movdqu(xmm_temp0, xmm_temp6);
3937 __ addptr(data, 16);
3938 __ jmp(L_ghash_loop);
3939
3940 __ BIND(L_exit);
3941 __ pshufb(xmm_temp6, xmm_temp10); // Byte swap 16-byte result
3942 __ movdqu(Address(state, 0), xmm_temp6); // store the result
3943
3944 #ifdef _WIN64
3945 // restore xmm regs belonging to calling function
3946 for (int i = 6; i <= XMM_REG_LAST; i++) {
3947 __ movdqu(as_XMMRegister(i), xmm_save(i));
3948 }
3949 #endif
3950 __ leave();
3951 __ ret(0);
3952 return start;
3953 }
3954
3955 /**
3956 * Arguments:
3957 *
3958 * Inputs:
3959 * c_rarg0 - int crc
3960 * c_rarg1 - byte* buf
3961 * c_rarg2 - int length
3962 *
3963 * Ouput:
3964 * rax - int crc result
3965 */
3966 address generate_updateBytesCRC32() {
3967 assert(UseCRC32Intrinsics, "need AVX and CLMUL instructions");
3968
3969 __ align(CodeEntryAlignment);
3970 StubCodeMark mark(this, "StubRoutines", "updateBytesCRC32");
3971
3972 address start = __ pc();
3973 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
3974 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
3975 // rscratch1: r10
3976 const Register crc = c_rarg0; // crc
3977 const Register buf = c_rarg1; // source java byte array address
3978 const Register len = c_rarg2; // length
3979 const Register table = c_rarg3; // crc_table address (reuse register)
3980 const Register tmp = r11;
3981 assert_different_registers(crc, buf, len, table, tmp, rax);
3982
3983 BLOCK_COMMENT("Entry:");
3984 __ enter(); // required for proper stackwalking of RuntimeStub frame
3985
3986 __ kernel_crc32(crc, buf, len, table, tmp);
3987
3988 __ movl(rax, crc);
3989 __ leave(); // required for proper stackwalking of RuntimeStub frame
3990 __ ret(0);
3991
3992 return start;
3993 }
3994
3995
3996 /**
3997 * Arguments:
3998 *
3999 * Input:
4000 * c_rarg0 - x address
4001 * c_rarg1 - x length
4002 * c_rarg2 - y address
4003 * c_rarg3 - y lenth
4004 * not Win64
4005 * c_rarg4 - z address
4006 * c_rarg5 - z length
4007 * Win64
4008 * rsp+40 - z address
4009 * rsp+48 - z length
4010 */
4011 address generate_multiplyToLen() {
4012 __ align(CodeEntryAlignment);
4013 StubCodeMark mark(this, "StubRoutines", "multiplyToLen");
4014
4015 address start = __ pc();
4016 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
4017 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
4018 const Register x = rdi;
4019 const Register xlen = rax;
4020 const Register y = rsi;
4021 const Register ylen = rcx;
4022 const Register z = r8;
4023 const Register zlen = r11;
4024
4025 // Next registers will be saved on stack in multiply_to_len().
4026 const Register tmp1 = r12;
4027 const Register tmp2 = r13;
4028 const Register tmp3 = r14;
4029 const Register tmp4 = r15;
4030 const Register tmp5 = rbx;
4031
4032 BLOCK_COMMENT("Entry:");
4033 __ enter(); // required for proper stackwalking of RuntimeStub frame
4034
4035 #ifndef _WIN64
4036 __ movptr(zlen, r9); // Save r9 in r11 - zlen
4037 #endif
4038 setup_arg_regs(4); // x => rdi, xlen => rsi, y => rdx
4039 // ylen => rcx, z => r8, zlen => r11
4040 // r9 and r10 may be used to save non-volatile registers
4041 #ifdef _WIN64
4042 // last 2 arguments (#4, #5) are on stack on Win64
4043 __ movptr(z, Address(rsp, 6 * wordSize));
4044 __ movptr(zlen, Address(rsp, 7 * wordSize));
4045 #endif
4046
4047 __ movptr(xlen, rsi);
4048 __ movptr(y, rdx);
4049 __ multiply_to_len(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5);
4050
4051 restore_arg_regs();
4052
4053 __ leave(); // required for proper stackwalking of RuntimeStub frame
4054 __ ret(0);
4055
4056 return start;
4057 }
4058
4059 /**
4060 * Arguments:
4061 *
4062 // Input:
4063 // c_rarg0 - x address
4064 // c_rarg1 - x length
4065 // c_rarg2 - z address
4066 // c_rarg3 - z lenth
4067 *
4068 */
4069 address generate_squareToLen() {
4070
4071 __ align(CodeEntryAlignment);
4072 StubCodeMark mark(this, "StubRoutines", "squareToLen");
4073
4074 address start = __ pc();
4075 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
4076 // Unix: rdi, rsi, rdx, rcx (c_rarg0, c_rarg1, ...)
4077 const Register x = rdi;
4078 const Register len = rsi;
4079 const Register z = r8;
4080 const Register zlen = rcx;
4081
4082 const Register tmp1 = r12;
4083 const Register tmp2 = r13;
4084 const Register tmp3 = r14;
4085 const Register tmp4 = r15;
4086 const Register tmp5 = rbx;
4087
4088 BLOCK_COMMENT("Entry:");
4089 __ enter(); // required for proper stackwalking of RuntimeStub frame
4090
4091 setup_arg_regs(4); // x => rdi, len => rsi, z => rdx
4092 // zlen => rcx
4093 // r9 and r10 may be used to save non-volatile registers
4094 __ movptr(r8, rdx);
4095 __ square_to_len(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx, rax);
4096
4097 restore_arg_regs();
4098
4099 __ leave(); // required for proper stackwalking of RuntimeStub frame
4100 __ ret(0);
4101
4102 return start;
4103 }
4104
4105 /**
4106 * Arguments:
4107 *
4108 * Input:
4109 * c_rarg0 - out address
4110 * c_rarg1 - in address
4111 * c_rarg2 - offset
4112 * c_rarg3 - len
4113 * not Win64
4114 * c_rarg4 - k
4115 * Win64
4116 * rsp+40 - k
4117 */
4118 address generate_mulAdd() {
4119 __ align(CodeEntryAlignment);
4120 StubCodeMark mark(this, "StubRoutines", "mulAdd");
4121
4122 address start = __ pc();
4123 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
4124 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
4125 const Register out = rdi;
4126 const Register in = rsi;
4127 const Register offset = r11;
4128 const Register len = rcx;
4129 const Register k = r8;
4130
4131 // Next registers will be saved on stack in mul_add().
4132 const Register tmp1 = r12;
4133 const Register tmp2 = r13;
4134 const Register tmp3 = r14;
4135 const Register tmp4 = r15;
4136 const Register tmp5 = rbx;
4137
4138 BLOCK_COMMENT("Entry:");
4139 __ enter(); // required for proper stackwalking of RuntimeStub frame
4140
4141 setup_arg_regs(4); // out => rdi, in => rsi, offset => rdx
4142 // len => rcx, k => r8
4143 // r9 and r10 may be used to save non-volatile registers
4144 #ifdef _WIN64
4145 // last argument is on stack on Win64
4146 __ movl(k, Address(rsp, 6 * wordSize));
4147 #endif
4148 __ movptr(r11, rdx); // move offset in rdx to offset(r11)
4149 __ mul_add(out, in, offset, len, k, tmp1, tmp2, tmp3, tmp4, tmp5, rdx, rax);
4150
4151 restore_arg_regs();
4152
4153 __ leave(); // required for proper stackwalking of RuntimeStub frame
4154 __ ret(0);
4155
4156 return start;
4157 }
4158
4159
4160 #undef __
4161 #define __ masm->
4162
4163 // Continuation point for throwing of implicit exceptions that are
4164 // not handled in the current activation. Fabricates an exception
4165 // oop and initiates normal exception dispatching in this
4166 // frame. Since we need to preserve callee-saved values (currently
4167 // only for C2, but done for C1 as well) we need a callee-saved oop
4168 // map and therefore have to make these stubs into RuntimeStubs
4169 // rather than BufferBlobs. If the compiler needs all registers to
4170 // be preserved between the fault point and the exception handler
4171 // then it must assume responsibility for that in
4172 // AbstractCompiler::continuation_for_implicit_null_exception or
4173 // continuation_for_implicit_division_by_zero_exception. All other
4174 // implicit exceptions (e.g., NullPointerException or
4175 // AbstractMethodError on entry) are either at call sites or
4176 // otherwise assume that stack unwinding will be initiated, so
4177 // caller saved registers were assumed volatile in the compiler.
4178 address generate_throw_exception(const char* name,
4179 address runtime_entry,
4180 Register arg1 = noreg,
4181 Register arg2 = noreg) {
4182 // Information about frame layout at time of blocking runtime call.
4183 // Note that we only have to preserve callee-saved registers since
4184 // the compilers are responsible for supplying a continuation point
4185 // if they expect all registers to be preserved.
4186 enum layout {
4187 rbp_off = frame::arg_reg_save_area_bytes/BytesPerInt,
4188 rbp_off2,
4189 return_off,
4190 return_off2,
4191 framesize // inclusive of return address
4192 };
4193
4194 int insts_size = 512;
4195 int locs_size = 64;
4196
4197 CodeBuffer code(name, insts_size, locs_size);
4198 OopMapSet* oop_maps = new OopMapSet();
4199 MacroAssembler* masm = new MacroAssembler(&code);
4200
4201 address start = __ pc();
4202
4203 // This is an inlined and slightly modified version of call_VM
4204 // which has the ability to fetch the return PC out of
4205 // thread-local storage and also sets up last_Java_sp slightly
4206 // differently than the real call_VM
4207
4208 __ enter(); // required for proper stackwalking of RuntimeStub frame
4209
4210 assert(is_even(framesize/2), "sp not 16-byte aligned");
4211
4212 // return address and rbp are already in place
4213 __ subptr(rsp, (framesize-4) << LogBytesPerInt); // prolog
4214
4215 int frame_complete = __ pc() - start;
4216
4217 // Set up last_Java_sp and last_Java_fp
4218 address the_pc = __ pc();
4219 __ set_last_Java_frame(rsp, rbp, the_pc);
4220 __ andptr(rsp, -(StackAlignmentInBytes)); // Align stack
4221
4222 // Call runtime
4223 if (arg1 != noreg) {
4224 assert(arg2 != c_rarg1, "clobbered");
4225 __ movptr(c_rarg1, arg1);
4226 }
4227 if (arg2 != noreg) {
4228 __ movptr(c_rarg2, arg2);
4229 }
4230 __ movptr(c_rarg0, r15_thread);
4231 BLOCK_COMMENT("call runtime_entry");
4232 __ call(RuntimeAddress(runtime_entry));
4233
4234 // Generate oop map
4235 OopMap* map = new OopMap(framesize, 0);
4236
4237 oop_maps->add_gc_map(the_pc - start, map);
4238
4239 __ reset_last_Java_frame(true, true);
4240
4241 __ leave(); // required for proper stackwalking of RuntimeStub frame
4242
4243 // check for pending exceptions
4244 #ifdef ASSERT
4245 Label L;
4246 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()),
4247 (int32_t) NULL_WORD);
4248 __ jcc(Assembler::notEqual, L);
4249 __ should_not_reach_here();
4250 __ bind(L);
4251 #endif // ASSERT
4252 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
4253
4254
4255 // codeBlob framesize is in words (not VMRegImpl::slot_size)
4256 RuntimeStub* stub =
4257 RuntimeStub::new_runtime_stub(name,
4258 &code,
4259 frame_complete,
4260 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
4261 oop_maps, false);
4262 return stub->entry_point();
4263 }
4264
4265 void create_control_words() {
4266 // Round to nearest, 53-bit mode, exceptions masked
4267 StubRoutines::_fpu_cntrl_wrd_std = 0x027F;
4268 // Round to zero, 53-bit mode, exception mased
4269 StubRoutines::_fpu_cntrl_wrd_trunc = 0x0D7F;
4270 // Round to nearest, 24-bit mode, exceptions masked
4271 StubRoutines::_fpu_cntrl_wrd_24 = 0x007F;
4272 // Round to nearest, 64-bit mode, exceptions masked
4273 StubRoutines::_fpu_cntrl_wrd_64 = 0x037F;
4274 // Round to nearest, 64-bit mode, exceptions masked
4275 StubRoutines::_mxcsr_std = 0x1F80;
4276 // Note: the following two constants are 80-bit values
4277 // layout is critical for correct loading by FPU.
4278 // Bias for strict fp multiply/divide
4279 StubRoutines::_fpu_subnormal_bias1[0]= 0x00000000; // 2^(-15360) == 0x03ff 8000 0000 0000 0000
4280 StubRoutines::_fpu_subnormal_bias1[1]= 0x80000000;
4281 StubRoutines::_fpu_subnormal_bias1[2]= 0x03ff;
4282 // Un-Bias for strict fp multiply/divide
4283 StubRoutines::_fpu_subnormal_bias2[0]= 0x00000000; // 2^(+15360) == 0x7bff 8000 0000 0000 0000
4284 StubRoutines::_fpu_subnormal_bias2[1]= 0x80000000;
4285 StubRoutines::_fpu_subnormal_bias2[2]= 0x7bff;
4286 }
4287
4288 // Initialization
4289 void generate_initial() {
4290 // Generates all stubs and initializes the entry points
4291
4292 // This platform-specific settings are needed by generate_call_stub()
4293 create_control_words();
4294
4295 // entry points that exist in all platforms Note: This is code
4296 // that could be shared among different platforms - however the
4297 // benefit seems to be smaller than the disadvantage of having a
4298 // much more complicated generator structure. See also comment in
4299 // stubRoutines.hpp.
4300
4301 StubRoutines::_forward_exception_entry = generate_forward_exception();
4302
4303 StubRoutines::_call_stub_entry =
4304 generate_call_stub(StubRoutines::_call_stub_return_address);
4305
4306 // is referenced by megamorphic call
4307 StubRoutines::_catch_exception_entry = generate_catch_exception();
4308
4309 // atomic calls
4310 StubRoutines::_atomic_xchg_entry = generate_atomic_xchg();
4311 StubRoutines::_atomic_xchg_ptr_entry = generate_atomic_xchg_ptr();
4312 StubRoutines::_atomic_cmpxchg_entry = generate_atomic_cmpxchg();
4313 StubRoutines::_atomic_cmpxchg_byte_entry = generate_atomic_cmpxchg_byte();
4314 StubRoutines::_atomic_cmpxchg_long_entry = generate_atomic_cmpxchg_long();
4315 StubRoutines::_atomic_add_entry = generate_atomic_add();
4316 StubRoutines::_atomic_add_ptr_entry = generate_atomic_add_ptr();
4317 StubRoutines::_fence_entry = generate_orderaccess_fence();
4318
4319 StubRoutines::_handler_for_unsafe_access_entry =
4320 generate_handler_for_unsafe_access();
4321
4322 // platform dependent
4323 StubRoutines::x86::_get_previous_fp_entry = generate_get_previous_fp();
4324 StubRoutines::x86::_get_previous_sp_entry = generate_get_previous_sp();
4325
4326 StubRoutines::x86::_verify_mxcsr_entry = generate_verify_mxcsr();
4327
4328 // Build this early so it's available for the interpreter.
4329 StubRoutines::_throw_StackOverflowError_entry =
4330 generate_throw_exception("StackOverflowError throw_exception",
4331 CAST_FROM_FN_PTR(address,
4332 SharedRuntime::
4333 throw_StackOverflowError));
4334 if (UseCRC32Intrinsics) {
4335 // set table address before stub generation which use it
4336 StubRoutines::_crc_table_adr = (address)StubRoutines::x86::_crc_table;
4337 StubRoutines::_updateBytesCRC32 = generate_updateBytesCRC32();
4338 }
4339 }
4340
4341 void generate_all() {
4342 // Generates all stubs and initializes the entry points
4343
4344 // These entry points require SharedInfo::stack0 to be set up in
4345 // non-core builds and need to be relocatable, so they each
4346 // fabricate a RuntimeStub internally.
4347 StubRoutines::_throw_AbstractMethodError_entry =
4348 generate_throw_exception("AbstractMethodError throw_exception",
4349 CAST_FROM_FN_PTR(address,
4350 SharedRuntime::
4351 throw_AbstractMethodError));
4352
4353 StubRoutines::_throw_IncompatibleClassChangeError_entry =
4354 generate_throw_exception("IncompatibleClassChangeError throw_exception",
4355 CAST_FROM_FN_PTR(address,
4356 SharedRuntime::
4357 throw_IncompatibleClassChangeError));
4358
4359 StubRoutines::_throw_NullPointerException_at_call_entry =
4360 generate_throw_exception("NullPointerException at call throw_exception",
4361 CAST_FROM_FN_PTR(address,
4362 SharedRuntime::
4363 throw_NullPointerException_at_call));
4364
4365 // entry points that are platform specific
4366 if (UseShenandoahGC) {
4367 StubRoutines::x86::_shenandoah_wb = generate_shenandoah_wb();
4368 }
4369 StubRoutines::x86::_f2i_fixup = generate_f2i_fixup();
4370 StubRoutines::x86::_f2l_fixup = generate_f2l_fixup();
4371 StubRoutines::x86::_d2i_fixup = generate_d2i_fixup();
4372 StubRoutines::x86::_d2l_fixup = generate_d2l_fixup();
4373
4374 StubRoutines::x86::_float_sign_mask = generate_fp_mask("float_sign_mask", 0x7FFFFFFF7FFFFFFF);
4375 StubRoutines::x86::_float_sign_flip = generate_fp_mask("float_sign_flip", 0x8000000080000000);
4376 StubRoutines::x86::_double_sign_mask = generate_fp_mask("double_sign_mask", 0x7FFFFFFFFFFFFFFF);
4377 StubRoutines::x86::_double_sign_flip = generate_fp_mask("double_sign_flip", 0x8000000000000000);
4378
4379 // support for verify_oop (must happen after universe_init)
4380 StubRoutines::_verify_oop_subroutine_entry = generate_verify_oop();
4381
4382 // arraycopy stubs used by compilers
4383 generate_arraycopy_stubs();
4384
4385 generate_math_stubs();
4386
4387 // don't bother generating these AES intrinsic stubs unless global flag is set
4388 if (UseAESIntrinsics) {
4389 StubRoutines::x86::_key_shuffle_mask_addr = generate_key_shuffle_mask(); // needed by the others
4390
4391 StubRoutines::_aescrypt_encryptBlock = generate_aescrypt_encryptBlock();
4392 StubRoutines::_aescrypt_decryptBlock = generate_aescrypt_decryptBlock();
4393 StubRoutines::_cipherBlockChaining_encryptAESCrypt = generate_cipherBlockChaining_encryptAESCrypt();
4394 StubRoutines::_cipherBlockChaining_decryptAESCrypt = generate_cipherBlockChaining_decryptAESCrypt_Parallel();
4395 }
4396
4397 // Generate GHASH intrinsics code
4398 if (UseGHASHIntrinsics) {
4399 StubRoutines::x86::_ghash_long_swap_mask_addr = generate_ghash_long_swap_mask();
4400 StubRoutines::x86::_ghash_byte_swap_mask_addr = generate_ghash_byte_swap_mask();
4401 StubRoutines::_ghash_processBlocks = generate_ghash_processBlocks();
4402 }
4403
4404 // Safefetch stubs.
4405 generate_safefetch("SafeFetch32", sizeof(int), &StubRoutines::_safefetch32_entry,
4406 &StubRoutines::_safefetch32_fault_pc,
4407 &StubRoutines::_safefetch32_continuation_pc);
4408 generate_safefetch("SafeFetchN", sizeof(intptr_t), &StubRoutines::_safefetchN_entry,
4409 &StubRoutines::_safefetchN_fault_pc,
4410 &StubRoutines::_safefetchN_continuation_pc);
4411 #ifdef COMPILER2
4412 if (UseMultiplyToLenIntrinsic) {
4413 StubRoutines::_multiplyToLen = generate_multiplyToLen();
4414 }
4415 if (UseSquareToLenIntrinsic) {
4416 StubRoutines::_squareToLen = generate_squareToLen();
4417 }
4418 if (UseMulAddIntrinsic) {
4419 StubRoutines::_mulAdd = generate_mulAdd();
4420 }
4421
4422 #ifndef _WINDOWS
4423 if (UseMontgomeryMultiplyIntrinsic) {
4424 StubRoutines::_montgomeryMultiply
4425 = CAST_FROM_FN_PTR(address, SharedRuntime::montgomery_multiply);
4426 }
4427 if (UseMontgomerySquareIntrinsic) {
4428 StubRoutines::_montgomerySquare
4429 = CAST_FROM_FN_PTR(address, SharedRuntime::montgomery_square);
4430 }
4431 #endif // WINDOWS
4432 #endif // COMPILER2
4433 }
4434
4435 public:
4436 StubGenerator(CodeBuffer* code, bool all) : StubCodeGenerator(code) {
4437 if (all) {
4438 generate_all();
4439 } else {
4440 generate_initial();
4441 }
4442 }
4443 }; // end class declaration
4444
4445 void StubGenerator_generate(CodeBuffer* code, bool all) {
4446 StubGenerator g(code, all);
4447 }