Thu, 03 Jan 2013 16:30:47 -0800
8005544: Use 256bit YMM registers in arraycopy stubs on x86
Summary: Use YMM registers in arraycopy and array_fill stubs.
Reviewed-by: roland, twisti
1 /*
2 * Copyright (c) 2003, 2012, 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 */
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
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
50 #define __ _masm->
51 #define TIMES_OOP (UseCompressedOops ? Address::times_4 : Address::times_8)
52 #define a__ ((Assembler*)_masm)->
54 #ifdef PRODUCT
55 #define BLOCK_COMMENT(str) /* nothing */
56 #else
57 #define BLOCK_COMMENT(str) __ block_comment(str)
58 #endif
60 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
61 const int MXCSR_MASK = 0xFFC0; // Mask out any pending exceptions
63 // Stub Code definitions
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);
73 // request an async exception
74 thread->set_pending_unsafe_access_error();
76 // return address of next instruction to execute
77 return npc;
78 }
80 class StubGenerator: public StubCodeGenerator {
81 private:
83 #ifdef PRODUCT
84 #define inc_counter_np(counter) (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
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 // -28 [ argument word 1 ]
141 // -27 [ saved xmm15 ] <--- rsp_after_call
142 // [ saved xmm7-xmm14 ]
143 // -9 [ saved xmm6 ] (each xmm register takes 2 slots)
144 // -7 [ saved r15 ]
145 // -6 [ saved r14 ]
146 // -5 [ saved r13 ]
147 // -4 [ saved r12 ]
148 // -3 [ saved rdi ]
149 // -2 [ saved rsi ]
150 // -1 [ saved rbx ]
151 // 0 [ saved rbp ] <--- rbp
152 // 1 [ return address ]
153 // 2 [ call wrapper ]
154 // 3 [ result ]
155 // 4 [ result type ]
156 // 5 [ method ]
157 // 6 [ entry point ]
158 // 7 [ parameters ]
159 // 8 [ parameter size ]
160 // 9 [ thread ]
161 //
162 // Windows reserves the callers stack space for arguments 1-4.
163 // We spill c_rarg0-c_rarg3 to this space.
165 // Call stub stack layout word offsets from rbp
166 enum call_stub_layout {
167 #ifdef _WIN64
168 xmm_save_first = 6, // save from xmm6
169 xmm_save_last = 15, // to xmm15
170 xmm_save_base = -9,
171 rsp_after_call_off = xmm_save_base - 2 * (xmm_save_last - xmm_save_first), // -27
172 r15_off = -7,
173 r14_off = -6,
174 r13_off = -5,
175 r12_off = -4,
176 rdi_off = -3,
177 rsi_off = -2,
178 rbx_off = -1,
179 rbp_off = 0,
180 retaddr_off = 1,
181 call_wrapper_off = 2,
182 result_off = 3,
183 result_type_off = 4,
184 method_off = 5,
185 entry_point_off = 6,
186 parameters_off = 7,
187 parameter_size_off = 8,
188 thread_off = 9
189 #else
190 rsp_after_call_off = -12,
191 mxcsr_off = rsp_after_call_off,
192 r15_off = -11,
193 r14_off = -10,
194 r13_off = -9,
195 r12_off = -8,
196 rbx_off = -7,
197 call_wrapper_off = -6,
198 result_off = -5,
199 result_type_off = -4,
200 method_off = -3,
201 entry_point_off = -2,
202 parameters_off = -1,
203 rbp_off = 0,
204 retaddr_off = 1,
205 parameter_size_off = 2,
206 thread_off = 3
207 #endif
208 };
210 #ifdef _WIN64
211 Address xmm_save(int reg) {
212 assert(reg >= xmm_save_first && reg <= xmm_save_last, "XMM register number out of range");
213 return Address(rbp, (xmm_save_base - (reg - xmm_save_first) * 2) * wordSize);
214 }
215 #endif
217 address generate_call_stub(address& return_address) {
218 assert((int)frame::entry_frame_after_call_words == -(int)rsp_after_call_off + 1 &&
219 (int)frame::entry_frame_call_wrapper_offset == (int)call_wrapper_off,
220 "adjust this code");
221 StubCodeMark mark(this, "StubRoutines", "call_stub");
222 address start = __ pc();
224 // same as in generate_catch_exception()!
225 const Address rsp_after_call(rbp, rsp_after_call_off * wordSize);
227 const Address call_wrapper (rbp, call_wrapper_off * wordSize);
228 const Address result (rbp, result_off * wordSize);
229 const Address result_type (rbp, result_type_off * wordSize);
230 const Address method (rbp, method_off * wordSize);
231 const Address entry_point (rbp, entry_point_off * wordSize);
232 const Address parameters (rbp, parameters_off * wordSize);
233 const Address parameter_size(rbp, parameter_size_off * wordSize);
235 // same as in generate_catch_exception()!
236 const Address thread (rbp, thread_off * wordSize);
238 const Address r15_save(rbp, r15_off * wordSize);
239 const Address r14_save(rbp, r14_off * wordSize);
240 const Address r13_save(rbp, r13_off * wordSize);
241 const Address r12_save(rbp, r12_off * wordSize);
242 const Address rbx_save(rbp, rbx_off * wordSize);
244 // stub code
245 __ enter();
246 __ subptr(rsp, -rsp_after_call_off * wordSize);
248 // save register parameters
249 #ifndef _WIN64
250 __ movptr(parameters, c_rarg5); // parameters
251 __ movptr(entry_point, c_rarg4); // entry_point
252 #endif
254 __ movptr(method, c_rarg3); // method
255 __ movl(result_type, c_rarg2); // result type
256 __ movptr(result, c_rarg1); // result
257 __ movptr(call_wrapper, c_rarg0); // call wrapper
259 // save regs belonging to calling function
260 __ movptr(rbx_save, rbx);
261 __ movptr(r12_save, r12);
262 __ movptr(r13_save, r13);
263 __ movptr(r14_save, r14);
264 __ movptr(r15_save, r15);
265 #ifdef _WIN64
266 for (int i = 6; i <= 15; i++) {
267 __ movdqu(xmm_save(i), as_XMMRegister(i));
268 }
270 const Address rdi_save(rbp, rdi_off * wordSize);
271 const Address rsi_save(rbp, rsi_off * wordSize);
273 __ movptr(rsi_save, rsi);
274 __ movptr(rdi_save, rdi);
275 #else
276 const Address mxcsr_save(rbp, mxcsr_off * wordSize);
277 {
278 Label skip_ldmx;
279 __ stmxcsr(mxcsr_save);
280 __ movl(rax, mxcsr_save);
281 __ andl(rax, MXCSR_MASK); // Only check control and mask bits
282 ExternalAddress mxcsr_std(StubRoutines::x86::mxcsr_std());
283 __ cmp32(rax, mxcsr_std);
284 __ jcc(Assembler::equal, skip_ldmx);
285 __ ldmxcsr(mxcsr_std);
286 __ bind(skip_ldmx);
287 }
288 #endif
290 // Load up thread register
291 __ movptr(r15_thread, thread);
292 __ reinit_heapbase();
294 #ifdef ASSERT
295 // make sure we have no pending exceptions
296 {
297 Label L;
298 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
299 __ jcc(Assembler::equal, L);
300 __ stop("StubRoutines::call_stub: entered with pending exception");
301 __ bind(L);
302 }
303 #endif
305 // pass parameters if any
306 BLOCK_COMMENT("pass parameters if any");
307 Label parameters_done;
308 __ movl(c_rarg3, parameter_size);
309 __ testl(c_rarg3, c_rarg3);
310 __ jcc(Assembler::zero, parameters_done);
312 Label loop;
313 __ movptr(c_rarg2, parameters); // parameter pointer
314 __ movl(c_rarg1, c_rarg3); // parameter counter is in c_rarg1
315 __ BIND(loop);
316 __ movptr(rax, Address(c_rarg2, 0));// get parameter
317 __ addptr(c_rarg2, wordSize); // advance to next parameter
318 __ decrementl(c_rarg1); // decrement counter
319 __ push(rax); // pass parameter
320 __ jcc(Assembler::notZero, loop);
322 // call Java function
323 __ BIND(parameters_done);
324 __ movptr(rbx, method); // get Method*
325 __ movptr(c_rarg1, entry_point); // get entry_point
326 __ mov(r13, rsp); // set sender sp
327 BLOCK_COMMENT("call Java function");
328 __ call(c_rarg1);
330 BLOCK_COMMENT("call_stub_return_address:");
331 return_address = __ pc();
333 // store result depending on type (everything that is not
334 // T_OBJECT, T_LONG, T_FLOAT or T_DOUBLE is treated as T_INT)
335 __ movptr(c_rarg0, result);
336 Label is_long, is_float, is_double, exit;
337 __ movl(c_rarg1, result_type);
338 __ cmpl(c_rarg1, T_OBJECT);
339 __ jcc(Assembler::equal, is_long);
340 __ cmpl(c_rarg1, T_LONG);
341 __ jcc(Assembler::equal, is_long);
342 __ cmpl(c_rarg1, T_FLOAT);
343 __ jcc(Assembler::equal, is_float);
344 __ cmpl(c_rarg1, T_DOUBLE);
345 __ jcc(Assembler::equal, is_double);
347 // handle T_INT case
348 __ movl(Address(c_rarg0, 0), rax);
350 __ BIND(exit);
352 // pop parameters
353 __ lea(rsp, rsp_after_call);
355 #ifdef ASSERT
356 // verify that threads correspond
357 {
358 Label L, S;
359 __ cmpptr(r15_thread, thread);
360 __ jcc(Assembler::notEqual, S);
361 __ get_thread(rbx);
362 __ cmpptr(r15_thread, rbx);
363 __ jcc(Assembler::equal, L);
364 __ bind(S);
365 __ jcc(Assembler::equal, L);
366 __ stop("StubRoutines::call_stub: threads must correspond");
367 __ bind(L);
368 }
369 #endif
371 // restore regs belonging to calling function
372 #ifdef _WIN64
373 for (int i = 15; i >= 6; i--) {
374 __ movdqu(as_XMMRegister(i), xmm_save(i));
375 }
376 #endif
377 __ movptr(r15, r15_save);
378 __ movptr(r14, r14_save);
379 __ movptr(r13, r13_save);
380 __ movptr(r12, r12_save);
381 __ movptr(rbx, rbx_save);
383 #ifdef _WIN64
384 __ movptr(rdi, rdi_save);
385 __ movptr(rsi, rsi_save);
386 #else
387 __ ldmxcsr(mxcsr_save);
388 #endif
390 // restore rsp
391 __ addptr(rsp, -rsp_after_call_off * wordSize);
393 // return
394 __ pop(rbp);
395 __ ret(0);
397 // handle return types different from T_INT
398 __ BIND(is_long);
399 __ movq(Address(c_rarg0, 0), rax);
400 __ jmp(exit);
402 __ BIND(is_float);
403 __ movflt(Address(c_rarg0, 0), xmm0);
404 __ jmp(exit);
406 __ BIND(is_double);
407 __ movdbl(Address(c_rarg0, 0), xmm0);
408 __ jmp(exit);
410 return start;
411 }
413 // Return point for a Java call if there's an exception thrown in
414 // Java code. The exception is caught and transformed into a
415 // pending exception stored in JavaThread that can be tested from
416 // within the VM.
417 //
418 // Note: Usually the parameters are removed by the callee. In case
419 // of an exception crossing an activation frame boundary, that is
420 // not the case if the callee is compiled code => need to setup the
421 // rsp.
422 //
423 // rax: exception oop
425 address generate_catch_exception() {
426 StubCodeMark mark(this, "StubRoutines", "catch_exception");
427 address start = __ pc();
429 // same as in generate_call_stub():
430 const Address rsp_after_call(rbp, rsp_after_call_off * wordSize);
431 const Address thread (rbp, thread_off * wordSize);
433 #ifdef ASSERT
434 // verify that threads correspond
435 {
436 Label L, S;
437 __ cmpptr(r15_thread, thread);
438 __ jcc(Assembler::notEqual, S);
439 __ get_thread(rbx);
440 __ cmpptr(r15_thread, rbx);
441 __ jcc(Assembler::equal, L);
442 __ bind(S);
443 __ stop("StubRoutines::catch_exception: threads must correspond");
444 __ bind(L);
445 }
446 #endif
448 // set pending exception
449 __ verify_oop(rax);
451 __ movptr(Address(r15_thread, Thread::pending_exception_offset()), rax);
452 __ lea(rscratch1, ExternalAddress((address)__FILE__));
453 __ movptr(Address(r15_thread, Thread::exception_file_offset()), rscratch1);
454 __ movl(Address(r15_thread, Thread::exception_line_offset()), (int) __LINE__);
456 // complete return to VM
457 assert(StubRoutines::_call_stub_return_address != NULL,
458 "_call_stub_return_address must have been generated before");
459 __ jump(RuntimeAddress(StubRoutines::_call_stub_return_address));
461 return start;
462 }
464 // Continuation point for runtime calls returning with a pending
465 // exception. The pending exception check happened in the runtime
466 // or native call stub. The pending exception in Thread is
467 // converted into a Java-level exception.
468 //
469 // Contract with Java-level exception handlers:
470 // rax: exception
471 // rdx: throwing pc
472 //
473 // NOTE: At entry of this stub, exception-pc must be on stack !!
475 address generate_forward_exception() {
476 StubCodeMark mark(this, "StubRoutines", "forward exception");
477 address start = __ pc();
479 // Upon entry, the sp points to the return address returning into
480 // Java (interpreted or compiled) code; i.e., the return address
481 // becomes the throwing pc.
482 //
483 // Arguments pushed before the runtime call are still on the stack
484 // but the exception handler will reset the stack pointer ->
485 // ignore them. A potential result in registers can be ignored as
486 // well.
488 #ifdef ASSERT
489 // make sure this code is only executed if there is a pending exception
490 {
491 Label L;
492 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t) NULL);
493 __ jcc(Assembler::notEqual, L);
494 __ stop("StubRoutines::forward exception: no pending exception (1)");
495 __ bind(L);
496 }
497 #endif
499 // compute exception handler into rbx
500 __ movptr(c_rarg0, Address(rsp, 0));
501 BLOCK_COMMENT("call exception_handler_for_return_address");
502 __ call_VM_leaf(CAST_FROM_FN_PTR(address,
503 SharedRuntime::exception_handler_for_return_address),
504 r15_thread, c_rarg0);
505 __ mov(rbx, rax);
507 // setup rax & rdx, remove return address & clear pending exception
508 __ pop(rdx);
509 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset()));
510 __ movptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
512 #ifdef ASSERT
513 // make sure exception is set
514 {
515 Label L;
516 __ testptr(rax, rax);
517 __ jcc(Assembler::notEqual, L);
518 __ stop("StubRoutines::forward exception: no pending exception (2)");
519 __ bind(L);
520 }
521 #endif
523 // continue at exception handler (return address removed)
524 // rax: exception
525 // rbx: exception handler
526 // rdx: throwing pc
527 __ verify_oop(rax);
528 __ jmp(rbx);
530 return start;
531 }
533 // Support for jint atomic::xchg(jint exchange_value, volatile jint* dest)
534 //
535 // Arguments :
536 // c_rarg0: exchange_value
537 // c_rarg0: dest
538 //
539 // Result:
540 // *dest <- ex, return (orig *dest)
541 address generate_atomic_xchg() {
542 StubCodeMark mark(this, "StubRoutines", "atomic_xchg");
543 address start = __ pc();
545 __ movl(rax, c_rarg0); // Copy to eax we need a return value anyhow
546 __ xchgl(rax, Address(c_rarg1, 0)); // automatic LOCK
547 __ ret(0);
549 return start;
550 }
552 // Support for intptr_t atomic::xchg_ptr(intptr_t exchange_value, volatile intptr_t* dest)
553 //
554 // Arguments :
555 // c_rarg0: exchange_value
556 // c_rarg1: dest
557 //
558 // Result:
559 // *dest <- ex, return (orig *dest)
560 address generate_atomic_xchg_ptr() {
561 StubCodeMark mark(this, "StubRoutines", "atomic_xchg_ptr");
562 address start = __ pc();
564 __ movptr(rax, c_rarg0); // Copy to eax we need a return value anyhow
565 __ xchgptr(rax, Address(c_rarg1, 0)); // automatic LOCK
566 __ ret(0);
568 return start;
569 }
571 // Support for jint atomic::atomic_cmpxchg(jint exchange_value, volatile jint* dest,
572 // jint compare_value)
573 //
574 // Arguments :
575 // c_rarg0: exchange_value
576 // c_rarg1: dest
577 // c_rarg2: compare_value
578 //
579 // Result:
580 // if ( compare_value == *dest ) {
581 // *dest = exchange_value
582 // return compare_value;
583 // else
584 // return *dest;
585 address generate_atomic_cmpxchg() {
586 StubCodeMark mark(this, "StubRoutines", "atomic_cmpxchg");
587 address start = __ pc();
589 __ movl(rax, c_rarg2);
590 if ( os::is_MP() ) __ lock();
591 __ cmpxchgl(c_rarg0, Address(c_rarg1, 0));
592 __ ret(0);
594 return start;
595 }
597 // Support for jint atomic::atomic_cmpxchg_long(jlong exchange_value,
598 // volatile jlong* dest,
599 // jlong compare_value)
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_long() {
612 StubCodeMark mark(this, "StubRoutines", "atomic_cmpxchg_long");
613 address start = __ pc();
615 __ movq(rax, c_rarg2);
616 if ( os::is_MP() ) __ lock();
617 __ cmpxchgq(c_rarg0, Address(c_rarg1, 0));
618 __ ret(0);
620 return start;
621 }
623 // Support for jint atomic::add(jint add_value, volatile jint* dest)
624 //
625 // Arguments :
626 // c_rarg0: add_value
627 // c_rarg1: dest
628 //
629 // Result:
630 // *dest += add_value
631 // return *dest;
632 address generate_atomic_add() {
633 StubCodeMark mark(this, "StubRoutines", "atomic_add");
634 address start = __ pc();
636 __ movl(rax, c_rarg0);
637 if ( os::is_MP() ) __ lock();
638 __ xaddl(Address(c_rarg1, 0), c_rarg0);
639 __ addl(rax, c_rarg0);
640 __ ret(0);
642 return start;
643 }
645 // Support for intptr_t atomic::add_ptr(intptr_t add_value, volatile intptr_t* dest)
646 //
647 // Arguments :
648 // c_rarg0: add_value
649 // c_rarg1: dest
650 //
651 // Result:
652 // *dest += add_value
653 // return *dest;
654 address generate_atomic_add_ptr() {
655 StubCodeMark mark(this, "StubRoutines", "atomic_add_ptr");
656 address start = __ pc();
658 __ movptr(rax, c_rarg0); // Copy to eax we need a return value anyhow
659 if ( os::is_MP() ) __ lock();
660 __ xaddptr(Address(c_rarg1, 0), c_rarg0);
661 __ addptr(rax, c_rarg0);
662 __ ret(0);
664 return start;
665 }
667 // Support for intptr_t OrderAccess::fence()
668 //
669 // Arguments :
670 //
671 // Result:
672 address generate_orderaccess_fence() {
673 StubCodeMark mark(this, "StubRoutines", "orderaccess_fence");
674 address start = __ pc();
675 __ membar(Assembler::StoreLoad);
676 __ ret(0);
678 return start;
679 }
681 // Support for intptr_t get_previous_fp()
682 //
683 // This routine is used to find the previous frame pointer for the
684 // caller (current_frame_guess). This is used as part of debugging
685 // ps() is seemingly lost trying to find frames.
686 // This code assumes that caller current_frame_guess) has a frame.
687 address generate_get_previous_fp() {
688 StubCodeMark mark(this, "StubRoutines", "get_previous_fp");
689 const Address old_fp(rbp, 0);
690 const Address older_fp(rax, 0);
691 address start = __ pc();
693 __ enter();
694 __ movptr(rax, old_fp); // callers fp
695 __ movptr(rax, older_fp); // the frame for ps()
696 __ pop(rbp);
697 __ ret(0);
699 return start;
700 }
702 // Support for intptr_t get_previous_sp()
703 //
704 // This routine is used to find the previous stack pointer for the
705 // caller.
706 address generate_get_previous_sp() {
707 StubCodeMark mark(this, "StubRoutines", "get_previous_sp");
708 address start = __ pc();
710 __ movptr(rax, rsp);
711 __ addptr(rax, 8); // return address is at the top of the stack.
712 __ ret(0);
714 return start;
715 }
717 //----------------------------------------------------------------------------------------------------
718 // Support for void verify_mxcsr()
719 //
720 // This routine is used with -Xcheck:jni to verify that native
721 // JNI code does not return to Java code without restoring the
722 // MXCSR register to our expected state.
724 address generate_verify_mxcsr() {
725 StubCodeMark mark(this, "StubRoutines", "verify_mxcsr");
726 address start = __ pc();
728 const Address mxcsr_save(rsp, 0);
730 if (CheckJNICalls) {
731 Label ok_ret;
732 __ push(rax);
733 __ subptr(rsp, wordSize); // allocate a temp location
734 __ stmxcsr(mxcsr_save);
735 __ movl(rax, mxcsr_save);
736 __ andl(rax, MXCSR_MASK); // Only check control and mask bits
737 __ cmpl(rax, *(int *)(StubRoutines::x86::mxcsr_std()));
738 __ jcc(Assembler::equal, ok_ret);
740 __ warn("MXCSR changed by native JNI code, use -XX:+RestoreMXCSROnJNICall");
742 __ ldmxcsr(ExternalAddress(StubRoutines::x86::mxcsr_std()));
744 __ bind(ok_ret);
745 __ addptr(rsp, wordSize);
746 __ pop(rax);
747 }
749 __ ret(0);
751 return start;
752 }
754 address generate_f2i_fixup() {
755 StubCodeMark mark(this, "StubRoutines", "f2i_fixup");
756 Address inout(rsp, 5 * wordSize); // return address + 4 saves
758 address start = __ pc();
760 Label L;
762 __ push(rax);
763 __ push(c_rarg3);
764 __ push(c_rarg2);
765 __ push(c_rarg1);
767 __ movl(rax, 0x7f800000);
768 __ xorl(c_rarg3, c_rarg3);
769 __ movl(c_rarg2, inout);
770 __ movl(c_rarg1, c_rarg2);
771 __ andl(c_rarg1, 0x7fffffff);
772 __ cmpl(rax, c_rarg1); // NaN? -> 0
773 __ jcc(Assembler::negative, L);
774 __ testl(c_rarg2, c_rarg2); // signed ? min_jint : max_jint
775 __ movl(c_rarg3, 0x80000000);
776 __ movl(rax, 0x7fffffff);
777 __ cmovl(Assembler::positive, c_rarg3, rax);
779 __ bind(L);
780 __ movptr(inout, c_rarg3);
782 __ pop(c_rarg1);
783 __ pop(c_rarg2);
784 __ pop(c_rarg3);
785 __ pop(rax);
787 __ ret(0);
789 return start;
790 }
792 address generate_f2l_fixup() {
793 StubCodeMark mark(this, "StubRoutines", "f2l_fixup");
794 Address inout(rsp, 5 * wordSize); // return address + 4 saves
795 address start = __ pc();
797 Label L;
799 __ push(rax);
800 __ push(c_rarg3);
801 __ push(c_rarg2);
802 __ push(c_rarg1);
804 __ movl(rax, 0x7f800000);
805 __ xorl(c_rarg3, c_rarg3);
806 __ movl(c_rarg2, inout);
807 __ movl(c_rarg1, c_rarg2);
808 __ andl(c_rarg1, 0x7fffffff);
809 __ cmpl(rax, c_rarg1); // NaN? -> 0
810 __ jcc(Assembler::negative, L);
811 __ testl(c_rarg2, c_rarg2); // signed ? min_jlong : max_jlong
812 __ mov64(c_rarg3, 0x8000000000000000);
813 __ mov64(rax, 0x7fffffffffffffff);
814 __ cmov(Assembler::positive, c_rarg3, rax);
816 __ bind(L);
817 __ movptr(inout, c_rarg3);
819 __ pop(c_rarg1);
820 __ pop(c_rarg2);
821 __ pop(c_rarg3);
822 __ pop(rax);
824 __ ret(0);
826 return start;
827 }
829 address generate_d2i_fixup() {
830 StubCodeMark mark(this, "StubRoutines", "d2i_fixup");
831 Address inout(rsp, 6 * wordSize); // return address + 5 saves
833 address start = __ pc();
835 Label L;
837 __ push(rax);
838 __ push(c_rarg3);
839 __ push(c_rarg2);
840 __ push(c_rarg1);
841 __ push(c_rarg0);
843 __ movl(rax, 0x7ff00000);
844 __ movq(c_rarg2, inout);
845 __ movl(c_rarg3, c_rarg2);
846 __ mov(c_rarg1, c_rarg2);
847 __ mov(c_rarg0, c_rarg2);
848 __ negl(c_rarg3);
849 __ shrptr(c_rarg1, 0x20);
850 __ orl(c_rarg3, c_rarg2);
851 __ andl(c_rarg1, 0x7fffffff);
852 __ xorl(c_rarg2, c_rarg2);
853 __ shrl(c_rarg3, 0x1f);
854 __ orl(c_rarg1, c_rarg3);
855 __ cmpl(rax, c_rarg1);
856 __ jcc(Assembler::negative, L); // NaN -> 0
857 __ testptr(c_rarg0, c_rarg0); // signed ? min_jint : max_jint
858 __ movl(c_rarg2, 0x80000000);
859 __ movl(rax, 0x7fffffff);
860 __ cmov(Assembler::positive, c_rarg2, rax);
862 __ bind(L);
863 __ movptr(inout, c_rarg2);
865 __ pop(c_rarg0);
866 __ pop(c_rarg1);
867 __ pop(c_rarg2);
868 __ pop(c_rarg3);
869 __ pop(rax);
871 __ ret(0);
873 return start;
874 }
876 address generate_d2l_fixup() {
877 StubCodeMark mark(this, "StubRoutines", "d2l_fixup");
878 Address inout(rsp, 6 * wordSize); // return address + 5 saves
880 address start = __ pc();
882 Label L;
884 __ push(rax);
885 __ push(c_rarg3);
886 __ push(c_rarg2);
887 __ push(c_rarg1);
888 __ push(c_rarg0);
890 __ movl(rax, 0x7ff00000);
891 __ movq(c_rarg2, inout);
892 __ movl(c_rarg3, c_rarg2);
893 __ mov(c_rarg1, c_rarg2);
894 __ mov(c_rarg0, c_rarg2);
895 __ negl(c_rarg3);
896 __ shrptr(c_rarg1, 0x20);
897 __ orl(c_rarg3, c_rarg2);
898 __ andl(c_rarg1, 0x7fffffff);
899 __ xorl(c_rarg2, c_rarg2);
900 __ shrl(c_rarg3, 0x1f);
901 __ orl(c_rarg1, c_rarg3);
902 __ cmpl(rax, c_rarg1);
903 __ jcc(Assembler::negative, L); // NaN -> 0
904 __ testq(c_rarg0, c_rarg0); // signed ? min_jlong : max_jlong
905 __ mov64(c_rarg2, 0x8000000000000000);
906 __ mov64(rax, 0x7fffffffffffffff);
907 __ cmovq(Assembler::positive, c_rarg2, rax);
909 __ bind(L);
910 __ movq(inout, c_rarg2);
912 __ pop(c_rarg0);
913 __ pop(c_rarg1);
914 __ pop(c_rarg2);
915 __ pop(c_rarg3);
916 __ pop(rax);
918 __ ret(0);
920 return start;
921 }
923 address generate_fp_mask(const char *stub_name, int64_t mask) {
924 __ align(CodeEntryAlignment);
925 StubCodeMark mark(this, "StubRoutines", stub_name);
926 address start = __ pc();
928 __ emit_data64( mask, relocInfo::none );
929 __ emit_data64( mask, relocInfo::none );
931 return start;
932 }
934 // The following routine generates a subroutine to throw an
935 // asynchronous UnknownError when an unsafe access gets a fault that
936 // could not be reasonably prevented by the programmer. (Example:
937 // SIGBUS/OBJERR.)
938 address generate_handler_for_unsafe_access() {
939 StubCodeMark mark(this, "StubRoutines", "handler_for_unsafe_access");
940 address start = __ pc();
942 __ push(0); // hole for return address-to-be
943 __ pusha(); // push registers
944 Address next_pc(rsp, RegisterImpl::number_of_registers * BytesPerWord);
946 // FIXME: this probably needs alignment logic
948 __ subptr(rsp, frame::arg_reg_save_area_bytes);
949 BLOCK_COMMENT("call handle_unsafe_access");
950 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, handle_unsafe_access)));
951 __ addptr(rsp, frame::arg_reg_save_area_bytes);
953 __ movptr(next_pc, rax); // stuff next address
954 __ popa();
955 __ ret(0); // jump to next address
957 return start;
958 }
960 // Non-destructive plausibility checks for oops
961 //
962 // Arguments:
963 // all args on stack!
964 //
965 // Stack after saving c_rarg3:
966 // [tos + 0]: saved c_rarg3
967 // [tos + 1]: saved c_rarg2
968 // [tos + 2]: saved r12 (several TemplateTable methods use it)
969 // [tos + 3]: saved flags
970 // [tos + 4]: return address
971 // * [tos + 5]: error message (char*)
972 // * [tos + 6]: object to verify (oop)
973 // * [tos + 7]: saved rax - saved by caller and bashed
974 // * [tos + 8]: saved r10 (rscratch1) - saved by caller
975 // * = popped on exit
976 address generate_verify_oop() {
977 StubCodeMark mark(this, "StubRoutines", "verify_oop");
978 address start = __ pc();
980 Label exit, error;
982 __ pushf();
983 __ incrementl(ExternalAddress((address) StubRoutines::verify_oop_count_addr()));
985 __ push(r12);
987 // save c_rarg2 and c_rarg3
988 __ push(c_rarg2);
989 __ push(c_rarg3);
991 enum {
992 // After previous pushes.
993 oop_to_verify = 6 * wordSize,
994 saved_rax = 7 * wordSize,
995 saved_r10 = 8 * wordSize,
997 // Before the call to MacroAssembler::debug(), see below.
998 return_addr = 16 * wordSize,
999 error_msg = 17 * wordSize
1000 };
1002 // get object
1003 __ movptr(rax, Address(rsp, oop_to_verify));
1005 // make sure object is 'reasonable'
1006 __ testptr(rax, rax);
1007 __ jcc(Assembler::zero, exit); // if obj is NULL it is OK
1008 // Check if the oop is in the right area of memory
1009 __ movptr(c_rarg2, rax);
1010 __ movptr(c_rarg3, (intptr_t) Universe::verify_oop_mask());
1011 __ andptr(c_rarg2, c_rarg3);
1012 __ movptr(c_rarg3, (intptr_t) Universe::verify_oop_bits());
1013 __ cmpptr(c_rarg2, c_rarg3);
1014 __ jcc(Assembler::notZero, error);
1016 // set r12 to heapbase for load_klass()
1017 __ reinit_heapbase();
1019 // make sure klass is 'reasonable', which is not zero.
1020 __ load_klass(rax, rax); // get klass
1021 __ testptr(rax, rax);
1022 __ jcc(Assembler::zero, error); // if klass is NULL it is broken
1023 // TODO: Future assert that klass is lower 4g memory for UseCompressedKlassPointers
1025 // return if everything seems ok
1026 __ bind(exit);
1027 __ movptr(rax, Address(rsp, saved_rax)); // get saved rax back
1028 __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back
1029 __ pop(c_rarg3); // restore c_rarg3
1030 __ pop(c_rarg2); // restore c_rarg2
1031 __ pop(r12); // restore r12
1032 __ popf(); // restore flags
1033 __ ret(4 * wordSize); // pop caller saved stuff
1035 // handle errors
1036 __ bind(error);
1037 __ movptr(rax, Address(rsp, saved_rax)); // get saved rax back
1038 __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back
1039 __ pop(c_rarg3); // get saved c_rarg3 back
1040 __ pop(c_rarg2); // get saved c_rarg2 back
1041 __ pop(r12); // get saved r12 back
1042 __ popf(); // get saved flags off stack --
1043 // will be ignored
1045 __ pusha(); // push registers
1046 // (rip is already
1047 // already pushed)
1048 // debug(char* msg, int64_t pc, int64_t regs[])
1049 // We've popped the registers we'd saved (c_rarg3, c_rarg2 and flags), and
1050 // pushed all the registers, so now the stack looks like:
1051 // [tos + 0] 16 saved registers
1052 // [tos + 16] return address
1053 // * [tos + 17] error message (char*)
1054 // * [tos + 18] object to verify (oop)
1055 // * [tos + 19] saved rax - saved by caller and bashed
1056 // * [tos + 20] saved r10 (rscratch1) - saved by caller
1057 // * = popped on exit
1059 __ movptr(c_rarg0, Address(rsp, error_msg)); // pass address of error message
1060 __ movptr(c_rarg1, Address(rsp, return_addr)); // pass return address
1061 __ movq(c_rarg2, rsp); // pass address of regs on stack
1062 __ mov(r12, rsp); // remember rsp
1063 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows
1064 __ andptr(rsp, -16); // align stack as required by ABI
1065 BLOCK_COMMENT("call MacroAssembler::debug");
1066 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64)));
1067 __ mov(rsp, r12); // restore rsp
1068 __ popa(); // pop registers (includes r12)
1069 __ ret(4 * wordSize); // pop caller saved stuff
1071 return start;
1072 }
1074 //
1075 // Verify that a register contains clean 32-bits positive value
1076 // (high 32-bits are 0) so it could be used in 64-bits shifts.
1077 //
1078 // Input:
1079 // Rint - 32-bits value
1080 // Rtmp - scratch
1081 //
1082 void assert_clean_int(Register Rint, Register Rtmp) {
1083 #ifdef ASSERT
1084 Label L;
1085 assert_different_registers(Rtmp, Rint);
1086 __ movslq(Rtmp, Rint);
1087 __ cmpq(Rtmp, Rint);
1088 __ jcc(Assembler::equal, L);
1089 __ stop("high 32-bits of int value are not 0");
1090 __ bind(L);
1091 #endif
1092 }
1094 // Generate overlap test for array copy stubs
1095 //
1096 // Input:
1097 // c_rarg0 - from
1098 // c_rarg1 - to
1099 // c_rarg2 - element count
1100 //
1101 // Output:
1102 // rax - &from[element count - 1]
1103 //
1104 void array_overlap_test(address no_overlap_target, Address::ScaleFactor sf) {
1105 assert(no_overlap_target != NULL, "must be generated");
1106 array_overlap_test(no_overlap_target, NULL, sf);
1107 }
1108 void array_overlap_test(Label& L_no_overlap, Address::ScaleFactor sf) {
1109 array_overlap_test(NULL, &L_no_overlap, sf);
1110 }
1111 void array_overlap_test(address no_overlap_target, Label* NOLp, Address::ScaleFactor sf) {
1112 const Register from = c_rarg0;
1113 const Register to = c_rarg1;
1114 const Register count = c_rarg2;
1115 const Register end_from = rax;
1117 __ cmpptr(to, from);
1118 __ lea(end_from, Address(from, count, sf, 0));
1119 if (NOLp == NULL) {
1120 ExternalAddress no_overlap(no_overlap_target);
1121 __ jump_cc(Assembler::belowEqual, no_overlap);
1122 __ cmpptr(to, end_from);
1123 __ jump_cc(Assembler::aboveEqual, no_overlap);
1124 } else {
1125 __ jcc(Assembler::belowEqual, (*NOLp));
1126 __ cmpptr(to, end_from);
1127 __ jcc(Assembler::aboveEqual, (*NOLp));
1128 }
1129 }
1131 // Shuffle first three arg regs on Windows into Linux/Solaris locations.
1132 //
1133 // Outputs:
1134 // rdi - rcx
1135 // rsi - rdx
1136 // rdx - r8
1137 // rcx - r9
1138 //
1139 // Registers r9 and r10 are used to save rdi and rsi on Windows, which latter
1140 // are non-volatile. r9 and r10 should not be used by the caller.
1141 //
1142 void setup_arg_regs(int nargs = 3) {
1143 const Register saved_rdi = r9;
1144 const Register saved_rsi = r10;
1145 assert(nargs == 3 || nargs == 4, "else fix");
1146 #ifdef _WIN64
1147 assert(c_rarg0 == rcx && c_rarg1 == rdx && c_rarg2 == r8 && c_rarg3 == r9,
1148 "unexpected argument registers");
1149 if (nargs >= 4)
1150 __ mov(rax, r9); // r9 is also saved_rdi
1151 __ movptr(saved_rdi, rdi);
1152 __ movptr(saved_rsi, rsi);
1153 __ mov(rdi, rcx); // c_rarg0
1154 __ mov(rsi, rdx); // c_rarg1
1155 __ mov(rdx, r8); // c_rarg2
1156 if (nargs >= 4)
1157 __ mov(rcx, rax); // c_rarg3 (via rax)
1158 #else
1159 assert(c_rarg0 == rdi && c_rarg1 == rsi && c_rarg2 == rdx && c_rarg3 == rcx,
1160 "unexpected argument registers");
1161 #endif
1162 }
1164 void restore_arg_regs() {
1165 const Register saved_rdi = r9;
1166 const Register saved_rsi = r10;
1167 #ifdef _WIN64
1168 __ movptr(rdi, saved_rdi);
1169 __ movptr(rsi, saved_rsi);
1170 #endif
1171 }
1173 // Generate code for an array write pre barrier
1174 //
1175 // addr - starting address
1176 // count - element count
1177 // tmp - scratch register
1178 //
1179 // Destroy no registers!
1180 //
1181 void gen_write_ref_array_pre_barrier(Register addr, Register count, bool dest_uninitialized) {
1182 BarrierSet* bs = Universe::heap()->barrier_set();
1183 switch (bs->kind()) {
1184 case BarrierSet::G1SATBCT:
1185 case BarrierSet::G1SATBCTLogging:
1186 // With G1, don't generate the call if we statically know that the target in uninitialized
1187 if (!dest_uninitialized) {
1188 __ pusha(); // push registers
1189 if (count == c_rarg0) {
1190 if (addr == c_rarg1) {
1191 // exactly backwards!!
1192 __ xchgptr(c_rarg1, c_rarg0);
1193 } else {
1194 __ movptr(c_rarg1, count);
1195 __ movptr(c_rarg0, addr);
1196 }
1197 } else {
1198 __ movptr(c_rarg0, addr);
1199 __ movptr(c_rarg1, count);
1200 }
1201 __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_pre), 2);
1202 __ popa();
1203 }
1204 break;
1205 case BarrierSet::CardTableModRef:
1206 case BarrierSet::CardTableExtension:
1207 case BarrierSet::ModRef:
1208 break;
1209 default:
1210 ShouldNotReachHere();
1212 }
1213 }
1215 //
1216 // Generate code for an array write post barrier
1217 //
1218 // Input:
1219 // start - register containing starting address of destination array
1220 // end - register containing ending address of destination array
1221 // scratch - scratch register
1222 //
1223 // The input registers are overwritten.
1224 // The ending address is inclusive.
1225 void gen_write_ref_array_post_barrier(Register start, Register end, Register scratch) {
1226 assert_different_registers(start, end, scratch);
1227 BarrierSet* bs = Universe::heap()->barrier_set();
1228 switch (bs->kind()) {
1229 case BarrierSet::G1SATBCT:
1230 case BarrierSet::G1SATBCTLogging:
1232 {
1233 __ pusha(); // push registers (overkill)
1234 // must compute element count unless barrier set interface is changed (other platforms supply count)
1235 assert_different_registers(start, end, scratch);
1236 __ lea(scratch, Address(end, BytesPerHeapOop));
1237 __ subptr(scratch, start); // subtract start to get #bytes
1238 __ shrptr(scratch, LogBytesPerHeapOop); // convert to element count
1239 __ mov(c_rarg0, start);
1240 __ mov(c_rarg1, scratch);
1241 __ call_VM_leaf(CAST_FROM_FN_PTR(address, BarrierSet::static_write_ref_array_post), 2);
1242 __ popa();
1243 }
1244 break;
1245 case BarrierSet::CardTableModRef:
1246 case BarrierSet::CardTableExtension:
1247 {
1248 CardTableModRefBS* ct = (CardTableModRefBS*)bs;
1249 assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "adjust this code");
1251 Label L_loop;
1253 __ shrptr(start, CardTableModRefBS::card_shift);
1254 __ addptr(end, BytesPerHeapOop);
1255 __ shrptr(end, CardTableModRefBS::card_shift);
1256 __ subptr(end, start); // number of bytes to copy
1258 intptr_t disp = (intptr_t) ct->byte_map_base;
1259 if (Assembler::is_simm32(disp)) {
1260 Address cardtable(noreg, noreg, Address::no_scale, disp);
1261 __ lea(scratch, cardtable);
1262 } else {
1263 ExternalAddress cardtable((address)disp);
1264 __ lea(scratch, cardtable);
1265 }
1267 const Register count = end; // 'end' register contains bytes count now
1268 __ addptr(start, scratch);
1269 __ BIND(L_loop);
1270 __ movb(Address(start, count, Address::times_1), 0);
1271 __ decrement(count);
1272 __ jcc(Assembler::greaterEqual, L_loop);
1273 }
1274 break;
1275 default:
1276 ShouldNotReachHere();
1278 }
1279 }
1282 // Copy big chunks forward
1283 //
1284 // Inputs:
1285 // end_from - source arrays end address
1286 // end_to - destination array end address
1287 // qword_count - 64-bits element count, negative
1288 // to - scratch
1289 // L_copy_bytes - entry label
1290 // L_copy_8_bytes - exit label
1291 //
1292 void copy_bytes_forward(Register end_from, Register end_to,
1293 Register qword_count, Register to,
1294 Label& L_copy_bytes, Label& L_copy_8_bytes) {
1295 DEBUG_ONLY(__ stop("enter at entry label, not here"));
1296 Label L_loop;
1297 __ align(OptoLoopAlignment);
1298 if (UseUnalignedLoadStores) {
1299 Label L_end;
1300 // Copy 64-bytes per iteration
1301 __ BIND(L_loop);
1302 if (UseAVX >= 2) {
1303 __ vmovdqu(xmm0, Address(end_from, qword_count, Address::times_8, -56));
1304 __ vmovdqu(Address(end_to, qword_count, Address::times_8, -56), xmm0);
1305 __ vmovdqu(xmm1, Address(end_from, qword_count, Address::times_8, -24));
1306 __ vmovdqu(Address(end_to, qword_count, Address::times_8, -24), xmm1);
1307 } else {
1308 __ movdqu(xmm0, Address(end_from, qword_count, Address::times_8, -56));
1309 __ movdqu(Address(end_to, qword_count, Address::times_8, -56), xmm0);
1310 __ movdqu(xmm1, Address(end_from, qword_count, Address::times_8, -40));
1311 __ movdqu(Address(end_to, qword_count, Address::times_8, -40), xmm1);
1312 __ movdqu(xmm2, Address(end_from, qword_count, Address::times_8, -24));
1313 __ movdqu(Address(end_to, qword_count, Address::times_8, -24), xmm2);
1314 __ movdqu(xmm3, Address(end_from, qword_count, Address::times_8, - 8));
1315 __ movdqu(Address(end_to, qword_count, Address::times_8, - 8), xmm3);
1316 }
1317 __ BIND(L_copy_bytes);
1318 __ addptr(qword_count, 8);
1319 __ jcc(Assembler::lessEqual, L_loop);
1320 __ subptr(qword_count, 4); // sub(8) and add(4)
1321 __ jccb(Assembler::greater, L_end);
1322 // Copy trailing 32 bytes
1323 if (UseAVX >= 2) {
1324 __ vmovdqu(xmm0, Address(end_from, qword_count, Address::times_8, -24));
1325 __ vmovdqu(Address(end_to, qword_count, Address::times_8, -24), xmm0);
1326 } else {
1327 __ movdqu(xmm0, Address(end_from, qword_count, Address::times_8, -24));
1328 __ movdqu(Address(end_to, qword_count, Address::times_8, -24), xmm0);
1329 __ movdqu(xmm1, Address(end_from, qword_count, Address::times_8, - 8));
1330 __ movdqu(Address(end_to, qword_count, Address::times_8, - 8), xmm1);
1331 }
1332 __ addptr(qword_count, 4);
1333 __ BIND(L_end);
1334 } else {
1335 // Copy 32-bytes per iteration
1336 __ BIND(L_loop);
1337 __ movq(to, Address(end_from, qword_count, Address::times_8, -24));
1338 __ movq(Address(end_to, qword_count, Address::times_8, -24), to);
1339 __ movq(to, Address(end_from, qword_count, Address::times_8, -16));
1340 __ movq(Address(end_to, qword_count, Address::times_8, -16), to);
1341 __ movq(to, Address(end_from, qword_count, Address::times_8, - 8));
1342 __ movq(Address(end_to, qword_count, Address::times_8, - 8), to);
1343 __ movq(to, Address(end_from, qword_count, Address::times_8, - 0));
1344 __ movq(Address(end_to, qword_count, Address::times_8, - 0), to);
1346 __ BIND(L_copy_bytes);
1347 __ addptr(qword_count, 4);
1348 __ jcc(Assembler::lessEqual, L_loop);
1349 }
1350 __ subptr(qword_count, 4);
1351 __ jcc(Assembler::less, L_copy_8_bytes); // Copy trailing qwords
1352 }
1354 // Copy big chunks backward
1355 //
1356 // Inputs:
1357 // from - source arrays address
1358 // dest - destination array address
1359 // qword_count - 64-bits element count
1360 // to - scratch
1361 // L_copy_bytes - entry label
1362 // L_copy_8_bytes - exit label
1363 //
1364 void copy_bytes_backward(Register from, Register dest,
1365 Register qword_count, Register to,
1366 Label& L_copy_bytes, Label& L_copy_8_bytes) {
1367 DEBUG_ONLY(__ stop("enter at entry label, not here"));
1368 Label L_loop;
1369 __ align(OptoLoopAlignment);
1370 if (UseUnalignedLoadStores) {
1371 Label L_end;
1372 // Copy 64-bytes per iteration
1373 __ BIND(L_loop);
1374 if (UseAVX >= 2) {
1375 __ vmovdqu(xmm0, Address(from, qword_count, Address::times_8, 32));
1376 __ vmovdqu(Address(dest, qword_count, Address::times_8, 32), xmm0);
1377 __ vmovdqu(xmm1, Address(from, qword_count, Address::times_8, 0));
1378 __ vmovdqu(Address(dest, qword_count, Address::times_8, 0), xmm1);
1379 } else {
1380 __ movdqu(xmm0, Address(from, qword_count, Address::times_8, 48));
1381 __ movdqu(Address(dest, qword_count, Address::times_8, 48), xmm0);
1382 __ movdqu(xmm1, Address(from, qword_count, Address::times_8, 32));
1383 __ movdqu(Address(dest, qword_count, Address::times_8, 32), xmm1);
1384 __ movdqu(xmm2, Address(from, qword_count, Address::times_8, 16));
1385 __ movdqu(Address(dest, qword_count, Address::times_8, 16), xmm2);
1386 __ movdqu(xmm3, Address(from, qword_count, Address::times_8, 0));
1387 __ movdqu(Address(dest, qword_count, Address::times_8, 0), xmm3);
1388 }
1389 __ BIND(L_copy_bytes);
1390 __ subptr(qword_count, 8);
1391 __ jcc(Assembler::greaterEqual, L_loop);
1393 __ addptr(qword_count, 4); // add(8) and sub(4)
1394 __ jccb(Assembler::less, L_end);
1395 // Copy trailing 32 bytes
1396 if (UseAVX >= 2) {
1397 __ vmovdqu(xmm0, Address(from, qword_count, Address::times_8, 0));
1398 __ vmovdqu(Address(dest, qword_count, Address::times_8, 0), xmm0);
1399 } else {
1400 __ movdqu(xmm0, Address(from, qword_count, Address::times_8, 16));
1401 __ movdqu(Address(dest, qword_count, Address::times_8, 16), xmm0);
1402 __ movdqu(xmm1, Address(from, qword_count, Address::times_8, 0));
1403 __ movdqu(Address(dest, qword_count, Address::times_8, 0), xmm1);
1404 }
1405 __ subptr(qword_count, 4);
1406 __ BIND(L_end);
1407 } else {
1408 // Copy 32-bytes per iteration
1409 __ BIND(L_loop);
1410 __ movq(to, Address(from, qword_count, Address::times_8, 24));
1411 __ movq(Address(dest, qword_count, Address::times_8, 24), to);
1412 __ movq(to, Address(from, qword_count, Address::times_8, 16));
1413 __ movq(Address(dest, qword_count, Address::times_8, 16), to);
1414 __ movq(to, Address(from, qword_count, Address::times_8, 8));
1415 __ movq(Address(dest, qword_count, Address::times_8, 8), to);
1416 __ movq(to, Address(from, qword_count, Address::times_8, 0));
1417 __ movq(Address(dest, qword_count, Address::times_8, 0), to);
1419 __ BIND(L_copy_bytes);
1420 __ subptr(qword_count, 4);
1421 __ jcc(Assembler::greaterEqual, L_loop);
1422 }
1423 __ addptr(qword_count, 4);
1424 __ jcc(Assembler::greater, L_copy_8_bytes); // Copy trailing qwords
1425 }
1428 // Arguments:
1429 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1430 // ignored
1431 // name - stub name string
1432 //
1433 // Inputs:
1434 // c_rarg0 - source array address
1435 // c_rarg1 - destination array address
1436 // c_rarg2 - element count, treated as ssize_t, can be zero
1437 //
1438 // If 'from' and/or 'to' are aligned on 4-, 2-, or 1-byte boundaries,
1439 // we let the hardware handle it. The one to eight bytes within words,
1440 // dwords or qwords that span cache line boundaries will still be loaded
1441 // and stored atomically.
1442 //
1443 // Side Effects:
1444 // disjoint_byte_copy_entry is set to the no-overlap entry point
1445 // used by generate_conjoint_byte_copy().
1446 //
1447 address generate_disjoint_byte_copy(bool aligned, address* entry, const char *name) {
1448 __ align(CodeEntryAlignment);
1449 StubCodeMark mark(this, "StubRoutines", name);
1450 address start = __ pc();
1452 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes, L_copy_2_bytes;
1453 Label L_copy_byte, L_exit;
1454 const Register from = rdi; // source array address
1455 const Register to = rsi; // destination array address
1456 const Register count = rdx; // elements count
1457 const Register byte_count = rcx;
1458 const Register qword_count = count;
1459 const Register end_from = from; // source array end address
1460 const Register end_to = to; // destination array end address
1461 // End pointers are inclusive, and if count is not zero they point
1462 // to the last unit copied: end_to[0] := end_from[0]
1464 __ enter(); // required for proper stackwalking of RuntimeStub frame
1465 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1467 if (entry != NULL) {
1468 *entry = __ pc();
1469 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1470 BLOCK_COMMENT("Entry:");
1471 }
1473 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1474 // r9 and r10 may be used to save non-volatile registers
1476 // 'from', 'to' and 'count' are now valid
1477 __ movptr(byte_count, count);
1478 __ shrptr(count, 3); // count => qword_count
1480 // Copy from low to high addresses. Use 'to' as scratch.
1481 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
1482 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
1483 __ negptr(qword_count); // make the count negative
1484 __ jmp(L_copy_bytes);
1486 // Copy trailing qwords
1487 __ BIND(L_copy_8_bytes);
1488 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
1489 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
1490 __ increment(qword_count);
1491 __ jcc(Assembler::notZero, L_copy_8_bytes);
1493 // Check for and copy trailing dword
1494 __ BIND(L_copy_4_bytes);
1495 __ testl(byte_count, 4);
1496 __ jccb(Assembler::zero, L_copy_2_bytes);
1497 __ movl(rax, Address(end_from, 8));
1498 __ movl(Address(end_to, 8), rax);
1500 __ addptr(end_from, 4);
1501 __ addptr(end_to, 4);
1503 // Check for and copy trailing word
1504 __ BIND(L_copy_2_bytes);
1505 __ testl(byte_count, 2);
1506 __ jccb(Assembler::zero, L_copy_byte);
1507 __ movw(rax, Address(end_from, 8));
1508 __ movw(Address(end_to, 8), rax);
1510 __ addptr(end_from, 2);
1511 __ addptr(end_to, 2);
1513 // Check for and copy trailing byte
1514 __ BIND(L_copy_byte);
1515 __ testl(byte_count, 1);
1516 __ jccb(Assembler::zero, L_exit);
1517 __ movb(rax, Address(end_from, 8));
1518 __ movb(Address(end_to, 8), rax);
1520 __ BIND(L_exit);
1521 restore_arg_regs();
1522 inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); // Update counter after rscratch1 is free
1523 __ xorptr(rax, rax); // return 0
1524 __ leave(); // required for proper stackwalking of RuntimeStub frame
1525 __ ret(0);
1527 // Copy in multi-bytes chunks
1528 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1529 __ jmp(L_copy_4_bytes);
1531 return start;
1532 }
1534 // Arguments:
1535 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1536 // ignored
1537 // name - stub name string
1538 //
1539 // Inputs:
1540 // c_rarg0 - source array address
1541 // c_rarg1 - destination array address
1542 // c_rarg2 - element count, treated as ssize_t, can be zero
1543 //
1544 // If 'from' and/or 'to' are aligned on 4-, 2-, or 1-byte boundaries,
1545 // we let the hardware handle it. The one to eight bytes within words,
1546 // dwords or qwords that span cache line boundaries will still be loaded
1547 // and stored atomically.
1548 //
1549 address generate_conjoint_byte_copy(bool aligned, address nooverlap_target,
1550 address* entry, const char *name) {
1551 __ align(CodeEntryAlignment);
1552 StubCodeMark mark(this, "StubRoutines", name);
1553 address start = __ pc();
1555 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes, L_copy_2_bytes;
1556 const Register from = rdi; // source array address
1557 const Register to = rsi; // destination array address
1558 const Register count = rdx; // elements count
1559 const Register byte_count = rcx;
1560 const Register qword_count = count;
1562 __ enter(); // required for proper stackwalking of RuntimeStub frame
1563 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1565 if (entry != NULL) {
1566 *entry = __ pc();
1567 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1568 BLOCK_COMMENT("Entry:");
1569 }
1571 array_overlap_test(nooverlap_target, Address::times_1);
1572 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1573 // r9 and r10 may be used to save non-volatile registers
1575 // 'from', 'to' and 'count' are now valid
1576 __ movptr(byte_count, count);
1577 __ shrptr(count, 3); // count => qword_count
1579 // Copy from high to low addresses.
1581 // Check for and copy trailing byte
1582 __ testl(byte_count, 1);
1583 __ jcc(Assembler::zero, L_copy_2_bytes);
1584 __ movb(rax, Address(from, byte_count, Address::times_1, -1));
1585 __ movb(Address(to, byte_count, Address::times_1, -1), rax);
1586 __ decrement(byte_count); // Adjust for possible trailing word
1588 // Check for and copy trailing word
1589 __ BIND(L_copy_2_bytes);
1590 __ testl(byte_count, 2);
1591 __ jcc(Assembler::zero, L_copy_4_bytes);
1592 __ movw(rax, Address(from, byte_count, Address::times_1, -2));
1593 __ movw(Address(to, byte_count, Address::times_1, -2), rax);
1595 // Check for and copy trailing dword
1596 __ BIND(L_copy_4_bytes);
1597 __ testl(byte_count, 4);
1598 __ jcc(Assembler::zero, L_copy_bytes);
1599 __ movl(rax, Address(from, qword_count, Address::times_8));
1600 __ movl(Address(to, qword_count, Address::times_8), rax);
1601 __ jmp(L_copy_bytes);
1603 // Copy trailing qwords
1604 __ BIND(L_copy_8_bytes);
1605 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
1606 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
1607 __ decrement(qword_count);
1608 __ jcc(Assembler::notZero, L_copy_8_bytes);
1610 restore_arg_regs();
1611 inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); // Update counter after rscratch1 is free
1612 __ xorptr(rax, rax); // return 0
1613 __ leave(); // required for proper stackwalking of RuntimeStub frame
1614 __ ret(0);
1616 // Copy in multi-bytes chunks
1617 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1619 restore_arg_regs();
1620 inc_counter_np(SharedRuntime::_jbyte_array_copy_ctr); // Update counter after rscratch1 is free
1621 __ xorptr(rax, rax); // return 0
1622 __ leave(); // required for proper stackwalking of RuntimeStub frame
1623 __ ret(0);
1625 return start;
1626 }
1628 // Arguments:
1629 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1630 // ignored
1631 // name - stub name string
1632 //
1633 // Inputs:
1634 // c_rarg0 - source array address
1635 // c_rarg1 - destination array address
1636 // c_rarg2 - element count, treated as ssize_t, can be zero
1637 //
1638 // If 'from' and/or 'to' are aligned on 4- or 2-byte boundaries, we
1639 // let the hardware handle it. The two or four words within dwords
1640 // or qwords that span cache line boundaries will still be loaded
1641 // and stored atomically.
1642 //
1643 // Side Effects:
1644 // disjoint_short_copy_entry is set to the no-overlap entry point
1645 // used by generate_conjoint_short_copy().
1646 //
1647 address generate_disjoint_short_copy(bool aligned, address *entry, const char *name) {
1648 __ align(CodeEntryAlignment);
1649 StubCodeMark mark(this, "StubRoutines", name);
1650 address start = __ pc();
1652 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes,L_copy_2_bytes,L_exit;
1653 const Register from = rdi; // source array address
1654 const Register to = rsi; // destination array address
1655 const Register count = rdx; // elements count
1656 const Register word_count = rcx;
1657 const Register qword_count = count;
1658 const Register end_from = from; // source array end address
1659 const Register end_to = to; // destination array end address
1660 // End pointers are inclusive, and if count is not zero they point
1661 // to the last unit copied: end_to[0] := end_from[0]
1663 __ enter(); // required for proper stackwalking of RuntimeStub frame
1664 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1666 if (entry != NULL) {
1667 *entry = __ pc();
1668 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1669 BLOCK_COMMENT("Entry:");
1670 }
1672 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1673 // r9 and r10 may be used to save non-volatile registers
1675 // 'from', 'to' and 'count' are now valid
1676 __ movptr(word_count, count);
1677 __ shrptr(count, 2); // count => qword_count
1679 // Copy from low to high addresses. Use 'to' as scratch.
1680 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
1681 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
1682 __ negptr(qword_count);
1683 __ jmp(L_copy_bytes);
1685 // Copy trailing qwords
1686 __ BIND(L_copy_8_bytes);
1687 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
1688 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
1689 __ increment(qword_count);
1690 __ jcc(Assembler::notZero, L_copy_8_bytes);
1692 // Original 'dest' is trashed, so we can't use it as a
1693 // base register for a possible trailing word copy
1695 // Check for and copy trailing dword
1696 __ BIND(L_copy_4_bytes);
1697 __ testl(word_count, 2);
1698 __ jccb(Assembler::zero, L_copy_2_bytes);
1699 __ movl(rax, Address(end_from, 8));
1700 __ movl(Address(end_to, 8), rax);
1702 __ addptr(end_from, 4);
1703 __ addptr(end_to, 4);
1705 // Check for and copy trailing word
1706 __ BIND(L_copy_2_bytes);
1707 __ testl(word_count, 1);
1708 __ jccb(Assembler::zero, L_exit);
1709 __ movw(rax, Address(end_from, 8));
1710 __ movw(Address(end_to, 8), rax);
1712 __ BIND(L_exit);
1713 restore_arg_regs();
1714 inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); // Update counter after rscratch1 is free
1715 __ xorptr(rax, rax); // return 0
1716 __ leave(); // required for proper stackwalking of RuntimeStub frame
1717 __ ret(0);
1719 // Copy in multi-bytes chunks
1720 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1721 __ jmp(L_copy_4_bytes);
1723 return start;
1724 }
1726 address generate_fill(BasicType t, bool aligned, const char *name) {
1727 __ align(CodeEntryAlignment);
1728 StubCodeMark mark(this, "StubRoutines", name);
1729 address start = __ pc();
1731 BLOCK_COMMENT("Entry:");
1733 const Register to = c_rarg0; // source array address
1734 const Register value = c_rarg1; // value
1735 const Register count = c_rarg2; // elements count
1737 __ enter(); // required for proper stackwalking of RuntimeStub frame
1739 __ generate_fill(t, aligned, to, value, count, rax, xmm0);
1741 __ leave(); // required for proper stackwalking of RuntimeStub frame
1742 __ ret(0);
1743 return start;
1744 }
1746 // Arguments:
1747 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1748 // ignored
1749 // name - stub name string
1750 //
1751 // Inputs:
1752 // c_rarg0 - source array address
1753 // c_rarg1 - destination array address
1754 // c_rarg2 - element count, treated as ssize_t, can be zero
1755 //
1756 // If 'from' and/or 'to' are aligned on 4- or 2-byte boundaries, we
1757 // let the hardware handle it. The two or four words within dwords
1758 // or qwords that span cache line boundaries will still be loaded
1759 // and stored atomically.
1760 //
1761 address generate_conjoint_short_copy(bool aligned, address nooverlap_target,
1762 address *entry, const char *name) {
1763 __ align(CodeEntryAlignment);
1764 StubCodeMark mark(this, "StubRoutines", name);
1765 address start = __ pc();
1767 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes;
1768 const Register from = rdi; // source array address
1769 const Register to = rsi; // destination array address
1770 const Register count = rdx; // elements count
1771 const Register word_count = rcx;
1772 const Register qword_count = count;
1774 __ enter(); // required for proper stackwalking of RuntimeStub frame
1775 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1777 if (entry != NULL) {
1778 *entry = __ pc();
1779 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1780 BLOCK_COMMENT("Entry:");
1781 }
1783 array_overlap_test(nooverlap_target, Address::times_2);
1784 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1785 // r9 and r10 may be used to save non-volatile registers
1787 // 'from', 'to' and 'count' are now valid
1788 __ movptr(word_count, count);
1789 __ shrptr(count, 2); // count => qword_count
1791 // Copy from high to low addresses. Use 'to' as scratch.
1793 // Check for and copy trailing word
1794 __ testl(word_count, 1);
1795 __ jccb(Assembler::zero, L_copy_4_bytes);
1796 __ movw(rax, Address(from, word_count, Address::times_2, -2));
1797 __ movw(Address(to, word_count, Address::times_2, -2), rax);
1799 // Check for and copy trailing dword
1800 __ BIND(L_copy_4_bytes);
1801 __ testl(word_count, 2);
1802 __ jcc(Assembler::zero, L_copy_bytes);
1803 __ movl(rax, Address(from, qword_count, Address::times_8));
1804 __ movl(Address(to, qword_count, Address::times_8), rax);
1805 __ jmp(L_copy_bytes);
1807 // Copy trailing qwords
1808 __ BIND(L_copy_8_bytes);
1809 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
1810 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
1811 __ decrement(qword_count);
1812 __ jcc(Assembler::notZero, L_copy_8_bytes);
1814 restore_arg_regs();
1815 inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); // Update counter after rscratch1 is free
1816 __ xorptr(rax, rax); // return 0
1817 __ leave(); // required for proper stackwalking of RuntimeStub frame
1818 __ ret(0);
1820 // Copy in multi-bytes chunks
1821 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1823 restore_arg_regs();
1824 inc_counter_np(SharedRuntime::_jshort_array_copy_ctr); // Update counter after rscratch1 is free
1825 __ xorptr(rax, rax); // return 0
1826 __ leave(); // required for proper stackwalking of RuntimeStub frame
1827 __ ret(0);
1829 return start;
1830 }
1832 // Arguments:
1833 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1834 // ignored
1835 // is_oop - true => oop array, so generate store check code
1836 // name - stub name string
1837 //
1838 // Inputs:
1839 // c_rarg0 - source array address
1840 // c_rarg1 - destination array address
1841 // c_rarg2 - element count, treated as ssize_t, can be zero
1842 //
1843 // If 'from' and/or 'to' are aligned on 4-byte boundaries, we let
1844 // the hardware handle it. The two dwords within qwords that span
1845 // cache line boundaries will still be loaded and stored atomicly.
1846 //
1847 // Side Effects:
1848 // disjoint_int_copy_entry is set to the no-overlap entry point
1849 // used by generate_conjoint_int_oop_copy().
1850 //
1851 address generate_disjoint_int_oop_copy(bool aligned, bool is_oop, address* entry,
1852 const char *name, bool dest_uninitialized = false) {
1853 __ align(CodeEntryAlignment);
1854 StubCodeMark mark(this, "StubRoutines", name);
1855 address start = __ pc();
1857 Label L_copy_bytes, L_copy_8_bytes, L_copy_4_bytes, L_exit;
1858 const Register from = rdi; // source array address
1859 const Register to = rsi; // destination array address
1860 const Register count = rdx; // elements count
1861 const Register dword_count = rcx;
1862 const Register qword_count = count;
1863 const Register end_from = from; // source array end address
1864 const Register end_to = to; // destination array end address
1865 const Register saved_to = r11; // saved destination array address
1866 // End pointers are inclusive, and if count is not zero they point
1867 // to the last unit copied: end_to[0] := end_from[0]
1869 __ enter(); // required for proper stackwalking of RuntimeStub frame
1870 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1872 if (entry != NULL) {
1873 *entry = __ pc();
1874 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1875 BLOCK_COMMENT("Entry:");
1876 }
1878 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1879 // r9 and r10 may be used to save non-volatile registers
1880 if (is_oop) {
1881 __ movq(saved_to, to);
1882 gen_write_ref_array_pre_barrier(to, count, dest_uninitialized);
1883 }
1885 // 'from', 'to' and 'count' are now valid
1886 __ movptr(dword_count, count);
1887 __ shrptr(count, 1); // count => qword_count
1889 // Copy from low to high addresses. Use 'to' as scratch.
1890 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
1891 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
1892 __ negptr(qword_count);
1893 __ jmp(L_copy_bytes);
1895 // Copy trailing qwords
1896 __ BIND(L_copy_8_bytes);
1897 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
1898 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
1899 __ increment(qword_count);
1900 __ jcc(Assembler::notZero, L_copy_8_bytes);
1902 // Check for and copy trailing dword
1903 __ BIND(L_copy_4_bytes);
1904 __ testl(dword_count, 1); // Only byte test since the value is 0 or 1
1905 __ jccb(Assembler::zero, L_exit);
1906 __ movl(rax, Address(end_from, 8));
1907 __ movl(Address(end_to, 8), rax);
1909 __ BIND(L_exit);
1910 if (is_oop) {
1911 __ leaq(end_to, Address(saved_to, dword_count, Address::times_4, -4));
1912 gen_write_ref_array_post_barrier(saved_to, end_to, rax);
1913 }
1914 restore_arg_regs();
1915 inc_counter_np(SharedRuntime::_jint_array_copy_ctr); // Update counter after rscratch1 is free
1916 __ xorptr(rax, rax); // return 0
1917 __ leave(); // required for proper stackwalking of RuntimeStub frame
1918 __ ret(0);
1920 // Copy in multi-bytes chunks
1921 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
1922 __ jmp(L_copy_4_bytes);
1924 return start;
1925 }
1927 // Arguments:
1928 // aligned - true => Input and output aligned on a HeapWord == 8-byte boundary
1929 // ignored
1930 // is_oop - true => oop array, so generate store check code
1931 // name - stub name string
1932 //
1933 // Inputs:
1934 // c_rarg0 - source array address
1935 // c_rarg1 - destination array address
1936 // c_rarg2 - element count, treated as ssize_t, can be zero
1937 //
1938 // If 'from' and/or 'to' are aligned on 4-byte boundaries, we let
1939 // the hardware handle it. The two dwords within qwords that span
1940 // cache line boundaries will still be loaded and stored atomicly.
1941 //
1942 address generate_conjoint_int_oop_copy(bool aligned, bool is_oop, address nooverlap_target,
1943 address *entry, const char *name,
1944 bool dest_uninitialized = false) {
1945 __ align(CodeEntryAlignment);
1946 StubCodeMark mark(this, "StubRoutines", name);
1947 address start = __ pc();
1949 Label L_copy_bytes, L_copy_8_bytes, L_copy_2_bytes, L_exit;
1950 const Register from = rdi; // source array address
1951 const Register to = rsi; // destination array address
1952 const Register count = rdx; // elements count
1953 const Register dword_count = rcx;
1954 const Register qword_count = count;
1956 __ enter(); // required for proper stackwalking of RuntimeStub frame
1957 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
1959 if (entry != NULL) {
1960 *entry = __ pc();
1961 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
1962 BLOCK_COMMENT("Entry:");
1963 }
1965 array_overlap_test(nooverlap_target, Address::times_4);
1966 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1967 // r9 and r10 may be used to save non-volatile registers
1969 if (is_oop) {
1970 // no registers are destroyed by this call
1971 gen_write_ref_array_pre_barrier(to, count, dest_uninitialized);
1972 }
1974 assert_clean_int(count, rax); // Make sure 'count' is clean int.
1975 // 'from', 'to' and 'count' are now valid
1976 __ movptr(dword_count, count);
1977 __ shrptr(count, 1); // count => qword_count
1979 // Copy from high to low addresses. Use 'to' as scratch.
1981 // Check for and copy trailing dword
1982 __ testl(dword_count, 1);
1983 __ jcc(Assembler::zero, L_copy_bytes);
1984 __ movl(rax, Address(from, dword_count, Address::times_4, -4));
1985 __ movl(Address(to, dword_count, Address::times_4, -4), rax);
1986 __ jmp(L_copy_bytes);
1988 // Copy trailing qwords
1989 __ BIND(L_copy_8_bytes);
1990 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
1991 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
1992 __ decrement(qword_count);
1993 __ jcc(Assembler::notZero, L_copy_8_bytes);
1995 if (is_oop) {
1996 __ jmp(L_exit);
1997 }
1998 restore_arg_regs();
1999 inc_counter_np(SharedRuntime::_jint_array_copy_ctr); // Update counter after rscratch1 is free
2000 __ xorptr(rax, rax); // return 0
2001 __ leave(); // required for proper stackwalking of RuntimeStub frame
2002 __ ret(0);
2004 // Copy in multi-bytes chunks
2005 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
2007 __ bind(L_exit);
2008 if (is_oop) {
2009 Register end_to = rdx;
2010 __ leaq(end_to, Address(to, dword_count, Address::times_4, -4));
2011 gen_write_ref_array_post_barrier(to, end_to, rax);
2012 }
2013 restore_arg_regs();
2014 inc_counter_np(SharedRuntime::_jint_array_copy_ctr); // Update counter after rscratch1 is free
2015 __ xorptr(rax, rax); // return 0
2016 __ leave(); // required for proper stackwalking of RuntimeStub frame
2017 __ ret(0);
2019 return start;
2020 }
2022 // Arguments:
2023 // aligned - true => Input and output aligned on a HeapWord boundary == 8 bytes
2024 // ignored
2025 // is_oop - true => oop array, so generate store check code
2026 // name - stub name string
2027 //
2028 // Inputs:
2029 // c_rarg0 - source array address
2030 // c_rarg1 - destination array address
2031 // c_rarg2 - element count, treated as ssize_t, can be zero
2032 //
2033 // Side Effects:
2034 // disjoint_oop_copy_entry or disjoint_long_copy_entry is set to the
2035 // no-overlap entry point used by generate_conjoint_long_oop_copy().
2036 //
2037 address generate_disjoint_long_oop_copy(bool aligned, bool is_oop, address *entry,
2038 const char *name, bool dest_uninitialized = false) {
2039 __ align(CodeEntryAlignment);
2040 StubCodeMark mark(this, "StubRoutines", name);
2041 address start = __ pc();
2043 Label L_copy_bytes, L_copy_8_bytes, L_exit;
2044 const Register from = rdi; // source array address
2045 const Register to = rsi; // destination array address
2046 const Register qword_count = rdx; // elements count
2047 const Register end_from = from; // source array end address
2048 const Register end_to = rcx; // destination array end address
2049 const Register saved_to = to;
2050 // End pointers are inclusive, and if count is not zero they point
2051 // to the last unit copied: end_to[0] := end_from[0]
2053 __ enter(); // required for proper stackwalking of RuntimeStub frame
2054 // Save no-overlap entry point for generate_conjoint_long_oop_copy()
2055 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
2057 if (entry != NULL) {
2058 *entry = __ pc();
2059 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
2060 BLOCK_COMMENT("Entry:");
2061 }
2063 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
2064 // r9 and r10 may be used to save non-volatile registers
2065 // 'from', 'to' and 'qword_count' are now valid
2066 if (is_oop) {
2067 // no registers are destroyed by this call
2068 gen_write_ref_array_pre_barrier(to, qword_count, dest_uninitialized);
2069 }
2071 // Copy from low to high addresses. Use 'to' as scratch.
2072 __ lea(end_from, Address(from, qword_count, Address::times_8, -8));
2073 __ lea(end_to, Address(to, qword_count, Address::times_8, -8));
2074 __ negptr(qword_count);
2075 __ jmp(L_copy_bytes);
2077 // Copy trailing qwords
2078 __ BIND(L_copy_8_bytes);
2079 __ movq(rax, Address(end_from, qword_count, Address::times_8, 8));
2080 __ movq(Address(end_to, qword_count, Address::times_8, 8), rax);
2081 __ increment(qword_count);
2082 __ jcc(Assembler::notZero, L_copy_8_bytes);
2084 if (is_oop) {
2085 __ jmp(L_exit);
2086 } else {
2087 restore_arg_regs();
2088 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2089 __ xorptr(rax, rax); // return 0
2090 __ leave(); // required for proper stackwalking of RuntimeStub frame
2091 __ ret(0);
2092 }
2094 // Copy in multi-bytes chunks
2095 copy_bytes_forward(end_from, end_to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
2097 if (is_oop) {
2098 __ BIND(L_exit);
2099 gen_write_ref_array_post_barrier(saved_to, end_to, rax);
2100 }
2101 restore_arg_regs();
2102 if (is_oop) {
2103 inc_counter_np(SharedRuntime::_oop_array_copy_ctr); // Update counter after rscratch1 is free
2104 } else {
2105 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2106 }
2107 __ xorptr(rax, rax); // return 0
2108 __ leave(); // required for proper stackwalking of RuntimeStub frame
2109 __ ret(0);
2111 return start;
2112 }
2114 // Arguments:
2115 // aligned - true => Input and output aligned on a HeapWord boundary == 8 bytes
2116 // ignored
2117 // is_oop - true => oop array, so generate store check code
2118 // name - stub name string
2119 //
2120 // Inputs:
2121 // c_rarg0 - source array address
2122 // c_rarg1 - destination array address
2123 // c_rarg2 - element count, treated as ssize_t, can be zero
2124 //
2125 address generate_conjoint_long_oop_copy(bool aligned, bool is_oop,
2126 address nooverlap_target, address *entry,
2127 const char *name, bool dest_uninitialized = false) {
2128 __ align(CodeEntryAlignment);
2129 StubCodeMark mark(this, "StubRoutines", name);
2130 address start = __ pc();
2132 Label L_copy_bytes, L_copy_8_bytes, L_exit;
2133 const Register from = rdi; // source array address
2134 const Register to = rsi; // destination array address
2135 const Register qword_count = rdx; // elements count
2136 const Register saved_count = rcx;
2138 __ enter(); // required for proper stackwalking of RuntimeStub frame
2139 assert_clean_int(c_rarg2, rax); // Make sure 'count' is clean int.
2141 if (entry != NULL) {
2142 *entry = __ pc();
2143 // caller can pass a 64-bit byte count here (from Unsafe.copyMemory)
2144 BLOCK_COMMENT("Entry:");
2145 }
2147 array_overlap_test(nooverlap_target, Address::times_8);
2148 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
2149 // r9 and r10 may be used to save non-volatile registers
2150 // 'from', 'to' and 'qword_count' are now valid
2151 if (is_oop) {
2152 // Save to and count for store barrier
2153 __ movptr(saved_count, qword_count);
2154 // No registers are destroyed by this call
2155 gen_write_ref_array_pre_barrier(to, saved_count, dest_uninitialized);
2156 }
2158 __ jmp(L_copy_bytes);
2160 // Copy trailing qwords
2161 __ BIND(L_copy_8_bytes);
2162 __ movq(rax, Address(from, qword_count, Address::times_8, -8));
2163 __ movq(Address(to, qword_count, Address::times_8, -8), rax);
2164 __ decrement(qword_count);
2165 __ jcc(Assembler::notZero, L_copy_8_bytes);
2167 if (is_oop) {
2168 __ jmp(L_exit);
2169 } else {
2170 restore_arg_regs();
2171 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2172 __ xorptr(rax, rax); // return 0
2173 __ leave(); // required for proper stackwalking of RuntimeStub frame
2174 __ ret(0);
2175 }
2177 // Copy in multi-bytes chunks
2178 copy_bytes_backward(from, to, qword_count, rax, L_copy_bytes, L_copy_8_bytes);
2180 if (is_oop) {
2181 __ BIND(L_exit);
2182 __ lea(rcx, Address(to, saved_count, Address::times_8, -8));
2183 gen_write_ref_array_post_barrier(to, rcx, rax);
2184 }
2185 restore_arg_regs();
2186 if (is_oop) {
2187 inc_counter_np(SharedRuntime::_oop_array_copy_ctr); // Update counter after rscratch1 is free
2188 } else {
2189 inc_counter_np(SharedRuntime::_jlong_array_copy_ctr); // Update counter after rscratch1 is free
2190 }
2191 __ xorptr(rax, rax); // return 0
2192 __ leave(); // required for proper stackwalking of RuntimeStub frame
2193 __ ret(0);
2195 return start;
2196 }
2199 // Helper for generating a dynamic type check.
2200 // Smashes no registers.
2201 void generate_type_check(Register sub_klass,
2202 Register super_check_offset,
2203 Register super_klass,
2204 Label& L_success) {
2205 assert_different_registers(sub_klass, super_check_offset, super_klass);
2207 BLOCK_COMMENT("type_check:");
2209 Label L_miss;
2211 __ check_klass_subtype_fast_path(sub_klass, super_klass, noreg, &L_success, &L_miss, NULL,
2212 super_check_offset);
2213 __ check_klass_subtype_slow_path(sub_klass, super_klass, noreg, noreg, &L_success, NULL);
2215 // Fall through on failure!
2216 __ BIND(L_miss);
2217 }
2219 //
2220 // Generate checkcasting array copy stub
2221 //
2222 // Input:
2223 // c_rarg0 - source array address
2224 // c_rarg1 - destination array address
2225 // c_rarg2 - element count, treated as ssize_t, can be zero
2226 // c_rarg3 - size_t ckoff (super_check_offset)
2227 // not Win64
2228 // c_rarg4 - oop ckval (super_klass)
2229 // Win64
2230 // rsp+40 - oop ckval (super_klass)
2231 //
2232 // Output:
2233 // rax == 0 - success
2234 // rax == -1^K - failure, where K is partial transfer count
2235 //
2236 address generate_checkcast_copy(const char *name, address *entry,
2237 bool dest_uninitialized = false) {
2239 Label L_load_element, L_store_element, L_do_card_marks, L_done;
2241 // Input registers (after setup_arg_regs)
2242 const Register from = rdi; // source array address
2243 const Register to = rsi; // destination array address
2244 const Register length = rdx; // elements count
2245 const Register ckoff = rcx; // super_check_offset
2246 const Register ckval = r8; // super_klass
2248 // Registers used as temps (r13, r14 are save-on-entry)
2249 const Register end_from = from; // source array end address
2250 const Register end_to = r13; // destination array end address
2251 const Register count = rdx; // -(count_remaining)
2252 const Register r14_length = r14; // saved copy of length
2253 // End pointers are inclusive, and if length is not zero they point
2254 // to the last unit copied: end_to[0] := end_from[0]
2256 const Register rax_oop = rax; // actual oop copied
2257 const Register r11_klass = r11; // oop._klass
2259 //---------------------------------------------------------------
2260 // Assembler stub will be used for this call to arraycopy
2261 // if the two arrays are subtypes of Object[] but the
2262 // destination array type is not equal to or a supertype
2263 // of the source type. Each element must be separately
2264 // checked.
2266 __ align(CodeEntryAlignment);
2267 StubCodeMark mark(this, "StubRoutines", name);
2268 address start = __ pc();
2270 __ enter(); // required for proper stackwalking of RuntimeStub frame
2272 #ifdef ASSERT
2273 // caller guarantees that the arrays really are different
2274 // otherwise, we would have to make conjoint checks
2275 { Label L;
2276 array_overlap_test(L, TIMES_OOP);
2277 __ stop("checkcast_copy within a single array");
2278 __ bind(L);
2279 }
2280 #endif //ASSERT
2282 setup_arg_regs(4); // from => rdi, to => rsi, length => rdx
2283 // ckoff => rcx, ckval => r8
2284 // r9 and r10 may be used to save non-volatile registers
2285 #ifdef _WIN64
2286 // last argument (#4) is on stack on Win64
2287 __ movptr(ckval, Address(rsp, 6 * wordSize));
2288 #endif
2290 // Caller of this entry point must set up the argument registers.
2291 if (entry != NULL) {
2292 *entry = __ pc();
2293 BLOCK_COMMENT("Entry:");
2294 }
2296 // allocate spill slots for r13, r14
2297 enum {
2298 saved_r13_offset,
2299 saved_r14_offset,
2300 saved_rbp_offset
2301 };
2302 __ subptr(rsp, saved_rbp_offset * wordSize);
2303 __ movptr(Address(rsp, saved_r13_offset * wordSize), r13);
2304 __ movptr(Address(rsp, saved_r14_offset * wordSize), r14);
2306 // check that int operands are properly extended to size_t
2307 assert_clean_int(length, rax);
2308 assert_clean_int(ckoff, rax);
2310 #ifdef ASSERT
2311 BLOCK_COMMENT("assert consistent ckoff/ckval");
2312 // The ckoff and ckval must be mutually consistent,
2313 // even though caller generates both.
2314 { Label L;
2315 int sco_offset = in_bytes(Klass::super_check_offset_offset());
2316 __ cmpl(ckoff, Address(ckval, sco_offset));
2317 __ jcc(Assembler::equal, L);
2318 __ stop("super_check_offset inconsistent");
2319 __ bind(L);
2320 }
2321 #endif //ASSERT
2323 // Loop-invariant addresses. They are exclusive end pointers.
2324 Address end_from_addr(from, length, TIMES_OOP, 0);
2325 Address end_to_addr(to, length, TIMES_OOP, 0);
2326 // Loop-variant addresses. They assume post-incremented count < 0.
2327 Address from_element_addr(end_from, count, TIMES_OOP, 0);
2328 Address to_element_addr(end_to, count, TIMES_OOP, 0);
2330 gen_write_ref_array_pre_barrier(to, count, dest_uninitialized);
2332 // Copy from low to high addresses, indexed from the end of each array.
2333 __ lea(end_from, end_from_addr);
2334 __ lea(end_to, end_to_addr);
2335 __ movptr(r14_length, length); // save a copy of the length
2336 assert(length == count, ""); // else fix next line:
2337 __ negptr(count); // negate and test the length
2338 __ jcc(Assembler::notZero, L_load_element);
2340 // Empty array: Nothing to do.
2341 __ xorptr(rax, rax); // return 0 on (trivial) success
2342 __ jmp(L_done);
2344 // ======== begin loop ========
2345 // (Loop is rotated; its entry is L_load_element.)
2346 // Loop control:
2347 // for (count = -count; count != 0; count++)
2348 // Base pointers src, dst are biased by 8*(count-1),to last element.
2349 __ align(OptoLoopAlignment);
2351 __ BIND(L_store_element);
2352 __ store_heap_oop(to_element_addr, rax_oop); // store the oop
2353 __ increment(count); // increment the count toward zero
2354 __ jcc(Assembler::zero, L_do_card_marks);
2356 // ======== loop entry is here ========
2357 __ BIND(L_load_element);
2358 __ load_heap_oop(rax_oop, from_element_addr); // load the oop
2359 __ testptr(rax_oop, rax_oop);
2360 __ jcc(Assembler::zero, L_store_element);
2362 __ load_klass(r11_klass, rax_oop);// query the object klass
2363 generate_type_check(r11_klass, ckoff, ckval, L_store_element);
2364 // ======== end loop ========
2366 // It was a real error; we must depend on the caller to finish the job.
2367 // Register rdx = -1 * number of *remaining* oops, r14 = *total* oops.
2368 // Emit GC store barriers for the oops we have copied (r14 + rdx),
2369 // and report their number to the caller.
2370 assert_different_registers(rax, r14_length, count, to, end_to, rcx);
2371 __ lea(end_to, to_element_addr);
2372 __ addptr(end_to, -heapOopSize); // make an inclusive end pointer
2373 gen_write_ref_array_post_barrier(to, end_to, rscratch1);
2374 __ movptr(rax, r14_length); // original oops
2375 __ addptr(rax, count); // K = (original - remaining) oops
2376 __ notptr(rax); // report (-1^K) to caller
2377 __ jmp(L_done);
2379 // Come here on success only.
2380 __ BIND(L_do_card_marks);
2381 __ addptr(end_to, -heapOopSize); // make an inclusive end pointer
2382 gen_write_ref_array_post_barrier(to, end_to, rscratch1);
2383 __ xorptr(rax, rax); // return 0 on success
2385 // Common exit point (success or failure).
2386 __ BIND(L_done);
2387 __ movptr(r13, Address(rsp, saved_r13_offset * wordSize));
2388 __ movptr(r14, Address(rsp, saved_r14_offset * wordSize));
2389 restore_arg_regs();
2390 inc_counter_np(SharedRuntime::_checkcast_array_copy_ctr); // Update counter after rscratch1 is free
2391 __ leave(); // required for proper stackwalking of RuntimeStub frame
2392 __ ret(0);
2394 return start;
2395 }
2397 //
2398 // Generate 'unsafe' array copy stub
2399 // Though just as safe as the other stubs, it takes an unscaled
2400 // size_t argument instead of an element count.
2401 //
2402 // Input:
2403 // c_rarg0 - source array address
2404 // c_rarg1 - destination array address
2405 // c_rarg2 - byte count, treated as ssize_t, can be zero
2406 //
2407 // Examines the alignment of the operands and dispatches
2408 // to a long, int, short, or byte copy loop.
2409 //
2410 address generate_unsafe_copy(const char *name,
2411 address byte_copy_entry, address short_copy_entry,
2412 address int_copy_entry, address long_copy_entry) {
2414 Label L_long_aligned, L_int_aligned, L_short_aligned;
2416 // Input registers (before setup_arg_regs)
2417 const Register from = c_rarg0; // source array address
2418 const Register to = c_rarg1; // destination array address
2419 const Register size = c_rarg2; // byte count (size_t)
2421 // Register used as a temp
2422 const Register bits = rax; // test copy of low bits
2424 __ align(CodeEntryAlignment);
2425 StubCodeMark mark(this, "StubRoutines", name);
2426 address start = __ pc();
2428 __ enter(); // required for proper stackwalking of RuntimeStub frame
2430 // bump this on entry, not on exit:
2431 inc_counter_np(SharedRuntime::_unsafe_array_copy_ctr);
2433 __ mov(bits, from);
2434 __ orptr(bits, to);
2435 __ orptr(bits, size);
2437 __ testb(bits, BytesPerLong-1);
2438 __ jccb(Assembler::zero, L_long_aligned);
2440 __ testb(bits, BytesPerInt-1);
2441 __ jccb(Assembler::zero, L_int_aligned);
2443 __ testb(bits, BytesPerShort-1);
2444 __ jump_cc(Assembler::notZero, RuntimeAddress(byte_copy_entry));
2446 __ BIND(L_short_aligned);
2447 __ shrptr(size, LogBytesPerShort); // size => short_count
2448 __ jump(RuntimeAddress(short_copy_entry));
2450 __ BIND(L_int_aligned);
2451 __ shrptr(size, LogBytesPerInt); // size => int_count
2452 __ jump(RuntimeAddress(int_copy_entry));
2454 __ BIND(L_long_aligned);
2455 __ shrptr(size, LogBytesPerLong); // size => qword_count
2456 __ jump(RuntimeAddress(long_copy_entry));
2458 return start;
2459 }
2461 // Perform range checks on the proposed arraycopy.
2462 // Kills temp, but nothing else.
2463 // Also, clean the sign bits of src_pos and dst_pos.
2464 void arraycopy_range_checks(Register src, // source array oop (c_rarg0)
2465 Register src_pos, // source position (c_rarg1)
2466 Register dst, // destination array oo (c_rarg2)
2467 Register dst_pos, // destination position (c_rarg3)
2468 Register length,
2469 Register temp,
2470 Label& L_failed) {
2471 BLOCK_COMMENT("arraycopy_range_checks:");
2473 // if (src_pos + length > arrayOop(src)->length()) FAIL;
2474 __ movl(temp, length);
2475 __ addl(temp, src_pos); // src_pos + length
2476 __ cmpl(temp, Address(src, arrayOopDesc::length_offset_in_bytes()));
2477 __ jcc(Assembler::above, L_failed);
2479 // if (dst_pos + length > arrayOop(dst)->length()) FAIL;
2480 __ movl(temp, length);
2481 __ addl(temp, dst_pos); // dst_pos + length
2482 __ cmpl(temp, Address(dst, arrayOopDesc::length_offset_in_bytes()));
2483 __ jcc(Assembler::above, L_failed);
2485 // Have to clean up high 32-bits of 'src_pos' and 'dst_pos'.
2486 // Move with sign extension can be used since they are positive.
2487 __ movslq(src_pos, src_pos);
2488 __ movslq(dst_pos, dst_pos);
2490 BLOCK_COMMENT("arraycopy_range_checks done");
2491 }
2493 //
2494 // Generate generic array copy stubs
2495 //
2496 // Input:
2497 // c_rarg0 - src oop
2498 // c_rarg1 - src_pos (32-bits)
2499 // c_rarg2 - dst oop
2500 // c_rarg3 - dst_pos (32-bits)
2501 // not Win64
2502 // c_rarg4 - element count (32-bits)
2503 // Win64
2504 // rsp+40 - element count (32-bits)
2505 //
2506 // Output:
2507 // rax == 0 - success
2508 // rax == -1^K - failure, where K is partial transfer count
2509 //
2510 address generate_generic_copy(const char *name,
2511 address byte_copy_entry, address short_copy_entry,
2512 address int_copy_entry, address oop_copy_entry,
2513 address long_copy_entry, address checkcast_copy_entry) {
2515 Label L_failed, L_failed_0, L_objArray;
2516 Label L_copy_bytes, L_copy_shorts, L_copy_ints, L_copy_longs;
2518 // Input registers
2519 const Register src = c_rarg0; // source array oop
2520 const Register src_pos = c_rarg1; // source position
2521 const Register dst = c_rarg2; // destination array oop
2522 const Register dst_pos = c_rarg3; // destination position
2523 #ifndef _WIN64
2524 const Register length = c_rarg4;
2525 #else
2526 const Address length(rsp, 6 * wordSize); // elements count is on stack on Win64
2527 #endif
2529 { int modulus = CodeEntryAlignment;
2530 int target = modulus - 5; // 5 = sizeof jmp(L_failed)
2531 int advance = target - (__ offset() % modulus);
2532 if (advance < 0) advance += modulus;
2533 if (advance > 0) __ nop(advance);
2534 }
2535 StubCodeMark mark(this, "StubRoutines", name);
2537 // Short-hop target to L_failed. Makes for denser prologue code.
2538 __ BIND(L_failed_0);
2539 __ jmp(L_failed);
2540 assert(__ offset() % CodeEntryAlignment == 0, "no further alignment needed");
2542 __ align(CodeEntryAlignment);
2543 address start = __ pc();
2545 __ enter(); // required for proper stackwalking of RuntimeStub frame
2547 // bump this on entry, not on exit:
2548 inc_counter_np(SharedRuntime::_generic_array_copy_ctr);
2550 //-----------------------------------------------------------------------
2551 // Assembler stub will be used for this call to arraycopy
2552 // if the following conditions are met:
2553 //
2554 // (1) src and dst must not be null.
2555 // (2) src_pos must not be negative.
2556 // (3) dst_pos must not be negative.
2557 // (4) length must not be negative.
2558 // (5) src klass and dst klass should be the same and not NULL.
2559 // (6) src and dst should be arrays.
2560 // (7) src_pos + length must not exceed length of src.
2561 // (8) dst_pos + length must not exceed length of dst.
2562 //
2564 // if (src == NULL) return -1;
2565 __ testptr(src, src); // src oop
2566 size_t j1off = __ offset();
2567 __ jccb(Assembler::zero, L_failed_0);
2569 // if (src_pos < 0) return -1;
2570 __ testl(src_pos, src_pos); // src_pos (32-bits)
2571 __ jccb(Assembler::negative, L_failed_0);
2573 // if (dst == NULL) return -1;
2574 __ testptr(dst, dst); // dst oop
2575 __ jccb(Assembler::zero, L_failed_0);
2577 // if (dst_pos < 0) return -1;
2578 __ testl(dst_pos, dst_pos); // dst_pos (32-bits)
2579 size_t j4off = __ offset();
2580 __ jccb(Assembler::negative, L_failed_0);
2582 // The first four tests are very dense code,
2583 // but not quite dense enough to put four
2584 // jumps in a 16-byte instruction fetch buffer.
2585 // That's good, because some branch predicters
2586 // do not like jumps so close together.
2587 // Make sure of this.
2588 guarantee(((j1off ^ j4off) & ~15) != 0, "I$ line of 1st & 4th jumps");
2590 // registers used as temp
2591 const Register r11_length = r11; // elements count to copy
2592 const Register r10_src_klass = r10; // array klass
2594 // if (length < 0) return -1;
2595 __ movl(r11_length, length); // length (elements count, 32-bits value)
2596 __ testl(r11_length, r11_length);
2597 __ jccb(Assembler::negative, L_failed_0);
2599 __ load_klass(r10_src_klass, src);
2600 #ifdef ASSERT
2601 // assert(src->klass() != NULL);
2602 {
2603 BLOCK_COMMENT("assert klasses not null {");
2604 Label L1, L2;
2605 __ testptr(r10_src_klass, r10_src_klass);
2606 __ jcc(Assembler::notZero, L2); // it is broken if klass is NULL
2607 __ bind(L1);
2608 __ stop("broken null klass");
2609 __ bind(L2);
2610 __ load_klass(rax, dst);
2611 __ cmpq(rax, 0);
2612 __ jcc(Assembler::equal, L1); // this would be broken also
2613 BLOCK_COMMENT("} assert klasses not null done");
2614 }
2615 #endif
2617 // Load layout helper (32-bits)
2618 //
2619 // |array_tag| | header_size | element_type | |log2_element_size|
2620 // 32 30 24 16 8 2 0
2621 //
2622 // array_tag: typeArray = 0x3, objArray = 0x2, non-array = 0x0
2623 //
2625 const int lh_offset = in_bytes(Klass::layout_helper_offset());
2627 // Handle objArrays completely differently...
2628 const jint objArray_lh = Klass::array_layout_helper(T_OBJECT);
2629 __ cmpl(Address(r10_src_klass, lh_offset), objArray_lh);
2630 __ jcc(Assembler::equal, L_objArray);
2632 // if (src->klass() != dst->klass()) return -1;
2633 __ load_klass(rax, dst);
2634 __ cmpq(r10_src_klass, rax);
2635 __ jcc(Assembler::notEqual, L_failed);
2637 const Register rax_lh = rax; // layout helper
2638 __ movl(rax_lh, Address(r10_src_klass, lh_offset));
2640 // if (!src->is_Array()) return -1;
2641 __ cmpl(rax_lh, Klass::_lh_neutral_value);
2642 __ jcc(Assembler::greaterEqual, L_failed);
2644 // At this point, it is known to be a typeArray (array_tag 0x3).
2645 #ifdef ASSERT
2646 {
2647 BLOCK_COMMENT("assert primitive array {");
2648 Label L;
2649 __ cmpl(rax_lh, (Klass::_lh_array_tag_type_value << Klass::_lh_array_tag_shift));
2650 __ jcc(Assembler::greaterEqual, L);
2651 __ stop("must be a primitive array");
2652 __ bind(L);
2653 BLOCK_COMMENT("} assert primitive array done");
2654 }
2655 #endif
2657 arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length,
2658 r10, L_failed);
2660 // TypeArrayKlass
2661 //
2662 // src_addr = (src + array_header_in_bytes()) + (src_pos << log2elemsize);
2663 // dst_addr = (dst + array_header_in_bytes()) + (dst_pos << log2elemsize);
2664 //
2666 const Register r10_offset = r10; // array offset
2667 const Register rax_elsize = rax_lh; // element size
2669 __ movl(r10_offset, rax_lh);
2670 __ shrl(r10_offset, Klass::_lh_header_size_shift);
2671 __ andptr(r10_offset, Klass::_lh_header_size_mask); // array_offset
2672 __ addptr(src, r10_offset); // src array offset
2673 __ addptr(dst, r10_offset); // dst array offset
2674 BLOCK_COMMENT("choose copy loop based on element size");
2675 __ andl(rax_lh, Klass::_lh_log2_element_size_mask); // rax_lh -> rax_elsize
2677 // next registers should be set before the jump to corresponding stub
2678 const Register from = c_rarg0; // source array address
2679 const Register to = c_rarg1; // destination array address
2680 const Register count = c_rarg2; // elements count
2682 // 'from', 'to', 'count' registers should be set in such order
2683 // since they are the same as 'src', 'src_pos', 'dst'.
2685 __ BIND(L_copy_bytes);
2686 __ cmpl(rax_elsize, 0);
2687 __ jccb(Assembler::notEqual, L_copy_shorts);
2688 __ lea(from, Address(src, src_pos, Address::times_1, 0));// src_addr
2689 __ lea(to, Address(dst, dst_pos, Address::times_1, 0));// dst_addr
2690 __ movl2ptr(count, r11_length); // length
2691 __ jump(RuntimeAddress(byte_copy_entry));
2693 __ BIND(L_copy_shorts);
2694 __ cmpl(rax_elsize, LogBytesPerShort);
2695 __ jccb(Assembler::notEqual, L_copy_ints);
2696 __ lea(from, Address(src, src_pos, Address::times_2, 0));// src_addr
2697 __ lea(to, Address(dst, dst_pos, Address::times_2, 0));// dst_addr
2698 __ movl2ptr(count, r11_length); // length
2699 __ jump(RuntimeAddress(short_copy_entry));
2701 __ BIND(L_copy_ints);
2702 __ cmpl(rax_elsize, LogBytesPerInt);
2703 __ jccb(Assembler::notEqual, L_copy_longs);
2704 __ lea(from, Address(src, src_pos, Address::times_4, 0));// src_addr
2705 __ lea(to, Address(dst, dst_pos, Address::times_4, 0));// dst_addr
2706 __ movl2ptr(count, r11_length); // length
2707 __ jump(RuntimeAddress(int_copy_entry));
2709 __ BIND(L_copy_longs);
2710 #ifdef ASSERT
2711 {
2712 BLOCK_COMMENT("assert long copy {");
2713 Label L;
2714 __ cmpl(rax_elsize, LogBytesPerLong);
2715 __ jcc(Assembler::equal, L);
2716 __ stop("must be long copy, but elsize is wrong");
2717 __ bind(L);
2718 BLOCK_COMMENT("} assert long copy done");
2719 }
2720 #endif
2721 __ lea(from, Address(src, src_pos, Address::times_8, 0));// src_addr
2722 __ lea(to, Address(dst, dst_pos, Address::times_8, 0));// dst_addr
2723 __ movl2ptr(count, r11_length); // length
2724 __ jump(RuntimeAddress(long_copy_entry));
2726 // ObjArrayKlass
2727 __ BIND(L_objArray);
2728 // live at this point: r10_src_klass, r11_length, src[_pos], dst[_pos]
2730 Label L_plain_copy, L_checkcast_copy;
2731 // test array classes for subtyping
2732 __ load_klass(rax, dst);
2733 __ cmpq(r10_src_klass, rax); // usual case is exact equality
2734 __ jcc(Assembler::notEqual, L_checkcast_copy);
2736 // Identically typed arrays can be copied without element-wise checks.
2737 arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length,
2738 r10, L_failed);
2740 __ lea(from, Address(src, src_pos, TIMES_OOP,
2741 arrayOopDesc::base_offset_in_bytes(T_OBJECT))); // src_addr
2742 __ lea(to, Address(dst, dst_pos, TIMES_OOP,
2743 arrayOopDesc::base_offset_in_bytes(T_OBJECT))); // dst_addr
2744 __ movl2ptr(count, r11_length); // length
2745 __ BIND(L_plain_copy);
2746 __ jump(RuntimeAddress(oop_copy_entry));
2748 __ BIND(L_checkcast_copy);
2749 // live at this point: r10_src_klass, r11_length, rax (dst_klass)
2750 {
2751 // Before looking at dst.length, make sure dst is also an objArray.
2752 __ cmpl(Address(rax, lh_offset), objArray_lh);
2753 __ jcc(Assembler::notEqual, L_failed);
2755 // It is safe to examine both src.length and dst.length.
2756 arraycopy_range_checks(src, src_pos, dst, dst_pos, r11_length,
2757 rax, L_failed);
2759 const Register r11_dst_klass = r11;
2760 __ load_klass(r11_dst_klass, dst); // reload
2762 // Marshal the base address arguments now, freeing registers.
2763 __ lea(from, Address(src, src_pos, TIMES_OOP,
2764 arrayOopDesc::base_offset_in_bytes(T_OBJECT)));
2765 __ lea(to, Address(dst, dst_pos, TIMES_OOP,
2766 arrayOopDesc::base_offset_in_bytes(T_OBJECT)));
2767 __ movl(count, length); // length (reloaded)
2768 Register sco_temp = c_rarg3; // this register is free now
2769 assert_different_registers(from, to, count, sco_temp,
2770 r11_dst_klass, r10_src_klass);
2771 assert_clean_int(count, sco_temp);
2773 // Generate the type check.
2774 const int sco_offset = in_bytes(Klass::super_check_offset_offset());
2775 __ movl(sco_temp, Address(r11_dst_klass, sco_offset));
2776 assert_clean_int(sco_temp, rax);
2777 generate_type_check(r10_src_klass, sco_temp, r11_dst_klass, L_plain_copy);
2779 // Fetch destination element klass from the ObjArrayKlass header.
2780 int ek_offset = in_bytes(ObjArrayKlass::element_klass_offset());
2781 __ movptr(r11_dst_klass, Address(r11_dst_klass, ek_offset));
2782 __ movl( sco_temp, Address(r11_dst_klass, sco_offset));
2783 assert_clean_int(sco_temp, rax);
2785 // the checkcast_copy loop needs two extra arguments:
2786 assert(c_rarg3 == sco_temp, "#3 already in place");
2787 // Set up arguments for checkcast_copy_entry.
2788 setup_arg_regs(4);
2789 __ movptr(r8, r11_dst_klass); // dst.klass.element_klass, r8 is c_rarg4 on Linux/Solaris
2790 __ jump(RuntimeAddress(checkcast_copy_entry));
2791 }
2793 __ BIND(L_failed);
2794 __ xorptr(rax, rax);
2795 __ notptr(rax); // return -1
2796 __ leave(); // required for proper stackwalking of RuntimeStub frame
2797 __ ret(0);
2799 return start;
2800 }
2802 void generate_arraycopy_stubs() {
2803 address entry;
2804 address entry_jbyte_arraycopy;
2805 address entry_jshort_arraycopy;
2806 address entry_jint_arraycopy;
2807 address entry_oop_arraycopy;
2808 address entry_jlong_arraycopy;
2809 address entry_checkcast_arraycopy;
2811 StubRoutines::_jbyte_disjoint_arraycopy = generate_disjoint_byte_copy(false, &entry,
2812 "jbyte_disjoint_arraycopy");
2813 StubRoutines::_jbyte_arraycopy = generate_conjoint_byte_copy(false, entry, &entry_jbyte_arraycopy,
2814 "jbyte_arraycopy");
2816 StubRoutines::_jshort_disjoint_arraycopy = generate_disjoint_short_copy(false, &entry,
2817 "jshort_disjoint_arraycopy");
2818 StubRoutines::_jshort_arraycopy = generate_conjoint_short_copy(false, entry, &entry_jshort_arraycopy,
2819 "jshort_arraycopy");
2821 StubRoutines::_jint_disjoint_arraycopy = generate_disjoint_int_oop_copy(false, false, &entry,
2822 "jint_disjoint_arraycopy");
2823 StubRoutines::_jint_arraycopy = generate_conjoint_int_oop_copy(false, false, entry,
2824 &entry_jint_arraycopy, "jint_arraycopy");
2826 StubRoutines::_jlong_disjoint_arraycopy = generate_disjoint_long_oop_copy(false, false, &entry,
2827 "jlong_disjoint_arraycopy");
2828 StubRoutines::_jlong_arraycopy = generate_conjoint_long_oop_copy(false, false, entry,
2829 &entry_jlong_arraycopy, "jlong_arraycopy");
2832 if (UseCompressedOops) {
2833 StubRoutines::_oop_disjoint_arraycopy = generate_disjoint_int_oop_copy(false, true, &entry,
2834 "oop_disjoint_arraycopy");
2835 StubRoutines::_oop_arraycopy = generate_conjoint_int_oop_copy(false, true, entry,
2836 &entry_oop_arraycopy, "oop_arraycopy");
2837 StubRoutines::_oop_disjoint_arraycopy_uninit = generate_disjoint_int_oop_copy(false, true, &entry,
2838 "oop_disjoint_arraycopy_uninit",
2839 /*dest_uninitialized*/true);
2840 StubRoutines::_oop_arraycopy_uninit = generate_conjoint_int_oop_copy(false, true, entry,
2841 NULL, "oop_arraycopy_uninit",
2842 /*dest_uninitialized*/true);
2843 } else {
2844 StubRoutines::_oop_disjoint_arraycopy = generate_disjoint_long_oop_copy(false, true, &entry,
2845 "oop_disjoint_arraycopy");
2846 StubRoutines::_oop_arraycopy = generate_conjoint_long_oop_copy(false, true, entry,
2847 &entry_oop_arraycopy, "oop_arraycopy");
2848 StubRoutines::_oop_disjoint_arraycopy_uninit = generate_disjoint_long_oop_copy(false, true, &entry,
2849 "oop_disjoint_arraycopy_uninit",
2850 /*dest_uninitialized*/true);
2851 StubRoutines::_oop_arraycopy_uninit = generate_conjoint_long_oop_copy(false, true, entry,
2852 NULL, "oop_arraycopy_uninit",
2853 /*dest_uninitialized*/true);
2854 }
2856 StubRoutines::_checkcast_arraycopy = generate_checkcast_copy("checkcast_arraycopy", &entry_checkcast_arraycopy);
2857 StubRoutines::_checkcast_arraycopy_uninit = generate_checkcast_copy("checkcast_arraycopy_uninit", NULL,
2858 /*dest_uninitialized*/true);
2860 StubRoutines::_unsafe_arraycopy = generate_unsafe_copy("unsafe_arraycopy",
2861 entry_jbyte_arraycopy,
2862 entry_jshort_arraycopy,
2863 entry_jint_arraycopy,
2864 entry_jlong_arraycopy);
2865 StubRoutines::_generic_arraycopy = generate_generic_copy("generic_arraycopy",
2866 entry_jbyte_arraycopy,
2867 entry_jshort_arraycopy,
2868 entry_jint_arraycopy,
2869 entry_oop_arraycopy,
2870 entry_jlong_arraycopy,
2871 entry_checkcast_arraycopy);
2873 StubRoutines::_jbyte_fill = generate_fill(T_BYTE, false, "jbyte_fill");
2874 StubRoutines::_jshort_fill = generate_fill(T_SHORT, false, "jshort_fill");
2875 StubRoutines::_jint_fill = generate_fill(T_INT, false, "jint_fill");
2876 StubRoutines::_arrayof_jbyte_fill = generate_fill(T_BYTE, true, "arrayof_jbyte_fill");
2877 StubRoutines::_arrayof_jshort_fill = generate_fill(T_SHORT, true, "arrayof_jshort_fill");
2878 StubRoutines::_arrayof_jint_fill = generate_fill(T_INT, true, "arrayof_jint_fill");
2880 // We don't generate specialized code for HeapWord-aligned source
2881 // arrays, so just use the code we've already generated
2882 StubRoutines::_arrayof_jbyte_disjoint_arraycopy = StubRoutines::_jbyte_disjoint_arraycopy;
2883 StubRoutines::_arrayof_jbyte_arraycopy = StubRoutines::_jbyte_arraycopy;
2885 StubRoutines::_arrayof_jshort_disjoint_arraycopy = StubRoutines::_jshort_disjoint_arraycopy;
2886 StubRoutines::_arrayof_jshort_arraycopy = StubRoutines::_jshort_arraycopy;
2888 StubRoutines::_arrayof_jint_disjoint_arraycopy = StubRoutines::_jint_disjoint_arraycopy;
2889 StubRoutines::_arrayof_jint_arraycopy = StubRoutines::_jint_arraycopy;
2891 StubRoutines::_arrayof_jlong_disjoint_arraycopy = StubRoutines::_jlong_disjoint_arraycopy;
2892 StubRoutines::_arrayof_jlong_arraycopy = StubRoutines::_jlong_arraycopy;
2894 StubRoutines::_arrayof_oop_disjoint_arraycopy = StubRoutines::_oop_disjoint_arraycopy;
2895 StubRoutines::_arrayof_oop_arraycopy = StubRoutines::_oop_arraycopy;
2897 StubRoutines::_arrayof_oop_disjoint_arraycopy_uninit = StubRoutines::_oop_disjoint_arraycopy_uninit;
2898 StubRoutines::_arrayof_oop_arraycopy_uninit = StubRoutines::_oop_arraycopy_uninit;
2899 }
2901 void generate_math_stubs() {
2902 {
2903 StubCodeMark mark(this, "StubRoutines", "log");
2904 StubRoutines::_intrinsic_log = (double (*)(double)) __ pc();
2906 __ subq(rsp, 8);
2907 __ movdbl(Address(rsp, 0), xmm0);
2908 __ fld_d(Address(rsp, 0));
2909 __ flog();
2910 __ fstp_d(Address(rsp, 0));
2911 __ movdbl(xmm0, Address(rsp, 0));
2912 __ addq(rsp, 8);
2913 __ ret(0);
2914 }
2915 {
2916 StubCodeMark mark(this, "StubRoutines", "log10");
2917 StubRoutines::_intrinsic_log10 = (double (*)(double)) __ pc();
2919 __ subq(rsp, 8);
2920 __ movdbl(Address(rsp, 0), xmm0);
2921 __ fld_d(Address(rsp, 0));
2922 __ flog10();
2923 __ fstp_d(Address(rsp, 0));
2924 __ movdbl(xmm0, Address(rsp, 0));
2925 __ addq(rsp, 8);
2926 __ ret(0);
2927 }
2928 {
2929 StubCodeMark mark(this, "StubRoutines", "sin");
2930 StubRoutines::_intrinsic_sin = (double (*)(double)) __ pc();
2932 __ subq(rsp, 8);
2933 __ movdbl(Address(rsp, 0), xmm0);
2934 __ fld_d(Address(rsp, 0));
2935 __ trigfunc('s');
2936 __ fstp_d(Address(rsp, 0));
2937 __ movdbl(xmm0, Address(rsp, 0));
2938 __ addq(rsp, 8);
2939 __ ret(0);
2940 }
2941 {
2942 StubCodeMark mark(this, "StubRoutines", "cos");
2943 StubRoutines::_intrinsic_cos = (double (*)(double)) __ pc();
2945 __ subq(rsp, 8);
2946 __ movdbl(Address(rsp, 0), xmm0);
2947 __ fld_d(Address(rsp, 0));
2948 __ trigfunc('c');
2949 __ fstp_d(Address(rsp, 0));
2950 __ movdbl(xmm0, Address(rsp, 0));
2951 __ addq(rsp, 8);
2952 __ ret(0);
2953 }
2954 {
2955 StubCodeMark mark(this, "StubRoutines", "tan");
2956 StubRoutines::_intrinsic_tan = (double (*)(double)) __ pc();
2958 __ subq(rsp, 8);
2959 __ movdbl(Address(rsp, 0), xmm0);
2960 __ fld_d(Address(rsp, 0));
2961 __ trigfunc('t');
2962 __ fstp_d(Address(rsp, 0));
2963 __ movdbl(xmm0, Address(rsp, 0));
2964 __ addq(rsp, 8);
2965 __ ret(0);
2966 }
2967 {
2968 StubCodeMark mark(this, "StubRoutines", "exp");
2969 StubRoutines::_intrinsic_exp = (double (*)(double)) __ pc();
2971 __ subq(rsp, 8);
2972 __ movdbl(Address(rsp, 0), xmm0);
2973 __ fld_d(Address(rsp, 0));
2974 __ exp_with_fallback(0);
2975 __ fstp_d(Address(rsp, 0));
2976 __ movdbl(xmm0, Address(rsp, 0));
2977 __ addq(rsp, 8);
2978 __ ret(0);
2979 }
2980 {
2981 StubCodeMark mark(this, "StubRoutines", "pow");
2982 StubRoutines::_intrinsic_pow = (double (*)(double,double)) __ pc();
2984 __ subq(rsp, 8);
2985 __ movdbl(Address(rsp, 0), xmm1);
2986 __ fld_d(Address(rsp, 0));
2987 __ movdbl(Address(rsp, 0), xmm0);
2988 __ fld_d(Address(rsp, 0));
2989 __ pow_with_fallback(0);
2990 __ fstp_d(Address(rsp, 0));
2991 __ movdbl(xmm0, Address(rsp, 0));
2992 __ addq(rsp, 8);
2993 __ ret(0);
2994 }
2995 }
2997 // AES intrinsic stubs
2998 enum {AESBlockSize = 16};
3000 address generate_key_shuffle_mask() {
3001 __ align(16);
3002 StubCodeMark mark(this, "StubRoutines", "key_shuffle_mask");
3003 address start = __ pc();
3004 __ emit_data64( 0x0405060700010203, relocInfo::none );
3005 __ emit_data64( 0x0c0d0e0f08090a0b, relocInfo::none );
3006 return start;
3007 }
3009 // Utility routine for loading a 128-bit key word in little endian format
3010 // can optionally specify that the shuffle mask is already in an xmmregister
3011 void load_key(XMMRegister xmmdst, Register key, int offset, XMMRegister xmm_shuf_mask=NULL) {
3012 __ movdqu(xmmdst, Address(key, offset));
3013 if (xmm_shuf_mask != NULL) {
3014 __ pshufb(xmmdst, xmm_shuf_mask);
3015 } else {
3016 __ pshufb(xmmdst, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3017 }
3018 }
3020 // Arguments:
3021 //
3022 // Inputs:
3023 // c_rarg0 - source byte array address
3024 // c_rarg1 - destination byte array address
3025 // c_rarg2 - K (key) in little endian int array
3026 //
3027 address generate_aescrypt_encryptBlock() {
3028 assert(UseAES, "need AES instructions and misaligned SSE support");
3029 __ align(CodeEntryAlignment);
3030 StubCodeMark mark(this, "StubRoutines", "aescrypt_encryptBlock");
3031 Label L_doLast;
3032 address start = __ pc();
3034 const Register from = c_rarg0; // source array address
3035 const Register to = c_rarg1; // destination array address
3036 const Register key = c_rarg2; // key array address
3037 const Register keylen = rax;
3039 const XMMRegister xmm_result = xmm0;
3040 const XMMRegister xmm_key_shuf_mask = xmm1;
3041 // On win64 xmm6-xmm15 must be preserved so don't use them.
3042 const XMMRegister xmm_temp1 = xmm2;
3043 const XMMRegister xmm_temp2 = xmm3;
3044 const XMMRegister xmm_temp3 = xmm4;
3045 const XMMRegister xmm_temp4 = xmm5;
3047 __ enter(); // required for proper stackwalking of RuntimeStub frame
3049 // keylen could be only {11, 13, 15} * 4 = {44, 52, 60}
3050 __ movl(keylen, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3052 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3053 __ movdqu(xmm_result, Address(from, 0)); // get 16 bytes of input
3055 // For encryption, the java expanded key ordering is just what we need
3056 // we don't know if the key is aligned, hence not using load-execute form
3058 load_key(xmm_temp1, key, 0x00, xmm_key_shuf_mask);
3059 __ pxor(xmm_result, xmm_temp1);
3061 load_key(xmm_temp1, key, 0x10, xmm_key_shuf_mask);
3062 load_key(xmm_temp2, key, 0x20, xmm_key_shuf_mask);
3063 load_key(xmm_temp3, key, 0x30, xmm_key_shuf_mask);
3064 load_key(xmm_temp4, key, 0x40, xmm_key_shuf_mask);
3066 __ aesenc(xmm_result, xmm_temp1);
3067 __ aesenc(xmm_result, xmm_temp2);
3068 __ aesenc(xmm_result, xmm_temp3);
3069 __ aesenc(xmm_result, xmm_temp4);
3071 load_key(xmm_temp1, key, 0x50, xmm_key_shuf_mask);
3072 load_key(xmm_temp2, key, 0x60, xmm_key_shuf_mask);
3073 load_key(xmm_temp3, key, 0x70, xmm_key_shuf_mask);
3074 load_key(xmm_temp4, key, 0x80, xmm_key_shuf_mask);
3076 __ aesenc(xmm_result, xmm_temp1);
3077 __ aesenc(xmm_result, xmm_temp2);
3078 __ aesenc(xmm_result, xmm_temp3);
3079 __ aesenc(xmm_result, xmm_temp4);
3081 load_key(xmm_temp1, key, 0x90, xmm_key_shuf_mask);
3082 load_key(xmm_temp2, key, 0xa0, xmm_key_shuf_mask);
3084 __ cmpl(keylen, 44);
3085 __ jccb(Assembler::equal, L_doLast);
3087 __ aesenc(xmm_result, xmm_temp1);
3088 __ aesenc(xmm_result, xmm_temp2);
3090 load_key(xmm_temp1, key, 0xb0, xmm_key_shuf_mask);
3091 load_key(xmm_temp2, key, 0xc0, xmm_key_shuf_mask);
3093 __ cmpl(keylen, 52);
3094 __ jccb(Assembler::equal, L_doLast);
3096 __ aesenc(xmm_result, xmm_temp1);
3097 __ aesenc(xmm_result, xmm_temp2);
3099 load_key(xmm_temp1, key, 0xd0, xmm_key_shuf_mask);
3100 load_key(xmm_temp2, key, 0xe0, xmm_key_shuf_mask);
3102 __ BIND(L_doLast);
3103 __ aesenc(xmm_result, xmm_temp1);
3104 __ aesenclast(xmm_result, xmm_temp2);
3105 __ movdqu(Address(to, 0), xmm_result); // store the result
3106 __ xorptr(rax, rax); // return 0
3107 __ leave(); // required for proper stackwalking of RuntimeStub frame
3108 __ ret(0);
3110 return start;
3111 }
3114 // Arguments:
3115 //
3116 // Inputs:
3117 // c_rarg0 - source byte array address
3118 // c_rarg1 - destination byte array address
3119 // c_rarg2 - K (key) in little endian int array
3120 //
3121 address generate_aescrypt_decryptBlock() {
3122 assert(UseAES, "need AES instructions and misaligned SSE support");
3123 __ align(CodeEntryAlignment);
3124 StubCodeMark mark(this, "StubRoutines", "aescrypt_decryptBlock");
3125 Label L_doLast;
3126 address start = __ pc();
3128 const Register from = c_rarg0; // source array address
3129 const Register to = c_rarg1; // destination array address
3130 const Register key = c_rarg2; // key array address
3131 const Register keylen = rax;
3133 const XMMRegister xmm_result = xmm0;
3134 const XMMRegister xmm_key_shuf_mask = xmm1;
3135 // On win64 xmm6-xmm15 must be preserved so don't use them.
3136 const XMMRegister xmm_temp1 = xmm2;
3137 const XMMRegister xmm_temp2 = xmm3;
3138 const XMMRegister xmm_temp3 = xmm4;
3139 const XMMRegister xmm_temp4 = xmm5;
3141 __ enter(); // required for proper stackwalking of RuntimeStub frame
3143 // keylen could be only {11, 13, 15} * 4 = {44, 52, 60}
3144 __ movl(keylen, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3146 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3147 __ movdqu(xmm_result, Address(from, 0));
3149 // for decryption java expanded key ordering is rotated one position from what we want
3150 // so we start from 0x10 here and hit 0x00 last
3151 // we don't know if the key is aligned, hence not using load-execute form
3152 load_key(xmm_temp1, key, 0x10, xmm_key_shuf_mask);
3153 load_key(xmm_temp2, key, 0x20, xmm_key_shuf_mask);
3154 load_key(xmm_temp3, key, 0x30, xmm_key_shuf_mask);
3155 load_key(xmm_temp4, key, 0x40, xmm_key_shuf_mask);
3157 __ pxor (xmm_result, xmm_temp1);
3158 __ aesdec(xmm_result, xmm_temp2);
3159 __ aesdec(xmm_result, xmm_temp3);
3160 __ aesdec(xmm_result, xmm_temp4);
3162 load_key(xmm_temp1, key, 0x50, xmm_key_shuf_mask);
3163 load_key(xmm_temp2, key, 0x60, xmm_key_shuf_mask);
3164 load_key(xmm_temp3, key, 0x70, xmm_key_shuf_mask);
3165 load_key(xmm_temp4, key, 0x80, xmm_key_shuf_mask);
3167 __ aesdec(xmm_result, xmm_temp1);
3168 __ aesdec(xmm_result, xmm_temp2);
3169 __ aesdec(xmm_result, xmm_temp3);
3170 __ aesdec(xmm_result, xmm_temp4);
3172 load_key(xmm_temp1, key, 0x90, xmm_key_shuf_mask);
3173 load_key(xmm_temp2, key, 0xa0, xmm_key_shuf_mask);
3174 load_key(xmm_temp3, key, 0x00, xmm_key_shuf_mask);
3176 __ cmpl(keylen, 44);
3177 __ jccb(Assembler::equal, L_doLast);
3179 __ aesdec(xmm_result, xmm_temp1);
3180 __ aesdec(xmm_result, xmm_temp2);
3182 load_key(xmm_temp1, key, 0xb0, xmm_key_shuf_mask);
3183 load_key(xmm_temp2, key, 0xc0, xmm_key_shuf_mask);
3185 __ cmpl(keylen, 52);
3186 __ jccb(Assembler::equal, L_doLast);
3188 __ aesdec(xmm_result, xmm_temp1);
3189 __ aesdec(xmm_result, xmm_temp2);
3191 load_key(xmm_temp1, key, 0xd0, xmm_key_shuf_mask);
3192 load_key(xmm_temp2, key, 0xe0, xmm_key_shuf_mask);
3194 __ BIND(L_doLast);
3195 __ aesdec(xmm_result, xmm_temp1);
3196 __ aesdec(xmm_result, xmm_temp2);
3198 // for decryption the aesdeclast operation is always on key+0x00
3199 __ aesdeclast(xmm_result, xmm_temp3);
3200 __ movdqu(Address(to, 0), xmm_result); // store the result
3201 __ xorptr(rax, rax); // return 0
3202 __ leave(); // required for proper stackwalking of RuntimeStub frame
3203 __ ret(0);
3205 return start;
3206 }
3209 // Arguments:
3210 //
3211 // Inputs:
3212 // c_rarg0 - source byte array address
3213 // c_rarg1 - destination byte array address
3214 // c_rarg2 - K (key) in little endian int array
3215 // c_rarg3 - r vector byte array address
3216 // c_rarg4 - input length
3217 //
3218 address generate_cipherBlockChaining_encryptAESCrypt() {
3219 assert(UseAES, "need AES instructions and misaligned SSE support");
3220 __ align(CodeEntryAlignment);
3221 StubCodeMark mark(this, "StubRoutines", "cipherBlockChaining_encryptAESCrypt");
3222 address start = __ pc();
3224 Label L_exit, L_key_192_256, L_key_256, L_loopTop_128, L_loopTop_192, L_loopTop_256;
3225 const Register from = c_rarg0; // source array address
3226 const Register to = c_rarg1; // destination array address
3227 const Register key = c_rarg2; // key array address
3228 const Register rvec = c_rarg3; // r byte array initialized from initvector array address
3229 // and left with the results of the last encryption block
3230 #ifndef _WIN64
3231 const Register len_reg = c_rarg4; // src len (must be multiple of blocksize 16)
3232 #else
3233 const Address len_mem(rsp, 6 * wordSize); // length is on stack on Win64
3234 const Register len_reg = r10; // pick the first volatile windows register
3235 #endif
3236 const Register pos = rax;
3238 // xmm register assignments for the loops below
3239 const XMMRegister xmm_result = xmm0;
3240 const XMMRegister xmm_temp = xmm1;
3241 // keys 0-10 preloaded into xmm2-xmm12
3242 const int XMM_REG_NUM_KEY_FIRST = 2;
3243 const int XMM_REG_NUM_KEY_LAST = 15;
3244 const XMMRegister xmm_key0 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST);
3245 const XMMRegister xmm_key10 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+10);
3246 const XMMRegister xmm_key11 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+11);
3247 const XMMRegister xmm_key12 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+12);
3248 const XMMRegister xmm_key13 = as_XMMRegister(XMM_REG_NUM_KEY_FIRST+13);
3250 __ enter(); // required for proper stackwalking of RuntimeStub frame
3252 #ifdef _WIN64
3253 // on win64, fill len_reg from stack position
3254 __ movl(len_reg, len_mem);
3255 // save the xmm registers which must be preserved 6-15
3256 __ subptr(rsp, -rsp_after_call_off * wordSize);
3257 for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3258 __ movdqu(xmm_save(i), as_XMMRegister(i));
3259 }
3260 #endif
3262 const XMMRegister xmm_key_shuf_mask = xmm_temp; // used temporarily to swap key bytes up front
3263 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3264 // load up xmm regs xmm2 thru xmm12 with key 0x00 - 0xa0
3265 for (int rnum = XMM_REG_NUM_KEY_FIRST, offset = 0x00; rnum <= XMM_REG_NUM_KEY_FIRST+10; rnum++) {
3266 load_key(as_XMMRegister(rnum), key, offset, xmm_key_shuf_mask);
3267 offset += 0x10;
3268 }
3269 __ movdqu(xmm_result, Address(rvec, 0x00)); // initialize xmm_result with r vec
3271 // now split to different paths depending on the keylen (len in ints of AESCrypt.KLE array (52=192, or 60=256))
3272 __ movl(rax, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3273 __ cmpl(rax, 44);
3274 __ jcc(Assembler::notEqual, L_key_192_256);
3276 // 128 bit code follows here
3277 __ movptr(pos, 0);
3278 __ align(OptoLoopAlignment);
3280 __ BIND(L_loopTop_128);
3281 __ movdqu(xmm_temp, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of input
3282 __ pxor (xmm_result, xmm_temp); // xor with the current r vector
3283 __ pxor (xmm_result, xmm_key0); // do the aes rounds
3284 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_FIRST + 9; rnum++) {
3285 __ aesenc(xmm_result, as_XMMRegister(rnum));
3286 }
3287 __ aesenclast(xmm_result, xmm_key10);
3288 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3289 // no need to store r to memory until we exit
3290 __ addptr(pos, AESBlockSize);
3291 __ subptr(len_reg, AESBlockSize);
3292 __ jcc(Assembler::notEqual, L_loopTop_128);
3294 __ BIND(L_exit);
3295 __ movdqu(Address(rvec, 0), xmm_result); // final value of r stored in rvec of CipherBlockChaining object
3297 #ifdef _WIN64
3298 // restore xmm regs belonging to calling function
3299 for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3300 __ movdqu(as_XMMRegister(i), xmm_save(i));
3301 }
3302 #endif
3303 __ movl(rax, 0); // return 0 (why?)
3304 __ leave(); // required for proper stackwalking of RuntimeStub frame
3305 __ ret(0);
3307 __ BIND(L_key_192_256);
3308 // here rax = len in ints of AESCrypt.KLE array (52=192, or 60=256)
3309 load_key(xmm_key11, key, 0xb0, xmm_key_shuf_mask);
3310 load_key(xmm_key12, key, 0xc0, xmm_key_shuf_mask);
3311 __ cmpl(rax, 52);
3312 __ jcc(Assembler::notEqual, L_key_256);
3314 // 192-bit code follows here (could be changed to use more xmm registers)
3315 __ movptr(pos, 0);
3316 __ align(OptoLoopAlignment);
3318 __ BIND(L_loopTop_192);
3319 __ movdqu(xmm_temp, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of input
3320 __ pxor (xmm_result, xmm_temp); // xor with the current r vector
3321 __ pxor (xmm_result, xmm_key0); // do the aes rounds
3322 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_FIRST + 11; rnum++) {
3323 __ aesenc(xmm_result, as_XMMRegister(rnum));
3324 }
3325 __ aesenclast(xmm_result, xmm_key12);
3326 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3327 // no need to store r to memory until we exit
3328 __ addptr(pos, AESBlockSize);
3329 __ subptr(len_reg, AESBlockSize);
3330 __ jcc(Assembler::notEqual, L_loopTop_192);
3331 __ jmp(L_exit);
3333 __ BIND(L_key_256);
3334 // 256-bit code follows here (could be changed to use more xmm registers)
3335 load_key(xmm_key13, key, 0xd0, xmm_key_shuf_mask);
3336 __ movptr(pos, 0);
3337 __ align(OptoLoopAlignment);
3339 __ BIND(L_loopTop_256);
3340 __ movdqu(xmm_temp, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of input
3341 __ pxor (xmm_result, xmm_temp); // xor with the current r vector
3342 __ pxor (xmm_result, xmm_key0); // do the aes rounds
3343 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_FIRST + 13; rnum++) {
3344 __ aesenc(xmm_result, as_XMMRegister(rnum));
3345 }
3346 load_key(xmm_temp, key, 0xe0);
3347 __ aesenclast(xmm_result, xmm_temp);
3348 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3349 // no need to store r to memory until we exit
3350 __ addptr(pos, AESBlockSize);
3351 __ subptr(len_reg, AESBlockSize);
3352 __ jcc(Assembler::notEqual, L_loopTop_256);
3353 __ jmp(L_exit);
3355 return start;
3356 }
3360 // This is a version of CBC/AES Decrypt which does 4 blocks in a loop at a time
3361 // to hide instruction latency
3362 //
3363 // Arguments:
3364 //
3365 // Inputs:
3366 // c_rarg0 - source byte array address
3367 // c_rarg1 - destination byte array address
3368 // c_rarg2 - K (key) in little endian int array
3369 // c_rarg3 - r vector byte array address
3370 // c_rarg4 - input length
3371 //
3373 address generate_cipherBlockChaining_decryptAESCrypt_Parallel() {
3374 assert(UseAES, "need AES instructions and misaligned SSE support");
3375 __ align(CodeEntryAlignment);
3376 StubCodeMark mark(this, "StubRoutines", "cipherBlockChaining_decryptAESCrypt");
3377 address start = __ pc();
3379 Label L_exit, L_key_192_256, L_key_256;
3380 Label L_singleBlock_loopTop_128, L_multiBlock_loopTop_128;
3381 Label L_singleBlock_loopTop_192, L_singleBlock_loopTop_256;
3382 const Register from = c_rarg0; // source array address
3383 const Register to = c_rarg1; // destination array address
3384 const Register key = c_rarg2; // key array address
3385 const Register rvec = c_rarg3; // r byte array initialized from initvector array address
3386 // and left with the results of the last encryption block
3387 #ifndef _WIN64
3388 const Register len_reg = c_rarg4; // src len (must be multiple of blocksize 16)
3389 #else
3390 const Address len_mem(rsp, 6 * wordSize); // length is on stack on Win64
3391 const Register len_reg = r10; // pick the first volatile windows register
3392 #endif
3393 const Register pos = rax;
3395 // keys 0-10 preloaded into xmm2-xmm12
3396 const int XMM_REG_NUM_KEY_FIRST = 5;
3397 const int XMM_REG_NUM_KEY_LAST = 15;
3398 const XMMRegister xmm_key_first = as_XMMRegister(XMM_REG_NUM_KEY_FIRST);
3399 const XMMRegister xmm_key_last = as_XMMRegister(XMM_REG_NUM_KEY_LAST);
3401 __ enter(); // required for proper stackwalking of RuntimeStub frame
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 #endif
3412 // the java expanded key ordering is rotated one position from what we want
3413 // so we start from 0x10 here and hit 0x00 last
3414 const XMMRegister xmm_key_shuf_mask = xmm1; // used temporarily to swap key bytes up front
3415 __ movdqu(xmm_key_shuf_mask, ExternalAddress(StubRoutines::x86::key_shuffle_mask_addr()));
3416 // load up xmm regs 5 thru 15 with key 0x10 - 0xa0 - 0x00
3417 for (int rnum = XMM_REG_NUM_KEY_FIRST, offset = 0x10; rnum < XMM_REG_NUM_KEY_LAST; rnum++) {
3418 load_key(as_XMMRegister(rnum), key, offset, xmm_key_shuf_mask);
3419 offset += 0x10;
3420 }
3421 load_key(xmm_key_last, key, 0x00, xmm_key_shuf_mask);
3423 const XMMRegister xmm_prev_block_cipher = xmm1; // holds cipher of previous block
3425 // registers holding the four results in the parallelized loop
3426 const XMMRegister xmm_result0 = xmm0;
3427 const XMMRegister xmm_result1 = xmm2;
3428 const XMMRegister xmm_result2 = xmm3;
3429 const XMMRegister xmm_result3 = xmm4;
3431 __ movdqu(xmm_prev_block_cipher, Address(rvec, 0x00)); // initialize with initial rvec
3433 // now split to different paths depending on the keylen (len in ints of AESCrypt.KLE array (52=192, or 60=256))
3434 __ movl(rax, Address(key, arrayOopDesc::length_offset_in_bytes() - arrayOopDesc::base_offset_in_bytes(T_INT)));
3435 __ cmpl(rax, 44);
3436 __ jcc(Assembler::notEqual, L_key_192_256);
3439 // 128-bit code follows here, parallelized
3440 __ movptr(pos, 0);
3441 __ align(OptoLoopAlignment);
3442 __ BIND(L_multiBlock_loopTop_128);
3443 __ cmpptr(len_reg, 4*AESBlockSize); // see if at least 4 blocks left
3444 __ jcc(Assembler::less, L_singleBlock_loopTop_128);
3446 __ movdqu(xmm_result0, Address(from, pos, Address::times_1, 0*AESBlockSize)); // get next 4 blocks into xmmresult registers
3447 __ movdqu(xmm_result1, Address(from, pos, Address::times_1, 1*AESBlockSize));
3448 __ movdqu(xmm_result2, Address(from, pos, Address::times_1, 2*AESBlockSize));
3449 __ movdqu(xmm_result3, Address(from, pos, Address::times_1, 3*AESBlockSize));
3451 #define DoFour(opc, src_reg) \
3452 __ opc(xmm_result0, src_reg); \
3453 __ opc(xmm_result1, src_reg); \
3454 __ opc(xmm_result2, src_reg); \
3455 __ opc(xmm_result3, src_reg);
3457 DoFour(pxor, xmm_key_first);
3458 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3459 DoFour(aesdec, as_XMMRegister(rnum));
3460 }
3461 DoFour(aesdeclast, xmm_key_last);
3462 // for each result, xor with the r vector of previous cipher block
3463 __ pxor(xmm_result0, xmm_prev_block_cipher);
3464 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 0*AESBlockSize));
3465 __ pxor(xmm_result1, xmm_prev_block_cipher);
3466 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 1*AESBlockSize));
3467 __ pxor(xmm_result2, xmm_prev_block_cipher);
3468 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 2*AESBlockSize));
3469 __ pxor(xmm_result3, xmm_prev_block_cipher);
3470 __ movdqu(xmm_prev_block_cipher, Address(from, pos, Address::times_1, 3*AESBlockSize)); // this will carry over to next set of blocks
3472 __ movdqu(Address(to, pos, Address::times_1, 0*AESBlockSize), xmm_result0); // store 4 results into the next 64 bytes of output
3473 __ movdqu(Address(to, pos, Address::times_1, 1*AESBlockSize), xmm_result1);
3474 __ movdqu(Address(to, pos, Address::times_1, 2*AESBlockSize), xmm_result2);
3475 __ movdqu(Address(to, pos, Address::times_1, 3*AESBlockSize), xmm_result3);
3477 __ addptr(pos, 4*AESBlockSize);
3478 __ subptr(len_reg, 4*AESBlockSize);
3479 __ jmp(L_multiBlock_loopTop_128);
3481 // registers used in the non-parallelized loops
3482 // xmm register assignments for the loops below
3483 const XMMRegister xmm_result = xmm0;
3484 const XMMRegister xmm_prev_block_cipher_save = xmm2;
3485 const XMMRegister xmm_key11 = xmm3;
3486 const XMMRegister xmm_key12 = xmm4;
3487 const XMMRegister xmm_temp = xmm4;
3489 __ align(OptoLoopAlignment);
3490 __ BIND(L_singleBlock_loopTop_128);
3491 __ cmpptr(len_reg, 0); // any blocks left??
3492 __ jcc(Assembler::equal, L_exit);
3493 __ movdqu(xmm_result, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of cipher input
3494 __ movdqa(xmm_prev_block_cipher_save, xmm_result); // save for next r vector
3495 __ pxor (xmm_result, xmm_key_first); // do the aes dec rounds
3496 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3497 __ aesdec(xmm_result, as_XMMRegister(rnum));
3498 }
3499 __ aesdeclast(xmm_result, xmm_key_last);
3500 __ pxor (xmm_result, xmm_prev_block_cipher); // xor with the current r vector
3501 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3502 // no need to store r to memory until we exit
3503 __ movdqa(xmm_prev_block_cipher, xmm_prev_block_cipher_save); // set up next r vector with cipher input from this block
3505 __ addptr(pos, AESBlockSize);
3506 __ subptr(len_reg, AESBlockSize);
3507 __ jmp(L_singleBlock_loopTop_128);
3510 __ BIND(L_exit);
3511 __ movdqu(Address(rvec, 0), xmm_prev_block_cipher); // final value of r stored in rvec of CipherBlockChaining object
3512 #ifdef _WIN64
3513 // restore regs belonging to calling function
3514 for (int i = 6; i <= XMM_REG_NUM_KEY_LAST; i++) {
3515 __ movdqu(as_XMMRegister(i), xmm_save(i));
3516 }
3517 #endif
3518 __ movl(rax, 0); // return 0 (why?)
3519 __ leave(); // required for proper stackwalking of RuntimeStub frame
3520 __ ret(0);
3523 __ BIND(L_key_192_256);
3524 // here rax = len in ints of AESCrypt.KLE array (52=192, or 60=256)
3525 load_key(xmm_key11, key, 0xb0);
3526 __ cmpl(rax, 52);
3527 __ jcc(Assembler::notEqual, L_key_256);
3529 // 192-bit code follows here (could be optimized to use parallelism)
3530 load_key(xmm_key12, key, 0xc0); // 192-bit key goes up to c0
3531 __ movptr(pos, 0);
3532 __ align(OptoLoopAlignment);
3534 __ BIND(L_singleBlock_loopTop_192);
3535 __ movdqu(xmm_result, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of cipher input
3536 __ movdqa(xmm_prev_block_cipher_save, xmm_result); // save for next r vector
3537 __ pxor (xmm_result, xmm_key_first); // do the aes dec rounds
3538 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3539 __ aesdec(xmm_result, as_XMMRegister(rnum));
3540 }
3541 __ aesdec(xmm_result, xmm_key11);
3542 __ aesdec(xmm_result, xmm_key12);
3543 __ aesdeclast(xmm_result, xmm_key_last); // xmm15 always came from key+0
3544 __ pxor (xmm_result, xmm_prev_block_cipher); // xor with the current r vector
3545 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3546 // no need to store r to memory until we exit
3547 __ movdqa(xmm_prev_block_cipher, xmm_prev_block_cipher_save); // set up next r vector with cipher input from this block
3548 __ addptr(pos, AESBlockSize);
3549 __ subptr(len_reg, AESBlockSize);
3550 __ jcc(Assembler::notEqual,L_singleBlock_loopTop_192);
3551 __ jmp(L_exit);
3553 __ BIND(L_key_256);
3554 // 256-bit code follows here (could be optimized to use parallelism)
3555 __ movptr(pos, 0);
3556 __ align(OptoLoopAlignment);
3558 __ BIND(L_singleBlock_loopTop_256);
3559 __ movdqu(xmm_result, Address(from, pos, Address::times_1, 0)); // get next 16 bytes of cipher input
3560 __ movdqa(xmm_prev_block_cipher_save, xmm_result); // save for next r vector
3561 __ pxor (xmm_result, xmm_key_first); // do the aes dec rounds
3562 for (int rnum = XMM_REG_NUM_KEY_FIRST + 1; rnum <= XMM_REG_NUM_KEY_LAST - 1; rnum++) {
3563 __ aesdec(xmm_result, as_XMMRegister(rnum));
3564 }
3565 __ aesdec(xmm_result, xmm_key11);
3566 load_key(xmm_temp, key, 0xc0);
3567 __ aesdec(xmm_result, xmm_temp);
3568 load_key(xmm_temp, key, 0xd0);
3569 __ aesdec(xmm_result, xmm_temp);
3570 load_key(xmm_temp, key, 0xe0); // 256-bit key goes up to e0
3571 __ aesdec(xmm_result, xmm_temp);
3572 __ aesdeclast(xmm_result, xmm_key_last); // xmm15 came from key+0
3573 __ pxor (xmm_result, xmm_prev_block_cipher); // xor with the current r vector
3574 __ movdqu(Address(to, pos, Address::times_1, 0), xmm_result); // store into the next 16 bytes of output
3575 // no need to store r to memory until we exit
3576 __ movdqa(xmm_prev_block_cipher, xmm_prev_block_cipher_save); // set up next r vector with cipher input from this block
3577 __ addptr(pos, AESBlockSize);
3578 __ subptr(len_reg, AESBlockSize);
3579 __ jcc(Assembler::notEqual,L_singleBlock_loopTop_256);
3580 __ jmp(L_exit);
3582 return start;
3583 }
3587 #undef __
3588 #define __ masm->
3590 // Continuation point for throwing of implicit exceptions that are
3591 // not handled in the current activation. Fabricates an exception
3592 // oop and initiates normal exception dispatching in this
3593 // frame. Since we need to preserve callee-saved values (currently
3594 // only for C2, but done for C1 as well) we need a callee-saved oop
3595 // map and therefore have to make these stubs into RuntimeStubs
3596 // rather than BufferBlobs. If the compiler needs all registers to
3597 // be preserved between the fault point and the exception handler
3598 // then it must assume responsibility for that in
3599 // AbstractCompiler::continuation_for_implicit_null_exception or
3600 // continuation_for_implicit_division_by_zero_exception. All other
3601 // implicit exceptions (e.g., NullPointerException or
3602 // AbstractMethodError on entry) are either at call sites or
3603 // otherwise assume that stack unwinding will be initiated, so
3604 // caller saved registers were assumed volatile in the compiler.
3605 address generate_throw_exception(const char* name,
3606 address runtime_entry,
3607 Register arg1 = noreg,
3608 Register arg2 = noreg) {
3609 // Information about frame layout at time of blocking runtime call.
3610 // Note that we only have to preserve callee-saved registers since
3611 // the compilers are responsible for supplying a continuation point
3612 // if they expect all registers to be preserved.
3613 enum layout {
3614 rbp_off = frame::arg_reg_save_area_bytes/BytesPerInt,
3615 rbp_off2,
3616 return_off,
3617 return_off2,
3618 framesize // inclusive of return address
3619 };
3621 int insts_size = 512;
3622 int locs_size = 64;
3624 CodeBuffer code(name, insts_size, locs_size);
3625 OopMapSet* oop_maps = new OopMapSet();
3626 MacroAssembler* masm = new MacroAssembler(&code);
3628 address start = __ pc();
3630 // This is an inlined and slightly modified version of call_VM
3631 // which has the ability to fetch the return PC out of
3632 // thread-local storage and also sets up last_Java_sp slightly
3633 // differently than the real call_VM
3635 __ enter(); // required for proper stackwalking of RuntimeStub frame
3637 assert(is_even(framesize/2), "sp not 16-byte aligned");
3639 // return address and rbp are already in place
3640 __ subptr(rsp, (framesize-4) << LogBytesPerInt); // prolog
3642 int frame_complete = __ pc() - start;
3644 // Set up last_Java_sp and last_Java_fp
3645 address the_pc = __ pc();
3646 __ set_last_Java_frame(rsp, rbp, the_pc);
3647 __ andptr(rsp, -(StackAlignmentInBytes)); // Align stack
3649 // Call runtime
3650 if (arg1 != noreg) {
3651 assert(arg2 != c_rarg1, "clobbered");
3652 __ movptr(c_rarg1, arg1);
3653 }
3654 if (arg2 != noreg) {
3655 __ movptr(c_rarg2, arg2);
3656 }
3657 __ movptr(c_rarg0, r15_thread);
3658 BLOCK_COMMENT("call runtime_entry");
3659 __ call(RuntimeAddress(runtime_entry));
3661 // Generate oop map
3662 OopMap* map = new OopMap(framesize, 0);
3664 oop_maps->add_gc_map(the_pc - start, map);
3666 __ reset_last_Java_frame(true, true);
3668 __ leave(); // required for proper stackwalking of RuntimeStub frame
3670 // check for pending exceptions
3671 #ifdef ASSERT
3672 Label L;
3673 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()),
3674 (int32_t) NULL_WORD);
3675 __ jcc(Assembler::notEqual, L);
3676 __ should_not_reach_here();
3677 __ bind(L);
3678 #endif // ASSERT
3679 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3682 // codeBlob framesize is in words (not VMRegImpl::slot_size)
3683 RuntimeStub* stub =
3684 RuntimeStub::new_runtime_stub(name,
3685 &code,
3686 frame_complete,
3687 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3688 oop_maps, false);
3689 return stub->entry_point();
3690 }
3692 // Initialization
3693 void generate_initial() {
3694 // Generates all stubs and initializes the entry points
3696 // This platform-specific stub is needed by generate_call_stub()
3697 StubRoutines::x86::_mxcsr_std = generate_fp_mask("mxcsr_std", 0x0000000000001F80);
3699 // entry points that exist in all platforms Note: This is code
3700 // that could be shared among different platforms - however the
3701 // benefit seems to be smaller than the disadvantage of having a
3702 // much more complicated generator structure. See also comment in
3703 // stubRoutines.hpp.
3705 StubRoutines::_forward_exception_entry = generate_forward_exception();
3707 StubRoutines::_call_stub_entry =
3708 generate_call_stub(StubRoutines::_call_stub_return_address);
3710 // is referenced by megamorphic call
3711 StubRoutines::_catch_exception_entry = generate_catch_exception();
3713 // atomic calls
3714 StubRoutines::_atomic_xchg_entry = generate_atomic_xchg();
3715 StubRoutines::_atomic_xchg_ptr_entry = generate_atomic_xchg_ptr();
3716 StubRoutines::_atomic_cmpxchg_entry = generate_atomic_cmpxchg();
3717 StubRoutines::_atomic_cmpxchg_long_entry = generate_atomic_cmpxchg_long();
3718 StubRoutines::_atomic_add_entry = generate_atomic_add();
3719 StubRoutines::_atomic_add_ptr_entry = generate_atomic_add_ptr();
3720 StubRoutines::_fence_entry = generate_orderaccess_fence();
3722 StubRoutines::_handler_for_unsafe_access_entry =
3723 generate_handler_for_unsafe_access();
3725 // platform dependent
3726 StubRoutines::x86::_get_previous_fp_entry = generate_get_previous_fp();
3727 StubRoutines::x86::_get_previous_sp_entry = generate_get_previous_sp();
3729 StubRoutines::x86::_verify_mxcsr_entry = generate_verify_mxcsr();
3731 // Build this early so it's available for the interpreter.
3732 StubRoutines::_throw_StackOverflowError_entry =
3733 generate_throw_exception("StackOverflowError throw_exception",
3734 CAST_FROM_FN_PTR(address,
3735 SharedRuntime::
3736 throw_StackOverflowError));
3737 }
3739 void generate_all() {
3740 // Generates all stubs and initializes the entry points
3742 // These entry points require SharedInfo::stack0 to be set up in
3743 // non-core builds and need to be relocatable, so they each
3744 // fabricate a RuntimeStub internally.
3745 StubRoutines::_throw_AbstractMethodError_entry =
3746 generate_throw_exception("AbstractMethodError throw_exception",
3747 CAST_FROM_FN_PTR(address,
3748 SharedRuntime::
3749 throw_AbstractMethodError));
3751 StubRoutines::_throw_IncompatibleClassChangeError_entry =
3752 generate_throw_exception("IncompatibleClassChangeError throw_exception",
3753 CAST_FROM_FN_PTR(address,
3754 SharedRuntime::
3755 throw_IncompatibleClassChangeError));
3757 StubRoutines::_throw_NullPointerException_at_call_entry =
3758 generate_throw_exception("NullPointerException at call throw_exception",
3759 CAST_FROM_FN_PTR(address,
3760 SharedRuntime::
3761 throw_NullPointerException_at_call));
3763 // entry points that are platform specific
3764 StubRoutines::x86::_f2i_fixup = generate_f2i_fixup();
3765 StubRoutines::x86::_f2l_fixup = generate_f2l_fixup();
3766 StubRoutines::x86::_d2i_fixup = generate_d2i_fixup();
3767 StubRoutines::x86::_d2l_fixup = generate_d2l_fixup();
3769 StubRoutines::x86::_float_sign_mask = generate_fp_mask("float_sign_mask", 0x7FFFFFFF7FFFFFFF);
3770 StubRoutines::x86::_float_sign_flip = generate_fp_mask("float_sign_flip", 0x8000000080000000);
3771 StubRoutines::x86::_double_sign_mask = generate_fp_mask("double_sign_mask", 0x7FFFFFFFFFFFFFFF);
3772 StubRoutines::x86::_double_sign_flip = generate_fp_mask("double_sign_flip", 0x8000000000000000);
3774 // support for verify_oop (must happen after universe_init)
3775 StubRoutines::_verify_oop_subroutine_entry = generate_verify_oop();
3777 // arraycopy stubs used by compilers
3778 generate_arraycopy_stubs();
3780 generate_math_stubs();
3782 // don't bother generating these AES intrinsic stubs unless global flag is set
3783 if (UseAESIntrinsics) {
3784 StubRoutines::x86::_key_shuffle_mask_addr = generate_key_shuffle_mask(); // needed by the others
3786 StubRoutines::_aescrypt_encryptBlock = generate_aescrypt_encryptBlock();
3787 StubRoutines::_aescrypt_decryptBlock = generate_aescrypt_decryptBlock();
3788 StubRoutines::_cipherBlockChaining_encryptAESCrypt = generate_cipherBlockChaining_encryptAESCrypt();
3789 StubRoutines::_cipherBlockChaining_decryptAESCrypt = generate_cipherBlockChaining_decryptAESCrypt_Parallel();
3790 }
3791 }
3793 public:
3794 StubGenerator(CodeBuffer* code, bool all) : StubCodeGenerator(code) {
3795 if (all) {
3796 generate_all();
3797 } else {
3798 generate_initial();
3799 }
3800 }
3801 }; // end class declaration
3803 void StubGenerator_generate(CodeBuffer* code, bool all) {
3804 StubGenerator g(code, all);
3805 }