1 /*
2 * Copyright (c) 2003, 2023, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/assembler.hpp"
27 #include "asm/macroAssembler.hpp"
28 #include "classfile/vmIntrinsics.hpp"
29 #include "compiler/oopMap.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/barrierSetAssembler.hpp"
32 #include "gc/shared/barrierSetNMethod.hpp"
33 #include "gc/shared/gc_globals.hpp"
34 #include "memory/universe.hpp"
35 #include "prims/jvmtiExport.hpp"
36 #include "prims/upcallLinker.hpp"
37 #include "runtime/arguments.hpp"
38 #include "runtime/javaThread.hpp"
39 #include "runtime/sharedRuntime.hpp"
40 #include "runtime/stubRoutines.hpp"
41 #include "utilities/macros.hpp"
42 #include "vmreg_x86.inline.hpp"
43 #include "stubGenerator_x86_64.hpp"
44 #ifdef COMPILER2
45 #include "opto/runtime.hpp"
46 #include "opto/c2_globals.hpp"
47 #endif
48 #if INCLUDE_JVMCI
49 #include "jvmci/jvmci_globals.hpp"
50 #endif
51 #if INCLUDE_JFR
52 #include "jfr/support/jfrIntrinsics.hpp"
53 #endif
54
55 // For a more detailed description of the stub routine structure
56 // see the comment in stubRoutines.hpp
57
58 #define __ _masm->
59 #define TIMES_OOP (UseCompressedOops ? Address::times_4 : Address::times_8)
60
61 #ifdef PRODUCT
62 #define BLOCK_COMMENT(str) /* nothing */
63 #else
64 #define BLOCK_COMMENT(str) __ block_comment(str)
65 #endif // PRODUCT
66
67 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
68
69 //
70 // Linux Arguments:
71 // c_rarg0: call wrapper address address
72 // c_rarg1: result address
73 // c_rarg2: result type BasicType
74 // c_rarg3: method Method*
75 // c_rarg4: (interpreter) entry point address
76 // c_rarg5: parameters intptr_t*
77 // 16(rbp): parameter size (in words) int
78 // 24(rbp): thread Thread*
79 //
80 // [ return_from_Java ] <--- rsp
81 // [ argument word n ]
82 // ...
83 // -12 [ argument word 1 ]
84 // -11 [ saved r15 ] <--- rsp_after_call
85 // -10 [ saved r14 ]
86 // -9 [ saved r13 ]
87 // -8 [ saved r12 ]
88 // -7 [ saved rbx ]
89 // -6 [ call wrapper ]
90 // -5 [ result ]
91 // -4 [ result type ]
92 // -3 [ method ]
93 // -2 [ entry point ]
94 // -1 [ parameters ]
95 // 0 [ saved rbp ] <--- rbp
96 // 1 [ return address ]
97 // 2 [ parameter size ]
98 // 3 [ thread ]
99 //
100 // Windows Arguments:
101 // c_rarg0: call wrapper address address
102 // c_rarg1: result address
103 // c_rarg2: result type BasicType
104 // c_rarg3: method Method*
105 // 48(rbp): (interpreter) entry point address
106 // 56(rbp): parameters intptr_t*
107 // 64(rbp): parameter size (in words) int
108 // 72(rbp): thread Thread*
109 //
110 // [ return_from_Java ] <--- rsp
111 // [ argument word n ]
112 // ...
113 // -28 [ argument word 1 ]
114 // -27 [ saved xmm15 ] <--- rsp after_call
115 // [ saved xmm7-xmm14 ]
116 // -9 [ saved xmm6 ] (each xmm register takes 2 slots)
117 // -7 [ saved r15 ]
118 // -6 [ saved r14 ]
119 // -5 [ saved r13 ]
120 // -4 [ saved r12 ]
121 // -3 [ saved rdi ]
122 // -2 [ saved rsi ]
123 // -1 [ saved rbx ]
124 // 0 [ saved rbp ] <--- rbp
125 // 1 [ return address ]
126 // 2 [ call wrapper ]
127 // 3 [ result ]
128 // 4 [ result type ]
129 // 5 [ method ]
130 // 6 [ entry point ]
131 // 7 [ parameters ]
132 // 8 [ parameter size ]
133 // 9 [ thread ]
134 //
135 // Windows reserves the callers stack space for arguments 1-4.
136 // We spill c_rarg0-c_rarg3 to this space.
137
138 // Call stub stack layout word offsets from rbp
139 #ifdef _WIN64
140 enum call_stub_layout {
141 xmm_save_first = 6, // save from xmm6
142 xmm_save_last = 15, // to xmm15
143 xmm_save_base = -9,
144 rsp_after_call_off = xmm_save_base - 2 * (xmm_save_last - xmm_save_first), // -27
145 r15_off = -7,
146 r14_off = -6,
147 r13_off = -5,
148 r12_off = -4,
149 rdi_off = -3,
150 rsi_off = -2,
151 rbx_off = -1,
152 rbp_off = 0,
153 retaddr_off = 1,
154 call_wrapper_off = 2,
155 result_off = 3,
156 result_type_off = 4,
157 method_off = 5,
158 entry_point_off = 6,
159 parameters_off = 7,
160 parameter_size_off = 8,
161 thread_off = 9
162 };
163
164 static Address xmm_save(int reg) {
165 assert(reg >= xmm_save_first && reg <= xmm_save_last, "XMM register number out of range");
166 return Address(rbp, (xmm_save_base - (reg - xmm_save_first) * 2) * wordSize);
167 }
168 #else // !_WIN64
169 enum call_stub_layout {
170 rsp_after_call_off = -12,
171 mxcsr_off = rsp_after_call_off,
172 r15_off = -11,
173 r14_off = -10,
174 r13_off = -9,
175 r12_off = -8,
176 rbx_off = -7,
177 call_wrapper_off = -6,
178 result_off = -5,
179 result_type_off = -4,
180 method_off = -3,
181 entry_point_off = -2,
182 parameters_off = -1,
183 rbp_off = 0,
184 retaddr_off = 1,
185 parameter_size_off = 2,
186 thread_off = 3
187 };
188 #endif // _WIN64
189
190 address StubGenerator::generate_call_stub(address& return_address) {
191
192 assert((int)frame::entry_frame_after_call_words == -(int)rsp_after_call_off + 1 &&
193 (int)frame::entry_frame_call_wrapper_offset == (int)call_wrapper_off,
194 "adjust this code");
195 StubCodeMark mark(this, "StubRoutines", "call_stub");
196 address start = __ pc();
197
198 // same as in generate_catch_exception()!
199 const Address rsp_after_call(rbp, rsp_after_call_off * wordSize);
200
201 const Address call_wrapper (rbp, call_wrapper_off * wordSize);
202 const Address result (rbp, result_off * wordSize);
203 const Address result_type (rbp, result_type_off * wordSize);
204 const Address method (rbp, method_off * wordSize);
205 const Address entry_point (rbp, entry_point_off * wordSize);
206 const Address parameters (rbp, parameters_off * wordSize);
207 const Address parameter_size(rbp, parameter_size_off * wordSize);
208
209 // same as in generate_catch_exception()!
210 const Address thread (rbp, thread_off * wordSize);
211
212 const Address r15_save(rbp, r15_off * wordSize);
213 const Address r14_save(rbp, r14_off * wordSize);
214 const Address r13_save(rbp, r13_off * wordSize);
215 const Address r12_save(rbp, r12_off * wordSize);
216 const Address rbx_save(rbp, rbx_off * wordSize);
217
218 // stub code
219 __ enter();
220 __ subptr(rsp, -rsp_after_call_off * wordSize);
221
222 // save register parameters
223 #ifndef _WIN64
224 __ movptr(parameters, c_rarg5); // parameters
225 __ movptr(entry_point, c_rarg4); // entry_point
226 #endif
227
228 __ movptr(method, c_rarg3); // method
229 __ movl(result_type, c_rarg2); // result type
230 __ movptr(result, c_rarg1); // result
231 __ movptr(call_wrapper, c_rarg0); // call wrapper
232
233 // save regs belonging to calling function
234 __ movptr(rbx_save, rbx);
235 __ movptr(r12_save, r12);
236 __ movptr(r13_save, r13);
237 __ movptr(r14_save, r14);
238 __ movptr(r15_save, r15);
239
240 #ifdef _WIN64
241 int last_reg = 15;
242 for (int i = xmm_save_first; i <= last_reg; i++) {
243 __ movdqu(xmm_save(i), as_XMMRegister(i));
244 }
245
246 const Address rdi_save(rbp, rdi_off * wordSize);
247 const Address rsi_save(rbp, rsi_off * wordSize);
248
249 __ movptr(rsi_save, rsi);
250 __ movptr(rdi_save, rdi);
251 #else
252 const Address mxcsr_save(rbp, mxcsr_off * wordSize);
253 {
254 Label skip_ldmx;
255 __ stmxcsr(mxcsr_save);
256 __ movl(rax, mxcsr_save);
257 __ andl(rax, 0xFFC0); // Mask out any pending exceptions (only check control and mask bits)
258 ExternalAddress mxcsr_std(StubRoutines::x86::addr_mxcsr_std());
259 __ cmp32(rax, mxcsr_std, rscratch1);
260 __ jcc(Assembler::equal, skip_ldmx);
261 __ ldmxcsr(mxcsr_std, rscratch1);
262 __ bind(skip_ldmx);
263 }
264 #endif
265
266 // Load up thread register
267 __ movptr(r15_thread, thread);
268 __ reinit_heapbase();
269
270 #ifdef ASSERT
271 // make sure we have no pending exceptions
272 {
273 Label L;
274 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), NULL_WORD);
275 __ jcc(Assembler::equal, L);
276 __ stop("StubRoutines::call_stub: entered with pending exception");
277 __ bind(L);
278 }
279 #endif
280
281 // pass parameters if any
282 BLOCK_COMMENT("pass parameters if any");
283 Label parameters_done;
284 __ movl(c_rarg3, parameter_size);
285 __ testl(c_rarg3, c_rarg3);
286 __ jcc(Assembler::zero, parameters_done);
287
288 Label loop;
289 __ movptr(c_rarg2, parameters); // parameter pointer
290 __ movl(c_rarg1, c_rarg3); // parameter counter is in c_rarg1
291 __ BIND(loop);
292 __ movptr(rax, Address(c_rarg2, 0));// get parameter
293 __ addptr(c_rarg2, wordSize); // advance to next parameter
294 __ decrementl(c_rarg1); // decrement counter
295 __ push(rax); // pass parameter
296 __ jcc(Assembler::notZero, loop);
297
298 // call Java function
299 __ BIND(parameters_done);
300 __ movptr(rbx, method); // get Method*
301 __ movptr(c_rarg1, entry_point); // get entry_point
302 __ mov(r13, rsp); // set sender sp
303 BLOCK_COMMENT("call Java function");
304 __ call(c_rarg1);
305
306 BLOCK_COMMENT("call_stub_return_address:");
307 return_address = __ pc();
308
309 // store result depending on type (everything that is not
310 // T_OBJECT, T_LONG, T_FLOAT or T_DOUBLE is treated as T_INT)
311 __ movptr(r13, result);
312 Label is_long, is_float, is_double, check_prim, exit;
313 __ movl(rbx, result_type);
314 __ cmpl(rbx, T_OBJECT);
315 __ jcc(Assembler::equal, check_prim);
316 __ cmpl(rbx, T_LONG);
317 __ jcc(Assembler::equal, is_long);
318 __ cmpl(rbx, T_FLOAT);
319 __ jcc(Assembler::equal, is_float);
320 __ cmpl(rbx, T_DOUBLE);
321 __ jcc(Assembler::equal, is_double);
322 #ifdef ASSERT
323 // make sure the type is INT
324 {
325 Label L;
326 __ cmpl(rbx, T_INT);
327 __ jcc(Assembler::equal, L);
328 __ stop("StubRoutines::call_stub: unexpected result type");
329 __ bind(L);
330 }
331 #endif
332
333 // handle T_INT case
334 __ movl(Address(r13, 0), rax);
335
336 __ BIND(exit);
337
338 // pop parameters
339 __ lea(rsp, rsp_after_call);
340
341 #ifdef ASSERT
342 // verify that threads correspond
343 {
344 Label L1, L2, L3;
345 __ cmpptr(r15_thread, thread);
346 __ jcc(Assembler::equal, L1);
347 __ stop("StubRoutines::call_stub: r15_thread is corrupted");
348 __ bind(L1);
349 __ get_thread(rbx);
350 __ cmpptr(r15_thread, thread);
351 __ jcc(Assembler::equal, L2);
352 __ stop("StubRoutines::call_stub: r15_thread is modified by call");
353 __ bind(L2);
354 __ cmpptr(r15_thread, rbx);
355 __ jcc(Assembler::equal, L3);
356 __ stop("StubRoutines::call_stub: threads must correspond");
357 __ bind(L3);
358 }
359 #endif
360
361 __ pop_cont_fastpath();
362
363 // restore regs belonging to calling function
364 #ifdef _WIN64
365 // emit the restores for xmm regs
366 for (int i = xmm_save_first; i <= last_reg; i++) {
367 __ movdqu(as_XMMRegister(i), xmm_save(i));
368 }
369 #endif
370 __ movptr(r15, r15_save);
371 __ movptr(r14, r14_save);
372 __ movptr(r13, r13_save);
373 __ movptr(r12, r12_save);
374 __ movptr(rbx, rbx_save);
375
376 #ifdef _WIN64
377 __ movptr(rdi, rdi_save);
378 __ movptr(rsi, rsi_save);
379 #else
380 __ ldmxcsr(mxcsr_save);
381 #endif
382
383 // restore rsp
384 __ addptr(rsp, -rsp_after_call_off * wordSize);
385
386 // return
387 __ vzeroupper();
388 __ pop(rbp);
389 __ ret(0);
390
391 // handle return types different from T_INT
392 __ BIND(check_prim);
393 if (InlineTypeReturnedAsFields) {
394 // Check for scalarized return value
395 __ testptr(rax, 1);
396 __ jcc(Assembler::zero, is_long);
397 // Load pack handler address
398 __ andptr(rax, -2);
399 __ movptr(rax, Address(rax, InstanceKlass::adr_inlineklass_fixed_block_offset()));
400 __ movptr(rbx, Address(rax, InlineKlass::pack_handler_jobject_offset()));
401 // Call pack handler to initialize the buffer
402 __ call(rbx);
403 __ jmp(exit);
404 }
405 __ BIND(is_long);
406 __ movq(Address(r13, 0), rax);
407 __ jmp(exit);
408
409 __ BIND(is_float);
410 __ movflt(Address(r13, 0), xmm0);
411 __ jmp(exit);
412
413 __ BIND(is_double);
414 __ movdbl(Address(r13, 0), xmm0);
415 __ jmp(exit);
416
417 return start;
418 }
419
420 // Return point for a Java call if there's an exception thrown in
421 // Java code. The exception is caught and transformed into a
422 // pending exception stored in JavaThread that can be tested from
423 // within the VM.
424 //
425 // Note: Usually the parameters are removed by the callee. In case
426 // of an exception crossing an activation frame boundary, that is
427 // not the case if the callee is compiled code => need to setup the
428 // rsp.
429 //
430 // rax: exception oop
431
432 address StubGenerator::generate_catch_exception() {
433 StubCodeMark mark(this, "StubRoutines", "catch_exception");
434 address start = __ pc();
435
436 // same as in generate_call_stub():
437 const Address rsp_after_call(rbp, rsp_after_call_off * wordSize);
438 const Address thread (rbp, thread_off * wordSize);
439
440 #ifdef ASSERT
441 // verify that threads correspond
442 {
443 Label L1, L2, L3;
444 __ cmpptr(r15_thread, thread);
445 __ jcc(Assembler::equal, L1);
446 __ stop("StubRoutines::catch_exception: r15_thread is corrupted");
447 __ bind(L1);
448 __ get_thread(rbx);
449 __ cmpptr(r15_thread, thread);
450 __ jcc(Assembler::equal, L2);
451 __ stop("StubRoutines::catch_exception: r15_thread is modified by call");
452 __ bind(L2);
453 __ cmpptr(r15_thread, rbx);
454 __ jcc(Assembler::equal, L3);
455 __ stop("StubRoutines::catch_exception: threads must correspond");
456 __ bind(L3);
457 }
458 #endif
459
460 // set pending exception
461 __ verify_oop(rax);
462
463 __ movptr(Address(r15_thread, Thread::pending_exception_offset()), rax);
464 __ lea(rscratch1, ExternalAddress((address)__FILE__));
465 __ movptr(Address(r15_thread, Thread::exception_file_offset()), rscratch1);
466 __ movl(Address(r15_thread, Thread::exception_line_offset()), (int) __LINE__);
467
468 // complete return to VM
469 assert(StubRoutines::_call_stub_return_address != nullptr,
470 "_call_stub_return_address must have been generated before");
471 __ jump(RuntimeAddress(StubRoutines::_call_stub_return_address));
472
473 return start;
474 }
475
476 // Continuation point for runtime calls returning with a pending
477 // exception. The pending exception check happened in the runtime
478 // or native call stub. The pending exception in Thread is
479 // converted into a Java-level exception.
480 //
481 // Contract with Java-level exception handlers:
482 // rax: exception
483 // rdx: throwing pc
484 //
485 // NOTE: At entry of this stub, exception-pc must be on stack !!
486
487 address StubGenerator::generate_forward_exception() {
488 StubCodeMark mark(this, "StubRoutines", "forward exception");
489 address start = __ pc();
490
491 // Upon entry, the sp points to the return address returning into
492 // Java (interpreted or compiled) code; i.e., the return address
493 // becomes the throwing pc.
494 //
495 // Arguments pushed before the runtime call are still on the stack
496 // but the exception handler will reset the stack pointer ->
497 // ignore them. A potential result in registers can be ignored as
498 // well.
499
500 #ifdef ASSERT
501 // make sure this code is only executed if there is a pending exception
502 {
503 Label L;
504 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), NULL_WORD);
505 __ jcc(Assembler::notEqual, L);
506 __ stop("StubRoutines::forward exception: no pending exception (1)");
507 __ bind(L);
508 }
509 #endif
510
511 // compute exception handler into rbx
512 __ movptr(c_rarg0, Address(rsp, 0));
513 BLOCK_COMMENT("call exception_handler_for_return_address");
514 __ call_VM_leaf(CAST_FROM_FN_PTR(address,
515 SharedRuntime::exception_handler_for_return_address),
516 r15_thread, c_rarg0);
517 __ mov(rbx, rax);
518
519 // setup rax & rdx, remove return address & clear pending exception
520 __ pop(rdx);
521 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset()));
522 __ movptr(Address(r15_thread, Thread::pending_exception_offset()), NULL_WORD);
523
524 #ifdef ASSERT
525 // make sure exception is set
526 {
527 Label L;
528 __ testptr(rax, rax);
529 __ jcc(Assembler::notEqual, L);
530 __ stop("StubRoutines::forward exception: no pending exception (2)");
531 __ bind(L);
532 }
533 #endif
534
535 // continue at exception handler (return address removed)
536 // rax: exception
537 // rbx: exception handler
538 // rdx: throwing pc
539 __ verify_oop(rax);
540 __ jmp(rbx);
541
542 return start;
543 }
544
545 // Support for intptr_t OrderAccess::fence()
546 //
547 // Arguments :
548 //
549 // Result:
550 address StubGenerator::generate_orderaccess_fence() {
551 StubCodeMark mark(this, "StubRoutines", "orderaccess_fence");
552 address start = __ pc();
553
554 __ membar(Assembler::StoreLoad);
555 __ ret(0);
556
557 return start;
558 }
559
560
561 // Support for intptr_t get_previous_sp()
562 //
563 // This routine is used to find the previous stack pointer for the
564 // caller.
565 address StubGenerator::generate_get_previous_sp() {
566 StubCodeMark mark(this, "StubRoutines", "get_previous_sp");
567 address start = __ pc();
568
569 __ movptr(rax, rsp);
570 __ addptr(rax, 8); // return address is at the top of the stack.
571 __ ret(0);
572
573 return start;
574 }
575
576 //----------------------------------------------------------------------------------------------------
577 // Support for void verify_mxcsr()
578 //
579 // This routine is used with -Xcheck:jni to verify that native
580 // JNI code does not return to Java code without restoring the
581 // MXCSR register to our expected state.
582
583 address StubGenerator::generate_verify_mxcsr() {
584 StubCodeMark mark(this, "StubRoutines", "verify_mxcsr");
585 address start = __ pc();
586
587 const Address mxcsr_save(rsp, 0);
588
589 if (CheckJNICalls) {
590 Label ok_ret;
591 ExternalAddress mxcsr_std(StubRoutines::x86::addr_mxcsr_std());
592 __ push(rax);
593 __ subptr(rsp, wordSize); // allocate a temp location
594 __ stmxcsr(mxcsr_save);
595 __ movl(rax, mxcsr_save);
596 __ andl(rax, 0xFFC0); // Mask out any pending exceptions (only check control and mask bits)
597 __ cmp32(rax, mxcsr_std, rscratch1);
598 __ jcc(Assembler::equal, ok_ret);
599
600 __ warn("MXCSR changed by native JNI code, use -XX:+RestoreMXCSROnJNICall");
601
602 __ ldmxcsr(mxcsr_std, rscratch1);
603
604 __ bind(ok_ret);
605 __ addptr(rsp, wordSize);
606 __ pop(rax);
607 }
608
609 __ ret(0);
610
611 return start;
612 }
613
614 address StubGenerator::generate_f2i_fixup() {
615 StubCodeMark mark(this, "StubRoutines", "f2i_fixup");
616 Address inout(rsp, 5 * wordSize); // return address + 4 saves
617
618 address start = __ pc();
619
620 Label L;
621
622 __ push(rax);
623 __ push(c_rarg3);
624 __ push(c_rarg2);
625 __ push(c_rarg1);
626
627 __ movl(rax, 0x7f800000);
628 __ xorl(c_rarg3, c_rarg3);
629 __ movl(c_rarg2, inout);
630 __ movl(c_rarg1, c_rarg2);
631 __ andl(c_rarg1, 0x7fffffff);
632 __ cmpl(rax, c_rarg1); // NaN? -> 0
633 __ jcc(Assembler::negative, L);
634 __ testl(c_rarg2, c_rarg2); // signed ? min_jint : max_jint
635 __ movl(c_rarg3, 0x80000000);
636 __ movl(rax, 0x7fffffff);
637 __ cmovl(Assembler::positive, c_rarg3, rax);
638
639 __ bind(L);
640 __ movptr(inout, c_rarg3);
641
642 __ pop(c_rarg1);
643 __ pop(c_rarg2);
644 __ pop(c_rarg3);
645 __ pop(rax);
646
647 __ ret(0);
648
649 return start;
650 }
651
652 address StubGenerator::generate_f2l_fixup() {
653 StubCodeMark mark(this, "StubRoutines", "f2l_fixup");
654 Address inout(rsp, 5 * wordSize); // return address + 4 saves
655 address start = __ pc();
656
657 Label L;
658
659 __ push(rax);
660 __ push(c_rarg3);
661 __ push(c_rarg2);
662 __ push(c_rarg1);
663
664 __ movl(rax, 0x7f800000);
665 __ xorl(c_rarg3, c_rarg3);
666 __ movl(c_rarg2, inout);
667 __ movl(c_rarg1, c_rarg2);
668 __ andl(c_rarg1, 0x7fffffff);
669 __ cmpl(rax, c_rarg1); // NaN? -> 0
670 __ jcc(Assembler::negative, L);
671 __ testl(c_rarg2, c_rarg2); // signed ? min_jlong : max_jlong
672 __ mov64(c_rarg3, 0x8000000000000000);
673 __ mov64(rax, 0x7fffffffffffffff);
674 __ cmov(Assembler::positive, c_rarg3, rax);
675
676 __ bind(L);
677 __ movptr(inout, c_rarg3);
678
679 __ pop(c_rarg1);
680 __ pop(c_rarg2);
681 __ pop(c_rarg3);
682 __ pop(rax);
683
684 __ ret(0);
685
686 return start;
687 }
688
689 address StubGenerator::generate_d2i_fixup() {
690 StubCodeMark mark(this, "StubRoutines", "d2i_fixup");
691 Address inout(rsp, 6 * wordSize); // return address + 5 saves
692
693 address start = __ pc();
694
695 Label L;
696
697 __ push(rax);
698 __ push(c_rarg3);
699 __ push(c_rarg2);
700 __ push(c_rarg1);
701 __ push(c_rarg0);
702
703 __ movl(rax, 0x7ff00000);
704 __ movq(c_rarg2, inout);
705 __ movl(c_rarg3, c_rarg2);
706 __ mov(c_rarg1, c_rarg2);
707 __ mov(c_rarg0, c_rarg2);
708 __ negl(c_rarg3);
709 __ shrptr(c_rarg1, 0x20);
710 __ orl(c_rarg3, c_rarg2);
711 __ andl(c_rarg1, 0x7fffffff);
712 __ xorl(c_rarg2, c_rarg2);
713 __ shrl(c_rarg3, 0x1f);
714 __ orl(c_rarg1, c_rarg3);
715 __ cmpl(rax, c_rarg1);
716 __ jcc(Assembler::negative, L); // NaN -> 0
717 __ testptr(c_rarg0, c_rarg0); // signed ? min_jint : max_jint
718 __ movl(c_rarg2, 0x80000000);
719 __ movl(rax, 0x7fffffff);
720 __ cmov(Assembler::positive, c_rarg2, rax);
721
722 __ bind(L);
723 __ movptr(inout, c_rarg2);
724
725 __ pop(c_rarg0);
726 __ pop(c_rarg1);
727 __ pop(c_rarg2);
728 __ pop(c_rarg3);
729 __ pop(rax);
730
731 __ ret(0);
732
733 return start;
734 }
735
736 address StubGenerator::generate_d2l_fixup() {
737 StubCodeMark mark(this, "StubRoutines", "d2l_fixup");
738 Address inout(rsp, 6 * wordSize); // return address + 5 saves
739
740 address start = __ pc();
741
742 Label L;
743
744 __ push(rax);
745 __ push(c_rarg3);
746 __ push(c_rarg2);
747 __ push(c_rarg1);
748 __ push(c_rarg0);
749
750 __ movl(rax, 0x7ff00000);
751 __ movq(c_rarg2, inout);
752 __ movl(c_rarg3, c_rarg2);
753 __ mov(c_rarg1, c_rarg2);
754 __ mov(c_rarg0, c_rarg2);
755 __ negl(c_rarg3);
756 __ shrptr(c_rarg1, 0x20);
757 __ orl(c_rarg3, c_rarg2);
758 __ andl(c_rarg1, 0x7fffffff);
759 __ xorl(c_rarg2, c_rarg2);
760 __ shrl(c_rarg3, 0x1f);
761 __ orl(c_rarg1, c_rarg3);
762 __ cmpl(rax, c_rarg1);
763 __ jcc(Assembler::negative, L); // NaN -> 0
764 __ testq(c_rarg0, c_rarg0); // signed ? min_jlong : max_jlong
765 __ mov64(c_rarg2, 0x8000000000000000);
766 __ mov64(rax, 0x7fffffffffffffff);
767 __ cmovq(Assembler::positive, c_rarg2, rax);
768
769 __ bind(L);
770 __ movq(inout, c_rarg2);
771
772 __ pop(c_rarg0);
773 __ pop(c_rarg1);
774 __ pop(c_rarg2);
775 __ pop(c_rarg3);
776 __ pop(rax);
777
778 __ ret(0);
779
780 return start;
781 }
782
783 address StubGenerator::generate_count_leading_zeros_lut(const char *stub_name) {
784 __ align64();
785 StubCodeMark mark(this, "StubRoutines", stub_name);
786 address start = __ pc();
787
788 __ emit_data64(0x0101010102020304, relocInfo::none);
789 __ emit_data64(0x0000000000000000, relocInfo::none);
790 __ emit_data64(0x0101010102020304, relocInfo::none);
791 __ emit_data64(0x0000000000000000, relocInfo::none);
792 __ emit_data64(0x0101010102020304, relocInfo::none);
793 __ emit_data64(0x0000000000000000, relocInfo::none);
794 __ emit_data64(0x0101010102020304, relocInfo::none);
795 __ emit_data64(0x0000000000000000, relocInfo::none);
796
797 return start;
798 }
799
800 address StubGenerator::generate_popcount_avx_lut(const char *stub_name) {
801 __ align64();
802 StubCodeMark mark(this, "StubRoutines", stub_name);
803 address start = __ pc();
804
805 __ emit_data64(0x0302020102010100, relocInfo::none);
806 __ emit_data64(0x0403030203020201, relocInfo::none);
807 __ emit_data64(0x0302020102010100, relocInfo::none);
808 __ emit_data64(0x0403030203020201, relocInfo::none);
809 __ emit_data64(0x0302020102010100, relocInfo::none);
810 __ emit_data64(0x0403030203020201, relocInfo::none);
811 __ emit_data64(0x0302020102010100, relocInfo::none);
812 __ emit_data64(0x0403030203020201, relocInfo::none);
813
814 return start;
815 }
816
817 address StubGenerator::generate_iota_indices(const char *stub_name) {
818 __ align(CodeEntryAlignment);
819 StubCodeMark mark(this, "StubRoutines", stub_name);
820 address start = __ pc();
821 // B
822 __ emit_data64(0x0706050403020100, relocInfo::none);
823 __ emit_data64(0x0F0E0D0C0B0A0908, relocInfo::none);
824 __ emit_data64(0x1716151413121110, relocInfo::none);
825 __ emit_data64(0x1F1E1D1C1B1A1918, relocInfo::none);
826 __ emit_data64(0x2726252423222120, relocInfo::none);
827 __ emit_data64(0x2F2E2D2C2B2A2928, relocInfo::none);
828 __ emit_data64(0x3736353433323130, relocInfo::none);
829 __ emit_data64(0x3F3E3D3C3B3A3938, relocInfo::none);
830 // W
831 __ emit_data64(0x0003000200010000, relocInfo::none);
832 __ emit_data64(0x0007000600050004, relocInfo::none);
833 __ emit_data64(0x000B000A00090008, relocInfo::none);
834 __ emit_data64(0x000F000E000D000C, relocInfo::none);
835 __ emit_data64(0x0013001200110010, relocInfo::none);
836 __ emit_data64(0x0017001600150014, relocInfo::none);
837 __ emit_data64(0x001B001A00190018, relocInfo::none);
838 __ emit_data64(0x001F001E001D001C, relocInfo::none);
839 // D
840 __ emit_data64(0x0000000100000000, relocInfo::none);
841 __ emit_data64(0x0000000300000002, relocInfo::none);
842 __ emit_data64(0x0000000500000004, relocInfo::none);
843 __ emit_data64(0x0000000700000006, relocInfo::none);
844 __ emit_data64(0x0000000900000008, relocInfo::none);
845 __ emit_data64(0x0000000B0000000A, relocInfo::none);
846 __ emit_data64(0x0000000D0000000C, relocInfo::none);
847 __ emit_data64(0x0000000F0000000E, relocInfo::none);
848 // Q
849 __ emit_data64(0x0000000000000000, relocInfo::none);
850 __ emit_data64(0x0000000000000001, relocInfo::none);
851 __ emit_data64(0x0000000000000002, relocInfo::none);
852 __ emit_data64(0x0000000000000003, relocInfo::none);
853 __ emit_data64(0x0000000000000004, relocInfo::none);
854 __ emit_data64(0x0000000000000005, relocInfo::none);
855 __ emit_data64(0x0000000000000006, relocInfo::none);
856 __ emit_data64(0x0000000000000007, relocInfo::none);
857 // D - FP
858 __ emit_data64(0x3F80000000000000, relocInfo::none); // 0.0f, 1.0f
859 __ emit_data64(0x4040000040000000, relocInfo::none); // 2.0f, 3.0f
860 __ emit_data64(0x40A0000040800000, relocInfo::none); // 4.0f, 5.0f
861 __ emit_data64(0x40E0000040C00000, relocInfo::none); // 6.0f, 7.0f
862 __ emit_data64(0x4110000041000000, relocInfo::none); // 8.0f, 9.0f
863 __ emit_data64(0x4130000041200000, relocInfo::none); // 10.0f, 11.0f
864 __ emit_data64(0x4150000041400000, relocInfo::none); // 12.0f, 13.0f
865 __ emit_data64(0x4170000041600000, relocInfo::none); // 14.0f, 15.0f
866 // Q - FP
867 __ emit_data64(0x0000000000000000, relocInfo::none); // 0.0d
868 __ emit_data64(0x3FF0000000000000, relocInfo::none); // 1.0d
869 __ emit_data64(0x4000000000000000, relocInfo::none); // 2.0d
870 __ emit_data64(0x4008000000000000, relocInfo::none); // 3.0d
871 __ emit_data64(0x4010000000000000, relocInfo::none); // 4.0d
872 __ emit_data64(0x4014000000000000, relocInfo::none); // 5.0d
873 __ emit_data64(0x4018000000000000, relocInfo::none); // 6.0d
874 __ emit_data64(0x401c000000000000, relocInfo::none); // 7.0d
875 return start;
876 }
877
878 address StubGenerator::generate_vector_reverse_bit_lut(const char *stub_name) {
879 __ align(CodeEntryAlignment);
880 StubCodeMark mark(this, "StubRoutines", stub_name);
881 address start = __ pc();
882
883 __ emit_data64(0x0E060A020C040800, relocInfo::none);
884 __ emit_data64(0x0F070B030D050901, relocInfo::none);
885 __ emit_data64(0x0E060A020C040800, relocInfo::none);
886 __ emit_data64(0x0F070B030D050901, relocInfo::none);
887 __ emit_data64(0x0E060A020C040800, relocInfo::none);
888 __ emit_data64(0x0F070B030D050901, relocInfo::none);
889 __ emit_data64(0x0E060A020C040800, relocInfo::none);
890 __ emit_data64(0x0F070B030D050901, relocInfo::none);
891
892 return start;
893 }
894
895 address StubGenerator::generate_vector_reverse_byte_perm_mask_long(const char *stub_name) {
896 __ align(CodeEntryAlignment);
897 StubCodeMark mark(this, "StubRoutines", stub_name);
898 address start = __ pc();
899
900 __ emit_data64(0x0001020304050607, relocInfo::none);
901 __ emit_data64(0x08090A0B0C0D0E0F, relocInfo::none);
902 __ emit_data64(0x0001020304050607, relocInfo::none);
903 __ emit_data64(0x08090A0B0C0D0E0F, relocInfo::none);
904 __ emit_data64(0x0001020304050607, relocInfo::none);
905 __ emit_data64(0x08090A0B0C0D0E0F, relocInfo::none);
906 __ emit_data64(0x0001020304050607, relocInfo::none);
907 __ emit_data64(0x08090A0B0C0D0E0F, relocInfo::none);
908
909 return start;
910 }
911
912 address StubGenerator::generate_vector_reverse_byte_perm_mask_int(const char *stub_name) {
913 __ align(CodeEntryAlignment);
914 StubCodeMark mark(this, "StubRoutines", stub_name);
915 address start = __ pc();
916
917 __ emit_data64(0x0405060700010203, relocInfo::none);
918 __ emit_data64(0x0C0D0E0F08090A0B, relocInfo::none);
919 __ emit_data64(0x0405060700010203, relocInfo::none);
920 __ emit_data64(0x0C0D0E0F08090A0B, relocInfo::none);
921 __ emit_data64(0x0405060700010203, relocInfo::none);
922 __ emit_data64(0x0C0D0E0F08090A0B, relocInfo::none);
923 __ emit_data64(0x0405060700010203, relocInfo::none);
924 __ emit_data64(0x0C0D0E0F08090A0B, relocInfo::none);
925
926 return start;
927 }
928
929 address StubGenerator::generate_vector_reverse_byte_perm_mask_short(const char *stub_name) {
930 __ align(CodeEntryAlignment);
931 StubCodeMark mark(this, "StubRoutines", stub_name);
932 address start = __ pc();
933
934 __ emit_data64(0x0607040502030001, relocInfo::none);
935 __ emit_data64(0x0E0F0C0D0A0B0809, relocInfo::none);
936 __ emit_data64(0x0607040502030001, relocInfo::none);
937 __ emit_data64(0x0E0F0C0D0A0B0809, relocInfo::none);
938 __ emit_data64(0x0607040502030001, relocInfo::none);
939 __ emit_data64(0x0E0F0C0D0A0B0809, relocInfo::none);
940 __ emit_data64(0x0607040502030001, relocInfo::none);
941 __ emit_data64(0x0E0F0C0D0A0B0809, relocInfo::none);
942
943 return start;
944 }
945
946 address StubGenerator::generate_vector_byte_shuffle_mask(const char *stub_name) {
947 __ align(CodeEntryAlignment);
948 StubCodeMark mark(this, "StubRoutines", stub_name);
949 address start = __ pc();
950
951 __ emit_data64(0x7070707070707070, relocInfo::none);
952 __ emit_data64(0x7070707070707070, relocInfo::none);
953 __ emit_data64(0xF0F0F0F0F0F0F0F0, relocInfo::none);
954 __ emit_data64(0xF0F0F0F0F0F0F0F0, relocInfo::none);
955
956 return start;
957 }
958
959 address StubGenerator::generate_fp_mask(const char *stub_name, int64_t mask) {
960 __ align(CodeEntryAlignment);
961 StubCodeMark mark(this, "StubRoutines", stub_name);
962 address start = __ pc();
963
964 __ emit_data64( mask, relocInfo::none );
965 __ emit_data64( mask, relocInfo::none );
966
967 return start;
968 }
969
970 address StubGenerator::generate_vector_mask(const char *stub_name, int64_t mask) {
971 __ align(CodeEntryAlignment);
972 StubCodeMark mark(this, "StubRoutines", stub_name);
973 address start = __ pc();
974
975 __ emit_data64(mask, relocInfo::none);
976 __ emit_data64(mask, relocInfo::none);
977 __ emit_data64(mask, relocInfo::none);
978 __ emit_data64(mask, relocInfo::none);
979 __ emit_data64(mask, relocInfo::none);
980 __ emit_data64(mask, relocInfo::none);
981 __ emit_data64(mask, relocInfo::none);
982 __ emit_data64(mask, relocInfo::none);
983
984 return start;
985 }
986
987 address StubGenerator::generate_vector_byte_perm_mask(const char *stub_name) {
988 __ align(CodeEntryAlignment);
989 StubCodeMark mark(this, "StubRoutines", stub_name);
990 address start = __ pc();
991
992 __ emit_data64(0x0000000000000001, relocInfo::none);
993 __ emit_data64(0x0000000000000003, relocInfo::none);
994 __ emit_data64(0x0000000000000005, relocInfo::none);
995 __ emit_data64(0x0000000000000007, relocInfo::none);
996 __ emit_data64(0x0000000000000000, relocInfo::none);
997 __ emit_data64(0x0000000000000002, relocInfo::none);
998 __ emit_data64(0x0000000000000004, relocInfo::none);
999 __ emit_data64(0x0000000000000006, relocInfo::none);
1000
1001 return start;
1002 }
1003
1004 address StubGenerator::generate_vector_fp_mask(const char *stub_name, int64_t mask) {
1005 __ align(CodeEntryAlignment);
1006 StubCodeMark mark(this, "StubRoutines", stub_name);
1007 address start = __ pc();
1008
1009 __ emit_data64(mask, relocInfo::none);
1010 __ emit_data64(mask, relocInfo::none);
1011 __ emit_data64(mask, relocInfo::none);
1012 __ emit_data64(mask, relocInfo::none);
1013 __ emit_data64(mask, relocInfo::none);
1014 __ emit_data64(mask, relocInfo::none);
1015 __ emit_data64(mask, relocInfo::none);
1016 __ emit_data64(mask, relocInfo::none);
1017
1018 return start;
1019 }
1020
1021 address StubGenerator::generate_vector_custom_i32(const char *stub_name, Assembler::AvxVectorLen len,
1022 int32_t val0, int32_t val1, int32_t val2, int32_t val3,
1023 int32_t val4, int32_t val5, int32_t val6, int32_t val7,
1024 int32_t val8, int32_t val9, int32_t val10, int32_t val11,
1025 int32_t val12, int32_t val13, int32_t val14, int32_t val15) {
1026 __ align(CodeEntryAlignment);
1027 StubCodeMark mark(this, "StubRoutines", stub_name);
1028 address start = __ pc();
1029
1030 assert(len != Assembler::AVX_NoVec, "vector len must be specified");
1031 __ emit_data(val0, relocInfo::none, 0);
1032 __ emit_data(val1, relocInfo::none, 0);
1033 __ emit_data(val2, relocInfo::none, 0);
1034 __ emit_data(val3, relocInfo::none, 0);
1035 if (len >= Assembler::AVX_256bit) {
1036 __ emit_data(val4, relocInfo::none, 0);
1037 __ emit_data(val5, relocInfo::none, 0);
1038 __ emit_data(val6, relocInfo::none, 0);
1039 __ emit_data(val7, relocInfo::none, 0);
1040 if (len >= Assembler::AVX_512bit) {
1041 __ emit_data(val8, relocInfo::none, 0);
1042 __ emit_data(val9, relocInfo::none, 0);
1043 __ emit_data(val10, relocInfo::none, 0);
1044 __ emit_data(val11, relocInfo::none, 0);
1045 __ emit_data(val12, relocInfo::none, 0);
1046 __ emit_data(val13, relocInfo::none, 0);
1047 __ emit_data(val14, relocInfo::none, 0);
1048 __ emit_data(val15, relocInfo::none, 0);
1049 }
1050 }
1051 return start;
1052 }
1053
1054 // Non-destructive plausibility checks for oops
1055 //
1056 // Arguments:
1057 // all args on stack!
1058 //
1059 // Stack after saving c_rarg3:
1060 // [tos + 0]: saved c_rarg3
1061 // [tos + 1]: saved c_rarg2
1062 // [tos + 2]: saved r12 (several TemplateTable methods use it)
1063 // [tos + 3]: saved flags
1064 // [tos + 4]: return address
1065 // * [tos + 5]: error message (char*)
1066 // * [tos + 6]: object to verify (oop)
1067 // * [tos + 7]: saved rax - saved by caller and bashed
1068 // * [tos + 8]: saved r10 (rscratch1) - saved by caller
1069 // * = popped on exit
1070 address StubGenerator::generate_verify_oop() {
1071 StubCodeMark mark(this, "StubRoutines", "verify_oop");
1072 address start = __ pc();
1073
1074 Label exit, error;
1075
1076 __ pushf();
1077 __ incrementl(ExternalAddress((address) StubRoutines::verify_oop_count_addr()), rscratch1);
1078
1079 __ push(r12);
1080
1081 // save c_rarg2 and c_rarg3
1082 __ push(c_rarg2);
1083 __ push(c_rarg3);
1084
1085 enum {
1086 // After previous pushes.
1087 oop_to_verify = 6 * wordSize,
1088 saved_rax = 7 * wordSize,
1089 saved_r10 = 8 * wordSize,
1090
1091 // Before the call to MacroAssembler::debug(), see below.
1092 return_addr = 16 * wordSize,
1093 error_msg = 17 * wordSize
1094 };
1095
1096 // get object
1097 __ movptr(rax, Address(rsp, oop_to_verify));
1098
1099 // make sure object is 'reasonable'
1100 __ testptr(rax, rax);
1101 __ jcc(Assembler::zero, exit); // if obj is null it is OK
1102
1103 BarrierSetAssembler* bs_asm = BarrierSet::barrier_set()->barrier_set_assembler();
1104 bs_asm->check_oop(_masm, rax, c_rarg2, c_rarg3, error);
1105
1106 // return if everything seems ok
1107 __ bind(exit);
1108 __ movptr(rax, Address(rsp, saved_rax)); // get saved rax back
1109 __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back
1110 __ pop(c_rarg3); // restore c_rarg3
1111 __ pop(c_rarg2); // restore c_rarg2
1112 __ pop(r12); // restore r12
1113 __ popf(); // restore flags
1114 __ ret(4 * wordSize); // pop caller saved stuff
1115
1116 // handle errors
1117 __ bind(error);
1118 __ movptr(rax, Address(rsp, saved_rax)); // get saved rax back
1119 __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back
1120 __ pop(c_rarg3); // get saved c_rarg3 back
1121 __ pop(c_rarg2); // get saved c_rarg2 back
1122 __ pop(r12); // get saved r12 back
1123 __ popf(); // get saved flags off stack --
1124 // will be ignored
1125
1126 __ pusha(); // push registers
1127 // (rip is already
1128 // already pushed)
1129 // debug(char* msg, int64_t pc, int64_t regs[])
1130 // We've popped the registers we'd saved (c_rarg3, c_rarg2 and flags), and
1131 // pushed all the registers, so now the stack looks like:
1132 // [tos + 0] 16 saved registers
1133 // [tos + 16] return address
1134 // * [tos + 17] error message (char*)
1135 // * [tos + 18] object to verify (oop)
1136 // * [tos + 19] saved rax - saved by caller and bashed
1137 // * [tos + 20] saved r10 (rscratch1) - saved by caller
1138 // * = popped on exit
1139
1140 __ movptr(c_rarg0, Address(rsp, error_msg)); // pass address of error message
1141 __ movptr(c_rarg1, Address(rsp, return_addr)); // pass return address
1142 __ movq(c_rarg2, rsp); // pass address of regs on stack
1143 __ mov(r12, rsp); // remember rsp
1144 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows
1145 __ andptr(rsp, -16); // align stack as required by ABI
1146 BLOCK_COMMENT("call MacroAssembler::debug");
1147 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64)));
1148 __ hlt();
1149
1150 return start;
1151 }
1152
1153
1154 // Shuffle first three arg regs on Windows into Linux/Solaris locations.
1155 //
1156 // Outputs:
1157 // rdi - rcx
1158 // rsi - rdx
1159 // rdx - r8
1160 // rcx - r9
1161 //
1162 // Registers r9 and r10 are used to save rdi and rsi on Windows, which latter
1163 // are non-volatile. r9 and r10 should not be used by the caller.
1164 //
1165 void StubGenerator::setup_arg_regs(int nargs) {
1166 const Register saved_rdi = r9;
1167 const Register saved_rsi = r10;
1168 assert(nargs == 3 || nargs == 4, "else fix");
1169 #ifdef _WIN64
1170 assert(c_rarg0 == rcx && c_rarg1 == rdx && c_rarg2 == r8 && c_rarg3 == r9,
1171 "unexpected argument registers");
1172 if (nargs == 4) {
1173 __ mov(rax, r9); // r9 is also saved_rdi
1174 }
1175 __ movptr(saved_rdi, rdi);
1176 __ movptr(saved_rsi, rsi);
1177 __ mov(rdi, rcx); // c_rarg0
1178 __ mov(rsi, rdx); // c_rarg1
1179 __ mov(rdx, r8); // c_rarg2
1180 if (nargs == 4) {
1181 __ mov(rcx, rax); // c_rarg3 (via rax)
1182 }
1183 #else
1184 assert(c_rarg0 == rdi && c_rarg1 == rsi && c_rarg2 == rdx && c_rarg3 == rcx,
1185 "unexpected argument registers");
1186 #endif
1187 DEBUG_ONLY(_regs_in_thread = false;)
1188 }
1189
1190
1191 void StubGenerator::restore_arg_regs() {
1192 assert(!_regs_in_thread, "wrong call to restore_arg_regs");
1193 const Register saved_rdi = r9;
1194 const Register saved_rsi = r10;
1195 #ifdef _WIN64
1196 __ movptr(rdi, saved_rdi);
1197 __ movptr(rsi, saved_rsi);
1198 #endif
1199 }
1200
1201
1202 // This is used in places where r10 is a scratch register, and can
1203 // be adapted if r9 is needed also.
1204 void StubGenerator::setup_arg_regs_using_thread(int nargs) {
1205 const Register saved_r15 = r9;
1206 assert(nargs == 3 || nargs == 4, "else fix");
1207 #ifdef _WIN64
1208 if (nargs == 4) {
1209 __ mov(rax, r9); // r9 is also saved_r15
1210 }
1211 __ mov(saved_r15, r15); // r15 is callee saved and needs to be restored
1212 __ get_thread(r15_thread);
1213 assert(c_rarg0 == rcx && c_rarg1 == rdx && c_rarg2 == r8 && c_rarg3 == r9,
1214 "unexpected argument registers");
1215 __ movptr(Address(r15_thread, in_bytes(JavaThread::windows_saved_rdi_offset())), rdi);
1216 __ movptr(Address(r15_thread, in_bytes(JavaThread::windows_saved_rsi_offset())), rsi);
1217
1218 __ mov(rdi, rcx); // c_rarg0
1219 __ mov(rsi, rdx); // c_rarg1
1220 __ mov(rdx, r8); // c_rarg2
1221 if (nargs == 4) {
1222 __ mov(rcx, rax); // c_rarg3 (via rax)
1223 }
1224 #else
1225 assert(c_rarg0 == rdi && c_rarg1 == rsi && c_rarg2 == rdx && c_rarg3 == rcx,
1226 "unexpected argument registers");
1227 #endif
1228 DEBUG_ONLY(_regs_in_thread = true;)
1229 }
1230
1231
1232 void StubGenerator::restore_arg_regs_using_thread() {
1233 assert(_regs_in_thread, "wrong call to restore_arg_regs");
1234 const Register saved_r15 = r9;
1235 #ifdef _WIN64
1236 __ get_thread(r15_thread);
1237 __ movptr(rsi, Address(r15_thread, in_bytes(JavaThread::windows_saved_rsi_offset())));
1238 __ movptr(rdi, Address(r15_thread, in_bytes(JavaThread::windows_saved_rdi_offset())));
1239 __ mov(r15, saved_r15); // r15 is callee saved and needs to be restored
1240 #endif
1241 }
1242
1243
1244 void StubGenerator::setup_argument_regs(BasicType type) {
1245 if (type == T_BYTE || type == T_SHORT) {
1246 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1247 // r9 and r10 may be used to save non-volatile registers
1248 } else {
1249 setup_arg_regs_using_thread(); // from => rdi, to => rsi, count => rdx
1250 // r9 is used to save r15_thread
1251 }
1252 }
1253
1254
1255 void StubGenerator::restore_argument_regs(BasicType type) {
1256 if (type == T_BYTE || type == T_SHORT) {
1257 restore_arg_regs();
1258 } else {
1259 restore_arg_regs_using_thread();
1260 }
1261 }
1262
1263 address StubGenerator::generate_data_cache_writeback() {
1264 const Register src = c_rarg0; // source address
1265
1266 __ align(CodeEntryAlignment);
1267
1268 StubCodeMark mark(this, "StubRoutines", "_data_cache_writeback");
1269
1270 address start = __ pc();
1271
1272 __ enter();
1273 __ cache_wb(Address(src, 0));
1274 __ leave();
1275 __ ret(0);
1276
1277 return start;
1278 }
1279
1280 address StubGenerator::generate_data_cache_writeback_sync() {
1281 const Register is_pre = c_rarg0; // pre or post sync
1282
1283 __ align(CodeEntryAlignment);
1284
1285 StubCodeMark mark(this, "StubRoutines", "_data_cache_writeback_sync");
1286
1287 // pre wbsync is a no-op
1288 // post wbsync translates to an sfence
1289
1290 Label skip;
1291 address start = __ pc();
1292
1293 __ enter();
1294 __ cmpl(is_pre, 0);
1295 __ jcc(Assembler::notEqual, skip);
1296 __ cache_wbsync(false);
1297 __ bind(skip);
1298 __ leave();
1299 __ ret(0);
1300
1301 return start;
1302 }
1303
1304 // ofs and limit are use for multi-block byte array.
1305 // int com.sun.security.provider.MD5.implCompress(byte[] b, int ofs)
1306 address StubGenerator::generate_md5_implCompress(bool multi_block, const char *name) {
1307 __ align(CodeEntryAlignment);
1308 StubCodeMark mark(this, "StubRoutines", name);
1309 address start = __ pc();
1310
1311 const Register buf_param = r15;
1312 const Address state_param(rsp, 0 * wordSize);
1313 const Address ofs_param (rsp, 1 * wordSize );
1314 const Address limit_param(rsp, 1 * wordSize + 4);
1315
1316 __ enter();
1317 __ push(rbx);
1318 __ push(rdi);
1319 __ push(rsi);
1320 __ push(r15);
1321 __ subptr(rsp, 2 * wordSize);
1322
1323 __ movptr(buf_param, c_rarg0);
1324 __ movptr(state_param, c_rarg1);
1325 if (multi_block) {
1326 __ movl(ofs_param, c_rarg2);
1327 __ movl(limit_param, c_rarg3);
1328 }
1329 __ fast_md5(buf_param, state_param, ofs_param, limit_param, multi_block);
1330
1331 __ addptr(rsp, 2 * wordSize);
1332 __ pop(r15);
1333 __ pop(rsi);
1334 __ pop(rdi);
1335 __ pop(rbx);
1336 __ leave();
1337 __ ret(0);
1338
1339 return start;
1340 }
1341
1342 address StubGenerator::generate_upper_word_mask() {
1343 __ align64();
1344 StubCodeMark mark(this, "StubRoutines", "upper_word_mask");
1345 address start = __ pc();
1346
1347 __ emit_data64(0x0000000000000000, relocInfo::none);
1348 __ emit_data64(0xFFFFFFFF00000000, relocInfo::none);
1349
1350 return start;
1351 }
1352
1353 address StubGenerator::generate_shuffle_byte_flip_mask() {
1354 __ align64();
1355 StubCodeMark mark(this, "StubRoutines", "shuffle_byte_flip_mask");
1356 address start = __ pc();
1357
1358 __ emit_data64(0x08090a0b0c0d0e0f, relocInfo::none);
1359 __ emit_data64(0x0001020304050607, relocInfo::none);
1360
1361 return start;
1362 }
1363
1364 // ofs and limit are use for multi-block byte array.
1365 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
1366 address StubGenerator::generate_sha1_implCompress(bool multi_block, const char *name) {
1367 __ align(CodeEntryAlignment);
1368 StubCodeMark mark(this, "StubRoutines", name);
1369 address start = __ pc();
1370
1371 Register buf = c_rarg0;
1372 Register state = c_rarg1;
1373 Register ofs = c_rarg2;
1374 Register limit = c_rarg3;
1375
1376 const XMMRegister abcd = xmm0;
1377 const XMMRegister e0 = xmm1;
1378 const XMMRegister e1 = xmm2;
1379 const XMMRegister msg0 = xmm3;
1380
1381 const XMMRegister msg1 = xmm4;
1382 const XMMRegister msg2 = xmm5;
1383 const XMMRegister msg3 = xmm6;
1384 const XMMRegister shuf_mask = xmm7;
1385
1386 __ enter();
1387
1388 __ subptr(rsp, 4 * wordSize);
1389
1390 __ fast_sha1(abcd, e0, e1, msg0, msg1, msg2, msg3, shuf_mask,
1391 buf, state, ofs, limit, rsp, multi_block);
1392
1393 __ addptr(rsp, 4 * wordSize);
1394
1395 __ leave();
1396 __ ret(0);
1397
1398 return start;
1399 }
1400
1401 address StubGenerator::generate_pshuffle_byte_flip_mask() {
1402 __ align64();
1403 StubCodeMark mark(this, "StubRoutines", "pshuffle_byte_flip_mask");
1404 address start = __ pc();
1405
1406 __ emit_data64(0x0405060700010203, relocInfo::none);
1407 __ emit_data64(0x0c0d0e0f08090a0b, relocInfo::none);
1408
1409 if (VM_Version::supports_avx2()) {
1410 __ emit_data64(0x0405060700010203, relocInfo::none); // second copy
1411 __ emit_data64(0x0c0d0e0f08090a0b, relocInfo::none);
1412 // _SHUF_00BA
1413 __ emit_data64(0x0b0a090803020100, relocInfo::none);
1414 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1415 __ emit_data64(0x0b0a090803020100, relocInfo::none);
1416 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1417 // _SHUF_DC00
1418 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1419 __ emit_data64(0x0b0a090803020100, relocInfo::none);
1420 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1421 __ emit_data64(0x0b0a090803020100, relocInfo::none);
1422 }
1423
1424 return start;
1425 }
1426
1427 //Mask for byte-swapping a couple of qwords in an XMM register using (v)pshufb.
1428 address StubGenerator::generate_pshuffle_byte_flip_mask_sha512() {
1429 __ align32();
1430 StubCodeMark mark(this, "StubRoutines", "pshuffle_byte_flip_mask_sha512");
1431 address start = __ pc();
1432
1433 if (VM_Version::supports_avx2()) {
1434 __ emit_data64(0x0001020304050607, relocInfo::none); // PSHUFFLE_BYTE_FLIP_MASK
1435 __ emit_data64(0x08090a0b0c0d0e0f, relocInfo::none);
1436 __ emit_data64(0x1011121314151617, relocInfo::none);
1437 __ emit_data64(0x18191a1b1c1d1e1f, relocInfo::none);
1438 __ emit_data64(0x0000000000000000, relocInfo::none); //MASK_YMM_LO
1439 __ emit_data64(0x0000000000000000, relocInfo::none);
1440 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1441 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1442 }
1443
1444 return start;
1445 }
1446
1447 // ofs and limit are use for multi-block byte array.
1448 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
1449 address StubGenerator::generate_sha256_implCompress(bool multi_block, const char *name) {
1450 assert(VM_Version::supports_sha() || VM_Version::supports_avx2(), "");
1451 __ align(CodeEntryAlignment);
1452 StubCodeMark mark(this, "StubRoutines", name);
1453 address start = __ pc();
1454
1455 Register buf = c_rarg0;
1456 Register state = c_rarg1;
1457 Register ofs = c_rarg2;
1458 Register limit = c_rarg3;
1459
1460 const XMMRegister msg = xmm0;
1461 const XMMRegister state0 = xmm1;
1462 const XMMRegister state1 = xmm2;
1463 const XMMRegister msgtmp0 = xmm3;
1464
1465 const XMMRegister msgtmp1 = xmm4;
1466 const XMMRegister msgtmp2 = xmm5;
1467 const XMMRegister msgtmp3 = xmm6;
1468 const XMMRegister msgtmp4 = xmm7;
1469
1470 const XMMRegister shuf_mask = xmm8;
1471
1472 __ enter();
1473
1474 __ subptr(rsp, 4 * wordSize);
1475
1476 if (VM_Version::supports_sha()) {
1477 __ fast_sha256(msg, state0, state1, msgtmp0, msgtmp1, msgtmp2, msgtmp3, msgtmp4,
1478 buf, state, ofs, limit, rsp, multi_block, shuf_mask);
1479 } else if (VM_Version::supports_avx2()) {
1480 __ sha256_AVX2(msg, state0, state1, msgtmp0, msgtmp1, msgtmp2, msgtmp3, msgtmp4,
1481 buf, state, ofs, limit, rsp, multi_block, shuf_mask);
1482 }
1483 __ addptr(rsp, 4 * wordSize);
1484 __ vzeroupper();
1485 __ leave();
1486 __ ret(0);
1487
1488 return start;
1489 }
1490
1491 address StubGenerator::generate_sha512_implCompress(bool multi_block, const char *name) {
1492 assert(VM_Version::supports_avx2(), "");
1493 assert(VM_Version::supports_bmi2(), "");
1494 __ align(CodeEntryAlignment);
1495 StubCodeMark mark(this, "StubRoutines", name);
1496 address start = __ pc();
1497
1498 Register buf = c_rarg0;
1499 Register state = c_rarg1;
1500 Register ofs = c_rarg2;
1501 Register limit = c_rarg3;
1502
1503 const XMMRegister msg = xmm0;
1504 const XMMRegister state0 = xmm1;
1505 const XMMRegister state1 = xmm2;
1506 const XMMRegister msgtmp0 = xmm3;
1507 const XMMRegister msgtmp1 = xmm4;
1508 const XMMRegister msgtmp2 = xmm5;
1509 const XMMRegister msgtmp3 = xmm6;
1510 const XMMRegister msgtmp4 = xmm7;
1511
1512 const XMMRegister shuf_mask = xmm8;
1513
1514 __ enter();
1515
1516 __ sha512_AVX2(msg, state0, state1, msgtmp0, msgtmp1, msgtmp2, msgtmp3, msgtmp4,
1517 buf, state, ofs, limit, rsp, multi_block, shuf_mask);
1518
1519 __ vzeroupper();
1520 __ leave();
1521 __ ret(0);
1522
1523 return start;
1524 }
1525
1526 address StubGenerator::base64_shuffle_addr() {
1527 __ align64();
1528 StubCodeMark mark(this, "StubRoutines", "shuffle_base64");
1529 address start = __ pc();
1530
1531 assert(((unsigned long long)start & 0x3f) == 0,
1532 "Alignment problem (0x%08llx)", (unsigned long long)start);
1533 __ emit_data64(0x0405030401020001, relocInfo::none);
1534 __ emit_data64(0x0a0b090a07080607, relocInfo::none);
1535 __ emit_data64(0x10110f100d0e0c0d, relocInfo::none);
1536 __ emit_data64(0x1617151613141213, relocInfo::none);
1537 __ emit_data64(0x1c1d1b1c191a1819, relocInfo::none);
1538 __ emit_data64(0x222321221f201e1f, relocInfo::none);
1539 __ emit_data64(0x2829272825262425, relocInfo::none);
1540 __ emit_data64(0x2e2f2d2e2b2c2a2b, relocInfo::none);
1541
1542 return start;
1543 }
1544
1545 address StubGenerator::base64_avx2_shuffle_addr() {
1546 __ align32();
1547 StubCodeMark mark(this, "StubRoutines", "avx2_shuffle_base64");
1548 address start = __ pc();
1549
1550 __ emit_data64(0x0809070805060405, relocInfo::none);
1551 __ emit_data64(0x0e0f0d0e0b0c0a0b, relocInfo::none);
1552 __ emit_data64(0x0405030401020001, relocInfo::none);
1553 __ emit_data64(0x0a0b090a07080607, relocInfo::none);
1554
1555 return start;
1556 }
1557
1558 address StubGenerator::base64_avx2_input_mask_addr() {
1559 __ align32();
1560 StubCodeMark mark(this, "StubRoutines", "avx2_input_mask_base64");
1561 address start = __ pc();
1562
1563 __ emit_data64(0x8000000000000000, relocInfo::none);
1564 __ emit_data64(0x8000000080000000, relocInfo::none);
1565 __ emit_data64(0x8000000080000000, relocInfo::none);
1566 __ emit_data64(0x8000000080000000, relocInfo::none);
1567
1568 return start;
1569 }
1570
1571 address StubGenerator::base64_avx2_lut_addr() {
1572 __ align32();
1573 StubCodeMark mark(this, "StubRoutines", "avx2_lut_base64");
1574 address start = __ pc();
1575
1576 __ emit_data64(0xfcfcfcfcfcfc4741, relocInfo::none);
1577 __ emit_data64(0x0000f0edfcfcfcfc, relocInfo::none);
1578 __ emit_data64(0xfcfcfcfcfcfc4741, relocInfo::none);
1579 __ emit_data64(0x0000f0edfcfcfcfc, relocInfo::none);
1580
1581 // URL LUT
1582 __ emit_data64(0xfcfcfcfcfcfc4741, relocInfo::none);
1583 __ emit_data64(0x000020effcfcfcfc, relocInfo::none);
1584 __ emit_data64(0xfcfcfcfcfcfc4741, relocInfo::none);
1585 __ emit_data64(0x000020effcfcfcfc, relocInfo::none);
1586
1587 return start;
1588 }
1589
1590 address StubGenerator::base64_encoding_table_addr() {
1591 __ align64();
1592 StubCodeMark mark(this, "StubRoutines", "encoding_table_base64");
1593 address start = __ pc();
1594
1595 assert(((unsigned long long)start & 0x3f) == 0, "Alignment problem (0x%08llx)", (unsigned long long)start);
1596 __ emit_data64(0x4847464544434241, relocInfo::none);
1597 __ emit_data64(0x504f4e4d4c4b4a49, relocInfo::none);
1598 __ emit_data64(0x5857565554535251, relocInfo::none);
1599 __ emit_data64(0x6665646362615a59, relocInfo::none);
1600 __ emit_data64(0x6e6d6c6b6a696867, relocInfo::none);
1601 __ emit_data64(0x767574737271706f, relocInfo::none);
1602 __ emit_data64(0x333231307a797877, relocInfo::none);
1603 __ emit_data64(0x2f2b393837363534, relocInfo::none);
1604
1605 // URL table
1606 __ emit_data64(0x4847464544434241, relocInfo::none);
1607 __ emit_data64(0x504f4e4d4c4b4a49, relocInfo::none);
1608 __ emit_data64(0x5857565554535251, relocInfo::none);
1609 __ emit_data64(0x6665646362615a59, relocInfo::none);
1610 __ emit_data64(0x6e6d6c6b6a696867, relocInfo::none);
1611 __ emit_data64(0x767574737271706f, relocInfo::none);
1612 __ emit_data64(0x333231307a797877, relocInfo::none);
1613 __ emit_data64(0x5f2d393837363534, relocInfo::none);
1614
1615 return start;
1616 }
1617
1618 // Code for generating Base64 encoding.
1619 // Intrinsic function prototype in Base64.java:
1620 // private void encodeBlock(byte[] src, int sp, int sl, byte[] dst, int dp,
1621 // boolean isURL) {
1622 address StubGenerator::generate_base64_encodeBlock()
1623 {
1624 __ align(CodeEntryAlignment);
1625 StubCodeMark mark(this, "StubRoutines", "implEncode");
1626 address start = __ pc();
1627
1628 __ enter();
1629
1630 // Save callee-saved registers before using them
1631 __ push(r12);
1632 __ push(r13);
1633 __ push(r14);
1634 __ push(r15);
1635
1636 // arguments
1637 const Register source = c_rarg0; // Source Array
1638 const Register start_offset = c_rarg1; // start offset
1639 const Register end_offset = c_rarg2; // end offset
1640 const Register dest = c_rarg3; // destination array
1641
1642 #ifndef _WIN64
1643 const Register dp = c_rarg4; // Position for writing to dest array
1644 const Register isURL = c_rarg5; // Base64 or URL character set
1645 #else
1646 const Address dp_mem(rbp, 6 * wordSize); // length is on stack on Win64
1647 const Address isURL_mem(rbp, 7 * wordSize);
1648 const Register isURL = r10; // pick the volatile windows register
1649 const Register dp = r12;
1650 __ movl(dp, dp_mem);
1651 __ movl(isURL, isURL_mem);
1652 #endif
1653
1654 const Register length = r14;
1655 const Register encode_table = r13;
1656 Label L_process3, L_exit, L_processdata, L_vbmiLoop, L_not512, L_32byteLoop;
1657
1658 // calculate length from offsets
1659 __ movl(length, end_offset);
1660 __ subl(length, start_offset);
1661 __ jcc(Assembler::lessEqual, L_exit);
1662
1663 // Code for 512-bit VBMI encoding. Encodes 48 input bytes into 64
1664 // output bytes. We read 64 input bytes and ignore the last 16, so be
1665 // sure not to read past the end of the input buffer.
1666 if (VM_Version::supports_avx512_vbmi()) {
1667 __ cmpl(length, 64); // Do not overrun input buffer.
1668 __ jcc(Assembler::below, L_not512);
1669
1670 __ shll(isURL, 6); // index into decode table based on isURL
1671 __ lea(encode_table, ExternalAddress(StubRoutines::x86::base64_encoding_table_addr()));
1672 __ addptr(encode_table, isURL);
1673 __ shrl(isURL, 6); // restore isURL
1674
1675 __ mov64(rax, 0x3036242a1016040aull); // Shifts
1676 __ evmovdquq(xmm3, ExternalAddress(StubRoutines::x86::base64_shuffle_addr()), Assembler::AVX_512bit, r15);
1677 __ evmovdquq(xmm2, Address(encode_table, 0), Assembler::AVX_512bit);
1678 __ evpbroadcastq(xmm1, rax, Assembler::AVX_512bit);
1679
1680 __ align32();
1681 __ BIND(L_vbmiLoop);
1682
1683 __ vpermb(xmm0, xmm3, Address(source, start_offset), Assembler::AVX_512bit);
1684 __ subl(length, 48);
1685
1686 // Put the input bytes into the proper lanes for writing, then
1687 // encode them.
1688 __ evpmultishiftqb(xmm0, xmm1, xmm0, Assembler::AVX_512bit);
1689 __ vpermb(xmm0, xmm0, xmm2, Assembler::AVX_512bit);
1690
1691 // Write to destination
1692 __ evmovdquq(Address(dest, dp), xmm0, Assembler::AVX_512bit);
1693
1694 __ addptr(dest, 64);
1695 __ addptr(source, 48);
1696 __ cmpl(length, 64);
1697 __ jcc(Assembler::aboveEqual, L_vbmiLoop);
1698
1699 __ vzeroupper();
1700 }
1701
1702 __ BIND(L_not512);
1703 if (VM_Version::supports_avx2()) {
1704 /*
1705 ** This AVX2 encoder is based off the paper at:
1706 ** https://dl.acm.org/doi/10.1145/3132709
1707 **
1708 ** We use AVX2 SIMD instructions to encode 24 bytes into 32
1709 ** output bytes.
1710 **
1711 */
1712 // Lengths under 32 bytes are done with scalar routine
1713 __ cmpl(length, 31);
1714 __ jcc(Assembler::belowEqual, L_process3);
1715
1716 // Set up supporting constant table data
1717 __ vmovdqu(xmm9, ExternalAddress(StubRoutines::x86::base64_avx2_shuffle_addr()), rax);
1718 // 6-bit mask for 2nd and 4th (and multiples) 6-bit values
1719 __ movl(rax, 0x0fc0fc00);
1720 __ movdl(xmm8, rax);
1721 __ vmovdqu(xmm1, ExternalAddress(StubRoutines::x86::base64_avx2_input_mask_addr()), rax);
1722 __ vpbroadcastd(xmm8, xmm8, Assembler::AVX_256bit);
1723
1724 // Multiplication constant for "shifting" right by 6 and 10
1725 // bits
1726 __ movl(rax, 0x04000040);
1727
1728 __ subl(length, 24);
1729 __ movdl(xmm7, rax);
1730 __ vpbroadcastd(xmm7, xmm7, Assembler::AVX_256bit);
1731
1732 // For the first load, we mask off reading of the first 4
1733 // bytes into the register. This is so we can get 4 3-byte
1734 // chunks into each lane of the register, avoiding having to
1735 // handle end conditions. We then shuffle these bytes into a
1736 // specific order so that manipulation is easier.
1737 //
1738 // The initial read loads the XMM register like this:
1739 //
1740 // Lower 128-bit lane:
1741 // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
1742 // | XX | XX | XX | XX | A0 | A1 | A2 | B0 | B1 | B2 | C0 | C1
1743 // | C2 | D0 | D1 | D2 |
1744 // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
1745 //
1746 // Upper 128-bit lane:
1747 // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
1748 // | E0 | E1 | E2 | F0 | F1 | F2 | G0 | G1 | G2 | H0 | H1 | H2
1749 // | XX | XX | XX | XX |
1750 // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
1751 //
1752 // Where A0 is the first input byte, B0 is the fourth, etc.
1753 // The alphabetical significance denotes the 3 bytes to be
1754 // consumed and encoded into 4 bytes.
1755 //
1756 // We then shuffle the register so each 32-bit word contains
1757 // the sequence:
1758 // A1 A0 A2 A1, B1, B0, B2, B1, etc.
1759 // Each of these byte sequences are then manipulated into 4
1760 // 6-bit values ready for encoding.
1761 //
1762 // If we focus on one set of 3-byte chunks, changing the
1763 // nomenclature such that A0 => a, A1 => b, and A2 => c, we
1764 // shuffle such that each 24-bit chunk contains:
1765 //
1766 // b7 b6 b5 b4 b3 b2 b1 b0 | a7 a6 a5 a4 a3 a2 a1 a0 | c7 c6
1767 // c5 c4 c3 c2 c1 c0 | b7 b6 b5 b4 b3 b2 b1 b0
1768 // Explain this step.
1769 // b3 b2 b1 b0 c5 c4 c3 c2 | c1 c0 d5 d4 d3 d2 d1 d0 | a5 a4
1770 // a3 a2 a1 a0 b5 b4 | b3 b2 b1 b0 c5 c4 c3 c2
1771 //
1772 // W first and off all but bits 4-9 and 16-21 (c5..c0 and
1773 // a5..a0) and shift them using a vector multiplication
1774 // operation (vpmulhuw) which effectively shifts c right by 6
1775 // bits and a right by 10 bits. We similarly mask bits 10-15
1776 // (d5..d0) and 22-27 (b5..b0) and shift them left by 8 and 4
1777 // bits respectively. This is done using vpmullw. We end up
1778 // with 4 6-bit values, thus splitting the 3 input bytes,
1779 // ready for encoding:
1780 // 0 0 d5..d0 0 0 c5..c0 0 0 b5..b0 0 0 a5..a0
1781 //
1782 // For translation, we recognize that there are 5 distinct
1783 // ranges of legal Base64 characters as below:
1784 //
1785 // +-------------+-------------+------------+
1786 // | 6-bit value | ASCII range | offset |
1787 // +-------------+-------------+------------+
1788 // | 0..25 | A..Z | 65 |
1789 // | 26..51 | a..z | 71 |
1790 // | 52..61 | 0..9 | -4 |
1791 // | 62 | + or - | -19 or -17 |
1792 // | 63 | / or _ | -16 or 32 |
1793 // +-------------+-------------+------------+
1794 //
1795 // We note that vpshufb does a parallel lookup in a
1796 // destination register using the lower 4 bits of bytes from a
1797 // source register. If we use a saturated subtraction and
1798 // subtract 51 from each 6-bit value, bytes from [0,51]
1799 // saturate to 0, and [52,63] map to a range of [1,12]. We
1800 // distinguish the [0,25] and [26,51] ranges by assigning a
1801 // value of 13 for all 6-bit values less than 26. We end up
1802 // with:
1803 //
1804 // +-------------+-------------+------------+
1805 // | 6-bit value | Reduced | offset |
1806 // +-------------+-------------+------------+
1807 // | 0..25 | 13 | 65 |
1808 // | 26..51 | 0 | 71 |
1809 // | 52..61 | 0..9 | -4 |
1810 // | 62 | 11 | -19 or -17 |
1811 // | 63 | 12 | -16 or 32 |
1812 // +-------------+-------------+------------+
1813 //
1814 // We then use a final vpshufb to add the appropriate offset,
1815 // translating the bytes.
1816 //
1817 // Load input bytes - only 28 bytes. Mask the first load to
1818 // not load into the full register.
1819 __ vpmaskmovd(xmm1, xmm1, Address(source, start_offset, Address::times_1, -4), Assembler::AVX_256bit);
1820
1821 // Move 3-byte chunks of input (12 bytes) into 16 bytes,
1822 // ordering by:
1823 // 1, 0, 2, 1; 4, 3, 5, 4; etc. This groups 6-bit chunks
1824 // for easy masking
1825 __ vpshufb(xmm1, xmm1, xmm9, Assembler::AVX_256bit);
1826
1827 __ addl(start_offset, 24);
1828
1829 // Load masking register for first and third (and multiples)
1830 // 6-bit values.
1831 __ movl(rax, 0x003f03f0);
1832 __ movdl(xmm6, rax);
1833 __ vpbroadcastd(xmm6, xmm6, Assembler::AVX_256bit);
1834 // Multiplication constant for "shifting" left by 4 and 8 bits
1835 __ movl(rax, 0x01000010);
1836 __ movdl(xmm5, rax);
1837 __ vpbroadcastd(xmm5, xmm5, Assembler::AVX_256bit);
1838
1839 // Isolate 6-bit chunks of interest
1840 __ vpand(xmm0, xmm8, xmm1, Assembler::AVX_256bit);
1841
1842 // Load constants for encoding
1843 __ movl(rax, 0x19191919);
1844 __ movdl(xmm3, rax);
1845 __ vpbroadcastd(xmm3, xmm3, Assembler::AVX_256bit);
1846 __ movl(rax, 0x33333333);
1847 __ movdl(xmm4, rax);
1848 __ vpbroadcastd(xmm4, xmm4, Assembler::AVX_256bit);
1849
1850 // Shift output bytes 0 and 2 into proper lanes
1851 __ vpmulhuw(xmm2, xmm0, xmm7, Assembler::AVX_256bit);
1852
1853 // Mask and shift output bytes 1 and 3 into proper lanes and
1854 // combine
1855 __ vpand(xmm0, xmm6, xmm1, Assembler::AVX_256bit);
1856 __ vpmullw(xmm0, xmm5, xmm0, Assembler::AVX_256bit);
1857 __ vpor(xmm0, xmm0, xmm2, Assembler::AVX_256bit);
1858
1859 // Find out which are 0..25. This indicates which input
1860 // values fall in the range of 'A'-'Z', which require an
1861 // additional offset (see comments above)
1862 __ vpcmpgtb(xmm2, xmm0, xmm3, Assembler::AVX_256bit);
1863 __ vpsubusb(xmm1, xmm0, xmm4, Assembler::AVX_256bit);
1864 __ vpsubb(xmm1, xmm1, xmm2, Assembler::AVX_256bit);
1865
1866 // Load the proper lookup table
1867 __ lea(r11, ExternalAddress(StubRoutines::x86::base64_avx2_lut_addr()));
1868 __ movl(r15, isURL);
1869 __ shll(r15, 5);
1870 __ vmovdqu(xmm2, Address(r11, r15));
1871
1872 // Shuffle the offsets based on the range calculation done
1873 // above. This allows us to add the correct offset to the
1874 // 6-bit value corresponding to the range documented above.
1875 __ vpshufb(xmm1, xmm2, xmm1, Assembler::AVX_256bit);
1876 __ vpaddb(xmm0, xmm1, xmm0, Assembler::AVX_256bit);
1877
1878 // Store the encoded bytes
1879 __ vmovdqu(Address(dest, dp), xmm0);
1880 __ addl(dp, 32);
1881
1882 __ cmpl(length, 31);
1883 __ jcc(Assembler::belowEqual, L_process3);
1884
1885 __ align32();
1886 __ BIND(L_32byteLoop);
1887
1888 // Get next 32 bytes
1889 __ vmovdqu(xmm1, Address(source, start_offset, Address::times_1, -4));
1890
1891 __ subl(length, 24);
1892 __ addl(start_offset, 24);
1893
1894 // This logic is identical to the above, with only constant
1895 // register loads removed. Shuffle the input, mask off 6-bit
1896 // chunks, shift them into place, then add the offset to
1897 // encode.
1898 __ vpshufb(xmm1, xmm1, xmm9, Assembler::AVX_256bit);
1899
1900 __ vpand(xmm0, xmm8, xmm1, Assembler::AVX_256bit);
1901 __ vpmulhuw(xmm10, xmm0, xmm7, Assembler::AVX_256bit);
1902 __ vpand(xmm0, xmm6, xmm1, Assembler::AVX_256bit);
1903 __ vpmullw(xmm0, xmm5, xmm0, Assembler::AVX_256bit);
1904 __ vpor(xmm0, xmm0, xmm10, Assembler::AVX_256bit);
1905 __ vpcmpgtb(xmm10, xmm0, xmm3, Assembler::AVX_256bit);
1906 __ vpsubusb(xmm1, xmm0, xmm4, Assembler::AVX_256bit);
1907 __ vpsubb(xmm1, xmm1, xmm10, Assembler::AVX_256bit);
1908 __ vpshufb(xmm1, xmm2, xmm1, Assembler::AVX_256bit);
1909 __ vpaddb(xmm0, xmm1, xmm0, Assembler::AVX_256bit);
1910
1911 // Store the encoded bytes
1912 __ vmovdqu(Address(dest, dp), xmm0);
1913 __ addl(dp, 32);
1914
1915 __ cmpl(length, 31);
1916 __ jcc(Assembler::above, L_32byteLoop);
1917
1918 __ BIND(L_process3);
1919 __ vzeroupper();
1920 } else {
1921 __ BIND(L_process3);
1922 }
1923
1924 __ cmpl(length, 3);
1925 __ jcc(Assembler::below, L_exit);
1926
1927 // Load the encoding table based on isURL
1928 __ lea(r11, ExternalAddress(StubRoutines::x86::base64_encoding_table_addr()));
1929 __ movl(r15, isURL);
1930 __ shll(r15, 6);
1931 __ addptr(r11, r15);
1932
1933 __ BIND(L_processdata);
1934
1935 // Load 3 bytes
1936 __ load_unsigned_byte(r15, Address(source, start_offset));
1937 __ load_unsigned_byte(r10, Address(source, start_offset, Address::times_1, 1));
1938 __ load_unsigned_byte(r13, Address(source, start_offset, Address::times_1, 2));
1939
1940 // Build a 32-bit word with bytes 1, 2, 0, 1
1941 __ movl(rax, r10);
1942 __ shll(r10, 24);
1943 __ orl(rax, r10);
1944
1945 __ subl(length, 3);
1946
1947 __ shll(r15, 8);
1948 __ shll(r13, 16);
1949 __ orl(rax, r15);
1950
1951 __ addl(start_offset, 3);
1952
1953 __ orl(rax, r13);
1954 // At this point, rax contains | byte1 | byte2 | byte0 | byte1
1955 // r13 has byte2 << 16 - need low-order 6 bits to translate.
1956 // This translated byte is the fourth output byte.
1957 __ shrl(r13, 16);
1958 __ andl(r13, 0x3f);
1959
1960 // The high-order 6 bits of r15 (byte0) is translated.
1961 // The translated byte is the first output byte.
1962 __ shrl(r15, 10);
1963
1964 __ load_unsigned_byte(r13, Address(r11, r13));
1965 __ load_unsigned_byte(r15, Address(r11, r15));
1966
1967 __ movb(Address(dest, dp, Address::times_1, 3), r13);
1968
1969 // Extract high-order 4 bits of byte1 and low-order 2 bits of byte0.
1970 // This translated byte is the second output byte.
1971 __ shrl(rax, 4);
1972 __ movl(r10, rax);
1973 __ andl(rax, 0x3f);
1974
1975 __ movb(Address(dest, dp, Address::times_1, 0), r15);
1976
1977 __ load_unsigned_byte(rax, Address(r11, rax));
1978
1979 // Extract low-order 2 bits of byte1 and high-order 4 bits of byte2.
1980 // This translated byte is the third output byte.
1981 __ shrl(r10, 18);
1982 __ andl(r10, 0x3f);
1983
1984 __ load_unsigned_byte(r10, Address(r11, r10));
1985
1986 __ movb(Address(dest, dp, Address::times_1, 1), rax);
1987 __ movb(Address(dest, dp, Address::times_1, 2), r10);
1988
1989 __ addl(dp, 4);
1990 __ cmpl(length, 3);
1991 __ jcc(Assembler::aboveEqual, L_processdata);
1992
1993 __ BIND(L_exit);
1994 __ pop(r15);
1995 __ pop(r14);
1996 __ pop(r13);
1997 __ pop(r12);
1998 __ leave();
1999 __ ret(0);
2000
2001 return start;
2002 }
2003
2004 // base64 AVX512vbmi tables
2005 address StubGenerator::base64_vbmi_lookup_lo_addr() {
2006 __ align64();
2007 StubCodeMark mark(this, "StubRoutines", "lookup_lo_base64");
2008 address start = __ pc();
2009
2010 assert(((unsigned long long)start & 0x3f) == 0,
2011 "Alignment problem (0x%08llx)", (unsigned long long)start);
2012 __ emit_data64(0x8080808080808080, relocInfo::none);
2013 __ emit_data64(0x8080808080808080, relocInfo::none);
2014 __ emit_data64(0x8080808080808080, relocInfo::none);
2015 __ emit_data64(0x8080808080808080, relocInfo::none);
2016 __ emit_data64(0x8080808080808080, relocInfo::none);
2017 __ emit_data64(0x3f8080803e808080, relocInfo::none);
2018 __ emit_data64(0x3b3a393837363534, relocInfo::none);
2019 __ emit_data64(0x8080808080803d3c, relocInfo::none);
2020
2021 return start;
2022 }
2023
2024 address StubGenerator::base64_vbmi_lookup_hi_addr() {
2025 __ align64();
2026 StubCodeMark mark(this, "StubRoutines", "lookup_hi_base64");
2027 address start = __ pc();
2028
2029 assert(((unsigned long long)start & 0x3f) == 0,
2030 "Alignment problem (0x%08llx)", (unsigned long long)start);
2031 __ emit_data64(0x0605040302010080, relocInfo::none);
2032 __ emit_data64(0x0e0d0c0b0a090807, relocInfo::none);
2033 __ emit_data64(0x161514131211100f, relocInfo::none);
2034 __ emit_data64(0x8080808080191817, relocInfo::none);
2035 __ emit_data64(0x201f1e1d1c1b1a80, relocInfo::none);
2036 __ emit_data64(0x2827262524232221, relocInfo::none);
2037 __ emit_data64(0x302f2e2d2c2b2a29, relocInfo::none);
2038 __ emit_data64(0x8080808080333231, relocInfo::none);
2039
2040 return start;
2041 }
2042 address StubGenerator::base64_vbmi_lookup_lo_url_addr() {
2043 __ align64();
2044 StubCodeMark mark(this, "StubRoutines", "lookup_lo_base64url");
2045 address start = __ pc();
2046
2047 assert(((unsigned long long)start & 0x3f) == 0,
2048 "Alignment problem (0x%08llx)", (unsigned long long)start);
2049 __ emit_data64(0x8080808080808080, relocInfo::none);
2050 __ emit_data64(0x8080808080808080, relocInfo::none);
2051 __ emit_data64(0x8080808080808080, relocInfo::none);
2052 __ emit_data64(0x8080808080808080, relocInfo::none);
2053 __ emit_data64(0x8080808080808080, relocInfo::none);
2054 __ emit_data64(0x80803e8080808080, relocInfo::none);
2055 __ emit_data64(0x3b3a393837363534, relocInfo::none);
2056 __ emit_data64(0x8080808080803d3c, relocInfo::none);
2057
2058 return start;
2059 }
2060
2061 address StubGenerator::base64_vbmi_lookup_hi_url_addr() {
2062 __ align64();
2063 StubCodeMark mark(this, "StubRoutines", "lookup_hi_base64url");
2064 address start = __ pc();
2065
2066 assert(((unsigned long long)start & 0x3f) == 0,
2067 "Alignment problem (0x%08llx)", (unsigned long long)start);
2068 __ emit_data64(0x0605040302010080, relocInfo::none);
2069 __ emit_data64(0x0e0d0c0b0a090807, relocInfo::none);
2070 __ emit_data64(0x161514131211100f, relocInfo::none);
2071 __ emit_data64(0x3f80808080191817, relocInfo::none);
2072 __ emit_data64(0x201f1e1d1c1b1a80, relocInfo::none);
2073 __ emit_data64(0x2827262524232221, relocInfo::none);
2074 __ emit_data64(0x302f2e2d2c2b2a29, relocInfo::none);
2075 __ emit_data64(0x8080808080333231, relocInfo::none);
2076
2077 return start;
2078 }
2079
2080 address StubGenerator::base64_vbmi_pack_vec_addr() {
2081 __ align64();
2082 StubCodeMark mark(this, "StubRoutines", "pack_vec_base64");
2083 address start = __ pc();
2084
2085 assert(((unsigned long long)start & 0x3f) == 0,
2086 "Alignment problem (0x%08llx)", (unsigned long long)start);
2087 __ emit_data64(0x090a040506000102, relocInfo::none);
2088 __ emit_data64(0x161011120c0d0e08, relocInfo::none);
2089 __ emit_data64(0x1c1d1e18191a1415, relocInfo::none);
2090 __ emit_data64(0x292a242526202122, relocInfo::none);
2091 __ emit_data64(0x363031322c2d2e28, relocInfo::none);
2092 __ emit_data64(0x3c3d3e38393a3435, relocInfo::none);
2093 __ emit_data64(0x0000000000000000, relocInfo::none);
2094 __ emit_data64(0x0000000000000000, relocInfo::none);
2095
2096 return start;
2097 }
2098
2099 address StubGenerator::base64_vbmi_join_0_1_addr() {
2100 __ align64();
2101 StubCodeMark mark(this, "StubRoutines", "join_0_1_base64");
2102 address start = __ pc();
2103
2104 assert(((unsigned long long)start & 0x3f) == 0,
2105 "Alignment problem (0x%08llx)", (unsigned long long)start);
2106 __ emit_data64(0x090a040506000102, relocInfo::none);
2107 __ emit_data64(0x161011120c0d0e08, relocInfo::none);
2108 __ emit_data64(0x1c1d1e18191a1415, relocInfo::none);
2109 __ emit_data64(0x292a242526202122, relocInfo::none);
2110 __ emit_data64(0x363031322c2d2e28, relocInfo::none);
2111 __ emit_data64(0x3c3d3e38393a3435, relocInfo::none);
2112 __ emit_data64(0x494a444546404142, relocInfo::none);
2113 __ emit_data64(0x565051524c4d4e48, relocInfo::none);
2114
2115 return start;
2116 }
2117
2118 address StubGenerator::base64_vbmi_join_1_2_addr() {
2119 __ align64();
2120 StubCodeMark mark(this, "StubRoutines", "join_1_2_base64");
2121 address start = __ pc();
2122
2123 assert(((unsigned long long)start & 0x3f) == 0,
2124 "Alignment problem (0x%08llx)", (unsigned long long)start);
2125 __ emit_data64(0x1c1d1e18191a1415, relocInfo::none);
2126 __ emit_data64(0x292a242526202122, relocInfo::none);
2127 __ emit_data64(0x363031322c2d2e28, relocInfo::none);
2128 __ emit_data64(0x3c3d3e38393a3435, relocInfo::none);
2129 __ emit_data64(0x494a444546404142, relocInfo::none);
2130 __ emit_data64(0x565051524c4d4e48, relocInfo::none);
2131 __ emit_data64(0x5c5d5e58595a5455, relocInfo::none);
2132 __ emit_data64(0x696a646566606162, relocInfo::none);
2133
2134 return start;
2135 }
2136
2137 address StubGenerator::base64_vbmi_join_2_3_addr() {
2138 __ align64();
2139 StubCodeMark mark(this, "StubRoutines", "join_2_3_base64");
2140 address start = __ pc();
2141
2142 assert(((unsigned long long)start & 0x3f) == 0,
2143 "Alignment problem (0x%08llx)", (unsigned long long)start);
2144 __ emit_data64(0x363031322c2d2e28, relocInfo::none);
2145 __ emit_data64(0x3c3d3e38393a3435, relocInfo::none);
2146 __ emit_data64(0x494a444546404142, relocInfo::none);
2147 __ emit_data64(0x565051524c4d4e48, relocInfo::none);
2148 __ emit_data64(0x5c5d5e58595a5455, relocInfo::none);
2149 __ emit_data64(0x696a646566606162, relocInfo::none);
2150 __ emit_data64(0x767071726c6d6e68, relocInfo::none);
2151 __ emit_data64(0x7c7d7e78797a7475, relocInfo::none);
2152
2153 return start;
2154 }
2155
2156 address StubGenerator::base64_AVX2_decode_tables_addr() {
2157 __ align64();
2158 StubCodeMark mark(this, "StubRoutines", "AVX2_tables_base64");
2159 address start = __ pc();
2160
2161 assert(((unsigned long long)start & 0x3f) == 0,
2162 "Alignment problem (0x%08llx)", (unsigned long long)start);
2163 __ emit_data(0x2f2f2f2f, relocInfo::none, 0);
2164 __ emit_data(0x5f5f5f5f, relocInfo::none, 0); // for URL
2165
2166 __ emit_data(0xffffffff, relocInfo::none, 0);
2167 __ emit_data(0xfcfcfcfc, relocInfo::none, 0); // for URL
2168
2169 // Permute table
2170 __ emit_data64(0x0000000100000000, relocInfo::none);
2171 __ emit_data64(0x0000000400000002, relocInfo::none);
2172 __ emit_data64(0x0000000600000005, relocInfo::none);
2173 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2174
2175 // Shuffle table
2176 __ emit_data64(0x090a040506000102, relocInfo::none);
2177 __ emit_data64(0xffffffff0c0d0e08, relocInfo::none);
2178 __ emit_data64(0x090a040506000102, relocInfo::none);
2179 __ emit_data64(0xffffffff0c0d0e08, relocInfo::none);
2180
2181 // merge table
2182 __ emit_data(0x01400140, relocInfo::none, 0);
2183
2184 // merge multiplier
2185 __ emit_data(0x00011000, relocInfo::none, 0);
2186
2187 return start;
2188 }
2189
2190 address StubGenerator::base64_AVX2_decode_LUT_tables_addr() {
2191 __ align64();
2192 StubCodeMark mark(this, "StubRoutines", "AVX2_tables_URL_base64");
2193 address start = __ pc();
2194
2195 assert(((unsigned long long)start & 0x3f) == 0,
2196 "Alignment problem (0x%08llx)", (unsigned long long)start);
2197 // lut_lo
2198 __ emit_data64(0x1111111111111115, relocInfo::none);
2199 __ emit_data64(0x1a1b1b1b1a131111, relocInfo::none);
2200 __ emit_data64(0x1111111111111115, relocInfo::none);
2201 __ emit_data64(0x1a1b1b1b1a131111, relocInfo::none);
2202
2203 // lut_roll
2204 __ emit_data64(0xb9b9bfbf04131000, relocInfo::none);
2205 __ emit_data64(0x0000000000000000, relocInfo::none);
2206 __ emit_data64(0xb9b9bfbf04131000, relocInfo::none);
2207 __ emit_data64(0x0000000000000000, relocInfo::none);
2208
2209 // lut_lo URL
2210 __ emit_data64(0x1111111111111115, relocInfo::none);
2211 __ emit_data64(0x1b1b1a1b1b131111, relocInfo::none);
2212 __ emit_data64(0x1111111111111115, relocInfo::none);
2213 __ emit_data64(0x1b1b1a1b1b131111, relocInfo::none);
2214
2215 // lut_roll URL
2216 __ emit_data64(0xb9b9bfbf0411e000, relocInfo::none);
2217 __ emit_data64(0x0000000000000000, relocInfo::none);
2218 __ emit_data64(0xb9b9bfbf0411e000, relocInfo::none);
2219 __ emit_data64(0x0000000000000000, relocInfo::none);
2220
2221 // lut_hi
2222 __ emit_data64(0x0804080402011010, relocInfo::none);
2223 __ emit_data64(0x1010101010101010, relocInfo::none);
2224 __ emit_data64(0x0804080402011010, relocInfo::none);
2225 __ emit_data64(0x1010101010101010, relocInfo::none);
2226
2227 return start;
2228 }
2229
2230 address StubGenerator::base64_decoding_table_addr() {
2231 StubCodeMark mark(this, "StubRoutines", "decoding_table_base64");
2232 address start = __ pc();
2233
2234 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2235 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2236 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2237 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2238 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2239 __ emit_data64(0x3fffffff3effffff, relocInfo::none);
2240 __ emit_data64(0x3b3a393837363534, relocInfo::none);
2241 __ emit_data64(0xffffffffffff3d3c, relocInfo::none);
2242 __ emit_data64(0x06050403020100ff, relocInfo::none);
2243 __ emit_data64(0x0e0d0c0b0a090807, relocInfo::none);
2244 __ emit_data64(0x161514131211100f, relocInfo::none);
2245 __ emit_data64(0xffffffffff191817, relocInfo::none);
2246 __ emit_data64(0x201f1e1d1c1b1aff, relocInfo::none);
2247 __ emit_data64(0x2827262524232221, relocInfo::none);
2248 __ emit_data64(0x302f2e2d2c2b2a29, relocInfo::none);
2249 __ emit_data64(0xffffffffff333231, relocInfo::none);
2250 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2251 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2252 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2253 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2254 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2255 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2256 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2257 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2258 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2259 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2260 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2261 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2262 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2263 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2264 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2265 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2266
2267 // URL table
2268 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2269 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2270 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2271 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2272 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2273 __ emit_data64(0xffff3effffffffff, relocInfo::none);
2274 __ emit_data64(0x3b3a393837363534, relocInfo::none);
2275 __ emit_data64(0xffffffffffff3d3c, relocInfo::none);
2276 __ emit_data64(0x06050403020100ff, relocInfo::none);
2277 __ emit_data64(0x0e0d0c0b0a090807, relocInfo::none);
2278 __ emit_data64(0x161514131211100f, relocInfo::none);
2279 __ emit_data64(0x3fffffffff191817, relocInfo::none);
2280 __ emit_data64(0x201f1e1d1c1b1aff, relocInfo::none);
2281 __ emit_data64(0x2827262524232221, relocInfo::none);
2282 __ emit_data64(0x302f2e2d2c2b2a29, relocInfo::none);
2283 __ emit_data64(0xffffffffff333231, relocInfo::none);
2284 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2285 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2286 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2287 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2288 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2289 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2290 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2291 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2292 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2293 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2294 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2295 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2296 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2297 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2298 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2299 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2300
2301 return start;
2302 }
2303
2304
2305 // Code for generating Base64 decoding.
2306 //
2307 // Based on the article (and associated code) from https://arxiv.org/abs/1910.05109.
2308 //
2309 // Intrinsic function prototype in Base64.java:
2310 // private void decodeBlock(byte[] src, int sp, int sl, byte[] dst, int dp, boolean isURL, isMIME) {
2311 address StubGenerator::generate_base64_decodeBlock() {
2312 __ align(CodeEntryAlignment);
2313 StubCodeMark mark(this, "StubRoutines", "implDecode");
2314 address start = __ pc();
2315
2316 __ enter();
2317
2318 // Save callee-saved registers before using them
2319 __ push(r12);
2320 __ push(r13);
2321 __ push(r14);
2322 __ push(r15);
2323 __ push(rbx);
2324
2325 // arguments
2326 const Register source = c_rarg0; // Source Array
2327 const Register start_offset = c_rarg1; // start offset
2328 const Register end_offset = c_rarg2; // end offset
2329 const Register dest = c_rarg3; // destination array
2330 const Register isMIME = rbx;
2331
2332 #ifndef _WIN64
2333 const Register dp = c_rarg4; // Position for writing to dest array
2334 const Register isURL = c_rarg5;// Base64 or URL character set
2335 __ movl(isMIME, Address(rbp, 2 * wordSize));
2336 #else
2337 const Address dp_mem(rbp, 6 * wordSize); // length is on stack on Win64
2338 const Address isURL_mem(rbp, 7 * wordSize);
2339 const Register isURL = r10; // pick the volatile windows register
2340 const Register dp = r12;
2341 __ movl(dp, dp_mem);
2342 __ movl(isURL, isURL_mem);
2343 __ movl(isMIME, Address(rbp, 8 * wordSize));
2344 #endif
2345
2346 const XMMRegister lookup_lo = xmm5;
2347 const XMMRegister lookup_hi = xmm6;
2348 const XMMRegister errorvec = xmm7;
2349 const XMMRegister pack16_op = xmm9;
2350 const XMMRegister pack32_op = xmm8;
2351 const XMMRegister input0 = xmm3;
2352 const XMMRegister input1 = xmm20;
2353 const XMMRegister input2 = xmm21;
2354 const XMMRegister input3 = xmm19;
2355 const XMMRegister join01 = xmm12;
2356 const XMMRegister join12 = xmm11;
2357 const XMMRegister join23 = xmm10;
2358 const XMMRegister translated0 = xmm2;
2359 const XMMRegister translated1 = xmm1;
2360 const XMMRegister translated2 = xmm0;
2361 const XMMRegister translated3 = xmm4;
2362
2363 const XMMRegister merged0 = xmm2;
2364 const XMMRegister merged1 = xmm1;
2365 const XMMRegister merged2 = xmm0;
2366 const XMMRegister merged3 = xmm4;
2367 const XMMRegister merge_ab_bc0 = xmm2;
2368 const XMMRegister merge_ab_bc1 = xmm1;
2369 const XMMRegister merge_ab_bc2 = xmm0;
2370 const XMMRegister merge_ab_bc3 = xmm4;
2371
2372 const XMMRegister pack24bits = xmm4;
2373
2374 const Register length = r14;
2375 const Register output_size = r13;
2376 const Register output_mask = r15;
2377 const KRegister input_mask = k1;
2378
2379 const XMMRegister input_initial_valid_b64 = xmm0;
2380 const XMMRegister tmp = xmm10;
2381 const XMMRegister mask = xmm0;
2382 const XMMRegister invalid_b64 = xmm1;
2383
2384 Label L_process256, L_process64, L_process64Loop, L_exit, L_processdata, L_loadURL;
2385 Label L_continue, L_finalBit, L_padding, L_donePadding, L_bruteForce;
2386 Label L_forceLoop, L_bottomLoop, L_checkMIME, L_exit_no_vzero, L_lastChunk;
2387
2388 // calculate length from offsets
2389 __ movl(length, end_offset);
2390 __ subl(length, start_offset);
2391 __ push(dest); // Save for return value calc
2392
2393 // If AVX512 VBMI not supported, just compile non-AVX code
2394 if(VM_Version::supports_avx512_vbmi() &&
2395 VM_Version::supports_avx512bw()) {
2396 __ cmpl(length, 31); // 32-bytes is break-even for AVX-512
2397 __ jcc(Assembler::lessEqual, L_lastChunk);
2398
2399 __ cmpl(isMIME, 0);
2400 __ jcc(Assembler::notEqual, L_lastChunk);
2401
2402 // Load lookup tables based on isURL
2403 __ cmpl(isURL, 0);
2404 __ jcc(Assembler::notZero, L_loadURL);
2405
2406 __ evmovdquq(lookup_lo, ExternalAddress(StubRoutines::x86::base64_vbmi_lookup_lo_addr()), Assembler::AVX_512bit, r13);
2407 __ evmovdquq(lookup_hi, ExternalAddress(StubRoutines::x86::base64_vbmi_lookup_hi_addr()), Assembler::AVX_512bit, r13);
2408
2409 __ BIND(L_continue);
2410
2411 __ movl(r15, 0x01400140);
2412 __ evpbroadcastd(pack16_op, r15, Assembler::AVX_512bit);
2413
2414 __ movl(r15, 0x00011000);
2415 __ evpbroadcastd(pack32_op, r15, Assembler::AVX_512bit);
2416
2417 __ cmpl(length, 0xff);
2418 __ jcc(Assembler::lessEqual, L_process64);
2419
2420 // load masks required for decoding data
2421 __ BIND(L_processdata);
2422 __ evmovdquq(join01, ExternalAddress(StubRoutines::x86::base64_vbmi_join_0_1_addr()), Assembler::AVX_512bit,r13);
2423 __ evmovdquq(join12, ExternalAddress(StubRoutines::x86::base64_vbmi_join_1_2_addr()), Assembler::AVX_512bit, r13);
2424 __ evmovdquq(join23, ExternalAddress(StubRoutines::x86::base64_vbmi_join_2_3_addr()), Assembler::AVX_512bit, r13);
2425
2426 __ align32();
2427 __ BIND(L_process256);
2428 // Grab input data
2429 __ evmovdquq(input0, Address(source, start_offset, Address::times_1, 0x00), Assembler::AVX_512bit);
2430 __ evmovdquq(input1, Address(source, start_offset, Address::times_1, 0x40), Assembler::AVX_512bit);
2431 __ evmovdquq(input2, Address(source, start_offset, Address::times_1, 0x80), Assembler::AVX_512bit);
2432 __ evmovdquq(input3, Address(source, start_offset, Address::times_1, 0xc0), Assembler::AVX_512bit);
2433
2434 // Copy the low part of the lookup table into the destination of the permutation
2435 __ evmovdquq(translated0, lookup_lo, Assembler::AVX_512bit);
2436 __ evmovdquq(translated1, lookup_lo, Assembler::AVX_512bit);
2437 __ evmovdquq(translated2, lookup_lo, Assembler::AVX_512bit);
2438 __ evmovdquq(translated3, lookup_lo, Assembler::AVX_512bit);
2439
2440 // Translate the base64 input into "decoded" bytes
2441 __ evpermt2b(translated0, input0, lookup_hi, Assembler::AVX_512bit);
2442 __ evpermt2b(translated1, input1, lookup_hi, Assembler::AVX_512bit);
2443 __ evpermt2b(translated2, input2, lookup_hi, Assembler::AVX_512bit);
2444 __ evpermt2b(translated3, input3, lookup_hi, Assembler::AVX_512bit);
2445
2446 // OR all of the translations together to check for errors (high-order bit of byte set)
2447 __ vpternlogd(input0, 0xfe, input1, input2, Assembler::AVX_512bit);
2448
2449 __ vpternlogd(input3, 0xfe, translated0, translated1, Assembler::AVX_512bit);
2450 __ vpternlogd(input0, 0xfe, translated2, translated3, Assembler::AVX_512bit);
2451 __ vpor(errorvec, input3, input0, Assembler::AVX_512bit);
2452
2453 // Check if there was an error - if so, try 64-byte chunks
2454 __ evpmovb2m(k3, errorvec, Assembler::AVX_512bit);
2455 __ kortestql(k3, k3);
2456 __ jcc(Assembler::notZero, L_process64);
2457
2458 // The merging and shuffling happens here
2459 // We multiply each byte pair [00dddddd | 00cccccc | 00bbbbbb | 00aaaaaa]
2460 // Multiply [00cccccc] by 2^6 added to [00dddddd] to get [0000cccc | ccdddddd]
2461 // The pack16_op is a vector of 0x01400140, so multiply D by 1 and C by 0x40
2462 __ vpmaddubsw(merge_ab_bc0, translated0, pack16_op, Assembler::AVX_512bit);
2463 __ vpmaddubsw(merge_ab_bc1, translated1, pack16_op, Assembler::AVX_512bit);
2464 __ vpmaddubsw(merge_ab_bc2, translated2, pack16_op, Assembler::AVX_512bit);
2465 __ vpmaddubsw(merge_ab_bc3, translated3, pack16_op, Assembler::AVX_512bit);
2466
2467 // Now do the same with packed 16-bit values.
2468 // We start with [0000cccc | ccdddddd | 0000aaaa | aabbbbbb]
2469 // pack32_op is 0x00011000 (2^12, 1), so this multiplies [0000aaaa | aabbbbbb] by 2^12
2470 // and adds [0000cccc | ccdddddd] to yield [00000000 | aaaaaabb | bbbbcccc | ccdddddd]
2471 __ vpmaddwd(merged0, merge_ab_bc0, pack32_op, Assembler::AVX_512bit);
2472 __ vpmaddwd(merged1, merge_ab_bc1, pack32_op, Assembler::AVX_512bit);
2473 __ vpmaddwd(merged2, merge_ab_bc2, pack32_op, Assembler::AVX_512bit);
2474 __ vpmaddwd(merged3, merge_ab_bc3, pack32_op, Assembler::AVX_512bit);
2475
2476 // The join vectors specify which byte from which vector goes into the outputs
2477 // One of every 4 bytes in the extended vector is zero, so we pack them into their
2478 // final positions in the register for storing (256 bytes in, 192 bytes out)
2479 __ evpermt2b(merged0, join01, merged1, Assembler::AVX_512bit);
2480 __ evpermt2b(merged1, join12, merged2, Assembler::AVX_512bit);
2481 __ evpermt2b(merged2, join23, merged3, Assembler::AVX_512bit);
2482
2483 // Store result
2484 __ evmovdquq(Address(dest, dp, Address::times_1, 0x00), merged0, Assembler::AVX_512bit);
2485 __ evmovdquq(Address(dest, dp, Address::times_1, 0x40), merged1, Assembler::AVX_512bit);
2486 __ evmovdquq(Address(dest, dp, Address::times_1, 0x80), merged2, Assembler::AVX_512bit);
2487
2488 __ addptr(source, 0x100);
2489 __ addptr(dest, 0xc0);
2490 __ subl(length, 0x100);
2491 __ cmpl(length, 64 * 4);
2492 __ jcc(Assembler::greaterEqual, L_process256);
2493
2494 // At this point, we've decoded 64 * 4 * n bytes.
2495 // The remaining length will be <= 64 * 4 - 1.
2496 // UNLESS there was an error decoding the first 256-byte chunk. In this
2497 // case, the length will be arbitrarily long.
2498 //
2499 // Note that this will be the path for MIME-encoded strings.
2500
2501 __ BIND(L_process64);
2502
2503 __ evmovdquq(pack24bits, ExternalAddress(StubRoutines::x86::base64_vbmi_pack_vec_addr()), Assembler::AVX_512bit, r13);
2504
2505 __ cmpl(length, 63);
2506 __ jcc(Assembler::lessEqual, L_finalBit);
2507
2508 __ mov64(rax, 0x0000ffffffffffff);
2509 __ kmovql(k2, rax);
2510
2511 __ align32();
2512 __ BIND(L_process64Loop);
2513
2514 // Handle first 64-byte block
2515
2516 __ evmovdquq(input0, Address(source, start_offset), Assembler::AVX_512bit);
2517 __ evmovdquq(translated0, lookup_lo, Assembler::AVX_512bit);
2518 __ evpermt2b(translated0, input0, lookup_hi, Assembler::AVX_512bit);
2519
2520 __ vpor(errorvec, translated0, input0, Assembler::AVX_512bit);
2521
2522 // Check for error and bomb out before updating dest
2523 __ evpmovb2m(k3, errorvec, Assembler::AVX_512bit);
2524 __ kortestql(k3, k3);
2525 __ jcc(Assembler::notZero, L_exit);
2526
2527 // Pack output register, selecting correct byte ordering
2528 __ vpmaddubsw(merge_ab_bc0, translated0, pack16_op, Assembler::AVX_512bit);
2529 __ vpmaddwd(merged0, merge_ab_bc0, pack32_op, Assembler::AVX_512bit);
2530 __ vpermb(merged0, pack24bits, merged0, Assembler::AVX_512bit);
2531
2532 __ evmovdqub(Address(dest, dp), k2, merged0, true, Assembler::AVX_512bit);
2533
2534 __ subl(length, 64);
2535 __ addptr(source, 64);
2536 __ addptr(dest, 48);
2537
2538 __ cmpl(length, 64);
2539 __ jcc(Assembler::greaterEqual, L_process64Loop);
2540
2541 __ cmpl(length, 0);
2542 __ jcc(Assembler::lessEqual, L_exit);
2543
2544 __ BIND(L_finalBit);
2545 // Now have 1 to 63 bytes left to decode
2546
2547 // I was going to let Java take care of the final fragment
2548 // however it will repeatedly call this routine for every 4 bytes
2549 // of input data, so handle the rest here.
2550 __ movq(rax, -1);
2551 __ bzhiq(rax, rax, length); // Input mask in rax
2552
2553 __ movl(output_size, length);
2554 __ shrl(output_size, 2); // Find (len / 4) * 3 (output length)
2555 __ lea(output_size, Address(output_size, output_size, Address::times_2, 0));
2556 // output_size in r13
2557
2558 // Strip pad characters, if any, and adjust length and mask
2559 __ cmpb(Address(source, length, Address::times_1, -1), '=');
2560 __ jcc(Assembler::equal, L_padding);
2561
2562 __ BIND(L_donePadding);
2563
2564 // Output size is (64 - output_size), output mask is (all 1s >> output_size).
2565 __ kmovql(input_mask, rax);
2566 __ movq(output_mask, -1);
2567 __ bzhiq(output_mask, output_mask, output_size);
2568
2569 // Load initial input with all valid base64 characters. Will be used
2570 // in merging source bytes to avoid masking when determining if an error occurred.
2571 __ movl(rax, 0x61616161);
2572 __ evpbroadcastd(input_initial_valid_b64, rax, Assembler::AVX_512bit);
2573
2574 // A register containing all invalid base64 decoded values
2575 __ movl(rax, 0x80808080);
2576 __ evpbroadcastd(invalid_b64, rax, Assembler::AVX_512bit);
2577
2578 // input_mask is in k1
2579 // output_size is in r13
2580 // output_mask is in r15
2581 // zmm0 - free
2582 // zmm1 - 0x00011000
2583 // zmm2 - 0x01400140
2584 // zmm3 - errorvec
2585 // zmm4 - pack vector
2586 // zmm5 - lookup_lo
2587 // zmm6 - lookup_hi
2588 // zmm7 - errorvec
2589 // zmm8 - 0x61616161
2590 // zmm9 - 0x80808080
2591
2592 // Load only the bytes from source, merging into our "fully-valid" register
2593 __ evmovdqub(input_initial_valid_b64, input_mask, Address(source, start_offset, Address::times_1, 0x0), true, Assembler::AVX_512bit);
2594
2595 // Decode all bytes within our merged input
2596 __ evmovdquq(tmp, lookup_lo, Assembler::AVX_512bit);
2597 __ evpermt2b(tmp, input_initial_valid_b64, lookup_hi, Assembler::AVX_512bit);
2598 __ evporq(mask, tmp, input_initial_valid_b64, Assembler::AVX_512bit);
2599
2600 // Check for error. Compare (decoded | initial) to all invalid.
2601 // If any bytes have their high-order bit set, then we have an error.
2602 __ evptestmb(k2, mask, invalid_b64, Assembler::AVX_512bit);
2603 __ kortestql(k2, k2);
2604
2605 // If we have an error, use the brute force loop to decode what we can (4-byte chunks).
2606 __ jcc(Assembler::notZero, L_bruteForce);
2607
2608 // Shuffle output bytes
2609 __ vpmaddubsw(tmp, tmp, pack16_op, Assembler::AVX_512bit);
2610 __ vpmaddwd(tmp, tmp, pack32_op, Assembler::AVX_512bit);
2611
2612 __ vpermb(tmp, pack24bits, tmp, Assembler::AVX_512bit);
2613 __ kmovql(k1, output_mask);
2614 __ evmovdqub(Address(dest, dp), k1, tmp, true, Assembler::AVX_512bit);
2615
2616 __ addptr(dest, output_size);
2617
2618 __ BIND(L_exit);
2619 __ vzeroupper();
2620 __ pop(rax); // Get original dest value
2621 __ subptr(dest, rax); // Number of bytes converted
2622 __ movptr(rax, dest);
2623 __ pop(rbx);
2624 __ pop(r15);
2625 __ pop(r14);
2626 __ pop(r13);
2627 __ pop(r12);
2628 __ leave();
2629 __ ret(0);
2630
2631 __ BIND(L_loadURL);
2632 __ evmovdquq(lookup_lo, ExternalAddress(StubRoutines::x86::base64_vbmi_lookup_lo_url_addr()), Assembler::AVX_512bit, r13);
2633 __ evmovdquq(lookup_hi, ExternalAddress(StubRoutines::x86::base64_vbmi_lookup_hi_url_addr()), Assembler::AVX_512bit, r13);
2634 __ jmp(L_continue);
2635
2636 __ BIND(L_padding);
2637 __ decrementq(output_size, 1);
2638 __ shrq(rax, 1);
2639
2640 __ cmpb(Address(source, length, Address::times_1, -2), '=');
2641 __ jcc(Assembler::notEqual, L_donePadding);
2642
2643 __ decrementq(output_size, 1);
2644 __ shrq(rax, 1);
2645 __ jmp(L_donePadding);
2646
2647 __ align32();
2648 __ BIND(L_bruteForce);
2649 } // End of if(avx512_vbmi)
2650
2651 if (VM_Version::supports_avx2()) {
2652 Label L_tailProc, L_topLoop, L_enterLoop;
2653
2654 __ cmpl(isMIME, 0);
2655 __ jcc(Assembler::notEqual, L_lastChunk);
2656
2657 // Check for buffer too small (for algorithm)
2658 __ subl(length, 0x2c);
2659 __ jcc(Assembler::less, L_tailProc);
2660
2661 __ shll(isURL, 2);
2662
2663 // Algorithm adapted from https://arxiv.org/abs/1704.00605, "Faster Base64
2664 // Encoding and Decoding using AVX2 Instructions". URL modifications added.
2665
2666 // Set up constants
2667 __ lea(r13, ExternalAddress(StubRoutines::x86::base64_AVX2_decode_tables_addr()));
2668 __ vpbroadcastd(xmm4, Address(r13, isURL, Address::times_1), Assembler::AVX_256bit); // 2F or 5F
2669 __ vpbroadcastd(xmm10, Address(r13, isURL, Address::times_1, 0x08), Assembler::AVX_256bit); // -1 or -4
2670 __ vmovdqu(xmm12, Address(r13, 0x10)); // permute
2671 __ vmovdqu(xmm13, Address(r13, 0x30)); // shuffle
2672 __ vpbroadcastd(xmm7, Address(r13, 0x50), Assembler::AVX_256bit); // merge
2673 __ vpbroadcastd(xmm6, Address(r13, 0x54), Assembler::AVX_256bit); // merge mult
2674
2675 __ lea(r13, ExternalAddress(StubRoutines::x86::base64_AVX2_decode_LUT_tables_addr()));
2676 __ shll(isURL, 4);
2677 __ vmovdqu(xmm11, Address(r13, isURL, Address::times_1, 0x00)); // lut_lo
2678 __ vmovdqu(xmm8, Address(r13, isURL, Address::times_1, 0x20)); // lut_roll
2679 __ shrl(isURL, 6); // restore isURL
2680 __ vmovdqu(xmm9, Address(r13, 0x80)); // lut_hi
2681 __ jmp(L_enterLoop);
2682
2683 __ align32();
2684 __ bind(L_topLoop);
2685 // Add in the offset value (roll) to get 6-bit out values
2686 __ vpaddb(xmm0, xmm0, xmm2, Assembler::AVX_256bit);
2687 // Merge and permute the output bits into appropriate output byte lanes
2688 __ vpmaddubsw(xmm0, xmm0, xmm7, Assembler::AVX_256bit);
2689 __ vpmaddwd(xmm0, xmm0, xmm6, Assembler::AVX_256bit);
2690 __ vpshufb(xmm0, xmm0, xmm13, Assembler::AVX_256bit);
2691 __ vpermd(xmm0, xmm12, xmm0, Assembler::AVX_256bit);
2692 // Store the output bytes
2693 __ vmovdqu(Address(dest, dp, Address::times_1, 0), xmm0);
2694 __ addptr(source, 0x20);
2695 __ addptr(dest, 0x18);
2696 __ subl(length, 0x20);
2697 __ jcc(Assembler::less, L_tailProc);
2698
2699 __ bind(L_enterLoop);
2700
2701 // Load in encoded string (32 bytes)
2702 __ vmovdqu(xmm2, Address(source, start_offset, Address::times_1, 0x0));
2703 // Extract the high nibble for indexing into the lut tables. High 4 bits are don't care.
2704 __ vpsrld(xmm1, xmm2, 0x4, Assembler::AVX_256bit);
2705 __ vpand(xmm1, xmm4, xmm1, Assembler::AVX_256bit);
2706 // Extract the low nibble. 5F/2F will isolate the low-order 4 bits. High 4 bits are don't care.
2707 __ vpand(xmm3, xmm2, xmm4, Assembler::AVX_256bit);
2708 // Check for special-case (0x2F or 0x5F (URL))
2709 __ vpcmpeqb(xmm0, xmm4, xmm2, Assembler::AVX_256bit);
2710 // Get the bitset based on the low nibble. vpshufb uses low-order 4 bits only.
2711 __ vpshufb(xmm3, xmm11, xmm3, Assembler::AVX_256bit);
2712 // Get the bit value of the high nibble
2713 __ vpshufb(xmm5, xmm9, xmm1, Assembler::AVX_256bit);
2714 // Make sure 2F / 5F shows as valid
2715 __ vpandn(xmm3, xmm0, xmm3, Assembler::AVX_256bit);
2716 // Make adjustment for roll index. For non-URL, this is a no-op,
2717 // for URL, this adjusts by -4. This is to properly index the
2718 // roll value for 2F / 5F.
2719 __ vpand(xmm0, xmm0, xmm10, Assembler::AVX_256bit);
2720 // If the and of the two is non-zero, we have an invalid input character
2721 __ vptest(xmm3, xmm5);
2722 // Extract the "roll" value - value to add to the input to get 6-bit out value
2723 __ vpaddb(xmm0, xmm0, xmm1, Assembler::AVX_256bit); // Handle 2F / 5F
2724 __ vpshufb(xmm0, xmm8, xmm0, Assembler::AVX_256bit);
2725 __ jcc(Assembler::equal, L_topLoop); // Fall through on error
2726
2727 __ bind(L_tailProc);
2728
2729 __ addl(length, 0x2c);
2730
2731 __ vzeroupper();
2732 }
2733
2734 // Use non-AVX code to decode 4-byte chunks into 3 bytes of output
2735
2736 // Register state (Linux):
2737 // r12-15 - saved on stack
2738 // rdi - src
2739 // rsi - sp
2740 // rdx - sl
2741 // rcx - dst
2742 // r8 - dp
2743 // r9 - isURL
2744
2745 // Register state (Windows):
2746 // r12-15 - saved on stack
2747 // rcx - src
2748 // rdx - sp
2749 // r8 - sl
2750 // r9 - dst
2751 // r12 - dp
2752 // r10 - isURL
2753
2754 // Registers (common):
2755 // length (r14) - bytes in src
2756
2757 const Register decode_table = r11;
2758 const Register out_byte_count = rbx;
2759 const Register byte1 = r13;
2760 const Register byte2 = r15;
2761 const Register byte3 = WIN64_ONLY(r8) NOT_WIN64(rdx);
2762 const Register byte4 = WIN64_ONLY(r10) NOT_WIN64(r9);
2763
2764 __ bind(L_lastChunk);
2765
2766 __ shrl(length, 2); // Multiple of 4 bytes only - length is # 4-byte chunks
2767 __ cmpl(length, 0);
2768 __ jcc(Assembler::lessEqual, L_exit_no_vzero);
2769
2770 __ shll(isURL, 8); // index into decode table based on isURL
2771 __ lea(decode_table, ExternalAddress(StubRoutines::x86::base64_decoding_table_addr()));
2772 __ addptr(decode_table, isURL);
2773
2774 __ jmp(L_bottomLoop);
2775
2776 __ align32();
2777 __ BIND(L_forceLoop);
2778 __ shll(byte1, 18);
2779 __ shll(byte2, 12);
2780 __ shll(byte3, 6);
2781 __ orl(byte1, byte2);
2782 __ orl(byte1, byte3);
2783 __ orl(byte1, byte4);
2784
2785 __ addptr(source, 4);
2786
2787 __ movb(Address(dest, dp, Address::times_1, 2), byte1);
2788 __ shrl(byte1, 8);
2789 __ movb(Address(dest, dp, Address::times_1, 1), byte1);
2790 __ shrl(byte1, 8);
2791 __ movb(Address(dest, dp, Address::times_1, 0), byte1);
2792
2793 __ addptr(dest, 3);
2794 __ decrementl(length, 1);
2795 __ jcc(Assembler::zero, L_exit_no_vzero);
2796
2797 __ BIND(L_bottomLoop);
2798 __ load_unsigned_byte(byte1, Address(source, start_offset, Address::times_1, 0x00));
2799 __ load_unsigned_byte(byte2, Address(source, start_offset, Address::times_1, 0x01));
2800 __ load_signed_byte(byte1, Address(decode_table, byte1));
2801 __ load_signed_byte(byte2, Address(decode_table, byte2));
2802 __ load_unsigned_byte(byte3, Address(source, start_offset, Address::times_1, 0x02));
2803 __ load_unsigned_byte(byte4, Address(source, start_offset, Address::times_1, 0x03));
2804 __ load_signed_byte(byte3, Address(decode_table, byte3));
2805 __ load_signed_byte(byte4, Address(decode_table, byte4));
2806
2807 __ mov(rax, byte1);
2808 __ orl(rax, byte2);
2809 __ orl(rax, byte3);
2810 __ orl(rax, byte4);
2811 __ jcc(Assembler::positive, L_forceLoop);
2812
2813 __ BIND(L_exit_no_vzero);
2814 __ pop(rax); // Get original dest value
2815 __ subptr(dest, rax); // Number of bytes converted
2816 __ movptr(rax, dest);
2817 __ pop(rbx);
2818 __ pop(r15);
2819 __ pop(r14);
2820 __ pop(r13);
2821 __ pop(r12);
2822 __ leave();
2823 __ ret(0);
2824
2825 return start;
2826 }
2827
2828
2829 /**
2830 * Arguments:
2831 *
2832 * Inputs:
2833 * c_rarg0 - int crc
2834 * c_rarg1 - byte* buf
2835 * c_rarg2 - int length
2836 *
2837 * Output:
2838 * rax - int crc result
2839 */
2840 address StubGenerator::generate_updateBytesCRC32() {
2841 assert(UseCRC32Intrinsics, "need AVX and CLMUL instructions");
2842
2843 __ align(CodeEntryAlignment);
2844 StubCodeMark mark(this, "StubRoutines", "updateBytesCRC32");
2845
2846 address start = __ pc();
2847
2848 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
2849 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
2850 // rscratch1: r10
2851 const Register crc = c_rarg0; // crc
2852 const Register buf = c_rarg1; // source java byte array address
2853 const Register len = c_rarg2; // length
2854 const Register table = c_rarg3; // crc_table address (reuse register)
2855 const Register tmp1 = r11;
2856 const Register tmp2 = r10;
2857 assert_different_registers(crc, buf, len, table, tmp1, tmp2, rax);
2858
2859 BLOCK_COMMENT("Entry:");
2860 __ enter(); // required for proper stackwalking of RuntimeStub frame
2861
2862 if (VM_Version::supports_sse4_1() && VM_Version::supports_avx512_vpclmulqdq() &&
2863 VM_Version::supports_avx512bw() &&
2864 VM_Version::supports_avx512vl()) {
2865 // The constants used in the CRC32 algorithm requires the 1's compliment of the initial crc value.
2866 // However, the constant table for CRC32-C assumes the original crc value. Account for this
2867 // difference before calling and after returning.
2868 __ lea(table, ExternalAddress(StubRoutines::x86::crc_table_avx512_addr()));
2869 __ notl(crc);
2870 __ kernel_crc32_avx512(crc, buf, len, table, tmp1, tmp2);
2871 __ notl(crc);
2872 } else {
2873 __ kernel_crc32(crc, buf, len, table, tmp1);
2874 }
2875
2876 __ movl(rax, crc);
2877 __ vzeroupper();
2878 __ leave(); // required for proper stackwalking of RuntimeStub frame
2879 __ ret(0);
2880
2881 return start;
2882 }
2883
2884 /**
2885 * Arguments:
2886 *
2887 * Inputs:
2888 * c_rarg0 - int crc
2889 * c_rarg1 - byte* buf
2890 * c_rarg2 - long length
2891 * c_rarg3 - table_start - optional (present only when doing a library_call,
2892 * not used by x86 algorithm)
2893 *
2894 * Output:
2895 * rax - int crc result
2896 */
2897 address StubGenerator::generate_updateBytesCRC32C(bool is_pclmulqdq_supported) {
2898 assert(UseCRC32CIntrinsics, "need SSE4_2");
2899 __ align(CodeEntryAlignment);
2900 StubCodeMark mark(this, "StubRoutines", "updateBytesCRC32C");
2901 address start = __ pc();
2902
2903 //reg.arg int#0 int#1 int#2 int#3 int#4 int#5 float regs
2904 //Windows RCX RDX R8 R9 none none XMM0..XMM3
2905 //Lin / Sol RDI RSI RDX RCX R8 R9 XMM0..XMM7
2906 const Register crc = c_rarg0; // crc
2907 const Register buf = c_rarg1; // source java byte array address
2908 const Register len = c_rarg2; // length
2909 const Register a = rax;
2910 const Register j = r9;
2911 const Register k = r10;
2912 const Register l = r11;
2913 #ifdef _WIN64
2914 const Register y = rdi;
2915 const Register z = rsi;
2916 #else
2917 const Register y = rcx;
2918 const Register z = r8;
2919 #endif
2920 assert_different_registers(crc, buf, len, a, j, k, l, y, z);
2921
2922 BLOCK_COMMENT("Entry:");
2923 __ enter(); // required for proper stackwalking of RuntimeStub frame
2924 Label L_continue;
2925
2926 if (VM_Version::supports_sse4_1() && VM_Version::supports_avx512_vpclmulqdq() &&
2927 VM_Version::supports_avx512bw() &&
2928 VM_Version::supports_avx512vl()) {
2929 Label L_doSmall;
2930
2931 __ cmpl(len, 384);
2932 __ jcc(Assembler::lessEqual, L_doSmall);
2933
2934 __ lea(j, ExternalAddress(StubRoutines::x86::crc32c_table_avx512_addr()));
2935 __ kernel_crc32_avx512(crc, buf, len, j, l, k);
2936
2937 __ jmp(L_continue);
2938
2939 __ bind(L_doSmall);
2940 }
2941 #ifdef _WIN64
2942 __ push(y);
2943 __ push(z);
2944 #endif
2945 __ crc32c_ipl_alg2_alt2(crc, buf, len,
2946 a, j, k,
2947 l, y, z,
2948 c_farg0, c_farg1, c_farg2,
2949 is_pclmulqdq_supported);
2950 #ifdef _WIN64
2951 __ pop(z);
2952 __ pop(y);
2953 #endif
2954
2955 __ bind(L_continue);
2956 __ movl(rax, crc);
2957 __ vzeroupper();
2958 __ leave(); // required for proper stackwalking of RuntimeStub frame
2959 __ ret(0);
2960
2961 return start;
2962 }
2963
2964
2965 /**
2966 * Arguments:
2967 *
2968 * Input:
2969 * c_rarg0 - x address
2970 * c_rarg1 - x length
2971 * c_rarg2 - y address
2972 * c_rarg3 - y length
2973 * not Win64
2974 * c_rarg4 - z address
2975 * c_rarg5 - z length
2976 * Win64
2977 * rsp+40 - z address
2978 * rsp+48 - z length
2979 */
2980 address StubGenerator::generate_multiplyToLen() {
2981 __ align(CodeEntryAlignment);
2982 StubCodeMark mark(this, "StubRoutines", "multiplyToLen");
2983 address start = __ pc();
2984
2985 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
2986 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
2987 const Register x = rdi;
2988 const Register xlen = rax;
2989 const Register y = rsi;
2990 const Register ylen = rcx;
2991 const Register z = r8;
2992 const Register zlen = r11;
2993
2994 // Next registers will be saved on stack in multiply_to_len().
2995 const Register tmp1 = r12;
2996 const Register tmp2 = r13;
2997 const Register tmp3 = r14;
2998 const Register tmp4 = r15;
2999 const Register tmp5 = rbx;
3000
3001 BLOCK_COMMENT("Entry:");
3002 __ enter(); // required for proper stackwalking of RuntimeStub frame
3003
3004 #ifndef _WIN64
3005 __ movptr(zlen, r9); // Save r9 in r11 - zlen
3006 #endif
3007 setup_arg_regs(4); // x => rdi, xlen => rsi, y => rdx
3008 // ylen => rcx, z => r8, zlen => r11
3009 // r9 and r10 may be used to save non-volatile registers
3010 #ifdef _WIN64
3011 // last 2 arguments (#4, #5) are on stack on Win64
3012 __ movptr(z, Address(rsp, 6 * wordSize));
3013 __ movptr(zlen, Address(rsp, 7 * wordSize));
3014 #endif
3015
3016 __ movptr(xlen, rsi);
3017 __ movptr(y, rdx);
3018 __ multiply_to_len(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5);
3019
3020 restore_arg_regs();
3021
3022 __ leave(); // required for proper stackwalking of RuntimeStub frame
3023 __ ret(0);
3024
3025 return start;
3026 }
3027
3028 /**
3029 * Arguments:
3030 *
3031 * Input:
3032 * c_rarg0 - obja address
3033 * c_rarg1 - objb address
3034 * c_rarg3 - length length
3035 * c_rarg4 - scale log2_array_indxscale
3036 *
3037 * Output:
3038 * rax - int >= mismatched index, < 0 bitwise complement of tail
3039 */
3040 address StubGenerator::generate_vectorizedMismatch() {
3041 __ align(CodeEntryAlignment);
3042 StubCodeMark mark(this, "StubRoutines", "vectorizedMismatch");
3043 address start = __ pc();
3044
3045 BLOCK_COMMENT("Entry:");
3046 __ enter();
3047
3048 #ifdef _WIN64 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
3049 const Register scale = c_rarg0; //rcx, will exchange with r9
3050 const Register objb = c_rarg1; //rdx
3051 const Register length = c_rarg2; //r8
3052 const Register obja = c_rarg3; //r9
3053 __ xchgq(obja, scale); //now obja and scale contains the correct contents
3054
3055 const Register tmp1 = r10;
3056 const Register tmp2 = r11;
3057 #endif
3058 #ifndef _WIN64 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
3059 const Register obja = c_rarg0; //U:rdi
3060 const Register objb = c_rarg1; //U:rsi
3061 const Register length = c_rarg2; //U:rdx
3062 const Register scale = c_rarg3; //U:rcx
3063 const Register tmp1 = r8;
3064 const Register tmp2 = r9;
3065 #endif
3066 const Register result = rax; //return value
3067 const XMMRegister vec0 = xmm0;
3068 const XMMRegister vec1 = xmm1;
3069 const XMMRegister vec2 = xmm2;
3070
3071 __ vectorized_mismatch(obja, objb, length, scale, result, tmp1, tmp2, vec0, vec1, vec2);
3072
3073 __ vzeroupper();
3074 __ leave();
3075 __ ret(0);
3076
3077 return start;
3078 }
3079
3080 /**
3081 * Arguments:
3082 *
3083 // Input:
3084 // c_rarg0 - x address
3085 // c_rarg1 - x length
3086 // c_rarg2 - z address
3087 // c_rarg3 - z length
3088 *
3089 */
3090 address StubGenerator::generate_squareToLen() {
3091
3092 __ align(CodeEntryAlignment);
3093 StubCodeMark mark(this, "StubRoutines", "squareToLen");
3094 address start = __ pc();
3095
3096 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
3097 // Unix: rdi, rsi, rdx, rcx (c_rarg0, c_rarg1, ...)
3098 const Register x = rdi;
3099 const Register len = rsi;
3100 const Register z = r8;
3101 const Register zlen = rcx;
3102
3103 const Register tmp1 = r12;
3104 const Register tmp2 = r13;
3105 const Register tmp3 = r14;
3106 const Register tmp4 = r15;
3107 const Register tmp5 = rbx;
3108
3109 BLOCK_COMMENT("Entry:");
3110 __ enter(); // required for proper stackwalking of RuntimeStub frame
3111
3112 setup_arg_regs(4); // x => rdi, len => rsi, z => rdx
3113 // zlen => rcx
3114 // r9 and r10 may be used to save non-volatile registers
3115 __ movptr(r8, rdx);
3116 __ square_to_len(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx, rax);
3117
3118 restore_arg_regs();
3119
3120 __ leave(); // required for proper stackwalking of RuntimeStub frame
3121 __ ret(0);
3122
3123 return start;
3124 }
3125
3126 address StubGenerator::generate_method_entry_barrier() {
3127 __ align(CodeEntryAlignment);
3128 StubCodeMark mark(this, "StubRoutines", "nmethod_entry_barrier");
3129 address start = __ pc();
3130
3131 Label deoptimize_label;
3132
3133 __ push(-1); // cookie, this is used for writing the new rsp when deoptimizing
3134
3135 BLOCK_COMMENT("Entry:");
3136 __ enter(); // save rbp
3137
3138 // save c_rarg0, because we want to use that value.
3139 // We could do without it but then we depend on the number of slots used by pusha
3140 __ push(c_rarg0);
3141
3142 __ lea(c_rarg0, Address(rsp, wordSize * 3)); // 1 for cookie, 1 for rbp, 1 for c_rarg0 - this should be the return address
3143
3144 __ pusha();
3145
3146 // The method may have floats as arguments, and we must spill them before calling
3147 // the VM runtime.
3148 assert(Argument::n_float_register_parameters_j == 8, "Assumption");
3149 const int xmm_size = wordSize * 2;
3150 const int xmm_spill_size = xmm_size * Argument::n_float_register_parameters_j;
3151 __ subptr(rsp, xmm_spill_size);
3152 __ movdqu(Address(rsp, xmm_size * 7), xmm7);
3153 __ movdqu(Address(rsp, xmm_size * 6), xmm6);
3154 __ movdqu(Address(rsp, xmm_size * 5), xmm5);
3155 __ movdqu(Address(rsp, xmm_size * 4), xmm4);
3156 __ movdqu(Address(rsp, xmm_size * 3), xmm3);
3157 __ movdqu(Address(rsp, xmm_size * 2), xmm2);
3158 __ movdqu(Address(rsp, xmm_size * 1), xmm1);
3159 __ movdqu(Address(rsp, xmm_size * 0), xmm0);
3160
3161 __ call_VM_leaf(CAST_FROM_FN_PTR(address, static_cast<int (*)(address*)>(BarrierSetNMethod::nmethod_stub_entry_barrier)), 1);
3162
3163 __ movdqu(xmm0, Address(rsp, xmm_size * 0));
3164 __ movdqu(xmm1, Address(rsp, xmm_size * 1));
3165 __ movdqu(xmm2, Address(rsp, xmm_size * 2));
3166 __ movdqu(xmm3, Address(rsp, xmm_size * 3));
3167 __ movdqu(xmm4, Address(rsp, xmm_size * 4));
3168 __ movdqu(xmm5, Address(rsp, xmm_size * 5));
3169 __ movdqu(xmm6, Address(rsp, xmm_size * 6));
3170 __ movdqu(xmm7, Address(rsp, xmm_size * 7));
3171 __ addptr(rsp, xmm_spill_size);
3172
3173 __ cmpl(rax, 1); // 1 means deoptimize
3174 __ jcc(Assembler::equal, deoptimize_label);
3175
3176 __ popa();
3177 __ pop(c_rarg0);
3178
3179 __ leave();
3180
3181 __ addptr(rsp, 1 * wordSize); // cookie
3182 __ ret(0);
3183
3184
3185 __ BIND(deoptimize_label);
3186
3187 __ popa();
3188 __ pop(c_rarg0);
3189
3190 __ leave();
3191
3192 // this can be taken out, but is good for verification purposes. getting a SIGSEGV
3193 // here while still having a correct stack is valuable
3194 __ testptr(rsp, Address(rsp, 0));
3195
3196 __ movptr(rsp, Address(rsp, 0)); // new rsp was written in the barrier
3197 __ jmp(Address(rsp, -1 * wordSize)); // jmp target should be callers verified_entry_point
3198
3199 return start;
3200 }
3201
3202 /**
3203 * Arguments:
3204 *
3205 * Input:
3206 * c_rarg0 - out address
3207 * c_rarg1 - in address
3208 * c_rarg2 - offset
3209 * c_rarg3 - len
3210 * not Win64
3211 * c_rarg4 - k
3212 * Win64
3213 * rsp+40 - k
3214 */
3215 address StubGenerator::generate_mulAdd() {
3216 __ align(CodeEntryAlignment);
3217 StubCodeMark mark(this, "StubRoutines", "mulAdd");
3218 address start = __ pc();
3219
3220 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
3221 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
3222 const Register out = rdi;
3223 const Register in = rsi;
3224 const Register offset = r11;
3225 const Register len = rcx;
3226 const Register k = r8;
3227
3228 // Next registers will be saved on stack in mul_add().
3229 const Register tmp1 = r12;
3230 const Register tmp2 = r13;
3231 const Register tmp3 = r14;
3232 const Register tmp4 = r15;
3233 const Register tmp5 = rbx;
3234
3235 BLOCK_COMMENT("Entry:");
3236 __ enter(); // required for proper stackwalking of RuntimeStub frame
3237
3238 setup_arg_regs(4); // out => rdi, in => rsi, offset => rdx
3239 // len => rcx, k => r8
3240 // r9 and r10 may be used to save non-volatile registers
3241 #ifdef _WIN64
3242 // last argument is on stack on Win64
3243 __ movl(k, Address(rsp, 6 * wordSize));
3244 #endif
3245 __ movptr(r11, rdx); // move offset in rdx to offset(r11)
3246 __ mul_add(out, in, offset, len, k, tmp1, tmp2, tmp3, tmp4, tmp5, rdx, rax);
3247
3248 restore_arg_regs();
3249
3250 __ leave(); // required for proper stackwalking of RuntimeStub frame
3251 __ ret(0);
3252
3253 return start;
3254 }
3255
3256 address StubGenerator::generate_bigIntegerRightShift() {
3257 __ align(CodeEntryAlignment);
3258 StubCodeMark mark(this, "StubRoutines", "bigIntegerRightShiftWorker");
3259 address start = __ pc();
3260
3261 Label Shift512Loop, ShiftTwo, ShiftTwoLoop, ShiftOne, Exit;
3262 // For Unix, the arguments are as follows: rdi, rsi, rdx, rcx, r8.
3263 const Register newArr = rdi;
3264 const Register oldArr = rsi;
3265 const Register newIdx = rdx;
3266 const Register shiftCount = rcx; // It was intentional to have shiftCount in rcx since it is used implicitly for shift.
3267 const Register totalNumIter = r8;
3268
3269 // For windows, we use r9 and r10 as temps to save rdi and rsi. Thus we cannot allocate them for our temps.
3270 // For everything else, we prefer using r9 and r10 since we do not have to save them before use.
3271 const Register tmp1 = r11; // Caller save.
3272 const Register tmp2 = rax; // Caller save.
3273 const Register tmp3 = WIN64_ONLY(r12) NOT_WIN64(r9); // Windows: Callee save. Linux: Caller save.
3274 const Register tmp4 = WIN64_ONLY(r13) NOT_WIN64(r10); // Windows: Callee save. Linux: Caller save.
3275 const Register tmp5 = r14; // Callee save.
3276 const Register tmp6 = r15;
3277
3278 const XMMRegister x0 = xmm0;
3279 const XMMRegister x1 = xmm1;
3280 const XMMRegister x2 = xmm2;
3281
3282 BLOCK_COMMENT("Entry:");
3283 __ enter(); // required for proper stackwalking of RuntimeStub frame
3284
3285 #ifdef _WIN64
3286 setup_arg_regs(4);
3287 // For windows, since last argument is on stack, we need to move it to the appropriate register.
3288 __ movl(totalNumIter, Address(rsp, 6 * wordSize));
3289 // Save callee save registers.
3290 __ push(tmp3);
3291 __ push(tmp4);
3292 #endif
3293 __ push(tmp5);
3294
3295 // Rename temps used throughout the code.
3296 const Register idx = tmp1;
3297 const Register nIdx = tmp2;
3298
3299 __ xorl(idx, idx);
3300
3301 // Start right shift from end of the array.
3302 // For example, if #iteration = 4 and newIdx = 1
3303 // then dest[4] = src[4] >> shiftCount | src[3] <<< (shiftCount - 32)
3304 // if #iteration = 4 and newIdx = 0
3305 // then dest[3] = src[4] >> shiftCount | src[3] <<< (shiftCount - 32)
3306 __ movl(idx, totalNumIter);
3307 __ movl(nIdx, idx);
3308 __ addl(nIdx, newIdx);
3309
3310 // If vectorization is enabled, check if the number of iterations is at least 64
3311 // If not, then go to ShifTwo processing 2 iterations
3312 if (VM_Version::supports_avx512_vbmi2()) {
3313 __ cmpptr(totalNumIter, (AVX3Threshold/64));
3314 __ jcc(Assembler::less, ShiftTwo);
3315
3316 if (AVX3Threshold < 16 * 64) {
3317 __ cmpl(totalNumIter, 16);
3318 __ jcc(Assembler::less, ShiftTwo);
3319 }
3320 __ evpbroadcastd(x0, shiftCount, Assembler::AVX_512bit);
3321 __ subl(idx, 16);
3322 __ subl(nIdx, 16);
3323 __ BIND(Shift512Loop);
3324 __ evmovdqul(x2, Address(oldArr, idx, Address::times_4, 4), Assembler::AVX_512bit);
3325 __ evmovdqul(x1, Address(oldArr, idx, Address::times_4), Assembler::AVX_512bit);
3326 __ vpshrdvd(x2, x1, x0, Assembler::AVX_512bit);
3327 __ evmovdqul(Address(newArr, nIdx, Address::times_4), x2, Assembler::AVX_512bit);
3328 __ subl(nIdx, 16);
3329 __ subl(idx, 16);
3330 __ jcc(Assembler::greaterEqual, Shift512Loop);
3331 __ addl(idx, 16);
3332 __ addl(nIdx, 16);
3333 }
3334 __ BIND(ShiftTwo);
3335 __ cmpl(idx, 2);
3336 __ jcc(Assembler::less, ShiftOne);
3337 __ subl(idx, 2);
3338 __ subl(nIdx, 2);
3339 __ BIND(ShiftTwoLoop);
3340 __ movl(tmp5, Address(oldArr, idx, Address::times_4, 8));
3341 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 4));
3342 __ movl(tmp3, Address(oldArr, idx, Address::times_4));
3343 __ shrdl(tmp5, tmp4);
3344 __ shrdl(tmp4, tmp3);
3345 __ movl(Address(newArr, nIdx, Address::times_4, 4), tmp5);
3346 __ movl(Address(newArr, nIdx, Address::times_4), tmp4);
3347 __ subl(nIdx, 2);
3348 __ subl(idx, 2);
3349 __ jcc(Assembler::greaterEqual, ShiftTwoLoop);
3350 __ addl(idx, 2);
3351 __ addl(nIdx, 2);
3352
3353 // Do the last iteration
3354 __ BIND(ShiftOne);
3355 __ cmpl(idx, 1);
3356 __ jcc(Assembler::less, Exit);
3357 __ subl(idx, 1);
3358 __ subl(nIdx, 1);
3359 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 4));
3360 __ movl(tmp3, Address(oldArr, idx, Address::times_4));
3361 __ shrdl(tmp4, tmp3);
3362 __ movl(Address(newArr, nIdx, Address::times_4), tmp4);
3363 __ BIND(Exit);
3364 __ vzeroupper();
3365 // Restore callee save registers.
3366 __ pop(tmp5);
3367 #ifdef _WIN64
3368 __ pop(tmp4);
3369 __ pop(tmp3);
3370 restore_arg_regs();
3371 #endif
3372 __ leave(); // required for proper stackwalking of RuntimeStub frame
3373 __ ret(0);
3374
3375 return start;
3376 }
3377
3378 /**
3379 * Arguments:
3380 *
3381 * Input:
3382 * c_rarg0 - newArr address
3383 * c_rarg1 - oldArr address
3384 * c_rarg2 - newIdx
3385 * c_rarg3 - shiftCount
3386 * not Win64
3387 * c_rarg4 - numIter
3388 * Win64
3389 * rsp40 - numIter
3390 */
3391 address StubGenerator::generate_bigIntegerLeftShift() {
3392 __ align(CodeEntryAlignment);
3393 StubCodeMark mark(this, "StubRoutines", "bigIntegerLeftShiftWorker");
3394 address start = __ pc();
3395
3396 Label Shift512Loop, ShiftTwo, ShiftTwoLoop, ShiftOne, Exit;
3397 // For Unix, the arguments are as follows: rdi, rsi, rdx, rcx, r8.
3398 const Register newArr = rdi;
3399 const Register oldArr = rsi;
3400 const Register newIdx = rdx;
3401 const Register shiftCount = rcx; // It was intentional to have shiftCount in rcx since it is used implicitly for shift.
3402 const Register totalNumIter = r8;
3403 // For windows, we use r9 and r10 as temps to save rdi and rsi. Thus we cannot allocate them for our temps.
3404 // For everything else, we prefer using r9 and r10 since we do not have to save them before use.
3405 const Register tmp1 = r11; // Caller save.
3406 const Register tmp2 = rax; // Caller save.
3407 const Register tmp3 = WIN64_ONLY(r12) NOT_WIN64(r9); // Windows: Callee save. Linux: Caller save.
3408 const Register tmp4 = WIN64_ONLY(r13) NOT_WIN64(r10); // Windows: Callee save. Linux: Caller save.
3409 const Register tmp5 = r14; // Callee save.
3410
3411 const XMMRegister x0 = xmm0;
3412 const XMMRegister x1 = xmm1;
3413 const XMMRegister x2 = xmm2;
3414 BLOCK_COMMENT("Entry:");
3415 __ enter(); // required for proper stackwalking of RuntimeStub frame
3416
3417 #ifdef _WIN64
3418 setup_arg_regs(4);
3419 // For windows, since last argument is on stack, we need to move it to the appropriate register.
3420 __ movl(totalNumIter, Address(rsp, 6 * wordSize));
3421 // Save callee save registers.
3422 __ push(tmp3);
3423 __ push(tmp4);
3424 #endif
3425 __ push(tmp5);
3426
3427 // Rename temps used throughout the code
3428 const Register idx = tmp1;
3429 const Register numIterTmp = tmp2;
3430
3431 // Start idx from zero.
3432 __ xorl(idx, idx);
3433 // Compute interior pointer for new array. We do this so that we can use same index for both old and new arrays.
3434 __ lea(newArr, Address(newArr, newIdx, Address::times_4));
3435 __ movl(numIterTmp, totalNumIter);
3436
3437 // If vectorization is enabled, check if the number of iterations is at least 64
3438 // If not, then go to ShiftTwo shifting two numbers at a time
3439 if (VM_Version::supports_avx512_vbmi2()) {
3440 __ cmpl(totalNumIter, (AVX3Threshold/64));
3441 __ jcc(Assembler::less, ShiftTwo);
3442
3443 if (AVX3Threshold < 16 * 64) {
3444 __ cmpl(totalNumIter, 16);
3445 __ jcc(Assembler::less, ShiftTwo);
3446 }
3447 __ evpbroadcastd(x0, shiftCount, Assembler::AVX_512bit);
3448 __ subl(numIterTmp, 16);
3449 __ BIND(Shift512Loop);
3450 __ evmovdqul(x1, Address(oldArr, idx, Address::times_4), Assembler::AVX_512bit);
3451 __ evmovdqul(x2, Address(oldArr, idx, Address::times_4, 0x4), Assembler::AVX_512bit);
3452 __ vpshldvd(x1, x2, x0, Assembler::AVX_512bit);
3453 __ evmovdqul(Address(newArr, idx, Address::times_4), x1, Assembler::AVX_512bit);
3454 __ addl(idx, 16);
3455 __ subl(numIterTmp, 16);
3456 __ jcc(Assembler::greaterEqual, Shift512Loop);
3457 __ addl(numIterTmp, 16);
3458 }
3459 __ BIND(ShiftTwo);
3460 __ cmpl(totalNumIter, 1);
3461 __ jcc(Assembler::less, Exit);
3462 __ movl(tmp3, Address(oldArr, idx, Address::times_4));
3463 __ subl(numIterTmp, 2);
3464 __ jcc(Assembler::less, ShiftOne);
3465
3466 __ BIND(ShiftTwoLoop);
3467 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 0x4));
3468 __ movl(tmp5, Address(oldArr, idx, Address::times_4, 0x8));
3469 __ shldl(tmp3, tmp4);
3470 __ shldl(tmp4, tmp5);
3471 __ movl(Address(newArr, idx, Address::times_4), tmp3);
3472 __ movl(Address(newArr, idx, Address::times_4, 0x4), tmp4);
3473 __ movl(tmp3, tmp5);
3474 __ addl(idx, 2);
3475 __ subl(numIterTmp, 2);
3476 __ jcc(Assembler::greaterEqual, ShiftTwoLoop);
3477
3478 // Do the last iteration
3479 __ BIND(ShiftOne);
3480 __ addl(numIterTmp, 2);
3481 __ cmpl(numIterTmp, 1);
3482 __ jcc(Assembler::less, Exit);
3483 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 0x4));
3484 __ shldl(tmp3, tmp4);
3485 __ movl(Address(newArr, idx, Address::times_4), tmp3);
3486
3487 __ BIND(Exit);
3488 __ vzeroupper();
3489 // Restore callee save registers.
3490 __ pop(tmp5);
3491 #ifdef _WIN64
3492 __ pop(tmp4);
3493 __ pop(tmp3);
3494 restore_arg_regs();
3495 #endif
3496 __ leave(); // required for proper stackwalking of RuntimeStub frame
3497 __ ret(0);
3498
3499 return start;
3500 }
3501
3502 void StubGenerator::generate_libm_stubs() {
3503 if (UseLibmIntrinsic && InlineIntrinsics) {
3504 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dsin)) {
3505 StubRoutines::_dsin = generate_libmSin(); // from stubGenerator_x86_64_sin.cpp
3506 }
3507 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dcos)) {
3508 StubRoutines::_dcos = generate_libmCos(); // from stubGenerator_x86_64_cos.cpp
3509 }
3510 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dtan)) {
3511 StubRoutines::_dtan = generate_libmTan(); // from stubGenerator_x86_64_tan.cpp
3512 }
3513 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dexp)) {
3514 StubRoutines::_dexp = generate_libmExp(); // from stubGenerator_x86_64_exp.cpp
3515 }
3516 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dpow)) {
3517 StubRoutines::_dpow = generate_libmPow(); // from stubGenerator_x86_64_pow.cpp
3518 }
3519 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dlog)) {
3520 StubRoutines::_dlog = generate_libmLog(); // from stubGenerator_x86_64_log.cpp
3521 }
3522 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dlog10)) {
3523 StubRoutines::_dlog10 = generate_libmLog10(); // from stubGenerator_x86_64_log.cpp
3524 }
3525 }
3526 }
3527
3528 /**
3529 * Arguments:
3530 *
3531 * Input:
3532 * c_rarg0 - float16 jshort
3533 *
3534 * Output:
3535 * xmm0 - float
3536 */
3537 address StubGenerator::generate_float16ToFloat() {
3538 StubCodeMark mark(this, "StubRoutines", "float16ToFloat");
3539
3540 address start = __ pc();
3541
3542 BLOCK_COMMENT("Entry:");
3543 // No need for RuntimeStub frame since it is called only during JIT compilation
3544
3545 // Load value into xmm0 and convert
3546 __ flt16_to_flt(xmm0, c_rarg0);
3547
3548 __ ret(0);
3549
3550 return start;
3551 }
3552
3553 /**
3554 * Arguments:
3555 *
3556 * Input:
3557 * xmm0 - float
3558 *
3559 * Output:
3560 * rax - float16 jshort
3561 */
3562 address StubGenerator::generate_floatToFloat16() {
3563 StubCodeMark mark(this, "StubRoutines", "floatToFloat16");
3564
3565 address start = __ pc();
3566
3567 BLOCK_COMMENT("Entry:");
3568 // No need for RuntimeStub frame since it is called only during JIT compilation
3569
3570 // Convert and put result into rax
3571 __ flt_to_flt16(rax, xmm0, xmm1);
3572
3573 __ ret(0);
3574
3575 return start;
3576 }
3577
3578 address StubGenerator::generate_cont_thaw(const char* label, Continuation::thaw_kind kind) {
3579 if (!Continuations::enabled()) return nullptr;
3580
3581 bool return_barrier = Continuation::is_thaw_return_barrier(kind);
3582 bool return_barrier_exception = Continuation::is_thaw_return_barrier_exception(kind);
3583
3584 StubCodeMark mark(this, "StubRoutines", label);
3585 address start = __ pc();
3586
3587 // TODO: Handle Valhalla return types. May require generating different return barriers.
3588
3589 if (!return_barrier) {
3590 // Pop return address. If we don't do this, we get a drift,
3591 // where the bottom-most frozen frame continuously grows.
3592 __ pop(c_rarg3);
3593 } else {
3594 __ movptr(rsp, Address(r15_thread, JavaThread::cont_entry_offset()));
3595 }
3596
3597 #ifdef ASSERT
3598 {
3599 Label L_good_sp;
3600 __ cmpptr(rsp, Address(r15_thread, JavaThread::cont_entry_offset()));
3601 __ jcc(Assembler::equal, L_good_sp);
3602 __ stop("Incorrect rsp at thaw entry");
3603 __ BIND(L_good_sp);
3604 }
3605 #endif // ASSERT
3606
3607 if (return_barrier) {
3608 // Preserve possible return value from a method returning to the return barrier.
3609 __ push(rax);
3610 __ push_d(xmm0);
3611 }
3612
3613 __ movptr(c_rarg0, r15_thread);
3614 __ movptr(c_rarg1, (return_barrier ? 1 : 0));
3615 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Continuation::prepare_thaw), 2);
3616 __ movptr(rbx, rax);
3617
3618 if (return_barrier) {
3619 // Restore return value from a method returning to the return barrier.
3620 // No safepoint in the call to thaw, so even an oop return value should be OK.
3621 __ pop_d(xmm0);
3622 __ pop(rax);
3623 }
3624
3625 #ifdef ASSERT
3626 {
3627 Label L_good_sp;
3628 __ cmpptr(rsp, Address(r15_thread, JavaThread::cont_entry_offset()));
3629 __ jcc(Assembler::equal, L_good_sp);
3630 __ stop("Incorrect rsp after prepare thaw");
3631 __ BIND(L_good_sp);
3632 }
3633 #endif // ASSERT
3634
3635 // rbx contains the size of the frames to thaw, 0 if overflow or no more frames
3636 Label L_thaw_success;
3637 __ testptr(rbx, rbx);
3638 __ jccb(Assembler::notZero, L_thaw_success);
3639 __ jump(ExternalAddress(StubRoutines::throw_StackOverflowError_entry()));
3640 __ bind(L_thaw_success);
3641
3642 // Make room for the thawed frames and align the stack.
3643 __ subptr(rsp, rbx);
3644 __ andptr(rsp, -StackAlignmentInBytes);
3645
3646 if (return_barrier) {
3647 // Preserve possible return value from a method returning to the return barrier. (Again.)
3648 __ push(rax);
3649 __ push_d(xmm0);
3650 }
3651
3652 // If we want, we can templatize thaw by kind, and have three different entries.
3653 __ movptr(c_rarg0, r15_thread);
3654 __ movptr(c_rarg1, kind);
3655 __ call_VM_leaf(Continuation::thaw_entry(), 2);
3656 __ movptr(rbx, rax);
3657
3658 if (return_barrier) {
3659 // Restore return value from a method returning to the return barrier. (Again.)
3660 // No safepoint in the call to thaw, so even an oop return value should be OK.
3661 __ pop_d(xmm0);
3662 __ pop(rax);
3663 } else {
3664 // Return 0 (success) from doYield.
3665 __ xorptr(rax, rax);
3666 }
3667
3668 // After thawing, rbx is the SP of the yielding frame.
3669 // Move there, and then to saved RBP slot.
3670 __ movptr(rsp, rbx);
3671 __ subptr(rsp, 2*wordSize);
3672
3673 if (return_barrier_exception) {
3674 __ movptr(c_rarg0, r15_thread);
3675 __ movptr(c_rarg1, Address(rsp, wordSize)); // return address
3676
3677 // rax still holds the original exception oop, save it before the call
3678 __ push(rax);
3679
3680 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), 2);
3681 __ movptr(rbx, rax);
3682
3683 // Continue at exception handler:
3684 // rax: exception oop
3685 // rbx: exception handler
3686 // rdx: exception pc
3687 __ pop(rax);
3688 __ verify_oop(rax);
3689 __ pop(rbp); // pop out RBP here too
3690 __ pop(rdx);
3691 __ jmp(rbx);
3692 } else {
3693 // We are "returning" into the topmost thawed frame; see Thaw::push_return_frame
3694 __ pop(rbp);
3695 __ ret(0);
3696 }
3697
3698 return start;
3699 }
3700
3701 address StubGenerator::generate_cont_thaw() {
3702 return generate_cont_thaw("Cont thaw", Continuation::thaw_top);
3703 }
3704
3705 // TODO: will probably need multiple return barriers depending on return type
3706
3707 address StubGenerator::generate_cont_returnBarrier() {
3708 return generate_cont_thaw("Cont thaw return barrier", Continuation::thaw_return_barrier);
3709 }
3710
3711 address StubGenerator::generate_cont_returnBarrier_exception() {
3712 return generate_cont_thaw("Cont thaw return barrier exception", Continuation::thaw_return_barrier_exception);
3713 }
3714
3715 #if INCLUDE_JFR
3716
3717 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
3718 // It returns a jobject handle to the event writer.
3719 // The handle is dereferenced and the return value is the event writer oop.
3720 RuntimeStub* StubGenerator::generate_jfr_write_checkpoint() {
3721 enum layout {
3722 rbp_off,
3723 rbpH_off,
3724 return_off,
3725 return_off2,
3726 framesize // inclusive of return address
3727 };
3728
3729 CodeBuffer code("jfr_write_checkpoint", 1024, 64);
3730 MacroAssembler* _masm = new MacroAssembler(&code);
3731 address start = __ pc();
3732
3733 __ enter();
3734 address the_pc = __ pc();
3735
3736 int frame_complete = the_pc - start;
3737
3738 __ set_last_Java_frame(rsp, rbp, the_pc, rscratch1);
3739 __ movptr(c_rarg0, r15_thread);
3740 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), 1);
3741 __ reset_last_Java_frame(true);
3742
3743 // rax is jobject handle result, unpack and process it through a barrier.
3744 __ resolve_global_jobject(rax, r15_thread, c_rarg0);
3745
3746 __ leave();
3747 __ ret(0);
3748
3749 OopMapSet* oop_maps = new OopMapSet();
3750 OopMap* map = new OopMap(framesize, 1);
3751 oop_maps->add_gc_map(frame_complete, map);
3752
3753 RuntimeStub* stub =
3754 RuntimeStub::new_runtime_stub(code.name(),
3755 &code,
3756 frame_complete,
3757 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3758 oop_maps,
3759 false);
3760 return stub;
3761 }
3762
3763 // For c2: call to return a leased buffer.
3764 RuntimeStub* StubGenerator::generate_jfr_return_lease() {
3765 enum layout {
3766 rbp_off,
3767 rbpH_off,
3768 return_off,
3769 return_off2,
3770 framesize // inclusive of return address
3771 };
3772
3773 CodeBuffer code("jfr_return_lease", 1024, 64);
3774 MacroAssembler* _masm = new MacroAssembler(&code);
3775 address start = __ pc();
3776
3777 __ enter();
3778 address the_pc = __ pc();
3779
3780 int frame_complete = the_pc - start;
3781
3782 __ set_last_Java_frame(rsp, rbp, the_pc, rscratch2);
3783 __ movptr(c_rarg0, r15_thread);
3784 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::return_lease), 1);
3785 __ reset_last_Java_frame(true);
3786
3787 __ leave();
3788 __ ret(0);
3789
3790 OopMapSet* oop_maps = new OopMapSet();
3791 OopMap* map = new OopMap(framesize, 1);
3792 oop_maps->add_gc_map(frame_complete, map);
3793
3794 RuntimeStub* stub =
3795 RuntimeStub::new_runtime_stub(code.name(),
3796 &code,
3797 frame_complete,
3798 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3799 oop_maps,
3800 false);
3801 return stub;
3802 }
3803
3804 #endif // INCLUDE_JFR
3805
3806 // Continuation point for throwing of implicit exceptions that are
3807 // not handled in the current activation. Fabricates an exception
3808 // oop and initiates normal exception dispatching in this
3809 // frame. Since we need to preserve callee-saved values (currently
3810 // only for C2, but done for C1 as well) we need a callee-saved oop
3811 // map and therefore have to make these stubs into RuntimeStubs
3812 // rather than BufferBlobs. If the compiler needs all registers to
3813 // be preserved between the fault point and the exception handler
3814 // then it must assume responsibility for that in
3815 // AbstractCompiler::continuation_for_implicit_null_exception or
3816 // continuation_for_implicit_division_by_zero_exception. All other
3817 // implicit exceptions (e.g., NullPointerException or
3818 // AbstractMethodError on entry) are either at call sites or
3819 // otherwise assume that stack unwinding will be initiated, so
3820 // caller saved registers were assumed volatile in the compiler.
3821 address StubGenerator::generate_throw_exception(const char* name,
3822 address runtime_entry,
3823 Register arg1,
3824 Register arg2) {
3825 // Information about frame layout at time of blocking runtime call.
3826 // Note that we only have to preserve callee-saved registers since
3827 // the compilers are responsible for supplying a continuation point
3828 // if they expect all registers to be preserved.
3829 enum layout {
3830 rbp_off = frame::arg_reg_save_area_bytes/BytesPerInt,
3831 rbp_off2,
3832 return_off,
3833 return_off2,
3834 framesize // inclusive of return address
3835 };
3836
3837 int insts_size = 512;
3838 int locs_size = 64;
3839
3840 CodeBuffer code(name, insts_size, locs_size);
3841 OopMapSet* oop_maps = new OopMapSet();
3842 MacroAssembler* _masm = new MacroAssembler(&code);
3843
3844 address start = __ pc();
3845
3846 // This is an inlined and slightly modified version of call_VM
3847 // which has the ability to fetch the return PC out of
3848 // thread-local storage and also sets up last_Java_sp slightly
3849 // differently than the real call_VM
3850
3851 __ enter(); // required for proper stackwalking of RuntimeStub frame
3852
3853 assert(is_even(framesize/2), "sp not 16-byte aligned");
3854
3855 // return address and rbp are already in place
3856 __ subptr(rsp, (framesize-4) << LogBytesPerInt); // prolog
3857
3858 int frame_complete = __ pc() - start;
3859
3860 // Set up last_Java_sp and last_Java_fp
3861 address the_pc = __ pc();
3862 __ set_last_Java_frame(rsp, rbp, the_pc, rscratch1);
3863 __ andptr(rsp, -(StackAlignmentInBytes)); // Align stack
3864
3865 // Call runtime
3866 if (arg1 != noreg) {
3867 assert(arg2 != c_rarg1, "clobbered");
3868 __ movptr(c_rarg1, arg1);
3869 }
3870 if (arg2 != noreg) {
3871 __ movptr(c_rarg2, arg2);
3872 }
3873 __ movptr(c_rarg0, r15_thread);
3874 BLOCK_COMMENT("call runtime_entry");
3875 __ call(RuntimeAddress(runtime_entry));
3876
3877 // Generate oop map
3878 OopMap* map = new OopMap(framesize, 0);
3879
3880 oop_maps->add_gc_map(the_pc - start, map);
3881
3882 __ reset_last_Java_frame(true);
3883
3884 __ leave(); // required for proper stackwalking of RuntimeStub frame
3885
3886 // check for pending exceptions
3887 #ifdef ASSERT
3888 Label L;
3889 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), NULL_WORD);
3890 __ jcc(Assembler::notEqual, L);
3891 __ should_not_reach_here();
3892 __ bind(L);
3893 #endif // ASSERT
3894 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3895
3896
3897 // codeBlob framesize is in words (not VMRegImpl::slot_size)
3898 RuntimeStub* stub =
3899 RuntimeStub::new_runtime_stub(name,
3900 &code,
3901 frame_complete,
3902 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3903 oop_maps, false);
3904 return stub->entry_point();
3905 }
3906
3907 // exception handler for upcall stubs
3908 address StubGenerator::generate_upcall_stub_exception_handler() {
3909 StubCodeMark mark(this, "StubRoutines", "upcall stub exception handler");
3910 address start = __ pc();
3911
3912 // native caller has no idea how to handle exceptions
3913 // we just crash here. Up to callee to catch exceptions.
3914 __ verify_oop(rax);
3915 __ vzeroupper();
3916 __ mov(c_rarg0, rax);
3917 __ andptr(rsp, -StackAlignmentInBytes); // align stack as required by ABI
3918 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows
3919 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, UpcallLinker::handle_uncaught_exception)));
3920 __ should_not_reach_here();
3921
3922 return start;
3923 }
3924
3925 void StubGenerator::create_control_words() {
3926 // Round to nearest, 64-bit mode, exceptions masked, flags specialized
3927 StubRoutines::x86::_mxcsr_std = EnableX86ECoreOpts ? 0x1FBF : 0x1F80;
3928 // Round to zero, 64-bit mode, exceptions masked, flags specialized
3929 StubRoutines::x86::_mxcsr_rz = EnableX86ECoreOpts ? 0x7FBF : 0x7F80;
3930 }
3931
3932 // Initialization
3933 void StubGenerator::generate_initial_stubs() {
3934 // Generates all stubs and initializes the entry points
3935
3936 // This platform-specific settings are needed by generate_call_stub()
3937 create_control_words();
3938
3939 // Initialize table for unsafe copy memeory check.
3940 if (UnsafeCopyMemory::_table == nullptr) {
3941 UnsafeCopyMemory::create_table(16);
3942 }
3943
3944 // entry points that exist in all platforms Note: This is code
3945 // that could be shared among different platforms - however the
3946 // benefit seems to be smaller than the disadvantage of having a
3947 // much more complicated generator structure. See also comment in
3948 // stubRoutines.hpp.
3949
3950 StubRoutines::_forward_exception_entry = generate_forward_exception();
3951
3952 // Generate these first because they are called from other stubs
3953 if (InlineTypeReturnedAsFields) {
3954 StubRoutines::_load_inline_type_fields_in_regs =
3955 generate_return_value_stub(CAST_FROM_FN_PTR(address, SharedRuntime::load_inline_type_fields_in_regs),
3956 "load_inline_type_fields_in_regs", false);
3957 StubRoutines::_store_inline_type_fields_to_buf =
3958 generate_return_value_stub(CAST_FROM_FN_PTR(address, SharedRuntime::store_inline_type_fields_to_buf),
3959 "store_inline_type_fields_to_buf", true);
3960 }
3961
3962 StubRoutines::_call_stub_entry =
3963 generate_call_stub(StubRoutines::_call_stub_return_address);
3964
3965 // is referenced by megamorphic call
3966 StubRoutines::_catch_exception_entry = generate_catch_exception();
3967
3968 // atomic calls
3969 StubRoutines::_fence_entry = generate_orderaccess_fence();
3970
3971 // platform dependent
3972 StubRoutines::x86::_get_previous_sp_entry = generate_get_previous_sp();
3973
3974 StubRoutines::x86::_verify_mxcsr_entry = generate_verify_mxcsr();
3975
3976 StubRoutines::x86::_f2i_fixup = generate_f2i_fixup();
3977 StubRoutines::x86::_f2l_fixup = generate_f2l_fixup();
3978 StubRoutines::x86::_d2i_fixup = generate_d2i_fixup();
3979 StubRoutines::x86::_d2l_fixup = generate_d2l_fixup();
3980
3981 StubRoutines::x86::_float_sign_mask = generate_fp_mask("float_sign_mask", 0x7FFFFFFF7FFFFFFF);
3982 StubRoutines::x86::_float_sign_flip = generate_fp_mask("float_sign_flip", 0x8000000080000000);
3983 StubRoutines::x86::_double_sign_mask = generate_fp_mask("double_sign_mask", 0x7FFFFFFFFFFFFFFF);
3984 StubRoutines::x86::_double_sign_flip = generate_fp_mask("double_sign_flip", 0x8000000000000000);
3985
3986 // Build this early so it's available for the interpreter.
3987 StubRoutines::_throw_StackOverflowError_entry =
3988 generate_throw_exception("StackOverflowError throw_exception",
3989 CAST_FROM_FN_PTR(address,
3990 SharedRuntime::
3991 throw_StackOverflowError));
3992 StubRoutines::_throw_delayed_StackOverflowError_entry =
3993 generate_throw_exception("delayed StackOverflowError throw_exception",
3994 CAST_FROM_FN_PTR(address,
3995 SharedRuntime::
3996 throw_delayed_StackOverflowError));
3997 if (UseCRC32Intrinsics) {
3998 // set table address before stub generation which use it
3999 StubRoutines::_crc_table_adr = (address)StubRoutines::x86::_crc_table;
4000 StubRoutines::_updateBytesCRC32 = generate_updateBytesCRC32();
4001 }
4002
4003 if (UseCRC32CIntrinsics) {
4004 bool supports_clmul = VM_Version::supports_clmul();
4005 StubRoutines::x86::generate_CRC32C_table(supports_clmul);
4006 StubRoutines::_crc32c_table_addr = (address)StubRoutines::x86::_crc32c_table;
4007 StubRoutines::_updateBytesCRC32C = generate_updateBytesCRC32C(supports_clmul);
4008 }
4009
4010 if (VM_Version::supports_float16()) {
4011 // For results consistency both intrinsics should be enabled.
4012 // vmIntrinsics checks InlineIntrinsics flag, no need to check it here.
4013 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_float16ToFloat) &&
4014 vmIntrinsics::is_intrinsic_available(vmIntrinsics::_floatToFloat16)) {
4015 StubRoutines::_hf2f = generate_float16ToFloat();
4016 StubRoutines::_f2hf = generate_floatToFloat16();
4017 }
4018 }
4019
4020 generate_libm_stubs();
4021
4022 StubRoutines::_fmod = generate_libmFmod(); // from stubGenerator_x86_64_fmod.cpp
4023 }
4024
4025 // Call here from the interpreter or compiled code to either load
4026 // multiple returned values from the inline type instance being
4027 // returned to registers or to store returned values to a newly
4028 // allocated inline type instance.
4029 // Register is a class, but it would be assigned numerical value.
4030 // "0" is assigned for xmm0. Thus we need to ignore -Wnonnull.
4031 PRAGMA_DIAG_PUSH
4032 PRAGMA_NONNULL_IGNORED
4033 address StubGenerator::generate_return_value_stub(address destination, const char* name, bool has_res) {
4034 // We need to save all registers the calling convention may use so
4035 // the runtime calls read or update those registers. This needs to
4036 // be in sync with SharedRuntime::java_return_convention().
4037 enum layout {
4038 pad_off = frame::arg_reg_save_area_bytes/BytesPerInt, pad_off_2,
4039 rax_off, rax_off_2,
4040 j_rarg5_off, j_rarg5_2,
4041 j_rarg4_off, j_rarg4_2,
4042 j_rarg3_off, j_rarg3_2,
4043 j_rarg2_off, j_rarg2_2,
4044 j_rarg1_off, j_rarg1_2,
4045 j_rarg0_off, j_rarg0_2,
4046 j_farg0_off, j_farg0_2,
4047 j_farg1_off, j_farg1_2,
4048 j_farg2_off, j_farg2_2,
4049 j_farg3_off, j_farg3_2,
4050 j_farg4_off, j_farg4_2,
4051 j_farg5_off, j_farg5_2,
4052 j_farg6_off, j_farg6_2,
4053 j_farg7_off, j_farg7_2,
4054 rbp_off, rbp_off_2,
4055 return_off, return_off_2,
4056
4057 framesize
4058 };
4059
4060 CodeBuffer buffer(name, 1000, 512);
4061 MacroAssembler* _masm = new MacroAssembler(&buffer);
4062
4063 int frame_size_in_bytes = align_up(framesize*BytesPerInt, 16);
4064 assert(frame_size_in_bytes == framesize*BytesPerInt, "misaligned");
4065 int frame_size_in_slots = frame_size_in_bytes / BytesPerInt;
4066 int frame_size_in_words = frame_size_in_bytes / wordSize;
4067
4068 OopMapSet *oop_maps = new OopMapSet();
4069 OopMap* map = new OopMap(frame_size_in_slots, 0);
4070
4071 map->set_callee_saved(VMRegImpl::stack2reg(rax_off), rax->as_VMReg());
4072 map->set_callee_saved(VMRegImpl::stack2reg(j_rarg5_off), j_rarg5->as_VMReg());
4073 map->set_callee_saved(VMRegImpl::stack2reg(j_rarg4_off), j_rarg4->as_VMReg());
4074 map->set_callee_saved(VMRegImpl::stack2reg(j_rarg3_off), j_rarg3->as_VMReg());
4075 map->set_callee_saved(VMRegImpl::stack2reg(j_rarg2_off), j_rarg2->as_VMReg());
4076 map->set_callee_saved(VMRegImpl::stack2reg(j_rarg1_off), j_rarg1->as_VMReg());
4077 map->set_callee_saved(VMRegImpl::stack2reg(j_rarg0_off), j_rarg0->as_VMReg());
4078 map->set_callee_saved(VMRegImpl::stack2reg(j_farg0_off), j_farg0->as_VMReg());
4079 map->set_callee_saved(VMRegImpl::stack2reg(j_farg1_off), j_farg1->as_VMReg());
4080 map->set_callee_saved(VMRegImpl::stack2reg(j_farg2_off), j_farg2->as_VMReg());
4081 map->set_callee_saved(VMRegImpl::stack2reg(j_farg3_off), j_farg3->as_VMReg());
4082 map->set_callee_saved(VMRegImpl::stack2reg(j_farg4_off), j_farg4->as_VMReg());
4083 map->set_callee_saved(VMRegImpl::stack2reg(j_farg5_off), j_farg5->as_VMReg());
4084 map->set_callee_saved(VMRegImpl::stack2reg(j_farg6_off), j_farg6->as_VMReg());
4085 map->set_callee_saved(VMRegImpl::stack2reg(j_farg7_off), j_farg7->as_VMReg());
4086
4087 int start = __ offset();
4088
4089 __ subptr(rsp, frame_size_in_bytes - 8 /* return address*/);
4090
4091 __ movptr(Address(rsp, rbp_off * BytesPerInt), rbp);
4092 __ movdbl(Address(rsp, j_farg7_off * BytesPerInt), j_farg7);
4093 __ movdbl(Address(rsp, j_farg6_off * BytesPerInt), j_farg6);
4094 __ movdbl(Address(rsp, j_farg5_off * BytesPerInt), j_farg5);
4095 __ movdbl(Address(rsp, j_farg4_off * BytesPerInt), j_farg4);
4096 __ movdbl(Address(rsp, j_farg3_off * BytesPerInt), j_farg3);
4097 __ movdbl(Address(rsp, j_farg2_off * BytesPerInt), j_farg2);
4098 __ movdbl(Address(rsp, j_farg1_off * BytesPerInt), j_farg1);
4099 __ movdbl(Address(rsp, j_farg0_off * BytesPerInt), j_farg0);
4100
4101 __ movptr(Address(rsp, j_rarg0_off * BytesPerInt), j_rarg0);
4102 __ movptr(Address(rsp, j_rarg1_off * BytesPerInt), j_rarg1);
4103 __ movptr(Address(rsp, j_rarg2_off * BytesPerInt), j_rarg2);
4104 __ movptr(Address(rsp, j_rarg3_off * BytesPerInt), j_rarg3);
4105 __ movptr(Address(rsp, j_rarg4_off * BytesPerInt), j_rarg4);
4106 __ movptr(Address(rsp, j_rarg5_off * BytesPerInt), j_rarg5);
4107 __ movptr(Address(rsp, rax_off * BytesPerInt), rax);
4108
4109 int frame_complete = __ offset();
4110
4111 __ set_last_Java_frame(noreg, noreg, nullptr, rscratch1);
4112
4113 __ mov(c_rarg0, r15_thread);
4114 __ mov(c_rarg1, rax);
4115
4116 __ call(RuntimeAddress(destination));
4117
4118 // Set an oopmap for the call site.
4119
4120 oop_maps->add_gc_map( __ offset() - start, map);
4121
4122 // clear last_Java_sp
4123 __ reset_last_Java_frame(false);
4124
4125 __ movptr(rbp, Address(rsp, rbp_off * BytesPerInt));
4126 __ movdbl(j_farg7, Address(rsp, j_farg7_off * BytesPerInt));
4127 __ movdbl(j_farg6, Address(rsp, j_farg6_off * BytesPerInt));
4128 __ movdbl(j_farg5, Address(rsp, j_farg5_off * BytesPerInt));
4129 __ movdbl(j_farg4, Address(rsp, j_farg4_off * BytesPerInt));
4130 __ movdbl(j_farg3, Address(rsp, j_farg3_off * BytesPerInt));
4131 __ movdbl(j_farg2, Address(rsp, j_farg2_off * BytesPerInt));
4132 __ movdbl(j_farg1, Address(rsp, j_farg1_off * BytesPerInt));
4133 __ movdbl(j_farg0, Address(rsp, j_farg0_off * BytesPerInt));
4134
4135 __ movptr(j_rarg0, Address(rsp, j_rarg0_off * BytesPerInt));
4136 __ movptr(j_rarg1, Address(rsp, j_rarg1_off * BytesPerInt));
4137 __ movptr(j_rarg2, Address(rsp, j_rarg2_off * BytesPerInt));
4138 __ movptr(j_rarg3, Address(rsp, j_rarg3_off * BytesPerInt));
4139 __ movptr(j_rarg4, Address(rsp, j_rarg4_off * BytesPerInt));
4140 __ movptr(j_rarg5, Address(rsp, j_rarg5_off * BytesPerInt));
4141 __ movptr(rax, Address(rsp, rax_off * BytesPerInt));
4142
4143 __ addptr(rsp, frame_size_in_bytes-8);
4144
4145 // check for pending exceptions
4146 Label pending;
4147 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
4148 __ jcc(Assembler::notEqual, pending);
4149
4150 if (has_res) {
4151 __ get_vm_result(rax, r15_thread);
4152 }
4153
4154 __ ret(0);
4155
4156 __ bind(pending);
4157
4158 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset()));
4159 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
4160
4161 // -------------
4162 // make sure all code is generated
4163 _masm->flush();
4164
4165 RuntimeStub* stub = RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, false);
4166 return stub->entry_point();
4167 }
4168
4169 void StubGenerator::generate_continuation_stubs() {
4170 // Continuation stubs:
4171 StubRoutines::_cont_thaw = generate_cont_thaw();
4172 StubRoutines::_cont_returnBarrier = generate_cont_returnBarrier();
4173 StubRoutines::_cont_returnBarrierExc = generate_cont_returnBarrier_exception();
4174
4175 JFR_ONLY(generate_jfr_stubs();)
4176 }
4177
4178 #if INCLUDE_JFR
4179 void StubGenerator::generate_jfr_stubs() {
4180 StubRoutines::_jfr_write_checkpoint_stub = generate_jfr_write_checkpoint();
4181 StubRoutines::_jfr_write_checkpoint = StubRoutines::_jfr_write_checkpoint_stub->entry_point();
4182 StubRoutines::_jfr_return_lease_stub = generate_jfr_return_lease();
4183 StubRoutines::_jfr_return_lease = StubRoutines::_jfr_return_lease_stub->entry_point();
4184 }
4185 #endif
4186
4187 void StubGenerator::generate_final_stubs() {
4188 // Generates the rest of stubs and initializes the entry points
4189
4190 // These entry points require SharedInfo::stack0 to be set up in
4191 // non-core builds and need to be relocatable, so they each
4192 // fabricate a RuntimeStub internally.
4193 StubRoutines::_throw_AbstractMethodError_entry =
4194 generate_throw_exception("AbstractMethodError throw_exception",
4195 CAST_FROM_FN_PTR(address,
4196 SharedRuntime::
4197 throw_AbstractMethodError));
4198
4199 StubRoutines::_throw_IncompatibleClassChangeError_entry =
4200 generate_throw_exception("IncompatibleClassChangeError throw_exception",
4201 CAST_FROM_FN_PTR(address,
4202 SharedRuntime::
4203 throw_IncompatibleClassChangeError));
4204
4205 StubRoutines::_throw_NullPointerException_at_call_entry =
4206 generate_throw_exception("NullPointerException at call throw_exception",
4207 CAST_FROM_FN_PTR(address,
4208 SharedRuntime::
4209 throw_NullPointerException_at_call));
4210
4211 // support for verify_oop (must happen after universe_init)
4212 if (VerifyOops) {
4213 StubRoutines::_verify_oop_subroutine_entry = generate_verify_oop();
4214 }
4215
4216 // data cache line writeback
4217 StubRoutines::_data_cache_writeback = generate_data_cache_writeback();
4218 StubRoutines::_data_cache_writeback_sync = generate_data_cache_writeback_sync();
4219
4220 // arraycopy stubs used by compilers
4221 generate_arraycopy_stubs();
4222
4223 BarrierSetNMethod* bs_nm = BarrierSet::barrier_set()->barrier_set_nmethod();
4224 if (bs_nm != nullptr) {
4225 StubRoutines::_method_entry_barrier = generate_method_entry_barrier();
4226 }
4227
4228 if (UseVectorizedMismatchIntrinsic) {
4229 StubRoutines::_vectorizedMismatch = generate_vectorizedMismatch();
4230 }
4231
4232 StubRoutines::_upcall_stub_exception_handler = generate_upcall_stub_exception_handler();
4233 }
4234
4235 void StubGenerator::generate_compiler_stubs() {
4236 #if COMPILER2_OR_JVMCI
4237
4238 // Entry points that are C2 compiler specific.
4239
4240 StubRoutines::x86::_vector_float_sign_mask = generate_vector_mask("vector_float_sign_mask", 0x7FFFFFFF7FFFFFFF);
4241 StubRoutines::x86::_vector_float_sign_flip = generate_vector_mask("vector_float_sign_flip", 0x8000000080000000);
4242 StubRoutines::x86::_vector_double_sign_mask = generate_vector_mask("vector_double_sign_mask", 0x7FFFFFFFFFFFFFFF);
4243 StubRoutines::x86::_vector_double_sign_flip = generate_vector_mask("vector_double_sign_flip", 0x8000000000000000);
4244 StubRoutines::x86::_vector_all_bits_set = generate_vector_mask("vector_all_bits_set", 0xFFFFFFFFFFFFFFFF);
4245 StubRoutines::x86::_vector_int_mask_cmp_bits = generate_vector_mask("vector_int_mask_cmp_bits", 0x0000000100000001);
4246 StubRoutines::x86::_vector_short_to_byte_mask = generate_vector_mask("vector_short_to_byte_mask", 0x00ff00ff00ff00ff);
4247 StubRoutines::x86::_vector_byte_perm_mask = generate_vector_byte_perm_mask("vector_byte_perm_mask");
4248 StubRoutines::x86::_vector_int_to_byte_mask = generate_vector_mask("vector_int_to_byte_mask", 0x000000ff000000ff);
4249 StubRoutines::x86::_vector_int_to_short_mask = generate_vector_mask("vector_int_to_short_mask", 0x0000ffff0000ffff);
4250 StubRoutines::x86::_vector_32_bit_mask = generate_vector_custom_i32("vector_32_bit_mask", Assembler::AVX_512bit,
4251 0xFFFFFFFF, 0, 0, 0);
4252 StubRoutines::x86::_vector_64_bit_mask = generate_vector_custom_i32("vector_64_bit_mask", Assembler::AVX_512bit,
4253 0xFFFFFFFF, 0xFFFFFFFF, 0, 0);
4254 StubRoutines::x86::_vector_int_shuffle_mask = generate_vector_mask("vector_int_shuffle_mask", 0x0302010003020100);
4255 StubRoutines::x86::_vector_byte_shuffle_mask = generate_vector_byte_shuffle_mask("vector_byte_shuffle_mask");
4256 StubRoutines::x86::_vector_short_shuffle_mask = generate_vector_mask("vector_short_shuffle_mask", 0x0100010001000100);
4257 StubRoutines::x86::_vector_long_shuffle_mask = generate_vector_mask("vector_long_shuffle_mask", 0x0000000100000000);
4258 StubRoutines::x86::_vector_long_sign_mask = generate_vector_mask("vector_long_sign_mask", 0x8000000000000000);
4259 StubRoutines::x86::_vector_iota_indices = generate_iota_indices("iota_indices");
4260 StubRoutines::x86::_vector_count_leading_zeros_lut = generate_count_leading_zeros_lut("count_leading_zeros_lut");
4261 StubRoutines::x86::_vector_reverse_bit_lut = generate_vector_reverse_bit_lut("reverse_bit_lut");
4262 StubRoutines::x86::_vector_reverse_byte_perm_mask_long = generate_vector_reverse_byte_perm_mask_long("perm_mask_long");
4263 StubRoutines::x86::_vector_reverse_byte_perm_mask_int = generate_vector_reverse_byte_perm_mask_int("perm_mask_int");
4264 StubRoutines::x86::_vector_reverse_byte_perm_mask_short = generate_vector_reverse_byte_perm_mask_short("perm_mask_short");
4265
4266 if (VM_Version::supports_avx2() && !VM_Version::supports_avx512_vpopcntdq()) {
4267 // lut implementation influenced by counting 1s algorithm from section 5-1 of Hackers' Delight.
4268 StubRoutines::x86::_vector_popcount_lut = generate_popcount_avx_lut("popcount_lut");
4269 }
4270
4271 generate_aes_stubs();
4272
4273 generate_ghash_stubs();
4274
4275 generate_chacha_stubs();
4276
4277 if (UseAdler32Intrinsics) {
4278 StubRoutines::_updateBytesAdler32 = generate_updateBytesAdler32();
4279 }
4280
4281 if (UsePoly1305Intrinsics) {
4282 StubRoutines::_poly1305_processBlocks = generate_poly1305_processBlocks();
4283 }
4284
4285 if (UseMD5Intrinsics) {
4286 StubRoutines::_md5_implCompress = generate_md5_implCompress(false, "md5_implCompress");
4287 StubRoutines::_md5_implCompressMB = generate_md5_implCompress(true, "md5_implCompressMB");
4288 }
4289
4290 if (UseSHA1Intrinsics) {
4291 StubRoutines::x86::_upper_word_mask_addr = generate_upper_word_mask();
4292 StubRoutines::x86::_shuffle_byte_flip_mask_addr = generate_shuffle_byte_flip_mask();
4293 StubRoutines::_sha1_implCompress = generate_sha1_implCompress(false, "sha1_implCompress");
4294 StubRoutines::_sha1_implCompressMB = generate_sha1_implCompress(true, "sha1_implCompressMB");
4295 }
4296
4297 if (UseSHA256Intrinsics) {
4298 StubRoutines::x86::_k256_adr = (address)StubRoutines::x86::_k256;
4299 char* dst = (char*)StubRoutines::x86::_k256_W;
4300 char* src = (char*)StubRoutines::x86::_k256;
4301 for (int ii = 0; ii < 16; ++ii) {
4302 memcpy(dst + 32 * ii, src + 16 * ii, 16);
4303 memcpy(dst + 32 * ii + 16, src + 16 * ii, 16);
4304 }
4305 StubRoutines::x86::_k256_W_adr = (address)StubRoutines::x86::_k256_W;
4306 StubRoutines::x86::_pshuffle_byte_flip_mask_addr = generate_pshuffle_byte_flip_mask();
4307 StubRoutines::_sha256_implCompress = generate_sha256_implCompress(false, "sha256_implCompress");
4308 StubRoutines::_sha256_implCompressMB = generate_sha256_implCompress(true, "sha256_implCompressMB");
4309 }
4310
4311 if (UseSHA512Intrinsics) {
4312 StubRoutines::x86::_k512_W_addr = (address)StubRoutines::x86::_k512_W;
4313 StubRoutines::x86::_pshuffle_byte_flip_mask_addr_sha512 = generate_pshuffle_byte_flip_mask_sha512();
4314 StubRoutines::_sha512_implCompress = generate_sha512_implCompress(false, "sha512_implCompress");
4315 StubRoutines::_sha512_implCompressMB = generate_sha512_implCompress(true, "sha512_implCompressMB");
4316 }
4317
4318 if (UseBASE64Intrinsics) {
4319 if(VM_Version::supports_avx2()) {
4320 StubRoutines::x86::_avx2_shuffle_base64 = base64_avx2_shuffle_addr();
4321 StubRoutines::x86::_avx2_input_mask_base64 = base64_avx2_input_mask_addr();
4322 StubRoutines::x86::_avx2_lut_base64 = base64_avx2_lut_addr();
4323 StubRoutines::x86::_avx2_decode_tables_base64 = base64_AVX2_decode_tables_addr();
4324 StubRoutines::x86::_avx2_decode_lut_tables_base64 = base64_AVX2_decode_LUT_tables_addr();
4325 }
4326 StubRoutines::x86::_encoding_table_base64 = base64_encoding_table_addr();
4327 if (VM_Version::supports_avx512_vbmi()) {
4328 StubRoutines::x86::_shuffle_base64 = base64_shuffle_addr();
4329 StubRoutines::x86::_lookup_lo_base64 = base64_vbmi_lookup_lo_addr();
4330 StubRoutines::x86::_lookup_hi_base64 = base64_vbmi_lookup_hi_addr();
4331 StubRoutines::x86::_lookup_lo_base64url = base64_vbmi_lookup_lo_url_addr();
4332 StubRoutines::x86::_lookup_hi_base64url = base64_vbmi_lookup_hi_url_addr();
4333 StubRoutines::x86::_pack_vec_base64 = base64_vbmi_pack_vec_addr();
4334 StubRoutines::x86::_join_0_1_base64 = base64_vbmi_join_0_1_addr();
4335 StubRoutines::x86::_join_1_2_base64 = base64_vbmi_join_1_2_addr();
4336 StubRoutines::x86::_join_2_3_base64 = base64_vbmi_join_2_3_addr();
4337 }
4338 StubRoutines::x86::_decoding_table_base64 = base64_decoding_table_addr();
4339 StubRoutines::_base64_encodeBlock = generate_base64_encodeBlock();
4340 StubRoutines::_base64_decodeBlock = generate_base64_decodeBlock();
4341 }
4342
4343 #ifdef COMPILER2
4344 if (UseMultiplyToLenIntrinsic) {
4345 StubRoutines::_multiplyToLen = generate_multiplyToLen();
4346 }
4347 if (UseSquareToLenIntrinsic) {
4348 StubRoutines::_squareToLen = generate_squareToLen();
4349 }
4350 if (UseMulAddIntrinsic) {
4351 StubRoutines::_mulAdd = generate_mulAdd();
4352 }
4353 if (VM_Version::supports_avx512_vbmi2()) {
4354 StubRoutines::_bigIntegerRightShiftWorker = generate_bigIntegerRightShift();
4355 StubRoutines::_bigIntegerLeftShiftWorker = generate_bigIntegerLeftShift();
4356 }
4357 if (UseMontgomeryMultiplyIntrinsic) {
4358 StubRoutines::_montgomeryMultiply
4359 = CAST_FROM_FN_PTR(address, SharedRuntime::montgomery_multiply);
4360 }
4361 if (UseMontgomerySquareIntrinsic) {
4362 StubRoutines::_montgomerySquare
4363 = CAST_FROM_FN_PTR(address, SharedRuntime::montgomery_square);
4364 }
4365
4366 // Load x86_64_sort library on supported hardware to enable avx512 sort and partition intrinsics
4367 if (VM_Version::is_intel() && VM_Version::supports_avx512dq()) {
4368 void *libsimdsort = nullptr;
4369 char ebuf_[1024];
4370 char dll_name_simd_sort[JVM_MAXPATHLEN];
4371 if (os::dll_locate_lib(dll_name_simd_sort, sizeof(dll_name_simd_sort), Arguments::get_dll_dir(), "simdsort")) {
4372 libsimdsort = os::dll_load(dll_name_simd_sort, ebuf_, sizeof ebuf_);
4373 }
4374 // Get addresses for avx512 sort and partition routines
4375 if (libsimdsort != nullptr) {
4376 log_info(library)("Loaded library %s, handle " INTPTR_FORMAT, JNI_LIB_PREFIX "simdsort" JNI_LIB_SUFFIX, p2i(libsimdsort));
4377
4378 snprintf(ebuf_, sizeof(ebuf_), "avx512_sort");
4379 StubRoutines::_array_sort = (address)os::dll_lookup(libsimdsort, ebuf_);
4380
4381 snprintf(ebuf_, sizeof(ebuf_), "avx512_partition");
4382 StubRoutines::_array_partition = (address)os::dll_lookup(libsimdsort, ebuf_);
4383 }
4384 }
4385
4386 // Get svml stub routine addresses
4387 void *libjsvml = nullptr;
4388 char ebuf[1024];
4389 char dll_name[JVM_MAXPATHLEN];
4390 if (os::dll_locate_lib(dll_name, sizeof(dll_name), Arguments::get_dll_dir(), "jsvml")) {
4391 libjsvml = os::dll_load(dll_name, ebuf, sizeof ebuf);
4392 }
4393 if (libjsvml != nullptr) {
4394 // SVML method naming convention
4395 // All the methods are named as __jsvml_op<T><N>_ha_<VV>
4396 // Where:
4397 // ha stands for high accuracy
4398 // <T> is optional to indicate float/double
4399 // Set to f for vector float operation
4400 // Omitted for vector double operation
4401 // <N> is the number of elements in the vector
4402 // 1, 2, 4, 8, 16
4403 // e.g. 128 bit float vector has 4 float elements
4404 // <VV> indicates the avx/sse level:
4405 // z0 is AVX512, l9 is AVX2, e9 is AVX1 and ex is for SSE2
4406 // e.g. __jsvml_expf16_ha_z0 is the method for computing 16 element vector float exp using AVX 512 insns
4407 // __jsvml_exp8_ha_z0 is the method for computing 8 element vector double exp using AVX 512 insns
4408
4409 log_info(library)("Loaded library %s, handle " INTPTR_FORMAT, JNI_LIB_PREFIX "jsvml" JNI_LIB_SUFFIX, p2i(libjsvml));
4410 if (UseAVX > 2) {
4411 for (int op = 0; op < VectorSupport::NUM_SVML_OP; op++) {
4412 int vop = VectorSupport::VECTOR_OP_SVML_START + op;
4413 if ((!VM_Version::supports_avx512dq()) &&
4414 (vop == VectorSupport::VECTOR_OP_LOG || vop == VectorSupport::VECTOR_OP_LOG10 || vop == VectorSupport::VECTOR_OP_POW)) {
4415 continue;
4416 }
4417 snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf16_ha_z0", VectorSupport::svmlname[op]);
4418 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_512][op] = (address)os::dll_lookup(libjsvml, ebuf);
4419
4420 snprintf(ebuf, sizeof(ebuf), "__jsvml_%s8_ha_z0", VectorSupport::svmlname[op]);
4421 StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_512][op] = (address)os::dll_lookup(libjsvml, ebuf);
4422 }
4423 }
4424 const char* avx_sse_str = (UseAVX >= 2) ? "l9" : ((UseAVX == 1) ? "e9" : "ex");
4425 for (int op = 0; op < VectorSupport::NUM_SVML_OP; op++) {
4426 int vop = VectorSupport::VECTOR_OP_SVML_START + op;
4427 if (vop == VectorSupport::VECTOR_OP_POW) {
4428 continue;
4429 }
4430 snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf4_ha_%s", VectorSupport::svmlname[op], avx_sse_str);
4431 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_64][op] = (address)os::dll_lookup(libjsvml, ebuf);
4432
4433 snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf4_ha_%s", VectorSupport::svmlname[op], avx_sse_str);
4434 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_128][op] = (address)os::dll_lookup(libjsvml, ebuf);
4435
4436 snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf8_ha_%s", VectorSupport::svmlname[op], avx_sse_str);
4437 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_256][op] = (address)os::dll_lookup(libjsvml, ebuf);
4438
4439 snprintf(ebuf, sizeof(ebuf), "__jsvml_%s1_ha_%s", VectorSupport::svmlname[op], avx_sse_str);
4440 StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_64][op] = (address)os::dll_lookup(libjsvml, ebuf);
4441
4442 snprintf(ebuf, sizeof(ebuf), "__jsvml_%s2_ha_%s", VectorSupport::svmlname[op], avx_sse_str);
4443 StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_128][op] = (address)os::dll_lookup(libjsvml, ebuf);
4444
4445 snprintf(ebuf, sizeof(ebuf), "__jsvml_%s4_ha_%s", VectorSupport::svmlname[op], avx_sse_str);
4446 StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_256][op] = (address)os::dll_lookup(libjsvml, ebuf);
4447 }
4448 }
4449 #endif // COMPILER2
4450 #endif // COMPILER2_OR_JVMCI
4451 }
4452
4453 StubGenerator::StubGenerator(CodeBuffer* code, StubsKind kind) : StubCodeGenerator(code) {
4454 DEBUG_ONLY( _regs_in_thread = false; )
4455 switch(kind) {
4456 case Initial_stubs:
4457 generate_initial_stubs();
4458 break;
4459 case Continuation_stubs:
4460 generate_continuation_stubs();
4461 break;
4462 case Compiler_stubs:
4463 generate_compiler_stubs();
4464 break;
4465 case Final_stubs:
4466 generate_final_stubs();
4467 break;
4468 default:
4469 fatal("unexpected stubs kind: %d", kind);
4470 break;
4471 };
4472 }
4473
4474 void StubGenerator_generate(CodeBuffer* code, StubCodeGenerator::StubsKind kind) {
4475 StubGenerator g(code, kind);
4476 }
4477
4478 #undef __