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