1 /*
2 * Copyright (c) 2003, 2024, 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/javaClasses.hpp"
28 #include "classfile/vmIntrinsics.hpp"
29 #include "code/SCCache.hpp"
30 #include "compiler/oopMap.hpp"
31 #include "gc/shared/barrierSet.hpp"
32 #include "gc/shared/barrierSetAssembler.hpp"
33 #include "gc/shared/barrierSetNMethod.hpp"
34 #include "gc/shared/gc_globals.hpp"
35 #include "memory/universe.hpp"
36 #include "prims/jvmtiExport.hpp"
37 #include "prims/upcallLinker.hpp"
38 #include "runtime/arguments.hpp"
39 #include "runtime/continuationEntry.hpp"
40 #include "runtime/javaThread.hpp"
41 #include "runtime/sharedRuntime.hpp"
42 #include "runtime/stubRoutines.hpp"
43 #include "stubGenerator_x86_64.hpp"
44 #ifdef COMPILER2
45 #include "opto/runtime.hpp"
46 #include "opto/c2_globals.hpp"
47 #endif
48 #if INCLUDE_JVMCI
49 #include "jvmci/jvmci_globals.hpp"
50 #endif
51
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_compress_perm_table(const char *stub_name, int32_t esize) {
955 __ align(CodeEntryAlignment);
956 StubCodeMark mark(this, "StubRoutines", stub_name);
957 address start = __ pc();
958 if (esize == 32) {
959 // Loop to generate 256 x 8 int compression permute index table. A row is
960 // accessed using 8 bit index computed using vector mask. An entry in
961 // a row holds either a valid permute index corresponding to set bit position
962 // or a -1 (default) value.
963 for (int mask = 0; mask < 256; mask++) {
964 int ctr = 0;
965 for (int j = 0; j < 8; j++) {
966 if (mask & (1 << j)) {
967 __ emit_data(j, relocInfo::none);
968 ctr++;
969 }
970 }
971 for (; ctr < 8; ctr++) {
972 __ emit_data(-1, relocInfo::none);
973 }
974 }
975 } else {
976 assert(esize == 64, "");
977 // Loop to generate 16 x 4 long compression permute index table. A row is
978 // accessed using 4 bit index computed using vector mask. An entry in
979 // a row holds either a valid permute index pair for a quadword corresponding
980 // to set bit position or a -1 (default) value.
981 for (int mask = 0; mask < 16; mask++) {
982 int ctr = 0;
983 for (int j = 0; j < 4; j++) {
984 if (mask & (1 << j)) {
985 __ emit_data(2 * j, relocInfo::none);
986 __ emit_data(2 * j + 1, relocInfo::none);
987 ctr++;
988 }
989 }
990 for (; ctr < 4; ctr++) {
991 __ emit_data64(-1L, relocInfo::none);
992 }
993 }
994 }
995 return start;
996 }
997
998 address StubGenerator::generate_expand_perm_table(const char *stub_name, int32_t esize) {
999 __ align(CodeEntryAlignment);
1000 StubCodeMark mark(this, "StubRoutines", stub_name);
1001 address start = __ pc();
1002 if (esize == 32) {
1003 // Loop to generate 256 x 8 int expand permute index table. A row is accessed
1004 // using 8 bit index computed using vector mask. An entry in a row holds either
1005 // a valid permute index (starting from least significant lane) placed at poisition
1006 // corresponding to set bit position or a -1 (default) value.
1007 for (int mask = 0; mask < 256; mask++) {
1008 int ctr = 0;
1009 for (int j = 0; j < 8; j++) {
1010 if (mask & (1 << j)) {
1011 __ emit_data(ctr++, relocInfo::none);
1012 } else {
1013 __ emit_data(-1, relocInfo::none);
1014 }
1015 }
1016 }
1017 } else {
1018 assert(esize == 64, "");
1019 // Loop to generate 16 x 4 long expand permute index table. A row is accessed
1020 // using 4 bit index computed using vector mask. An entry in a row holds either
1021 // a valid doubleword permute index pair representing a quadword index (starting
1022 // from least significant lane) placed at poisition corresponding to set bit
1023 // position or a -1 (default) value.
1024 for (int mask = 0; mask < 16; mask++) {
1025 int ctr = 0;
1026 for (int j = 0; j < 4; j++) {
1027 if (mask & (1 << j)) {
1028 __ emit_data(2 * ctr, relocInfo::none);
1029 __ emit_data(2 * ctr + 1, relocInfo::none);
1030 ctr++;
1031 } else {
1032 __ emit_data64(-1L, relocInfo::none);
1033 }
1034 }
1035 }
1036 }
1037 return start;
1038 }
1039
1040 address StubGenerator::generate_vector_mask(const char *stub_name, int64_t mask) {
1041 __ align(CodeEntryAlignment);
1042 StubCodeMark mark(this, "StubRoutines", stub_name);
1043 address start = __ pc();
1044
1045 __ emit_data64(mask, relocInfo::none);
1046 __ emit_data64(mask, relocInfo::none);
1047 __ emit_data64(mask, relocInfo::none);
1048 __ emit_data64(mask, relocInfo::none);
1049 __ emit_data64(mask, relocInfo::none);
1050 __ emit_data64(mask, relocInfo::none);
1051 __ emit_data64(mask, relocInfo::none);
1052 __ emit_data64(mask, relocInfo::none);
1053
1054 return start;
1055 }
1056
1057 address StubGenerator::generate_vector_byte_perm_mask(const char *stub_name) {
1058 __ align(CodeEntryAlignment);
1059 StubCodeMark mark(this, "StubRoutines", stub_name);
1060 address start = __ pc();
1061
1062 __ emit_data64(0x0000000000000001, relocInfo::none);
1063 __ emit_data64(0x0000000000000003, relocInfo::none);
1064 __ emit_data64(0x0000000000000005, relocInfo::none);
1065 __ emit_data64(0x0000000000000007, relocInfo::none);
1066 __ emit_data64(0x0000000000000000, relocInfo::none);
1067 __ emit_data64(0x0000000000000002, relocInfo::none);
1068 __ emit_data64(0x0000000000000004, relocInfo::none);
1069 __ emit_data64(0x0000000000000006, relocInfo::none);
1070
1071 return start;
1072 }
1073
1074 address StubGenerator::generate_vector_fp_mask(const char *stub_name, int64_t mask) {
1075 __ align(CodeEntryAlignment);
1076 StubCodeMark mark(this, "StubRoutines", stub_name);
1077 address start = __ pc();
1078
1079 __ emit_data64(mask, relocInfo::none);
1080 __ emit_data64(mask, relocInfo::none);
1081 __ emit_data64(mask, relocInfo::none);
1082 __ emit_data64(mask, relocInfo::none);
1083 __ emit_data64(mask, relocInfo::none);
1084 __ emit_data64(mask, relocInfo::none);
1085 __ emit_data64(mask, relocInfo::none);
1086 __ emit_data64(mask, relocInfo::none);
1087
1088 return start;
1089 }
1090
1091 address StubGenerator::generate_vector_custom_i32(const char *stub_name, Assembler::AvxVectorLen len,
1092 int32_t val0, int32_t val1, int32_t val2, int32_t val3,
1093 int32_t val4, int32_t val5, int32_t val6, int32_t val7,
1094 int32_t val8, int32_t val9, int32_t val10, int32_t val11,
1095 int32_t val12, int32_t val13, int32_t val14, int32_t val15) {
1096 __ align(CodeEntryAlignment);
1097 StubCodeMark mark(this, "StubRoutines", stub_name);
1098 address start = __ pc();
1099
1100 assert(len != Assembler::AVX_NoVec, "vector len must be specified");
1101 __ emit_data(val0, relocInfo::none, 0);
1102 __ emit_data(val1, relocInfo::none, 0);
1103 __ emit_data(val2, relocInfo::none, 0);
1104 __ emit_data(val3, relocInfo::none, 0);
1105 if (len >= Assembler::AVX_256bit) {
1106 __ emit_data(val4, relocInfo::none, 0);
1107 __ emit_data(val5, relocInfo::none, 0);
1108 __ emit_data(val6, relocInfo::none, 0);
1109 __ emit_data(val7, relocInfo::none, 0);
1110 if (len >= Assembler::AVX_512bit) {
1111 __ emit_data(val8, relocInfo::none, 0);
1112 __ emit_data(val9, relocInfo::none, 0);
1113 __ emit_data(val10, relocInfo::none, 0);
1114 __ emit_data(val11, relocInfo::none, 0);
1115 __ emit_data(val12, relocInfo::none, 0);
1116 __ emit_data(val13, relocInfo::none, 0);
1117 __ emit_data(val14, relocInfo::none, 0);
1118 __ emit_data(val15, relocInfo::none, 0);
1119 }
1120 }
1121 return start;
1122 }
1123
1124 // Non-destructive plausibility checks for oops
1125 //
1126 // Arguments:
1127 // all args on stack!
1128 //
1129 // Stack after saving c_rarg3:
1130 // [tos + 0]: saved c_rarg3
1131 // [tos + 1]: saved c_rarg2
1132 // [tos + 2]: saved r12 (several TemplateTable methods use it)
1133 // [tos + 3]: saved flags
1134 // [tos + 4]: return address
1135 // * [tos + 5]: error message (char*)
1136 // * [tos + 6]: object to verify (oop)
1137 // * [tos + 7]: saved rax - saved by caller and bashed
1138 // * [tos + 8]: saved r10 (rscratch1) - saved by caller
1139 // * = popped on exit
1140 address StubGenerator::generate_verify_oop() {
1141 StubCodeMark mark(this, "StubRoutines", "verify_oop");
1142 address start = __ pc();
1143
1144 Label exit, error;
1145
1146 __ pushf();
1147 __ incrementl(ExternalAddress((address) StubRoutines::verify_oop_count_addr()), rscratch1);
1148
1149 __ push(r12);
1150
1151 // save c_rarg2 and c_rarg3
1152 __ push(c_rarg2);
1153 __ push(c_rarg3);
1154
1155 enum {
1156 // After previous pushes.
1157 oop_to_verify = 6 * wordSize,
1158 saved_rax = 7 * wordSize,
1159 saved_r10 = 8 * wordSize,
1160
1161 // Before the call to MacroAssembler::debug(), see below.
1162 return_addr = 16 * wordSize,
1163 error_msg = 17 * wordSize
1164 };
1165
1166 // get object
1167 __ movptr(rax, Address(rsp, oop_to_verify));
1168
1169 // make sure object is 'reasonable'
1170 __ testptr(rax, rax);
1171 __ jcc(Assembler::zero, exit); // if obj is null it is OK
1172
1173 BarrierSetAssembler* bs_asm = BarrierSet::barrier_set()->barrier_set_assembler();
1174 bs_asm->check_oop(_masm, rax, c_rarg2, c_rarg3, error);
1175
1176 // return if everything seems ok
1177 __ bind(exit);
1178 __ movptr(rax, Address(rsp, saved_rax)); // get saved rax back
1179 __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back
1180 __ pop(c_rarg3); // restore c_rarg3
1181 __ pop(c_rarg2); // restore c_rarg2
1182 __ pop(r12); // restore r12
1183 __ popf(); // restore flags
1184 __ ret(4 * wordSize); // pop caller saved stuff
1185
1186 // handle errors
1187 __ bind(error);
1188 __ movptr(rax, Address(rsp, saved_rax)); // get saved rax back
1189 __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back
1190 __ pop(c_rarg3); // get saved c_rarg3 back
1191 __ pop(c_rarg2); // get saved c_rarg2 back
1192 __ pop(r12); // get saved r12 back
1193 __ popf(); // get saved flags off stack --
1194 // will be ignored
1195
1196 __ pusha(); // push registers
1197 // (rip is already
1198 // already pushed)
1199 // debug(char* msg, int64_t pc, int64_t regs[])
1200 // We've popped the registers we'd saved (c_rarg3, c_rarg2 and flags), and
1201 // pushed all the registers, so now the stack looks like:
1202 // [tos + 0] 16 saved registers
1203 // [tos + 16] return address
1204 // * [tos + 17] error message (char*)
1205 // * [tos + 18] object to verify (oop)
1206 // * [tos + 19] saved rax - saved by caller and bashed
1207 // * [tos + 20] saved r10 (rscratch1) - saved by caller
1208 // * = popped on exit
1209
1210 __ movptr(c_rarg0, Address(rsp, error_msg)); // pass address of error message
1211 __ movptr(c_rarg1, Address(rsp, return_addr)); // pass return address
1212 __ movq(c_rarg2, rsp); // pass address of regs on stack
1213 __ mov(r12, rsp); // remember rsp
1214 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows
1215 __ andptr(rsp, -16); // align stack as required by ABI
1216 BLOCK_COMMENT("call MacroAssembler::debug");
1217 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64)));
1218 __ hlt();
1219
1220 return start;
1221 }
1222
1223
1224 // Shuffle first three arg regs on Windows into Linux/Solaris locations.
1225 //
1226 // Outputs:
1227 // rdi - rcx
1228 // rsi - rdx
1229 // rdx - r8
1230 // rcx - r9
1231 //
1232 // Registers r9 and r10 are used to save rdi and rsi on Windows, which latter
1233 // are non-volatile. r9 and r10 should not be used by the caller.
1234 //
1235 void StubGenerator::setup_arg_regs(int nargs) {
1236 const Register saved_rdi = r9;
1237 const Register saved_rsi = r10;
1238 assert(nargs == 3 || nargs == 4, "else fix");
1239 #ifdef _WIN64
1240 assert(c_rarg0 == rcx && c_rarg1 == rdx && c_rarg2 == r8 && c_rarg3 == r9,
1241 "unexpected argument registers");
1242 if (nargs == 4) {
1243 __ mov(rax, r9); // r9 is also saved_rdi
1244 }
1245 __ movptr(saved_rdi, rdi);
1246 __ movptr(saved_rsi, rsi);
1247 __ mov(rdi, rcx); // c_rarg0
1248 __ mov(rsi, rdx); // c_rarg1
1249 __ mov(rdx, r8); // c_rarg2
1250 if (nargs == 4) {
1251 __ mov(rcx, rax); // c_rarg3 (via rax)
1252 }
1253 #else
1254 assert(c_rarg0 == rdi && c_rarg1 == rsi && c_rarg2 == rdx && c_rarg3 == rcx,
1255 "unexpected argument registers");
1256 #endif
1257 DEBUG_ONLY(_regs_in_thread = false;)
1258 }
1259
1260
1261 void StubGenerator::restore_arg_regs() {
1262 assert(!_regs_in_thread, "wrong call to restore_arg_regs");
1263 const Register saved_rdi = r9;
1264 const Register saved_rsi = r10;
1265 #ifdef _WIN64
1266 __ movptr(rdi, saved_rdi);
1267 __ movptr(rsi, saved_rsi);
1268 #endif
1269 }
1270
1271
1272 // This is used in places where r10 is a scratch register, and can
1273 // be adapted if r9 is needed also.
1274 void StubGenerator::setup_arg_regs_using_thread(int nargs) {
1275 const Register saved_r15 = r9;
1276 assert(nargs == 3 || nargs == 4, "else fix");
1277 #ifdef _WIN64
1278 if (nargs == 4) {
1279 __ mov(rax, r9); // r9 is also saved_r15
1280 }
1281 __ mov(saved_r15, r15); // r15 is callee saved and needs to be restored
1282 __ get_thread(r15_thread);
1283 assert(c_rarg0 == rcx && c_rarg1 == rdx && c_rarg2 == r8 && c_rarg3 == r9,
1284 "unexpected argument registers");
1285 __ movptr(Address(r15_thread, in_bytes(JavaThread::windows_saved_rdi_offset())), rdi);
1286 __ movptr(Address(r15_thread, in_bytes(JavaThread::windows_saved_rsi_offset())), rsi);
1287
1288 __ mov(rdi, rcx); // c_rarg0
1289 __ mov(rsi, rdx); // c_rarg1
1290 __ mov(rdx, r8); // c_rarg2
1291 if (nargs == 4) {
1292 __ mov(rcx, rax); // c_rarg3 (via rax)
1293 }
1294 #else
1295 assert(c_rarg0 == rdi && c_rarg1 == rsi && c_rarg2 == rdx && c_rarg3 == rcx,
1296 "unexpected argument registers");
1297 #endif
1298 DEBUG_ONLY(_regs_in_thread = true;)
1299 }
1300
1301
1302 void StubGenerator::restore_arg_regs_using_thread() {
1303 assert(_regs_in_thread, "wrong call to restore_arg_regs");
1304 const Register saved_r15 = r9;
1305 #ifdef _WIN64
1306 __ get_thread(r15_thread);
1307 __ movptr(rsi, Address(r15_thread, in_bytes(JavaThread::windows_saved_rsi_offset())));
1308 __ movptr(rdi, Address(r15_thread, in_bytes(JavaThread::windows_saved_rdi_offset())));
1309 __ mov(r15, saved_r15); // r15 is callee saved and needs to be restored
1310 #endif
1311 }
1312
1313
1314 void StubGenerator::setup_argument_regs(BasicType type) {
1315 if (type == T_BYTE || type == T_SHORT) {
1316 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1317 // r9 and r10 may be used to save non-volatile registers
1318 } else {
1319 setup_arg_regs_using_thread(); // from => rdi, to => rsi, count => rdx
1320 // r9 is used to save r15_thread
1321 }
1322 }
1323
1324
1325 void StubGenerator::restore_argument_regs(BasicType type) {
1326 if (type == T_BYTE || type == T_SHORT) {
1327 restore_arg_regs();
1328 } else {
1329 restore_arg_regs_using_thread();
1330 }
1331 }
1332
1333 address StubGenerator::generate_data_cache_writeback() {
1334 const Register src = c_rarg0; // source address
1335
1336 __ align(CodeEntryAlignment);
1337
1338 StubCodeMark mark(this, "StubRoutines", "_data_cache_writeback");
1339
1340 address start = __ pc();
1341
1342 __ enter();
1343 __ cache_wb(Address(src, 0));
1344 __ leave();
1345 __ ret(0);
1346
1347 return start;
1348 }
1349
1350 address StubGenerator::generate_data_cache_writeback_sync() {
1351 const Register is_pre = c_rarg0; // pre or post sync
1352
1353 __ align(CodeEntryAlignment);
1354
1355 StubCodeMark mark(this, "StubRoutines", "_data_cache_writeback_sync");
1356
1357 // pre wbsync is a no-op
1358 // post wbsync translates to an sfence
1359
1360 Label skip;
1361 address start = __ pc();
1362
1363 __ enter();
1364 __ cmpl(is_pre, 0);
1365 __ jcc(Assembler::notEqual, skip);
1366 __ cache_wbsync(false);
1367 __ bind(skip);
1368 __ leave();
1369 __ ret(0);
1370
1371 return start;
1372 }
1373
1374 // ofs and limit are use for multi-block byte array.
1375 // int com.sun.security.provider.MD5.implCompress(byte[] b, int ofs)
1376 address StubGenerator::generate_md5_implCompress(bool multi_block, const char *name) {
1377 __ align(CodeEntryAlignment);
1378 StubCodeMark mark(this, "StubRoutines", name);
1379 address start = __ pc();
1380
1381 const Register buf_param = r15;
1382 const Address state_param(rsp, 0 * wordSize);
1383 const Address ofs_param (rsp, 1 * wordSize );
1384 const Address limit_param(rsp, 1 * wordSize + 4);
1385
1386 __ enter();
1387 __ push(rbx);
1388 __ push(rdi);
1389 __ push(rsi);
1390 __ push(r15);
1391 __ subptr(rsp, 2 * wordSize);
1392
1393 __ movptr(buf_param, c_rarg0);
1394 __ movptr(state_param, c_rarg1);
1395 if (multi_block) {
1396 __ movl(ofs_param, c_rarg2);
1397 __ movl(limit_param, c_rarg3);
1398 }
1399 __ fast_md5(buf_param, state_param, ofs_param, limit_param, multi_block);
1400
1401 __ addptr(rsp, 2 * wordSize);
1402 __ pop(r15);
1403 __ pop(rsi);
1404 __ pop(rdi);
1405 __ pop(rbx);
1406 __ leave();
1407 __ ret(0);
1408
1409 return start;
1410 }
1411
1412 address StubGenerator::generate_upper_word_mask() {
1413 __ align64();
1414 StubCodeMark mark(this, "StubRoutines", "upper_word_mask");
1415 address start = __ pc();
1416
1417 __ emit_data64(0x0000000000000000, relocInfo::none);
1418 __ emit_data64(0xFFFFFFFF00000000, relocInfo::none);
1419
1420 return start;
1421 }
1422
1423 address StubGenerator::generate_shuffle_byte_flip_mask() {
1424 __ align64();
1425 StubCodeMark mark(this, "StubRoutines", "shuffle_byte_flip_mask");
1426 address start = __ pc();
1427
1428 __ emit_data64(0x08090a0b0c0d0e0f, relocInfo::none);
1429 __ emit_data64(0x0001020304050607, relocInfo::none);
1430
1431 return start;
1432 }
1433
1434 // ofs and limit are use for multi-block byte array.
1435 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
1436 address StubGenerator::generate_sha1_implCompress(bool multi_block, const char *name) {
1437 __ align(CodeEntryAlignment);
1438 StubCodeMark mark(this, "StubRoutines", name);
1439 address start = __ pc();
1440
1441 Register buf = c_rarg0;
1442 Register state = c_rarg1;
1443 Register ofs = c_rarg2;
1444 Register limit = c_rarg3;
1445
1446 const XMMRegister abcd = xmm0;
1447 const XMMRegister e0 = xmm1;
1448 const XMMRegister e1 = xmm2;
1449 const XMMRegister msg0 = xmm3;
1450
1451 const XMMRegister msg1 = xmm4;
1452 const XMMRegister msg2 = xmm5;
1453 const XMMRegister msg3 = xmm6;
1454 const XMMRegister shuf_mask = xmm7;
1455
1456 __ enter();
1457
1458 __ subptr(rsp, 4 * wordSize);
1459
1460 __ fast_sha1(abcd, e0, e1, msg0, msg1, msg2, msg3, shuf_mask,
1461 buf, state, ofs, limit, rsp, multi_block);
1462
1463 __ addptr(rsp, 4 * wordSize);
1464
1465 __ leave();
1466 __ ret(0);
1467
1468 return start;
1469 }
1470
1471 address StubGenerator::generate_pshuffle_byte_flip_mask() {
1472 __ align64();
1473 StubCodeMark mark(this, "StubRoutines", "pshuffle_byte_flip_mask");
1474 address start = __ pc();
1475
1476 __ emit_data64(0x0405060700010203, relocInfo::none);
1477 __ emit_data64(0x0c0d0e0f08090a0b, relocInfo::none);
1478
1479 if (VM_Version::supports_avx2()) {
1480 __ emit_data64(0x0405060700010203, relocInfo::none); // second copy
1481 __ emit_data64(0x0c0d0e0f08090a0b, relocInfo::none);
1482 // _SHUF_00BA
1483 __ emit_data64(0x0b0a090803020100, relocInfo::none);
1484 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1485 __ emit_data64(0x0b0a090803020100, relocInfo::none);
1486 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1487 // _SHUF_DC00
1488 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1489 __ emit_data64(0x0b0a090803020100, relocInfo::none);
1490 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1491 __ emit_data64(0x0b0a090803020100, relocInfo::none);
1492 }
1493
1494 return start;
1495 }
1496
1497 //Mask for byte-swapping a couple of qwords in an XMM register using (v)pshufb.
1498 address StubGenerator::generate_pshuffle_byte_flip_mask_sha512() {
1499 __ align32();
1500 StubCodeMark mark(this, "StubRoutines", "pshuffle_byte_flip_mask_sha512");
1501 address start = __ pc();
1502
1503 if (VM_Version::supports_avx2()) {
1504 __ emit_data64(0x0001020304050607, relocInfo::none); // PSHUFFLE_BYTE_FLIP_MASK
1505 __ emit_data64(0x08090a0b0c0d0e0f, relocInfo::none);
1506 __ emit_data64(0x1011121314151617, relocInfo::none);
1507 __ emit_data64(0x18191a1b1c1d1e1f, relocInfo::none);
1508 __ emit_data64(0x0000000000000000, relocInfo::none); //MASK_YMM_LO
1509 __ emit_data64(0x0000000000000000, relocInfo::none);
1510 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1511 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1512 }
1513
1514 return start;
1515 }
1516
1517 // ofs and limit are use for multi-block byte array.
1518 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
1519 address StubGenerator::generate_sha256_implCompress(bool multi_block, const char *name) {
1520 assert(VM_Version::supports_sha() || VM_Version::supports_avx2(), "");
1521 __ align(CodeEntryAlignment);
1522 StubCodeMark mark(this, "StubRoutines", name);
1523 address start = __ pc();
1524
1525 Register buf = c_rarg0;
1526 Register state = c_rarg1;
1527 Register ofs = c_rarg2;
1528 Register limit = c_rarg3;
1529
1530 const XMMRegister msg = xmm0;
1531 const XMMRegister state0 = xmm1;
1532 const XMMRegister state1 = xmm2;
1533 const XMMRegister msgtmp0 = xmm3;
1534
1535 const XMMRegister msgtmp1 = xmm4;
1536 const XMMRegister msgtmp2 = xmm5;
1537 const XMMRegister msgtmp3 = xmm6;
1538 const XMMRegister msgtmp4 = xmm7;
1539
1540 const XMMRegister shuf_mask = xmm8;
1541
1542 __ enter();
1543
1544 __ subptr(rsp, 4 * wordSize);
1545
1546 if (VM_Version::supports_sha()) {
1547 __ fast_sha256(msg, state0, state1, msgtmp0, msgtmp1, msgtmp2, msgtmp3, msgtmp4,
1548 buf, state, ofs, limit, rsp, multi_block, shuf_mask);
1549 } else if (VM_Version::supports_avx2()) {
1550 __ sha256_AVX2(msg, state0, state1, msgtmp0, msgtmp1, msgtmp2, msgtmp3, msgtmp4,
1551 buf, state, ofs, limit, rsp, multi_block, shuf_mask);
1552 }
1553 __ addptr(rsp, 4 * wordSize);
1554 __ vzeroupper();
1555 __ leave();
1556 __ ret(0);
1557
1558 return start;
1559 }
1560
1561 address StubGenerator::generate_sha512_implCompress(bool multi_block, const char *name) {
1562 assert(VM_Version::supports_avx2(), "");
1563 assert(VM_Version::supports_bmi2() || VM_Version::supports_sha512(), "");
1564 __ align(CodeEntryAlignment);
1565 StubCodeMark mark(this, "StubRoutines", name);
1566 address start = __ pc();
1567
1568 Register buf = c_rarg0;
1569 Register state = c_rarg1;
1570 Register ofs = c_rarg2;
1571 Register limit = c_rarg3;
1572
1573 __ enter();
1574
1575 if (VM_Version::supports_sha512()) {
1576 __ sha512_update_ni_x1(state, buf, ofs, limit, multi_block);
1577 } else {
1578 const XMMRegister msg = xmm0;
1579 const XMMRegister state0 = xmm1;
1580 const XMMRegister state1 = xmm2;
1581 const XMMRegister msgtmp0 = xmm3;
1582 const XMMRegister msgtmp1 = xmm4;
1583 const XMMRegister msgtmp2 = xmm5;
1584 const XMMRegister msgtmp3 = xmm6;
1585 const XMMRegister msgtmp4 = xmm7;
1586
1587 const XMMRegister shuf_mask = xmm8;
1588 __ sha512_AVX2(msg, state0, state1, msgtmp0, msgtmp1, msgtmp2, msgtmp3, msgtmp4,
1589 buf, state, ofs, limit, rsp, multi_block, shuf_mask);
1590 }
1591 __ vzeroupper();
1592 __ leave();
1593 __ ret(0);
1594
1595 return start;
1596 }
1597
1598 address StubGenerator::base64_shuffle_addr() {
1599 __ align64();
1600 StubCodeMark mark(this, "StubRoutines", "shuffle_base64");
1601 address start = __ pc();
1602
1603 assert(((unsigned long long)start & 0x3f) == 0,
1604 "Alignment problem (0x%08llx)", (unsigned long long)start);
1605 __ emit_data64(0x0405030401020001, relocInfo::none);
1606 __ emit_data64(0x0a0b090a07080607, relocInfo::none);
1607 __ emit_data64(0x10110f100d0e0c0d, relocInfo::none);
1608 __ emit_data64(0x1617151613141213, relocInfo::none);
1609 __ emit_data64(0x1c1d1b1c191a1819, relocInfo::none);
1610 __ emit_data64(0x222321221f201e1f, relocInfo::none);
1611 __ emit_data64(0x2829272825262425, relocInfo::none);
1612 __ emit_data64(0x2e2f2d2e2b2c2a2b, relocInfo::none);
1613
1614 return start;
1615 }
1616
1617 address StubGenerator::base64_avx2_shuffle_addr() {
1618 __ align32();
1619 StubCodeMark mark(this, "StubRoutines", "avx2_shuffle_base64");
1620 address start = __ pc();
1621
1622 __ emit_data64(0x0809070805060405, relocInfo::none);
1623 __ emit_data64(0x0e0f0d0e0b0c0a0b, relocInfo::none);
1624 __ emit_data64(0x0405030401020001, relocInfo::none);
1625 __ emit_data64(0x0a0b090a07080607, relocInfo::none);
1626
1627 return start;
1628 }
1629
1630 address StubGenerator::base64_avx2_input_mask_addr() {
1631 __ align32();
1632 StubCodeMark mark(this, "StubRoutines", "avx2_input_mask_base64");
1633 address start = __ pc();
1634
1635 __ emit_data64(0x8000000000000000, relocInfo::none);
1636 __ emit_data64(0x8000000080000000, relocInfo::none);
1637 __ emit_data64(0x8000000080000000, relocInfo::none);
1638 __ emit_data64(0x8000000080000000, relocInfo::none);
1639
1640 return start;
1641 }
1642
1643 address StubGenerator::base64_avx2_lut_addr() {
1644 __ align32();
1645 StubCodeMark mark(this, "StubRoutines", "avx2_lut_base64");
1646 address start = __ pc();
1647
1648 __ emit_data64(0xfcfcfcfcfcfc4741, relocInfo::none);
1649 __ emit_data64(0x0000f0edfcfcfcfc, relocInfo::none);
1650 __ emit_data64(0xfcfcfcfcfcfc4741, relocInfo::none);
1651 __ emit_data64(0x0000f0edfcfcfcfc, relocInfo::none);
1652
1653 // URL LUT
1654 __ emit_data64(0xfcfcfcfcfcfc4741, relocInfo::none);
1655 __ emit_data64(0x000020effcfcfcfc, relocInfo::none);
1656 __ emit_data64(0xfcfcfcfcfcfc4741, relocInfo::none);
1657 __ emit_data64(0x000020effcfcfcfc, relocInfo::none);
1658
1659 return start;
1660 }
1661
1662 address StubGenerator::base64_encoding_table_addr() {
1663 __ align64();
1664 StubCodeMark mark(this, "StubRoutines", "encoding_table_base64");
1665 address start = __ pc();
1666
1667 assert(((unsigned long long)start & 0x3f) == 0, "Alignment problem (0x%08llx)", (unsigned long long)start);
1668 __ emit_data64(0x4847464544434241, relocInfo::none);
1669 __ emit_data64(0x504f4e4d4c4b4a49, relocInfo::none);
1670 __ emit_data64(0x5857565554535251, relocInfo::none);
1671 __ emit_data64(0x6665646362615a59, relocInfo::none);
1672 __ emit_data64(0x6e6d6c6b6a696867, relocInfo::none);
1673 __ emit_data64(0x767574737271706f, relocInfo::none);
1674 __ emit_data64(0x333231307a797877, relocInfo::none);
1675 __ emit_data64(0x2f2b393837363534, relocInfo::none);
1676
1677 // URL table
1678 __ emit_data64(0x4847464544434241, relocInfo::none);
1679 __ emit_data64(0x504f4e4d4c4b4a49, relocInfo::none);
1680 __ emit_data64(0x5857565554535251, relocInfo::none);
1681 __ emit_data64(0x6665646362615a59, relocInfo::none);
1682 __ emit_data64(0x6e6d6c6b6a696867, relocInfo::none);
1683 __ emit_data64(0x767574737271706f, relocInfo::none);
1684 __ emit_data64(0x333231307a797877, relocInfo::none);
1685 __ emit_data64(0x5f2d393837363534, relocInfo::none);
1686
1687 return start;
1688 }
1689
1690 // Code for generating Base64 encoding.
1691 // Intrinsic function prototype in Base64.java:
1692 // private void encodeBlock(byte[] src, int sp, int sl, byte[] dst, int dp,
1693 // boolean isURL) {
1694 address StubGenerator::generate_base64_encodeBlock()
1695 {
1696 __ align(CodeEntryAlignment);
1697 StubCodeMark mark(this, "StubRoutines", "implEncode");
1698 address start = __ pc();
1699
1700 __ enter();
1701
1702 // Save callee-saved registers before using them
1703 __ push(r12);
1704 __ push(r13);
1705 __ push(r14);
1706 __ push(r15);
1707
1708 // arguments
1709 const Register source = c_rarg0; // Source Array
1710 const Register start_offset = c_rarg1; // start offset
1711 const Register end_offset = c_rarg2; // end offset
1712 const Register dest = c_rarg3; // destination array
1713
1714 #ifndef _WIN64
1715 const Register dp = c_rarg4; // Position for writing to dest array
1716 const Register isURL = c_rarg5; // Base64 or URL character set
1717 #else
1718 const Address dp_mem(rbp, 6 * wordSize); // length is on stack on Win64
1719 const Address isURL_mem(rbp, 7 * wordSize);
1720 const Register isURL = r10; // pick the volatile windows register
1721 const Register dp = r12;
1722 __ movl(dp, dp_mem);
1723 __ movl(isURL, isURL_mem);
1724 #endif
1725
1726 const Register length = r14;
1727 const Register encode_table = r13;
1728 Label L_process3, L_exit, L_processdata, L_vbmiLoop, L_not512, L_32byteLoop;
1729
1730 // calculate length from offsets
1731 __ movl(length, end_offset);
1732 __ subl(length, start_offset);
1733 __ jcc(Assembler::lessEqual, L_exit);
1734
1735 // Code for 512-bit VBMI encoding. Encodes 48 input bytes into 64
1736 // output bytes. We read 64 input bytes and ignore the last 16, so be
1737 // sure not to read past the end of the input buffer.
1738 if (VM_Version::supports_avx512_vbmi()) {
1739 __ cmpl(length, 64); // Do not overrun input buffer.
1740 __ jcc(Assembler::below, L_not512);
1741
1742 __ shll(isURL, 6); // index into decode table based on isURL
1743 __ lea(encode_table, ExternalAddress(StubRoutines::x86::base64_encoding_table_addr()));
1744 __ addptr(encode_table, isURL);
1745 __ shrl(isURL, 6); // restore isURL
1746
1747 __ mov64(rax, 0x3036242a1016040aull); // Shifts
1748 __ evmovdquq(xmm3, ExternalAddress(StubRoutines::x86::base64_shuffle_addr()), Assembler::AVX_512bit, r15);
1749 __ evmovdquq(xmm2, Address(encode_table, 0), Assembler::AVX_512bit);
1750 __ evpbroadcastq(xmm1, rax, Assembler::AVX_512bit);
1751
1752 __ align32();
1753 __ BIND(L_vbmiLoop);
1754
1755 __ vpermb(xmm0, xmm3, Address(source, start_offset), Assembler::AVX_512bit);
1756 __ subl(length, 48);
1757
1758 // Put the input bytes into the proper lanes for writing, then
1759 // encode them.
1760 __ evpmultishiftqb(xmm0, xmm1, xmm0, Assembler::AVX_512bit);
1761 __ vpermb(xmm0, xmm0, xmm2, Assembler::AVX_512bit);
1762
1763 // Write to destination
1764 __ evmovdquq(Address(dest, dp), xmm0, Assembler::AVX_512bit);
1765
1766 __ addptr(dest, 64);
1767 __ addptr(source, 48);
1768 __ cmpl(length, 64);
1769 __ jcc(Assembler::aboveEqual, L_vbmiLoop);
1770
1771 __ vzeroupper();
1772 }
1773
1774 __ BIND(L_not512);
1775 if (VM_Version::supports_avx2()) {
1776 /*
1777 ** This AVX2 encoder is based off the paper at:
1778 ** https://dl.acm.org/doi/10.1145/3132709
1779 **
1780 ** We use AVX2 SIMD instructions to encode 24 bytes into 32
1781 ** output bytes.
1782 **
1783 */
1784 // Lengths under 32 bytes are done with scalar routine
1785 __ cmpl(length, 31);
1786 __ jcc(Assembler::belowEqual, L_process3);
1787
1788 // Set up supporting constant table data
1789 __ vmovdqu(xmm9, ExternalAddress(StubRoutines::x86::base64_avx2_shuffle_addr()), rax);
1790 // 6-bit mask for 2nd and 4th (and multiples) 6-bit values
1791 __ movl(rax, 0x0fc0fc00);
1792 __ movdl(xmm8, rax);
1793 __ vmovdqu(xmm1, ExternalAddress(StubRoutines::x86::base64_avx2_input_mask_addr()), rax);
1794 __ vpbroadcastd(xmm8, xmm8, Assembler::AVX_256bit);
1795
1796 // Multiplication constant for "shifting" right by 6 and 10
1797 // bits
1798 __ movl(rax, 0x04000040);
1799
1800 __ subl(length, 24);
1801 __ movdl(xmm7, rax);
1802 __ vpbroadcastd(xmm7, xmm7, Assembler::AVX_256bit);
1803
1804 // For the first load, we mask off reading of the first 4
1805 // bytes into the register. This is so we can get 4 3-byte
1806 // chunks into each lane of the register, avoiding having to
1807 // handle end conditions. We then shuffle these bytes into a
1808 // specific order so that manipulation is easier.
1809 //
1810 // The initial read loads the XMM register like this:
1811 //
1812 // Lower 128-bit lane:
1813 // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
1814 // | XX | XX | XX | XX | A0 | A1 | A2 | B0 | B1 | B2 | C0 | C1
1815 // | C2 | D0 | D1 | D2 |
1816 // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
1817 //
1818 // Upper 128-bit lane:
1819 // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
1820 // | E0 | E1 | E2 | F0 | F1 | F2 | G0 | G1 | G2 | H0 | H1 | H2
1821 // | XX | XX | XX | XX |
1822 // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
1823 //
1824 // Where A0 is the first input byte, B0 is the fourth, etc.
1825 // The alphabetical significance denotes the 3 bytes to be
1826 // consumed and encoded into 4 bytes.
1827 //
1828 // We then shuffle the register so each 32-bit word contains
1829 // the sequence:
1830 // A1 A0 A2 A1, B1, B0, B2, B1, etc.
1831 // Each of these byte sequences are then manipulated into 4
1832 // 6-bit values ready for encoding.
1833 //
1834 // If we focus on one set of 3-byte chunks, changing the
1835 // nomenclature such that A0 => a, A1 => b, and A2 => c, we
1836 // shuffle such that each 24-bit chunk contains:
1837 //
1838 // b7 b6 b5 b4 b3 b2 b1 b0 | a7 a6 a5 a4 a3 a2 a1 a0 | c7 c6
1839 // c5 c4 c3 c2 c1 c0 | b7 b6 b5 b4 b3 b2 b1 b0
1840 // Explain this step.
1841 // b3 b2 b1 b0 c5 c4 c3 c2 | c1 c0 d5 d4 d3 d2 d1 d0 | a5 a4
1842 // a3 a2 a1 a0 b5 b4 | b3 b2 b1 b0 c5 c4 c3 c2
1843 //
1844 // W first and off all but bits 4-9 and 16-21 (c5..c0 and
1845 // a5..a0) and shift them using a vector multiplication
1846 // operation (vpmulhuw) which effectively shifts c right by 6
1847 // bits and a right by 10 bits. We similarly mask bits 10-15
1848 // (d5..d0) and 22-27 (b5..b0) and shift them left by 8 and 4
1849 // bits respectively. This is done using vpmullw. We end up
1850 // with 4 6-bit values, thus splitting the 3 input bytes,
1851 // ready for encoding:
1852 // 0 0 d5..d0 0 0 c5..c0 0 0 b5..b0 0 0 a5..a0
1853 //
1854 // For translation, we recognize that there are 5 distinct
1855 // ranges of legal Base64 characters as below:
1856 //
1857 // +-------------+-------------+------------+
1858 // | 6-bit value | ASCII range | offset |
1859 // +-------------+-------------+------------+
1860 // | 0..25 | A..Z | 65 |
1861 // | 26..51 | a..z | 71 |
1862 // | 52..61 | 0..9 | -4 |
1863 // | 62 | + or - | -19 or -17 |
1864 // | 63 | / or _ | -16 or 32 |
1865 // +-------------+-------------+------------+
1866 //
1867 // We note that vpshufb does a parallel lookup in a
1868 // destination register using the lower 4 bits of bytes from a
1869 // source register. If we use a saturated subtraction and
1870 // subtract 51 from each 6-bit value, bytes from [0,51]
1871 // saturate to 0, and [52,63] map to a range of [1,12]. We
1872 // distinguish the [0,25] and [26,51] ranges by assigning a
1873 // value of 13 for all 6-bit values less than 26. We end up
1874 // with:
1875 //
1876 // +-------------+-------------+------------+
1877 // | 6-bit value | Reduced | offset |
1878 // +-------------+-------------+------------+
1879 // | 0..25 | 13 | 65 |
1880 // | 26..51 | 0 | 71 |
1881 // | 52..61 | 0..9 | -4 |
1882 // | 62 | 11 | -19 or -17 |
1883 // | 63 | 12 | -16 or 32 |
1884 // +-------------+-------------+------------+
1885 //
1886 // We then use a final vpshufb to add the appropriate offset,
1887 // translating the bytes.
1888 //
1889 // Load input bytes - only 28 bytes. Mask the first load to
1890 // not load into the full register.
1891 __ vpmaskmovd(xmm1, xmm1, Address(source, start_offset, Address::times_1, -4), Assembler::AVX_256bit);
1892
1893 // Move 3-byte chunks of input (12 bytes) into 16 bytes,
1894 // ordering by:
1895 // 1, 0, 2, 1; 4, 3, 5, 4; etc. This groups 6-bit chunks
1896 // for easy masking
1897 __ vpshufb(xmm1, xmm1, xmm9, Assembler::AVX_256bit);
1898
1899 __ addl(start_offset, 24);
1900
1901 // Load masking register for first and third (and multiples)
1902 // 6-bit values.
1903 __ movl(rax, 0x003f03f0);
1904 __ movdl(xmm6, rax);
1905 __ vpbroadcastd(xmm6, xmm6, Assembler::AVX_256bit);
1906 // Multiplication constant for "shifting" left by 4 and 8 bits
1907 __ movl(rax, 0x01000010);
1908 __ movdl(xmm5, rax);
1909 __ vpbroadcastd(xmm5, xmm5, Assembler::AVX_256bit);
1910
1911 // Isolate 6-bit chunks of interest
1912 __ vpand(xmm0, xmm8, xmm1, Assembler::AVX_256bit);
1913
1914 // Load constants for encoding
1915 __ movl(rax, 0x19191919);
1916 __ movdl(xmm3, rax);
1917 __ vpbroadcastd(xmm3, xmm3, Assembler::AVX_256bit);
1918 __ movl(rax, 0x33333333);
1919 __ movdl(xmm4, rax);
1920 __ vpbroadcastd(xmm4, xmm4, Assembler::AVX_256bit);
1921
1922 // Shift output bytes 0 and 2 into proper lanes
1923 __ vpmulhuw(xmm2, xmm0, xmm7, Assembler::AVX_256bit);
1924
1925 // Mask and shift output bytes 1 and 3 into proper lanes and
1926 // combine
1927 __ vpand(xmm0, xmm6, xmm1, Assembler::AVX_256bit);
1928 __ vpmullw(xmm0, xmm5, xmm0, Assembler::AVX_256bit);
1929 __ vpor(xmm0, xmm0, xmm2, Assembler::AVX_256bit);
1930
1931 // Find out which are 0..25. This indicates which input
1932 // values fall in the range of 'A'-'Z', which require an
1933 // additional offset (see comments above)
1934 __ vpcmpgtb(xmm2, xmm0, xmm3, Assembler::AVX_256bit);
1935 __ vpsubusb(xmm1, xmm0, xmm4, Assembler::AVX_256bit);
1936 __ vpsubb(xmm1, xmm1, xmm2, Assembler::AVX_256bit);
1937
1938 // Load the proper lookup table
1939 __ lea(r11, ExternalAddress(StubRoutines::x86::base64_avx2_lut_addr()));
1940 __ movl(r15, isURL);
1941 __ shll(r15, 5);
1942 __ vmovdqu(xmm2, Address(r11, r15));
1943
1944 // Shuffle the offsets based on the range calculation done
1945 // above. This allows us to add the correct offset to the
1946 // 6-bit value corresponding to the range documented above.
1947 __ vpshufb(xmm1, xmm2, xmm1, Assembler::AVX_256bit);
1948 __ vpaddb(xmm0, xmm1, xmm0, Assembler::AVX_256bit);
1949
1950 // Store the encoded bytes
1951 __ vmovdqu(Address(dest, dp), xmm0);
1952 __ addl(dp, 32);
1953
1954 __ cmpl(length, 31);
1955 __ jcc(Assembler::belowEqual, L_process3);
1956
1957 __ align32();
1958 __ BIND(L_32byteLoop);
1959
1960 // Get next 32 bytes
1961 __ vmovdqu(xmm1, Address(source, start_offset, Address::times_1, -4));
1962
1963 __ subl(length, 24);
1964 __ addl(start_offset, 24);
1965
1966 // This logic is identical to the above, with only constant
1967 // register loads removed. Shuffle the input, mask off 6-bit
1968 // chunks, shift them into place, then add the offset to
1969 // encode.
1970 __ vpshufb(xmm1, xmm1, xmm9, Assembler::AVX_256bit);
1971
1972 __ vpand(xmm0, xmm8, xmm1, Assembler::AVX_256bit);
1973 __ vpmulhuw(xmm10, xmm0, xmm7, Assembler::AVX_256bit);
1974 __ vpand(xmm0, xmm6, xmm1, Assembler::AVX_256bit);
1975 __ vpmullw(xmm0, xmm5, xmm0, Assembler::AVX_256bit);
1976 __ vpor(xmm0, xmm0, xmm10, Assembler::AVX_256bit);
1977 __ vpcmpgtb(xmm10, xmm0, xmm3, Assembler::AVX_256bit);
1978 __ vpsubusb(xmm1, xmm0, xmm4, Assembler::AVX_256bit);
1979 __ vpsubb(xmm1, xmm1, xmm10, Assembler::AVX_256bit);
1980 __ vpshufb(xmm1, xmm2, xmm1, Assembler::AVX_256bit);
1981 __ vpaddb(xmm0, xmm1, xmm0, Assembler::AVX_256bit);
1982
1983 // Store the encoded bytes
1984 __ vmovdqu(Address(dest, dp), xmm0);
1985 __ addl(dp, 32);
1986
1987 __ cmpl(length, 31);
1988 __ jcc(Assembler::above, L_32byteLoop);
1989
1990 __ BIND(L_process3);
1991 __ vzeroupper();
1992 } else {
1993 __ BIND(L_process3);
1994 }
1995
1996 __ cmpl(length, 3);
1997 __ jcc(Assembler::below, L_exit);
1998
1999 // Load the encoding table based on isURL
2000 __ lea(r11, ExternalAddress(StubRoutines::x86::base64_encoding_table_addr()));
2001 __ movl(r15, isURL);
2002 __ shll(r15, 6);
2003 __ addptr(r11, r15);
2004
2005 __ BIND(L_processdata);
2006
2007 // Load 3 bytes
2008 __ load_unsigned_byte(r15, Address(source, start_offset));
2009 __ load_unsigned_byte(r10, Address(source, start_offset, Address::times_1, 1));
2010 __ load_unsigned_byte(r13, Address(source, start_offset, Address::times_1, 2));
2011
2012 // Build a 32-bit word with bytes 1, 2, 0, 1
2013 __ movl(rax, r10);
2014 __ shll(r10, 24);
2015 __ orl(rax, r10);
2016
2017 __ subl(length, 3);
2018
2019 __ shll(r15, 8);
2020 __ shll(r13, 16);
2021 __ orl(rax, r15);
2022
2023 __ addl(start_offset, 3);
2024
2025 __ orl(rax, r13);
2026 // At this point, rax contains | byte1 | byte2 | byte0 | byte1
2027 // r13 has byte2 << 16 - need low-order 6 bits to translate.
2028 // This translated byte is the fourth output byte.
2029 __ shrl(r13, 16);
2030 __ andl(r13, 0x3f);
2031
2032 // The high-order 6 bits of r15 (byte0) is translated.
2033 // The translated byte is the first output byte.
2034 __ shrl(r15, 10);
2035
2036 __ load_unsigned_byte(r13, Address(r11, r13));
2037 __ load_unsigned_byte(r15, Address(r11, r15));
2038
2039 __ movb(Address(dest, dp, Address::times_1, 3), r13);
2040
2041 // Extract high-order 4 bits of byte1 and low-order 2 bits of byte0.
2042 // This translated byte is the second output byte.
2043 __ shrl(rax, 4);
2044 __ movl(r10, rax);
2045 __ andl(rax, 0x3f);
2046
2047 __ movb(Address(dest, dp, Address::times_1, 0), r15);
2048
2049 __ load_unsigned_byte(rax, Address(r11, rax));
2050
2051 // Extract low-order 2 bits of byte1 and high-order 4 bits of byte2.
2052 // This translated byte is the third output byte.
2053 __ shrl(r10, 18);
2054 __ andl(r10, 0x3f);
2055
2056 __ load_unsigned_byte(r10, Address(r11, r10));
2057
2058 __ movb(Address(dest, dp, Address::times_1, 1), rax);
2059 __ movb(Address(dest, dp, Address::times_1, 2), r10);
2060
2061 __ addl(dp, 4);
2062 __ cmpl(length, 3);
2063 __ jcc(Assembler::aboveEqual, L_processdata);
2064
2065 __ BIND(L_exit);
2066 __ pop(r15);
2067 __ pop(r14);
2068 __ pop(r13);
2069 __ pop(r12);
2070 __ leave();
2071 __ ret(0);
2072
2073 return start;
2074 }
2075
2076 // base64 AVX512vbmi tables
2077 address StubGenerator::base64_vbmi_lookup_lo_addr() {
2078 __ align64();
2079 StubCodeMark mark(this, "StubRoutines", "lookup_lo_base64");
2080 address start = __ pc();
2081
2082 assert(((unsigned long long)start & 0x3f) == 0,
2083 "Alignment problem (0x%08llx)", (unsigned long long)start);
2084 __ emit_data64(0x8080808080808080, relocInfo::none);
2085 __ emit_data64(0x8080808080808080, relocInfo::none);
2086 __ emit_data64(0x8080808080808080, relocInfo::none);
2087 __ emit_data64(0x8080808080808080, relocInfo::none);
2088 __ emit_data64(0x8080808080808080, relocInfo::none);
2089 __ emit_data64(0x3f8080803e808080, relocInfo::none);
2090 __ emit_data64(0x3b3a393837363534, relocInfo::none);
2091 __ emit_data64(0x8080808080803d3c, relocInfo::none);
2092
2093 return start;
2094 }
2095
2096 address StubGenerator::base64_vbmi_lookup_hi_addr() {
2097 __ align64();
2098 StubCodeMark mark(this, "StubRoutines", "lookup_hi_base64");
2099 address start = __ pc();
2100
2101 assert(((unsigned long long)start & 0x3f) == 0,
2102 "Alignment problem (0x%08llx)", (unsigned long long)start);
2103 __ emit_data64(0x0605040302010080, relocInfo::none);
2104 __ emit_data64(0x0e0d0c0b0a090807, relocInfo::none);
2105 __ emit_data64(0x161514131211100f, relocInfo::none);
2106 __ emit_data64(0x8080808080191817, relocInfo::none);
2107 __ emit_data64(0x201f1e1d1c1b1a80, relocInfo::none);
2108 __ emit_data64(0x2827262524232221, relocInfo::none);
2109 __ emit_data64(0x302f2e2d2c2b2a29, relocInfo::none);
2110 __ emit_data64(0x8080808080333231, relocInfo::none);
2111
2112 return start;
2113 }
2114 address StubGenerator::base64_vbmi_lookup_lo_url_addr() {
2115 __ align64();
2116 StubCodeMark mark(this, "StubRoutines", "lookup_lo_base64url");
2117 address start = __ pc();
2118
2119 assert(((unsigned long long)start & 0x3f) == 0,
2120 "Alignment problem (0x%08llx)", (unsigned long long)start);
2121 __ emit_data64(0x8080808080808080, relocInfo::none);
2122 __ emit_data64(0x8080808080808080, relocInfo::none);
2123 __ emit_data64(0x8080808080808080, relocInfo::none);
2124 __ emit_data64(0x8080808080808080, relocInfo::none);
2125 __ emit_data64(0x8080808080808080, relocInfo::none);
2126 __ emit_data64(0x80803e8080808080, relocInfo::none);
2127 __ emit_data64(0x3b3a393837363534, relocInfo::none);
2128 __ emit_data64(0x8080808080803d3c, relocInfo::none);
2129
2130 return start;
2131 }
2132
2133 address StubGenerator::base64_vbmi_lookup_hi_url_addr() {
2134 __ align64();
2135 StubCodeMark mark(this, "StubRoutines", "lookup_hi_base64url");
2136 address start = __ pc();
2137
2138 assert(((unsigned long long)start & 0x3f) == 0,
2139 "Alignment problem (0x%08llx)", (unsigned long long)start);
2140 __ emit_data64(0x0605040302010080, relocInfo::none);
2141 __ emit_data64(0x0e0d0c0b0a090807, relocInfo::none);
2142 __ emit_data64(0x161514131211100f, relocInfo::none);
2143 __ emit_data64(0x3f80808080191817, relocInfo::none);
2144 __ emit_data64(0x201f1e1d1c1b1a80, relocInfo::none);
2145 __ emit_data64(0x2827262524232221, relocInfo::none);
2146 __ emit_data64(0x302f2e2d2c2b2a29, relocInfo::none);
2147 __ emit_data64(0x8080808080333231, relocInfo::none);
2148
2149 return start;
2150 }
2151
2152 address StubGenerator::base64_vbmi_pack_vec_addr() {
2153 __ align64();
2154 StubCodeMark mark(this, "StubRoutines", "pack_vec_base64");
2155 address start = __ pc();
2156
2157 assert(((unsigned long long)start & 0x3f) == 0,
2158 "Alignment problem (0x%08llx)", (unsigned long long)start);
2159 __ emit_data64(0x090a040506000102, relocInfo::none);
2160 __ emit_data64(0x161011120c0d0e08, relocInfo::none);
2161 __ emit_data64(0x1c1d1e18191a1415, relocInfo::none);
2162 __ emit_data64(0x292a242526202122, relocInfo::none);
2163 __ emit_data64(0x363031322c2d2e28, relocInfo::none);
2164 __ emit_data64(0x3c3d3e38393a3435, relocInfo::none);
2165 __ emit_data64(0x0000000000000000, relocInfo::none);
2166 __ emit_data64(0x0000000000000000, relocInfo::none);
2167
2168 return start;
2169 }
2170
2171 address StubGenerator::base64_vbmi_join_0_1_addr() {
2172 __ align64();
2173 StubCodeMark mark(this, "StubRoutines", "join_0_1_base64");
2174 address start = __ pc();
2175
2176 assert(((unsigned long long)start & 0x3f) == 0,
2177 "Alignment problem (0x%08llx)", (unsigned long long)start);
2178 __ emit_data64(0x090a040506000102, relocInfo::none);
2179 __ emit_data64(0x161011120c0d0e08, relocInfo::none);
2180 __ emit_data64(0x1c1d1e18191a1415, relocInfo::none);
2181 __ emit_data64(0x292a242526202122, relocInfo::none);
2182 __ emit_data64(0x363031322c2d2e28, relocInfo::none);
2183 __ emit_data64(0x3c3d3e38393a3435, relocInfo::none);
2184 __ emit_data64(0x494a444546404142, relocInfo::none);
2185 __ emit_data64(0x565051524c4d4e48, relocInfo::none);
2186
2187 return start;
2188 }
2189
2190 address StubGenerator::base64_vbmi_join_1_2_addr() {
2191 __ align64();
2192 StubCodeMark mark(this, "StubRoutines", "join_1_2_base64");
2193 address start = __ pc();
2194
2195 assert(((unsigned long long)start & 0x3f) == 0,
2196 "Alignment problem (0x%08llx)", (unsigned long long)start);
2197 __ emit_data64(0x1c1d1e18191a1415, relocInfo::none);
2198 __ emit_data64(0x292a242526202122, relocInfo::none);
2199 __ emit_data64(0x363031322c2d2e28, relocInfo::none);
2200 __ emit_data64(0x3c3d3e38393a3435, relocInfo::none);
2201 __ emit_data64(0x494a444546404142, relocInfo::none);
2202 __ emit_data64(0x565051524c4d4e48, relocInfo::none);
2203 __ emit_data64(0x5c5d5e58595a5455, relocInfo::none);
2204 __ emit_data64(0x696a646566606162, relocInfo::none);
2205
2206 return start;
2207 }
2208
2209 address StubGenerator::base64_vbmi_join_2_3_addr() {
2210 __ align64();
2211 StubCodeMark mark(this, "StubRoutines", "join_2_3_base64");
2212 address start = __ pc();
2213
2214 assert(((unsigned long long)start & 0x3f) == 0,
2215 "Alignment problem (0x%08llx)", (unsigned long long)start);
2216 __ emit_data64(0x363031322c2d2e28, relocInfo::none);
2217 __ emit_data64(0x3c3d3e38393a3435, relocInfo::none);
2218 __ emit_data64(0x494a444546404142, relocInfo::none);
2219 __ emit_data64(0x565051524c4d4e48, relocInfo::none);
2220 __ emit_data64(0x5c5d5e58595a5455, relocInfo::none);
2221 __ emit_data64(0x696a646566606162, relocInfo::none);
2222 __ emit_data64(0x767071726c6d6e68, relocInfo::none);
2223 __ emit_data64(0x7c7d7e78797a7475, relocInfo::none);
2224
2225 return start;
2226 }
2227
2228 address StubGenerator::base64_AVX2_decode_tables_addr() {
2229 __ align64();
2230 StubCodeMark mark(this, "StubRoutines", "AVX2_tables_base64");
2231 address start = __ pc();
2232
2233 assert(((unsigned long long)start & 0x3f) == 0,
2234 "Alignment problem (0x%08llx)", (unsigned long long)start);
2235 __ emit_data(0x2f2f2f2f, relocInfo::none, 0);
2236 __ emit_data(0x5f5f5f5f, relocInfo::none, 0); // for URL
2237
2238 __ emit_data(0xffffffff, relocInfo::none, 0);
2239 __ emit_data(0xfcfcfcfc, relocInfo::none, 0); // for URL
2240
2241 // Permute table
2242 __ emit_data64(0x0000000100000000, relocInfo::none);
2243 __ emit_data64(0x0000000400000002, relocInfo::none);
2244 __ emit_data64(0x0000000600000005, relocInfo::none);
2245 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2246
2247 // Shuffle table
2248 __ emit_data64(0x090a040506000102, relocInfo::none);
2249 __ emit_data64(0xffffffff0c0d0e08, relocInfo::none);
2250 __ emit_data64(0x090a040506000102, relocInfo::none);
2251 __ emit_data64(0xffffffff0c0d0e08, relocInfo::none);
2252
2253 // merge table
2254 __ emit_data(0x01400140, relocInfo::none, 0);
2255
2256 // merge multiplier
2257 __ emit_data(0x00011000, relocInfo::none, 0);
2258
2259 return start;
2260 }
2261
2262 address StubGenerator::base64_AVX2_decode_LUT_tables_addr() {
2263 __ align64();
2264 StubCodeMark mark(this, "StubRoutines", "AVX2_tables_URL_base64");
2265 address start = __ pc();
2266
2267 assert(((unsigned long long)start & 0x3f) == 0,
2268 "Alignment problem (0x%08llx)", (unsigned long long)start);
2269 // lut_lo
2270 __ emit_data64(0x1111111111111115, relocInfo::none);
2271 __ emit_data64(0x1a1b1b1b1a131111, relocInfo::none);
2272 __ emit_data64(0x1111111111111115, relocInfo::none);
2273 __ emit_data64(0x1a1b1b1b1a131111, relocInfo::none);
2274
2275 // lut_roll
2276 __ emit_data64(0xb9b9bfbf04131000, relocInfo::none);
2277 __ emit_data64(0x0000000000000000, relocInfo::none);
2278 __ emit_data64(0xb9b9bfbf04131000, relocInfo::none);
2279 __ emit_data64(0x0000000000000000, relocInfo::none);
2280
2281 // lut_lo URL
2282 __ emit_data64(0x1111111111111115, relocInfo::none);
2283 __ emit_data64(0x1b1b1a1b1b131111, relocInfo::none);
2284 __ emit_data64(0x1111111111111115, relocInfo::none);
2285 __ emit_data64(0x1b1b1a1b1b131111, relocInfo::none);
2286
2287 // lut_roll URL
2288 __ emit_data64(0xb9b9bfbf0411e000, relocInfo::none);
2289 __ emit_data64(0x0000000000000000, relocInfo::none);
2290 __ emit_data64(0xb9b9bfbf0411e000, relocInfo::none);
2291 __ emit_data64(0x0000000000000000, relocInfo::none);
2292
2293 // lut_hi
2294 __ emit_data64(0x0804080402011010, relocInfo::none);
2295 __ emit_data64(0x1010101010101010, relocInfo::none);
2296 __ emit_data64(0x0804080402011010, relocInfo::none);
2297 __ emit_data64(0x1010101010101010, relocInfo::none);
2298
2299 return start;
2300 }
2301
2302 address StubGenerator::base64_decoding_table_addr() {
2303 StubCodeMark mark(this, "StubRoutines", "decoding_table_base64");
2304 address start = __ pc();
2305
2306 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2307 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2308 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2309 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2310 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2311 __ emit_data64(0x3fffffff3effffff, relocInfo::none);
2312 __ emit_data64(0x3b3a393837363534, relocInfo::none);
2313 __ emit_data64(0xffffffffffff3d3c, relocInfo::none);
2314 __ emit_data64(0x06050403020100ff, relocInfo::none);
2315 __ emit_data64(0x0e0d0c0b0a090807, relocInfo::none);
2316 __ emit_data64(0x161514131211100f, relocInfo::none);
2317 __ emit_data64(0xffffffffff191817, relocInfo::none);
2318 __ emit_data64(0x201f1e1d1c1b1aff, relocInfo::none);
2319 __ emit_data64(0x2827262524232221, relocInfo::none);
2320 __ emit_data64(0x302f2e2d2c2b2a29, relocInfo::none);
2321 __ emit_data64(0xffffffffff333231, relocInfo::none);
2322 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2323 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2324 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2325 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2326 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2327 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2328 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2329 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2330 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2331 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2332 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2333 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2334 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2335 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2336 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2337 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2338
2339 // URL table
2340 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2341 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2342 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2343 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2344 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2345 __ emit_data64(0xffff3effffffffff, relocInfo::none);
2346 __ emit_data64(0x3b3a393837363534, relocInfo::none);
2347 __ emit_data64(0xffffffffffff3d3c, relocInfo::none);
2348 __ emit_data64(0x06050403020100ff, relocInfo::none);
2349 __ emit_data64(0x0e0d0c0b0a090807, relocInfo::none);
2350 __ emit_data64(0x161514131211100f, relocInfo::none);
2351 __ emit_data64(0x3fffffffff191817, relocInfo::none);
2352 __ emit_data64(0x201f1e1d1c1b1aff, relocInfo::none);
2353 __ emit_data64(0x2827262524232221, relocInfo::none);
2354 __ emit_data64(0x302f2e2d2c2b2a29, relocInfo::none);
2355 __ emit_data64(0xffffffffff333231, relocInfo::none);
2356 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2357 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2358 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2359 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2360 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2361 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2362 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2363 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2364 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2365 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2366 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2367 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2368 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2369 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2370 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2371 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2372
2373 return start;
2374 }
2375
2376
2377 // Code for generating Base64 decoding.
2378 //
2379 // Based on the article (and associated code) from https://arxiv.org/abs/1910.05109.
2380 //
2381 // Intrinsic function prototype in Base64.java:
2382 // private void decodeBlock(byte[] src, int sp, int sl, byte[] dst, int dp, boolean isURL, isMIME) {
2383 address StubGenerator::generate_base64_decodeBlock() {
2384 __ align(CodeEntryAlignment);
2385 StubCodeMark mark(this, "StubRoutines", "implDecode");
2386 address start = __ pc();
2387
2388 __ enter();
2389
2390 // Save callee-saved registers before using them
2391 __ push(r12);
2392 __ push(r13);
2393 __ push(r14);
2394 __ push(r15);
2395 __ push(rbx);
2396
2397 // arguments
2398 const Register source = c_rarg0; // Source Array
2399 const Register start_offset = c_rarg1; // start offset
2400 const Register end_offset = c_rarg2; // end offset
2401 const Register dest = c_rarg3; // destination array
2402 const Register isMIME = rbx;
2403
2404 #ifndef _WIN64
2405 const Register dp = c_rarg4; // Position for writing to dest array
2406 const Register isURL = c_rarg5;// Base64 or URL character set
2407 __ movl(isMIME, Address(rbp, 2 * wordSize));
2408 #else
2409 const Address dp_mem(rbp, 6 * wordSize); // length is on stack on Win64
2410 const Address isURL_mem(rbp, 7 * wordSize);
2411 const Register isURL = r10; // pick the volatile windows register
2412 const Register dp = r12;
2413 __ movl(dp, dp_mem);
2414 __ movl(isURL, isURL_mem);
2415 __ movl(isMIME, Address(rbp, 8 * wordSize));
2416 #endif
2417
2418 const XMMRegister lookup_lo = xmm5;
2419 const XMMRegister lookup_hi = xmm6;
2420 const XMMRegister errorvec = xmm7;
2421 const XMMRegister pack16_op = xmm9;
2422 const XMMRegister pack32_op = xmm8;
2423 const XMMRegister input0 = xmm3;
2424 const XMMRegister input1 = xmm20;
2425 const XMMRegister input2 = xmm21;
2426 const XMMRegister input3 = xmm19;
2427 const XMMRegister join01 = xmm12;
2428 const XMMRegister join12 = xmm11;
2429 const XMMRegister join23 = xmm10;
2430 const XMMRegister translated0 = xmm2;
2431 const XMMRegister translated1 = xmm1;
2432 const XMMRegister translated2 = xmm0;
2433 const XMMRegister translated3 = xmm4;
2434
2435 const XMMRegister merged0 = xmm2;
2436 const XMMRegister merged1 = xmm1;
2437 const XMMRegister merged2 = xmm0;
2438 const XMMRegister merged3 = xmm4;
2439 const XMMRegister merge_ab_bc0 = xmm2;
2440 const XMMRegister merge_ab_bc1 = xmm1;
2441 const XMMRegister merge_ab_bc2 = xmm0;
2442 const XMMRegister merge_ab_bc3 = xmm4;
2443
2444 const XMMRegister pack24bits = xmm4;
2445
2446 const Register length = r14;
2447 const Register output_size = r13;
2448 const Register output_mask = r15;
2449 const KRegister input_mask = k1;
2450
2451 const XMMRegister input_initial_valid_b64 = xmm0;
2452 const XMMRegister tmp = xmm10;
2453 const XMMRegister mask = xmm0;
2454 const XMMRegister invalid_b64 = xmm1;
2455
2456 Label L_process256, L_process64, L_process64Loop, L_exit, L_processdata, L_loadURL;
2457 Label L_continue, L_finalBit, L_padding, L_donePadding, L_bruteForce;
2458 Label L_forceLoop, L_bottomLoop, L_checkMIME, L_exit_no_vzero, L_lastChunk;
2459
2460 // calculate length from offsets
2461 __ movl(length, end_offset);
2462 __ subl(length, start_offset);
2463 __ push(dest); // Save for return value calc
2464
2465 // If AVX512 VBMI not supported, just compile non-AVX code
2466 if(VM_Version::supports_avx512_vbmi() &&
2467 VM_Version::supports_avx512bw()) {
2468 __ cmpl(length, 31); // 32-bytes is break-even for AVX-512
2469 __ jcc(Assembler::lessEqual, L_lastChunk);
2470
2471 __ cmpl(isMIME, 0);
2472 __ jcc(Assembler::notEqual, L_lastChunk);
2473
2474 // Load lookup tables based on isURL
2475 __ cmpl(isURL, 0);
2476 __ jcc(Assembler::notZero, L_loadURL);
2477
2478 __ evmovdquq(lookup_lo, ExternalAddress(StubRoutines::x86::base64_vbmi_lookup_lo_addr()), Assembler::AVX_512bit, r13);
2479 __ evmovdquq(lookup_hi, ExternalAddress(StubRoutines::x86::base64_vbmi_lookup_hi_addr()), Assembler::AVX_512bit, r13);
2480
2481 __ BIND(L_continue);
2482
2483 __ movl(r15, 0x01400140);
2484 __ evpbroadcastd(pack16_op, r15, Assembler::AVX_512bit);
2485
2486 __ movl(r15, 0x00011000);
2487 __ evpbroadcastd(pack32_op, r15, Assembler::AVX_512bit);
2488
2489 __ cmpl(length, 0xff);
2490 __ jcc(Assembler::lessEqual, L_process64);
2491
2492 // load masks required for decoding data
2493 __ BIND(L_processdata);
2494 __ evmovdquq(join01, ExternalAddress(StubRoutines::x86::base64_vbmi_join_0_1_addr()), Assembler::AVX_512bit,r13);
2495 __ evmovdquq(join12, ExternalAddress(StubRoutines::x86::base64_vbmi_join_1_2_addr()), Assembler::AVX_512bit, r13);
2496 __ evmovdquq(join23, ExternalAddress(StubRoutines::x86::base64_vbmi_join_2_3_addr()), Assembler::AVX_512bit, r13);
2497
2498 __ align32();
2499 __ BIND(L_process256);
2500 // Grab input data
2501 __ evmovdquq(input0, Address(source, start_offset, Address::times_1, 0x00), Assembler::AVX_512bit);
2502 __ evmovdquq(input1, Address(source, start_offset, Address::times_1, 0x40), Assembler::AVX_512bit);
2503 __ evmovdquq(input2, Address(source, start_offset, Address::times_1, 0x80), Assembler::AVX_512bit);
2504 __ evmovdquq(input3, Address(source, start_offset, Address::times_1, 0xc0), Assembler::AVX_512bit);
2505
2506 // Copy the low part of the lookup table into the destination of the permutation
2507 __ evmovdquq(translated0, lookup_lo, Assembler::AVX_512bit);
2508 __ evmovdquq(translated1, lookup_lo, Assembler::AVX_512bit);
2509 __ evmovdquq(translated2, lookup_lo, Assembler::AVX_512bit);
2510 __ evmovdquq(translated3, lookup_lo, Assembler::AVX_512bit);
2511
2512 // Translate the base64 input into "decoded" bytes
2513 __ evpermt2b(translated0, input0, lookup_hi, Assembler::AVX_512bit);
2514 __ evpermt2b(translated1, input1, lookup_hi, Assembler::AVX_512bit);
2515 __ evpermt2b(translated2, input2, lookup_hi, Assembler::AVX_512bit);
2516 __ evpermt2b(translated3, input3, lookup_hi, Assembler::AVX_512bit);
2517
2518 // OR all of the translations together to check for errors (high-order bit of byte set)
2519 __ vpternlogd(input0, 0xfe, input1, input2, Assembler::AVX_512bit);
2520
2521 __ vpternlogd(input3, 0xfe, translated0, translated1, Assembler::AVX_512bit);
2522 __ vpternlogd(input0, 0xfe, translated2, translated3, Assembler::AVX_512bit);
2523 __ vpor(errorvec, input3, input0, Assembler::AVX_512bit);
2524
2525 // Check if there was an error - if so, try 64-byte chunks
2526 __ evpmovb2m(k3, errorvec, Assembler::AVX_512bit);
2527 __ kortestql(k3, k3);
2528 __ jcc(Assembler::notZero, L_process64);
2529
2530 // The merging and shuffling happens here
2531 // We multiply each byte pair [00dddddd | 00cccccc | 00bbbbbb | 00aaaaaa]
2532 // Multiply [00cccccc] by 2^6 added to [00dddddd] to get [0000cccc | ccdddddd]
2533 // The pack16_op is a vector of 0x01400140, so multiply D by 1 and C by 0x40
2534 __ vpmaddubsw(merge_ab_bc0, translated0, pack16_op, Assembler::AVX_512bit);
2535 __ vpmaddubsw(merge_ab_bc1, translated1, pack16_op, Assembler::AVX_512bit);
2536 __ vpmaddubsw(merge_ab_bc2, translated2, pack16_op, Assembler::AVX_512bit);
2537 __ vpmaddubsw(merge_ab_bc3, translated3, pack16_op, Assembler::AVX_512bit);
2538
2539 // Now do the same with packed 16-bit values.
2540 // We start with [0000cccc | ccdddddd | 0000aaaa | aabbbbbb]
2541 // pack32_op is 0x00011000 (2^12, 1), so this multiplies [0000aaaa | aabbbbbb] by 2^12
2542 // and adds [0000cccc | ccdddddd] to yield [00000000 | aaaaaabb | bbbbcccc | ccdddddd]
2543 __ vpmaddwd(merged0, merge_ab_bc0, pack32_op, Assembler::AVX_512bit);
2544 __ vpmaddwd(merged1, merge_ab_bc1, pack32_op, Assembler::AVX_512bit);
2545 __ vpmaddwd(merged2, merge_ab_bc2, pack32_op, Assembler::AVX_512bit);
2546 __ vpmaddwd(merged3, merge_ab_bc3, pack32_op, Assembler::AVX_512bit);
2547
2548 // The join vectors specify which byte from which vector goes into the outputs
2549 // One of every 4 bytes in the extended vector is zero, so we pack them into their
2550 // final positions in the register for storing (256 bytes in, 192 bytes out)
2551 __ evpermt2b(merged0, join01, merged1, Assembler::AVX_512bit);
2552 __ evpermt2b(merged1, join12, merged2, Assembler::AVX_512bit);
2553 __ evpermt2b(merged2, join23, merged3, Assembler::AVX_512bit);
2554
2555 // Store result
2556 __ evmovdquq(Address(dest, dp, Address::times_1, 0x00), merged0, Assembler::AVX_512bit);
2557 __ evmovdquq(Address(dest, dp, Address::times_1, 0x40), merged1, Assembler::AVX_512bit);
2558 __ evmovdquq(Address(dest, dp, Address::times_1, 0x80), merged2, Assembler::AVX_512bit);
2559
2560 __ addptr(source, 0x100);
2561 __ addptr(dest, 0xc0);
2562 __ subl(length, 0x100);
2563 __ cmpl(length, 64 * 4);
2564 __ jcc(Assembler::greaterEqual, L_process256);
2565
2566 // At this point, we've decoded 64 * 4 * n bytes.
2567 // The remaining length will be <= 64 * 4 - 1.
2568 // UNLESS there was an error decoding the first 256-byte chunk. In this
2569 // case, the length will be arbitrarily long.
2570 //
2571 // Note that this will be the path for MIME-encoded strings.
2572
2573 __ BIND(L_process64);
2574
2575 __ evmovdquq(pack24bits, ExternalAddress(StubRoutines::x86::base64_vbmi_pack_vec_addr()), Assembler::AVX_512bit, r13);
2576
2577 __ cmpl(length, 63);
2578 __ jcc(Assembler::lessEqual, L_finalBit);
2579
2580 __ mov64(rax, 0x0000ffffffffffff);
2581 __ kmovql(k2, rax);
2582
2583 __ align32();
2584 __ BIND(L_process64Loop);
2585
2586 // Handle first 64-byte block
2587
2588 __ evmovdquq(input0, Address(source, start_offset), Assembler::AVX_512bit);
2589 __ evmovdquq(translated0, lookup_lo, Assembler::AVX_512bit);
2590 __ evpermt2b(translated0, input0, lookup_hi, Assembler::AVX_512bit);
2591
2592 __ vpor(errorvec, translated0, input0, Assembler::AVX_512bit);
2593
2594 // Check for error and bomb out before updating dest
2595 __ evpmovb2m(k3, errorvec, Assembler::AVX_512bit);
2596 __ kortestql(k3, k3);
2597 __ jcc(Assembler::notZero, L_exit);
2598
2599 // Pack output register, selecting correct byte ordering
2600 __ vpmaddubsw(merge_ab_bc0, translated0, pack16_op, Assembler::AVX_512bit);
2601 __ vpmaddwd(merged0, merge_ab_bc0, pack32_op, Assembler::AVX_512bit);
2602 __ vpermb(merged0, pack24bits, merged0, Assembler::AVX_512bit);
2603
2604 __ evmovdqub(Address(dest, dp), k2, merged0, true, Assembler::AVX_512bit);
2605
2606 __ subl(length, 64);
2607 __ addptr(source, 64);
2608 __ addptr(dest, 48);
2609
2610 __ cmpl(length, 64);
2611 __ jcc(Assembler::greaterEqual, L_process64Loop);
2612
2613 __ cmpl(length, 0);
2614 __ jcc(Assembler::lessEqual, L_exit);
2615
2616 __ BIND(L_finalBit);
2617 // Now have 1 to 63 bytes left to decode
2618
2619 // I was going to let Java take care of the final fragment
2620 // however it will repeatedly call this routine for every 4 bytes
2621 // of input data, so handle the rest here.
2622 __ movq(rax, -1);
2623 __ bzhiq(rax, rax, length); // Input mask in rax
2624
2625 __ movl(output_size, length);
2626 __ shrl(output_size, 2); // Find (len / 4) * 3 (output length)
2627 __ lea(output_size, Address(output_size, output_size, Address::times_2, 0));
2628 // output_size in r13
2629
2630 // Strip pad characters, if any, and adjust length and mask
2631 __ addq(length, start_offset);
2632 __ cmpb(Address(source, length, Address::times_1, -1), '=');
2633 __ jcc(Assembler::equal, L_padding);
2634
2635 __ BIND(L_donePadding);
2636 __ subq(length, start_offset);
2637
2638 // Output size is (64 - output_size), output mask is (all 1s >> output_size).
2639 __ kmovql(input_mask, rax);
2640 __ movq(output_mask, -1);
2641 __ bzhiq(output_mask, output_mask, output_size);
2642
2643 // Load initial input with all valid base64 characters. Will be used
2644 // in merging source bytes to avoid masking when determining if an error occurred.
2645 __ movl(rax, 0x61616161);
2646 __ evpbroadcastd(input_initial_valid_b64, rax, Assembler::AVX_512bit);
2647
2648 // A register containing all invalid base64 decoded values
2649 __ movl(rax, 0x80808080);
2650 __ evpbroadcastd(invalid_b64, rax, Assembler::AVX_512bit);
2651
2652 // input_mask is in k1
2653 // output_size is in r13
2654 // output_mask is in r15
2655 // zmm0 - free
2656 // zmm1 - 0x00011000
2657 // zmm2 - 0x01400140
2658 // zmm3 - errorvec
2659 // zmm4 - pack vector
2660 // zmm5 - lookup_lo
2661 // zmm6 - lookup_hi
2662 // zmm7 - errorvec
2663 // zmm8 - 0x61616161
2664 // zmm9 - 0x80808080
2665
2666 // Load only the bytes from source, merging into our "fully-valid" register
2667 __ evmovdqub(input_initial_valid_b64, input_mask, Address(source, start_offset, Address::times_1, 0x0), true, Assembler::AVX_512bit);
2668
2669 // Decode all bytes within our merged input
2670 __ evmovdquq(tmp, lookup_lo, Assembler::AVX_512bit);
2671 __ evpermt2b(tmp, input_initial_valid_b64, lookup_hi, Assembler::AVX_512bit);
2672 __ evporq(mask, tmp, input_initial_valid_b64, Assembler::AVX_512bit);
2673
2674 // Check for error. Compare (decoded | initial) to all invalid.
2675 // If any bytes have their high-order bit set, then we have an error.
2676 __ evptestmb(k2, mask, invalid_b64, Assembler::AVX_512bit);
2677 __ kortestql(k2, k2);
2678
2679 // If we have an error, use the brute force loop to decode what we can (4-byte chunks).
2680 __ jcc(Assembler::notZero, L_bruteForce);
2681
2682 // Shuffle output bytes
2683 __ vpmaddubsw(tmp, tmp, pack16_op, Assembler::AVX_512bit);
2684 __ vpmaddwd(tmp, tmp, pack32_op, Assembler::AVX_512bit);
2685
2686 __ vpermb(tmp, pack24bits, tmp, Assembler::AVX_512bit);
2687 __ kmovql(k1, output_mask);
2688 __ evmovdqub(Address(dest, dp), k1, tmp, true, Assembler::AVX_512bit);
2689
2690 __ addptr(dest, output_size);
2691
2692 __ BIND(L_exit);
2693 __ vzeroupper();
2694 __ pop(rax); // Get original dest value
2695 __ subptr(dest, rax); // Number of bytes converted
2696 __ movptr(rax, dest);
2697 __ pop(rbx);
2698 __ pop(r15);
2699 __ pop(r14);
2700 __ pop(r13);
2701 __ pop(r12);
2702 __ leave();
2703 __ ret(0);
2704
2705 __ BIND(L_loadURL);
2706 __ evmovdquq(lookup_lo, ExternalAddress(StubRoutines::x86::base64_vbmi_lookup_lo_url_addr()), Assembler::AVX_512bit, r13);
2707 __ evmovdquq(lookup_hi, ExternalAddress(StubRoutines::x86::base64_vbmi_lookup_hi_url_addr()), Assembler::AVX_512bit, r13);
2708 __ jmp(L_continue);
2709
2710 __ BIND(L_padding);
2711 __ decrementq(output_size, 1);
2712 __ shrq(rax, 1);
2713
2714 __ cmpb(Address(source, length, Address::times_1, -2), '=');
2715 __ jcc(Assembler::notEqual, L_donePadding);
2716
2717 __ decrementq(output_size, 1);
2718 __ shrq(rax, 1);
2719 __ jmp(L_donePadding);
2720
2721 __ align32();
2722 __ BIND(L_bruteForce);
2723 } // End of if(avx512_vbmi)
2724
2725 if (VM_Version::supports_avx2()) {
2726 Label L_tailProc, L_topLoop, L_enterLoop;
2727
2728 __ cmpl(isMIME, 0);
2729 __ jcc(Assembler::notEqual, L_lastChunk);
2730
2731 // Check for buffer too small (for algorithm)
2732 __ subl(length, 0x2c);
2733 __ jcc(Assembler::less, L_tailProc);
2734
2735 __ shll(isURL, 2);
2736
2737 // Algorithm adapted from https://arxiv.org/abs/1704.00605, "Faster Base64
2738 // Encoding and Decoding using AVX2 Instructions". URL modifications added.
2739
2740 // Set up constants
2741 __ lea(r13, ExternalAddress(StubRoutines::x86::base64_AVX2_decode_tables_addr()));
2742 __ vpbroadcastd(xmm4, Address(r13, isURL, Address::times_1), Assembler::AVX_256bit); // 2F or 5F
2743 __ vpbroadcastd(xmm10, Address(r13, isURL, Address::times_1, 0x08), Assembler::AVX_256bit); // -1 or -4
2744 __ vmovdqu(xmm12, Address(r13, 0x10)); // permute
2745 __ vmovdqu(xmm13, Address(r13, 0x30)); // shuffle
2746 __ vpbroadcastd(xmm7, Address(r13, 0x50), Assembler::AVX_256bit); // merge
2747 __ vpbroadcastd(xmm6, Address(r13, 0x54), Assembler::AVX_256bit); // merge mult
2748
2749 __ lea(r13, ExternalAddress(StubRoutines::x86::base64_AVX2_decode_LUT_tables_addr()));
2750 __ shll(isURL, 4);
2751 __ vmovdqu(xmm11, Address(r13, isURL, Address::times_1, 0x00)); // lut_lo
2752 __ vmovdqu(xmm8, Address(r13, isURL, Address::times_1, 0x20)); // lut_roll
2753 __ shrl(isURL, 6); // restore isURL
2754 __ vmovdqu(xmm9, Address(r13, 0x80)); // lut_hi
2755 __ jmp(L_enterLoop);
2756
2757 __ align32();
2758 __ bind(L_topLoop);
2759 // Add in the offset value (roll) to get 6-bit out values
2760 __ vpaddb(xmm0, xmm0, xmm2, Assembler::AVX_256bit);
2761 // Merge and permute the output bits into appropriate output byte lanes
2762 __ vpmaddubsw(xmm0, xmm0, xmm7, Assembler::AVX_256bit);
2763 __ vpmaddwd(xmm0, xmm0, xmm6, Assembler::AVX_256bit);
2764 __ vpshufb(xmm0, xmm0, xmm13, Assembler::AVX_256bit);
2765 __ vpermd(xmm0, xmm12, xmm0, Assembler::AVX_256bit);
2766 // Store the output bytes
2767 __ vmovdqu(Address(dest, dp, Address::times_1, 0), xmm0);
2768 __ addptr(source, 0x20);
2769 __ addptr(dest, 0x18);
2770 __ subl(length, 0x20);
2771 __ jcc(Assembler::less, L_tailProc);
2772
2773 __ bind(L_enterLoop);
2774
2775 // Load in encoded string (32 bytes)
2776 __ vmovdqu(xmm2, Address(source, start_offset, Address::times_1, 0x0));
2777 // Extract the high nibble for indexing into the lut tables. High 4 bits are don't care.
2778 __ vpsrld(xmm1, xmm2, 0x4, Assembler::AVX_256bit);
2779 __ vpand(xmm1, xmm4, xmm1, Assembler::AVX_256bit);
2780 // Extract the low nibble. 5F/2F will isolate the low-order 4 bits. High 4 bits are don't care.
2781 __ vpand(xmm3, xmm2, xmm4, Assembler::AVX_256bit);
2782 // Check for special-case (0x2F or 0x5F (URL))
2783 __ vpcmpeqb(xmm0, xmm4, xmm2, Assembler::AVX_256bit);
2784 // Get the bitset based on the low nibble. vpshufb uses low-order 4 bits only.
2785 __ vpshufb(xmm3, xmm11, xmm3, Assembler::AVX_256bit);
2786 // Get the bit value of the high nibble
2787 __ vpshufb(xmm5, xmm9, xmm1, Assembler::AVX_256bit);
2788 // Make sure 2F / 5F shows as valid
2789 __ vpandn(xmm3, xmm0, xmm3, Assembler::AVX_256bit);
2790 // Make adjustment for roll index. For non-URL, this is a no-op,
2791 // for URL, this adjusts by -4. This is to properly index the
2792 // roll value for 2F / 5F.
2793 __ vpand(xmm0, xmm0, xmm10, Assembler::AVX_256bit);
2794 // If the and of the two is non-zero, we have an invalid input character
2795 __ vptest(xmm3, xmm5);
2796 // Extract the "roll" value - value to add to the input to get 6-bit out value
2797 __ vpaddb(xmm0, xmm0, xmm1, Assembler::AVX_256bit); // Handle 2F / 5F
2798 __ vpshufb(xmm0, xmm8, xmm0, Assembler::AVX_256bit);
2799 __ jcc(Assembler::equal, L_topLoop); // Fall through on error
2800
2801 __ bind(L_tailProc);
2802
2803 __ addl(length, 0x2c);
2804
2805 __ vzeroupper();
2806 }
2807
2808 // Use non-AVX code to decode 4-byte chunks into 3 bytes of output
2809
2810 // Register state (Linux):
2811 // r12-15 - saved on stack
2812 // rdi - src
2813 // rsi - sp
2814 // rdx - sl
2815 // rcx - dst
2816 // r8 - dp
2817 // r9 - isURL
2818
2819 // Register state (Windows):
2820 // r12-15 - saved on stack
2821 // rcx - src
2822 // rdx - sp
2823 // r8 - sl
2824 // r9 - dst
2825 // r12 - dp
2826 // r10 - isURL
2827
2828 // Registers (common):
2829 // length (r14) - bytes in src
2830
2831 const Register decode_table = r11;
2832 const Register out_byte_count = rbx;
2833 const Register byte1 = r13;
2834 const Register byte2 = r15;
2835 const Register byte3 = WIN64_ONLY(r8) NOT_WIN64(rdx);
2836 const Register byte4 = WIN64_ONLY(r10) NOT_WIN64(r9);
2837
2838 __ bind(L_lastChunk);
2839
2840 __ shrl(length, 2); // Multiple of 4 bytes only - length is # 4-byte chunks
2841 __ cmpl(length, 0);
2842 __ jcc(Assembler::lessEqual, L_exit_no_vzero);
2843
2844 __ shll(isURL, 8); // index into decode table based on isURL
2845 __ lea(decode_table, ExternalAddress(StubRoutines::x86::base64_decoding_table_addr()));
2846 __ addptr(decode_table, isURL);
2847
2848 __ jmp(L_bottomLoop);
2849
2850 __ align32();
2851 __ BIND(L_forceLoop);
2852 __ shll(byte1, 18);
2853 __ shll(byte2, 12);
2854 __ shll(byte3, 6);
2855 __ orl(byte1, byte2);
2856 __ orl(byte1, byte3);
2857 __ orl(byte1, byte4);
2858
2859 __ addptr(source, 4);
2860
2861 __ movb(Address(dest, dp, Address::times_1, 2), byte1);
2862 __ shrl(byte1, 8);
2863 __ movb(Address(dest, dp, Address::times_1, 1), byte1);
2864 __ shrl(byte1, 8);
2865 __ movb(Address(dest, dp, Address::times_1, 0), byte1);
2866
2867 __ addptr(dest, 3);
2868 __ decrementl(length, 1);
2869 __ jcc(Assembler::zero, L_exit_no_vzero);
2870
2871 __ BIND(L_bottomLoop);
2872 __ load_unsigned_byte(byte1, Address(source, start_offset, Address::times_1, 0x00));
2873 __ load_unsigned_byte(byte2, Address(source, start_offset, Address::times_1, 0x01));
2874 __ load_signed_byte(byte1, Address(decode_table, byte1));
2875 __ load_signed_byte(byte2, Address(decode_table, byte2));
2876 __ load_unsigned_byte(byte3, Address(source, start_offset, Address::times_1, 0x02));
2877 __ load_unsigned_byte(byte4, Address(source, start_offset, Address::times_1, 0x03));
2878 __ load_signed_byte(byte3, Address(decode_table, byte3));
2879 __ load_signed_byte(byte4, Address(decode_table, byte4));
2880
2881 __ mov(rax, byte1);
2882 __ orl(rax, byte2);
2883 __ orl(rax, byte3);
2884 __ orl(rax, byte4);
2885 __ jcc(Assembler::positive, L_forceLoop);
2886
2887 __ BIND(L_exit_no_vzero);
2888 __ pop(rax); // Get original dest value
2889 __ subptr(dest, rax); // Number of bytes converted
2890 __ movptr(rax, dest);
2891 __ pop(rbx);
2892 __ pop(r15);
2893 __ pop(r14);
2894 __ pop(r13);
2895 __ pop(r12);
2896 __ leave();
2897 __ ret(0);
2898
2899 return start;
2900 }
2901
2902
2903 /**
2904 * Arguments:
2905 *
2906 * Inputs:
2907 * c_rarg0 - int crc
2908 * c_rarg1 - byte* buf
2909 * c_rarg2 - int length
2910 *
2911 * Output:
2912 * rax - int crc result
2913 */
2914 address StubGenerator::generate_updateBytesCRC32() {
2915 assert(UseCRC32Intrinsics, "need AVX and CLMUL instructions");
2916
2917 __ align(CodeEntryAlignment);
2918 StubCodeMark mark(this, "StubRoutines", "updateBytesCRC32");
2919
2920 address start = __ pc();
2921
2922 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
2923 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
2924 // rscratch1: r10
2925 const Register crc = c_rarg0; // crc
2926 const Register buf = c_rarg1; // source java byte array address
2927 const Register len = c_rarg2; // length
2928 const Register table = c_rarg3; // crc_table address (reuse register)
2929 const Register tmp1 = r11;
2930 const Register tmp2 = r10;
2931 assert_different_registers(crc, buf, len, table, tmp1, tmp2, rax);
2932
2933 BLOCK_COMMENT("Entry:");
2934 __ enter(); // required for proper stackwalking of RuntimeStub frame
2935
2936 if (VM_Version::supports_sse4_1() && VM_Version::supports_avx512_vpclmulqdq() &&
2937 VM_Version::supports_avx512bw() &&
2938 VM_Version::supports_avx512vl()) {
2939 // The constants used in the CRC32 algorithm requires the 1's compliment of the initial crc value.
2940 // However, the constant table for CRC32-C assumes the original crc value. Account for this
2941 // difference before calling and after returning.
2942 __ lea(table, ExternalAddress(StubRoutines::x86::crc_table_avx512_addr()));
2943 __ notl(crc);
2944 __ kernel_crc32_avx512(crc, buf, len, table, tmp1, tmp2);
2945 __ notl(crc);
2946 } else {
2947 __ kernel_crc32(crc, buf, len, table, tmp1);
2948 }
2949
2950 __ movl(rax, crc);
2951 __ vzeroupper();
2952 __ leave(); // required for proper stackwalking of RuntimeStub frame
2953 __ ret(0);
2954
2955 return start;
2956 }
2957
2958 /**
2959 * Arguments:
2960 *
2961 * Inputs:
2962 * c_rarg0 - int crc
2963 * c_rarg1 - byte* buf
2964 * c_rarg2 - long length
2965 * c_rarg3 - table_start - optional (present only when doing a library_call,
2966 * not used by x86 algorithm)
2967 *
2968 * Output:
2969 * rax - int crc result
2970 */
2971 address StubGenerator::generate_updateBytesCRC32C(bool is_pclmulqdq_supported) {
2972 assert(UseCRC32CIntrinsics, "need SSE4_2");
2973 __ align(CodeEntryAlignment);
2974 StubCodeMark mark(this, "StubRoutines", "updateBytesCRC32C");
2975 address start = __ pc();
2976
2977 //reg.arg int#0 int#1 int#2 int#3 int#4 int#5 float regs
2978 //Windows RCX RDX R8 R9 none none XMM0..XMM3
2979 //Lin / Sol RDI RSI RDX RCX R8 R9 XMM0..XMM7
2980 const Register crc = c_rarg0; // crc
2981 const Register buf = c_rarg1; // source java byte array address
2982 const Register len = c_rarg2; // length
2983 const Register a = rax;
2984 const Register j = r9;
2985 const Register k = r10;
2986 const Register l = r11;
2987 #ifdef _WIN64
2988 const Register y = rdi;
2989 const Register z = rsi;
2990 #else
2991 const Register y = rcx;
2992 const Register z = r8;
2993 #endif
2994 assert_different_registers(crc, buf, len, a, j, k, l, y, z);
2995
2996 BLOCK_COMMENT("Entry:");
2997 __ enter(); // required for proper stackwalking of RuntimeStub frame
2998 Label L_continue;
2999
3000 if (VM_Version::supports_sse4_1() && VM_Version::supports_avx512_vpclmulqdq() &&
3001 VM_Version::supports_avx512bw() &&
3002 VM_Version::supports_avx512vl()) {
3003 Label L_doSmall;
3004
3005 __ cmpl(len, 384);
3006 __ jcc(Assembler::lessEqual, L_doSmall);
3007
3008 __ lea(j, ExternalAddress(StubRoutines::x86::crc32c_table_avx512_addr()));
3009 __ kernel_crc32_avx512(crc, buf, len, j, l, k);
3010
3011 __ jmp(L_continue);
3012
3013 __ bind(L_doSmall);
3014 }
3015 #ifdef _WIN64
3016 __ push(y);
3017 __ push(z);
3018 #endif
3019 __ crc32c_ipl_alg2_alt2(crc, buf, len,
3020 a, j, k,
3021 l, y, z,
3022 c_farg0, c_farg1, c_farg2,
3023 is_pclmulqdq_supported);
3024 #ifdef _WIN64
3025 __ pop(z);
3026 __ pop(y);
3027 #endif
3028
3029 __ bind(L_continue);
3030 __ movl(rax, crc);
3031 __ vzeroupper();
3032 __ leave(); // required for proper stackwalking of RuntimeStub frame
3033 __ ret(0);
3034
3035 return start;
3036 }
3037
3038
3039 /**
3040 * Arguments:
3041 *
3042 * Input:
3043 * c_rarg0 - x address
3044 * c_rarg1 - x length
3045 * c_rarg2 - y address
3046 * c_rarg3 - y length
3047 * not Win64
3048 * c_rarg4 - z address
3049 * Win64
3050 * rsp+40 - z address
3051 */
3052 address StubGenerator::generate_multiplyToLen() {
3053 __ align(CodeEntryAlignment);
3054 StubCodeMark mark(this, "StubRoutines", "multiplyToLen");
3055 address start = __ pc();
3056
3057 if (SCCache::load_stub(this, vmIntrinsics::_multiplyToLen, "multiplyToLen", start)) {
3058 return start;
3059 }
3060
3061 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
3062 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
3063 const Register x = rdi;
3064 const Register xlen = rax;
3065 const Register y = rsi;
3066 const Register ylen = rcx;
3067 const Register z = r8;
3068
3069 // Next registers will be saved on stack in multiply_to_len().
3070 const Register tmp0 = r11;
3071 const Register tmp1 = r12;
3072 const Register tmp2 = r13;
3073 const Register tmp3 = r14;
3074 const Register tmp4 = r15;
3075 const Register tmp5 = rbx;
3076
3077 BLOCK_COMMENT("Entry:");
3078 __ enter(); // required for proper stackwalking of RuntimeStub frame
3079
3080 setup_arg_regs(4); // x => rdi, xlen => rsi, y => rdx
3081 // ylen => rcx, z => r8
3082 // r9 and r10 may be used to save non-volatile registers
3083 #ifdef _WIN64
3084 // last argument (#4) is on stack on Win64
3085 __ movptr(z, Address(rsp, 6 * wordSize));
3086 #endif
3087
3088 __ movptr(xlen, rsi);
3089 __ movptr(y, rdx);
3090 __ multiply_to_len(x, xlen, y, ylen, z, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5);
3091
3092 restore_arg_regs();
3093
3094 __ leave(); // required for proper stackwalking of RuntimeStub frame
3095 __ ret(0);
3096
3097 SCCache::store_stub(this, vmIntrinsics::_multiplyToLen, "multiplyToLen", start);
3098 return start;
3099 }
3100
3101 /**
3102 * Arguments:
3103 *
3104 * Input:
3105 * c_rarg0 - obja address
3106 * c_rarg1 - objb address
3107 * c_rarg3 - length length
3108 * c_rarg4 - scale log2_array_indxscale
3109 *
3110 * Output:
3111 * rax - int >= mismatched index, < 0 bitwise complement of tail
3112 */
3113 address StubGenerator::generate_vectorizedMismatch() {
3114 __ align(CodeEntryAlignment);
3115 StubCodeMark mark(this, "StubRoutines", "vectorizedMismatch");
3116 address start = __ pc();
3117
3118 BLOCK_COMMENT("Entry:");
3119 __ enter();
3120
3121 #ifdef _WIN64 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
3122 const Register scale = c_rarg0; //rcx, will exchange with r9
3123 const Register objb = c_rarg1; //rdx
3124 const Register length = c_rarg2; //r8
3125 const Register obja = c_rarg3; //r9
3126 __ xchgq(obja, scale); //now obja and scale contains the correct contents
3127
3128 const Register tmp1 = r10;
3129 const Register tmp2 = r11;
3130 #endif
3131 #ifndef _WIN64 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
3132 const Register obja = c_rarg0; //U:rdi
3133 const Register objb = c_rarg1; //U:rsi
3134 const Register length = c_rarg2; //U:rdx
3135 const Register scale = c_rarg3; //U:rcx
3136 const Register tmp1 = r8;
3137 const Register tmp2 = r9;
3138 #endif
3139 const Register result = rax; //return value
3140 const XMMRegister vec0 = xmm0;
3141 const XMMRegister vec1 = xmm1;
3142 const XMMRegister vec2 = xmm2;
3143
3144 __ vectorized_mismatch(obja, objb, length, scale, result, tmp1, tmp2, vec0, vec1, vec2);
3145
3146 __ vzeroupper();
3147 __ leave();
3148 __ ret(0);
3149
3150 return start;
3151 }
3152
3153 /**
3154 * Arguments:
3155 *
3156 // Input:
3157 // c_rarg0 - x address
3158 // c_rarg1 - x length
3159 // c_rarg2 - z address
3160 // c_rarg3 - z length
3161 *
3162 */
3163 address StubGenerator::generate_squareToLen() {
3164
3165 __ align(CodeEntryAlignment);
3166 StubCodeMark mark(this, "StubRoutines", "squareToLen");
3167 address start = __ pc();
3168
3169 if (SCCache::load_stub(this, vmIntrinsics::_squareToLen, "squareToLen", start)) {
3170 return start;
3171 }
3172
3173 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
3174 // Unix: rdi, rsi, rdx, rcx (c_rarg0, c_rarg1, ...)
3175 const Register x = rdi;
3176 const Register len = rsi;
3177 const Register z = r8;
3178 const Register zlen = rcx;
3179
3180 const Register tmp1 = r12;
3181 const Register tmp2 = r13;
3182 const Register tmp3 = r14;
3183 const Register tmp4 = r15;
3184 const Register tmp5 = rbx;
3185
3186 BLOCK_COMMENT("Entry:");
3187 __ enter(); // required for proper stackwalking of RuntimeStub frame
3188
3189 setup_arg_regs(4); // x => rdi, len => rsi, z => rdx
3190 // zlen => rcx
3191 // r9 and r10 may be used to save non-volatile registers
3192 __ movptr(r8, rdx);
3193 __ square_to_len(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx, rax);
3194
3195 restore_arg_regs();
3196
3197 __ leave(); // required for proper stackwalking of RuntimeStub frame
3198 __ ret(0);
3199
3200 SCCache::store_stub(this, vmIntrinsics::_squareToLen, "squareToLen", start);
3201 return start;
3202 }
3203
3204 address StubGenerator::generate_method_entry_barrier() {
3205 __ align(CodeEntryAlignment);
3206 StubCodeMark mark(this, "StubRoutines", "nmethod_entry_barrier");
3207 address start = __ pc();
3208
3209 Label deoptimize_label;
3210
3211 __ push(-1); // cookie, this is used for writing the new rsp when deoptimizing
3212
3213 BLOCK_COMMENT("Entry:");
3214 __ enter(); // save rbp
3215
3216 // save c_rarg0, because we want to use that value.
3217 // We could do without it but then we depend on the number of slots used by pusha
3218 __ push(c_rarg0);
3219
3220 __ lea(c_rarg0, Address(rsp, wordSize * 3)); // 1 for cookie, 1 for rbp, 1 for c_rarg0 - this should be the return address
3221
3222 __ pusha();
3223
3224 // The method may have floats as arguments, and we must spill them before calling
3225 // the VM runtime.
3226 assert(Argument::n_float_register_parameters_j == 8, "Assumption");
3227 const int xmm_size = wordSize * 2;
3228 const int xmm_spill_size = xmm_size * Argument::n_float_register_parameters_j;
3229 __ subptr(rsp, xmm_spill_size);
3230 __ movdqu(Address(rsp, xmm_size * 7), xmm7);
3231 __ movdqu(Address(rsp, xmm_size * 6), xmm6);
3232 __ movdqu(Address(rsp, xmm_size * 5), xmm5);
3233 __ movdqu(Address(rsp, xmm_size * 4), xmm4);
3234 __ movdqu(Address(rsp, xmm_size * 3), xmm3);
3235 __ movdqu(Address(rsp, xmm_size * 2), xmm2);
3236 __ movdqu(Address(rsp, xmm_size * 1), xmm1);
3237 __ movdqu(Address(rsp, xmm_size * 0), xmm0);
3238
3239 __ call_VM_leaf(CAST_FROM_FN_PTR(address, static_cast<int (*)(address*)>(BarrierSetNMethod::nmethod_stub_entry_barrier)), 1);
3240
3241 __ movdqu(xmm0, Address(rsp, xmm_size * 0));
3242 __ movdqu(xmm1, Address(rsp, xmm_size * 1));
3243 __ movdqu(xmm2, Address(rsp, xmm_size * 2));
3244 __ movdqu(xmm3, Address(rsp, xmm_size * 3));
3245 __ movdqu(xmm4, Address(rsp, xmm_size * 4));
3246 __ movdqu(xmm5, Address(rsp, xmm_size * 5));
3247 __ movdqu(xmm6, Address(rsp, xmm_size * 6));
3248 __ movdqu(xmm7, Address(rsp, xmm_size * 7));
3249 __ addptr(rsp, xmm_spill_size);
3250
3251 __ cmpl(rax, 1); // 1 means deoptimize
3252 __ jcc(Assembler::equal, deoptimize_label);
3253
3254 __ popa();
3255 __ pop(c_rarg0);
3256
3257 __ leave();
3258
3259 __ addptr(rsp, 1 * wordSize); // cookie
3260 __ ret(0);
3261
3262
3263 __ BIND(deoptimize_label);
3264
3265 __ popa();
3266 __ pop(c_rarg0);
3267
3268 __ leave();
3269
3270 // this can be taken out, but is good for verification purposes. getting a SIGSEGV
3271 // here while still having a correct stack is valuable
3272 __ testptr(rsp, Address(rsp, 0));
3273
3274 __ movptr(rsp, Address(rsp, 0)); // new rsp was written in the barrier
3275 __ jmp(Address(rsp, -1 * wordSize)); // jmp target should be callers verified_entry_point
3276
3277 return start;
3278 }
3279
3280 /**
3281 * Arguments:
3282 *
3283 * Input:
3284 * c_rarg0 - out address
3285 * c_rarg1 - in address
3286 * c_rarg2 - offset
3287 * c_rarg3 - len
3288 * not Win64
3289 * c_rarg4 - k
3290 * Win64
3291 * rsp+40 - k
3292 */
3293 address StubGenerator::generate_mulAdd() {
3294 __ align(CodeEntryAlignment);
3295 StubCodeMark mark(this, "StubRoutines", "mulAdd");
3296 address start = __ pc();
3297
3298 if (SCCache::load_stub(this, vmIntrinsics::_mulAdd, "mulAdd", start)) {
3299 return start;
3300 }
3301
3302 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
3303 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
3304 const Register out = rdi;
3305 const Register in = rsi;
3306 const Register offset = r11;
3307 const Register len = rcx;
3308 const Register k = r8;
3309
3310 // Next registers will be saved on stack in mul_add().
3311 const Register tmp1 = r12;
3312 const Register tmp2 = r13;
3313 const Register tmp3 = r14;
3314 const Register tmp4 = r15;
3315 const Register tmp5 = rbx;
3316
3317 BLOCK_COMMENT("Entry:");
3318 __ enter(); // required for proper stackwalking of RuntimeStub frame
3319
3320 setup_arg_regs(4); // out => rdi, in => rsi, offset => rdx
3321 // len => rcx, k => r8
3322 // r9 and r10 may be used to save non-volatile registers
3323 #ifdef _WIN64
3324 // last argument is on stack on Win64
3325 __ movl(k, Address(rsp, 6 * wordSize));
3326 #endif
3327 __ movptr(r11, rdx); // move offset in rdx to offset(r11)
3328 __ mul_add(out, in, offset, len, k, tmp1, tmp2, tmp3, tmp4, tmp5, rdx, rax);
3329
3330 restore_arg_regs();
3331
3332 __ leave(); // required for proper stackwalking of RuntimeStub frame
3333 __ ret(0);
3334
3335 SCCache::store_stub(this, vmIntrinsics::_mulAdd, "mulAdd", start);
3336 return start;
3337 }
3338
3339 address StubGenerator::generate_bigIntegerRightShift() {
3340 __ align(CodeEntryAlignment);
3341 StubCodeMark mark(this, "StubRoutines", "bigIntegerRightShiftWorker");
3342 address start = __ pc();
3343
3344 Label Shift512Loop, ShiftTwo, ShiftTwoLoop, ShiftOne, Exit;
3345 // For Unix, the arguments are as follows: rdi, rsi, rdx, rcx, r8.
3346 const Register newArr = rdi;
3347 const Register oldArr = rsi;
3348 const Register newIdx = rdx;
3349 const Register shiftCount = rcx; // It was intentional to have shiftCount in rcx since it is used implicitly for shift.
3350 const Register totalNumIter = r8;
3351
3352 // For windows, we use r9 and r10 as temps to save rdi and rsi. Thus we cannot allocate them for our temps.
3353 // For everything else, we prefer using r9 and r10 since we do not have to save them before use.
3354 const Register tmp1 = r11; // Caller save.
3355 const Register tmp2 = rax; // Caller save.
3356 const Register tmp3 = WIN64_ONLY(r12) NOT_WIN64(r9); // Windows: Callee save. Linux: Caller save.
3357 const Register tmp4 = WIN64_ONLY(r13) NOT_WIN64(r10); // Windows: Callee save. Linux: Caller save.
3358 const Register tmp5 = r14; // Callee save.
3359 const Register tmp6 = r15;
3360
3361 const XMMRegister x0 = xmm0;
3362 const XMMRegister x1 = xmm1;
3363 const XMMRegister x2 = xmm2;
3364
3365 BLOCK_COMMENT("Entry:");
3366 __ enter(); // required for proper stackwalking of RuntimeStub frame
3367
3368 #ifdef _WIN64
3369 setup_arg_regs(4);
3370 // For windows, since last argument is on stack, we need to move it to the appropriate register.
3371 __ movl(totalNumIter, Address(rsp, 6 * wordSize));
3372 // Save callee save registers.
3373 __ push(tmp3);
3374 __ push(tmp4);
3375 #endif
3376 __ push(tmp5);
3377
3378 // Rename temps used throughout the code.
3379 const Register idx = tmp1;
3380 const Register nIdx = tmp2;
3381
3382 __ xorl(idx, idx);
3383
3384 // Start right shift from end of the array.
3385 // For example, if #iteration = 4 and newIdx = 1
3386 // then dest[4] = src[4] >> shiftCount | src[3] <<< (shiftCount - 32)
3387 // if #iteration = 4 and newIdx = 0
3388 // then dest[3] = src[4] >> shiftCount | src[3] <<< (shiftCount - 32)
3389 __ movl(idx, totalNumIter);
3390 __ movl(nIdx, idx);
3391 __ addl(nIdx, newIdx);
3392
3393 // If vectorization is enabled, check if the number of iterations is at least 64
3394 // If not, then go to ShifTwo processing 2 iterations
3395 if (VM_Version::supports_avx512_vbmi2()) {
3396 __ cmpptr(totalNumIter, (AVX3Threshold/64));
3397 __ jcc(Assembler::less, ShiftTwo);
3398
3399 if (AVX3Threshold < 16 * 64) {
3400 __ cmpl(totalNumIter, 16);
3401 __ jcc(Assembler::less, ShiftTwo);
3402 }
3403 __ evpbroadcastd(x0, shiftCount, Assembler::AVX_512bit);
3404 __ subl(idx, 16);
3405 __ subl(nIdx, 16);
3406 __ BIND(Shift512Loop);
3407 __ evmovdqul(x2, Address(oldArr, idx, Address::times_4, 4), Assembler::AVX_512bit);
3408 __ evmovdqul(x1, Address(oldArr, idx, Address::times_4), Assembler::AVX_512bit);
3409 __ vpshrdvd(x2, x1, x0, Assembler::AVX_512bit);
3410 __ evmovdqul(Address(newArr, nIdx, Address::times_4), x2, Assembler::AVX_512bit);
3411 __ subl(nIdx, 16);
3412 __ subl(idx, 16);
3413 __ jcc(Assembler::greaterEqual, Shift512Loop);
3414 __ addl(idx, 16);
3415 __ addl(nIdx, 16);
3416 }
3417 __ BIND(ShiftTwo);
3418 __ cmpl(idx, 2);
3419 __ jcc(Assembler::less, ShiftOne);
3420 __ subl(idx, 2);
3421 __ subl(nIdx, 2);
3422 __ BIND(ShiftTwoLoop);
3423 __ movl(tmp5, Address(oldArr, idx, Address::times_4, 8));
3424 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 4));
3425 __ movl(tmp3, Address(oldArr, idx, Address::times_4));
3426 __ shrdl(tmp5, tmp4);
3427 __ shrdl(tmp4, tmp3);
3428 __ movl(Address(newArr, nIdx, Address::times_4, 4), tmp5);
3429 __ movl(Address(newArr, nIdx, Address::times_4), tmp4);
3430 __ subl(nIdx, 2);
3431 __ subl(idx, 2);
3432 __ jcc(Assembler::greaterEqual, ShiftTwoLoop);
3433 __ addl(idx, 2);
3434 __ addl(nIdx, 2);
3435
3436 // Do the last iteration
3437 __ BIND(ShiftOne);
3438 __ cmpl(idx, 1);
3439 __ jcc(Assembler::less, Exit);
3440 __ subl(idx, 1);
3441 __ subl(nIdx, 1);
3442 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 4));
3443 __ movl(tmp3, Address(oldArr, idx, Address::times_4));
3444 __ shrdl(tmp4, tmp3);
3445 __ movl(Address(newArr, nIdx, Address::times_4), tmp4);
3446 __ BIND(Exit);
3447 __ vzeroupper();
3448 // Restore callee save registers.
3449 __ pop(tmp5);
3450 #ifdef _WIN64
3451 __ pop(tmp4);
3452 __ pop(tmp3);
3453 restore_arg_regs();
3454 #endif
3455 __ leave(); // required for proper stackwalking of RuntimeStub frame
3456 __ ret(0);
3457
3458 return start;
3459 }
3460
3461 /**
3462 * Arguments:
3463 *
3464 * Input:
3465 * c_rarg0 - newArr address
3466 * c_rarg1 - oldArr address
3467 * c_rarg2 - newIdx
3468 * c_rarg3 - shiftCount
3469 * not Win64
3470 * c_rarg4 - numIter
3471 * Win64
3472 * rsp40 - numIter
3473 */
3474 address StubGenerator::generate_bigIntegerLeftShift() {
3475 __ align(CodeEntryAlignment);
3476 StubCodeMark mark(this, "StubRoutines", "bigIntegerLeftShiftWorker");
3477 address start = __ pc();
3478
3479 Label Shift512Loop, ShiftTwo, ShiftTwoLoop, ShiftOne, Exit;
3480 // For Unix, the arguments are as follows: rdi, rsi, rdx, rcx, r8.
3481 const Register newArr = rdi;
3482 const Register oldArr = rsi;
3483 const Register newIdx = rdx;
3484 const Register shiftCount = rcx; // It was intentional to have shiftCount in rcx since it is used implicitly for shift.
3485 const Register totalNumIter = r8;
3486 // For windows, we use r9 and r10 as temps to save rdi and rsi. Thus we cannot allocate them for our temps.
3487 // For everything else, we prefer using r9 and r10 since we do not have to save them before use.
3488 const Register tmp1 = r11; // Caller save.
3489 const Register tmp2 = rax; // Caller save.
3490 const Register tmp3 = WIN64_ONLY(r12) NOT_WIN64(r9); // Windows: Callee save. Linux: Caller save.
3491 const Register tmp4 = WIN64_ONLY(r13) NOT_WIN64(r10); // Windows: Callee save. Linux: Caller save.
3492 const Register tmp5 = r14; // Callee save.
3493
3494 const XMMRegister x0 = xmm0;
3495 const XMMRegister x1 = xmm1;
3496 const XMMRegister x2 = xmm2;
3497 BLOCK_COMMENT("Entry:");
3498 __ enter(); // required for proper stackwalking of RuntimeStub frame
3499
3500 #ifdef _WIN64
3501 setup_arg_regs(4);
3502 // For windows, since last argument is on stack, we need to move it to the appropriate register.
3503 __ movl(totalNumIter, Address(rsp, 6 * wordSize));
3504 // Save callee save registers.
3505 __ push(tmp3);
3506 __ push(tmp4);
3507 #endif
3508 __ push(tmp5);
3509
3510 // Rename temps used throughout the code
3511 const Register idx = tmp1;
3512 const Register numIterTmp = tmp2;
3513
3514 // Start idx from zero.
3515 __ xorl(idx, idx);
3516 // Compute interior pointer for new array. We do this so that we can use same index for both old and new arrays.
3517 __ lea(newArr, Address(newArr, newIdx, Address::times_4));
3518 __ movl(numIterTmp, totalNumIter);
3519
3520 // If vectorization is enabled, check if the number of iterations is at least 64
3521 // If not, then go to ShiftTwo shifting two numbers at a time
3522 if (VM_Version::supports_avx512_vbmi2()) {
3523 __ cmpl(totalNumIter, (AVX3Threshold/64));
3524 __ jcc(Assembler::less, ShiftTwo);
3525
3526 if (AVX3Threshold < 16 * 64) {
3527 __ cmpl(totalNumIter, 16);
3528 __ jcc(Assembler::less, ShiftTwo);
3529 }
3530 __ evpbroadcastd(x0, shiftCount, Assembler::AVX_512bit);
3531 __ subl(numIterTmp, 16);
3532 __ BIND(Shift512Loop);
3533 __ evmovdqul(x1, Address(oldArr, idx, Address::times_4), Assembler::AVX_512bit);
3534 __ evmovdqul(x2, Address(oldArr, idx, Address::times_4, 0x4), Assembler::AVX_512bit);
3535 __ vpshldvd(x1, x2, x0, Assembler::AVX_512bit);
3536 __ evmovdqul(Address(newArr, idx, Address::times_4), x1, Assembler::AVX_512bit);
3537 __ addl(idx, 16);
3538 __ subl(numIterTmp, 16);
3539 __ jcc(Assembler::greaterEqual, Shift512Loop);
3540 __ addl(numIterTmp, 16);
3541 }
3542 __ BIND(ShiftTwo);
3543 __ cmpl(totalNumIter, 1);
3544 __ jcc(Assembler::less, Exit);
3545 __ movl(tmp3, Address(oldArr, idx, Address::times_4));
3546 __ subl(numIterTmp, 2);
3547 __ jcc(Assembler::less, ShiftOne);
3548
3549 __ BIND(ShiftTwoLoop);
3550 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 0x4));
3551 __ movl(tmp5, Address(oldArr, idx, Address::times_4, 0x8));
3552 __ shldl(tmp3, tmp4);
3553 __ shldl(tmp4, tmp5);
3554 __ movl(Address(newArr, idx, Address::times_4), tmp3);
3555 __ movl(Address(newArr, idx, Address::times_4, 0x4), tmp4);
3556 __ movl(tmp3, tmp5);
3557 __ addl(idx, 2);
3558 __ subl(numIterTmp, 2);
3559 __ jcc(Assembler::greaterEqual, ShiftTwoLoop);
3560
3561 // Do the last iteration
3562 __ BIND(ShiftOne);
3563 __ addl(numIterTmp, 2);
3564 __ cmpl(numIterTmp, 1);
3565 __ jcc(Assembler::less, Exit);
3566 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 0x4));
3567 __ shldl(tmp3, tmp4);
3568 __ movl(Address(newArr, idx, Address::times_4), tmp3);
3569
3570 __ BIND(Exit);
3571 __ vzeroupper();
3572 // Restore callee save registers.
3573 __ pop(tmp5);
3574 #ifdef _WIN64
3575 __ pop(tmp4);
3576 __ pop(tmp3);
3577 restore_arg_regs();
3578 #endif
3579 __ leave(); // required for proper stackwalking of RuntimeStub frame
3580 __ ret(0);
3581
3582 return start;
3583 }
3584
3585 void StubGenerator::generate_libm_stubs() {
3586 if (UseLibmIntrinsic && InlineIntrinsics) {
3587 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dsin)) {
3588 StubRoutines::_dsin = generate_libmSin(); // from stubGenerator_x86_64_sin.cpp
3589 }
3590 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dcos)) {
3591 StubRoutines::_dcos = generate_libmCos(); // from stubGenerator_x86_64_cos.cpp
3592 }
3593 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dtan)) {
3594 StubRoutines::_dtan = generate_libmTan(); // from stubGenerator_x86_64_tan.cpp
3595 }
3596 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dtanh)) {
3597 StubRoutines::_dtanh = generate_libmTanh(); // from stubGenerator_x86_64_tanh.cpp
3598 }
3599 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dexp)) {
3600 StubRoutines::_dexp = generate_libmExp(); // from stubGenerator_x86_64_exp.cpp
3601 }
3602 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dpow)) {
3603 StubRoutines::_dpow = generate_libmPow(); // from stubGenerator_x86_64_pow.cpp
3604 }
3605 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dlog)) {
3606 StubRoutines::_dlog = generate_libmLog(); // from stubGenerator_x86_64_log.cpp
3607 }
3608 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dlog10)) {
3609 StubRoutines::_dlog10 = generate_libmLog10(); // from stubGenerator_x86_64_log.cpp
3610 }
3611 }
3612 }
3613
3614 /**
3615 * Arguments:
3616 *
3617 * Input:
3618 * c_rarg0 - float16 jshort
3619 *
3620 * Output:
3621 * xmm0 - float
3622 */
3623 address StubGenerator::generate_float16ToFloat() {
3624 StubCodeMark mark(this, "StubRoutines", "float16ToFloat");
3625
3626 address start = __ pc();
3627
3628 BLOCK_COMMENT("Entry:");
3629 // No need for RuntimeStub frame since it is called only during JIT compilation
3630
3631 // Load value into xmm0 and convert
3632 __ flt16_to_flt(xmm0, c_rarg0);
3633
3634 __ ret(0);
3635
3636 return start;
3637 }
3638
3639 /**
3640 * Arguments:
3641 *
3642 * Input:
3643 * xmm0 - float
3644 *
3645 * Output:
3646 * rax - float16 jshort
3647 */
3648 address StubGenerator::generate_floatToFloat16() {
3649 StubCodeMark mark(this, "StubRoutines", "floatToFloat16");
3650
3651 address start = __ pc();
3652
3653 BLOCK_COMMENT("Entry:");
3654 // No need for RuntimeStub frame since it is called only during JIT compilation
3655
3656 // Convert and put result into rax
3657 __ flt_to_flt16(rax, xmm0, xmm1);
3658
3659 __ ret(0);
3660
3661 return start;
3662 }
3663
3664 address StubGenerator::generate_cont_thaw(const char* label, Continuation::thaw_kind kind) {
3665 if (!Continuations::enabled()) return nullptr;
3666
3667 bool return_barrier = Continuation::is_thaw_return_barrier(kind);
3668 bool return_barrier_exception = Continuation::is_thaw_return_barrier_exception(kind);
3669
3670 StubCodeMark mark(this, "StubRoutines", label);
3671 address start = __ pc();
3672
3673 // TODO: Handle Valhalla return types. May require generating different return barriers.
3674
3675 if (!return_barrier) {
3676 // Pop return address. If we don't do this, we get a drift,
3677 // where the bottom-most frozen frame continuously grows.
3678 __ pop(c_rarg3);
3679 } else {
3680 __ movptr(rsp, Address(r15_thread, JavaThread::cont_entry_offset()));
3681 }
3682
3683 #ifdef ASSERT
3684 {
3685 Label L_good_sp;
3686 __ cmpptr(rsp, Address(r15_thread, JavaThread::cont_entry_offset()));
3687 __ jcc(Assembler::equal, L_good_sp);
3688 __ stop("Incorrect rsp at thaw entry");
3689 __ BIND(L_good_sp);
3690 }
3691 #endif // ASSERT
3692
3693 if (return_barrier) {
3694 // Preserve possible return value from a method returning to the return barrier.
3695 __ push(rax);
3696 __ push_d(xmm0);
3697 }
3698
3699 __ movptr(c_rarg0, r15_thread);
3700 __ movptr(c_rarg1, (return_barrier ? 1 : 0));
3701 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Continuation::prepare_thaw), 2);
3702 __ movptr(rbx, rax);
3703
3704 if (return_barrier) {
3705 // Restore return value from a method returning to the return barrier.
3706 // No safepoint in the call to thaw, so even an oop return value should be OK.
3707 __ pop_d(xmm0);
3708 __ pop(rax);
3709 }
3710
3711 #ifdef ASSERT
3712 {
3713 Label L_good_sp;
3714 __ cmpptr(rsp, Address(r15_thread, JavaThread::cont_entry_offset()));
3715 __ jcc(Assembler::equal, L_good_sp);
3716 __ stop("Incorrect rsp after prepare thaw");
3717 __ BIND(L_good_sp);
3718 }
3719 #endif // ASSERT
3720
3721 // rbx contains the size of the frames to thaw, 0 if overflow or no more frames
3722 Label L_thaw_success;
3723 __ testptr(rbx, rbx);
3724 __ jccb(Assembler::notZero, L_thaw_success);
3725 __ jump(RuntimeAddress(SharedRuntime::throw_StackOverflowError_entry()));
3726 __ bind(L_thaw_success);
3727
3728 // Make room for the thawed frames and align the stack.
3729 __ subptr(rsp, rbx);
3730 __ andptr(rsp, -StackAlignmentInBytes);
3731
3732 if (return_barrier) {
3733 // Preserve possible return value from a method returning to the return barrier. (Again.)
3734 __ push(rax);
3735 __ push_d(xmm0);
3736 }
3737
3738 // If we want, we can templatize thaw by kind, and have three different entries.
3739 __ movptr(c_rarg0, r15_thread);
3740 __ movptr(c_rarg1, kind);
3741 __ call_VM_leaf(Continuation::thaw_entry(), 2);
3742 __ movptr(rbx, rax);
3743
3744 if (return_barrier) {
3745 // Restore return value from a method returning to the return barrier. (Again.)
3746 // No safepoint in the call to thaw, so even an oop return value should be OK.
3747 __ pop_d(xmm0);
3748 __ pop(rax);
3749 } else {
3750 // Return 0 (success) from doYield.
3751 __ xorptr(rax, rax);
3752 }
3753
3754 // After thawing, rbx is the SP of the yielding frame.
3755 // Move there, and then to saved RBP slot.
3756 __ movptr(rsp, rbx);
3757 __ subptr(rsp, 2*wordSize);
3758
3759 if (return_barrier_exception) {
3760 __ movptr(c_rarg0, r15_thread);
3761 __ movptr(c_rarg1, Address(rsp, wordSize)); // return address
3762
3763 // rax still holds the original exception oop, save it before the call
3764 __ push(rax);
3765
3766 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), 2);
3767 __ movptr(rbx, rax);
3768
3769 // Continue at exception handler:
3770 // rax: exception oop
3771 // rbx: exception handler
3772 // rdx: exception pc
3773 __ pop(rax);
3774 __ verify_oop(rax);
3775 __ pop(rbp); // pop out RBP here too
3776 __ pop(rdx);
3777 __ jmp(rbx);
3778 } else {
3779 // We are "returning" into the topmost thawed frame; see Thaw::push_return_frame
3780 __ pop(rbp);
3781 __ ret(0);
3782 }
3783
3784 return start;
3785 }
3786
3787 address StubGenerator::generate_cont_thaw() {
3788 return generate_cont_thaw("Cont thaw", Continuation::thaw_top);
3789 }
3790
3791 // TODO: will probably need multiple return barriers depending on return type
3792
3793 address StubGenerator::generate_cont_returnBarrier() {
3794 return generate_cont_thaw("Cont thaw return barrier", Continuation::thaw_return_barrier);
3795 }
3796
3797 address StubGenerator::generate_cont_returnBarrier_exception() {
3798 return generate_cont_thaw("Cont thaw return barrier exception", Continuation::thaw_return_barrier_exception);
3799 }
3800
3801 address StubGenerator::generate_cont_preempt_stub() {
3802 if (!Continuations::enabled()) return nullptr;
3803 StubCodeMark mark(this, "StubRoutines","Continuation preempt stub");
3804 address start = __ pc();
3805
3806 __ reset_last_Java_frame(true);
3807
3808 // Set rsp to enterSpecial frame, i.e. remove all frames copied into the heap.
3809 __ movptr(rsp, Address(r15_thread, JavaThread::cont_entry_offset()));
3810
3811 Label preemption_cancelled;
3812 __ movbool(rscratch1, Address(r15_thread, JavaThread::preemption_cancelled_offset()));
3813 __ testbool(rscratch1);
3814 __ jcc(Assembler::notZero, preemption_cancelled);
3815
3816 // Remove enterSpecial frame from the stack and return to Continuation.run() to unmount.
3817 SharedRuntime::continuation_enter_cleanup(_masm);
3818 __ pop(rbp);
3819 __ ret(0);
3820
3821 // We acquired the monitor after freezing the frames so call thaw to continue execution.
3822 __ bind(preemption_cancelled);
3823 __ movbool(Address(r15_thread, JavaThread::preemption_cancelled_offset()), false);
3824 __ lea(rbp, Address(rsp, checked_cast<int32_t>(ContinuationEntry::size())));
3825 __ movptr(rscratch1, ExternalAddress(ContinuationEntry::thaw_call_pc_address()));
3826 __ jmp(rscratch1);
3827
3828 return start;
3829 }
3830
3831 // exception handler for upcall stubs
3832 address StubGenerator::generate_upcall_stub_exception_handler() {
3833 StubCodeMark mark(this, "StubRoutines", "upcall stub exception handler");
3834 address start = __ pc();
3835
3836 // native caller has no idea how to handle exceptions
3837 // we just crash here. Up to callee to catch exceptions.
3838 __ verify_oop(rax);
3839 __ vzeroupper();
3840 __ mov(c_rarg0, rax);
3841 __ andptr(rsp, -StackAlignmentInBytes); // align stack as required by ABI
3842 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows
3843 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, UpcallLinker::handle_uncaught_exception)));
3844 __ should_not_reach_here();
3845
3846 return start;
3847 }
3848
3849 // load Method* target of MethodHandle
3850 // j_rarg0 = jobject receiver
3851 // rbx = result
3852 address StubGenerator::generate_upcall_stub_load_target() {
3853 StubCodeMark mark(this, "StubRoutines", "upcall_stub_load_target");
3854 address start = __ pc();
3855
3856 __ resolve_global_jobject(j_rarg0, r15_thread, rscratch1);
3857 // Load target method from receiver
3858 __ load_heap_oop(rbx, Address(j_rarg0, java_lang_invoke_MethodHandle::form_offset()), rscratch1);
3859 __ load_heap_oop(rbx, Address(rbx, java_lang_invoke_LambdaForm::vmentry_offset()), rscratch1);
3860 __ load_heap_oop(rbx, Address(rbx, java_lang_invoke_MemberName::method_offset()), rscratch1);
3861 __ access_load_at(T_ADDRESS, IN_HEAP, rbx,
3862 Address(rbx, java_lang_invoke_ResolvedMethodName::vmtarget_offset()),
3863 noreg, noreg);
3864 __ movptr(Address(r15_thread, JavaThread::callee_target_offset()), rbx); // just in case callee is deoptimized
3865
3866 __ ret(0);
3867
3868 return start;
3869 }
3870
3871 address StubGenerator::generate_lookup_secondary_supers_table_stub(u1 super_klass_index) {
3872 StubCodeMark mark(this, "StubRoutines", "lookup_secondary_supers_table");
3873
3874 address start = __ pc();
3875
3876 const Register
3877 r_super_klass = rax,
3878 r_sub_klass = rsi,
3879 result = rdi;
3880
3881 __ lookup_secondary_supers_table_const(r_sub_klass, r_super_klass,
3882 rdx, rcx, rbx, r11, // temps
3883 result,
3884 super_klass_index);
3885 __ ret(0);
3886
3887 return start;
3888 }
3889
3890 // Slow path implementation for UseSecondarySupersTable.
3891 address StubGenerator::generate_lookup_secondary_supers_table_slow_path_stub() {
3892 StubCodeMark mark(this, "StubRoutines", "lookup_secondary_supers_table");
3893
3894 address start = __ pc();
3895
3896 const Register
3897 r_super_klass = rax,
3898 r_array_base = rbx,
3899 r_array_index = rdx,
3900 r_sub_klass = rsi,
3901 r_bitmap = r11,
3902 result = rdi;
3903
3904 Label L_success;
3905 __ lookup_secondary_supers_table_slow_path(r_super_klass, r_array_base, r_array_index, r_bitmap,
3906 rcx, rdi, // temps
3907 &L_success);
3908 // bind(L_failure);
3909 __ movl(result, 1);
3910 __ ret(0);
3911
3912 __ bind(L_success);
3913 __ movl(result, 0);
3914 __ ret(0);
3915
3916 return start;
3917 }
3918
3919 void StubGenerator::create_control_words() {
3920 // Round to nearest, 64-bit mode, exceptions masked, flags specialized
3921 StubRoutines::x86::_mxcsr_std = EnableX86ECoreOpts ? 0x1FBF : 0x1F80;
3922 // Round to zero, 64-bit mode, exceptions masked, flags specialized
3923 StubRoutines::x86::_mxcsr_rz = EnableX86ECoreOpts ? 0x7FBF : 0x7F80;
3924 }
3925
3926 // Initialization
3927 void StubGenerator::generate_initial_stubs() {
3928 // Generates all stubs and initializes the entry points
3929
3930 // This platform-specific settings are needed by generate_call_stub()
3931 create_control_words();
3932
3933 // Initialize table for unsafe copy memeory check.
3934 if (UnsafeMemoryAccess::_table == nullptr) {
3935 UnsafeMemoryAccess::create_table(16 + 4); // 16 for copyMemory; 4 for setMemory
3936 }
3937
3938 // entry points that exist in all platforms Note: This is code
3939 // that could be shared among different platforms - however the
3940 // benefit seems to be smaller than the disadvantage of having a
3941 // much more complicated generator structure. See also comment in
3942 // stubRoutines.hpp.
3943
3944 StubRoutines::_forward_exception_entry = generate_forward_exception();
3945
3946 StubRoutines::_call_stub_entry =
3947 generate_call_stub(StubRoutines::_call_stub_return_address);
3948
3949 // is referenced by megamorphic call
3950 StubRoutines::_catch_exception_entry = generate_catch_exception();
3951
3952 // atomic calls
3953 StubRoutines::_fence_entry = generate_orderaccess_fence();
3954
3955 // platform dependent
3956 StubRoutines::x86::_get_previous_sp_entry = generate_get_previous_sp();
3957
3958 StubRoutines::x86::_verify_mxcsr_entry = generate_verify_mxcsr();
3959
3960 StubRoutines::x86::_f2i_fixup = generate_f2i_fixup();
3961 StubRoutines::x86::_f2l_fixup = generate_f2l_fixup();
3962 StubRoutines::x86::_d2i_fixup = generate_d2i_fixup();
3963 StubRoutines::x86::_d2l_fixup = generate_d2l_fixup();
3964
3965 StubRoutines::x86::_float_sign_mask = generate_fp_mask("float_sign_mask", 0x7FFFFFFF7FFFFFFF);
3966 StubRoutines::x86::_float_sign_flip = generate_fp_mask("float_sign_flip", 0x8000000080000000);
3967 StubRoutines::x86::_double_sign_mask = generate_fp_mask("double_sign_mask", 0x7FFFFFFFFFFFFFFF);
3968 StubRoutines::x86::_double_sign_flip = generate_fp_mask("double_sign_flip", 0x8000000000000000);
3969
3970 if (UseCRC32Intrinsics) {
3971 // set table address before stub generation which use it
3972 StubRoutines::_crc_table_adr = (address)StubRoutines::x86::_crc_table;
3973 StubRoutines::_updateBytesCRC32 = generate_updateBytesCRC32();
3974 }
3975
3976 if (UseCRC32CIntrinsics) {
3977 bool supports_clmul = VM_Version::supports_clmul();
3978 StubRoutines::x86::generate_CRC32C_table(supports_clmul);
3979 StubRoutines::_crc32c_table_addr = (address)StubRoutines::x86::_crc32c_table;
3980 StubRoutines::_updateBytesCRC32C = generate_updateBytesCRC32C(supports_clmul);
3981 }
3982
3983 if (VM_Version::supports_float16()) {
3984 // For results consistency both intrinsics should be enabled.
3985 // vmIntrinsics checks InlineIntrinsics flag, no need to check it here.
3986 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_float16ToFloat) &&
3987 vmIntrinsics::is_intrinsic_available(vmIntrinsics::_floatToFloat16)) {
3988 StubRoutines::_hf2f = generate_float16ToFloat();
3989 StubRoutines::_f2hf = generate_floatToFloat16();
3990 }
3991 }
3992
3993 generate_libm_stubs();
3994
3995 StubRoutines::_fmod = generate_libmFmod(); // from stubGenerator_x86_64_fmod.cpp
3996 }
3997
3998 void StubGenerator::generate_continuation_stubs() {
3999 // Continuation stubs:
4000 StubRoutines::_cont_thaw = generate_cont_thaw();
4001 StubRoutines::_cont_returnBarrier = generate_cont_returnBarrier();
4002 StubRoutines::_cont_returnBarrierExc = generate_cont_returnBarrier_exception();
4003 StubRoutines::_cont_preempt_stub = generate_cont_preempt_stub();
4004 }
4005
4006 void StubGenerator::generate_final_stubs() {
4007 // Generates the rest of stubs and initializes the entry points
4008
4009 // support for verify_oop (must happen after universe_init)
4010 if (VerifyOops) {
4011 StubRoutines::_verify_oop_subroutine_entry = generate_verify_oop();
4012 }
4013
4014 // data cache line writeback
4015 StubRoutines::_data_cache_writeback = generate_data_cache_writeback();
4016 StubRoutines::_data_cache_writeback_sync = generate_data_cache_writeback_sync();
4017
4018 // arraycopy stubs used by compilers
4019 generate_arraycopy_stubs();
4020
4021 BarrierSetNMethod* bs_nm = BarrierSet::barrier_set()->barrier_set_nmethod();
4022 if (bs_nm != nullptr) {
4023 StubRoutines::_method_entry_barrier = generate_method_entry_barrier();
4024 }
4025
4026 if (UseVectorizedMismatchIntrinsic) {
4027 StubRoutines::_vectorizedMismatch = generate_vectorizedMismatch();
4028 }
4029
4030 StubRoutines::_upcall_stub_exception_handler = generate_upcall_stub_exception_handler();
4031 StubRoutines::_upcall_stub_load_target = generate_upcall_stub_load_target();
4032 }
4033
4034 void StubGenerator::generate_compiler_stubs() {
4035 #if COMPILER2_OR_JVMCI
4036
4037 // Entry points that are C2 compiler specific.
4038
4039 StubRoutines::x86::_vector_float_sign_mask = generate_vector_mask("vector_float_sign_mask", 0x7FFFFFFF7FFFFFFF);
4040 StubRoutines::x86::_vector_float_sign_flip = generate_vector_mask("vector_float_sign_flip", 0x8000000080000000);
4041 StubRoutines::x86::_vector_double_sign_mask = generate_vector_mask("vector_double_sign_mask", 0x7FFFFFFFFFFFFFFF);
4042 StubRoutines::x86::_vector_double_sign_flip = generate_vector_mask("vector_double_sign_flip", 0x8000000000000000);
4043 StubRoutines::x86::_vector_all_bits_set = generate_vector_mask("vector_all_bits_set", 0xFFFFFFFFFFFFFFFF);
4044 StubRoutines::x86::_vector_int_mask_cmp_bits = generate_vector_mask("vector_int_mask_cmp_bits", 0x0000000100000001);
4045 StubRoutines::x86::_vector_short_to_byte_mask = generate_vector_mask("vector_short_to_byte_mask", 0x00ff00ff00ff00ff);
4046 StubRoutines::x86::_vector_byte_perm_mask = generate_vector_byte_perm_mask("vector_byte_perm_mask");
4047 StubRoutines::x86::_vector_int_to_byte_mask = generate_vector_mask("vector_int_to_byte_mask", 0x000000ff000000ff);
4048 StubRoutines::x86::_vector_int_to_short_mask = generate_vector_mask("vector_int_to_short_mask", 0x0000ffff0000ffff);
4049 StubRoutines::x86::_vector_32_bit_mask = generate_vector_custom_i32("vector_32_bit_mask", Assembler::AVX_512bit,
4050 0xFFFFFFFF, 0, 0, 0);
4051 StubRoutines::x86::_vector_64_bit_mask = generate_vector_custom_i32("vector_64_bit_mask", Assembler::AVX_512bit,
4052 0xFFFFFFFF, 0xFFFFFFFF, 0, 0);
4053 StubRoutines::x86::_vector_int_shuffle_mask = generate_vector_mask("vector_int_shuffle_mask", 0x0302010003020100);
4054 StubRoutines::x86::_vector_byte_shuffle_mask = generate_vector_byte_shuffle_mask("vector_byte_shuffle_mask");
4055 StubRoutines::x86::_vector_short_shuffle_mask = generate_vector_mask("vector_short_shuffle_mask", 0x0100010001000100);
4056 StubRoutines::x86::_vector_long_shuffle_mask = generate_vector_mask("vector_long_shuffle_mask", 0x0000000100000000);
4057 StubRoutines::x86::_vector_long_sign_mask = generate_vector_mask("vector_long_sign_mask", 0x8000000000000000);
4058 StubRoutines::x86::_vector_iota_indices = generate_iota_indices("iota_indices");
4059 StubRoutines::x86::_vector_count_leading_zeros_lut = generate_count_leading_zeros_lut("count_leading_zeros_lut");
4060 StubRoutines::x86::_vector_reverse_bit_lut = generate_vector_reverse_bit_lut("reverse_bit_lut");
4061 StubRoutines::x86::_vector_reverse_byte_perm_mask_long = generate_vector_reverse_byte_perm_mask_long("perm_mask_long");
4062 StubRoutines::x86::_vector_reverse_byte_perm_mask_int = generate_vector_reverse_byte_perm_mask_int("perm_mask_int");
4063 StubRoutines::x86::_vector_reverse_byte_perm_mask_short = generate_vector_reverse_byte_perm_mask_short("perm_mask_short");
4064
4065 if (VM_Version::supports_avx2() && !VM_Version::supports_avx512vl()) {
4066 StubRoutines::x86::_compress_perm_table32 = generate_compress_perm_table("compress_perm_table32", 32);
4067 StubRoutines::x86::_compress_perm_table64 = generate_compress_perm_table("compress_perm_table64", 64);
4068 StubRoutines::x86::_expand_perm_table32 = generate_expand_perm_table("expand_perm_table32", 32);
4069 StubRoutines::x86::_expand_perm_table64 = generate_expand_perm_table("expand_perm_table64", 64);
4070 }
4071
4072 if (VM_Version::supports_avx2() && !VM_Version::supports_avx512_vpopcntdq()) {
4073 // lut implementation influenced by counting 1s algorithm from section 5-1 of Hackers' Delight.
4074 StubRoutines::x86::_vector_popcount_lut = generate_popcount_avx_lut("popcount_lut");
4075 }
4076
4077 generate_aes_stubs();
4078
4079 generate_ghash_stubs();
4080
4081 generate_chacha_stubs();
4082
4083 generate_sha3_stubs();
4084
4085 #ifdef COMPILER2
4086 if ((UseAVX == 2) && EnableX86ECoreOpts) {
4087 generate_string_indexof(StubRoutines::_string_indexof_array);
4088 }
4089 #endif
4090
4091 if (UseAdler32Intrinsics) {
4092 StubRoutines::_updateBytesAdler32 = generate_updateBytesAdler32();
4093 }
4094
4095 if (UsePoly1305Intrinsics) {
4096 StubRoutines::_poly1305_processBlocks = generate_poly1305_processBlocks();
4097 }
4098
4099 if (UseIntPolyIntrinsics) {
4100 StubRoutines::_intpoly_montgomeryMult_P256 = generate_intpoly_montgomeryMult_P256();
4101 StubRoutines::_intpoly_assign = generate_intpoly_assign();
4102 }
4103
4104 if (UseMD5Intrinsics) {
4105 StubRoutines::_md5_implCompress = generate_md5_implCompress(false, "md5_implCompress");
4106 StubRoutines::_md5_implCompressMB = generate_md5_implCompress(true, "md5_implCompressMB");
4107 }
4108
4109 if (UseSHA1Intrinsics) {
4110 StubRoutines::x86::_upper_word_mask_addr = generate_upper_word_mask();
4111 StubRoutines::x86::_shuffle_byte_flip_mask_addr = generate_shuffle_byte_flip_mask();
4112 StubRoutines::_sha1_implCompress = generate_sha1_implCompress(false, "sha1_implCompress");
4113 StubRoutines::_sha1_implCompressMB = generate_sha1_implCompress(true, "sha1_implCompressMB");
4114 }
4115
4116 if (UseSHA256Intrinsics) {
4117 StubRoutines::x86::_k256_adr = (address)StubRoutines::x86::_k256;
4118 char* dst = (char*)StubRoutines::x86::_k256_W;
4119 char* src = (char*)StubRoutines::x86::_k256;
4120 for (int ii = 0; ii < 16; ++ii) {
4121 memcpy(dst + 32 * ii, src + 16 * ii, 16);
4122 memcpy(dst + 32 * ii + 16, src + 16 * ii, 16);
4123 }
4124 StubRoutines::x86::_k256_W_adr = (address)StubRoutines::x86::_k256_W;
4125 StubRoutines::x86::_pshuffle_byte_flip_mask_addr = generate_pshuffle_byte_flip_mask();
4126 StubRoutines::_sha256_implCompress = generate_sha256_implCompress(false, "sha256_implCompress");
4127 StubRoutines::_sha256_implCompressMB = generate_sha256_implCompress(true, "sha256_implCompressMB");
4128 }
4129
4130 if (UseSHA512Intrinsics) {
4131 StubRoutines::x86::_k512_W_addr = (address)StubRoutines::x86::_k512_W;
4132 StubRoutines::x86::_pshuffle_byte_flip_mask_addr_sha512 = generate_pshuffle_byte_flip_mask_sha512();
4133 StubRoutines::_sha512_implCompress = generate_sha512_implCompress(false, "sha512_implCompress");
4134 StubRoutines::_sha512_implCompressMB = generate_sha512_implCompress(true, "sha512_implCompressMB");
4135 }
4136
4137 if (UseBASE64Intrinsics) {
4138 if(VM_Version::supports_avx2()) {
4139 StubRoutines::x86::_avx2_shuffle_base64 = base64_avx2_shuffle_addr();
4140 StubRoutines::x86::_avx2_input_mask_base64 = base64_avx2_input_mask_addr();
4141 StubRoutines::x86::_avx2_lut_base64 = base64_avx2_lut_addr();
4142 StubRoutines::x86::_avx2_decode_tables_base64 = base64_AVX2_decode_tables_addr();
4143 StubRoutines::x86::_avx2_decode_lut_tables_base64 = base64_AVX2_decode_LUT_tables_addr();
4144 }
4145 StubRoutines::x86::_encoding_table_base64 = base64_encoding_table_addr();
4146 if (VM_Version::supports_avx512_vbmi()) {
4147 StubRoutines::x86::_shuffle_base64 = base64_shuffle_addr();
4148 StubRoutines::x86::_lookup_lo_base64 = base64_vbmi_lookup_lo_addr();
4149 StubRoutines::x86::_lookup_hi_base64 = base64_vbmi_lookup_hi_addr();
4150 StubRoutines::x86::_lookup_lo_base64url = base64_vbmi_lookup_lo_url_addr();
4151 StubRoutines::x86::_lookup_hi_base64url = base64_vbmi_lookup_hi_url_addr();
4152 StubRoutines::x86::_pack_vec_base64 = base64_vbmi_pack_vec_addr();
4153 StubRoutines::x86::_join_0_1_base64 = base64_vbmi_join_0_1_addr();
4154 StubRoutines::x86::_join_1_2_base64 = base64_vbmi_join_1_2_addr();
4155 StubRoutines::x86::_join_2_3_base64 = base64_vbmi_join_2_3_addr();
4156 }
4157 StubRoutines::x86::_decoding_table_base64 = base64_decoding_table_addr();
4158 StubRoutines::_base64_encodeBlock = generate_base64_encodeBlock();
4159 StubRoutines::_base64_decodeBlock = generate_base64_decodeBlock();
4160 }
4161
4162 #ifdef COMPILER2
4163 if (UseMultiplyToLenIntrinsic) {
4164 StubRoutines::_multiplyToLen = generate_multiplyToLen();
4165 }
4166 if (UseSquareToLenIntrinsic) {
4167 StubRoutines::_squareToLen = generate_squareToLen();
4168 }
4169 if (UseMulAddIntrinsic) {
4170 StubRoutines::_mulAdd = generate_mulAdd();
4171 }
4172 if (VM_Version::supports_avx512_vbmi2()) {
4173 StubRoutines::_bigIntegerRightShiftWorker = generate_bigIntegerRightShift();
4174 StubRoutines::_bigIntegerLeftShiftWorker = generate_bigIntegerLeftShift();
4175 }
4176 if (UseSecondarySupersTable) {
4177 StubRoutines::_lookup_secondary_supers_table_slow_path_stub = generate_lookup_secondary_supers_table_slow_path_stub();
4178 if (! InlineSecondarySupersTest) {
4179 for (int slot = 0; slot < Klass::SECONDARY_SUPERS_TABLE_SIZE; slot++) {
4180 StubRoutines::_lookup_secondary_supers_table_stubs[slot] = generate_lookup_secondary_supers_table_stub(slot);
4181 }
4182 }
4183 }
4184 if (UseMontgomeryMultiplyIntrinsic) {
4185 StubRoutines::_montgomeryMultiply
4186 = CAST_FROM_FN_PTR(address, SharedRuntime::montgomery_multiply);
4187 }
4188 if (UseMontgomerySquareIntrinsic) {
4189 StubRoutines::_montgomerySquare
4190 = CAST_FROM_FN_PTR(address, SharedRuntime::montgomery_square);
4191 }
4192
4193 // Load x86_64_sort library on supported hardware to enable SIMD sort and partition intrinsics
4194
4195 if (VM_Version::is_intel() && (VM_Version::supports_avx512dq() || VM_Version::supports_avx2())) {
4196 void *libsimdsort = nullptr;
4197 char ebuf_[1024];
4198 char dll_name_simd_sort[JVM_MAXPATHLEN];
4199 if (os::dll_locate_lib(dll_name_simd_sort, sizeof(dll_name_simd_sort), Arguments::get_dll_dir(), "simdsort")) {
4200 libsimdsort = os::dll_load(dll_name_simd_sort, ebuf_, sizeof ebuf_);
4201 }
4202 // Get addresses for SIMD sort and partition routines
4203 if (libsimdsort != nullptr) {
4204 log_info(library)("Loaded library %s, handle " INTPTR_FORMAT, JNI_LIB_PREFIX "simdsort" JNI_LIB_SUFFIX, p2i(libsimdsort));
4205
4206 snprintf(ebuf_, sizeof(ebuf_), VM_Version::supports_avx512dq() ? "avx512_sort" : "avx2_sort");
4207 StubRoutines::_array_sort = (address)os::dll_lookup(libsimdsort, ebuf_);
4208
4209 snprintf(ebuf_, sizeof(ebuf_), VM_Version::supports_avx512dq() ? "avx512_partition" : "avx2_partition");
4210 StubRoutines::_array_partition = (address)os::dll_lookup(libsimdsort, ebuf_);
4211 }
4212 }
4213
4214 // Get svml stub routine addresses
4215 void *libjsvml = nullptr;
4216 char ebuf[1024];
4217 char dll_name[JVM_MAXPATHLEN];
4218 if (os::dll_locate_lib(dll_name, sizeof(dll_name), Arguments::get_dll_dir(), "jsvml")) {
4219 libjsvml = os::dll_load(dll_name, ebuf, sizeof ebuf);
4220 }
4221 if (libjsvml != nullptr) {
4222 // SVML method naming convention
4223 // All the methods are named as __jsvml_op<T><N>_ha_<VV>
4224 // Where:
4225 // ha stands for high accuracy
4226 // <T> is optional to indicate float/double
4227 // Set to f for vector float operation
4228 // Omitted for vector double operation
4229 // <N> is the number of elements in the vector
4230 // 1, 2, 4, 8, 16
4231 // e.g. 128 bit float vector has 4 float elements
4232 // <VV> indicates the avx/sse level:
4233 // z0 is AVX512, l9 is AVX2, e9 is AVX1 and ex is for SSE2
4234 // e.g. __jsvml_expf16_ha_z0 is the method for computing 16 element vector float exp using AVX 512 insns
4235 // __jsvml_exp8_ha_z0 is the method for computing 8 element vector double exp using AVX 512 insns
4236
4237 log_info(library)("Loaded library %s, handle " INTPTR_FORMAT, JNI_LIB_PREFIX "jsvml" JNI_LIB_SUFFIX, p2i(libjsvml));
4238 if (UseAVX > 2) {
4239 for (int op = 0; op < VectorSupport::NUM_VECTOR_OP_MATH; op++) {
4240 int vop = VectorSupport::VECTOR_OP_MATH_START + op;
4241 if ((!VM_Version::supports_avx512dq()) &&
4242 (vop == VectorSupport::VECTOR_OP_LOG || vop == VectorSupport::VECTOR_OP_LOG10 || vop == VectorSupport::VECTOR_OP_POW)) {
4243 continue;
4244 }
4245 snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf16_ha_z0", VectorSupport::mathname[op]);
4246 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_512][op] = (address)os::dll_lookup(libjsvml, ebuf);
4247
4248 snprintf(ebuf, sizeof(ebuf), "__jsvml_%s8_ha_z0", VectorSupport::mathname[op]);
4249 StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_512][op] = (address)os::dll_lookup(libjsvml, ebuf);
4250 }
4251 }
4252 const char* avx_sse_str = (UseAVX >= 2) ? "l9" : ((UseAVX == 1) ? "e9" : "ex");
4253 for (int op = 0; op < VectorSupport::NUM_VECTOR_OP_MATH; op++) {
4254 int vop = VectorSupport::VECTOR_OP_MATH_START + op;
4255 if (vop == VectorSupport::VECTOR_OP_POW) {
4256 continue;
4257 }
4258 snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf4_ha_%s", VectorSupport::mathname[op], avx_sse_str);
4259 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_64][op] = (address)os::dll_lookup(libjsvml, ebuf);
4260
4261 snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf4_ha_%s", VectorSupport::mathname[op], avx_sse_str);
4262 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_128][op] = (address)os::dll_lookup(libjsvml, ebuf);
4263
4264 snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf8_ha_%s", VectorSupport::mathname[op], avx_sse_str);
4265 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_256][op] = (address)os::dll_lookup(libjsvml, ebuf);
4266
4267 snprintf(ebuf, sizeof(ebuf), "__jsvml_%s1_ha_%s", VectorSupport::mathname[op], avx_sse_str);
4268 StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_64][op] = (address)os::dll_lookup(libjsvml, ebuf);
4269
4270 snprintf(ebuf, sizeof(ebuf), "__jsvml_%s2_ha_%s", VectorSupport::mathname[op], avx_sse_str);
4271 StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_128][op] = (address)os::dll_lookup(libjsvml, ebuf);
4272
4273 snprintf(ebuf, sizeof(ebuf), "__jsvml_%s4_ha_%s", VectorSupport::mathname[op], avx_sse_str);
4274 StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_256][op] = (address)os::dll_lookup(libjsvml, ebuf);
4275 }
4276 }
4277
4278 #endif // COMPILER2
4279 #endif // COMPILER2_OR_JVMCI
4280 }
4281
4282 StubGenerator::StubGenerator(CodeBuffer* code, StubsKind kind) : StubCodeGenerator(code) {
4283 DEBUG_ONLY( _regs_in_thread = false; )
4284 switch(kind) {
4285 case Initial_stubs:
4286 generate_initial_stubs();
4287 break;
4288 case Continuation_stubs:
4289 generate_continuation_stubs();
4290 break;
4291 case Compiler_stubs:
4292 generate_compiler_stubs();
4293 break;
4294 case Final_stubs:
4295 generate_final_stubs();
4296 break;
4297 default:
4298 fatal("unexpected stubs kind: %d", kind);
4299 break;
4300 };
4301 }
4302
4303 void StubGenerator_generate(CodeBuffer* code, StubCodeGenerator::StubsKind kind) {
4304 StubGenerator g(code, kind);
4305 }
4306
4307 #undef __