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/assembler.hpp"
27 #include "asm/macroAssembler.hpp"
28 #include "classfile/javaClasses.hpp"
29 #include "classfile/vmIntrinsics.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/javaThread.hpp"
40 #include "runtime/sharedRuntime.hpp"
41 #include "runtime/stubRoutines.hpp"
42 #include "utilities/macros.hpp"
43 #include "vmreg_x86.inline.hpp"
44 #include "stubGenerator_x86_64.hpp"
45 #ifdef COMPILER2
46 #include "opto/runtime.hpp"
47 #include "opto/c2_globals.hpp"
48 #endif
49 #if INCLUDE_JVMCI
50 #include "jvmci/jvmci_globals.hpp"
51 #endif
52
53 // For a more detailed description of the stub routine structure
54 // see the comment in stubRoutines.hpp
55
56 #define __ _masm->
57 #define TIMES_OOP (UseCompressedOops ? Address::times_4 : Address::times_8)
58
59 #ifdef PRODUCT
60 #define BLOCK_COMMENT(str) /* nothing */
61 #else
62 #define BLOCK_COMMENT(str) __ block_comment(str)
63 #endif // PRODUCT
64
65 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
66
67 //
68 // Linux Arguments:
69 // c_rarg0: call wrapper address address
70 // c_rarg1: result address
71 // c_rarg2: result type BasicType
72 // c_rarg3: method Method*
73 // c_rarg4: (interpreter) entry point address
74 // c_rarg5: parameters intptr_t*
75 // 16(rbp): parameter size (in words) int
76 // 24(rbp): thread Thread*
77 //
78 // [ return_from_Java ] <--- rsp
79 // [ argument word n ]
80 // ...
81 // -12 [ argument word 1 ]
82 // -11 [ saved r15 ] <--- rsp_after_call
83 // -10 [ saved r14 ]
84 // -9 [ saved r13 ]
85 // -8 [ saved r12 ]
86 // -7 [ saved rbx ]
87 // -6 [ call wrapper ]
88 // -5 [ result ]
89 // -4 [ result type ]
90 // -3 [ method ]
91 // -2 [ entry point ]
92 // -1 [ parameters ]
93 // 0 [ saved rbp ] <--- rbp
94 // 1 [ return address ]
95 // 2 [ parameter size ]
96 // 3 [ thread ]
97 //
98 // Windows Arguments:
99 // c_rarg0: call wrapper address address
100 // c_rarg1: result address
101 // c_rarg2: result type BasicType
102 // c_rarg3: method Method*
103 // 48(rbp): (interpreter) entry point address
104 // 56(rbp): parameters intptr_t*
105 // 64(rbp): parameter size (in words) int
106 // 72(rbp): thread Thread*
107 //
108 // [ return_from_Java ] <--- rsp
109 // [ argument word n ]
110 // ...
111 // -28 [ argument word 1 ]
112 // -27 [ saved xmm15 ] <--- rsp after_call
113 // [ saved xmm7-xmm14 ]
114 // -9 [ saved xmm6 ] (each xmm register takes 2 slots)
115 // -7 [ saved r15 ]
116 // -6 [ saved r14 ]
117 // -5 [ saved r13 ]
118 // -4 [ saved r12 ]
119 // -3 [ saved rdi ]
120 // -2 [ saved rsi ]
121 // -1 [ saved rbx ]
122 // 0 [ saved rbp ] <--- rbp
123 // 1 [ return address ]
124 // 2 [ call wrapper ]
125 // 3 [ result ]
126 // 4 [ result type ]
127 // 5 [ method ]
128 // 6 [ entry point ]
129 // 7 [ parameters ]
130 // 8 [ parameter size ]
131 // 9 [ thread ]
132 //
133 // Windows reserves the callers stack space for arguments 1-4.
134 // We spill c_rarg0-c_rarg3 to this space.
135
136 // Call stub stack layout word offsets from rbp
137 #ifdef _WIN64
138 enum call_stub_layout {
139 xmm_save_first = 6, // save from xmm6
140 xmm_save_last = 15, // to xmm15
141 xmm_save_base = -9,
142 rsp_after_call_off = xmm_save_base - 2 * (xmm_save_last - xmm_save_first), // -27
143 r15_off = -7,
144 r14_off = -6,
145 r13_off = -5,
146 r12_off = -4,
147 rdi_off = -3,
148 rsi_off = -2,
149 rbx_off = -1,
150 rbp_off = 0,
151 retaddr_off = 1,
152 call_wrapper_off = 2,
153 result_off = 3,
154 result_type_off = 4,
155 method_off = 5,
156 entry_point_off = 6,
157 parameters_off = 7,
158 parameter_size_off = 8,
159 thread_off = 9
160 };
161
162 static Address xmm_save(int reg) {
163 assert(reg >= xmm_save_first && reg <= xmm_save_last, "XMM register number out of range");
164 return Address(rbp, (xmm_save_base - (reg - xmm_save_first) * 2) * wordSize);
165 }
166 #else // !_WIN64
167 enum call_stub_layout {
168 rsp_after_call_off = -12,
169 mxcsr_off = rsp_after_call_off,
170 r15_off = -11,
171 r14_off = -10,
172 r13_off = -9,
173 r12_off = -8,
174 rbx_off = -7,
175 call_wrapper_off = -6,
176 result_off = -5,
177 result_type_off = -4,
178 method_off = -3,
179 entry_point_off = -2,
180 parameters_off = -1,
181 rbp_off = 0,
182 retaddr_off = 1,
183 parameter_size_off = 2,
184 thread_off = 3
185 };
186 #endif // _WIN64
187
188 address StubGenerator::generate_call_stub(address& return_address) {
189
190 assert((int)frame::entry_frame_after_call_words == -(int)rsp_after_call_off + 1 &&
191 (int)frame::entry_frame_call_wrapper_offset == (int)call_wrapper_off,
192 "adjust this code");
193 StubCodeMark mark(this, "StubRoutines", "call_stub");
194 address start = __ pc();
195
196 // same as in generate_catch_exception()!
197 const Address rsp_after_call(rbp, rsp_after_call_off * wordSize);
198
199 const Address call_wrapper (rbp, call_wrapper_off * wordSize);
200 const Address result (rbp, result_off * wordSize);
201 const Address result_type (rbp, result_type_off * wordSize);
202 const Address method (rbp, method_off * wordSize);
203 const Address entry_point (rbp, entry_point_off * wordSize);
204 const Address parameters (rbp, parameters_off * wordSize);
205 const Address parameter_size(rbp, parameter_size_off * wordSize);
206
207 // same as in generate_catch_exception()!
208 const Address thread (rbp, thread_off * wordSize);
209
210 const Address r15_save(rbp, r15_off * wordSize);
211 const Address r14_save(rbp, r14_off * wordSize);
212 const Address r13_save(rbp, r13_off * wordSize);
213 const Address r12_save(rbp, r12_off * wordSize);
214 const Address rbx_save(rbp, rbx_off * wordSize);
215
216 // stub code
217 __ enter();
218 __ subptr(rsp, -rsp_after_call_off * wordSize);
219
220 // save register parameters
221 #ifndef _WIN64
222 __ movptr(parameters, c_rarg5); // parameters
223 __ movptr(entry_point, c_rarg4); // entry_point
224 #endif
225
226 __ movptr(method, c_rarg3); // method
227 __ movl(result_type, c_rarg2); // result type
228 __ movptr(result, c_rarg1); // result
229 __ movptr(call_wrapper, c_rarg0); // call wrapper
230
231 // save regs belonging to calling function
232 __ movptr(rbx_save, rbx);
233 __ movptr(r12_save, r12);
234 __ movptr(r13_save, r13);
235 __ movptr(r14_save, r14);
236 __ movptr(r15_save, r15);
237
238 #ifdef _WIN64
239 int last_reg = 15;
240 for (int i = xmm_save_first; i <= last_reg; i++) {
241 __ movdqu(xmm_save(i), as_XMMRegister(i));
242 }
243
244 const Address rdi_save(rbp, rdi_off * wordSize);
245 const Address rsi_save(rbp, rsi_off * wordSize);
246
247 __ movptr(rsi_save, rsi);
248 __ movptr(rdi_save, rdi);
249 #else
250 const Address mxcsr_save(rbp, mxcsr_off * wordSize);
251 {
252 Label skip_ldmx;
253 __ stmxcsr(mxcsr_save);
254 __ movl(rax, mxcsr_save);
255 __ andl(rax, 0xFFC0); // Mask out any pending exceptions (only check control and mask bits)
256 ExternalAddress mxcsr_std(StubRoutines::x86::addr_mxcsr_std());
257 __ cmp32(rax, mxcsr_std, rscratch1);
258 __ jcc(Assembler::equal, skip_ldmx);
259 __ ldmxcsr(mxcsr_std, rscratch1);
260 __ bind(skip_ldmx);
261 }
262 #endif
263
264 // Load up thread register
265 __ movptr(r15_thread, thread);
266 __ reinit_heapbase();
267
268 #ifdef ASSERT
269 // make sure we have no pending exceptions
270 {
271 Label L;
272 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), NULL_WORD);
273 __ jcc(Assembler::equal, L);
274 __ stop("StubRoutines::call_stub: entered with pending exception");
275 __ bind(L);
276 }
277 #endif
278
279 // pass parameters if any
280 BLOCK_COMMENT("pass parameters if any");
281 Label parameters_done;
282 __ movl(c_rarg3, parameter_size);
283 __ testl(c_rarg3, c_rarg3);
284 __ jcc(Assembler::zero, parameters_done);
285
286 Label loop;
287 __ movptr(c_rarg2, parameters); // parameter pointer
288 __ movl(c_rarg1, c_rarg3); // parameter counter is in c_rarg1
289 __ BIND(loop);
290 __ movptr(rax, Address(c_rarg2, 0));// get parameter
291 __ addptr(c_rarg2, wordSize); // advance to next parameter
292 __ decrementl(c_rarg1); // decrement counter
293 __ push(rax); // pass parameter
294 __ jcc(Assembler::notZero, loop);
295
296 // call Java function
297 __ BIND(parameters_done);
298 __ movptr(rbx, method); // get Method*
299 __ movptr(c_rarg1, entry_point); // get entry_point
300 __ mov(r13, rsp); // set sender sp
301 BLOCK_COMMENT("call Java function");
302 __ call(c_rarg1);
303
304 BLOCK_COMMENT("call_stub_return_address:");
305 return_address = __ pc();
306
307 // store result depending on type (everything that is not
308 // T_OBJECT, T_LONG, T_FLOAT or T_DOUBLE is treated as T_INT)
309 __ movptr(r13, result);
310 Label is_long, is_float, is_double, check_prim, exit;
311 __ movl(rbx, result_type);
312 __ cmpl(rbx, T_OBJECT);
313 __ jcc(Assembler::equal, check_prim);
314 __ cmpl(rbx, T_LONG);
315 __ jcc(Assembler::equal, is_long);
316 __ cmpl(rbx, T_FLOAT);
317 __ jcc(Assembler::equal, is_float);
318 __ cmpl(rbx, T_DOUBLE);
319 __ jcc(Assembler::equal, is_double);
320 #ifdef ASSERT
321 // make sure the type is INT
322 {
323 Label L;
324 __ cmpl(rbx, T_INT);
325 __ jcc(Assembler::equal, L);
326 __ stop("StubRoutines::call_stub: unexpected result type");
327 __ bind(L);
328 }
329 #endif
330
331 // handle T_INT case
332 __ movl(Address(r13, 0), rax);
333
334 __ BIND(exit);
335
336 // pop parameters
337 __ lea(rsp, rsp_after_call);
338
339 #ifdef ASSERT
340 // verify that threads correspond
341 {
342 Label L1, L2, L3;
343 __ cmpptr(r15_thread, thread);
344 __ jcc(Assembler::equal, L1);
345 __ stop("StubRoutines::call_stub: r15_thread is corrupted");
346 __ bind(L1);
347 __ get_thread(rbx);
348 __ cmpptr(r15_thread, thread);
349 __ jcc(Assembler::equal, L2);
350 __ stop("StubRoutines::call_stub: r15_thread is modified by call");
351 __ bind(L2);
352 __ cmpptr(r15_thread, rbx);
353 __ jcc(Assembler::equal, L3);
354 __ stop("StubRoutines::call_stub: threads must correspond");
355 __ bind(L3);
356 }
357 #endif
358
359 __ pop_cont_fastpath();
360
361 // restore regs belonging to calling function
362 #ifdef _WIN64
363 // emit the restores for xmm regs
364 for (int i = xmm_save_first; i <= last_reg; i++) {
365 __ movdqu(as_XMMRegister(i), xmm_save(i));
366 }
367 #endif
368 __ movptr(r15, r15_save);
369 __ movptr(r14, r14_save);
370 __ movptr(r13, r13_save);
371 __ movptr(r12, r12_save);
372 __ movptr(rbx, rbx_save);
373
374 #ifdef _WIN64
375 __ movptr(rdi, rdi_save);
376 __ movptr(rsi, rsi_save);
377 #else
378 __ ldmxcsr(mxcsr_save);
379 #endif
380
381 // restore rsp
382 __ addptr(rsp, -rsp_after_call_off * wordSize);
383
384 // return
385 __ vzeroupper();
386 __ pop(rbp);
387 __ ret(0);
388
389 // handle return types different from T_INT
390 __ BIND(check_prim);
391 if (InlineTypeReturnedAsFields) {
392 // Check for scalarized return value
393 __ testptr(rax, 1);
394 __ jcc(Assembler::zero, is_long);
395 // Load pack handler address
396 __ andptr(rax, -2);
397 __ movptr(rax, Address(rax, InstanceKlass::adr_inlineklass_fixed_block_offset()));
398 __ movptr(rbx, Address(rax, InlineKlass::pack_handler_jobject_offset()));
399 // Call pack handler to initialize the buffer
400 __ call(rbx);
401 __ jmp(exit);
402 }
403 __ BIND(is_long);
404 __ movq(Address(r13, 0), rax);
405 __ jmp(exit);
406
407 __ BIND(is_float);
408 __ movflt(Address(r13, 0), xmm0);
409 __ jmp(exit);
410
411 __ BIND(is_double);
412 __ movdbl(Address(r13, 0), xmm0);
413 __ jmp(exit);
414
415 return start;
416 }
417
418 // Return point for a Java call if there's an exception thrown in
419 // Java code. The exception is caught and transformed into a
420 // pending exception stored in JavaThread that can be tested from
421 // within the VM.
422 //
423 // Note: Usually the parameters are removed by the callee. In case
424 // of an exception crossing an activation frame boundary, that is
425 // not the case if the callee is compiled code => need to setup the
426 // rsp.
427 //
428 // rax: exception oop
429
430 address StubGenerator::generate_catch_exception() {
431 StubCodeMark mark(this, "StubRoutines", "catch_exception");
432 address start = __ pc();
433
434 // same as in generate_call_stub():
435 const Address rsp_after_call(rbp, rsp_after_call_off * wordSize);
436 const Address thread (rbp, thread_off * wordSize);
437
438 #ifdef ASSERT
439 // verify that threads correspond
440 {
441 Label L1, L2, L3;
442 __ cmpptr(r15_thread, thread);
443 __ jcc(Assembler::equal, L1);
444 __ stop("StubRoutines::catch_exception: r15_thread is corrupted");
445 __ bind(L1);
446 __ get_thread(rbx);
447 __ cmpptr(r15_thread, thread);
448 __ jcc(Assembler::equal, L2);
449 __ stop("StubRoutines::catch_exception: r15_thread is modified by call");
450 __ bind(L2);
451 __ cmpptr(r15_thread, rbx);
452 __ jcc(Assembler::equal, L3);
453 __ stop("StubRoutines::catch_exception: threads must correspond");
454 __ bind(L3);
455 }
456 #endif
457
458 // set pending exception
459 __ verify_oop(rax);
460
461 __ movptr(Address(r15_thread, Thread::pending_exception_offset()), rax);
462 __ lea(rscratch1, ExternalAddress((address)__FILE__));
463 __ movptr(Address(r15_thread, Thread::exception_file_offset()), rscratch1);
464 __ movl(Address(r15_thread, Thread::exception_line_offset()), (int) __LINE__);
465
466 // complete return to VM
467 assert(StubRoutines::_call_stub_return_address != nullptr,
468 "_call_stub_return_address must have been generated before");
469 __ jump(RuntimeAddress(StubRoutines::_call_stub_return_address));
470
471 return start;
472 }
473
474 // Continuation point for runtime calls returning with a pending
475 // exception. The pending exception check happened in the runtime
476 // or native call stub. The pending exception in Thread is
477 // converted into a Java-level exception.
478 //
479 // Contract with Java-level exception handlers:
480 // rax: exception
481 // rdx: throwing pc
482 //
483 // NOTE: At entry of this stub, exception-pc must be on stack !!
484
485 address StubGenerator::generate_forward_exception() {
486 StubCodeMark mark(this, "StubRoutines", "forward exception");
487 address start = __ pc();
488
489 // Upon entry, the sp points to the return address returning into
490 // Java (interpreted or compiled) code; i.e., the return address
491 // becomes the throwing pc.
492 //
493 // Arguments pushed before the runtime call are still on the stack
494 // but the exception handler will reset the stack pointer ->
495 // ignore them. A potential result in registers can be ignored as
496 // well.
497
498 #ifdef ASSERT
499 // make sure this code is only executed if there is a pending exception
500 {
501 Label L;
502 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), NULL_WORD);
503 __ jcc(Assembler::notEqual, L);
504 __ stop("StubRoutines::forward exception: no pending exception (1)");
505 __ bind(L);
506 }
507 #endif
508
509 // compute exception handler into rbx
510 __ movptr(c_rarg0, Address(rsp, 0));
511 BLOCK_COMMENT("call exception_handler_for_return_address");
512 __ call_VM_leaf(CAST_FROM_FN_PTR(address,
513 SharedRuntime::exception_handler_for_return_address),
514 r15_thread, c_rarg0);
515 __ mov(rbx, rax);
516
517 // setup rax & rdx, remove return address & clear pending exception
518 __ pop(rdx);
519 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset()));
520 __ movptr(Address(r15_thread, Thread::pending_exception_offset()), NULL_WORD);
521
522 #ifdef ASSERT
523 // make sure exception is set
524 {
525 Label L;
526 __ testptr(rax, rax);
527 __ jcc(Assembler::notEqual, L);
528 __ stop("StubRoutines::forward exception: no pending exception (2)");
529 __ bind(L);
530 }
531 #endif
532
533 // continue at exception handler (return address removed)
534 // rax: exception
535 // rbx: exception handler
536 // rdx: throwing pc
537 __ verify_oop(rax);
538 __ jmp(rbx);
539
540 return start;
541 }
542
543 // Support for intptr_t OrderAccess::fence()
544 //
545 // Arguments :
546 //
547 // Result:
548 address StubGenerator::generate_orderaccess_fence() {
549 StubCodeMark mark(this, "StubRoutines", "orderaccess_fence");
550 address start = __ pc();
551
552 __ membar(Assembler::StoreLoad);
553 __ ret(0);
554
555 return start;
556 }
557
558
559 // Support for intptr_t get_previous_sp()
560 //
561 // This routine is used to find the previous stack pointer for the
562 // caller.
563 address StubGenerator::generate_get_previous_sp() {
564 StubCodeMark mark(this, "StubRoutines", "get_previous_sp");
565 address start = __ pc();
566
567 __ movptr(rax, rsp);
568 __ addptr(rax, 8); // return address is at the top of the stack.
569 __ ret(0);
570
571 return start;
572 }
573
574 //----------------------------------------------------------------------------------------------------
575 // Support for void verify_mxcsr()
576 //
577 // This routine is used with -Xcheck:jni to verify that native
578 // JNI code does not return to Java code without restoring the
579 // MXCSR register to our expected state.
580
581 address StubGenerator::generate_verify_mxcsr() {
582 StubCodeMark mark(this, "StubRoutines", "verify_mxcsr");
583 address start = __ pc();
584
585 const Address mxcsr_save(rsp, 0);
586
587 if (CheckJNICalls) {
588 Label ok_ret;
589 ExternalAddress mxcsr_std(StubRoutines::x86::addr_mxcsr_std());
590 __ push(rax);
591 __ subptr(rsp, wordSize); // allocate a temp location
592 __ stmxcsr(mxcsr_save);
593 __ movl(rax, mxcsr_save);
594 __ andl(rax, 0xFFC0); // Mask out any pending exceptions (only check control and mask bits)
595 __ cmp32(rax, mxcsr_std, rscratch1);
596 __ jcc(Assembler::equal, ok_ret);
597
598 __ warn("MXCSR changed by native JNI code, use -XX:+RestoreMXCSROnJNICall");
599
600 __ ldmxcsr(mxcsr_std, rscratch1);
601
602 __ bind(ok_ret);
603 __ addptr(rsp, wordSize);
604 __ pop(rax);
605 }
606
607 __ ret(0);
608
609 return start;
610 }
611
612 address StubGenerator::generate_f2i_fixup() {
613 StubCodeMark mark(this, "StubRoutines", "f2i_fixup");
614 Address inout(rsp, 5 * wordSize); // return address + 4 saves
615
616 address start = __ pc();
617
618 Label L;
619
620 __ push(rax);
621 __ push(c_rarg3);
622 __ push(c_rarg2);
623 __ push(c_rarg1);
624
625 __ movl(rax, 0x7f800000);
626 __ xorl(c_rarg3, c_rarg3);
627 __ movl(c_rarg2, inout);
628 __ movl(c_rarg1, c_rarg2);
629 __ andl(c_rarg1, 0x7fffffff);
630 __ cmpl(rax, c_rarg1); // NaN? -> 0
631 __ jcc(Assembler::negative, L);
632 __ testl(c_rarg2, c_rarg2); // signed ? min_jint : max_jint
633 __ movl(c_rarg3, 0x80000000);
634 __ movl(rax, 0x7fffffff);
635 __ cmovl(Assembler::positive, c_rarg3, rax);
636
637 __ bind(L);
638 __ movptr(inout, c_rarg3);
639
640 __ pop(c_rarg1);
641 __ pop(c_rarg2);
642 __ pop(c_rarg3);
643 __ pop(rax);
644
645 __ ret(0);
646
647 return start;
648 }
649
650 address StubGenerator::generate_f2l_fixup() {
651 StubCodeMark mark(this, "StubRoutines", "f2l_fixup");
652 Address inout(rsp, 5 * wordSize); // return address + 4 saves
653 address start = __ pc();
654
655 Label L;
656
657 __ push(rax);
658 __ push(c_rarg3);
659 __ push(c_rarg2);
660 __ push(c_rarg1);
661
662 __ movl(rax, 0x7f800000);
663 __ xorl(c_rarg3, c_rarg3);
664 __ movl(c_rarg2, inout);
665 __ movl(c_rarg1, c_rarg2);
666 __ andl(c_rarg1, 0x7fffffff);
667 __ cmpl(rax, c_rarg1); // NaN? -> 0
668 __ jcc(Assembler::negative, L);
669 __ testl(c_rarg2, c_rarg2); // signed ? min_jlong : max_jlong
670 __ mov64(c_rarg3, 0x8000000000000000);
671 __ mov64(rax, 0x7fffffffffffffff);
672 __ cmov(Assembler::positive, c_rarg3, rax);
673
674 __ bind(L);
675 __ movptr(inout, c_rarg3);
676
677 __ pop(c_rarg1);
678 __ pop(c_rarg2);
679 __ pop(c_rarg3);
680 __ pop(rax);
681
682 __ ret(0);
683
684 return start;
685 }
686
687 address StubGenerator::generate_d2i_fixup() {
688 StubCodeMark mark(this, "StubRoutines", "d2i_fixup");
689 Address inout(rsp, 6 * wordSize); // return address + 5 saves
690
691 address start = __ pc();
692
693 Label L;
694
695 __ push(rax);
696 __ push(c_rarg3);
697 __ push(c_rarg2);
698 __ push(c_rarg1);
699 __ push(c_rarg0);
700
701 __ movl(rax, 0x7ff00000);
702 __ movq(c_rarg2, inout);
703 __ movl(c_rarg3, c_rarg2);
704 __ mov(c_rarg1, c_rarg2);
705 __ mov(c_rarg0, c_rarg2);
706 __ negl(c_rarg3);
707 __ shrptr(c_rarg1, 0x20);
708 __ orl(c_rarg3, c_rarg2);
709 __ andl(c_rarg1, 0x7fffffff);
710 __ xorl(c_rarg2, c_rarg2);
711 __ shrl(c_rarg3, 0x1f);
712 __ orl(c_rarg1, c_rarg3);
713 __ cmpl(rax, c_rarg1);
714 __ jcc(Assembler::negative, L); // NaN -> 0
715 __ testptr(c_rarg0, c_rarg0); // signed ? min_jint : max_jint
716 __ movl(c_rarg2, 0x80000000);
717 __ movl(rax, 0x7fffffff);
718 __ cmov(Assembler::positive, c_rarg2, rax);
719
720 __ bind(L);
721 __ movptr(inout, c_rarg2);
722
723 __ pop(c_rarg0);
724 __ pop(c_rarg1);
725 __ pop(c_rarg2);
726 __ pop(c_rarg3);
727 __ pop(rax);
728
729 __ ret(0);
730
731 return start;
732 }
733
734 address StubGenerator::generate_d2l_fixup() {
735 StubCodeMark mark(this, "StubRoutines", "d2l_fixup");
736 Address inout(rsp, 6 * wordSize); // return address + 5 saves
737
738 address start = __ pc();
739
740 Label L;
741
742 __ push(rax);
743 __ push(c_rarg3);
744 __ push(c_rarg2);
745 __ push(c_rarg1);
746 __ push(c_rarg0);
747
748 __ movl(rax, 0x7ff00000);
749 __ movq(c_rarg2, inout);
750 __ movl(c_rarg3, c_rarg2);
751 __ mov(c_rarg1, c_rarg2);
752 __ mov(c_rarg0, c_rarg2);
753 __ negl(c_rarg3);
754 __ shrptr(c_rarg1, 0x20);
755 __ orl(c_rarg3, c_rarg2);
756 __ andl(c_rarg1, 0x7fffffff);
757 __ xorl(c_rarg2, c_rarg2);
758 __ shrl(c_rarg3, 0x1f);
759 __ orl(c_rarg1, c_rarg3);
760 __ cmpl(rax, c_rarg1);
761 __ jcc(Assembler::negative, L); // NaN -> 0
762 __ testq(c_rarg0, c_rarg0); // signed ? min_jlong : max_jlong
763 __ mov64(c_rarg2, 0x8000000000000000);
764 __ mov64(rax, 0x7fffffffffffffff);
765 __ cmovq(Assembler::positive, c_rarg2, rax);
766
767 __ bind(L);
768 __ movq(inout, c_rarg2);
769
770 __ pop(c_rarg0);
771 __ pop(c_rarg1);
772 __ pop(c_rarg2);
773 __ pop(c_rarg3);
774 __ pop(rax);
775
776 __ ret(0);
777
778 return start;
779 }
780
781 address StubGenerator::generate_count_leading_zeros_lut(const char *stub_name) {
782 __ align64();
783 StubCodeMark mark(this, "StubRoutines", stub_name);
784 address start = __ pc();
785
786 __ emit_data64(0x0101010102020304, relocInfo::none);
787 __ emit_data64(0x0000000000000000, relocInfo::none);
788 __ emit_data64(0x0101010102020304, relocInfo::none);
789 __ emit_data64(0x0000000000000000, relocInfo::none);
790 __ emit_data64(0x0101010102020304, relocInfo::none);
791 __ emit_data64(0x0000000000000000, relocInfo::none);
792 __ emit_data64(0x0101010102020304, relocInfo::none);
793 __ emit_data64(0x0000000000000000, relocInfo::none);
794
795 return start;
796 }
797
798 address StubGenerator::generate_popcount_avx_lut(const char *stub_name) {
799 __ align64();
800 StubCodeMark mark(this, "StubRoutines", stub_name);
801 address start = __ pc();
802
803 __ emit_data64(0x0302020102010100, relocInfo::none);
804 __ emit_data64(0x0403030203020201, relocInfo::none);
805 __ emit_data64(0x0302020102010100, relocInfo::none);
806 __ emit_data64(0x0403030203020201, relocInfo::none);
807 __ emit_data64(0x0302020102010100, relocInfo::none);
808 __ emit_data64(0x0403030203020201, relocInfo::none);
809 __ emit_data64(0x0302020102010100, relocInfo::none);
810 __ emit_data64(0x0403030203020201, relocInfo::none);
811
812 return start;
813 }
814
815 address StubGenerator::generate_iota_indices(const char *stub_name) {
816 __ align(CodeEntryAlignment);
817 StubCodeMark mark(this, "StubRoutines", stub_name);
818 address start = __ pc();
819 // B
820 __ emit_data64(0x0706050403020100, relocInfo::none);
821 __ emit_data64(0x0F0E0D0C0B0A0908, relocInfo::none);
822 __ emit_data64(0x1716151413121110, relocInfo::none);
823 __ emit_data64(0x1F1E1D1C1B1A1918, relocInfo::none);
824 __ emit_data64(0x2726252423222120, relocInfo::none);
825 __ emit_data64(0x2F2E2D2C2B2A2928, relocInfo::none);
826 __ emit_data64(0x3736353433323130, relocInfo::none);
827 __ emit_data64(0x3F3E3D3C3B3A3938, relocInfo::none);
828 // W
829 __ emit_data64(0x0003000200010000, relocInfo::none);
830 __ emit_data64(0x0007000600050004, relocInfo::none);
831 __ emit_data64(0x000B000A00090008, relocInfo::none);
832 __ emit_data64(0x000F000E000D000C, relocInfo::none);
833 __ emit_data64(0x0013001200110010, relocInfo::none);
834 __ emit_data64(0x0017001600150014, relocInfo::none);
835 __ emit_data64(0x001B001A00190018, relocInfo::none);
836 __ emit_data64(0x001F001E001D001C, relocInfo::none);
837 // D
838 __ emit_data64(0x0000000100000000, relocInfo::none);
839 __ emit_data64(0x0000000300000002, relocInfo::none);
840 __ emit_data64(0x0000000500000004, relocInfo::none);
841 __ emit_data64(0x0000000700000006, relocInfo::none);
842 __ emit_data64(0x0000000900000008, relocInfo::none);
843 __ emit_data64(0x0000000B0000000A, relocInfo::none);
844 __ emit_data64(0x0000000D0000000C, relocInfo::none);
845 __ emit_data64(0x0000000F0000000E, relocInfo::none);
846 // Q
847 __ emit_data64(0x0000000000000000, relocInfo::none);
848 __ emit_data64(0x0000000000000001, relocInfo::none);
849 __ emit_data64(0x0000000000000002, relocInfo::none);
850 __ emit_data64(0x0000000000000003, relocInfo::none);
851 __ emit_data64(0x0000000000000004, relocInfo::none);
852 __ emit_data64(0x0000000000000005, relocInfo::none);
853 __ emit_data64(0x0000000000000006, relocInfo::none);
854 __ emit_data64(0x0000000000000007, relocInfo::none);
855 // D - FP
856 __ emit_data64(0x3F80000000000000, relocInfo::none); // 0.0f, 1.0f
857 __ emit_data64(0x4040000040000000, relocInfo::none); // 2.0f, 3.0f
858 __ emit_data64(0x40A0000040800000, relocInfo::none); // 4.0f, 5.0f
859 __ emit_data64(0x40E0000040C00000, relocInfo::none); // 6.0f, 7.0f
860 __ emit_data64(0x4110000041000000, relocInfo::none); // 8.0f, 9.0f
861 __ emit_data64(0x4130000041200000, relocInfo::none); // 10.0f, 11.0f
862 __ emit_data64(0x4150000041400000, relocInfo::none); // 12.0f, 13.0f
863 __ emit_data64(0x4170000041600000, relocInfo::none); // 14.0f, 15.0f
864 // Q - FP
865 __ emit_data64(0x0000000000000000, relocInfo::none); // 0.0d
866 __ emit_data64(0x3FF0000000000000, relocInfo::none); // 1.0d
867 __ emit_data64(0x4000000000000000, relocInfo::none); // 2.0d
868 __ emit_data64(0x4008000000000000, relocInfo::none); // 3.0d
869 __ emit_data64(0x4010000000000000, relocInfo::none); // 4.0d
870 __ emit_data64(0x4014000000000000, relocInfo::none); // 5.0d
871 __ emit_data64(0x4018000000000000, relocInfo::none); // 6.0d
872 __ emit_data64(0x401c000000000000, relocInfo::none); // 7.0d
873 return start;
874 }
875
876 address StubGenerator::generate_vector_reverse_bit_lut(const char *stub_name) {
877 __ align(CodeEntryAlignment);
878 StubCodeMark mark(this, "StubRoutines", stub_name);
879 address start = __ pc();
880
881 __ emit_data64(0x0E060A020C040800, relocInfo::none);
882 __ emit_data64(0x0F070B030D050901, relocInfo::none);
883 __ emit_data64(0x0E060A020C040800, relocInfo::none);
884 __ emit_data64(0x0F070B030D050901, relocInfo::none);
885 __ emit_data64(0x0E060A020C040800, relocInfo::none);
886 __ emit_data64(0x0F070B030D050901, relocInfo::none);
887 __ emit_data64(0x0E060A020C040800, relocInfo::none);
888 __ emit_data64(0x0F070B030D050901, relocInfo::none);
889
890 return start;
891 }
892
893 address StubGenerator::generate_vector_reverse_byte_perm_mask_long(const char *stub_name) {
894 __ align(CodeEntryAlignment);
895 StubCodeMark mark(this, "StubRoutines", stub_name);
896 address start = __ pc();
897
898 __ emit_data64(0x0001020304050607, relocInfo::none);
899 __ emit_data64(0x08090A0B0C0D0E0F, relocInfo::none);
900 __ emit_data64(0x0001020304050607, relocInfo::none);
901 __ emit_data64(0x08090A0B0C0D0E0F, relocInfo::none);
902 __ emit_data64(0x0001020304050607, relocInfo::none);
903 __ emit_data64(0x08090A0B0C0D0E0F, relocInfo::none);
904 __ emit_data64(0x0001020304050607, relocInfo::none);
905 __ emit_data64(0x08090A0B0C0D0E0F, relocInfo::none);
906
907 return start;
908 }
909
910 address StubGenerator::generate_vector_reverse_byte_perm_mask_int(const char *stub_name) {
911 __ align(CodeEntryAlignment);
912 StubCodeMark mark(this, "StubRoutines", stub_name);
913 address start = __ pc();
914
915 __ emit_data64(0x0405060700010203, relocInfo::none);
916 __ emit_data64(0x0C0D0E0F08090A0B, relocInfo::none);
917 __ emit_data64(0x0405060700010203, relocInfo::none);
918 __ emit_data64(0x0C0D0E0F08090A0B, relocInfo::none);
919 __ emit_data64(0x0405060700010203, relocInfo::none);
920 __ emit_data64(0x0C0D0E0F08090A0B, relocInfo::none);
921 __ emit_data64(0x0405060700010203, relocInfo::none);
922 __ emit_data64(0x0C0D0E0F08090A0B, relocInfo::none);
923
924 return start;
925 }
926
927 address StubGenerator::generate_vector_reverse_byte_perm_mask_short(const char *stub_name) {
928 __ align(CodeEntryAlignment);
929 StubCodeMark mark(this, "StubRoutines", stub_name);
930 address start = __ pc();
931
932 __ emit_data64(0x0607040502030001, relocInfo::none);
933 __ emit_data64(0x0E0F0C0D0A0B0809, relocInfo::none);
934 __ emit_data64(0x0607040502030001, relocInfo::none);
935 __ emit_data64(0x0E0F0C0D0A0B0809, relocInfo::none);
936 __ emit_data64(0x0607040502030001, relocInfo::none);
937 __ emit_data64(0x0E0F0C0D0A0B0809, relocInfo::none);
938 __ emit_data64(0x0607040502030001, relocInfo::none);
939 __ emit_data64(0x0E0F0C0D0A0B0809, relocInfo::none);
940
941 return start;
942 }
943
944 address StubGenerator::generate_vector_byte_shuffle_mask(const char *stub_name) {
945 __ align(CodeEntryAlignment);
946 StubCodeMark mark(this, "StubRoutines", stub_name);
947 address start = __ pc();
948
949 __ emit_data64(0x7070707070707070, relocInfo::none);
950 __ emit_data64(0x7070707070707070, relocInfo::none);
951 __ emit_data64(0xF0F0F0F0F0F0F0F0, relocInfo::none);
952 __ emit_data64(0xF0F0F0F0F0F0F0F0, relocInfo::none);
953
954 return start;
955 }
956
957 address StubGenerator::generate_fp_mask(const char *stub_name, int64_t mask) {
958 __ align(CodeEntryAlignment);
959 StubCodeMark mark(this, "StubRoutines", stub_name);
960 address start = __ pc();
961
962 __ emit_data64( mask, relocInfo::none );
963 __ emit_data64( mask, relocInfo::none );
964
965 return start;
966 }
967
968 address StubGenerator::generate_compress_perm_table(const char *stub_name, int32_t esize) {
969 __ align(CodeEntryAlignment);
970 StubCodeMark mark(this, "StubRoutines", stub_name);
971 address start = __ pc();
972 if (esize == 32) {
973 // Loop to generate 256 x 8 int compression permute index table. A row is
974 // accessed using 8 bit index computed using vector mask. An entry in
975 // a row holds either a valid permute index corresponding to set bit position
976 // or a -1 (default) value.
977 for (int mask = 0; mask < 256; mask++) {
978 int ctr = 0;
979 for (int j = 0; j < 8; j++) {
980 if (mask & (1 << j)) {
981 __ emit_data(j, relocInfo::none);
982 ctr++;
983 }
984 }
985 for (; ctr < 8; ctr++) {
986 __ emit_data(-1, relocInfo::none);
987 }
988 }
989 } else {
990 assert(esize == 64, "");
991 // Loop to generate 16 x 4 long compression permute index table. A row is
992 // accessed using 4 bit index computed using vector mask. An entry in
993 // a row holds either a valid permute index pair for a quadword corresponding
994 // to set bit position or a -1 (default) value.
995 for (int mask = 0; mask < 16; mask++) {
996 int ctr = 0;
997 for (int j = 0; j < 4; j++) {
998 if (mask & (1 << j)) {
999 __ emit_data(2 * j, relocInfo::none);
1000 __ emit_data(2 * j + 1, relocInfo::none);
1001 ctr++;
1002 }
1003 }
1004 for (; ctr < 4; ctr++) {
1005 __ emit_data64(-1L, relocInfo::none);
1006 }
1007 }
1008 }
1009 return start;
1010 }
1011
1012 address StubGenerator::generate_expand_perm_table(const char *stub_name, int32_t esize) {
1013 __ align(CodeEntryAlignment);
1014 StubCodeMark mark(this, "StubRoutines", stub_name);
1015 address start = __ pc();
1016 if (esize == 32) {
1017 // Loop to generate 256 x 8 int expand permute index table. A row is accessed
1018 // using 8 bit index computed using vector mask. An entry in a row holds either
1019 // a valid permute index (starting from least significant lane) placed at poisition
1020 // corresponding to set bit position or a -1 (default) value.
1021 for (int mask = 0; mask < 256; mask++) {
1022 int ctr = 0;
1023 for (int j = 0; j < 8; j++) {
1024 if (mask & (1 << j)) {
1025 __ emit_data(ctr++, relocInfo::none);
1026 } else {
1027 __ emit_data(-1, relocInfo::none);
1028 }
1029 }
1030 }
1031 } else {
1032 assert(esize == 64, "");
1033 // Loop to generate 16 x 4 long expand permute index table. A row is accessed
1034 // using 4 bit index computed using vector mask. An entry in a row holds either
1035 // a valid doubleword permute index pair representing a quadword index (starting
1036 // from least significant lane) placed at poisition corresponding to set bit
1037 // position or a -1 (default) value.
1038 for (int mask = 0; mask < 16; mask++) {
1039 int ctr = 0;
1040 for (int j = 0; j < 4; j++) {
1041 if (mask & (1 << j)) {
1042 __ emit_data(2 * ctr, relocInfo::none);
1043 __ emit_data(2 * ctr + 1, relocInfo::none);
1044 ctr++;
1045 } else {
1046 __ emit_data64(-1L, relocInfo::none);
1047 }
1048 }
1049 }
1050 }
1051 return start;
1052 }
1053
1054 address StubGenerator::generate_vector_mask(const char *stub_name, int64_t mask) {
1055 __ align(CodeEntryAlignment);
1056 StubCodeMark mark(this, "StubRoutines", stub_name);
1057 address start = __ pc();
1058
1059 __ emit_data64(mask, relocInfo::none);
1060 __ emit_data64(mask, relocInfo::none);
1061 __ emit_data64(mask, relocInfo::none);
1062 __ emit_data64(mask, relocInfo::none);
1063 __ emit_data64(mask, relocInfo::none);
1064 __ emit_data64(mask, relocInfo::none);
1065 __ emit_data64(mask, relocInfo::none);
1066 __ emit_data64(mask, relocInfo::none);
1067
1068 return start;
1069 }
1070
1071 address StubGenerator::generate_vector_byte_perm_mask(const char *stub_name) {
1072 __ align(CodeEntryAlignment);
1073 StubCodeMark mark(this, "StubRoutines", stub_name);
1074 address start = __ pc();
1075
1076 __ emit_data64(0x0000000000000001, relocInfo::none);
1077 __ emit_data64(0x0000000000000003, relocInfo::none);
1078 __ emit_data64(0x0000000000000005, relocInfo::none);
1079 __ emit_data64(0x0000000000000007, relocInfo::none);
1080 __ emit_data64(0x0000000000000000, relocInfo::none);
1081 __ emit_data64(0x0000000000000002, relocInfo::none);
1082 __ emit_data64(0x0000000000000004, relocInfo::none);
1083 __ emit_data64(0x0000000000000006, relocInfo::none);
1084
1085 return start;
1086 }
1087
1088 address StubGenerator::generate_vector_fp_mask(const char *stub_name, int64_t mask) {
1089 __ align(CodeEntryAlignment);
1090 StubCodeMark mark(this, "StubRoutines", stub_name);
1091 address start = __ pc();
1092
1093 __ emit_data64(mask, relocInfo::none);
1094 __ emit_data64(mask, relocInfo::none);
1095 __ emit_data64(mask, relocInfo::none);
1096 __ emit_data64(mask, relocInfo::none);
1097 __ emit_data64(mask, relocInfo::none);
1098 __ emit_data64(mask, relocInfo::none);
1099 __ emit_data64(mask, relocInfo::none);
1100 __ emit_data64(mask, relocInfo::none);
1101
1102 return start;
1103 }
1104
1105 address StubGenerator::generate_vector_custom_i32(const char *stub_name, Assembler::AvxVectorLen len,
1106 int32_t val0, int32_t val1, int32_t val2, int32_t val3,
1107 int32_t val4, int32_t val5, int32_t val6, int32_t val7,
1108 int32_t val8, int32_t val9, int32_t val10, int32_t val11,
1109 int32_t val12, int32_t val13, int32_t val14, int32_t val15) {
1110 __ align(CodeEntryAlignment);
1111 StubCodeMark mark(this, "StubRoutines", stub_name);
1112 address start = __ pc();
1113
1114 assert(len != Assembler::AVX_NoVec, "vector len must be specified");
1115 __ emit_data(val0, relocInfo::none, 0);
1116 __ emit_data(val1, relocInfo::none, 0);
1117 __ emit_data(val2, relocInfo::none, 0);
1118 __ emit_data(val3, relocInfo::none, 0);
1119 if (len >= Assembler::AVX_256bit) {
1120 __ emit_data(val4, relocInfo::none, 0);
1121 __ emit_data(val5, relocInfo::none, 0);
1122 __ emit_data(val6, relocInfo::none, 0);
1123 __ emit_data(val7, relocInfo::none, 0);
1124 if (len >= Assembler::AVX_512bit) {
1125 __ emit_data(val8, relocInfo::none, 0);
1126 __ emit_data(val9, relocInfo::none, 0);
1127 __ emit_data(val10, relocInfo::none, 0);
1128 __ emit_data(val11, relocInfo::none, 0);
1129 __ emit_data(val12, relocInfo::none, 0);
1130 __ emit_data(val13, relocInfo::none, 0);
1131 __ emit_data(val14, relocInfo::none, 0);
1132 __ emit_data(val15, relocInfo::none, 0);
1133 }
1134 }
1135 return start;
1136 }
1137
1138 // Non-destructive plausibility checks for oops
1139 //
1140 // Arguments:
1141 // all args on stack!
1142 //
1143 // Stack after saving c_rarg3:
1144 // [tos + 0]: saved c_rarg3
1145 // [tos + 1]: saved c_rarg2
1146 // [tos + 2]: saved r12 (several TemplateTable methods use it)
1147 // [tos + 3]: saved flags
1148 // [tos + 4]: return address
1149 // * [tos + 5]: error message (char*)
1150 // * [tos + 6]: object to verify (oop)
1151 // * [tos + 7]: saved rax - saved by caller and bashed
1152 // * [tos + 8]: saved r10 (rscratch1) - saved by caller
1153 // * = popped on exit
1154 address StubGenerator::generate_verify_oop() {
1155 StubCodeMark mark(this, "StubRoutines", "verify_oop");
1156 address start = __ pc();
1157
1158 Label exit, error;
1159
1160 __ pushf();
1161 __ incrementl(ExternalAddress((address) StubRoutines::verify_oop_count_addr()), rscratch1);
1162
1163 __ push(r12);
1164
1165 // save c_rarg2 and c_rarg3
1166 __ push(c_rarg2);
1167 __ push(c_rarg3);
1168
1169 enum {
1170 // After previous pushes.
1171 oop_to_verify = 6 * wordSize,
1172 saved_rax = 7 * wordSize,
1173 saved_r10 = 8 * wordSize,
1174
1175 // Before the call to MacroAssembler::debug(), see below.
1176 return_addr = 16 * wordSize,
1177 error_msg = 17 * wordSize
1178 };
1179
1180 // get object
1181 __ movptr(rax, Address(rsp, oop_to_verify));
1182
1183 // make sure object is 'reasonable'
1184 __ testptr(rax, rax);
1185 __ jcc(Assembler::zero, exit); // if obj is null it is OK
1186
1187 BarrierSetAssembler* bs_asm = BarrierSet::barrier_set()->barrier_set_assembler();
1188 bs_asm->check_oop(_masm, rax, c_rarg2, c_rarg3, error);
1189
1190 // return if everything seems ok
1191 __ bind(exit);
1192 __ movptr(rax, Address(rsp, saved_rax)); // get saved rax back
1193 __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back
1194 __ pop(c_rarg3); // restore c_rarg3
1195 __ pop(c_rarg2); // restore c_rarg2
1196 __ pop(r12); // restore r12
1197 __ popf(); // restore flags
1198 __ ret(4 * wordSize); // pop caller saved stuff
1199
1200 // handle errors
1201 __ bind(error);
1202 __ movptr(rax, Address(rsp, saved_rax)); // get saved rax back
1203 __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back
1204 __ pop(c_rarg3); // get saved c_rarg3 back
1205 __ pop(c_rarg2); // get saved c_rarg2 back
1206 __ pop(r12); // get saved r12 back
1207 __ popf(); // get saved flags off stack --
1208 // will be ignored
1209
1210 __ pusha(); // push registers
1211 // (rip is already
1212 // already pushed)
1213 // debug(char* msg, int64_t pc, int64_t regs[])
1214 // We've popped the registers we'd saved (c_rarg3, c_rarg2 and flags), and
1215 // pushed all the registers, so now the stack looks like:
1216 // [tos + 0] 16 saved registers
1217 // [tos + 16] return address
1218 // * [tos + 17] error message (char*)
1219 // * [tos + 18] object to verify (oop)
1220 // * [tos + 19] saved rax - saved by caller and bashed
1221 // * [tos + 20] saved r10 (rscratch1) - saved by caller
1222 // * = popped on exit
1223
1224 __ movptr(c_rarg0, Address(rsp, error_msg)); // pass address of error message
1225 __ movptr(c_rarg1, Address(rsp, return_addr)); // pass return address
1226 __ movq(c_rarg2, rsp); // pass address of regs on stack
1227 __ mov(r12, rsp); // remember rsp
1228 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows
1229 __ andptr(rsp, -16); // align stack as required by ABI
1230 BLOCK_COMMENT("call MacroAssembler::debug");
1231 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64)));
1232 __ hlt();
1233
1234 return start;
1235 }
1236
1237
1238 // Shuffle first three arg regs on Windows into Linux/Solaris locations.
1239 //
1240 // Outputs:
1241 // rdi - rcx
1242 // rsi - rdx
1243 // rdx - r8
1244 // rcx - r9
1245 //
1246 // Registers r9 and r10 are used to save rdi and rsi on Windows, which latter
1247 // are non-volatile. r9 and r10 should not be used by the caller.
1248 //
1249 void StubGenerator::setup_arg_regs(int nargs) {
1250 const Register saved_rdi = r9;
1251 const Register saved_rsi = r10;
1252 assert(nargs == 3 || nargs == 4, "else fix");
1253 #ifdef _WIN64
1254 assert(c_rarg0 == rcx && c_rarg1 == rdx && c_rarg2 == r8 && c_rarg3 == r9,
1255 "unexpected argument registers");
1256 if (nargs == 4) {
1257 __ mov(rax, r9); // r9 is also saved_rdi
1258 }
1259 __ movptr(saved_rdi, rdi);
1260 __ movptr(saved_rsi, rsi);
1261 __ mov(rdi, rcx); // c_rarg0
1262 __ mov(rsi, rdx); // c_rarg1
1263 __ mov(rdx, r8); // c_rarg2
1264 if (nargs == 4) {
1265 __ mov(rcx, rax); // c_rarg3 (via rax)
1266 }
1267 #else
1268 assert(c_rarg0 == rdi && c_rarg1 == rsi && c_rarg2 == rdx && c_rarg3 == rcx,
1269 "unexpected argument registers");
1270 #endif
1271 DEBUG_ONLY(_regs_in_thread = false;)
1272 }
1273
1274
1275 void StubGenerator::restore_arg_regs() {
1276 assert(!_regs_in_thread, "wrong call to restore_arg_regs");
1277 const Register saved_rdi = r9;
1278 const Register saved_rsi = r10;
1279 #ifdef _WIN64
1280 __ movptr(rdi, saved_rdi);
1281 __ movptr(rsi, saved_rsi);
1282 #endif
1283 }
1284
1285
1286 // This is used in places where r10 is a scratch register, and can
1287 // be adapted if r9 is needed also.
1288 void StubGenerator::setup_arg_regs_using_thread(int nargs) {
1289 const Register saved_r15 = r9;
1290 assert(nargs == 3 || nargs == 4, "else fix");
1291 #ifdef _WIN64
1292 if (nargs == 4) {
1293 __ mov(rax, r9); // r9 is also saved_r15
1294 }
1295 __ mov(saved_r15, r15); // r15 is callee saved and needs to be restored
1296 __ get_thread(r15_thread);
1297 assert(c_rarg0 == rcx && c_rarg1 == rdx && c_rarg2 == r8 && c_rarg3 == r9,
1298 "unexpected argument registers");
1299 __ movptr(Address(r15_thread, in_bytes(JavaThread::windows_saved_rdi_offset())), rdi);
1300 __ movptr(Address(r15_thread, in_bytes(JavaThread::windows_saved_rsi_offset())), rsi);
1301
1302 __ mov(rdi, rcx); // c_rarg0
1303 __ mov(rsi, rdx); // c_rarg1
1304 __ mov(rdx, r8); // c_rarg2
1305 if (nargs == 4) {
1306 __ mov(rcx, rax); // c_rarg3 (via rax)
1307 }
1308 #else
1309 assert(c_rarg0 == rdi && c_rarg1 == rsi && c_rarg2 == rdx && c_rarg3 == rcx,
1310 "unexpected argument registers");
1311 #endif
1312 DEBUG_ONLY(_regs_in_thread = true;)
1313 }
1314
1315
1316 void StubGenerator::restore_arg_regs_using_thread() {
1317 assert(_regs_in_thread, "wrong call to restore_arg_regs");
1318 const Register saved_r15 = r9;
1319 #ifdef _WIN64
1320 __ get_thread(r15_thread);
1321 __ movptr(rsi, Address(r15_thread, in_bytes(JavaThread::windows_saved_rsi_offset())));
1322 __ movptr(rdi, Address(r15_thread, in_bytes(JavaThread::windows_saved_rdi_offset())));
1323 __ mov(r15, saved_r15); // r15 is callee saved and needs to be restored
1324 #endif
1325 }
1326
1327
1328 void StubGenerator::setup_argument_regs(BasicType type) {
1329 if (type == T_BYTE || type == T_SHORT) {
1330 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1331 // r9 and r10 may be used to save non-volatile registers
1332 } else {
1333 setup_arg_regs_using_thread(); // from => rdi, to => rsi, count => rdx
1334 // r9 is used to save r15_thread
1335 }
1336 }
1337
1338
1339 void StubGenerator::restore_argument_regs(BasicType type) {
1340 if (type == T_BYTE || type == T_SHORT) {
1341 restore_arg_regs();
1342 } else {
1343 restore_arg_regs_using_thread();
1344 }
1345 }
1346
1347 address StubGenerator::generate_data_cache_writeback() {
1348 const Register src = c_rarg0; // source address
1349
1350 __ align(CodeEntryAlignment);
1351
1352 StubCodeMark mark(this, "StubRoutines", "_data_cache_writeback");
1353
1354 address start = __ pc();
1355
1356 __ enter();
1357 __ cache_wb(Address(src, 0));
1358 __ leave();
1359 __ ret(0);
1360
1361 return start;
1362 }
1363
1364 address StubGenerator::generate_data_cache_writeback_sync() {
1365 const Register is_pre = c_rarg0; // pre or post sync
1366
1367 __ align(CodeEntryAlignment);
1368
1369 StubCodeMark mark(this, "StubRoutines", "_data_cache_writeback_sync");
1370
1371 // pre wbsync is a no-op
1372 // post wbsync translates to an sfence
1373
1374 Label skip;
1375 address start = __ pc();
1376
1377 __ enter();
1378 __ cmpl(is_pre, 0);
1379 __ jcc(Assembler::notEqual, skip);
1380 __ cache_wbsync(false);
1381 __ bind(skip);
1382 __ leave();
1383 __ ret(0);
1384
1385 return start;
1386 }
1387
1388 // ofs and limit are use for multi-block byte array.
1389 // int com.sun.security.provider.MD5.implCompress(byte[] b, int ofs)
1390 address StubGenerator::generate_md5_implCompress(bool multi_block, const char *name) {
1391 __ align(CodeEntryAlignment);
1392 StubCodeMark mark(this, "StubRoutines", name);
1393 address start = __ pc();
1394
1395 const Register buf_param = r15;
1396 const Address state_param(rsp, 0 * wordSize);
1397 const Address ofs_param (rsp, 1 * wordSize );
1398 const Address limit_param(rsp, 1 * wordSize + 4);
1399
1400 __ enter();
1401 __ push(rbx);
1402 __ push(rdi);
1403 __ push(rsi);
1404 __ push(r15);
1405 __ subptr(rsp, 2 * wordSize);
1406
1407 __ movptr(buf_param, c_rarg0);
1408 __ movptr(state_param, c_rarg1);
1409 if (multi_block) {
1410 __ movl(ofs_param, c_rarg2);
1411 __ movl(limit_param, c_rarg3);
1412 }
1413 __ fast_md5(buf_param, state_param, ofs_param, limit_param, multi_block);
1414
1415 __ addptr(rsp, 2 * wordSize);
1416 __ pop(r15);
1417 __ pop(rsi);
1418 __ pop(rdi);
1419 __ pop(rbx);
1420 __ leave();
1421 __ ret(0);
1422
1423 return start;
1424 }
1425
1426 address StubGenerator::generate_upper_word_mask() {
1427 __ align64();
1428 StubCodeMark mark(this, "StubRoutines", "upper_word_mask");
1429 address start = __ pc();
1430
1431 __ emit_data64(0x0000000000000000, relocInfo::none);
1432 __ emit_data64(0xFFFFFFFF00000000, relocInfo::none);
1433
1434 return start;
1435 }
1436
1437 address StubGenerator::generate_shuffle_byte_flip_mask() {
1438 __ align64();
1439 StubCodeMark mark(this, "StubRoutines", "shuffle_byte_flip_mask");
1440 address start = __ pc();
1441
1442 __ emit_data64(0x08090a0b0c0d0e0f, relocInfo::none);
1443 __ emit_data64(0x0001020304050607, relocInfo::none);
1444
1445 return start;
1446 }
1447
1448 // ofs and limit are use for multi-block byte array.
1449 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
1450 address StubGenerator::generate_sha1_implCompress(bool multi_block, const char *name) {
1451 __ align(CodeEntryAlignment);
1452 StubCodeMark mark(this, "StubRoutines", name);
1453 address start = __ pc();
1454
1455 Register buf = c_rarg0;
1456 Register state = c_rarg1;
1457 Register ofs = c_rarg2;
1458 Register limit = c_rarg3;
1459
1460 const XMMRegister abcd = xmm0;
1461 const XMMRegister e0 = xmm1;
1462 const XMMRegister e1 = xmm2;
1463 const XMMRegister msg0 = xmm3;
1464
1465 const XMMRegister msg1 = xmm4;
1466 const XMMRegister msg2 = xmm5;
1467 const XMMRegister msg3 = xmm6;
1468 const XMMRegister shuf_mask = xmm7;
1469
1470 __ enter();
1471
1472 __ subptr(rsp, 4 * wordSize);
1473
1474 __ fast_sha1(abcd, e0, e1, msg0, msg1, msg2, msg3, shuf_mask,
1475 buf, state, ofs, limit, rsp, multi_block);
1476
1477 __ addptr(rsp, 4 * wordSize);
1478
1479 __ leave();
1480 __ ret(0);
1481
1482 return start;
1483 }
1484
1485 address StubGenerator::generate_pshuffle_byte_flip_mask() {
1486 __ align64();
1487 StubCodeMark mark(this, "StubRoutines", "pshuffle_byte_flip_mask");
1488 address start = __ pc();
1489
1490 __ emit_data64(0x0405060700010203, relocInfo::none);
1491 __ emit_data64(0x0c0d0e0f08090a0b, relocInfo::none);
1492
1493 if (VM_Version::supports_avx2()) {
1494 __ emit_data64(0x0405060700010203, relocInfo::none); // second copy
1495 __ emit_data64(0x0c0d0e0f08090a0b, relocInfo::none);
1496 // _SHUF_00BA
1497 __ emit_data64(0x0b0a090803020100, relocInfo::none);
1498 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1499 __ emit_data64(0x0b0a090803020100, relocInfo::none);
1500 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1501 // _SHUF_DC00
1502 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1503 __ emit_data64(0x0b0a090803020100, relocInfo::none);
1504 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1505 __ emit_data64(0x0b0a090803020100, relocInfo::none);
1506 }
1507
1508 return start;
1509 }
1510
1511 //Mask for byte-swapping a couple of qwords in an XMM register using (v)pshufb.
1512 address StubGenerator::generate_pshuffle_byte_flip_mask_sha512() {
1513 __ align32();
1514 StubCodeMark mark(this, "StubRoutines", "pshuffle_byte_flip_mask_sha512");
1515 address start = __ pc();
1516
1517 if (VM_Version::supports_avx2()) {
1518 __ emit_data64(0x0001020304050607, relocInfo::none); // PSHUFFLE_BYTE_FLIP_MASK
1519 __ emit_data64(0x08090a0b0c0d0e0f, relocInfo::none);
1520 __ emit_data64(0x1011121314151617, relocInfo::none);
1521 __ emit_data64(0x18191a1b1c1d1e1f, relocInfo::none);
1522 __ emit_data64(0x0000000000000000, relocInfo::none); //MASK_YMM_LO
1523 __ emit_data64(0x0000000000000000, relocInfo::none);
1524 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1525 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1526 }
1527
1528 return start;
1529 }
1530
1531 // ofs and limit are use for multi-block byte array.
1532 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
1533 address StubGenerator::generate_sha256_implCompress(bool multi_block, const char *name) {
1534 assert(VM_Version::supports_sha() || VM_Version::supports_avx2(), "");
1535 __ align(CodeEntryAlignment);
1536 StubCodeMark mark(this, "StubRoutines", name);
1537 address start = __ pc();
1538
1539 Register buf = c_rarg0;
1540 Register state = c_rarg1;
1541 Register ofs = c_rarg2;
1542 Register limit = c_rarg3;
1543
1544 const XMMRegister msg = xmm0;
1545 const XMMRegister state0 = xmm1;
1546 const XMMRegister state1 = xmm2;
1547 const XMMRegister msgtmp0 = xmm3;
1548
1549 const XMMRegister msgtmp1 = xmm4;
1550 const XMMRegister msgtmp2 = xmm5;
1551 const XMMRegister msgtmp3 = xmm6;
1552 const XMMRegister msgtmp4 = xmm7;
1553
1554 const XMMRegister shuf_mask = xmm8;
1555
1556 __ enter();
1557
1558 __ subptr(rsp, 4 * wordSize);
1559
1560 if (VM_Version::supports_sha()) {
1561 __ fast_sha256(msg, state0, state1, msgtmp0, msgtmp1, msgtmp2, msgtmp3, msgtmp4,
1562 buf, state, ofs, limit, rsp, multi_block, shuf_mask);
1563 } else if (VM_Version::supports_avx2()) {
1564 __ sha256_AVX2(msg, state0, state1, msgtmp0, msgtmp1, msgtmp2, msgtmp3, msgtmp4,
1565 buf, state, ofs, limit, rsp, multi_block, shuf_mask);
1566 }
1567 __ addptr(rsp, 4 * wordSize);
1568 __ vzeroupper();
1569 __ leave();
1570 __ ret(0);
1571
1572 return start;
1573 }
1574
1575 address StubGenerator::generate_sha512_implCompress(bool multi_block, const char *name) {
1576 assert(VM_Version::supports_avx2(), "");
1577 assert(VM_Version::supports_bmi2() || VM_Version::supports_sha512(), "");
1578 __ align(CodeEntryAlignment);
1579 StubCodeMark mark(this, "StubRoutines", name);
1580 address start = __ pc();
1581
1582 Register buf = c_rarg0;
1583 Register state = c_rarg1;
1584 Register ofs = c_rarg2;
1585 Register limit = c_rarg3;
1586
1587 __ enter();
1588
1589 if (VM_Version::supports_sha512()) {
1590 __ sha512_update_ni_x1(state, buf, ofs, limit, multi_block);
1591 } else {
1592 const XMMRegister msg = xmm0;
1593 const XMMRegister state0 = xmm1;
1594 const XMMRegister state1 = xmm2;
1595 const XMMRegister msgtmp0 = xmm3;
1596 const XMMRegister msgtmp1 = xmm4;
1597 const XMMRegister msgtmp2 = xmm5;
1598 const XMMRegister msgtmp3 = xmm6;
1599 const XMMRegister msgtmp4 = xmm7;
1600
1601 const XMMRegister shuf_mask = xmm8;
1602 __ sha512_AVX2(msg, state0, state1, msgtmp0, msgtmp1, msgtmp2, msgtmp3, msgtmp4,
1603 buf, state, ofs, limit, rsp, multi_block, shuf_mask);
1604 }
1605 __ vzeroupper();
1606 __ leave();
1607 __ ret(0);
1608
1609 return start;
1610 }
1611
1612 address StubGenerator::base64_shuffle_addr() {
1613 __ align64();
1614 StubCodeMark mark(this, "StubRoutines", "shuffle_base64");
1615 address start = __ pc();
1616
1617 assert(((unsigned long long)start & 0x3f) == 0,
1618 "Alignment problem (0x%08llx)", (unsigned long long)start);
1619 __ emit_data64(0x0405030401020001, relocInfo::none);
1620 __ emit_data64(0x0a0b090a07080607, relocInfo::none);
1621 __ emit_data64(0x10110f100d0e0c0d, relocInfo::none);
1622 __ emit_data64(0x1617151613141213, relocInfo::none);
1623 __ emit_data64(0x1c1d1b1c191a1819, relocInfo::none);
1624 __ emit_data64(0x222321221f201e1f, relocInfo::none);
1625 __ emit_data64(0x2829272825262425, relocInfo::none);
1626 __ emit_data64(0x2e2f2d2e2b2c2a2b, relocInfo::none);
1627
1628 return start;
1629 }
1630
1631 address StubGenerator::base64_avx2_shuffle_addr() {
1632 __ align32();
1633 StubCodeMark mark(this, "StubRoutines", "avx2_shuffle_base64");
1634 address start = __ pc();
1635
1636 __ emit_data64(0x0809070805060405, relocInfo::none);
1637 __ emit_data64(0x0e0f0d0e0b0c0a0b, relocInfo::none);
1638 __ emit_data64(0x0405030401020001, relocInfo::none);
1639 __ emit_data64(0x0a0b090a07080607, relocInfo::none);
1640
1641 return start;
1642 }
1643
1644 address StubGenerator::base64_avx2_input_mask_addr() {
1645 __ align32();
1646 StubCodeMark mark(this, "StubRoutines", "avx2_input_mask_base64");
1647 address start = __ pc();
1648
1649 __ emit_data64(0x8000000000000000, relocInfo::none);
1650 __ emit_data64(0x8000000080000000, relocInfo::none);
1651 __ emit_data64(0x8000000080000000, relocInfo::none);
1652 __ emit_data64(0x8000000080000000, relocInfo::none);
1653
1654 return start;
1655 }
1656
1657 address StubGenerator::base64_avx2_lut_addr() {
1658 __ align32();
1659 StubCodeMark mark(this, "StubRoutines", "avx2_lut_base64");
1660 address start = __ pc();
1661
1662 __ emit_data64(0xfcfcfcfcfcfc4741, relocInfo::none);
1663 __ emit_data64(0x0000f0edfcfcfcfc, relocInfo::none);
1664 __ emit_data64(0xfcfcfcfcfcfc4741, relocInfo::none);
1665 __ emit_data64(0x0000f0edfcfcfcfc, relocInfo::none);
1666
1667 // URL LUT
1668 __ emit_data64(0xfcfcfcfcfcfc4741, relocInfo::none);
1669 __ emit_data64(0x000020effcfcfcfc, relocInfo::none);
1670 __ emit_data64(0xfcfcfcfcfcfc4741, relocInfo::none);
1671 __ emit_data64(0x000020effcfcfcfc, relocInfo::none);
1672
1673 return start;
1674 }
1675
1676 address StubGenerator::base64_encoding_table_addr() {
1677 __ align64();
1678 StubCodeMark mark(this, "StubRoutines", "encoding_table_base64");
1679 address start = __ pc();
1680
1681 assert(((unsigned long long)start & 0x3f) == 0, "Alignment problem (0x%08llx)", (unsigned long long)start);
1682 __ emit_data64(0x4847464544434241, relocInfo::none);
1683 __ emit_data64(0x504f4e4d4c4b4a49, relocInfo::none);
1684 __ emit_data64(0x5857565554535251, relocInfo::none);
1685 __ emit_data64(0x6665646362615a59, relocInfo::none);
1686 __ emit_data64(0x6e6d6c6b6a696867, relocInfo::none);
1687 __ emit_data64(0x767574737271706f, relocInfo::none);
1688 __ emit_data64(0x333231307a797877, relocInfo::none);
1689 __ emit_data64(0x2f2b393837363534, relocInfo::none);
1690
1691 // URL table
1692 __ emit_data64(0x4847464544434241, relocInfo::none);
1693 __ emit_data64(0x504f4e4d4c4b4a49, relocInfo::none);
1694 __ emit_data64(0x5857565554535251, relocInfo::none);
1695 __ emit_data64(0x6665646362615a59, relocInfo::none);
1696 __ emit_data64(0x6e6d6c6b6a696867, relocInfo::none);
1697 __ emit_data64(0x767574737271706f, relocInfo::none);
1698 __ emit_data64(0x333231307a797877, relocInfo::none);
1699 __ emit_data64(0x5f2d393837363534, relocInfo::none);
1700
1701 return start;
1702 }
1703
1704 // Code for generating Base64 encoding.
1705 // Intrinsic function prototype in Base64.java:
1706 // private void encodeBlock(byte[] src, int sp, int sl, byte[] dst, int dp,
1707 // boolean isURL) {
1708 address StubGenerator::generate_base64_encodeBlock()
1709 {
1710 __ align(CodeEntryAlignment);
1711 StubCodeMark mark(this, "StubRoutines", "implEncode");
1712 address start = __ pc();
1713
1714 __ enter();
1715
1716 // Save callee-saved registers before using them
1717 __ push(r12);
1718 __ push(r13);
1719 __ push(r14);
1720 __ push(r15);
1721
1722 // arguments
1723 const Register source = c_rarg0; // Source Array
1724 const Register start_offset = c_rarg1; // start offset
1725 const Register end_offset = c_rarg2; // end offset
1726 const Register dest = c_rarg3; // destination array
1727
1728 #ifndef _WIN64
1729 const Register dp = c_rarg4; // Position for writing to dest array
1730 const Register isURL = c_rarg5; // Base64 or URL character set
1731 #else
1732 const Address dp_mem(rbp, 6 * wordSize); // length is on stack on Win64
1733 const Address isURL_mem(rbp, 7 * wordSize);
1734 const Register isURL = r10; // pick the volatile windows register
1735 const Register dp = r12;
1736 __ movl(dp, dp_mem);
1737 __ movl(isURL, isURL_mem);
1738 #endif
1739
1740 const Register length = r14;
1741 const Register encode_table = r13;
1742 Label L_process3, L_exit, L_processdata, L_vbmiLoop, L_not512, L_32byteLoop;
1743
1744 // calculate length from offsets
1745 __ movl(length, end_offset);
1746 __ subl(length, start_offset);
1747 __ jcc(Assembler::lessEqual, L_exit);
1748
1749 // Code for 512-bit VBMI encoding. Encodes 48 input bytes into 64
1750 // output bytes. We read 64 input bytes and ignore the last 16, so be
1751 // sure not to read past the end of the input buffer.
1752 if (VM_Version::supports_avx512_vbmi()) {
1753 __ cmpl(length, 64); // Do not overrun input buffer.
1754 __ jcc(Assembler::below, L_not512);
1755
1756 __ shll(isURL, 6); // index into decode table based on isURL
1757 __ lea(encode_table, ExternalAddress(StubRoutines::x86::base64_encoding_table_addr()));
1758 __ addptr(encode_table, isURL);
1759 __ shrl(isURL, 6); // restore isURL
1760
1761 __ mov64(rax, 0x3036242a1016040aull); // Shifts
1762 __ evmovdquq(xmm3, ExternalAddress(StubRoutines::x86::base64_shuffle_addr()), Assembler::AVX_512bit, r15);
1763 __ evmovdquq(xmm2, Address(encode_table, 0), Assembler::AVX_512bit);
1764 __ evpbroadcastq(xmm1, rax, Assembler::AVX_512bit);
1765
1766 __ align32();
1767 __ BIND(L_vbmiLoop);
1768
1769 __ vpermb(xmm0, xmm3, Address(source, start_offset), Assembler::AVX_512bit);
1770 __ subl(length, 48);
1771
1772 // Put the input bytes into the proper lanes for writing, then
1773 // encode them.
1774 __ evpmultishiftqb(xmm0, xmm1, xmm0, Assembler::AVX_512bit);
1775 __ vpermb(xmm0, xmm0, xmm2, Assembler::AVX_512bit);
1776
1777 // Write to destination
1778 __ evmovdquq(Address(dest, dp), xmm0, Assembler::AVX_512bit);
1779
1780 __ addptr(dest, 64);
1781 __ addptr(source, 48);
1782 __ cmpl(length, 64);
1783 __ jcc(Assembler::aboveEqual, L_vbmiLoop);
1784
1785 __ vzeroupper();
1786 }
1787
1788 __ BIND(L_not512);
1789 if (VM_Version::supports_avx2()) {
1790 /*
1791 ** This AVX2 encoder is based off the paper at:
1792 ** https://dl.acm.org/doi/10.1145/3132709
1793 **
1794 ** We use AVX2 SIMD instructions to encode 24 bytes into 32
1795 ** output bytes.
1796 **
1797 */
1798 // Lengths under 32 bytes are done with scalar routine
1799 __ cmpl(length, 31);
1800 __ jcc(Assembler::belowEqual, L_process3);
1801
1802 // Set up supporting constant table data
1803 __ vmovdqu(xmm9, ExternalAddress(StubRoutines::x86::base64_avx2_shuffle_addr()), rax);
1804 // 6-bit mask for 2nd and 4th (and multiples) 6-bit values
1805 __ movl(rax, 0x0fc0fc00);
1806 __ movdl(xmm8, rax);
1807 __ vmovdqu(xmm1, ExternalAddress(StubRoutines::x86::base64_avx2_input_mask_addr()), rax);
1808 __ vpbroadcastd(xmm8, xmm8, Assembler::AVX_256bit);
1809
1810 // Multiplication constant for "shifting" right by 6 and 10
1811 // bits
1812 __ movl(rax, 0x04000040);
1813
1814 __ subl(length, 24);
1815 __ movdl(xmm7, rax);
1816 __ vpbroadcastd(xmm7, xmm7, Assembler::AVX_256bit);
1817
1818 // For the first load, we mask off reading of the first 4
1819 // bytes into the register. This is so we can get 4 3-byte
1820 // chunks into each lane of the register, avoiding having to
1821 // handle end conditions. We then shuffle these bytes into a
1822 // specific order so that manipulation is easier.
1823 //
1824 // The initial read loads the XMM register like this:
1825 //
1826 // Lower 128-bit lane:
1827 // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
1828 // | XX | XX | XX | XX | A0 | A1 | A2 | B0 | B1 | B2 | C0 | C1
1829 // | C2 | D0 | D1 | D2 |
1830 // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
1831 //
1832 // Upper 128-bit lane:
1833 // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
1834 // | E0 | E1 | E2 | F0 | F1 | F2 | G0 | G1 | G2 | H0 | H1 | H2
1835 // | XX | XX | XX | XX |
1836 // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
1837 //
1838 // Where A0 is the first input byte, B0 is the fourth, etc.
1839 // The alphabetical significance denotes the 3 bytes to be
1840 // consumed and encoded into 4 bytes.
1841 //
1842 // We then shuffle the register so each 32-bit word contains
1843 // the sequence:
1844 // A1 A0 A2 A1, B1, B0, B2, B1, etc.
1845 // Each of these byte sequences are then manipulated into 4
1846 // 6-bit values ready for encoding.
1847 //
1848 // If we focus on one set of 3-byte chunks, changing the
1849 // nomenclature such that A0 => a, A1 => b, and A2 => c, we
1850 // shuffle such that each 24-bit chunk contains:
1851 //
1852 // b7 b6 b5 b4 b3 b2 b1 b0 | a7 a6 a5 a4 a3 a2 a1 a0 | c7 c6
1853 // c5 c4 c3 c2 c1 c0 | b7 b6 b5 b4 b3 b2 b1 b0
1854 // Explain this step.
1855 // b3 b2 b1 b0 c5 c4 c3 c2 | c1 c0 d5 d4 d3 d2 d1 d0 | a5 a4
1856 // a3 a2 a1 a0 b5 b4 | b3 b2 b1 b0 c5 c4 c3 c2
1857 //
1858 // W first and off all but bits 4-9 and 16-21 (c5..c0 and
1859 // a5..a0) and shift them using a vector multiplication
1860 // operation (vpmulhuw) which effectively shifts c right by 6
1861 // bits and a right by 10 bits. We similarly mask bits 10-15
1862 // (d5..d0) and 22-27 (b5..b0) and shift them left by 8 and 4
1863 // bits respectively. This is done using vpmullw. We end up
1864 // with 4 6-bit values, thus splitting the 3 input bytes,
1865 // ready for encoding:
1866 // 0 0 d5..d0 0 0 c5..c0 0 0 b5..b0 0 0 a5..a0
1867 //
1868 // For translation, we recognize that there are 5 distinct
1869 // ranges of legal Base64 characters as below:
1870 //
1871 // +-------------+-------------+------------+
1872 // | 6-bit value | ASCII range | offset |
1873 // +-------------+-------------+------------+
1874 // | 0..25 | A..Z | 65 |
1875 // | 26..51 | a..z | 71 |
1876 // | 52..61 | 0..9 | -4 |
1877 // | 62 | + or - | -19 or -17 |
1878 // | 63 | / or _ | -16 or 32 |
1879 // +-------------+-------------+------------+
1880 //
1881 // We note that vpshufb does a parallel lookup in a
1882 // destination register using the lower 4 bits of bytes from a
1883 // source register. If we use a saturated subtraction and
1884 // subtract 51 from each 6-bit value, bytes from [0,51]
1885 // saturate to 0, and [52,63] map to a range of [1,12]. We
1886 // distinguish the [0,25] and [26,51] ranges by assigning a
1887 // value of 13 for all 6-bit values less than 26. We end up
1888 // with:
1889 //
1890 // +-------------+-------------+------------+
1891 // | 6-bit value | Reduced | offset |
1892 // +-------------+-------------+------------+
1893 // | 0..25 | 13 | 65 |
1894 // | 26..51 | 0 | 71 |
1895 // | 52..61 | 0..9 | -4 |
1896 // | 62 | 11 | -19 or -17 |
1897 // | 63 | 12 | -16 or 32 |
1898 // +-------------+-------------+------------+
1899 //
1900 // We then use a final vpshufb to add the appropriate offset,
1901 // translating the bytes.
1902 //
1903 // Load input bytes - only 28 bytes. Mask the first load to
1904 // not load into the full register.
1905 __ vpmaskmovd(xmm1, xmm1, Address(source, start_offset, Address::times_1, -4), Assembler::AVX_256bit);
1906
1907 // Move 3-byte chunks of input (12 bytes) into 16 bytes,
1908 // ordering by:
1909 // 1, 0, 2, 1; 4, 3, 5, 4; etc. This groups 6-bit chunks
1910 // for easy masking
1911 __ vpshufb(xmm1, xmm1, xmm9, Assembler::AVX_256bit);
1912
1913 __ addl(start_offset, 24);
1914
1915 // Load masking register for first and third (and multiples)
1916 // 6-bit values.
1917 __ movl(rax, 0x003f03f0);
1918 __ movdl(xmm6, rax);
1919 __ vpbroadcastd(xmm6, xmm6, Assembler::AVX_256bit);
1920 // Multiplication constant for "shifting" left by 4 and 8 bits
1921 __ movl(rax, 0x01000010);
1922 __ movdl(xmm5, rax);
1923 __ vpbroadcastd(xmm5, xmm5, Assembler::AVX_256bit);
1924
1925 // Isolate 6-bit chunks of interest
1926 __ vpand(xmm0, xmm8, xmm1, Assembler::AVX_256bit);
1927
1928 // Load constants for encoding
1929 __ movl(rax, 0x19191919);
1930 __ movdl(xmm3, rax);
1931 __ vpbroadcastd(xmm3, xmm3, Assembler::AVX_256bit);
1932 __ movl(rax, 0x33333333);
1933 __ movdl(xmm4, rax);
1934 __ vpbroadcastd(xmm4, xmm4, Assembler::AVX_256bit);
1935
1936 // Shift output bytes 0 and 2 into proper lanes
1937 __ vpmulhuw(xmm2, xmm0, xmm7, Assembler::AVX_256bit);
1938
1939 // Mask and shift output bytes 1 and 3 into proper lanes and
1940 // combine
1941 __ vpand(xmm0, xmm6, xmm1, Assembler::AVX_256bit);
1942 __ vpmullw(xmm0, xmm5, xmm0, Assembler::AVX_256bit);
1943 __ vpor(xmm0, xmm0, xmm2, Assembler::AVX_256bit);
1944
1945 // Find out which are 0..25. This indicates which input
1946 // values fall in the range of 'A'-'Z', which require an
1947 // additional offset (see comments above)
1948 __ vpcmpgtb(xmm2, xmm0, xmm3, Assembler::AVX_256bit);
1949 __ vpsubusb(xmm1, xmm0, xmm4, Assembler::AVX_256bit);
1950 __ vpsubb(xmm1, xmm1, xmm2, Assembler::AVX_256bit);
1951
1952 // Load the proper lookup table
1953 __ lea(r11, ExternalAddress(StubRoutines::x86::base64_avx2_lut_addr()));
1954 __ movl(r15, isURL);
1955 __ shll(r15, 5);
1956 __ vmovdqu(xmm2, Address(r11, r15));
1957
1958 // Shuffle the offsets based on the range calculation done
1959 // above. This allows us to add the correct offset to the
1960 // 6-bit value corresponding to the range documented above.
1961 __ vpshufb(xmm1, xmm2, xmm1, Assembler::AVX_256bit);
1962 __ vpaddb(xmm0, xmm1, xmm0, Assembler::AVX_256bit);
1963
1964 // Store the encoded bytes
1965 __ vmovdqu(Address(dest, dp), xmm0);
1966 __ addl(dp, 32);
1967
1968 __ cmpl(length, 31);
1969 __ jcc(Assembler::belowEqual, L_process3);
1970
1971 __ align32();
1972 __ BIND(L_32byteLoop);
1973
1974 // Get next 32 bytes
1975 __ vmovdqu(xmm1, Address(source, start_offset, Address::times_1, -4));
1976
1977 __ subl(length, 24);
1978 __ addl(start_offset, 24);
1979
1980 // This logic is identical to the above, with only constant
1981 // register loads removed. Shuffle the input, mask off 6-bit
1982 // chunks, shift them into place, then add the offset to
1983 // encode.
1984 __ vpshufb(xmm1, xmm1, xmm9, Assembler::AVX_256bit);
1985
1986 __ vpand(xmm0, xmm8, xmm1, Assembler::AVX_256bit);
1987 __ vpmulhuw(xmm10, xmm0, xmm7, Assembler::AVX_256bit);
1988 __ vpand(xmm0, xmm6, xmm1, Assembler::AVX_256bit);
1989 __ vpmullw(xmm0, xmm5, xmm0, Assembler::AVX_256bit);
1990 __ vpor(xmm0, xmm0, xmm10, Assembler::AVX_256bit);
1991 __ vpcmpgtb(xmm10, xmm0, xmm3, Assembler::AVX_256bit);
1992 __ vpsubusb(xmm1, xmm0, xmm4, Assembler::AVX_256bit);
1993 __ vpsubb(xmm1, xmm1, xmm10, Assembler::AVX_256bit);
1994 __ vpshufb(xmm1, xmm2, xmm1, Assembler::AVX_256bit);
1995 __ vpaddb(xmm0, xmm1, xmm0, Assembler::AVX_256bit);
1996
1997 // Store the encoded bytes
1998 __ vmovdqu(Address(dest, dp), xmm0);
1999 __ addl(dp, 32);
2000
2001 __ cmpl(length, 31);
2002 __ jcc(Assembler::above, L_32byteLoop);
2003
2004 __ BIND(L_process3);
2005 __ vzeroupper();
2006 } else {
2007 __ BIND(L_process3);
2008 }
2009
2010 __ cmpl(length, 3);
2011 __ jcc(Assembler::below, L_exit);
2012
2013 // Load the encoding table based on isURL
2014 __ lea(r11, ExternalAddress(StubRoutines::x86::base64_encoding_table_addr()));
2015 __ movl(r15, isURL);
2016 __ shll(r15, 6);
2017 __ addptr(r11, r15);
2018
2019 __ BIND(L_processdata);
2020
2021 // Load 3 bytes
2022 __ load_unsigned_byte(r15, Address(source, start_offset));
2023 __ load_unsigned_byte(r10, Address(source, start_offset, Address::times_1, 1));
2024 __ load_unsigned_byte(r13, Address(source, start_offset, Address::times_1, 2));
2025
2026 // Build a 32-bit word with bytes 1, 2, 0, 1
2027 __ movl(rax, r10);
2028 __ shll(r10, 24);
2029 __ orl(rax, r10);
2030
2031 __ subl(length, 3);
2032
2033 __ shll(r15, 8);
2034 __ shll(r13, 16);
2035 __ orl(rax, r15);
2036
2037 __ addl(start_offset, 3);
2038
2039 __ orl(rax, r13);
2040 // At this point, rax contains | byte1 | byte2 | byte0 | byte1
2041 // r13 has byte2 << 16 - need low-order 6 bits to translate.
2042 // This translated byte is the fourth output byte.
2043 __ shrl(r13, 16);
2044 __ andl(r13, 0x3f);
2045
2046 // The high-order 6 bits of r15 (byte0) is translated.
2047 // The translated byte is the first output byte.
2048 __ shrl(r15, 10);
2049
2050 __ load_unsigned_byte(r13, Address(r11, r13));
2051 __ load_unsigned_byte(r15, Address(r11, r15));
2052
2053 __ movb(Address(dest, dp, Address::times_1, 3), r13);
2054
2055 // Extract high-order 4 bits of byte1 and low-order 2 bits of byte0.
2056 // This translated byte is the second output byte.
2057 __ shrl(rax, 4);
2058 __ movl(r10, rax);
2059 __ andl(rax, 0x3f);
2060
2061 __ movb(Address(dest, dp, Address::times_1, 0), r15);
2062
2063 __ load_unsigned_byte(rax, Address(r11, rax));
2064
2065 // Extract low-order 2 bits of byte1 and high-order 4 bits of byte2.
2066 // This translated byte is the third output byte.
2067 __ shrl(r10, 18);
2068 __ andl(r10, 0x3f);
2069
2070 __ load_unsigned_byte(r10, Address(r11, r10));
2071
2072 __ movb(Address(dest, dp, Address::times_1, 1), rax);
2073 __ movb(Address(dest, dp, Address::times_1, 2), r10);
2074
2075 __ addl(dp, 4);
2076 __ cmpl(length, 3);
2077 __ jcc(Assembler::aboveEqual, L_processdata);
2078
2079 __ BIND(L_exit);
2080 __ pop(r15);
2081 __ pop(r14);
2082 __ pop(r13);
2083 __ pop(r12);
2084 __ leave();
2085 __ ret(0);
2086
2087 return start;
2088 }
2089
2090 // base64 AVX512vbmi tables
2091 address StubGenerator::base64_vbmi_lookup_lo_addr() {
2092 __ align64();
2093 StubCodeMark mark(this, "StubRoutines", "lookup_lo_base64");
2094 address start = __ pc();
2095
2096 assert(((unsigned long long)start & 0x3f) == 0,
2097 "Alignment problem (0x%08llx)", (unsigned long long)start);
2098 __ emit_data64(0x8080808080808080, relocInfo::none);
2099 __ emit_data64(0x8080808080808080, relocInfo::none);
2100 __ emit_data64(0x8080808080808080, relocInfo::none);
2101 __ emit_data64(0x8080808080808080, relocInfo::none);
2102 __ emit_data64(0x8080808080808080, relocInfo::none);
2103 __ emit_data64(0x3f8080803e808080, relocInfo::none);
2104 __ emit_data64(0x3b3a393837363534, relocInfo::none);
2105 __ emit_data64(0x8080808080803d3c, relocInfo::none);
2106
2107 return start;
2108 }
2109
2110 address StubGenerator::base64_vbmi_lookup_hi_addr() {
2111 __ align64();
2112 StubCodeMark mark(this, "StubRoutines", "lookup_hi_base64");
2113 address start = __ pc();
2114
2115 assert(((unsigned long long)start & 0x3f) == 0,
2116 "Alignment problem (0x%08llx)", (unsigned long long)start);
2117 __ emit_data64(0x0605040302010080, relocInfo::none);
2118 __ emit_data64(0x0e0d0c0b0a090807, relocInfo::none);
2119 __ emit_data64(0x161514131211100f, relocInfo::none);
2120 __ emit_data64(0x8080808080191817, relocInfo::none);
2121 __ emit_data64(0x201f1e1d1c1b1a80, relocInfo::none);
2122 __ emit_data64(0x2827262524232221, relocInfo::none);
2123 __ emit_data64(0x302f2e2d2c2b2a29, relocInfo::none);
2124 __ emit_data64(0x8080808080333231, relocInfo::none);
2125
2126 return start;
2127 }
2128 address StubGenerator::base64_vbmi_lookup_lo_url_addr() {
2129 __ align64();
2130 StubCodeMark mark(this, "StubRoutines", "lookup_lo_base64url");
2131 address start = __ pc();
2132
2133 assert(((unsigned long long)start & 0x3f) == 0,
2134 "Alignment problem (0x%08llx)", (unsigned long long)start);
2135 __ emit_data64(0x8080808080808080, relocInfo::none);
2136 __ emit_data64(0x8080808080808080, relocInfo::none);
2137 __ emit_data64(0x8080808080808080, relocInfo::none);
2138 __ emit_data64(0x8080808080808080, relocInfo::none);
2139 __ emit_data64(0x8080808080808080, relocInfo::none);
2140 __ emit_data64(0x80803e8080808080, relocInfo::none);
2141 __ emit_data64(0x3b3a393837363534, relocInfo::none);
2142 __ emit_data64(0x8080808080803d3c, relocInfo::none);
2143
2144 return start;
2145 }
2146
2147 address StubGenerator::base64_vbmi_lookup_hi_url_addr() {
2148 __ align64();
2149 StubCodeMark mark(this, "StubRoutines", "lookup_hi_base64url");
2150 address start = __ pc();
2151
2152 assert(((unsigned long long)start & 0x3f) == 0,
2153 "Alignment problem (0x%08llx)", (unsigned long long)start);
2154 __ emit_data64(0x0605040302010080, relocInfo::none);
2155 __ emit_data64(0x0e0d0c0b0a090807, relocInfo::none);
2156 __ emit_data64(0x161514131211100f, relocInfo::none);
2157 __ emit_data64(0x3f80808080191817, relocInfo::none);
2158 __ emit_data64(0x201f1e1d1c1b1a80, relocInfo::none);
2159 __ emit_data64(0x2827262524232221, relocInfo::none);
2160 __ emit_data64(0x302f2e2d2c2b2a29, relocInfo::none);
2161 __ emit_data64(0x8080808080333231, relocInfo::none);
2162
2163 return start;
2164 }
2165
2166 address StubGenerator::base64_vbmi_pack_vec_addr() {
2167 __ align64();
2168 StubCodeMark mark(this, "StubRoutines", "pack_vec_base64");
2169 address start = __ pc();
2170
2171 assert(((unsigned long long)start & 0x3f) == 0,
2172 "Alignment problem (0x%08llx)", (unsigned long long)start);
2173 __ emit_data64(0x090a040506000102, relocInfo::none);
2174 __ emit_data64(0x161011120c0d0e08, relocInfo::none);
2175 __ emit_data64(0x1c1d1e18191a1415, relocInfo::none);
2176 __ emit_data64(0x292a242526202122, relocInfo::none);
2177 __ emit_data64(0x363031322c2d2e28, relocInfo::none);
2178 __ emit_data64(0x3c3d3e38393a3435, relocInfo::none);
2179 __ emit_data64(0x0000000000000000, relocInfo::none);
2180 __ emit_data64(0x0000000000000000, relocInfo::none);
2181
2182 return start;
2183 }
2184
2185 address StubGenerator::base64_vbmi_join_0_1_addr() {
2186 __ align64();
2187 StubCodeMark mark(this, "StubRoutines", "join_0_1_base64");
2188 address start = __ pc();
2189
2190 assert(((unsigned long long)start & 0x3f) == 0,
2191 "Alignment problem (0x%08llx)", (unsigned long long)start);
2192 __ emit_data64(0x090a040506000102, relocInfo::none);
2193 __ emit_data64(0x161011120c0d0e08, relocInfo::none);
2194 __ emit_data64(0x1c1d1e18191a1415, relocInfo::none);
2195 __ emit_data64(0x292a242526202122, relocInfo::none);
2196 __ emit_data64(0x363031322c2d2e28, relocInfo::none);
2197 __ emit_data64(0x3c3d3e38393a3435, relocInfo::none);
2198 __ emit_data64(0x494a444546404142, relocInfo::none);
2199 __ emit_data64(0x565051524c4d4e48, relocInfo::none);
2200
2201 return start;
2202 }
2203
2204 address StubGenerator::base64_vbmi_join_1_2_addr() {
2205 __ align64();
2206 StubCodeMark mark(this, "StubRoutines", "join_1_2_base64");
2207 address start = __ pc();
2208
2209 assert(((unsigned long long)start & 0x3f) == 0,
2210 "Alignment problem (0x%08llx)", (unsigned long long)start);
2211 __ emit_data64(0x1c1d1e18191a1415, relocInfo::none);
2212 __ emit_data64(0x292a242526202122, relocInfo::none);
2213 __ emit_data64(0x363031322c2d2e28, relocInfo::none);
2214 __ emit_data64(0x3c3d3e38393a3435, relocInfo::none);
2215 __ emit_data64(0x494a444546404142, relocInfo::none);
2216 __ emit_data64(0x565051524c4d4e48, relocInfo::none);
2217 __ emit_data64(0x5c5d5e58595a5455, relocInfo::none);
2218 __ emit_data64(0x696a646566606162, relocInfo::none);
2219
2220 return start;
2221 }
2222
2223 address StubGenerator::base64_vbmi_join_2_3_addr() {
2224 __ align64();
2225 StubCodeMark mark(this, "StubRoutines", "join_2_3_base64");
2226 address start = __ pc();
2227
2228 assert(((unsigned long long)start & 0x3f) == 0,
2229 "Alignment problem (0x%08llx)", (unsigned long long)start);
2230 __ emit_data64(0x363031322c2d2e28, relocInfo::none);
2231 __ emit_data64(0x3c3d3e38393a3435, relocInfo::none);
2232 __ emit_data64(0x494a444546404142, relocInfo::none);
2233 __ emit_data64(0x565051524c4d4e48, relocInfo::none);
2234 __ emit_data64(0x5c5d5e58595a5455, relocInfo::none);
2235 __ emit_data64(0x696a646566606162, relocInfo::none);
2236 __ emit_data64(0x767071726c6d6e68, relocInfo::none);
2237 __ emit_data64(0x7c7d7e78797a7475, relocInfo::none);
2238
2239 return start;
2240 }
2241
2242 address StubGenerator::base64_AVX2_decode_tables_addr() {
2243 __ align64();
2244 StubCodeMark mark(this, "StubRoutines", "AVX2_tables_base64");
2245 address start = __ pc();
2246
2247 assert(((unsigned long long)start & 0x3f) == 0,
2248 "Alignment problem (0x%08llx)", (unsigned long long)start);
2249 __ emit_data(0x2f2f2f2f, relocInfo::none, 0);
2250 __ emit_data(0x5f5f5f5f, relocInfo::none, 0); // for URL
2251
2252 __ emit_data(0xffffffff, relocInfo::none, 0);
2253 __ emit_data(0xfcfcfcfc, relocInfo::none, 0); // for URL
2254
2255 // Permute table
2256 __ emit_data64(0x0000000100000000, relocInfo::none);
2257 __ emit_data64(0x0000000400000002, relocInfo::none);
2258 __ emit_data64(0x0000000600000005, relocInfo::none);
2259 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2260
2261 // Shuffle table
2262 __ emit_data64(0x090a040506000102, relocInfo::none);
2263 __ emit_data64(0xffffffff0c0d0e08, relocInfo::none);
2264 __ emit_data64(0x090a040506000102, relocInfo::none);
2265 __ emit_data64(0xffffffff0c0d0e08, relocInfo::none);
2266
2267 // merge table
2268 __ emit_data(0x01400140, relocInfo::none, 0);
2269
2270 // merge multiplier
2271 __ emit_data(0x00011000, relocInfo::none, 0);
2272
2273 return start;
2274 }
2275
2276 address StubGenerator::base64_AVX2_decode_LUT_tables_addr() {
2277 __ align64();
2278 StubCodeMark mark(this, "StubRoutines", "AVX2_tables_URL_base64");
2279 address start = __ pc();
2280
2281 assert(((unsigned long long)start & 0x3f) == 0,
2282 "Alignment problem (0x%08llx)", (unsigned long long)start);
2283 // lut_lo
2284 __ emit_data64(0x1111111111111115, relocInfo::none);
2285 __ emit_data64(0x1a1b1b1b1a131111, relocInfo::none);
2286 __ emit_data64(0x1111111111111115, relocInfo::none);
2287 __ emit_data64(0x1a1b1b1b1a131111, relocInfo::none);
2288
2289 // lut_roll
2290 __ emit_data64(0xb9b9bfbf04131000, relocInfo::none);
2291 __ emit_data64(0x0000000000000000, relocInfo::none);
2292 __ emit_data64(0xb9b9bfbf04131000, relocInfo::none);
2293 __ emit_data64(0x0000000000000000, relocInfo::none);
2294
2295 // lut_lo URL
2296 __ emit_data64(0x1111111111111115, relocInfo::none);
2297 __ emit_data64(0x1b1b1a1b1b131111, relocInfo::none);
2298 __ emit_data64(0x1111111111111115, relocInfo::none);
2299 __ emit_data64(0x1b1b1a1b1b131111, relocInfo::none);
2300
2301 // lut_roll URL
2302 __ emit_data64(0xb9b9bfbf0411e000, relocInfo::none);
2303 __ emit_data64(0x0000000000000000, relocInfo::none);
2304 __ emit_data64(0xb9b9bfbf0411e000, relocInfo::none);
2305 __ emit_data64(0x0000000000000000, relocInfo::none);
2306
2307 // lut_hi
2308 __ emit_data64(0x0804080402011010, relocInfo::none);
2309 __ emit_data64(0x1010101010101010, relocInfo::none);
2310 __ emit_data64(0x0804080402011010, relocInfo::none);
2311 __ emit_data64(0x1010101010101010, relocInfo::none);
2312
2313 return start;
2314 }
2315
2316 address StubGenerator::base64_decoding_table_addr() {
2317 StubCodeMark mark(this, "StubRoutines", "decoding_table_base64");
2318 address start = __ pc();
2319
2320 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2321 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2322 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2323 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2324 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2325 __ emit_data64(0x3fffffff3effffff, relocInfo::none);
2326 __ emit_data64(0x3b3a393837363534, relocInfo::none);
2327 __ emit_data64(0xffffffffffff3d3c, relocInfo::none);
2328 __ emit_data64(0x06050403020100ff, relocInfo::none);
2329 __ emit_data64(0x0e0d0c0b0a090807, relocInfo::none);
2330 __ emit_data64(0x161514131211100f, relocInfo::none);
2331 __ emit_data64(0xffffffffff191817, relocInfo::none);
2332 __ emit_data64(0x201f1e1d1c1b1aff, relocInfo::none);
2333 __ emit_data64(0x2827262524232221, relocInfo::none);
2334 __ emit_data64(0x302f2e2d2c2b2a29, relocInfo::none);
2335 __ emit_data64(0xffffffffff333231, relocInfo::none);
2336 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2337 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2338 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2339 __ emit_data64(0xffffffffffffffff, relocInfo::none);
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(0xffffffffffffffff, relocInfo::none);
2346 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2347 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2348 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2349 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2350 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2351 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2352
2353 // URL table
2354 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2355 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2356 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2357 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2358 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2359 __ emit_data64(0xffff3effffffffff, relocInfo::none);
2360 __ emit_data64(0x3b3a393837363534, relocInfo::none);
2361 __ emit_data64(0xffffffffffff3d3c, relocInfo::none);
2362 __ emit_data64(0x06050403020100ff, relocInfo::none);
2363 __ emit_data64(0x0e0d0c0b0a090807, relocInfo::none);
2364 __ emit_data64(0x161514131211100f, relocInfo::none);
2365 __ emit_data64(0x3fffffffff191817, relocInfo::none);
2366 __ emit_data64(0x201f1e1d1c1b1aff, relocInfo::none);
2367 __ emit_data64(0x2827262524232221, relocInfo::none);
2368 __ emit_data64(0x302f2e2d2c2b2a29, relocInfo::none);
2369 __ emit_data64(0xffffffffff333231, relocInfo::none);
2370 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2371 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2372 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2373 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2374 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2375 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2376 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2377 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2378 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2379 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2380 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2381 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2382 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2383 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2384 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2385 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2386
2387 return start;
2388 }
2389
2390
2391 // Code for generating Base64 decoding.
2392 //
2393 // Based on the article (and associated code) from https://arxiv.org/abs/1910.05109.
2394 //
2395 // Intrinsic function prototype in Base64.java:
2396 // private void decodeBlock(byte[] src, int sp, int sl, byte[] dst, int dp, boolean isURL, isMIME) {
2397 address StubGenerator::generate_base64_decodeBlock() {
2398 __ align(CodeEntryAlignment);
2399 StubCodeMark mark(this, "StubRoutines", "implDecode");
2400 address start = __ pc();
2401
2402 __ enter();
2403
2404 // Save callee-saved registers before using them
2405 __ push(r12);
2406 __ push(r13);
2407 __ push(r14);
2408 __ push(r15);
2409 __ push(rbx);
2410
2411 // arguments
2412 const Register source = c_rarg0; // Source Array
2413 const Register start_offset = c_rarg1; // start offset
2414 const Register end_offset = c_rarg2; // end offset
2415 const Register dest = c_rarg3; // destination array
2416 const Register isMIME = rbx;
2417
2418 #ifndef _WIN64
2419 const Register dp = c_rarg4; // Position for writing to dest array
2420 const Register isURL = c_rarg5;// Base64 or URL character set
2421 __ movl(isMIME, Address(rbp, 2 * wordSize));
2422 #else
2423 const Address dp_mem(rbp, 6 * wordSize); // length is on stack on Win64
2424 const Address isURL_mem(rbp, 7 * wordSize);
2425 const Register isURL = r10; // pick the volatile windows register
2426 const Register dp = r12;
2427 __ movl(dp, dp_mem);
2428 __ movl(isURL, isURL_mem);
2429 __ movl(isMIME, Address(rbp, 8 * wordSize));
2430 #endif
2431
2432 const XMMRegister lookup_lo = xmm5;
2433 const XMMRegister lookup_hi = xmm6;
2434 const XMMRegister errorvec = xmm7;
2435 const XMMRegister pack16_op = xmm9;
2436 const XMMRegister pack32_op = xmm8;
2437 const XMMRegister input0 = xmm3;
2438 const XMMRegister input1 = xmm20;
2439 const XMMRegister input2 = xmm21;
2440 const XMMRegister input3 = xmm19;
2441 const XMMRegister join01 = xmm12;
2442 const XMMRegister join12 = xmm11;
2443 const XMMRegister join23 = xmm10;
2444 const XMMRegister translated0 = xmm2;
2445 const XMMRegister translated1 = xmm1;
2446 const XMMRegister translated2 = xmm0;
2447 const XMMRegister translated3 = xmm4;
2448
2449 const XMMRegister merged0 = xmm2;
2450 const XMMRegister merged1 = xmm1;
2451 const XMMRegister merged2 = xmm0;
2452 const XMMRegister merged3 = xmm4;
2453 const XMMRegister merge_ab_bc0 = xmm2;
2454 const XMMRegister merge_ab_bc1 = xmm1;
2455 const XMMRegister merge_ab_bc2 = xmm0;
2456 const XMMRegister merge_ab_bc3 = xmm4;
2457
2458 const XMMRegister pack24bits = xmm4;
2459
2460 const Register length = r14;
2461 const Register output_size = r13;
2462 const Register output_mask = r15;
2463 const KRegister input_mask = k1;
2464
2465 const XMMRegister input_initial_valid_b64 = xmm0;
2466 const XMMRegister tmp = xmm10;
2467 const XMMRegister mask = xmm0;
2468 const XMMRegister invalid_b64 = xmm1;
2469
2470 Label L_process256, L_process64, L_process64Loop, L_exit, L_processdata, L_loadURL;
2471 Label L_continue, L_finalBit, L_padding, L_donePadding, L_bruteForce;
2472 Label L_forceLoop, L_bottomLoop, L_checkMIME, L_exit_no_vzero, L_lastChunk;
2473
2474 // calculate length from offsets
2475 __ movl(length, end_offset);
2476 __ subl(length, start_offset);
2477 __ push(dest); // Save for return value calc
2478
2479 // If AVX512 VBMI not supported, just compile non-AVX code
2480 if(VM_Version::supports_avx512_vbmi() &&
2481 VM_Version::supports_avx512bw()) {
2482 __ cmpl(length, 31); // 32-bytes is break-even for AVX-512
2483 __ jcc(Assembler::lessEqual, L_lastChunk);
2484
2485 __ cmpl(isMIME, 0);
2486 __ jcc(Assembler::notEqual, L_lastChunk);
2487
2488 // Load lookup tables based on isURL
2489 __ cmpl(isURL, 0);
2490 __ jcc(Assembler::notZero, L_loadURL);
2491
2492 __ evmovdquq(lookup_lo, ExternalAddress(StubRoutines::x86::base64_vbmi_lookup_lo_addr()), Assembler::AVX_512bit, r13);
2493 __ evmovdquq(lookup_hi, ExternalAddress(StubRoutines::x86::base64_vbmi_lookup_hi_addr()), Assembler::AVX_512bit, r13);
2494
2495 __ BIND(L_continue);
2496
2497 __ movl(r15, 0x01400140);
2498 __ evpbroadcastd(pack16_op, r15, Assembler::AVX_512bit);
2499
2500 __ movl(r15, 0x00011000);
2501 __ evpbroadcastd(pack32_op, r15, Assembler::AVX_512bit);
2502
2503 __ cmpl(length, 0xff);
2504 __ jcc(Assembler::lessEqual, L_process64);
2505
2506 // load masks required for decoding data
2507 __ BIND(L_processdata);
2508 __ evmovdquq(join01, ExternalAddress(StubRoutines::x86::base64_vbmi_join_0_1_addr()), Assembler::AVX_512bit,r13);
2509 __ evmovdquq(join12, ExternalAddress(StubRoutines::x86::base64_vbmi_join_1_2_addr()), Assembler::AVX_512bit, r13);
2510 __ evmovdquq(join23, ExternalAddress(StubRoutines::x86::base64_vbmi_join_2_3_addr()), Assembler::AVX_512bit, r13);
2511
2512 __ align32();
2513 __ BIND(L_process256);
2514 // Grab input data
2515 __ evmovdquq(input0, Address(source, start_offset, Address::times_1, 0x00), Assembler::AVX_512bit);
2516 __ evmovdquq(input1, Address(source, start_offset, Address::times_1, 0x40), Assembler::AVX_512bit);
2517 __ evmovdquq(input2, Address(source, start_offset, Address::times_1, 0x80), Assembler::AVX_512bit);
2518 __ evmovdquq(input3, Address(source, start_offset, Address::times_1, 0xc0), Assembler::AVX_512bit);
2519
2520 // Copy the low part of the lookup table into the destination of the permutation
2521 __ evmovdquq(translated0, lookup_lo, Assembler::AVX_512bit);
2522 __ evmovdquq(translated1, lookup_lo, Assembler::AVX_512bit);
2523 __ evmovdquq(translated2, lookup_lo, Assembler::AVX_512bit);
2524 __ evmovdquq(translated3, lookup_lo, Assembler::AVX_512bit);
2525
2526 // Translate the base64 input into "decoded" bytes
2527 __ evpermt2b(translated0, input0, lookup_hi, Assembler::AVX_512bit);
2528 __ evpermt2b(translated1, input1, lookup_hi, Assembler::AVX_512bit);
2529 __ evpermt2b(translated2, input2, lookup_hi, Assembler::AVX_512bit);
2530 __ evpermt2b(translated3, input3, lookup_hi, Assembler::AVX_512bit);
2531
2532 // OR all of the translations together to check for errors (high-order bit of byte set)
2533 __ vpternlogd(input0, 0xfe, input1, input2, Assembler::AVX_512bit);
2534
2535 __ vpternlogd(input3, 0xfe, translated0, translated1, Assembler::AVX_512bit);
2536 __ vpternlogd(input0, 0xfe, translated2, translated3, Assembler::AVX_512bit);
2537 __ vpor(errorvec, input3, input0, Assembler::AVX_512bit);
2538
2539 // Check if there was an error - if so, try 64-byte chunks
2540 __ evpmovb2m(k3, errorvec, Assembler::AVX_512bit);
2541 __ kortestql(k3, k3);
2542 __ jcc(Assembler::notZero, L_process64);
2543
2544 // The merging and shuffling happens here
2545 // We multiply each byte pair [00dddddd | 00cccccc | 00bbbbbb | 00aaaaaa]
2546 // Multiply [00cccccc] by 2^6 added to [00dddddd] to get [0000cccc | ccdddddd]
2547 // The pack16_op is a vector of 0x01400140, so multiply D by 1 and C by 0x40
2548 __ vpmaddubsw(merge_ab_bc0, translated0, pack16_op, Assembler::AVX_512bit);
2549 __ vpmaddubsw(merge_ab_bc1, translated1, pack16_op, Assembler::AVX_512bit);
2550 __ vpmaddubsw(merge_ab_bc2, translated2, pack16_op, Assembler::AVX_512bit);
2551 __ vpmaddubsw(merge_ab_bc3, translated3, pack16_op, Assembler::AVX_512bit);
2552
2553 // Now do the same with packed 16-bit values.
2554 // We start with [0000cccc | ccdddddd | 0000aaaa | aabbbbbb]
2555 // pack32_op is 0x00011000 (2^12, 1), so this multiplies [0000aaaa | aabbbbbb] by 2^12
2556 // and adds [0000cccc | ccdddddd] to yield [00000000 | aaaaaabb | bbbbcccc | ccdddddd]
2557 __ vpmaddwd(merged0, merge_ab_bc0, pack32_op, Assembler::AVX_512bit);
2558 __ vpmaddwd(merged1, merge_ab_bc1, pack32_op, Assembler::AVX_512bit);
2559 __ vpmaddwd(merged2, merge_ab_bc2, pack32_op, Assembler::AVX_512bit);
2560 __ vpmaddwd(merged3, merge_ab_bc3, pack32_op, Assembler::AVX_512bit);
2561
2562 // The join vectors specify which byte from which vector goes into the outputs
2563 // One of every 4 bytes in the extended vector is zero, so we pack them into their
2564 // final positions in the register for storing (256 bytes in, 192 bytes out)
2565 __ evpermt2b(merged0, join01, merged1, Assembler::AVX_512bit);
2566 __ evpermt2b(merged1, join12, merged2, Assembler::AVX_512bit);
2567 __ evpermt2b(merged2, join23, merged3, Assembler::AVX_512bit);
2568
2569 // Store result
2570 __ evmovdquq(Address(dest, dp, Address::times_1, 0x00), merged0, Assembler::AVX_512bit);
2571 __ evmovdquq(Address(dest, dp, Address::times_1, 0x40), merged1, Assembler::AVX_512bit);
2572 __ evmovdquq(Address(dest, dp, Address::times_1, 0x80), merged2, Assembler::AVX_512bit);
2573
2574 __ addptr(source, 0x100);
2575 __ addptr(dest, 0xc0);
2576 __ subl(length, 0x100);
2577 __ cmpl(length, 64 * 4);
2578 __ jcc(Assembler::greaterEqual, L_process256);
2579
2580 // At this point, we've decoded 64 * 4 * n bytes.
2581 // The remaining length will be <= 64 * 4 - 1.
2582 // UNLESS there was an error decoding the first 256-byte chunk. In this
2583 // case, the length will be arbitrarily long.
2584 //
2585 // Note that this will be the path for MIME-encoded strings.
2586
2587 __ BIND(L_process64);
2588
2589 __ evmovdquq(pack24bits, ExternalAddress(StubRoutines::x86::base64_vbmi_pack_vec_addr()), Assembler::AVX_512bit, r13);
2590
2591 __ cmpl(length, 63);
2592 __ jcc(Assembler::lessEqual, L_finalBit);
2593
2594 __ mov64(rax, 0x0000ffffffffffff);
2595 __ kmovql(k2, rax);
2596
2597 __ align32();
2598 __ BIND(L_process64Loop);
2599
2600 // Handle first 64-byte block
2601
2602 __ evmovdquq(input0, Address(source, start_offset), Assembler::AVX_512bit);
2603 __ evmovdquq(translated0, lookup_lo, Assembler::AVX_512bit);
2604 __ evpermt2b(translated0, input0, lookup_hi, Assembler::AVX_512bit);
2605
2606 __ vpor(errorvec, translated0, input0, Assembler::AVX_512bit);
2607
2608 // Check for error and bomb out before updating dest
2609 __ evpmovb2m(k3, errorvec, Assembler::AVX_512bit);
2610 __ kortestql(k3, k3);
2611 __ jcc(Assembler::notZero, L_exit);
2612
2613 // Pack output register, selecting correct byte ordering
2614 __ vpmaddubsw(merge_ab_bc0, translated0, pack16_op, Assembler::AVX_512bit);
2615 __ vpmaddwd(merged0, merge_ab_bc0, pack32_op, Assembler::AVX_512bit);
2616 __ vpermb(merged0, pack24bits, merged0, Assembler::AVX_512bit);
2617
2618 __ evmovdqub(Address(dest, dp), k2, merged0, true, Assembler::AVX_512bit);
2619
2620 __ subl(length, 64);
2621 __ addptr(source, 64);
2622 __ addptr(dest, 48);
2623
2624 __ cmpl(length, 64);
2625 __ jcc(Assembler::greaterEqual, L_process64Loop);
2626
2627 __ cmpl(length, 0);
2628 __ jcc(Assembler::lessEqual, L_exit);
2629
2630 __ BIND(L_finalBit);
2631 // Now have 1 to 63 bytes left to decode
2632
2633 // I was going to let Java take care of the final fragment
2634 // however it will repeatedly call this routine for every 4 bytes
2635 // of input data, so handle the rest here.
2636 __ movq(rax, -1);
2637 __ bzhiq(rax, rax, length); // Input mask in rax
2638
2639 __ movl(output_size, length);
2640 __ shrl(output_size, 2); // Find (len / 4) * 3 (output length)
2641 __ lea(output_size, Address(output_size, output_size, Address::times_2, 0));
2642 // output_size in r13
2643
2644 // Strip pad characters, if any, and adjust length and mask
2645 __ addq(length, start_offset);
2646 __ cmpb(Address(source, length, Address::times_1, -1), '=');
2647 __ jcc(Assembler::equal, L_padding);
2648
2649 __ BIND(L_donePadding);
2650 __ subq(length, start_offset);
2651
2652 // Output size is (64 - output_size), output mask is (all 1s >> output_size).
2653 __ kmovql(input_mask, rax);
2654 __ movq(output_mask, -1);
2655 __ bzhiq(output_mask, output_mask, output_size);
2656
2657 // Load initial input with all valid base64 characters. Will be used
2658 // in merging source bytes to avoid masking when determining if an error occurred.
2659 __ movl(rax, 0x61616161);
2660 __ evpbroadcastd(input_initial_valid_b64, rax, Assembler::AVX_512bit);
2661
2662 // A register containing all invalid base64 decoded values
2663 __ movl(rax, 0x80808080);
2664 __ evpbroadcastd(invalid_b64, rax, Assembler::AVX_512bit);
2665
2666 // input_mask is in k1
2667 // output_size is in r13
2668 // output_mask is in r15
2669 // zmm0 - free
2670 // zmm1 - 0x00011000
2671 // zmm2 - 0x01400140
2672 // zmm3 - errorvec
2673 // zmm4 - pack vector
2674 // zmm5 - lookup_lo
2675 // zmm6 - lookup_hi
2676 // zmm7 - errorvec
2677 // zmm8 - 0x61616161
2678 // zmm9 - 0x80808080
2679
2680 // Load only the bytes from source, merging into our "fully-valid" register
2681 __ evmovdqub(input_initial_valid_b64, input_mask, Address(source, start_offset, Address::times_1, 0x0), true, Assembler::AVX_512bit);
2682
2683 // Decode all bytes within our merged input
2684 __ evmovdquq(tmp, lookup_lo, Assembler::AVX_512bit);
2685 __ evpermt2b(tmp, input_initial_valid_b64, lookup_hi, Assembler::AVX_512bit);
2686 __ evporq(mask, tmp, input_initial_valid_b64, Assembler::AVX_512bit);
2687
2688 // Check for error. Compare (decoded | initial) to all invalid.
2689 // If any bytes have their high-order bit set, then we have an error.
2690 __ evptestmb(k2, mask, invalid_b64, Assembler::AVX_512bit);
2691 __ kortestql(k2, k2);
2692
2693 // If we have an error, use the brute force loop to decode what we can (4-byte chunks).
2694 __ jcc(Assembler::notZero, L_bruteForce);
2695
2696 // Shuffle output bytes
2697 __ vpmaddubsw(tmp, tmp, pack16_op, Assembler::AVX_512bit);
2698 __ vpmaddwd(tmp, tmp, pack32_op, Assembler::AVX_512bit);
2699
2700 __ vpermb(tmp, pack24bits, tmp, Assembler::AVX_512bit);
2701 __ kmovql(k1, output_mask);
2702 __ evmovdqub(Address(dest, dp), k1, tmp, true, Assembler::AVX_512bit);
2703
2704 __ addptr(dest, output_size);
2705
2706 __ BIND(L_exit);
2707 __ vzeroupper();
2708 __ pop(rax); // Get original dest value
2709 __ subptr(dest, rax); // Number of bytes converted
2710 __ movptr(rax, dest);
2711 __ pop(rbx);
2712 __ pop(r15);
2713 __ pop(r14);
2714 __ pop(r13);
2715 __ pop(r12);
2716 __ leave();
2717 __ ret(0);
2718
2719 __ BIND(L_loadURL);
2720 __ evmovdquq(lookup_lo, ExternalAddress(StubRoutines::x86::base64_vbmi_lookup_lo_url_addr()), Assembler::AVX_512bit, r13);
2721 __ evmovdquq(lookup_hi, ExternalAddress(StubRoutines::x86::base64_vbmi_lookup_hi_url_addr()), Assembler::AVX_512bit, r13);
2722 __ jmp(L_continue);
2723
2724 __ BIND(L_padding);
2725 __ decrementq(output_size, 1);
2726 __ shrq(rax, 1);
2727
2728 __ cmpb(Address(source, length, Address::times_1, -2), '=');
2729 __ jcc(Assembler::notEqual, L_donePadding);
2730
2731 __ decrementq(output_size, 1);
2732 __ shrq(rax, 1);
2733 __ jmp(L_donePadding);
2734
2735 __ align32();
2736 __ BIND(L_bruteForce);
2737 } // End of if(avx512_vbmi)
2738
2739 if (VM_Version::supports_avx2()) {
2740 Label L_tailProc, L_topLoop, L_enterLoop;
2741
2742 __ cmpl(isMIME, 0);
2743 __ jcc(Assembler::notEqual, L_lastChunk);
2744
2745 // Check for buffer too small (for algorithm)
2746 __ subl(length, 0x2c);
2747 __ jcc(Assembler::less, L_tailProc);
2748
2749 __ shll(isURL, 2);
2750
2751 // Algorithm adapted from https://arxiv.org/abs/1704.00605, "Faster Base64
2752 // Encoding and Decoding using AVX2 Instructions". URL modifications added.
2753
2754 // Set up constants
2755 __ lea(r13, ExternalAddress(StubRoutines::x86::base64_AVX2_decode_tables_addr()));
2756 __ vpbroadcastd(xmm4, Address(r13, isURL, Address::times_1), Assembler::AVX_256bit); // 2F or 5F
2757 __ vpbroadcastd(xmm10, Address(r13, isURL, Address::times_1, 0x08), Assembler::AVX_256bit); // -1 or -4
2758 __ vmovdqu(xmm12, Address(r13, 0x10)); // permute
2759 __ vmovdqu(xmm13, Address(r13, 0x30)); // shuffle
2760 __ vpbroadcastd(xmm7, Address(r13, 0x50), Assembler::AVX_256bit); // merge
2761 __ vpbroadcastd(xmm6, Address(r13, 0x54), Assembler::AVX_256bit); // merge mult
2762
2763 __ lea(r13, ExternalAddress(StubRoutines::x86::base64_AVX2_decode_LUT_tables_addr()));
2764 __ shll(isURL, 4);
2765 __ vmovdqu(xmm11, Address(r13, isURL, Address::times_1, 0x00)); // lut_lo
2766 __ vmovdqu(xmm8, Address(r13, isURL, Address::times_1, 0x20)); // lut_roll
2767 __ shrl(isURL, 6); // restore isURL
2768 __ vmovdqu(xmm9, Address(r13, 0x80)); // lut_hi
2769 __ jmp(L_enterLoop);
2770
2771 __ align32();
2772 __ bind(L_topLoop);
2773 // Add in the offset value (roll) to get 6-bit out values
2774 __ vpaddb(xmm0, xmm0, xmm2, Assembler::AVX_256bit);
2775 // Merge and permute the output bits into appropriate output byte lanes
2776 __ vpmaddubsw(xmm0, xmm0, xmm7, Assembler::AVX_256bit);
2777 __ vpmaddwd(xmm0, xmm0, xmm6, Assembler::AVX_256bit);
2778 __ vpshufb(xmm0, xmm0, xmm13, Assembler::AVX_256bit);
2779 __ vpermd(xmm0, xmm12, xmm0, Assembler::AVX_256bit);
2780 // Store the output bytes
2781 __ vmovdqu(Address(dest, dp, Address::times_1, 0), xmm0);
2782 __ addptr(source, 0x20);
2783 __ addptr(dest, 0x18);
2784 __ subl(length, 0x20);
2785 __ jcc(Assembler::less, L_tailProc);
2786
2787 __ bind(L_enterLoop);
2788
2789 // Load in encoded string (32 bytes)
2790 __ vmovdqu(xmm2, Address(source, start_offset, Address::times_1, 0x0));
2791 // Extract the high nibble for indexing into the lut tables. High 4 bits are don't care.
2792 __ vpsrld(xmm1, xmm2, 0x4, Assembler::AVX_256bit);
2793 __ vpand(xmm1, xmm4, xmm1, Assembler::AVX_256bit);
2794 // Extract the low nibble. 5F/2F will isolate the low-order 4 bits. High 4 bits are don't care.
2795 __ vpand(xmm3, xmm2, xmm4, Assembler::AVX_256bit);
2796 // Check for special-case (0x2F or 0x5F (URL))
2797 __ vpcmpeqb(xmm0, xmm4, xmm2, Assembler::AVX_256bit);
2798 // Get the bitset based on the low nibble. vpshufb uses low-order 4 bits only.
2799 __ vpshufb(xmm3, xmm11, xmm3, Assembler::AVX_256bit);
2800 // Get the bit value of the high nibble
2801 __ vpshufb(xmm5, xmm9, xmm1, Assembler::AVX_256bit);
2802 // Make sure 2F / 5F shows as valid
2803 __ vpandn(xmm3, xmm0, xmm3, Assembler::AVX_256bit);
2804 // Make adjustment for roll index. For non-URL, this is a no-op,
2805 // for URL, this adjusts by -4. This is to properly index the
2806 // roll value for 2F / 5F.
2807 __ vpand(xmm0, xmm0, xmm10, Assembler::AVX_256bit);
2808 // If the and of the two is non-zero, we have an invalid input character
2809 __ vptest(xmm3, xmm5);
2810 // Extract the "roll" value - value to add to the input to get 6-bit out value
2811 __ vpaddb(xmm0, xmm0, xmm1, Assembler::AVX_256bit); // Handle 2F / 5F
2812 __ vpshufb(xmm0, xmm8, xmm0, Assembler::AVX_256bit);
2813 __ jcc(Assembler::equal, L_topLoop); // Fall through on error
2814
2815 __ bind(L_tailProc);
2816
2817 __ addl(length, 0x2c);
2818
2819 __ vzeroupper();
2820 }
2821
2822 // Use non-AVX code to decode 4-byte chunks into 3 bytes of output
2823
2824 // Register state (Linux):
2825 // r12-15 - saved on stack
2826 // rdi - src
2827 // rsi - sp
2828 // rdx - sl
2829 // rcx - dst
2830 // r8 - dp
2831 // r9 - isURL
2832
2833 // Register state (Windows):
2834 // r12-15 - saved on stack
2835 // rcx - src
2836 // rdx - sp
2837 // r8 - sl
2838 // r9 - dst
2839 // r12 - dp
2840 // r10 - isURL
2841
2842 // Registers (common):
2843 // length (r14) - bytes in src
2844
2845 const Register decode_table = r11;
2846 const Register out_byte_count = rbx;
2847 const Register byte1 = r13;
2848 const Register byte2 = r15;
2849 const Register byte3 = WIN64_ONLY(r8) NOT_WIN64(rdx);
2850 const Register byte4 = WIN64_ONLY(r10) NOT_WIN64(r9);
2851
2852 __ bind(L_lastChunk);
2853
2854 __ shrl(length, 2); // Multiple of 4 bytes only - length is # 4-byte chunks
2855 __ cmpl(length, 0);
2856 __ jcc(Assembler::lessEqual, L_exit_no_vzero);
2857
2858 __ shll(isURL, 8); // index into decode table based on isURL
2859 __ lea(decode_table, ExternalAddress(StubRoutines::x86::base64_decoding_table_addr()));
2860 __ addptr(decode_table, isURL);
2861
2862 __ jmp(L_bottomLoop);
2863
2864 __ align32();
2865 __ BIND(L_forceLoop);
2866 __ shll(byte1, 18);
2867 __ shll(byte2, 12);
2868 __ shll(byte3, 6);
2869 __ orl(byte1, byte2);
2870 __ orl(byte1, byte3);
2871 __ orl(byte1, byte4);
2872
2873 __ addptr(source, 4);
2874
2875 __ movb(Address(dest, dp, Address::times_1, 2), byte1);
2876 __ shrl(byte1, 8);
2877 __ movb(Address(dest, dp, Address::times_1, 1), byte1);
2878 __ shrl(byte1, 8);
2879 __ movb(Address(dest, dp, Address::times_1, 0), byte1);
2880
2881 __ addptr(dest, 3);
2882 __ decrementl(length, 1);
2883 __ jcc(Assembler::zero, L_exit_no_vzero);
2884
2885 __ BIND(L_bottomLoop);
2886 __ load_unsigned_byte(byte1, Address(source, start_offset, Address::times_1, 0x00));
2887 __ load_unsigned_byte(byte2, Address(source, start_offset, Address::times_1, 0x01));
2888 __ load_signed_byte(byte1, Address(decode_table, byte1));
2889 __ load_signed_byte(byte2, Address(decode_table, byte2));
2890 __ load_unsigned_byte(byte3, Address(source, start_offset, Address::times_1, 0x02));
2891 __ load_unsigned_byte(byte4, Address(source, start_offset, Address::times_1, 0x03));
2892 __ load_signed_byte(byte3, Address(decode_table, byte3));
2893 __ load_signed_byte(byte4, Address(decode_table, byte4));
2894
2895 __ mov(rax, byte1);
2896 __ orl(rax, byte2);
2897 __ orl(rax, byte3);
2898 __ orl(rax, byte4);
2899 __ jcc(Assembler::positive, L_forceLoop);
2900
2901 __ BIND(L_exit_no_vzero);
2902 __ pop(rax); // Get original dest value
2903 __ subptr(dest, rax); // Number of bytes converted
2904 __ movptr(rax, dest);
2905 __ pop(rbx);
2906 __ pop(r15);
2907 __ pop(r14);
2908 __ pop(r13);
2909 __ pop(r12);
2910 __ leave();
2911 __ ret(0);
2912
2913 return start;
2914 }
2915
2916
2917 /**
2918 * Arguments:
2919 *
2920 * Inputs:
2921 * c_rarg0 - int crc
2922 * c_rarg1 - byte* buf
2923 * c_rarg2 - int length
2924 *
2925 * Output:
2926 * rax - int crc result
2927 */
2928 address StubGenerator::generate_updateBytesCRC32() {
2929 assert(UseCRC32Intrinsics, "need AVX and CLMUL instructions");
2930
2931 __ align(CodeEntryAlignment);
2932 StubCodeMark mark(this, "StubRoutines", "updateBytesCRC32");
2933
2934 address start = __ pc();
2935
2936 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
2937 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
2938 // rscratch1: r10
2939 const Register crc = c_rarg0; // crc
2940 const Register buf = c_rarg1; // source java byte array address
2941 const Register len = c_rarg2; // length
2942 const Register table = c_rarg3; // crc_table address (reuse register)
2943 const Register tmp1 = r11;
2944 const Register tmp2 = r10;
2945 assert_different_registers(crc, buf, len, table, tmp1, tmp2, rax);
2946
2947 BLOCK_COMMENT("Entry:");
2948 __ enter(); // required for proper stackwalking of RuntimeStub frame
2949
2950 if (VM_Version::supports_sse4_1() && VM_Version::supports_avx512_vpclmulqdq() &&
2951 VM_Version::supports_avx512bw() &&
2952 VM_Version::supports_avx512vl()) {
2953 // The constants used in the CRC32 algorithm requires the 1's compliment of the initial crc value.
2954 // However, the constant table for CRC32-C assumes the original crc value. Account for this
2955 // difference before calling and after returning.
2956 __ lea(table, ExternalAddress(StubRoutines::x86::crc_table_avx512_addr()));
2957 __ notl(crc);
2958 __ kernel_crc32_avx512(crc, buf, len, table, tmp1, tmp2);
2959 __ notl(crc);
2960 } else {
2961 __ kernel_crc32(crc, buf, len, table, tmp1);
2962 }
2963
2964 __ movl(rax, crc);
2965 __ vzeroupper();
2966 __ leave(); // required for proper stackwalking of RuntimeStub frame
2967 __ ret(0);
2968
2969 return start;
2970 }
2971
2972 /**
2973 * Arguments:
2974 *
2975 * Inputs:
2976 * c_rarg0 - int crc
2977 * c_rarg1 - byte* buf
2978 * c_rarg2 - long length
2979 * c_rarg3 - table_start - optional (present only when doing a library_call,
2980 * not used by x86 algorithm)
2981 *
2982 * Output:
2983 * rax - int crc result
2984 */
2985 address StubGenerator::generate_updateBytesCRC32C(bool is_pclmulqdq_supported) {
2986 assert(UseCRC32CIntrinsics, "need SSE4_2");
2987 __ align(CodeEntryAlignment);
2988 StubCodeMark mark(this, "StubRoutines", "updateBytesCRC32C");
2989 address start = __ pc();
2990
2991 //reg.arg int#0 int#1 int#2 int#3 int#4 int#5 float regs
2992 //Windows RCX RDX R8 R9 none none XMM0..XMM3
2993 //Lin / Sol RDI RSI RDX RCX R8 R9 XMM0..XMM7
2994 const Register crc = c_rarg0; // crc
2995 const Register buf = c_rarg1; // source java byte array address
2996 const Register len = c_rarg2; // length
2997 const Register a = rax;
2998 const Register j = r9;
2999 const Register k = r10;
3000 const Register l = r11;
3001 #ifdef _WIN64
3002 const Register y = rdi;
3003 const Register z = rsi;
3004 #else
3005 const Register y = rcx;
3006 const Register z = r8;
3007 #endif
3008 assert_different_registers(crc, buf, len, a, j, k, l, y, z);
3009
3010 BLOCK_COMMENT("Entry:");
3011 __ enter(); // required for proper stackwalking of RuntimeStub frame
3012 Label L_continue;
3013
3014 if (VM_Version::supports_sse4_1() && VM_Version::supports_avx512_vpclmulqdq() &&
3015 VM_Version::supports_avx512bw() &&
3016 VM_Version::supports_avx512vl()) {
3017 Label L_doSmall;
3018
3019 __ cmpl(len, 384);
3020 __ jcc(Assembler::lessEqual, L_doSmall);
3021
3022 __ lea(j, ExternalAddress(StubRoutines::x86::crc32c_table_avx512_addr()));
3023 __ kernel_crc32_avx512(crc, buf, len, j, l, k);
3024
3025 __ jmp(L_continue);
3026
3027 __ bind(L_doSmall);
3028 }
3029 #ifdef _WIN64
3030 __ push(y);
3031 __ push(z);
3032 #endif
3033 __ crc32c_ipl_alg2_alt2(crc, buf, len,
3034 a, j, k,
3035 l, y, z,
3036 c_farg0, c_farg1, c_farg2,
3037 is_pclmulqdq_supported);
3038 #ifdef _WIN64
3039 __ pop(z);
3040 __ pop(y);
3041 #endif
3042
3043 __ bind(L_continue);
3044 __ movl(rax, crc);
3045 __ vzeroupper();
3046 __ leave(); // required for proper stackwalking of RuntimeStub frame
3047 __ ret(0);
3048
3049 return start;
3050 }
3051
3052
3053 /**
3054 * Arguments:
3055 *
3056 * Input:
3057 * c_rarg0 - x address
3058 * c_rarg1 - x length
3059 * c_rarg2 - y address
3060 * c_rarg3 - y length
3061 * not Win64
3062 * c_rarg4 - z address
3063 * Win64
3064 * rsp+40 - z address
3065 */
3066 address StubGenerator::generate_multiplyToLen() {
3067 __ align(CodeEntryAlignment);
3068 StubCodeMark mark(this, "StubRoutines", "multiplyToLen");
3069 address start = __ pc();
3070
3071 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
3072 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
3073 const Register x = rdi;
3074 const Register xlen = rax;
3075 const Register y = rsi;
3076 const Register ylen = rcx;
3077 const Register z = r8;
3078
3079 // Next registers will be saved on stack in multiply_to_len().
3080 const Register tmp0 = r11;
3081 const Register tmp1 = r12;
3082 const Register tmp2 = r13;
3083 const Register tmp3 = r14;
3084 const Register tmp4 = r15;
3085 const Register tmp5 = rbx;
3086
3087 BLOCK_COMMENT("Entry:");
3088 __ enter(); // required for proper stackwalking of RuntimeStub frame
3089
3090 setup_arg_regs(4); // x => rdi, xlen => rsi, y => rdx
3091 // ylen => rcx, z => r8
3092 // r9 and r10 may be used to save non-volatile registers
3093 #ifdef _WIN64
3094 // last argument (#4) is on stack on Win64
3095 __ movptr(z, Address(rsp, 6 * wordSize));
3096 #endif
3097
3098 __ movptr(xlen, rsi);
3099 __ movptr(y, rdx);
3100 __ multiply_to_len(x, xlen, y, ylen, z, tmp0, tmp1, tmp2, tmp3, tmp4, tmp5);
3101
3102 restore_arg_regs();
3103
3104 __ leave(); // required for proper stackwalking of RuntimeStub frame
3105 __ ret(0);
3106
3107 return start;
3108 }
3109
3110 /**
3111 * Arguments:
3112 *
3113 * Input:
3114 * c_rarg0 - obja address
3115 * c_rarg1 - objb address
3116 * c_rarg3 - length length
3117 * c_rarg4 - scale log2_array_indxscale
3118 *
3119 * Output:
3120 * rax - int >= mismatched index, < 0 bitwise complement of tail
3121 */
3122 address StubGenerator::generate_vectorizedMismatch() {
3123 __ align(CodeEntryAlignment);
3124 StubCodeMark mark(this, "StubRoutines", "vectorizedMismatch");
3125 address start = __ pc();
3126
3127 BLOCK_COMMENT("Entry:");
3128 __ enter();
3129
3130 #ifdef _WIN64 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
3131 const Register scale = c_rarg0; //rcx, will exchange with r9
3132 const Register objb = c_rarg1; //rdx
3133 const Register length = c_rarg2; //r8
3134 const Register obja = c_rarg3; //r9
3135 __ xchgq(obja, scale); //now obja and scale contains the correct contents
3136
3137 const Register tmp1 = r10;
3138 const Register tmp2 = r11;
3139 #endif
3140 #ifndef _WIN64 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
3141 const Register obja = c_rarg0; //U:rdi
3142 const Register objb = c_rarg1; //U:rsi
3143 const Register length = c_rarg2; //U:rdx
3144 const Register scale = c_rarg3; //U:rcx
3145 const Register tmp1 = r8;
3146 const Register tmp2 = r9;
3147 #endif
3148 const Register result = rax; //return value
3149 const XMMRegister vec0 = xmm0;
3150 const XMMRegister vec1 = xmm1;
3151 const XMMRegister vec2 = xmm2;
3152
3153 __ vectorized_mismatch(obja, objb, length, scale, result, tmp1, tmp2, vec0, vec1, vec2);
3154
3155 __ vzeroupper();
3156 __ leave();
3157 __ ret(0);
3158
3159 return start;
3160 }
3161
3162 /**
3163 * Arguments:
3164 *
3165 // Input:
3166 // c_rarg0 - x address
3167 // c_rarg1 - x length
3168 // c_rarg2 - z address
3169 // c_rarg3 - z length
3170 *
3171 */
3172 address StubGenerator::generate_squareToLen() {
3173
3174 __ align(CodeEntryAlignment);
3175 StubCodeMark mark(this, "StubRoutines", "squareToLen");
3176 address start = __ pc();
3177
3178 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
3179 // Unix: rdi, rsi, rdx, rcx (c_rarg0, c_rarg1, ...)
3180 const Register x = rdi;
3181 const Register len = rsi;
3182 const Register z = r8;
3183 const Register zlen = rcx;
3184
3185 const Register tmp1 = r12;
3186 const Register tmp2 = r13;
3187 const Register tmp3 = r14;
3188 const Register tmp4 = r15;
3189 const Register tmp5 = rbx;
3190
3191 BLOCK_COMMENT("Entry:");
3192 __ enter(); // required for proper stackwalking of RuntimeStub frame
3193
3194 setup_arg_regs(4); // x => rdi, len => rsi, z => rdx
3195 // zlen => rcx
3196 // r9 and r10 may be used to save non-volatile registers
3197 __ movptr(r8, rdx);
3198 __ square_to_len(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx, rax);
3199
3200 restore_arg_regs();
3201
3202 __ leave(); // required for proper stackwalking of RuntimeStub frame
3203 __ ret(0);
3204
3205 return start;
3206 }
3207
3208 address StubGenerator::generate_method_entry_barrier() {
3209 __ align(CodeEntryAlignment);
3210 StubCodeMark mark(this, "StubRoutines", "nmethod_entry_barrier");
3211 address start = __ pc();
3212
3213 Label deoptimize_label;
3214
3215 __ push(-1); // cookie, this is used for writing the new rsp when deoptimizing
3216
3217 BLOCK_COMMENT("Entry:");
3218 __ enter(); // save rbp
3219
3220 // save c_rarg0, because we want to use that value.
3221 // We could do without it but then we depend on the number of slots used by pusha
3222 __ push(c_rarg0);
3223
3224 __ lea(c_rarg0, Address(rsp, wordSize * 3)); // 1 for cookie, 1 for rbp, 1 for c_rarg0 - this should be the return address
3225
3226 __ pusha();
3227
3228 // The method may have floats as arguments, and we must spill them before calling
3229 // the VM runtime.
3230 assert(Argument::n_float_register_parameters_j == 8, "Assumption");
3231 const int xmm_size = wordSize * 2;
3232 const int xmm_spill_size = xmm_size * Argument::n_float_register_parameters_j;
3233 __ subptr(rsp, xmm_spill_size);
3234 __ movdqu(Address(rsp, xmm_size * 7), xmm7);
3235 __ movdqu(Address(rsp, xmm_size * 6), xmm6);
3236 __ movdqu(Address(rsp, xmm_size * 5), xmm5);
3237 __ movdqu(Address(rsp, xmm_size * 4), xmm4);
3238 __ movdqu(Address(rsp, xmm_size * 3), xmm3);
3239 __ movdqu(Address(rsp, xmm_size * 2), xmm2);
3240 __ movdqu(Address(rsp, xmm_size * 1), xmm1);
3241 __ movdqu(Address(rsp, xmm_size * 0), xmm0);
3242
3243 __ call_VM_leaf(CAST_FROM_FN_PTR(address, static_cast<int (*)(address*)>(BarrierSetNMethod::nmethod_stub_entry_barrier)), 1);
3244
3245 __ movdqu(xmm0, Address(rsp, xmm_size * 0));
3246 __ movdqu(xmm1, Address(rsp, xmm_size * 1));
3247 __ movdqu(xmm2, Address(rsp, xmm_size * 2));
3248 __ movdqu(xmm3, Address(rsp, xmm_size * 3));
3249 __ movdqu(xmm4, Address(rsp, xmm_size * 4));
3250 __ movdqu(xmm5, Address(rsp, xmm_size * 5));
3251 __ movdqu(xmm6, Address(rsp, xmm_size * 6));
3252 __ movdqu(xmm7, Address(rsp, xmm_size * 7));
3253 __ addptr(rsp, xmm_spill_size);
3254
3255 __ cmpl(rax, 1); // 1 means deoptimize
3256 __ jcc(Assembler::equal, deoptimize_label);
3257
3258 __ popa();
3259 __ pop(c_rarg0);
3260
3261 __ leave();
3262
3263 __ addptr(rsp, 1 * wordSize); // cookie
3264 __ ret(0);
3265
3266
3267 __ BIND(deoptimize_label);
3268
3269 __ popa();
3270 __ pop(c_rarg0);
3271
3272 __ leave();
3273
3274 // this can be taken out, but is good for verification purposes. getting a SIGSEGV
3275 // here while still having a correct stack is valuable
3276 __ testptr(rsp, Address(rsp, 0));
3277
3278 __ movptr(rsp, Address(rsp, 0)); // new rsp was written in the barrier
3279 __ jmp(Address(rsp, -1 * wordSize)); // jmp target should be callers verified_entry_point
3280
3281 return start;
3282 }
3283
3284 /**
3285 * Arguments:
3286 *
3287 * Input:
3288 * c_rarg0 - out address
3289 * c_rarg1 - in address
3290 * c_rarg2 - offset
3291 * c_rarg3 - len
3292 * not Win64
3293 * c_rarg4 - k
3294 * Win64
3295 * rsp+40 - k
3296 */
3297 address StubGenerator::generate_mulAdd() {
3298 __ align(CodeEntryAlignment);
3299 StubCodeMark mark(this, "StubRoutines", "mulAdd");
3300 address start = __ pc();
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 return start;
3336 }
3337
3338 address StubGenerator::generate_bigIntegerRightShift() {
3339 __ align(CodeEntryAlignment);
3340 StubCodeMark mark(this, "StubRoutines", "bigIntegerRightShiftWorker");
3341 address start = __ pc();
3342
3343 Label Shift512Loop, ShiftTwo, ShiftTwoLoop, ShiftOne, Exit;
3344 // For Unix, the arguments are as follows: rdi, rsi, rdx, rcx, r8.
3345 const Register newArr = rdi;
3346 const Register oldArr = rsi;
3347 const Register newIdx = rdx;
3348 const Register shiftCount = rcx; // It was intentional to have shiftCount in rcx since it is used implicitly for shift.
3349 const Register totalNumIter = r8;
3350
3351 // For windows, we use r9 and r10 as temps to save rdi and rsi. Thus we cannot allocate them for our temps.
3352 // For everything else, we prefer using r9 and r10 since we do not have to save them before use.
3353 const Register tmp1 = r11; // Caller save.
3354 const Register tmp2 = rax; // Caller save.
3355 const Register tmp3 = WIN64_ONLY(r12) NOT_WIN64(r9); // Windows: Callee save. Linux: Caller save.
3356 const Register tmp4 = WIN64_ONLY(r13) NOT_WIN64(r10); // Windows: Callee save. Linux: Caller save.
3357 const Register tmp5 = r14; // Callee save.
3358 const Register tmp6 = r15;
3359
3360 const XMMRegister x0 = xmm0;
3361 const XMMRegister x1 = xmm1;
3362 const XMMRegister x2 = xmm2;
3363
3364 BLOCK_COMMENT("Entry:");
3365 __ enter(); // required for proper stackwalking of RuntimeStub frame
3366
3367 #ifdef _WIN64
3368 setup_arg_regs(4);
3369 // For windows, since last argument is on stack, we need to move it to the appropriate register.
3370 __ movl(totalNumIter, Address(rsp, 6 * wordSize));
3371 // Save callee save registers.
3372 __ push(tmp3);
3373 __ push(tmp4);
3374 #endif
3375 __ push(tmp5);
3376
3377 // Rename temps used throughout the code.
3378 const Register idx = tmp1;
3379 const Register nIdx = tmp2;
3380
3381 __ xorl(idx, idx);
3382
3383 // Start right shift from end of the array.
3384 // For example, if #iteration = 4 and newIdx = 1
3385 // then dest[4] = src[4] >> shiftCount | src[3] <<< (shiftCount - 32)
3386 // if #iteration = 4 and newIdx = 0
3387 // then dest[3] = src[4] >> shiftCount | src[3] <<< (shiftCount - 32)
3388 __ movl(idx, totalNumIter);
3389 __ movl(nIdx, idx);
3390 __ addl(nIdx, newIdx);
3391
3392 // If vectorization is enabled, check if the number of iterations is at least 64
3393 // If not, then go to ShifTwo processing 2 iterations
3394 if (VM_Version::supports_avx512_vbmi2()) {
3395 __ cmpptr(totalNumIter, (AVX3Threshold/64));
3396 __ jcc(Assembler::less, ShiftTwo);
3397
3398 if (AVX3Threshold < 16 * 64) {
3399 __ cmpl(totalNumIter, 16);
3400 __ jcc(Assembler::less, ShiftTwo);
3401 }
3402 __ evpbroadcastd(x0, shiftCount, Assembler::AVX_512bit);
3403 __ subl(idx, 16);
3404 __ subl(nIdx, 16);
3405 __ BIND(Shift512Loop);
3406 __ evmovdqul(x2, Address(oldArr, idx, Address::times_4, 4), Assembler::AVX_512bit);
3407 __ evmovdqul(x1, Address(oldArr, idx, Address::times_4), Assembler::AVX_512bit);
3408 __ vpshrdvd(x2, x1, x0, Assembler::AVX_512bit);
3409 __ evmovdqul(Address(newArr, nIdx, Address::times_4), x2, Assembler::AVX_512bit);
3410 __ subl(nIdx, 16);
3411 __ subl(idx, 16);
3412 __ jcc(Assembler::greaterEqual, Shift512Loop);
3413 __ addl(idx, 16);
3414 __ addl(nIdx, 16);
3415 }
3416 __ BIND(ShiftTwo);
3417 __ cmpl(idx, 2);
3418 __ jcc(Assembler::less, ShiftOne);
3419 __ subl(idx, 2);
3420 __ subl(nIdx, 2);
3421 __ BIND(ShiftTwoLoop);
3422 __ movl(tmp5, Address(oldArr, idx, Address::times_4, 8));
3423 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 4));
3424 __ movl(tmp3, Address(oldArr, idx, Address::times_4));
3425 __ shrdl(tmp5, tmp4);
3426 __ shrdl(tmp4, tmp3);
3427 __ movl(Address(newArr, nIdx, Address::times_4, 4), tmp5);
3428 __ movl(Address(newArr, nIdx, Address::times_4), tmp4);
3429 __ subl(nIdx, 2);
3430 __ subl(idx, 2);
3431 __ jcc(Assembler::greaterEqual, ShiftTwoLoop);
3432 __ addl(idx, 2);
3433 __ addl(nIdx, 2);
3434
3435 // Do the last iteration
3436 __ BIND(ShiftOne);
3437 __ cmpl(idx, 1);
3438 __ jcc(Assembler::less, Exit);
3439 __ subl(idx, 1);
3440 __ subl(nIdx, 1);
3441 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 4));
3442 __ movl(tmp3, Address(oldArr, idx, Address::times_4));
3443 __ shrdl(tmp4, tmp3);
3444 __ movl(Address(newArr, nIdx, Address::times_4), tmp4);
3445 __ BIND(Exit);
3446 __ vzeroupper();
3447 // Restore callee save registers.
3448 __ pop(tmp5);
3449 #ifdef _WIN64
3450 __ pop(tmp4);
3451 __ pop(tmp3);
3452 restore_arg_regs();
3453 #endif
3454 __ leave(); // required for proper stackwalking of RuntimeStub frame
3455 __ ret(0);
3456
3457 return start;
3458 }
3459
3460 /**
3461 * Arguments:
3462 *
3463 * Input:
3464 * c_rarg0 - newArr address
3465 * c_rarg1 - oldArr address
3466 * c_rarg2 - newIdx
3467 * c_rarg3 - shiftCount
3468 * not Win64
3469 * c_rarg4 - numIter
3470 * Win64
3471 * rsp40 - numIter
3472 */
3473 address StubGenerator::generate_bigIntegerLeftShift() {
3474 __ align(CodeEntryAlignment);
3475 StubCodeMark mark(this, "StubRoutines", "bigIntegerLeftShiftWorker");
3476 address start = __ pc();
3477
3478 Label Shift512Loop, ShiftTwo, ShiftTwoLoop, ShiftOne, Exit;
3479 // For Unix, the arguments are as follows: rdi, rsi, rdx, rcx, r8.
3480 const Register newArr = rdi;
3481 const Register oldArr = rsi;
3482 const Register newIdx = rdx;
3483 const Register shiftCount = rcx; // It was intentional to have shiftCount in rcx since it is used implicitly for shift.
3484 const Register totalNumIter = r8;
3485 // For windows, we use r9 and r10 as temps to save rdi and rsi. Thus we cannot allocate them for our temps.
3486 // For everything else, we prefer using r9 and r10 since we do not have to save them before use.
3487 const Register tmp1 = r11; // Caller save.
3488 const Register tmp2 = rax; // Caller save.
3489 const Register tmp3 = WIN64_ONLY(r12) NOT_WIN64(r9); // Windows: Callee save. Linux: Caller save.
3490 const Register tmp4 = WIN64_ONLY(r13) NOT_WIN64(r10); // Windows: Callee save. Linux: Caller save.
3491 const Register tmp5 = r14; // Callee save.
3492
3493 const XMMRegister x0 = xmm0;
3494 const XMMRegister x1 = xmm1;
3495 const XMMRegister x2 = xmm2;
3496 BLOCK_COMMENT("Entry:");
3497 __ enter(); // required for proper stackwalking of RuntimeStub frame
3498
3499 #ifdef _WIN64
3500 setup_arg_regs(4);
3501 // For windows, since last argument is on stack, we need to move it to the appropriate register.
3502 __ movl(totalNumIter, Address(rsp, 6 * wordSize));
3503 // Save callee save registers.
3504 __ push(tmp3);
3505 __ push(tmp4);
3506 #endif
3507 __ push(tmp5);
3508
3509 // Rename temps used throughout the code
3510 const Register idx = tmp1;
3511 const Register numIterTmp = tmp2;
3512
3513 // Start idx from zero.
3514 __ xorl(idx, idx);
3515 // Compute interior pointer for new array. We do this so that we can use same index for both old and new arrays.
3516 __ lea(newArr, Address(newArr, newIdx, Address::times_4));
3517 __ movl(numIterTmp, totalNumIter);
3518
3519 // If vectorization is enabled, check if the number of iterations is at least 64
3520 // If not, then go to ShiftTwo shifting two numbers at a time
3521 if (VM_Version::supports_avx512_vbmi2()) {
3522 __ cmpl(totalNumIter, (AVX3Threshold/64));
3523 __ jcc(Assembler::less, ShiftTwo);
3524
3525 if (AVX3Threshold < 16 * 64) {
3526 __ cmpl(totalNumIter, 16);
3527 __ jcc(Assembler::less, ShiftTwo);
3528 }
3529 __ evpbroadcastd(x0, shiftCount, Assembler::AVX_512bit);
3530 __ subl(numIterTmp, 16);
3531 __ BIND(Shift512Loop);
3532 __ evmovdqul(x1, Address(oldArr, idx, Address::times_4), Assembler::AVX_512bit);
3533 __ evmovdqul(x2, Address(oldArr, idx, Address::times_4, 0x4), Assembler::AVX_512bit);
3534 __ vpshldvd(x1, x2, x0, Assembler::AVX_512bit);
3535 __ evmovdqul(Address(newArr, idx, Address::times_4), x1, Assembler::AVX_512bit);
3536 __ addl(idx, 16);
3537 __ subl(numIterTmp, 16);
3538 __ jcc(Assembler::greaterEqual, Shift512Loop);
3539 __ addl(numIterTmp, 16);
3540 }
3541 __ BIND(ShiftTwo);
3542 __ cmpl(totalNumIter, 1);
3543 __ jcc(Assembler::less, Exit);
3544 __ movl(tmp3, Address(oldArr, idx, Address::times_4));
3545 __ subl(numIterTmp, 2);
3546 __ jcc(Assembler::less, ShiftOne);
3547
3548 __ BIND(ShiftTwoLoop);
3549 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 0x4));
3550 __ movl(tmp5, Address(oldArr, idx, Address::times_4, 0x8));
3551 __ shldl(tmp3, tmp4);
3552 __ shldl(tmp4, tmp5);
3553 __ movl(Address(newArr, idx, Address::times_4), tmp3);
3554 __ movl(Address(newArr, idx, Address::times_4, 0x4), tmp4);
3555 __ movl(tmp3, tmp5);
3556 __ addl(idx, 2);
3557 __ subl(numIterTmp, 2);
3558 __ jcc(Assembler::greaterEqual, ShiftTwoLoop);
3559
3560 // Do the last iteration
3561 __ BIND(ShiftOne);
3562 __ addl(numIterTmp, 2);
3563 __ cmpl(numIterTmp, 1);
3564 __ jcc(Assembler::less, Exit);
3565 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 0x4));
3566 __ shldl(tmp3, tmp4);
3567 __ movl(Address(newArr, idx, Address::times_4), tmp3);
3568
3569 __ BIND(Exit);
3570 __ vzeroupper();
3571 // Restore callee save registers.
3572 __ pop(tmp5);
3573 #ifdef _WIN64
3574 __ pop(tmp4);
3575 __ pop(tmp3);
3576 restore_arg_regs();
3577 #endif
3578 __ leave(); // required for proper stackwalking of RuntimeStub frame
3579 __ ret(0);
3580
3581 return start;
3582 }
3583
3584 void StubGenerator::generate_libm_stubs() {
3585 if (UseLibmIntrinsic && InlineIntrinsics) {
3586 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dsin)) {
3587 StubRoutines::_dsin = generate_libmSin(); // from stubGenerator_x86_64_sin.cpp
3588 }
3589 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dcos)) {
3590 StubRoutines::_dcos = generate_libmCos(); // from stubGenerator_x86_64_cos.cpp
3591 }
3592 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dtan)) {
3593 StubRoutines::_dtan = generate_libmTan(); // from stubGenerator_x86_64_tan.cpp
3594 }
3595 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dtanh)) {
3596 StubRoutines::_dtanh = generate_libmTanh(); // from stubGenerator_x86_64_tanh.cpp
3597 }
3598 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dexp)) {
3599 StubRoutines::_dexp = generate_libmExp(); // from stubGenerator_x86_64_exp.cpp
3600 }
3601 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dpow)) {
3602 StubRoutines::_dpow = generate_libmPow(); // from stubGenerator_x86_64_pow.cpp
3603 }
3604 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dlog)) {
3605 StubRoutines::_dlog = generate_libmLog(); // from stubGenerator_x86_64_log.cpp
3606 }
3607 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dlog10)) {
3608 StubRoutines::_dlog10 = generate_libmLog10(); // from stubGenerator_x86_64_log.cpp
3609 }
3610 }
3611 }
3612
3613 /**
3614 * Arguments:
3615 *
3616 * Input:
3617 * c_rarg0 - float16 jshort
3618 *
3619 * Output:
3620 * xmm0 - float
3621 */
3622 address StubGenerator::generate_float16ToFloat() {
3623 StubCodeMark mark(this, "StubRoutines", "float16ToFloat");
3624
3625 address start = __ pc();
3626
3627 BLOCK_COMMENT("Entry:");
3628 // No need for RuntimeStub frame since it is called only during JIT compilation
3629
3630 // Load value into xmm0 and convert
3631 __ flt16_to_flt(xmm0, c_rarg0);
3632
3633 __ ret(0);
3634
3635 return start;
3636 }
3637
3638 /**
3639 * Arguments:
3640 *
3641 * Input:
3642 * xmm0 - float
3643 *
3644 * Output:
3645 * rax - float16 jshort
3646 */
3647 address StubGenerator::generate_floatToFloat16() {
3648 StubCodeMark mark(this, "StubRoutines", "floatToFloat16");
3649
3650 address start = __ pc();
3651
3652 BLOCK_COMMENT("Entry:");
3653 // No need for RuntimeStub frame since it is called only during JIT compilation
3654
3655 // Convert and put result into rax
3656 __ flt_to_flt16(rax, xmm0, xmm1);
3657
3658 __ ret(0);
3659
3660 return start;
3661 }
3662
3663 address StubGenerator::generate_cont_thaw(const char* label, Continuation::thaw_kind kind) {
3664 if (!Continuations::enabled()) return nullptr;
3665
3666 bool return_barrier = Continuation::is_thaw_return_barrier(kind);
3667 bool return_barrier_exception = Continuation::is_thaw_return_barrier_exception(kind);
3668
3669 StubCodeMark mark(this, "StubRoutines", label);
3670 address start = __ pc();
3671
3672 // TODO: Handle Valhalla return types. May require generating different return barriers.
3673
3674 if (!return_barrier) {
3675 // Pop return address. If we don't do this, we get a drift,
3676 // where the bottom-most frozen frame continuously grows.
3677 __ pop(c_rarg3);
3678 } else {
3679 __ movptr(rsp, Address(r15_thread, JavaThread::cont_entry_offset()));
3680 }
3681
3682 #ifdef ASSERT
3683 {
3684 Label L_good_sp;
3685 __ cmpptr(rsp, Address(r15_thread, JavaThread::cont_entry_offset()));
3686 __ jcc(Assembler::equal, L_good_sp);
3687 __ stop("Incorrect rsp at thaw entry");
3688 __ BIND(L_good_sp);
3689 }
3690 #endif // ASSERT
3691
3692 if (return_barrier) {
3693 // Preserve possible return value from a method returning to the return barrier.
3694 __ push(rax);
3695 __ push_d(xmm0);
3696 }
3697
3698 __ movptr(c_rarg0, r15_thread);
3699 __ movptr(c_rarg1, (return_barrier ? 1 : 0));
3700 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Continuation::prepare_thaw), 2);
3701 __ movptr(rbx, rax);
3702
3703 if (return_barrier) {
3704 // Restore return value from a method returning to the return barrier.
3705 // No safepoint in the call to thaw, so even an oop return value should be OK.
3706 __ pop_d(xmm0);
3707 __ pop(rax);
3708 }
3709
3710 #ifdef ASSERT
3711 {
3712 Label L_good_sp;
3713 __ cmpptr(rsp, Address(r15_thread, JavaThread::cont_entry_offset()));
3714 __ jcc(Assembler::equal, L_good_sp);
3715 __ stop("Incorrect rsp after prepare thaw");
3716 __ BIND(L_good_sp);
3717 }
3718 #endif // ASSERT
3719
3720 // rbx contains the size of the frames to thaw, 0 if overflow or no more frames
3721 Label L_thaw_success;
3722 __ testptr(rbx, rbx);
3723 __ jccb(Assembler::notZero, L_thaw_success);
3724 __ jump(RuntimeAddress(SharedRuntime::throw_StackOverflowError_entry()));
3725 __ bind(L_thaw_success);
3726
3727 // Make room for the thawed frames and align the stack.
3728 __ subptr(rsp, rbx);
3729 __ andptr(rsp, -StackAlignmentInBytes);
3730
3731 if (return_barrier) {
3732 // Preserve possible return value from a method returning to the return barrier. (Again.)
3733 __ push(rax);
3734 __ push_d(xmm0);
3735 }
3736
3737 // If we want, we can templatize thaw by kind, and have three different entries.
3738 __ movptr(c_rarg0, r15_thread);
3739 __ movptr(c_rarg1, kind);
3740 __ call_VM_leaf(Continuation::thaw_entry(), 2);
3741 __ movptr(rbx, rax);
3742
3743 if (return_barrier) {
3744 // Restore return value from a method returning to the return barrier. (Again.)
3745 // No safepoint in the call to thaw, so even an oop return value should be OK.
3746 __ pop_d(xmm0);
3747 __ pop(rax);
3748 } else {
3749 // Return 0 (success) from doYield.
3750 __ xorptr(rax, rax);
3751 }
3752
3753 // After thawing, rbx is the SP of the yielding frame.
3754 // Move there, and then to saved RBP slot.
3755 __ movptr(rsp, rbx);
3756 __ subptr(rsp, 2*wordSize);
3757
3758 if (return_barrier_exception) {
3759 __ movptr(c_rarg0, r15_thread);
3760 __ movptr(c_rarg1, Address(rsp, wordSize)); // return address
3761
3762 // rax still holds the original exception oop, save it before the call
3763 __ push(rax);
3764
3765 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), 2);
3766 __ movptr(rbx, rax);
3767
3768 // Continue at exception handler:
3769 // rax: exception oop
3770 // rbx: exception handler
3771 // rdx: exception pc
3772 __ pop(rax);
3773 __ verify_oop(rax);
3774 __ pop(rbp); // pop out RBP here too
3775 __ pop(rdx);
3776 __ jmp(rbx);
3777 } else {
3778 // We are "returning" into the topmost thawed frame; see Thaw::push_return_frame
3779 __ pop(rbp);
3780 __ ret(0);
3781 }
3782
3783 return start;
3784 }
3785
3786 address StubGenerator::generate_cont_thaw() {
3787 return generate_cont_thaw("Cont thaw", Continuation::thaw_top);
3788 }
3789
3790 // TODO: will probably need multiple return barriers depending on return type
3791
3792 address StubGenerator::generate_cont_returnBarrier() {
3793 return generate_cont_thaw("Cont thaw return barrier", Continuation::thaw_return_barrier);
3794 }
3795
3796 address StubGenerator::generate_cont_returnBarrier_exception() {
3797 return generate_cont_thaw("Cont thaw return barrier exception", Continuation::thaw_return_barrier_exception);
3798 }
3799
3800 // exception handler for upcall stubs
3801 address StubGenerator::generate_upcall_stub_exception_handler() {
3802 StubCodeMark mark(this, "StubRoutines", "upcall stub exception handler");
3803 address start = __ pc();
3804
3805 // native caller has no idea how to handle exceptions
3806 // we just crash here. Up to callee to catch exceptions.
3807 __ verify_oop(rax);
3808 __ vzeroupper();
3809 __ mov(c_rarg0, rax);
3810 __ andptr(rsp, -StackAlignmentInBytes); // align stack as required by ABI
3811 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows
3812 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, UpcallLinker::handle_uncaught_exception)));
3813 __ should_not_reach_here();
3814
3815 return start;
3816 }
3817
3818 // load Method* target of MethodHandle
3819 // j_rarg0 = jobject receiver
3820 // rbx = result
3821 address StubGenerator::generate_upcall_stub_load_target() {
3822 StubCodeMark mark(this, "StubRoutines", "upcall_stub_load_target");
3823 address start = __ pc();
3824
3825 __ resolve_global_jobject(j_rarg0, r15_thread, rscratch1);
3826 // Load target method from receiver
3827 __ load_heap_oop(rbx, Address(j_rarg0, java_lang_invoke_MethodHandle::form_offset()), rscratch1);
3828 __ load_heap_oop(rbx, Address(rbx, java_lang_invoke_LambdaForm::vmentry_offset()), rscratch1);
3829 __ load_heap_oop(rbx, Address(rbx, java_lang_invoke_MemberName::method_offset()), rscratch1);
3830 __ access_load_at(T_ADDRESS, IN_HEAP, rbx,
3831 Address(rbx, java_lang_invoke_ResolvedMethodName::vmtarget_offset()),
3832 noreg, noreg);
3833 __ movptr(Address(r15_thread, JavaThread::callee_target_offset()), rbx); // just in case callee is deoptimized
3834
3835 __ ret(0);
3836
3837 return start;
3838 }
3839
3840 address StubGenerator::generate_lookup_secondary_supers_table_stub(u1 super_klass_index) {
3841 StubCodeMark mark(this, "StubRoutines", "lookup_secondary_supers_table");
3842
3843 address start = __ pc();
3844
3845 const Register
3846 r_super_klass = rax,
3847 r_sub_klass = rsi,
3848 result = rdi;
3849
3850 __ lookup_secondary_supers_table_const(r_sub_klass, r_super_klass,
3851 rdx, rcx, rbx, r11, // temps
3852 result,
3853 super_klass_index);
3854 __ ret(0);
3855
3856 return start;
3857 }
3858
3859 // Slow path implementation for UseSecondarySupersTable.
3860 address StubGenerator::generate_lookup_secondary_supers_table_slow_path_stub() {
3861 StubCodeMark mark(this, "StubRoutines", "lookup_secondary_supers_table");
3862
3863 address start = __ pc();
3864
3865 const Register
3866 r_super_klass = rax,
3867 r_array_base = rbx,
3868 r_array_index = rdx,
3869 r_sub_klass = rsi,
3870 r_bitmap = r11,
3871 result = rdi;
3872
3873 Label L_success;
3874 __ lookup_secondary_supers_table_slow_path(r_super_klass, r_array_base, r_array_index, r_bitmap,
3875 rcx, rdi, // temps
3876 &L_success);
3877 // bind(L_failure);
3878 __ movl(result, 1);
3879 __ ret(0);
3880
3881 __ bind(L_success);
3882 __ movl(result, 0);
3883 __ ret(0);
3884
3885 return start;
3886 }
3887
3888 void StubGenerator::create_control_words() {
3889 // Round to nearest, 64-bit mode, exceptions masked, flags specialized
3890 StubRoutines::x86::_mxcsr_std = EnableX86ECoreOpts ? 0x1FBF : 0x1F80;
3891 // Round to zero, 64-bit mode, exceptions masked, flags specialized
3892 StubRoutines::x86::_mxcsr_rz = EnableX86ECoreOpts ? 0x7FBF : 0x7F80;
3893 }
3894
3895 // Initialization
3896 void StubGenerator::generate_initial_stubs() {
3897 // Generates all stubs and initializes the entry points
3898
3899 // This platform-specific settings are needed by generate_call_stub()
3900 create_control_words();
3901
3902 // Initialize table for unsafe copy memeory check.
3903 if (UnsafeMemoryAccess::_table == nullptr) {
3904 UnsafeMemoryAccess::create_table(16 + 4); // 16 for copyMemory; 4 for setMemory
3905 }
3906
3907 // entry points that exist in all platforms Note: This is code
3908 // that could be shared among different platforms - however the
3909 // benefit seems to be smaller than the disadvantage of having a
3910 // much more complicated generator structure. See also comment in
3911 // stubRoutines.hpp.
3912
3913 StubRoutines::_forward_exception_entry = generate_forward_exception();
3914
3915 // Generate these first because they are called from other stubs
3916 if (InlineTypeReturnedAsFields) {
3917 StubRoutines::_load_inline_type_fields_in_regs =
3918 generate_return_value_stub(CAST_FROM_FN_PTR(address, SharedRuntime::load_inline_type_fields_in_regs),
3919 "load_inline_type_fields_in_regs", false);
3920 StubRoutines::_store_inline_type_fields_to_buf =
3921 generate_return_value_stub(CAST_FROM_FN_PTR(address, SharedRuntime::store_inline_type_fields_to_buf),
3922 "store_inline_type_fields_to_buf", true);
3923 }
3924
3925 StubRoutines::_call_stub_entry =
3926 generate_call_stub(StubRoutines::_call_stub_return_address);
3927
3928 // is referenced by megamorphic call
3929 StubRoutines::_catch_exception_entry = generate_catch_exception();
3930
3931 // atomic calls
3932 StubRoutines::_fence_entry = generate_orderaccess_fence();
3933
3934 // platform dependent
3935 StubRoutines::x86::_get_previous_sp_entry = generate_get_previous_sp();
3936
3937 StubRoutines::x86::_verify_mxcsr_entry = generate_verify_mxcsr();
3938
3939 StubRoutines::x86::_f2i_fixup = generate_f2i_fixup();
3940 StubRoutines::x86::_f2l_fixup = generate_f2l_fixup();
3941 StubRoutines::x86::_d2i_fixup = generate_d2i_fixup();
3942 StubRoutines::x86::_d2l_fixup = generate_d2l_fixup();
3943
3944 StubRoutines::x86::_float_sign_mask = generate_fp_mask("float_sign_mask", 0x7FFFFFFF7FFFFFFF);
3945 StubRoutines::x86::_float_sign_flip = generate_fp_mask("float_sign_flip", 0x8000000080000000);
3946 StubRoutines::x86::_double_sign_mask = generate_fp_mask("double_sign_mask", 0x7FFFFFFFFFFFFFFF);
3947 StubRoutines::x86::_double_sign_flip = generate_fp_mask("double_sign_flip", 0x8000000000000000);
3948
3949 if (UseCRC32Intrinsics) {
3950 // set table address before stub generation which use it
3951 StubRoutines::_crc_table_adr = (address)StubRoutines::x86::_crc_table;
3952 StubRoutines::_updateBytesCRC32 = generate_updateBytesCRC32();
3953 }
3954
3955 if (UseCRC32CIntrinsics) {
3956 bool supports_clmul = VM_Version::supports_clmul();
3957 StubRoutines::x86::generate_CRC32C_table(supports_clmul);
3958 StubRoutines::_crc32c_table_addr = (address)StubRoutines::x86::_crc32c_table;
3959 StubRoutines::_updateBytesCRC32C = generate_updateBytesCRC32C(supports_clmul);
3960 }
3961
3962 if (VM_Version::supports_float16()) {
3963 // For results consistency both intrinsics should be enabled.
3964 // vmIntrinsics checks InlineIntrinsics flag, no need to check it here.
3965 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_float16ToFloat) &&
3966 vmIntrinsics::is_intrinsic_available(vmIntrinsics::_floatToFloat16)) {
3967 StubRoutines::_hf2f = generate_float16ToFloat();
3968 StubRoutines::_f2hf = generate_floatToFloat16();
3969 }
3970 }
3971
3972 generate_libm_stubs();
3973
3974 StubRoutines::_fmod = generate_libmFmod(); // from stubGenerator_x86_64_fmod.cpp
3975 }
3976
3977 // Call here from the interpreter or compiled code to either load
3978 // multiple returned values from the inline type instance being
3979 // returned to registers or to store returned values to a newly
3980 // allocated inline type instance.
3981 // Register is a class, but it would be assigned numerical value.
3982 // "0" is assigned for xmm0. Thus we need to ignore -Wnonnull.
3983 PRAGMA_DIAG_PUSH
3984 PRAGMA_NONNULL_IGNORED
3985 address StubGenerator::generate_return_value_stub(address destination, const char* name, bool has_res) {
3986 // We need to save all registers the calling convention may use so
3987 // the runtime calls read or update those registers. This needs to
3988 // be in sync with SharedRuntime::java_return_convention().
3989 enum layout {
3990 pad_off = frame::arg_reg_save_area_bytes/BytesPerInt, pad_off_2,
3991 rax_off, rax_off_2,
3992 j_rarg5_off, j_rarg5_2,
3993 j_rarg4_off, j_rarg4_2,
3994 j_rarg3_off, j_rarg3_2,
3995 j_rarg2_off, j_rarg2_2,
3996 j_rarg1_off, j_rarg1_2,
3997 j_rarg0_off, j_rarg0_2,
3998 j_farg0_off, j_farg0_2,
3999 j_farg1_off, j_farg1_2,
4000 j_farg2_off, j_farg2_2,
4001 j_farg3_off, j_farg3_2,
4002 j_farg4_off, j_farg4_2,
4003 j_farg5_off, j_farg5_2,
4004 j_farg6_off, j_farg6_2,
4005 j_farg7_off, j_farg7_2,
4006 rbp_off, rbp_off_2,
4007 return_off, return_off_2,
4008
4009 framesize
4010 };
4011
4012 CodeBuffer buffer(name, 1000, 512);
4013 MacroAssembler* _masm = new MacroAssembler(&buffer);
4014
4015 int frame_size_in_bytes = align_up(framesize*BytesPerInt, 16);
4016 assert(frame_size_in_bytes == framesize*BytesPerInt, "misaligned");
4017 int frame_size_in_slots = frame_size_in_bytes / BytesPerInt;
4018 int frame_size_in_words = frame_size_in_bytes / wordSize;
4019
4020 OopMapSet *oop_maps = new OopMapSet();
4021 OopMap* map = new OopMap(frame_size_in_slots, 0);
4022
4023 map->set_callee_saved(VMRegImpl::stack2reg(rax_off), rax->as_VMReg());
4024 map->set_callee_saved(VMRegImpl::stack2reg(j_rarg5_off), j_rarg5->as_VMReg());
4025 map->set_callee_saved(VMRegImpl::stack2reg(j_rarg4_off), j_rarg4->as_VMReg());
4026 map->set_callee_saved(VMRegImpl::stack2reg(j_rarg3_off), j_rarg3->as_VMReg());
4027 map->set_callee_saved(VMRegImpl::stack2reg(j_rarg2_off), j_rarg2->as_VMReg());
4028 map->set_callee_saved(VMRegImpl::stack2reg(j_rarg1_off), j_rarg1->as_VMReg());
4029 map->set_callee_saved(VMRegImpl::stack2reg(j_rarg0_off), j_rarg0->as_VMReg());
4030 map->set_callee_saved(VMRegImpl::stack2reg(j_farg0_off), j_farg0->as_VMReg());
4031 map->set_callee_saved(VMRegImpl::stack2reg(j_farg1_off), j_farg1->as_VMReg());
4032 map->set_callee_saved(VMRegImpl::stack2reg(j_farg2_off), j_farg2->as_VMReg());
4033 map->set_callee_saved(VMRegImpl::stack2reg(j_farg3_off), j_farg3->as_VMReg());
4034 map->set_callee_saved(VMRegImpl::stack2reg(j_farg4_off), j_farg4->as_VMReg());
4035 map->set_callee_saved(VMRegImpl::stack2reg(j_farg5_off), j_farg5->as_VMReg());
4036 map->set_callee_saved(VMRegImpl::stack2reg(j_farg6_off), j_farg6->as_VMReg());
4037 map->set_callee_saved(VMRegImpl::stack2reg(j_farg7_off), j_farg7->as_VMReg());
4038
4039 int start = __ offset();
4040
4041 __ subptr(rsp, frame_size_in_bytes - 8 /* return address*/);
4042
4043 __ movptr(Address(rsp, rbp_off * BytesPerInt), rbp);
4044 __ movdbl(Address(rsp, j_farg7_off * BytesPerInt), j_farg7);
4045 __ movdbl(Address(rsp, j_farg6_off * BytesPerInt), j_farg6);
4046 __ movdbl(Address(rsp, j_farg5_off * BytesPerInt), j_farg5);
4047 __ movdbl(Address(rsp, j_farg4_off * BytesPerInt), j_farg4);
4048 __ movdbl(Address(rsp, j_farg3_off * BytesPerInt), j_farg3);
4049 __ movdbl(Address(rsp, j_farg2_off * BytesPerInt), j_farg2);
4050 __ movdbl(Address(rsp, j_farg1_off * BytesPerInt), j_farg1);
4051 __ movdbl(Address(rsp, j_farg0_off * BytesPerInt), j_farg0);
4052
4053 __ movptr(Address(rsp, j_rarg0_off * BytesPerInt), j_rarg0);
4054 __ movptr(Address(rsp, j_rarg1_off * BytesPerInt), j_rarg1);
4055 __ movptr(Address(rsp, j_rarg2_off * BytesPerInt), j_rarg2);
4056 __ movptr(Address(rsp, j_rarg3_off * BytesPerInt), j_rarg3);
4057 __ movptr(Address(rsp, j_rarg4_off * BytesPerInt), j_rarg4);
4058 __ movptr(Address(rsp, j_rarg5_off * BytesPerInt), j_rarg5);
4059 __ movptr(Address(rsp, rax_off * BytesPerInt), rax);
4060
4061 int frame_complete = __ offset();
4062
4063 __ set_last_Java_frame(noreg, noreg, nullptr, rscratch1);
4064
4065 __ mov(c_rarg0, r15_thread);
4066 __ mov(c_rarg1, rax);
4067
4068 __ call(RuntimeAddress(destination));
4069
4070 // Set an oopmap for the call site.
4071
4072 oop_maps->add_gc_map( __ offset() - start, map);
4073
4074 // clear last_Java_sp
4075 __ reset_last_Java_frame(false);
4076
4077 __ movptr(rbp, Address(rsp, rbp_off * BytesPerInt));
4078 __ movdbl(j_farg7, Address(rsp, j_farg7_off * BytesPerInt));
4079 __ movdbl(j_farg6, Address(rsp, j_farg6_off * BytesPerInt));
4080 __ movdbl(j_farg5, Address(rsp, j_farg5_off * BytesPerInt));
4081 __ movdbl(j_farg4, Address(rsp, j_farg4_off * BytesPerInt));
4082 __ movdbl(j_farg3, Address(rsp, j_farg3_off * BytesPerInt));
4083 __ movdbl(j_farg2, Address(rsp, j_farg2_off * BytesPerInt));
4084 __ movdbl(j_farg1, Address(rsp, j_farg1_off * BytesPerInt));
4085 __ movdbl(j_farg0, Address(rsp, j_farg0_off * BytesPerInt));
4086
4087 __ movptr(j_rarg0, Address(rsp, j_rarg0_off * BytesPerInt));
4088 __ movptr(j_rarg1, Address(rsp, j_rarg1_off * BytesPerInt));
4089 __ movptr(j_rarg2, Address(rsp, j_rarg2_off * BytesPerInt));
4090 __ movptr(j_rarg3, Address(rsp, j_rarg3_off * BytesPerInt));
4091 __ movptr(j_rarg4, Address(rsp, j_rarg4_off * BytesPerInt));
4092 __ movptr(j_rarg5, Address(rsp, j_rarg5_off * BytesPerInt));
4093 __ movptr(rax, Address(rsp, rax_off * BytesPerInt));
4094
4095 __ addptr(rsp, frame_size_in_bytes-8);
4096
4097 // check for pending exceptions
4098 Label pending;
4099 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
4100 __ jcc(Assembler::notEqual, pending);
4101
4102 if (has_res) {
4103 __ get_vm_result(rax, r15_thread);
4104 }
4105
4106 __ ret(0);
4107
4108 __ bind(pending);
4109
4110 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset()));
4111 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
4112
4113 // -------------
4114 // make sure all code is generated
4115 _masm->flush();
4116
4117 RuntimeStub* stub = RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, false);
4118 return stub->entry_point();
4119 }
4120
4121 void StubGenerator::generate_continuation_stubs() {
4122 // Continuation stubs:
4123 StubRoutines::_cont_thaw = generate_cont_thaw();
4124 StubRoutines::_cont_returnBarrier = generate_cont_returnBarrier();
4125 StubRoutines::_cont_returnBarrierExc = generate_cont_returnBarrier_exception();
4126 }
4127
4128 void StubGenerator::generate_final_stubs() {
4129 // Generates the rest of stubs and initializes the entry points
4130
4131 // support for verify_oop (must happen after universe_init)
4132 if (VerifyOops) {
4133 StubRoutines::_verify_oop_subroutine_entry = generate_verify_oop();
4134 }
4135
4136 // data cache line writeback
4137 StubRoutines::_data_cache_writeback = generate_data_cache_writeback();
4138 StubRoutines::_data_cache_writeback_sync = generate_data_cache_writeback_sync();
4139
4140 // arraycopy stubs used by compilers
4141 generate_arraycopy_stubs();
4142
4143 BarrierSetNMethod* bs_nm = BarrierSet::barrier_set()->barrier_set_nmethod();
4144 if (bs_nm != nullptr) {
4145 StubRoutines::_method_entry_barrier = generate_method_entry_barrier();
4146 }
4147
4148 if (UseVectorizedMismatchIntrinsic) {
4149 StubRoutines::_vectorizedMismatch = generate_vectorizedMismatch();
4150 }
4151
4152 StubRoutines::_upcall_stub_exception_handler = generate_upcall_stub_exception_handler();
4153 StubRoutines::_upcall_stub_load_target = generate_upcall_stub_load_target();
4154 }
4155
4156 void StubGenerator::generate_compiler_stubs() {
4157 #if COMPILER2_OR_JVMCI
4158
4159 // Entry points that are C2 compiler specific.
4160
4161 StubRoutines::x86::_vector_float_sign_mask = generate_vector_mask("vector_float_sign_mask", 0x7FFFFFFF7FFFFFFF);
4162 StubRoutines::x86::_vector_float_sign_flip = generate_vector_mask("vector_float_sign_flip", 0x8000000080000000);
4163 StubRoutines::x86::_vector_double_sign_mask = generate_vector_mask("vector_double_sign_mask", 0x7FFFFFFFFFFFFFFF);
4164 StubRoutines::x86::_vector_double_sign_flip = generate_vector_mask("vector_double_sign_flip", 0x8000000000000000);
4165 StubRoutines::x86::_vector_all_bits_set = generate_vector_mask("vector_all_bits_set", 0xFFFFFFFFFFFFFFFF);
4166 StubRoutines::x86::_vector_int_mask_cmp_bits = generate_vector_mask("vector_int_mask_cmp_bits", 0x0000000100000001);
4167 StubRoutines::x86::_vector_short_to_byte_mask = generate_vector_mask("vector_short_to_byte_mask", 0x00ff00ff00ff00ff);
4168 StubRoutines::x86::_vector_byte_perm_mask = generate_vector_byte_perm_mask("vector_byte_perm_mask");
4169 StubRoutines::x86::_vector_int_to_byte_mask = generate_vector_mask("vector_int_to_byte_mask", 0x000000ff000000ff);
4170 StubRoutines::x86::_vector_int_to_short_mask = generate_vector_mask("vector_int_to_short_mask", 0x0000ffff0000ffff);
4171 StubRoutines::x86::_vector_32_bit_mask = generate_vector_custom_i32("vector_32_bit_mask", Assembler::AVX_512bit,
4172 0xFFFFFFFF, 0, 0, 0);
4173 StubRoutines::x86::_vector_64_bit_mask = generate_vector_custom_i32("vector_64_bit_mask", Assembler::AVX_512bit,
4174 0xFFFFFFFF, 0xFFFFFFFF, 0, 0);
4175 StubRoutines::x86::_vector_int_shuffle_mask = generate_vector_mask("vector_int_shuffle_mask", 0x0302010003020100);
4176 StubRoutines::x86::_vector_byte_shuffle_mask = generate_vector_byte_shuffle_mask("vector_byte_shuffle_mask");
4177 StubRoutines::x86::_vector_short_shuffle_mask = generate_vector_mask("vector_short_shuffle_mask", 0x0100010001000100);
4178 StubRoutines::x86::_vector_long_shuffle_mask = generate_vector_mask("vector_long_shuffle_mask", 0x0000000100000000);
4179 StubRoutines::x86::_vector_long_sign_mask = generate_vector_mask("vector_long_sign_mask", 0x8000000000000000);
4180 StubRoutines::x86::_vector_iota_indices = generate_iota_indices("iota_indices");
4181 StubRoutines::x86::_vector_count_leading_zeros_lut = generate_count_leading_zeros_lut("count_leading_zeros_lut");
4182 StubRoutines::x86::_vector_reverse_bit_lut = generate_vector_reverse_bit_lut("reverse_bit_lut");
4183 StubRoutines::x86::_vector_reverse_byte_perm_mask_long = generate_vector_reverse_byte_perm_mask_long("perm_mask_long");
4184 StubRoutines::x86::_vector_reverse_byte_perm_mask_int = generate_vector_reverse_byte_perm_mask_int("perm_mask_int");
4185 StubRoutines::x86::_vector_reverse_byte_perm_mask_short = generate_vector_reverse_byte_perm_mask_short("perm_mask_short");
4186
4187 if (VM_Version::supports_avx2() && !VM_Version::supports_avx512vl()) {
4188 StubRoutines::x86::_compress_perm_table32 = generate_compress_perm_table("compress_perm_table32", 32);
4189 StubRoutines::x86::_compress_perm_table64 = generate_compress_perm_table("compress_perm_table64", 64);
4190 StubRoutines::x86::_expand_perm_table32 = generate_expand_perm_table("expand_perm_table32", 32);
4191 StubRoutines::x86::_expand_perm_table64 = generate_expand_perm_table("expand_perm_table64", 64);
4192 }
4193
4194 if (VM_Version::supports_avx2() && !VM_Version::supports_avx512_vpopcntdq()) {
4195 // lut implementation influenced by counting 1s algorithm from section 5-1 of Hackers' Delight.
4196 StubRoutines::x86::_vector_popcount_lut = generate_popcount_avx_lut("popcount_lut");
4197 }
4198
4199 generate_aes_stubs();
4200
4201 generate_ghash_stubs();
4202
4203 generate_chacha_stubs();
4204
4205 generate_sha3_stubs();
4206
4207 #ifdef COMPILER2
4208 if ((UseAVX == 2) && EnableX86ECoreOpts) {
4209 generate_string_indexof(StubRoutines::_string_indexof_array);
4210 }
4211 #endif
4212
4213 if (UseAdler32Intrinsics) {
4214 StubRoutines::_updateBytesAdler32 = generate_updateBytesAdler32();
4215 }
4216
4217 if (UsePoly1305Intrinsics) {
4218 StubRoutines::_poly1305_processBlocks = generate_poly1305_processBlocks();
4219 }
4220
4221 if (UseIntPolyIntrinsics) {
4222 StubRoutines::_intpoly_montgomeryMult_P256 = generate_intpoly_montgomeryMult_P256();
4223 StubRoutines::_intpoly_assign = generate_intpoly_assign();
4224 }
4225
4226 if (UseMD5Intrinsics) {
4227 StubRoutines::_md5_implCompress = generate_md5_implCompress(false, "md5_implCompress");
4228 StubRoutines::_md5_implCompressMB = generate_md5_implCompress(true, "md5_implCompressMB");
4229 }
4230
4231 if (UseSHA1Intrinsics) {
4232 StubRoutines::x86::_upper_word_mask_addr = generate_upper_word_mask();
4233 StubRoutines::x86::_shuffle_byte_flip_mask_addr = generate_shuffle_byte_flip_mask();
4234 StubRoutines::_sha1_implCompress = generate_sha1_implCompress(false, "sha1_implCompress");
4235 StubRoutines::_sha1_implCompressMB = generate_sha1_implCompress(true, "sha1_implCompressMB");
4236 }
4237
4238 if (UseSHA256Intrinsics) {
4239 StubRoutines::x86::_k256_adr = (address)StubRoutines::x86::_k256;
4240 char* dst = (char*)StubRoutines::x86::_k256_W;
4241 char* src = (char*)StubRoutines::x86::_k256;
4242 for (int ii = 0; ii < 16; ++ii) {
4243 memcpy(dst + 32 * ii, src + 16 * ii, 16);
4244 memcpy(dst + 32 * ii + 16, src + 16 * ii, 16);
4245 }
4246 StubRoutines::x86::_k256_W_adr = (address)StubRoutines::x86::_k256_W;
4247 StubRoutines::x86::_pshuffle_byte_flip_mask_addr = generate_pshuffle_byte_flip_mask();
4248 StubRoutines::_sha256_implCompress = generate_sha256_implCompress(false, "sha256_implCompress");
4249 StubRoutines::_sha256_implCompressMB = generate_sha256_implCompress(true, "sha256_implCompressMB");
4250 }
4251
4252 if (UseSHA512Intrinsics) {
4253 StubRoutines::x86::_k512_W_addr = (address)StubRoutines::x86::_k512_W;
4254 StubRoutines::x86::_pshuffle_byte_flip_mask_addr_sha512 = generate_pshuffle_byte_flip_mask_sha512();
4255 StubRoutines::_sha512_implCompress = generate_sha512_implCompress(false, "sha512_implCompress");
4256 StubRoutines::_sha512_implCompressMB = generate_sha512_implCompress(true, "sha512_implCompressMB");
4257 }
4258
4259 if (UseBASE64Intrinsics) {
4260 if(VM_Version::supports_avx2()) {
4261 StubRoutines::x86::_avx2_shuffle_base64 = base64_avx2_shuffle_addr();
4262 StubRoutines::x86::_avx2_input_mask_base64 = base64_avx2_input_mask_addr();
4263 StubRoutines::x86::_avx2_lut_base64 = base64_avx2_lut_addr();
4264 StubRoutines::x86::_avx2_decode_tables_base64 = base64_AVX2_decode_tables_addr();
4265 StubRoutines::x86::_avx2_decode_lut_tables_base64 = base64_AVX2_decode_LUT_tables_addr();
4266 }
4267 StubRoutines::x86::_encoding_table_base64 = base64_encoding_table_addr();
4268 if (VM_Version::supports_avx512_vbmi()) {
4269 StubRoutines::x86::_shuffle_base64 = base64_shuffle_addr();
4270 StubRoutines::x86::_lookup_lo_base64 = base64_vbmi_lookup_lo_addr();
4271 StubRoutines::x86::_lookup_hi_base64 = base64_vbmi_lookup_hi_addr();
4272 StubRoutines::x86::_lookup_lo_base64url = base64_vbmi_lookup_lo_url_addr();
4273 StubRoutines::x86::_lookup_hi_base64url = base64_vbmi_lookup_hi_url_addr();
4274 StubRoutines::x86::_pack_vec_base64 = base64_vbmi_pack_vec_addr();
4275 StubRoutines::x86::_join_0_1_base64 = base64_vbmi_join_0_1_addr();
4276 StubRoutines::x86::_join_1_2_base64 = base64_vbmi_join_1_2_addr();
4277 StubRoutines::x86::_join_2_3_base64 = base64_vbmi_join_2_3_addr();
4278 }
4279 StubRoutines::x86::_decoding_table_base64 = base64_decoding_table_addr();
4280 StubRoutines::_base64_encodeBlock = generate_base64_encodeBlock();
4281 StubRoutines::_base64_decodeBlock = generate_base64_decodeBlock();
4282 }
4283
4284 #ifdef COMPILER2
4285 if (UseMultiplyToLenIntrinsic) {
4286 StubRoutines::_multiplyToLen = generate_multiplyToLen();
4287 }
4288 if (UseSquareToLenIntrinsic) {
4289 StubRoutines::_squareToLen = generate_squareToLen();
4290 }
4291 if (UseMulAddIntrinsic) {
4292 StubRoutines::_mulAdd = generate_mulAdd();
4293 }
4294 if (VM_Version::supports_avx512_vbmi2()) {
4295 StubRoutines::_bigIntegerRightShiftWorker = generate_bigIntegerRightShift();
4296 StubRoutines::_bigIntegerLeftShiftWorker = generate_bigIntegerLeftShift();
4297 }
4298 if (UseSecondarySupersTable) {
4299 StubRoutines::_lookup_secondary_supers_table_slow_path_stub = generate_lookup_secondary_supers_table_slow_path_stub();
4300 if (! InlineSecondarySupersTest) {
4301 for (int slot = 0; slot < Klass::SECONDARY_SUPERS_TABLE_SIZE; slot++) {
4302 StubRoutines::_lookup_secondary_supers_table_stubs[slot] = generate_lookup_secondary_supers_table_stub(slot);
4303 }
4304 }
4305 }
4306 if (UseMontgomeryMultiplyIntrinsic) {
4307 StubRoutines::_montgomeryMultiply
4308 = CAST_FROM_FN_PTR(address, SharedRuntime::montgomery_multiply);
4309 }
4310 if (UseMontgomerySquareIntrinsic) {
4311 StubRoutines::_montgomerySquare
4312 = CAST_FROM_FN_PTR(address, SharedRuntime::montgomery_square);
4313 }
4314
4315 // Load x86_64_sort library on supported hardware to enable SIMD sort and partition intrinsics
4316
4317 if (VM_Version::is_intel() && (VM_Version::supports_avx512dq() || VM_Version::supports_avx2())) {
4318 void *libsimdsort = nullptr;
4319 char ebuf_[1024];
4320 char dll_name_simd_sort[JVM_MAXPATHLEN];
4321 if (os::dll_locate_lib(dll_name_simd_sort, sizeof(dll_name_simd_sort), Arguments::get_dll_dir(), "simdsort")) {
4322 libsimdsort = os::dll_load(dll_name_simd_sort, ebuf_, sizeof ebuf_);
4323 }
4324 // Get addresses for SIMD sort and partition routines
4325 if (libsimdsort != nullptr) {
4326 log_info(library)("Loaded library %s, handle " INTPTR_FORMAT, JNI_LIB_PREFIX "simdsort" JNI_LIB_SUFFIX, p2i(libsimdsort));
4327
4328 snprintf(ebuf_, sizeof(ebuf_), VM_Version::supports_avx512dq() ? "avx512_sort" : "avx2_sort");
4329 StubRoutines::_array_sort = (address)os::dll_lookup(libsimdsort, ebuf_);
4330
4331 snprintf(ebuf_, sizeof(ebuf_), VM_Version::supports_avx512dq() ? "avx512_partition" : "avx2_partition");
4332 StubRoutines::_array_partition = (address)os::dll_lookup(libsimdsort, ebuf_);
4333 }
4334 }
4335
4336 // Get svml stub routine addresses
4337 void *libjsvml = nullptr;
4338 char ebuf[1024];
4339 char dll_name[JVM_MAXPATHLEN];
4340 if (os::dll_locate_lib(dll_name, sizeof(dll_name), Arguments::get_dll_dir(), "jsvml")) {
4341 libjsvml = os::dll_load(dll_name, ebuf, sizeof ebuf);
4342 }
4343 if (libjsvml != nullptr) {
4344 // SVML method naming convention
4345 // All the methods are named as __jsvml_op<T><N>_ha_<VV>
4346 // Where:
4347 // ha stands for high accuracy
4348 // <T> is optional to indicate float/double
4349 // Set to f for vector float operation
4350 // Omitted for vector double operation
4351 // <N> is the number of elements in the vector
4352 // 1, 2, 4, 8, 16
4353 // e.g. 128 bit float vector has 4 float elements
4354 // <VV> indicates the avx/sse level:
4355 // z0 is AVX512, l9 is AVX2, e9 is AVX1 and ex is for SSE2
4356 // e.g. __jsvml_expf16_ha_z0 is the method for computing 16 element vector float exp using AVX 512 insns
4357 // __jsvml_exp8_ha_z0 is the method for computing 8 element vector double exp using AVX 512 insns
4358
4359 log_info(library)("Loaded library %s, handle " INTPTR_FORMAT, JNI_LIB_PREFIX "jsvml" JNI_LIB_SUFFIX, p2i(libjsvml));
4360 if (UseAVX > 2) {
4361 for (int op = 0; op < VectorSupport::NUM_VECTOR_OP_MATH; op++) {
4362 int vop = VectorSupport::VECTOR_OP_MATH_START + op;
4363 if ((!VM_Version::supports_avx512dq()) &&
4364 (vop == VectorSupport::VECTOR_OP_LOG || vop == VectorSupport::VECTOR_OP_LOG10 || vop == VectorSupport::VECTOR_OP_POW)) {
4365 continue;
4366 }
4367 snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf16_ha_z0", VectorSupport::mathname[op]);
4368 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_512][op] = (address)os::dll_lookup(libjsvml, ebuf);
4369
4370 snprintf(ebuf, sizeof(ebuf), "__jsvml_%s8_ha_z0", VectorSupport::mathname[op]);
4371 StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_512][op] = (address)os::dll_lookup(libjsvml, ebuf);
4372 }
4373 }
4374 const char* avx_sse_str = (UseAVX >= 2) ? "l9" : ((UseAVX == 1) ? "e9" : "ex");
4375 for (int op = 0; op < VectorSupport::NUM_VECTOR_OP_MATH; op++) {
4376 int vop = VectorSupport::VECTOR_OP_MATH_START + op;
4377 if (vop == VectorSupport::VECTOR_OP_POW) {
4378 continue;
4379 }
4380 snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf4_ha_%s", VectorSupport::mathname[op], avx_sse_str);
4381 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_64][op] = (address)os::dll_lookup(libjsvml, ebuf);
4382
4383 snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf4_ha_%s", VectorSupport::mathname[op], avx_sse_str);
4384 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_128][op] = (address)os::dll_lookup(libjsvml, ebuf);
4385
4386 snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf8_ha_%s", VectorSupport::mathname[op], avx_sse_str);
4387 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_256][op] = (address)os::dll_lookup(libjsvml, ebuf);
4388
4389 snprintf(ebuf, sizeof(ebuf), "__jsvml_%s1_ha_%s", VectorSupport::mathname[op], avx_sse_str);
4390 StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_64][op] = (address)os::dll_lookup(libjsvml, ebuf);
4391
4392 snprintf(ebuf, sizeof(ebuf), "__jsvml_%s2_ha_%s", VectorSupport::mathname[op], avx_sse_str);
4393 StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_128][op] = (address)os::dll_lookup(libjsvml, ebuf);
4394
4395 snprintf(ebuf, sizeof(ebuf), "__jsvml_%s4_ha_%s", VectorSupport::mathname[op], avx_sse_str);
4396 StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_256][op] = (address)os::dll_lookup(libjsvml, ebuf);
4397 }
4398 }
4399
4400 #endif // COMPILER2
4401 #endif // COMPILER2_OR_JVMCI
4402 }
4403
4404 StubGenerator::StubGenerator(CodeBuffer* code, StubsKind kind) : StubCodeGenerator(code) {
4405 DEBUG_ONLY( _regs_in_thread = false; )
4406 switch(kind) {
4407 case Initial_stubs:
4408 generate_initial_stubs();
4409 break;
4410 case Continuation_stubs:
4411 generate_continuation_stubs();
4412 break;
4413 case Compiler_stubs:
4414 generate_compiler_stubs();
4415 break;
4416 case Final_stubs:
4417 generate_final_stubs();
4418 break;
4419 default:
4420 fatal("unexpected stubs kind: %d", kind);
4421 break;
4422 };
4423 }
4424
4425 void StubGenerator_generate(CodeBuffer* code, StubCodeGenerator::StubsKind kind) {
4426 StubGenerator g(code, kind);
4427 }
4428
4429 #undef __