1 /*
2 * Copyright (c) 2003, 2022, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/assembler.hpp"
27 #include "asm/macroAssembler.hpp"
28 #include "classfile/vmIntrinsics.hpp"
29 #include "compiler/oopMap.hpp"
30 #include "gc/shared/barrierSet.hpp"
31 #include "gc/shared/barrierSetAssembler.hpp"
32 #include "gc/shared/barrierSetNMethod.hpp"
33 #include "gc/shared/gc_globals.hpp"
34 #include "memory/universe.hpp"
35 #include "prims/jvmtiExport.hpp"
36 #include "runtime/arguments.hpp"
37 #include "runtime/javaThread.hpp"
38 #include "runtime/sharedRuntime.hpp"
39 #include "runtime/stubRoutines.hpp"
40 #include "utilities/macros.hpp"
41 #include "vmreg_x86.inline.hpp"
42 #include "stubGenerator_x86_64.hpp"
43 #ifdef COMPILER2
44 #include "opto/runtime.hpp"
45 #include "opto/c2_globals.hpp"
46 #endif
47 #if INCLUDE_JVMCI
48 #include "jvmci/jvmci_globals.hpp"
49 #endif
50 #if INCLUDE_ZGC
51 #include "gc/z/zThreadLocalData.hpp"
52 #endif
53 #if INCLUDE_JFR
54 #include "jfr/support/jfrIntrinsics.hpp"
55 #endif
56
57 // For a more detailed description of the stub routine structure
58 // see the comment in stubRoutines.hpp
59
60 #define __ _masm->
61 #define TIMES_OOP (UseCompressedOops ? Address::times_4 : Address::times_8)
62
63 #ifdef PRODUCT
64 #define BLOCK_COMMENT(str) /* nothing */
65 #else
66 #define BLOCK_COMMENT(str) __ block_comment(str)
67 #endif // PRODUCT
68
69 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
70
71 //
72 // Linux Arguments:
73 // c_rarg0: call wrapper address address
74 // c_rarg1: result address
75 // c_rarg2: result type BasicType
76 // c_rarg3: method Method*
77 // c_rarg4: (interpreter) entry point address
78 // c_rarg5: parameters intptr_t*
79 // 16(rbp): parameter size (in words) int
80 // 24(rbp): thread Thread*
81 //
82 // [ return_from_Java ] <--- rsp
83 // [ argument word n ]
84 // ...
85 // -12 [ argument word 1 ]
86 // -11 [ saved r15 ] <--- rsp_after_call
87 // -10 [ saved r14 ]
88 // -9 [ saved r13 ]
89 // -8 [ saved r12 ]
90 // -7 [ saved rbx ]
91 // -6 [ call wrapper ]
92 // -5 [ result ]
93 // -4 [ result type ]
94 // -3 [ method ]
95 // -2 [ entry point ]
96 // -1 [ parameters ]
97 // 0 [ saved rbp ] <--- rbp
98 // 1 [ return address ]
99 // 2 [ parameter size ]
100 // 3 [ thread ]
101 //
102 // Windows Arguments:
103 // c_rarg0: call wrapper address address
104 // c_rarg1: result address
105 // c_rarg2: result type BasicType
106 // c_rarg3: method Method*
107 // 48(rbp): (interpreter) entry point address
108 // 56(rbp): parameters intptr_t*
109 // 64(rbp): parameter size (in words) int
110 // 72(rbp): thread Thread*
111 //
112 // [ return_from_Java ] <--- rsp
113 // [ argument word n ]
114 // ...
115 // -60 [ argument word 1 ]
116 // -59 [ saved xmm31 ] <--- rsp after_call
117 // [ saved xmm16-xmm30 ] (EVEX enabled, else the space is blank)
118 // -27 [ saved xmm15 ]
119 // [ saved xmm7-xmm14 ]
120 // -9 [ saved xmm6 ] (each xmm register takes 2 slots)
121 // -7 [ saved r15 ]
122 // -6 [ saved r14 ]
123 // -5 [ saved r13 ]
124 // -4 [ saved r12 ]
125 // -3 [ saved rdi ]
126 // -2 [ saved rsi ]
127 // -1 [ saved rbx ]
128 // 0 [ saved rbp ] <--- rbp
129 // 1 [ return address ]
130 // 2 [ call wrapper ]
131 // 3 [ result ]
132 // 4 [ result type ]
133 // 5 [ method ]
134 // 6 [ entry point ]
135 // 7 [ parameters ]
136 // 8 [ parameter size ]
137 // 9 [ thread ]
138 //
139 // Windows reserves the callers stack space for arguments 1-4.
140 // We spill c_rarg0-c_rarg3 to this space.
141
142 // Call stub stack layout word offsets from rbp
143 #ifdef _WIN64
144 enum call_stub_layout {
145 xmm_save_first = 6, // save from xmm6
146 xmm_save_last = 31, // to xmm31
147 xmm_save_base = -9,
148 rsp_after_call_off = xmm_save_base - 2 * (xmm_save_last - xmm_save_first), // -27
149 r15_off = -7,
150 r14_off = -6,
151 r13_off = -5,
152 r12_off = -4,
153 rdi_off = -3,
154 rsi_off = -2,
155 rbx_off = -1,
156 rbp_off = 0,
157 retaddr_off = 1,
158 call_wrapper_off = 2,
159 result_off = 3,
160 result_type_off = 4,
161 method_off = 5,
162 entry_point_off = 6,
163 parameters_off = 7,
164 parameter_size_off = 8,
165 thread_off = 9
166 };
167
168 static Address xmm_save(int reg) {
169 assert(reg >= xmm_save_first && reg <= xmm_save_last, "XMM register number out of range");
170 return Address(rbp, (xmm_save_base - (reg - xmm_save_first) * 2) * wordSize);
171 }
172 #else // !_WIN64
173 enum call_stub_layout {
174 rsp_after_call_off = -12,
175 mxcsr_off = rsp_after_call_off,
176 r15_off = -11,
177 r14_off = -10,
178 r13_off = -9,
179 r12_off = -8,
180 rbx_off = -7,
181 call_wrapper_off = -6,
182 result_off = -5,
183 result_type_off = -4,
184 method_off = -3,
185 entry_point_off = -2,
186 parameters_off = -1,
187 rbp_off = 0,
188 retaddr_off = 1,
189 parameter_size_off = 2,
190 thread_off = 3
191 };
192 #endif // _WIN64
193
194 address StubGenerator::generate_call_stub(address& return_address) {
195
196 assert((int)frame::entry_frame_after_call_words == -(int)rsp_after_call_off + 1 &&
197 (int)frame::entry_frame_call_wrapper_offset == (int)call_wrapper_off,
198 "adjust this code");
199 StubCodeMark mark(this, "StubRoutines", "call_stub");
200 address start = __ pc();
201
202 // same as in generate_catch_exception()!
203 const Address rsp_after_call(rbp, rsp_after_call_off * wordSize);
204
205 const Address call_wrapper (rbp, call_wrapper_off * wordSize);
206 const Address result (rbp, result_off * wordSize);
207 const Address result_type (rbp, result_type_off * wordSize);
208 const Address method (rbp, method_off * wordSize);
209 const Address entry_point (rbp, entry_point_off * wordSize);
210 const Address parameters (rbp, parameters_off * wordSize);
211 const Address parameter_size(rbp, parameter_size_off * wordSize);
212
213 // same as in generate_catch_exception()!
214 const Address thread (rbp, thread_off * wordSize);
215
216 const Address r15_save(rbp, r15_off * wordSize);
217 const Address r14_save(rbp, r14_off * wordSize);
218 const Address r13_save(rbp, r13_off * wordSize);
219 const Address r12_save(rbp, r12_off * wordSize);
220 const Address rbx_save(rbp, rbx_off * wordSize);
221
222 // stub code
223 __ enter();
224 __ subptr(rsp, -rsp_after_call_off * wordSize);
225
226 // save register parameters
227 #ifndef _WIN64
228 __ movptr(parameters, c_rarg5); // parameters
229 __ movptr(entry_point, c_rarg4); // entry_point
230 #endif
231
232 __ movptr(method, c_rarg3); // method
233 __ movl(result_type, c_rarg2); // result type
234 __ movptr(result, c_rarg1); // result
235 __ movptr(call_wrapper, c_rarg0); // call wrapper
236
237 // save regs belonging to calling function
238 __ movptr(rbx_save, rbx);
239 __ movptr(r12_save, r12);
240 __ movptr(r13_save, r13);
241 __ movptr(r14_save, r14);
242 __ movptr(r15_save, r15);
243
244 #ifdef _WIN64
245 int last_reg = 15;
246 if (UseAVX > 2) {
247 last_reg = 31;
248 }
249 if (VM_Version::supports_evex()) {
250 for (int i = xmm_save_first; i <= last_reg; i++) {
251 __ vextractf32x4(xmm_save(i), as_XMMRegister(i), 0);
252 }
253 } else {
254 for (int i = xmm_save_first; i <= last_reg; i++) {
255 __ movdqu(xmm_save(i), as_XMMRegister(i));
256 }
257 }
258
259 const Address rdi_save(rbp, rdi_off * wordSize);
260 const Address rsi_save(rbp, rsi_off * wordSize);
261
262 __ movptr(rsi_save, rsi);
263 __ movptr(rdi_save, rdi);
264 #else
265 const Address mxcsr_save(rbp, mxcsr_off * wordSize);
266 {
267 Label skip_ldmx;
268 __ stmxcsr(mxcsr_save);
269 __ movl(rax, mxcsr_save);
270 __ andl(rax, 0xFFC0); // Mask out any pending exceptions (only check control and mask bits)
271 ExternalAddress mxcsr_std(StubRoutines::x86::addr_mxcsr_std());
272 __ cmp32(rax, mxcsr_std, rscratch1);
273 __ jcc(Assembler::equal, skip_ldmx);
274 __ ldmxcsr(mxcsr_std, rscratch1);
275 __ bind(skip_ldmx);
276 }
277 #endif
278
279 // Load up thread register
280 __ movptr(r15_thread, thread);
281 __ reinit_heapbase();
282
283 #ifdef ASSERT
284 // make sure we have no pending exceptions
285 {
286 Label L;
287 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), NULL_WORD);
288 __ jcc(Assembler::equal, L);
289 __ stop("StubRoutines::call_stub: entered with pending exception");
290 __ bind(L);
291 }
292 #endif
293
294 // pass parameters if any
295 BLOCK_COMMENT("pass parameters if any");
296 Label parameters_done;
297 __ movl(c_rarg3, parameter_size);
298 __ testl(c_rarg3, c_rarg3);
299 __ jcc(Assembler::zero, parameters_done);
300
301 Label loop;
302 __ movptr(c_rarg2, parameters); // parameter pointer
303 __ movl(c_rarg1, c_rarg3); // parameter counter is in c_rarg1
304 __ BIND(loop);
305 __ movptr(rax, Address(c_rarg2, 0));// get parameter
306 __ addptr(c_rarg2, wordSize); // advance to next parameter
307 __ decrementl(c_rarg1); // decrement counter
308 __ push(rax); // pass parameter
309 __ jcc(Assembler::notZero, loop);
310
311 // call Java function
312 __ BIND(parameters_done);
313 __ movptr(rbx, method); // get Method*
314 __ movptr(c_rarg1, entry_point); // get entry_point
315 __ mov(r13, rsp); // set sender sp
316 BLOCK_COMMENT("call Java function");
317 __ call(c_rarg1);
318
319 BLOCK_COMMENT("call_stub_return_address:");
320 return_address = __ pc();
321
322 // store result depending on type (everything that is not
323 // T_OBJECT, T_PRIMITIVE_OBJECT, T_LONG, T_FLOAT or T_DOUBLE is treated as T_INT)
324 __ movptr(r13, result);
325 Label is_long, is_float, is_double, check_prim, exit;
326 __ movl(rbx, result_type);
327 __ cmpl(rbx, T_OBJECT);
328 __ jcc(Assembler::equal, check_prim);
329 __ cmpl(rbx, T_PRIMITIVE_OBJECT);
330 __ jcc(Assembler::equal, check_prim);
331 __ cmpl(rbx, T_LONG);
332 __ jcc(Assembler::equal, is_long);
333 __ cmpl(rbx, T_FLOAT);
334 __ jcc(Assembler::equal, is_float);
335 __ cmpl(rbx, T_DOUBLE);
336 __ jcc(Assembler::equal, is_double);
337
338 // handle T_INT case
339 __ movl(Address(r13, 0), rax);
340
341 __ BIND(exit);
342
343 // pop parameters
344 __ lea(rsp, rsp_after_call);
345
346 #ifdef ASSERT
347 // verify that threads correspond
348 {
349 Label L1, L2, L3;
350 __ cmpptr(r15_thread, thread);
351 __ jcc(Assembler::equal, L1);
352 __ stop("StubRoutines::call_stub: r15_thread is corrupted");
353 __ bind(L1);
354 __ get_thread(rbx);
355 __ cmpptr(r15_thread, thread);
356 __ jcc(Assembler::equal, L2);
357 __ stop("StubRoutines::call_stub: r15_thread is modified by call");
358 __ bind(L2);
359 __ cmpptr(r15_thread, rbx);
360 __ jcc(Assembler::equal, L3);
361 __ stop("StubRoutines::call_stub: threads must correspond");
362 __ bind(L3);
363 }
364 #endif
365
366 __ pop_cont_fastpath();
367
368 // restore regs belonging to calling function
369 #ifdef _WIN64
370 // emit the restores for xmm regs
371 if (VM_Version::supports_evex()) {
372 for (int i = xmm_save_first; i <= last_reg; i++) {
373 __ vinsertf32x4(as_XMMRegister(i), as_XMMRegister(i), xmm_save(i), 0);
374 }
375 } else {
376 for (int i = xmm_save_first; i <= last_reg; i++) {
377 __ movdqu(as_XMMRegister(i), xmm_save(i));
378 }
379 }
380 #endif
381 __ movptr(r15, r15_save);
382 __ movptr(r14, r14_save);
383 __ movptr(r13, r13_save);
384 __ movptr(r12, r12_save);
385 __ movptr(rbx, rbx_save);
386
387 #ifdef _WIN64
388 __ movptr(rdi, rdi_save);
389 __ movptr(rsi, rsi_save);
390 #else
391 __ ldmxcsr(mxcsr_save);
392 #endif
393
394 // restore rsp
395 __ addptr(rsp, -rsp_after_call_off * wordSize);
396
397 // return
398 __ vzeroupper();
399 __ pop(rbp);
400 __ ret(0);
401
402 // handle return types different from T_INT
403 __ BIND(check_prim);
404 if (InlineTypeReturnedAsFields) {
405 // Check for scalarized return value
406 __ testptr(rax, 1);
407 __ jcc(Assembler::zero, is_long);
408 // Load pack handler address
409 __ andptr(rax, -2);
410 __ movptr(rax, Address(rax, InstanceKlass::adr_inlineklass_fixed_block_offset()));
411 __ movptr(rbx, Address(rax, InlineKlass::pack_handler_jobject_offset()));
412 // Call pack handler to initialize the buffer
413 __ call(rbx);
414 __ jmp(exit);
415 }
416 __ BIND(is_long);
417 __ movq(Address(r13, 0), rax);
418 __ jmp(exit);
419
420 __ BIND(is_float);
421 __ movflt(Address(r13, 0), xmm0);
422 __ jmp(exit);
423
424 __ BIND(is_double);
425 __ movdbl(Address(r13, 0), xmm0);
426 __ jmp(exit);
427
428 return start;
429 }
430
431 // Return point for a Java call if there's an exception thrown in
432 // Java code. The exception is caught and transformed into a
433 // pending exception stored in JavaThread that can be tested from
434 // within the VM.
435 //
436 // Note: Usually the parameters are removed by the callee. In case
437 // of an exception crossing an activation frame boundary, that is
438 // not the case if the callee is compiled code => need to setup the
439 // rsp.
440 //
441 // rax: exception oop
442
443 address StubGenerator::generate_catch_exception() {
444 StubCodeMark mark(this, "StubRoutines", "catch_exception");
445 address start = __ pc();
446
447 // same as in generate_call_stub():
448 const Address rsp_after_call(rbp, rsp_after_call_off * wordSize);
449 const Address thread (rbp, thread_off * wordSize);
450
451 #ifdef ASSERT
452 // verify that threads correspond
453 {
454 Label L1, L2, L3;
455 __ cmpptr(r15_thread, thread);
456 __ jcc(Assembler::equal, L1);
457 __ stop("StubRoutines::catch_exception: r15_thread is corrupted");
458 __ bind(L1);
459 __ get_thread(rbx);
460 __ cmpptr(r15_thread, thread);
461 __ jcc(Assembler::equal, L2);
462 __ stop("StubRoutines::catch_exception: r15_thread is modified by call");
463 __ bind(L2);
464 __ cmpptr(r15_thread, rbx);
465 __ jcc(Assembler::equal, L3);
466 __ stop("StubRoutines::catch_exception: threads must correspond");
467 __ bind(L3);
468 }
469 #endif
470
471 // set pending exception
472 __ verify_oop(rax);
473
474 __ movptr(Address(r15_thread, Thread::pending_exception_offset()), rax);
475 __ lea(rscratch1, ExternalAddress((address)__FILE__));
476 __ movptr(Address(r15_thread, Thread::exception_file_offset()), rscratch1);
477 __ movl(Address(r15_thread, Thread::exception_line_offset()), (int) __LINE__);
478
479 // complete return to VM
480 assert(StubRoutines::_call_stub_return_address != NULL,
481 "_call_stub_return_address must have been generated before");
482 __ jump(RuntimeAddress(StubRoutines::_call_stub_return_address));
483
484 return start;
485 }
486
487 // Continuation point for runtime calls returning with a pending
488 // exception. The pending exception check happened in the runtime
489 // or native call stub. The pending exception in Thread is
490 // converted into a Java-level exception.
491 //
492 // Contract with Java-level exception handlers:
493 // rax: exception
494 // rdx: throwing pc
495 //
496 // NOTE: At entry of this stub, exception-pc must be on stack !!
497
498 address StubGenerator::generate_forward_exception() {
499 StubCodeMark mark(this, "StubRoutines", "forward exception");
500 address start = __ pc();
501
502 // Upon entry, the sp points to the return address returning into
503 // Java (interpreted or compiled) code; i.e., the return address
504 // becomes the throwing pc.
505 //
506 // Arguments pushed before the runtime call are still on the stack
507 // but the exception handler will reset the stack pointer ->
508 // ignore them. A potential result in registers can be ignored as
509 // well.
510
511 #ifdef ASSERT
512 // make sure this code is only executed if there is a pending exception
513 {
514 Label L;
515 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), NULL_WORD);
516 __ jcc(Assembler::notEqual, L);
517 __ stop("StubRoutines::forward exception: no pending exception (1)");
518 __ bind(L);
519 }
520 #endif
521
522 // compute exception handler into rbx
523 __ movptr(c_rarg0, Address(rsp, 0));
524 BLOCK_COMMENT("call exception_handler_for_return_address");
525 __ call_VM_leaf(CAST_FROM_FN_PTR(address,
526 SharedRuntime::exception_handler_for_return_address),
527 r15_thread, c_rarg0);
528 __ mov(rbx, rax);
529
530 // setup rax & rdx, remove return address & clear pending exception
531 __ pop(rdx);
532 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset()));
533 __ movptr(Address(r15_thread, Thread::pending_exception_offset()), NULL_WORD);
534
535 #ifdef ASSERT
536 // make sure exception is set
537 {
538 Label L;
539 __ testptr(rax, rax);
540 __ jcc(Assembler::notEqual, L);
541 __ stop("StubRoutines::forward exception: no pending exception (2)");
542 __ bind(L);
543 }
544 #endif
545
546 // continue at exception handler (return address removed)
547 // rax: exception
548 // rbx: exception handler
549 // rdx: throwing pc
550 __ verify_oop(rax);
551 __ jmp(rbx);
552
553 return start;
554 }
555
556 // Support for intptr_t OrderAccess::fence()
557 //
558 // Arguments :
559 //
560 // Result:
561 address StubGenerator::generate_orderaccess_fence() {
562 StubCodeMark mark(this, "StubRoutines", "orderaccess_fence");
563 address start = __ pc();
564
565 __ membar(Assembler::StoreLoad);
566 __ ret(0);
567
568 return start;
569 }
570
571
572 // Support for intptr_t get_previous_sp()
573 //
574 // This routine is used to find the previous stack pointer for the
575 // caller.
576 address StubGenerator::generate_get_previous_sp() {
577 StubCodeMark mark(this, "StubRoutines", "get_previous_sp");
578 address start = __ pc();
579
580 __ movptr(rax, rsp);
581 __ addptr(rax, 8); // return address is at the top of the stack.
582 __ ret(0);
583
584 return start;
585 }
586
587 //----------------------------------------------------------------------------------------------------
588 // Support for void verify_mxcsr()
589 //
590 // This routine is used with -Xcheck:jni to verify that native
591 // JNI code does not return to Java code without restoring the
592 // MXCSR register to our expected state.
593
594 address StubGenerator::generate_verify_mxcsr() {
595 StubCodeMark mark(this, "StubRoutines", "verify_mxcsr");
596 address start = __ pc();
597
598 const Address mxcsr_save(rsp, 0);
599
600 if (CheckJNICalls) {
601 Label ok_ret;
602 ExternalAddress mxcsr_std(StubRoutines::x86::addr_mxcsr_std());
603 __ push(rax);
604 __ subptr(rsp, wordSize); // allocate a temp location
605 __ stmxcsr(mxcsr_save);
606 __ movl(rax, mxcsr_save);
607 __ andl(rax, 0xFFC0); // Mask out any pending exceptions (only check control and mask bits)
608 __ cmp32(rax, mxcsr_std, rscratch1);
609 __ jcc(Assembler::equal, ok_ret);
610
611 __ warn("MXCSR changed by native JNI code, use -XX:+RestoreMXCSROnJNICall");
612
613 __ ldmxcsr(mxcsr_std, rscratch1);
614
615 __ bind(ok_ret);
616 __ addptr(rsp, wordSize);
617 __ pop(rax);
618 }
619
620 __ ret(0);
621
622 return start;
623 }
624
625 address StubGenerator::generate_f2i_fixup() {
626 StubCodeMark mark(this, "StubRoutines", "f2i_fixup");
627 Address inout(rsp, 5 * wordSize); // return address + 4 saves
628
629 address start = __ pc();
630
631 Label L;
632
633 __ push(rax);
634 __ push(c_rarg3);
635 __ push(c_rarg2);
636 __ push(c_rarg1);
637
638 __ movl(rax, 0x7f800000);
639 __ xorl(c_rarg3, c_rarg3);
640 __ movl(c_rarg2, inout);
641 __ movl(c_rarg1, c_rarg2);
642 __ andl(c_rarg1, 0x7fffffff);
643 __ cmpl(rax, c_rarg1); // NaN? -> 0
644 __ jcc(Assembler::negative, L);
645 __ testl(c_rarg2, c_rarg2); // signed ? min_jint : max_jint
646 __ movl(c_rarg3, 0x80000000);
647 __ movl(rax, 0x7fffffff);
648 __ cmovl(Assembler::positive, c_rarg3, rax);
649
650 __ bind(L);
651 __ movptr(inout, c_rarg3);
652
653 __ pop(c_rarg1);
654 __ pop(c_rarg2);
655 __ pop(c_rarg3);
656 __ pop(rax);
657
658 __ ret(0);
659
660 return start;
661 }
662
663 address StubGenerator::generate_f2l_fixup() {
664 StubCodeMark mark(this, "StubRoutines", "f2l_fixup");
665 Address inout(rsp, 5 * wordSize); // return address + 4 saves
666 address start = __ pc();
667
668 Label L;
669
670 __ push(rax);
671 __ push(c_rarg3);
672 __ push(c_rarg2);
673 __ push(c_rarg1);
674
675 __ movl(rax, 0x7f800000);
676 __ xorl(c_rarg3, c_rarg3);
677 __ movl(c_rarg2, inout);
678 __ movl(c_rarg1, c_rarg2);
679 __ andl(c_rarg1, 0x7fffffff);
680 __ cmpl(rax, c_rarg1); // NaN? -> 0
681 __ jcc(Assembler::negative, L);
682 __ testl(c_rarg2, c_rarg2); // signed ? min_jlong : max_jlong
683 __ mov64(c_rarg3, 0x8000000000000000);
684 __ mov64(rax, 0x7fffffffffffffff);
685 __ cmov(Assembler::positive, c_rarg3, rax);
686
687 __ bind(L);
688 __ movptr(inout, c_rarg3);
689
690 __ pop(c_rarg1);
691 __ pop(c_rarg2);
692 __ pop(c_rarg3);
693 __ pop(rax);
694
695 __ ret(0);
696
697 return start;
698 }
699
700 address StubGenerator::generate_d2i_fixup() {
701 StubCodeMark mark(this, "StubRoutines", "d2i_fixup");
702 Address inout(rsp, 6 * wordSize); // return address + 5 saves
703
704 address start = __ pc();
705
706 Label L;
707
708 __ push(rax);
709 __ push(c_rarg3);
710 __ push(c_rarg2);
711 __ push(c_rarg1);
712 __ push(c_rarg0);
713
714 __ movl(rax, 0x7ff00000);
715 __ movq(c_rarg2, inout);
716 __ movl(c_rarg3, c_rarg2);
717 __ mov(c_rarg1, c_rarg2);
718 __ mov(c_rarg0, c_rarg2);
719 __ negl(c_rarg3);
720 __ shrptr(c_rarg1, 0x20);
721 __ orl(c_rarg3, c_rarg2);
722 __ andl(c_rarg1, 0x7fffffff);
723 __ xorl(c_rarg2, c_rarg2);
724 __ shrl(c_rarg3, 0x1f);
725 __ orl(c_rarg1, c_rarg3);
726 __ cmpl(rax, c_rarg1);
727 __ jcc(Assembler::negative, L); // NaN -> 0
728 __ testptr(c_rarg0, c_rarg0); // signed ? min_jint : max_jint
729 __ movl(c_rarg2, 0x80000000);
730 __ movl(rax, 0x7fffffff);
731 __ cmov(Assembler::positive, c_rarg2, rax);
732
733 __ bind(L);
734 __ movptr(inout, c_rarg2);
735
736 __ pop(c_rarg0);
737 __ pop(c_rarg1);
738 __ pop(c_rarg2);
739 __ pop(c_rarg3);
740 __ pop(rax);
741
742 __ ret(0);
743
744 return start;
745 }
746
747 address StubGenerator::generate_d2l_fixup() {
748 StubCodeMark mark(this, "StubRoutines", "d2l_fixup");
749 Address inout(rsp, 6 * wordSize); // return address + 5 saves
750
751 address start = __ pc();
752
753 Label L;
754
755 __ push(rax);
756 __ push(c_rarg3);
757 __ push(c_rarg2);
758 __ push(c_rarg1);
759 __ push(c_rarg0);
760
761 __ movl(rax, 0x7ff00000);
762 __ movq(c_rarg2, inout);
763 __ movl(c_rarg3, c_rarg2);
764 __ mov(c_rarg1, c_rarg2);
765 __ mov(c_rarg0, c_rarg2);
766 __ negl(c_rarg3);
767 __ shrptr(c_rarg1, 0x20);
768 __ orl(c_rarg3, c_rarg2);
769 __ andl(c_rarg1, 0x7fffffff);
770 __ xorl(c_rarg2, c_rarg2);
771 __ shrl(c_rarg3, 0x1f);
772 __ orl(c_rarg1, c_rarg3);
773 __ cmpl(rax, c_rarg1);
774 __ jcc(Assembler::negative, L); // NaN -> 0
775 __ testq(c_rarg0, c_rarg0); // signed ? min_jlong : max_jlong
776 __ mov64(c_rarg2, 0x8000000000000000);
777 __ mov64(rax, 0x7fffffffffffffff);
778 __ cmovq(Assembler::positive, c_rarg2, rax);
779
780 __ bind(L);
781 __ movq(inout, c_rarg2);
782
783 __ pop(c_rarg0);
784 __ pop(c_rarg1);
785 __ pop(c_rarg2);
786 __ pop(c_rarg3);
787 __ pop(rax);
788
789 __ ret(0);
790
791 return start;
792 }
793
794 address StubGenerator::generate_count_leading_zeros_lut(const char *stub_name) {
795 __ align64();
796 StubCodeMark mark(this, "StubRoutines", stub_name);
797 address start = __ pc();
798
799 __ emit_data64(0x0101010102020304, relocInfo::none);
800 __ emit_data64(0x0000000000000000, relocInfo::none);
801 __ emit_data64(0x0101010102020304, relocInfo::none);
802 __ emit_data64(0x0000000000000000, relocInfo::none);
803 __ emit_data64(0x0101010102020304, relocInfo::none);
804 __ emit_data64(0x0000000000000000, relocInfo::none);
805 __ emit_data64(0x0101010102020304, relocInfo::none);
806 __ emit_data64(0x0000000000000000, relocInfo::none);
807
808 return start;
809 }
810
811 address StubGenerator::generate_popcount_avx_lut(const char *stub_name) {
812 __ align64();
813 StubCodeMark mark(this, "StubRoutines", stub_name);
814 address start = __ pc();
815
816 __ emit_data64(0x0302020102010100, relocInfo::none);
817 __ emit_data64(0x0403030203020201, relocInfo::none);
818 __ emit_data64(0x0302020102010100, relocInfo::none);
819 __ emit_data64(0x0403030203020201, relocInfo::none);
820 __ emit_data64(0x0302020102010100, relocInfo::none);
821 __ emit_data64(0x0403030203020201, relocInfo::none);
822 __ emit_data64(0x0302020102010100, relocInfo::none);
823 __ emit_data64(0x0403030203020201, relocInfo::none);
824
825 return start;
826 }
827
828 address StubGenerator::generate_iota_indices(const char *stub_name) {
829 __ align(CodeEntryAlignment);
830 StubCodeMark mark(this, "StubRoutines", stub_name);
831 address start = __ pc();
832 // B
833 __ emit_data64(0x0706050403020100, relocInfo::none);
834 __ emit_data64(0x0F0E0D0C0B0A0908, relocInfo::none);
835 __ emit_data64(0x1716151413121110, relocInfo::none);
836 __ emit_data64(0x1F1E1D1C1B1A1918, relocInfo::none);
837 __ emit_data64(0x2726252423222120, relocInfo::none);
838 __ emit_data64(0x2F2E2D2C2B2A2928, relocInfo::none);
839 __ emit_data64(0x3736353433323130, relocInfo::none);
840 __ emit_data64(0x3F3E3D3C3B3A3938, relocInfo::none);
841 // W
842 __ emit_data64(0x0003000200010000, relocInfo::none);
843 __ emit_data64(0x0007000600050004, relocInfo::none);
844 __ emit_data64(0x000B000A00090008, relocInfo::none);
845 __ emit_data64(0x000F000E000D000C, relocInfo::none);
846 __ emit_data64(0x0013001200110010, relocInfo::none);
847 __ emit_data64(0x0017001600150014, relocInfo::none);
848 __ emit_data64(0x001B001A00190018, relocInfo::none);
849 __ emit_data64(0x001F001E001D001C, relocInfo::none);
850 // D
851 __ emit_data64(0x0000000100000000, relocInfo::none);
852 __ emit_data64(0x0000000300000002, relocInfo::none);
853 __ emit_data64(0x0000000500000004, relocInfo::none);
854 __ emit_data64(0x0000000700000006, relocInfo::none);
855 __ emit_data64(0x0000000900000008, relocInfo::none);
856 __ emit_data64(0x0000000B0000000A, relocInfo::none);
857 __ emit_data64(0x0000000D0000000C, relocInfo::none);
858 __ emit_data64(0x0000000F0000000E, relocInfo::none);
859 // Q
860 __ emit_data64(0x0000000000000000, relocInfo::none);
861 __ emit_data64(0x0000000000000001, relocInfo::none);
862 __ emit_data64(0x0000000000000002, relocInfo::none);
863 __ emit_data64(0x0000000000000003, relocInfo::none);
864 __ emit_data64(0x0000000000000004, relocInfo::none);
865 __ emit_data64(0x0000000000000005, relocInfo::none);
866 __ emit_data64(0x0000000000000006, relocInfo::none);
867 __ emit_data64(0x0000000000000007, relocInfo::none);
868 // D - FP
869 __ emit_data64(0x3F80000000000000, relocInfo::none); // 0.0f, 1.0f
870 __ emit_data64(0x4040000040000000, relocInfo::none); // 2.0f, 3.0f
871 __ emit_data64(0x40A0000040800000, relocInfo::none); // 4.0f, 5.0f
872 __ emit_data64(0x40E0000040C00000, relocInfo::none); // 6.0f, 7.0f
873 __ emit_data64(0x4110000041000000, relocInfo::none); // 8.0f, 9.0f
874 __ emit_data64(0x4130000041200000, relocInfo::none); // 10.0f, 11.0f
875 __ emit_data64(0x4150000041400000, relocInfo::none); // 12.0f, 13.0f
876 __ emit_data64(0x4170000041600000, relocInfo::none); // 14.0f, 15.0f
877 // Q - FP
878 __ emit_data64(0x0000000000000000, relocInfo::none); // 0.0d
879 __ emit_data64(0x3FF0000000000000, relocInfo::none); // 1.0d
880 __ emit_data64(0x4000000000000000, relocInfo::none); // 2.0d
881 __ emit_data64(0x4008000000000000, relocInfo::none); // 3.0d
882 __ emit_data64(0x4010000000000000, relocInfo::none); // 4.0d
883 __ emit_data64(0x4014000000000000, relocInfo::none); // 5.0d
884 __ emit_data64(0x4018000000000000, relocInfo::none); // 6.0d
885 __ emit_data64(0x401c000000000000, relocInfo::none); // 7.0d
886 return start;
887 }
888
889 address StubGenerator::generate_vector_reverse_bit_lut(const char *stub_name) {
890 __ align(CodeEntryAlignment);
891 StubCodeMark mark(this, "StubRoutines", stub_name);
892 address start = __ pc();
893
894 __ emit_data64(0x0E060A020C040800, relocInfo::none);
895 __ emit_data64(0x0F070B030D050901, relocInfo::none);
896 __ emit_data64(0x0E060A020C040800, relocInfo::none);
897 __ emit_data64(0x0F070B030D050901, relocInfo::none);
898 __ emit_data64(0x0E060A020C040800, relocInfo::none);
899 __ emit_data64(0x0F070B030D050901, relocInfo::none);
900 __ emit_data64(0x0E060A020C040800, relocInfo::none);
901 __ emit_data64(0x0F070B030D050901, relocInfo::none);
902
903 return start;
904 }
905
906 address StubGenerator::generate_vector_reverse_byte_perm_mask_long(const char *stub_name) {
907 __ align(CodeEntryAlignment);
908 StubCodeMark mark(this, "StubRoutines", stub_name);
909 address start = __ pc();
910
911 __ emit_data64(0x0001020304050607, relocInfo::none);
912 __ emit_data64(0x08090A0B0C0D0E0F, relocInfo::none);
913 __ emit_data64(0x0001020304050607, relocInfo::none);
914 __ emit_data64(0x08090A0B0C0D0E0F, relocInfo::none);
915 __ emit_data64(0x0001020304050607, relocInfo::none);
916 __ emit_data64(0x08090A0B0C0D0E0F, relocInfo::none);
917 __ emit_data64(0x0001020304050607, relocInfo::none);
918 __ emit_data64(0x08090A0B0C0D0E0F, relocInfo::none);
919
920 return start;
921 }
922
923 address StubGenerator::generate_vector_reverse_byte_perm_mask_int(const char *stub_name) {
924 __ align(CodeEntryAlignment);
925 StubCodeMark mark(this, "StubRoutines", stub_name);
926 address start = __ pc();
927
928 __ emit_data64(0x0405060700010203, relocInfo::none);
929 __ emit_data64(0x0C0D0E0F08090A0B, relocInfo::none);
930 __ emit_data64(0x0405060700010203, relocInfo::none);
931 __ emit_data64(0x0C0D0E0F08090A0B, relocInfo::none);
932 __ emit_data64(0x0405060700010203, relocInfo::none);
933 __ emit_data64(0x0C0D0E0F08090A0B, relocInfo::none);
934 __ emit_data64(0x0405060700010203, relocInfo::none);
935 __ emit_data64(0x0C0D0E0F08090A0B, relocInfo::none);
936
937 return start;
938 }
939
940 address StubGenerator::generate_vector_reverse_byte_perm_mask_short(const char *stub_name) {
941 __ align(CodeEntryAlignment);
942 StubCodeMark mark(this, "StubRoutines", stub_name);
943 address start = __ pc();
944
945 __ emit_data64(0x0607040502030001, relocInfo::none);
946 __ emit_data64(0x0E0F0C0D0A0B0809, relocInfo::none);
947 __ emit_data64(0x0607040502030001, relocInfo::none);
948 __ emit_data64(0x0E0F0C0D0A0B0809, relocInfo::none);
949 __ emit_data64(0x0607040502030001, relocInfo::none);
950 __ emit_data64(0x0E0F0C0D0A0B0809, relocInfo::none);
951 __ emit_data64(0x0607040502030001, relocInfo::none);
952 __ emit_data64(0x0E0F0C0D0A0B0809, relocInfo::none);
953
954 return start;
955 }
956
957 address StubGenerator::generate_vector_byte_shuffle_mask(const char *stub_name) {
958 __ align(CodeEntryAlignment);
959 StubCodeMark mark(this, "StubRoutines", stub_name);
960 address start = __ pc();
961
962 __ emit_data64(0x7070707070707070, relocInfo::none);
963 __ emit_data64(0x7070707070707070, relocInfo::none);
964 __ emit_data64(0xF0F0F0F0F0F0F0F0, relocInfo::none);
965 __ emit_data64(0xF0F0F0F0F0F0F0F0, relocInfo::none);
966
967 return start;
968 }
969
970 address StubGenerator::generate_fp_mask(const char *stub_name, int64_t mask) {
971 __ align(CodeEntryAlignment);
972 StubCodeMark mark(this, "StubRoutines", stub_name);
973 address start = __ pc();
974
975 __ emit_data64( mask, relocInfo::none );
976 __ emit_data64( mask, relocInfo::none );
977
978 return start;
979 }
980
981 address StubGenerator::generate_vector_mask(const char *stub_name, int64_t mask) {
982 __ align(CodeEntryAlignment);
983 StubCodeMark mark(this, "StubRoutines", stub_name);
984 address start = __ pc();
985
986 __ emit_data64(mask, relocInfo::none);
987 __ emit_data64(mask, relocInfo::none);
988 __ emit_data64(mask, relocInfo::none);
989 __ emit_data64(mask, relocInfo::none);
990 __ emit_data64(mask, relocInfo::none);
991 __ emit_data64(mask, relocInfo::none);
992 __ emit_data64(mask, relocInfo::none);
993 __ emit_data64(mask, relocInfo::none);
994
995 return start;
996 }
997
998 address StubGenerator::generate_vector_byte_perm_mask(const char *stub_name) {
999 __ align(CodeEntryAlignment);
1000 StubCodeMark mark(this, "StubRoutines", stub_name);
1001 address start = __ pc();
1002
1003 __ emit_data64(0x0000000000000001, relocInfo::none);
1004 __ emit_data64(0x0000000000000003, relocInfo::none);
1005 __ emit_data64(0x0000000000000005, relocInfo::none);
1006 __ emit_data64(0x0000000000000007, relocInfo::none);
1007 __ emit_data64(0x0000000000000000, relocInfo::none);
1008 __ emit_data64(0x0000000000000002, relocInfo::none);
1009 __ emit_data64(0x0000000000000004, relocInfo::none);
1010 __ emit_data64(0x0000000000000006, relocInfo::none);
1011
1012 return start;
1013 }
1014
1015 address StubGenerator::generate_vector_fp_mask(const char *stub_name, int64_t mask) {
1016 __ align(CodeEntryAlignment);
1017 StubCodeMark mark(this, "StubRoutines", stub_name);
1018 address start = __ pc();
1019
1020 __ emit_data64(mask, relocInfo::none);
1021 __ emit_data64(mask, relocInfo::none);
1022 __ emit_data64(mask, relocInfo::none);
1023 __ emit_data64(mask, relocInfo::none);
1024 __ emit_data64(mask, relocInfo::none);
1025 __ emit_data64(mask, relocInfo::none);
1026 __ emit_data64(mask, relocInfo::none);
1027 __ emit_data64(mask, relocInfo::none);
1028
1029 return start;
1030 }
1031
1032 address StubGenerator::generate_vector_custom_i32(const char *stub_name, Assembler::AvxVectorLen len,
1033 int32_t val0, int32_t val1, int32_t val2, int32_t val3,
1034 int32_t val4, int32_t val5, int32_t val6, int32_t val7,
1035 int32_t val8, int32_t val9, int32_t val10, int32_t val11,
1036 int32_t val12, int32_t val13, int32_t val14, int32_t val15) {
1037 __ align(CodeEntryAlignment);
1038 StubCodeMark mark(this, "StubRoutines", stub_name);
1039 address start = __ pc();
1040
1041 assert(len != Assembler::AVX_NoVec, "vector len must be specified");
1042 __ emit_data(val0, relocInfo::none, 0);
1043 __ emit_data(val1, relocInfo::none, 0);
1044 __ emit_data(val2, relocInfo::none, 0);
1045 __ emit_data(val3, relocInfo::none, 0);
1046 if (len >= Assembler::AVX_256bit) {
1047 __ emit_data(val4, relocInfo::none, 0);
1048 __ emit_data(val5, relocInfo::none, 0);
1049 __ emit_data(val6, relocInfo::none, 0);
1050 __ emit_data(val7, relocInfo::none, 0);
1051 if (len >= Assembler::AVX_512bit) {
1052 __ emit_data(val8, relocInfo::none, 0);
1053 __ emit_data(val9, relocInfo::none, 0);
1054 __ emit_data(val10, relocInfo::none, 0);
1055 __ emit_data(val11, relocInfo::none, 0);
1056 __ emit_data(val12, relocInfo::none, 0);
1057 __ emit_data(val13, relocInfo::none, 0);
1058 __ emit_data(val14, relocInfo::none, 0);
1059 __ emit_data(val15, relocInfo::none, 0);
1060 }
1061 }
1062 return start;
1063 }
1064
1065 // Non-destructive plausibility checks for oops
1066 //
1067 // Arguments:
1068 // all args on stack!
1069 //
1070 // Stack after saving c_rarg3:
1071 // [tos + 0]: saved c_rarg3
1072 // [tos + 1]: saved c_rarg2
1073 // [tos + 2]: saved r12 (several TemplateTable methods use it)
1074 // [tos + 3]: saved flags
1075 // [tos + 4]: return address
1076 // * [tos + 5]: error message (char*)
1077 // * [tos + 6]: object to verify (oop)
1078 // * [tos + 7]: saved rax - saved by caller and bashed
1079 // * [tos + 8]: saved r10 (rscratch1) - saved by caller
1080 // * = popped on exit
1081 address StubGenerator::generate_verify_oop() {
1082 StubCodeMark mark(this, "StubRoutines", "verify_oop");
1083 address start = __ pc();
1084
1085 Label exit, error;
1086
1087 __ pushf();
1088 __ incrementl(ExternalAddress((address) StubRoutines::verify_oop_count_addr()), rscratch1);
1089
1090 __ push(r12);
1091
1092 // save c_rarg2 and c_rarg3
1093 __ push(c_rarg2);
1094 __ push(c_rarg3);
1095
1096 enum {
1097 // After previous pushes.
1098 oop_to_verify = 6 * wordSize,
1099 saved_rax = 7 * wordSize,
1100 saved_r10 = 8 * wordSize,
1101
1102 // Before the call to MacroAssembler::debug(), see below.
1103 return_addr = 16 * wordSize,
1104 error_msg = 17 * wordSize
1105 };
1106
1107 // get object
1108 __ movptr(rax, Address(rsp, oop_to_verify));
1109
1110 // make sure object is 'reasonable'
1111 __ testptr(rax, rax);
1112 __ jcc(Assembler::zero, exit); // if obj is NULL it is OK
1113
1114 #if INCLUDE_ZGC
1115 if (UseZGC) {
1116 // Check if metadata bits indicate a bad oop
1117 __ testptr(rax, Address(r15_thread, ZThreadLocalData::address_bad_mask_offset()));
1118 __ jcc(Assembler::notZero, error);
1119 }
1120 #endif
1121
1122 // Check if the oop is in the right area of memory
1123 __ movptr(c_rarg2, rax);
1124 __ movptr(c_rarg3, (intptr_t) Universe::verify_oop_mask());
1125 __ andptr(c_rarg2, c_rarg3);
1126 __ movptr(c_rarg3, (intptr_t) Universe::verify_oop_bits());
1127 __ cmpptr(c_rarg2, c_rarg3);
1128 __ jcc(Assembler::notZero, error);
1129
1130 // make sure klass is 'reasonable', which is not zero.
1131 __ load_klass(rax, rax, rscratch1); // get klass
1132 __ testptr(rax, rax);
1133 __ jcc(Assembler::zero, error); // if klass is NULL it is broken
1134
1135 // return if everything seems ok
1136 __ bind(exit);
1137 __ movptr(rax, Address(rsp, saved_rax)); // get saved rax back
1138 __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back
1139 __ pop(c_rarg3); // restore c_rarg3
1140 __ pop(c_rarg2); // restore c_rarg2
1141 __ pop(r12); // restore r12
1142 __ popf(); // restore flags
1143 __ ret(4 * wordSize); // pop caller saved stuff
1144
1145 // handle errors
1146 __ bind(error);
1147 __ movptr(rax, Address(rsp, saved_rax)); // get saved rax back
1148 __ movptr(rscratch1, Address(rsp, saved_r10)); // get saved r10 back
1149 __ pop(c_rarg3); // get saved c_rarg3 back
1150 __ pop(c_rarg2); // get saved c_rarg2 back
1151 __ pop(r12); // get saved r12 back
1152 __ popf(); // get saved flags off stack --
1153 // will be ignored
1154
1155 __ pusha(); // push registers
1156 // (rip is already
1157 // already pushed)
1158 // debug(char* msg, int64_t pc, int64_t regs[])
1159 // We've popped the registers we'd saved (c_rarg3, c_rarg2 and flags), and
1160 // pushed all the registers, so now the stack looks like:
1161 // [tos + 0] 16 saved registers
1162 // [tos + 16] return address
1163 // * [tos + 17] error message (char*)
1164 // * [tos + 18] object to verify (oop)
1165 // * [tos + 19] saved rax - saved by caller and bashed
1166 // * [tos + 20] saved r10 (rscratch1) - saved by caller
1167 // * = popped on exit
1168
1169 __ movptr(c_rarg0, Address(rsp, error_msg)); // pass address of error message
1170 __ movptr(c_rarg1, Address(rsp, return_addr)); // pass return address
1171 __ movq(c_rarg2, rsp); // pass address of regs on stack
1172 __ mov(r12, rsp); // remember rsp
1173 __ subptr(rsp, frame::arg_reg_save_area_bytes); // windows
1174 __ andptr(rsp, -16); // align stack as required by ABI
1175 BLOCK_COMMENT("call MacroAssembler::debug");
1176 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, MacroAssembler::debug64)));
1177 __ hlt();
1178
1179 return start;
1180 }
1181
1182
1183 // Shuffle first three arg regs on Windows into Linux/Solaris locations.
1184 //
1185 // Outputs:
1186 // rdi - rcx
1187 // rsi - rdx
1188 // rdx - r8
1189 // rcx - r9
1190 //
1191 // Registers r9 and r10 are used to save rdi and rsi on Windows, which latter
1192 // are non-volatile. r9 and r10 should not be used by the caller.
1193 //
1194 void StubGenerator::setup_arg_regs(int nargs) {
1195 const Register saved_rdi = r9;
1196 const Register saved_rsi = r10;
1197 assert(nargs == 3 || nargs == 4, "else fix");
1198 #ifdef _WIN64
1199 assert(c_rarg0 == rcx && c_rarg1 == rdx && c_rarg2 == r8 && c_rarg3 == r9,
1200 "unexpected argument registers");
1201 if (nargs >= 4)
1202 __ mov(rax, r9); // r9 is also saved_rdi
1203 __ movptr(saved_rdi, rdi);
1204 __ movptr(saved_rsi, rsi);
1205 __ mov(rdi, rcx); // c_rarg0
1206 __ mov(rsi, rdx); // c_rarg1
1207 __ mov(rdx, r8); // c_rarg2
1208 if (nargs >= 4)
1209 __ mov(rcx, rax); // c_rarg3 (via rax)
1210 #else
1211 assert(c_rarg0 == rdi && c_rarg1 == rsi && c_rarg2 == rdx && c_rarg3 == rcx,
1212 "unexpected argument registers");
1213 #endif
1214 DEBUG_ONLY(_regs_in_thread = false;)
1215 }
1216
1217
1218 void StubGenerator::restore_arg_regs() {
1219 assert(!_regs_in_thread, "wrong call to restore_arg_regs");
1220 const Register saved_rdi = r9;
1221 const Register saved_rsi = r10;
1222 #ifdef _WIN64
1223 __ movptr(rdi, saved_rdi);
1224 __ movptr(rsi, saved_rsi);
1225 #endif
1226 }
1227
1228
1229 // This is used in places where r10 is a scratch register, and can
1230 // be adapted if r9 is needed also.
1231 void StubGenerator::setup_arg_regs_using_thread() {
1232 const Register saved_r15 = r9;
1233 #ifdef _WIN64
1234 __ mov(saved_r15, r15); // r15 is callee saved and needs to be restored
1235 __ get_thread(r15_thread);
1236 assert(c_rarg0 == rcx && c_rarg1 == rdx && c_rarg2 == r8 && c_rarg3 == r9,
1237 "unexpected argument registers");
1238 __ movptr(Address(r15_thread, in_bytes(JavaThread::windows_saved_rdi_offset())), rdi);
1239 __ movptr(Address(r15_thread, in_bytes(JavaThread::windows_saved_rsi_offset())), rsi);
1240
1241 __ mov(rdi, rcx); // c_rarg0
1242 __ mov(rsi, rdx); // c_rarg1
1243 __ mov(rdx, r8); // c_rarg2
1244 #else
1245 assert(c_rarg0 == rdi && c_rarg1 == rsi && c_rarg2 == rdx && c_rarg3 == rcx,
1246 "unexpected argument registers");
1247 #endif
1248 DEBUG_ONLY(_regs_in_thread = true;)
1249 }
1250
1251
1252 void StubGenerator::restore_arg_regs_using_thread() {
1253 assert(_regs_in_thread, "wrong call to restore_arg_regs");
1254 const Register saved_r15 = r9;
1255 #ifdef _WIN64
1256 __ get_thread(r15_thread);
1257 __ movptr(rsi, Address(r15_thread, in_bytes(JavaThread::windows_saved_rsi_offset())));
1258 __ movptr(rdi, Address(r15_thread, in_bytes(JavaThread::windows_saved_rdi_offset())));
1259 __ mov(r15, saved_r15); // r15 is callee saved and needs to be restored
1260 #endif
1261 }
1262
1263
1264 void StubGenerator::setup_argument_regs(BasicType type) {
1265 if (type == T_BYTE || type == T_SHORT) {
1266 setup_arg_regs(); // from => rdi, to => rsi, count => rdx
1267 // r9 and r10 may be used to save non-volatile registers
1268 } else {
1269 setup_arg_regs_using_thread(); // from => rdi, to => rsi, count => rdx
1270 // r9 is used to save r15_thread
1271 }
1272 }
1273
1274
1275 void StubGenerator::restore_argument_regs(BasicType type) {
1276 if (type == T_BYTE || type == T_SHORT) {
1277 restore_arg_regs();
1278 } else {
1279 restore_arg_regs_using_thread();
1280 }
1281 }
1282
1283 address StubGenerator::generate_data_cache_writeback() {
1284 const Register src = c_rarg0; // source address
1285
1286 __ align(CodeEntryAlignment);
1287
1288 StubCodeMark mark(this, "StubRoutines", "_data_cache_writeback");
1289
1290 address start = __ pc();
1291
1292 __ enter();
1293 __ cache_wb(Address(src, 0));
1294 __ leave();
1295 __ ret(0);
1296
1297 return start;
1298 }
1299
1300 address StubGenerator::generate_data_cache_writeback_sync() {
1301 const Register is_pre = c_rarg0; // pre or post sync
1302
1303 __ align(CodeEntryAlignment);
1304
1305 StubCodeMark mark(this, "StubRoutines", "_data_cache_writeback_sync");
1306
1307 // pre wbsync is a no-op
1308 // post wbsync translates to an sfence
1309
1310 Label skip;
1311 address start = __ pc();
1312
1313 __ enter();
1314 __ cmpl(is_pre, 0);
1315 __ jcc(Assembler::notEqual, skip);
1316 __ cache_wbsync(false);
1317 __ bind(skip);
1318 __ leave();
1319 __ ret(0);
1320
1321 return start;
1322 }
1323
1324 // ofs and limit are use for multi-block byte array.
1325 // int com.sun.security.provider.MD5.implCompress(byte[] b, int ofs)
1326 address StubGenerator::generate_md5_implCompress(bool multi_block, const char *name) {
1327 __ align(CodeEntryAlignment);
1328 StubCodeMark mark(this, "StubRoutines", name);
1329 address start = __ pc();
1330
1331 const Register buf_param = r15;
1332 const Address state_param(rsp, 0 * wordSize);
1333 const Address ofs_param (rsp, 1 * wordSize );
1334 const Address limit_param(rsp, 1 * wordSize + 4);
1335
1336 __ enter();
1337 __ push(rbx);
1338 __ push(rdi);
1339 __ push(rsi);
1340 __ push(r15);
1341 __ subptr(rsp, 2 * wordSize);
1342
1343 __ movptr(buf_param, c_rarg0);
1344 __ movptr(state_param, c_rarg1);
1345 if (multi_block) {
1346 __ movl(ofs_param, c_rarg2);
1347 __ movl(limit_param, c_rarg3);
1348 }
1349 __ fast_md5(buf_param, state_param, ofs_param, limit_param, multi_block);
1350
1351 __ addptr(rsp, 2 * wordSize);
1352 __ pop(r15);
1353 __ pop(rsi);
1354 __ pop(rdi);
1355 __ pop(rbx);
1356 __ leave();
1357 __ ret(0);
1358
1359 return start;
1360 }
1361
1362 address StubGenerator::generate_upper_word_mask() {
1363 __ align64();
1364 StubCodeMark mark(this, "StubRoutines", "upper_word_mask");
1365 address start = __ pc();
1366
1367 __ emit_data64(0x0000000000000000, relocInfo::none);
1368 __ emit_data64(0xFFFFFFFF00000000, relocInfo::none);
1369
1370 return start;
1371 }
1372
1373 address StubGenerator::generate_shuffle_byte_flip_mask() {
1374 __ align64();
1375 StubCodeMark mark(this, "StubRoutines", "shuffle_byte_flip_mask");
1376 address start = __ pc();
1377
1378 __ emit_data64(0x08090a0b0c0d0e0f, relocInfo::none);
1379 __ emit_data64(0x0001020304050607, relocInfo::none);
1380
1381 return start;
1382 }
1383
1384 // ofs and limit are use for multi-block byte array.
1385 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
1386 address StubGenerator::generate_sha1_implCompress(bool multi_block, const char *name) {
1387 __ align(CodeEntryAlignment);
1388 StubCodeMark mark(this, "StubRoutines", name);
1389 address start = __ pc();
1390
1391 Register buf = c_rarg0;
1392 Register state = c_rarg1;
1393 Register ofs = c_rarg2;
1394 Register limit = c_rarg3;
1395
1396 const XMMRegister abcd = xmm0;
1397 const XMMRegister e0 = xmm1;
1398 const XMMRegister e1 = xmm2;
1399 const XMMRegister msg0 = xmm3;
1400
1401 const XMMRegister msg1 = xmm4;
1402 const XMMRegister msg2 = xmm5;
1403 const XMMRegister msg3 = xmm6;
1404 const XMMRegister shuf_mask = xmm7;
1405
1406 __ enter();
1407
1408 __ subptr(rsp, 4 * wordSize);
1409
1410 __ fast_sha1(abcd, e0, e1, msg0, msg1, msg2, msg3, shuf_mask,
1411 buf, state, ofs, limit, rsp, multi_block);
1412
1413 __ addptr(rsp, 4 * wordSize);
1414
1415 __ leave();
1416 __ ret(0);
1417
1418 return start;
1419 }
1420
1421 address StubGenerator::generate_pshuffle_byte_flip_mask() {
1422 __ align64();
1423 StubCodeMark mark(this, "StubRoutines", "pshuffle_byte_flip_mask");
1424 address start = __ pc();
1425
1426 __ emit_data64(0x0405060700010203, relocInfo::none);
1427 __ emit_data64(0x0c0d0e0f08090a0b, relocInfo::none);
1428
1429 if (VM_Version::supports_avx2()) {
1430 __ emit_data64(0x0405060700010203, relocInfo::none); // second copy
1431 __ emit_data64(0x0c0d0e0f08090a0b, relocInfo::none);
1432 // _SHUF_00BA
1433 __ emit_data64(0x0b0a090803020100, relocInfo::none);
1434 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1435 __ emit_data64(0x0b0a090803020100, relocInfo::none);
1436 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1437 // _SHUF_DC00
1438 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1439 __ emit_data64(0x0b0a090803020100, relocInfo::none);
1440 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1441 __ emit_data64(0x0b0a090803020100, relocInfo::none);
1442 }
1443
1444 return start;
1445 }
1446
1447 //Mask for byte-swapping a couple of qwords in an XMM register using (v)pshufb.
1448 address StubGenerator::generate_pshuffle_byte_flip_mask_sha512() {
1449 __ align32();
1450 StubCodeMark mark(this, "StubRoutines", "pshuffle_byte_flip_mask_sha512");
1451 address start = __ pc();
1452
1453 if (VM_Version::supports_avx2()) {
1454 __ emit_data64(0x0001020304050607, relocInfo::none); // PSHUFFLE_BYTE_FLIP_MASK
1455 __ emit_data64(0x08090a0b0c0d0e0f, relocInfo::none);
1456 __ emit_data64(0x1011121314151617, relocInfo::none);
1457 __ emit_data64(0x18191a1b1c1d1e1f, relocInfo::none);
1458 __ emit_data64(0x0000000000000000, relocInfo::none); //MASK_YMM_LO
1459 __ emit_data64(0x0000000000000000, relocInfo::none);
1460 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1461 __ emit_data64(0xFFFFFFFFFFFFFFFF, relocInfo::none);
1462 }
1463
1464 return start;
1465 }
1466
1467 // ofs and limit are use for multi-block byte array.
1468 // int com.sun.security.provider.DigestBase.implCompressMultiBlock(byte[] b, int ofs, int limit)
1469 address StubGenerator::generate_sha256_implCompress(bool multi_block, const char *name) {
1470 assert(VM_Version::supports_sha() || VM_Version::supports_avx2(), "");
1471 __ align(CodeEntryAlignment);
1472 StubCodeMark mark(this, "StubRoutines", name);
1473 address start = __ pc();
1474
1475 Register buf = c_rarg0;
1476 Register state = c_rarg1;
1477 Register ofs = c_rarg2;
1478 Register limit = c_rarg3;
1479
1480 const XMMRegister msg = xmm0;
1481 const XMMRegister state0 = xmm1;
1482 const XMMRegister state1 = xmm2;
1483 const XMMRegister msgtmp0 = xmm3;
1484
1485 const XMMRegister msgtmp1 = xmm4;
1486 const XMMRegister msgtmp2 = xmm5;
1487 const XMMRegister msgtmp3 = xmm6;
1488 const XMMRegister msgtmp4 = xmm7;
1489
1490 const XMMRegister shuf_mask = xmm8;
1491
1492 __ enter();
1493
1494 __ subptr(rsp, 4 * wordSize);
1495
1496 if (VM_Version::supports_sha()) {
1497 __ fast_sha256(msg, state0, state1, msgtmp0, msgtmp1, msgtmp2, msgtmp3, msgtmp4,
1498 buf, state, ofs, limit, rsp, multi_block, shuf_mask);
1499 } else if (VM_Version::supports_avx2()) {
1500 __ sha256_AVX2(msg, state0, state1, msgtmp0, msgtmp1, msgtmp2, msgtmp3, msgtmp4,
1501 buf, state, ofs, limit, rsp, multi_block, shuf_mask);
1502 }
1503 __ addptr(rsp, 4 * wordSize);
1504 __ vzeroupper();
1505 __ leave();
1506 __ ret(0);
1507
1508 return start;
1509 }
1510
1511 address StubGenerator::generate_sha512_implCompress(bool multi_block, const char *name) {
1512 assert(VM_Version::supports_avx2(), "");
1513 assert(VM_Version::supports_bmi2(), "");
1514 __ align(CodeEntryAlignment);
1515 StubCodeMark mark(this, "StubRoutines", name);
1516 address start = __ pc();
1517
1518 Register buf = c_rarg0;
1519 Register state = c_rarg1;
1520 Register ofs = c_rarg2;
1521 Register limit = c_rarg3;
1522
1523 const XMMRegister msg = xmm0;
1524 const XMMRegister state0 = xmm1;
1525 const XMMRegister state1 = xmm2;
1526 const XMMRegister msgtmp0 = xmm3;
1527 const XMMRegister msgtmp1 = xmm4;
1528 const XMMRegister msgtmp2 = xmm5;
1529 const XMMRegister msgtmp3 = xmm6;
1530 const XMMRegister msgtmp4 = xmm7;
1531
1532 const XMMRegister shuf_mask = xmm8;
1533
1534 __ enter();
1535
1536 __ sha512_AVX2(msg, state0, state1, msgtmp0, msgtmp1, msgtmp2, msgtmp3, msgtmp4,
1537 buf, state, ofs, limit, rsp, multi_block, shuf_mask);
1538
1539 __ vzeroupper();
1540 __ leave();
1541 __ ret(0);
1542
1543 return start;
1544 }
1545
1546 address StubGenerator::base64_shuffle_addr() {
1547 __ align64();
1548 StubCodeMark mark(this, "StubRoutines", "shuffle_base64");
1549 address start = __ pc();
1550
1551 assert(((unsigned long long)start & 0x3f) == 0,
1552 "Alignment problem (0x%08llx)", (unsigned long long)start);
1553 __ emit_data64(0x0405030401020001, relocInfo::none);
1554 __ emit_data64(0x0a0b090a07080607, relocInfo::none);
1555 __ emit_data64(0x10110f100d0e0c0d, relocInfo::none);
1556 __ emit_data64(0x1617151613141213, relocInfo::none);
1557 __ emit_data64(0x1c1d1b1c191a1819, relocInfo::none);
1558 __ emit_data64(0x222321221f201e1f, relocInfo::none);
1559 __ emit_data64(0x2829272825262425, relocInfo::none);
1560 __ emit_data64(0x2e2f2d2e2b2c2a2b, relocInfo::none);
1561
1562 return start;
1563 }
1564
1565 address StubGenerator::base64_avx2_shuffle_addr() {
1566 __ align32();
1567 StubCodeMark mark(this, "StubRoutines", "avx2_shuffle_base64");
1568 address start = __ pc();
1569
1570 __ emit_data64(0x0809070805060405, relocInfo::none);
1571 __ emit_data64(0x0e0f0d0e0b0c0a0b, relocInfo::none);
1572 __ emit_data64(0x0405030401020001, relocInfo::none);
1573 __ emit_data64(0x0a0b090a07080607, relocInfo::none);
1574
1575 return start;
1576 }
1577
1578 address StubGenerator::base64_avx2_input_mask_addr() {
1579 __ align32();
1580 StubCodeMark mark(this, "StubRoutines", "avx2_input_mask_base64");
1581 address start = __ pc();
1582
1583 __ emit_data64(0x8000000000000000, relocInfo::none);
1584 __ emit_data64(0x8000000080000000, relocInfo::none);
1585 __ emit_data64(0x8000000080000000, relocInfo::none);
1586 __ emit_data64(0x8000000080000000, relocInfo::none);
1587
1588 return start;
1589 }
1590
1591 address StubGenerator::base64_avx2_lut_addr() {
1592 __ align32();
1593 StubCodeMark mark(this, "StubRoutines", "avx2_lut_base64");
1594 address start = __ pc();
1595
1596 __ emit_data64(0xfcfcfcfcfcfc4741, relocInfo::none);
1597 __ emit_data64(0x0000f0edfcfcfcfc, relocInfo::none);
1598 __ emit_data64(0xfcfcfcfcfcfc4741, relocInfo::none);
1599 __ emit_data64(0x0000f0edfcfcfcfc, relocInfo::none);
1600
1601 // URL LUT
1602 __ emit_data64(0xfcfcfcfcfcfc4741, relocInfo::none);
1603 __ emit_data64(0x000020effcfcfcfc, relocInfo::none);
1604 __ emit_data64(0xfcfcfcfcfcfc4741, relocInfo::none);
1605 __ emit_data64(0x000020effcfcfcfc, relocInfo::none);
1606
1607 return start;
1608 }
1609
1610 address StubGenerator::base64_encoding_table_addr() {
1611 __ align64();
1612 StubCodeMark mark(this, "StubRoutines", "encoding_table_base64");
1613 address start = __ pc();
1614
1615 assert(((unsigned long long)start & 0x3f) == 0, "Alignment problem (0x%08llx)", (unsigned long long)start);
1616 __ emit_data64(0x4847464544434241, relocInfo::none);
1617 __ emit_data64(0x504f4e4d4c4b4a49, relocInfo::none);
1618 __ emit_data64(0x5857565554535251, relocInfo::none);
1619 __ emit_data64(0x6665646362615a59, relocInfo::none);
1620 __ emit_data64(0x6e6d6c6b6a696867, relocInfo::none);
1621 __ emit_data64(0x767574737271706f, relocInfo::none);
1622 __ emit_data64(0x333231307a797877, relocInfo::none);
1623 __ emit_data64(0x2f2b393837363534, relocInfo::none);
1624
1625 // URL table
1626 __ emit_data64(0x4847464544434241, relocInfo::none);
1627 __ emit_data64(0x504f4e4d4c4b4a49, relocInfo::none);
1628 __ emit_data64(0x5857565554535251, relocInfo::none);
1629 __ emit_data64(0x6665646362615a59, relocInfo::none);
1630 __ emit_data64(0x6e6d6c6b6a696867, relocInfo::none);
1631 __ emit_data64(0x767574737271706f, relocInfo::none);
1632 __ emit_data64(0x333231307a797877, relocInfo::none);
1633 __ emit_data64(0x5f2d393837363534, relocInfo::none);
1634
1635 return start;
1636 }
1637
1638 // Code for generating Base64 encoding.
1639 // Intrinsic function prototype in Base64.java:
1640 // private void encodeBlock(byte[] src, int sp, int sl, byte[] dst, int dp,
1641 // boolean isURL) {
1642 address StubGenerator::generate_base64_encodeBlock()
1643 {
1644 __ align(CodeEntryAlignment);
1645 StubCodeMark mark(this, "StubRoutines", "implEncode");
1646 address start = __ pc();
1647
1648 __ enter();
1649
1650 // Save callee-saved registers before using them
1651 __ push(r12);
1652 __ push(r13);
1653 __ push(r14);
1654 __ push(r15);
1655
1656 // arguments
1657 const Register source = c_rarg0; // Source Array
1658 const Register start_offset = c_rarg1; // start offset
1659 const Register end_offset = c_rarg2; // end offset
1660 const Register dest = c_rarg3; // destination array
1661
1662 #ifndef _WIN64
1663 const Register dp = c_rarg4; // Position for writing to dest array
1664 const Register isURL = c_rarg5; // Base64 or URL character set
1665 #else
1666 const Address dp_mem(rbp, 6 * wordSize); // length is on stack on Win64
1667 const Address isURL_mem(rbp, 7 * wordSize);
1668 const Register isURL = r10; // pick the volatile windows register
1669 const Register dp = r12;
1670 __ movl(dp, dp_mem);
1671 __ movl(isURL, isURL_mem);
1672 #endif
1673
1674 const Register length = r14;
1675 const Register encode_table = r13;
1676 Label L_process3, L_exit, L_processdata, L_vbmiLoop, L_not512, L_32byteLoop;
1677
1678 // calculate length from offsets
1679 __ movl(length, end_offset);
1680 __ subl(length, start_offset);
1681 __ cmpl(length, 0);
1682 __ jcc(Assembler::lessEqual, L_exit);
1683
1684 // Code for 512-bit VBMI encoding. Encodes 48 input bytes into 64
1685 // output bytes. We read 64 input bytes and ignore the last 16, so be
1686 // sure not to read past the end of the input buffer.
1687 if (VM_Version::supports_avx512_vbmi()) {
1688 __ cmpl(length, 64); // Do not overrun input buffer.
1689 __ jcc(Assembler::below, L_not512);
1690
1691 __ shll(isURL, 6); // index into decode table based on isURL
1692 __ lea(encode_table, ExternalAddress(StubRoutines::x86::base64_encoding_table_addr()));
1693 __ addptr(encode_table, isURL);
1694 __ shrl(isURL, 6); // restore isURL
1695
1696 __ mov64(rax, 0x3036242a1016040aull); // Shifts
1697 __ evmovdquq(xmm3, ExternalAddress(StubRoutines::x86::base64_shuffle_addr()), Assembler::AVX_512bit, r15);
1698 __ evmovdquq(xmm2, Address(encode_table, 0), Assembler::AVX_512bit);
1699 __ evpbroadcastq(xmm1, rax, Assembler::AVX_512bit);
1700
1701 __ align32();
1702 __ BIND(L_vbmiLoop);
1703
1704 __ vpermb(xmm0, xmm3, Address(source, start_offset), Assembler::AVX_512bit);
1705 __ subl(length, 48);
1706
1707 // Put the input bytes into the proper lanes for writing, then
1708 // encode them.
1709 __ evpmultishiftqb(xmm0, xmm1, xmm0, Assembler::AVX_512bit);
1710 __ vpermb(xmm0, xmm0, xmm2, Assembler::AVX_512bit);
1711
1712 // Write to destination
1713 __ evmovdquq(Address(dest, dp), xmm0, Assembler::AVX_512bit);
1714
1715 __ addptr(dest, 64);
1716 __ addptr(source, 48);
1717 __ cmpl(length, 64);
1718 __ jcc(Assembler::aboveEqual, L_vbmiLoop);
1719
1720 __ vzeroupper();
1721 }
1722
1723 __ BIND(L_not512);
1724 if (VM_Version::supports_avx2()
1725 && VM_Version::supports_avx512vlbw()) {
1726 /*
1727 ** This AVX2 encoder is based off the paper at:
1728 ** https://dl.acm.org/doi/10.1145/3132709
1729 **
1730 ** We use AVX2 SIMD instructions to encode 24 bytes into 32
1731 ** output bytes.
1732 **
1733 */
1734 // Lengths under 32 bytes are done with scalar routine
1735 __ cmpl(length, 31);
1736 __ jcc(Assembler::belowEqual, L_process3);
1737
1738 // Set up supporting constant table data
1739 __ vmovdqu(xmm9, ExternalAddress(StubRoutines::x86::base64_avx2_shuffle_addr()), rax);
1740 // 6-bit mask for 2nd and 4th (and multiples) 6-bit values
1741 __ movl(rax, 0x0fc0fc00);
1742 __ vmovdqu(xmm1, ExternalAddress(StubRoutines::x86::base64_avx2_input_mask_addr()), rax);
1743 __ evpbroadcastd(xmm8, rax, Assembler::AVX_256bit);
1744
1745 // Multiplication constant for "shifting" right by 6 and 10
1746 // bits
1747 __ movl(rax, 0x04000040);
1748
1749 __ subl(length, 24);
1750 __ evpbroadcastd(xmm7, rax, Assembler::AVX_256bit);
1751
1752 // For the first load, we mask off reading of the first 4
1753 // bytes into the register. This is so we can get 4 3-byte
1754 // chunks into each lane of the register, avoiding having to
1755 // handle end conditions. We then shuffle these bytes into a
1756 // specific order so that manipulation is easier.
1757 //
1758 // The initial read loads the XMM register like this:
1759 //
1760 // Lower 128-bit lane:
1761 // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
1762 // | XX | XX | XX | XX | A0 | A1 | A2 | B0 | B1 | B2 | C0 | C1
1763 // | C2 | D0 | D1 | D2 |
1764 // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
1765 //
1766 // Upper 128-bit lane:
1767 // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
1768 // | E0 | E1 | E2 | F0 | F1 | F2 | G0 | G1 | G2 | H0 | H1 | H2
1769 // | XX | XX | XX | XX |
1770 // +----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+----+
1771 //
1772 // Where A0 is the first input byte, B0 is the fourth, etc.
1773 // The alphabetical significance denotes the 3 bytes to be
1774 // consumed and encoded into 4 bytes.
1775 //
1776 // We then shuffle the register so each 32-bit word contains
1777 // the sequence:
1778 // A1 A0 A2 A1, B1, B0, B2, B1, etc.
1779 // Each of these byte sequences are then manipulated into 4
1780 // 6-bit values ready for encoding.
1781 //
1782 // If we focus on one set of 3-byte chunks, changing the
1783 // nomenclature such that A0 => a, A1 => b, and A2 => c, we
1784 // shuffle such that each 24-bit chunk contains:
1785 //
1786 // b7 b6 b5 b4 b3 b2 b1 b0 | a7 a6 a5 a4 a3 a2 a1 a0 | c7 c6
1787 // c5 c4 c3 c2 c1 c0 | b7 b6 b5 b4 b3 b2 b1 b0
1788 // Explain this step.
1789 // b3 b2 b1 b0 c5 c4 c3 c2 | c1 c0 d5 d4 d3 d2 d1 d0 | a5 a4
1790 // a3 a2 a1 a0 b5 b4 | b3 b2 b1 b0 c5 c4 c3 c2
1791 //
1792 // W first and off all but bits 4-9 and 16-21 (c5..c0 and
1793 // a5..a0) and shift them using a vector multiplication
1794 // operation (vpmulhuw) which effectively shifts c right by 6
1795 // bits and a right by 10 bits. We similarly mask bits 10-15
1796 // (d5..d0) and 22-27 (b5..b0) and shift them left by 8 and 4
1797 // bits respectively. This is done using vpmullw. We end up
1798 // with 4 6-bit values, thus splitting the 3 input bytes,
1799 // ready for encoding:
1800 // 0 0 d5..d0 0 0 c5..c0 0 0 b5..b0 0 0 a5..a0
1801 //
1802 // For translation, we recognize that there are 5 distinct
1803 // ranges of legal Base64 characters as below:
1804 //
1805 // +-------------+-------------+------------+
1806 // | 6-bit value | ASCII range | offset |
1807 // +-------------+-------------+------------+
1808 // | 0..25 | A..Z | 65 |
1809 // | 26..51 | a..z | 71 |
1810 // | 52..61 | 0..9 | -4 |
1811 // | 62 | + or - | -19 or -17 |
1812 // | 63 | / or _ | -16 or 32 |
1813 // +-------------+-------------+------------+
1814 //
1815 // We note that vpshufb does a parallel lookup in a
1816 // destination register using the lower 4 bits of bytes from a
1817 // source register. If we use a saturated subtraction and
1818 // subtract 51 from each 6-bit value, bytes from [0,51]
1819 // saturate to 0, and [52,63] map to a range of [1,12]. We
1820 // distinguish the [0,25] and [26,51] ranges by assigning a
1821 // value of 13 for all 6-bit values less than 26. We end up
1822 // with:
1823 //
1824 // +-------------+-------------+------------+
1825 // | 6-bit value | Reduced | offset |
1826 // +-------------+-------------+------------+
1827 // | 0..25 | 13 | 65 |
1828 // | 26..51 | 0 | 71 |
1829 // | 52..61 | 0..9 | -4 |
1830 // | 62 | 11 | -19 or -17 |
1831 // | 63 | 12 | -16 or 32 |
1832 // +-------------+-------------+------------+
1833 //
1834 // We then use a final vpshufb to add the appropriate offset,
1835 // translating the bytes.
1836 //
1837 // Load input bytes - only 28 bytes. Mask the first load to
1838 // not load into the full register.
1839 __ vpmaskmovd(xmm1, xmm1, Address(source, start_offset, Address::times_1, -4), Assembler::AVX_256bit);
1840
1841 // Move 3-byte chunks of input (12 bytes) into 16 bytes,
1842 // ordering by:
1843 // 1, 0, 2, 1; 4, 3, 5, 4; etc. This groups 6-bit chunks
1844 // for easy masking
1845 __ vpshufb(xmm1, xmm1, xmm9, Assembler::AVX_256bit);
1846
1847 __ addl(start_offset, 24);
1848
1849 // Load masking register for first and third (and multiples)
1850 // 6-bit values.
1851 __ movl(rax, 0x003f03f0);
1852 __ evpbroadcastd(xmm6, rax, Assembler::AVX_256bit);
1853 // Multiplication constant for "shifting" left by 4 and 8 bits
1854 __ movl(rax, 0x01000010);
1855 __ evpbroadcastd(xmm5, rax, Assembler::AVX_256bit);
1856
1857 // Isolate 6-bit chunks of interest
1858 __ vpand(xmm0, xmm8, xmm1, Assembler::AVX_256bit);
1859
1860 // Load constants for encoding
1861 __ movl(rax, 0x19191919);
1862 __ evpbroadcastd(xmm3, rax, Assembler::AVX_256bit);
1863 __ movl(rax, 0x33333333);
1864 __ evpbroadcastd(xmm4, rax, Assembler::AVX_256bit);
1865
1866 // Shift output bytes 0 and 2 into proper lanes
1867 __ vpmulhuw(xmm2, xmm0, xmm7, Assembler::AVX_256bit);
1868
1869 // Mask and shift output bytes 1 and 3 into proper lanes and
1870 // combine
1871 __ vpand(xmm0, xmm6, xmm1, Assembler::AVX_256bit);
1872 __ vpmullw(xmm0, xmm5, xmm0, Assembler::AVX_256bit);
1873 __ vpor(xmm0, xmm0, xmm2, Assembler::AVX_256bit);
1874
1875 // Find out which are 0..25. This indicates which input
1876 // values fall in the range of 'A'-'Z', which require an
1877 // additional offset (see comments above)
1878 __ vpcmpgtb(xmm2, xmm0, xmm3, Assembler::AVX_256bit);
1879 __ vpsubusb(xmm1, xmm0, xmm4, Assembler::AVX_256bit);
1880 __ vpsubb(xmm1, xmm1, xmm2, Assembler::AVX_256bit);
1881
1882 // Load the proper lookup table
1883 __ lea(r11, ExternalAddress(StubRoutines::x86::base64_avx2_lut_addr()));
1884 __ movl(r15, isURL);
1885 __ shll(r15, 5);
1886 __ vmovdqu(xmm2, Address(r11, r15));
1887
1888 // Shuffle the offsets based on the range calculation done
1889 // above. This allows us to add the correct offset to the
1890 // 6-bit value corresponding to the range documented above.
1891 __ vpshufb(xmm1, xmm2, xmm1, Assembler::AVX_256bit);
1892 __ vpaddb(xmm0, xmm1, xmm0, Assembler::AVX_256bit);
1893
1894 // Store the encoded bytes
1895 __ vmovdqu(Address(dest, dp), xmm0);
1896 __ addl(dp, 32);
1897
1898 __ cmpl(length, 31);
1899 __ jcc(Assembler::belowEqual, L_process3);
1900
1901 __ align32();
1902 __ BIND(L_32byteLoop);
1903
1904 // Get next 32 bytes
1905 __ vmovdqu(xmm1, Address(source, start_offset, Address::times_1, -4));
1906
1907 __ subl(length, 24);
1908 __ addl(start_offset, 24);
1909
1910 // This logic is identical to the above, with only constant
1911 // register loads removed. Shuffle the input, mask off 6-bit
1912 // chunks, shift them into place, then add the offset to
1913 // encode.
1914 __ vpshufb(xmm1, xmm1, xmm9, Assembler::AVX_256bit);
1915
1916 __ vpand(xmm0, xmm8, xmm1, Assembler::AVX_256bit);
1917 __ vpmulhuw(xmm10, xmm0, xmm7, Assembler::AVX_256bit);
1918 __ vpand(xmm0, xmm6, xmm1, Assembler::AVX_256bit);
1919 __ vpmullw(xmm0, xmm5, xmm0, Assembler::AVX_256bit);
1920 __ vpor(xmm0, xmm0, xmm10, Assembler::AVX_256bit);
1921 __ vpcmpgtb(xmm10, xmm0, xmm3, Assembler::AVX_256bit);
1922 __ vpsubusb(xmm1, xmm0, xmm4, Assembler::AVX_256bit);
1923 __ vpsubb(xmm1, xmm1, xmm10, Assembler::AVX_256bit);
1924 __ vpshufb(xmm1, xmm2, xmm1, Assembler::AVX_256bit);
1925 __ vpaddb(xmm0, xmm1, xmm0, Assembler::AVX_256bit);
1926
1927 // Store the encoded bytes
1928 __ vmovdqu(Address(dest, dp), xmm0);
1929 __ addl(dp, 32);
1930
1931 __ cmpl(length, 31);
1932 __ jcc(Assembler::above, L_32byteLoop);
1933
1934 __ BIND(L_process3);
1935 __ vzeroupper();
1936 } else {
1937 __ BIND(L_process3);
1938 }
1939
1940 __ cmpl(length, 3);
1941 __ jcc(Assembler::below, L_exit);
1942
1943 // Load the encoding table based on isURL
1944 __ lea(r11, ExternalAddress(StubRoutines::x86::base64_encoding_table_addr()));
1945 __ movl(r15, isURL);
1946 __ shll(r15, 6);
1947 __ addptr(r11, r15);
1948
1949 __ BIND(L_processdata);
1950
1951 // Load 3 bytes
1952 __ load_unsigned_byte(r15, Address(source, start_offset));
1953 __ load_unsigned_byte(r10, Address(source, start_offset, Address::times_1, 1));
1954 __ load_unsigned_byte(r13, Address(source, start_offset, Address::times_1, 2));
1955
1956 // Build a 32-bit word with bytes 1, 2, 0, 1
1957 __ movl(rax, r10);
1958 __ shll(r10, 24);
1959 __ orl(rax, r10);
1960
1961 __ subl(length, 3);
1962
1963 __ shll(r15, 8);
1964 __ shll(r13, 16);
1965 __ orl(rax, r15);
1966
1967 __ addl(start_offset, 3);
1968
1969 __ orl(rax, r13);
1970 // At this point, rax contains | byte1 | byte2 | byte0 | byte1
1971 // r13 has byte2 << 16 - need low-order 6 bits to translate.
1972 // This translated byte is the fourth output byte.
1973 __ shrl(r13, 16);
1974 __ andl(r13, 0x3f);
1975
1976 // The high-order 6 bits of r15 (byte0) is translated.
1977 // The translated byte is the first output byte.
1978 __ shrl(r15, 10);
1979
1980 __ load_unsigned_byte(r13, Address(r11, r13));
1981 __ load_unsigned_byte(r15, Address(r11, r15));
1982
1983 __ movb(Address(dest, dp, Address::times_1, 3), r13);
1984
1985 // Extract high-order 4 bits of byte1 and low-order 2 bits of byte0.
1986 // This translated byte is the second output byte.
1987 __ shrl(rax, 4);
1988 __ movl(r10, rax);
1989 __ andl(rax, 0x3f);
1990
1991 __ movb(Address(dest, dp, Address::times_1, 0), r15);
1992
1993 __ load_unsigned_byte(rax, Address(r11, rax));
1994
1995 // Extract low-order 2 bits of byte1 and high-order 4 bits of byte2.
1996 // This translated byte is the third output byte.
1997 __ shrl(r10, 18);
1998 __ andl(r10, 0x3f);
1999
2000 __ load_unsigned_byte(r10, Address(r11, r10));
2001
2002 __ movb(Address(dest, dp, Address::times_1, 1), rax);
2003 __ movb(Address(dest, dp, Address::times_1, 2), r10);
2004
2005 __ addl(dp, 4);
2006 __ cmpl(length, 3);
2007 __ jcc(Assembler::aboveEqual, L_processdata);
2008
2009 __ BIND(L_exit);
2010 __ pop(r15);
2011 __ pop(r14);
2012 __ pop(r13);
2013 __ pop(r12);
2014 __ leave();
2015 __ ret(0);
2016
2017 return start;
2018 }
2019
2020 // base64 AVX512vbmi tables
2021 address StubGenerator::base64_vbmi_lookup_lo_addr() {
2022 __ align64();
2023 StubCodeMark mark(this, "StubRoutines", "lookup_lo_base64");
2024 address start = __ pc();
2025
2026 assert(((unsigned long long)start & 0x3f) == 0,
2027 "Alignment problem (0x%08llx)", (unsigned long long)start);
2028 __ emit_data64(0x8080808080808080, relocInfo::none);
2029 __ emit_data64(0x8080808080808080, relocInfo::none);
2030 __ emit_data64(0x8080808080808080, relocInfo::none);
2031 __ emit_data64(0x8080808080808080, relocInfo::none);
2032 __ emit_data64(0x8080808080808080, relocInfo::none);
2033 __ emit_data64(0x3f8080803e808080, relocInfo::none);
2034 __ emit_data64(0x3b3a393837363534, relocInfo::none);
2035 __ emit_data64(0x8080808080803d3c, relocInfo::none);
2036
2037 return start;
2038 }
2039
2040 address StubGenerator::base64_vbmi_lookup_hi_addr() {
2041 __ align64();
2042 StubCodeMark mark(this, "StubRoutines", "lookup_hi_base64");
2043 address start = __ pc();
2044
2045 assert(((unsigned long long)start & 0x3f) == 0,
2046 "Alignment problem (0x%08llx)", (unsigned long long)start);
2047 __ emit_data64(0x0605040302010080, relocInfo::none);
2048 __ emit_data64(0x0e0d0c0b0a090807, relocInfo::none);
2049 __ emit_data64(0x161514131211100f, relocInfo::none);
2050 __ emit_data64(0x8080808080191817, relocInfo::none);
2051 __ emit_data64(0x201f1e1d1c1b1a80, relocInfo::none);
2052 __ emit_data64(0x2827262524232221, relocInfo::none);
2053 __ emit_data64(0x302f2e2d2c2b2a29, relocInfo::none);
2054 __ emit_data64(0x8080808080333231, relocInfo::none);
2055
2056 return start;
2057 }
2058 address StubGenerator::base64_vbmi_lookup_lo_url_addr() {
2059 __ align64();
2060 StubCodeMark mark(this, "StubRoutines", "lookup_lo_base64url");
2061 address start = __ pc();
2062
2063 assert(((unsigned long long)start & 0x3f) == 0,
2064 "Alignment problem (0x%08llx)", (unsigned long long)start);
2065 __ emit_data64(0x8080808080808080, relocInfo::none);
2066 __ emit_data64(0x8080808080808080, relocInfo::none);
2067 __ emit_data64(0x8080808080808080, relocInfo::none);
2068 __ emit_data64(0x8080808080808080, relocInfo::none);
2069 __ emit_data64(0x8080808080808080, relocInfo::none);
2070 __ emit_data64(0x80803e8080808080, relocInfo::none);
2071 __ emit_data64(0x3b3a393837363534, relocInfo::none);
2072 __ emit_data64(0x8080808080803d3c, relocInfo::none);
2073
2074 return start;
2075 }
2076
2077 address StubGenerator::base64_vbmi_lookup_hi_url_addr() {
2078 __ align64();
2079 StubCodeMark mark(this, "StubRoutines", "lookup_hi_base64url");
2080 address start = __ pc();
2081
2082 assert(((unsigned long long)start & 0x3f) == 0,
2083 "Alignment problem (0x%08llx)", (unsigned long long)start);
2084 __ emit_data64(0x0605040302010080, relocInfo::none);
2085 __ emit_data64(0x0e0d0c0b0a090807, relocInfo::none);
2086 __ emit_data64(0x161514131211100f, relocInfo::none);
2087 __ emit_data64(0x3f80808080191817, relocInfo::none);
2088 __ emit_data64(0x201f1e1d1c1b1a80, relocInfo::none);
2089 __ emit_data64(0x2827262524232221, relocInfo::none);
2090 __ emit_data64(0x302f2e2d2c2b2a29, relocInfo::none);
2091 __ emit_data64(0x8080808080333231, relocInfo::none);
2092
2093 return start;
2094 }
2095
2096 address StubGenerator::base64_vbmi_pack_vec_addr() {
2097 __ align64();
2098 StubCodeMark mark(this, "StubRoutines", "pack_vec_base64");
2099 address start = __ pc();
2100
2101 assert(((unsigned long long)start & 0x3f) == 0,
2102 "Alignment problem (0x%08llx)", (unsigned long long)start);
2103 __ emit_data64(0x090a040506000102, relocInfo::none);
2104 __ emit_data64(0x161011120c0d0e08, relocInfo::none);
2105 __ emit_data64(0x1c1d1e18191a1415, relocInfo::none);
2106 __ emit_data64(0x292a242526202122, relocInfo::none);
2107 __ emit_data64(0x363031322c2d2e28, relocInfo::none);
2108 __ emit_data64(0x3c3d3e38393a3435, relocInfo::none);
2109 __ emit_data64(0x0000000000000000, relocInfo::none);
2110 __ emit_data64(0x0000000000000000, relocInfo::none);
2111
2112 return start;
2113 }
2114
2115 address StubGenerator::base64_vbmi_join_0_1_addr() {
2116 __ align64();
2117 StubCodeMark mark(this, "StubRoutines", "join_0_1_base64");
2118 address start = __ pc();
2119
2120 assert(((unsigned long long)start & 0x3f) == 0,
2121 "Alignment problem (0x%08llx)", (unsigned long long)start);
2122 __ emit_data64(0x090a040506000102, relocInfo::none);
2123 __ emit_data64(0x161011120c0d0e08, relocInfo::none);
2124 __ emit_data64(0x1c1d1e18191a1415, relocInfo::none);
2125 __ emit_data64(0x292a242526202122, relocInfo::none);
2126 __ emit_data64(0x363031322c2d2e28, relocInfo::none);
2127 __ emit_data64(0x3c3d3e38393a3435, relocInfo::none);
2128 __ emit_data64(0x494a444546404142, relocInfo::none);
2129 __ emit_data64(0x565051524c4d4e48, relocInfo::none);
2130
2131 return start;
2132 }
2133
2134 address StubGenerator::base64_vbmi_join_1_2_addr() {
2135 __ align64();
2136 StubCodeMark mark(this, "StubRoutines", "join_1_2_base64");
2137 address start = __ pc();
2138
2139 assert(((unsigned long long)start & 0x3f) == 0,
2140 "Alignment problem (0x%08llx)", (unsigned long long)start);
2141 __ emit_data64(0x1c1d1e18191a1415, relocInfo::none);
2142 __ emit_data64(0x292a242526202122, relocInfo::none);
2143 __ emit_data64(0x363031322c2d2e28, relocInfo::none);
2144 __ emit_data64(0x3c3d3e38393a3435, relocInfo::none);
2145 __ emit_data64(0x494a444546404142, relocInfo::none);
2146 __ emit_data64(0x565051524c4d4e48, relocInfo::none);
2147 __ emit_data64(0x5c5d5e58595a5455, relocInfo::none);
2148 __ emit_data64(0x696a646566606162, relocInfo::none);
2149
2150 return start;
2151 }
2152
2153 address StubGenerator::base64_vbmi_join_2_3_addr() {
2154 __ align64();
2155 StubCodeMark mark(this, "StubRoutines", "join_2_3_base64");
2156 address start = __ pc();
2157
2158 assert(((unsigned long long)start & 0x3f) == 0,
2159 "Alignment problem (0x%08llx)", (unsigned long long)start);
2160 __ emit_data64(0x363031322c2d2e28, relocInfo::none);
2161 __ emit_data64(0x3c3d3e38393a3435, relocInfo::none);
2162 __ emit_data64(0x494a444546404142, relocInfo::none);
2163 __ emit_data64(0x565051524c4d4e48, relocInfo::none);
2164 __ emit_data64(0x5c5d5e58595a5455, relocInfo::none);
2165 __ emit_data64(0x696a646566606162, relocInfo::none);
2166 __ emit_data64(0x767071726c6d6e68, relocInfo::none);
2167 __ emit_data64(0x7c7d7e78797a7475, relocInfo::none);
2168
2169 return start;
2170 }
2171
2172 address StubGenerator::base64_decoding_table_addr() {
2173 StubCodeMark mark(this, "StubRoutines", "decoding_table_base64");
2174 address start = __ pc();
2175
2176 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2177 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2178 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2179 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2180 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2181 __ emit_data64(0x3fffffff3effffff, relocInfo::none);
2182 __ emit_data64(0x3b3a393837363534, relocInfo::none);
2183 __ emit_data64(0xffffffffffff3d3c, relocInfo::none);
2184 __ emit_data64(0x06050403020100ff, relocInfo::none);
2185 __ emit_data64(0x0e0d0c0b0a090807, relocInfo::none);
2186 __ emit_data64(0x161514131211100f, relocInfo::none);
2187 __ emit_data64(0xffffffffff191817, relocInfo::none);
2188 __ emit_data64(0x201f1e1d1c1b1aff, relocInfo::none);
2189 __ emit_data64(0x2827262524232221, relocInfo::none);
2190 __ emit_data64(0x302f2e2d2c2b2a29, relocInfo::none);
2191 __ emit_data64(0xffffffffff333231, relocInfo::none);
2192 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2193 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2194 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2195 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2196 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2197 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2198 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2199 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2200 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2201 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2202 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2203 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2204 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2205 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2206 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2207 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2208
2209 // URL table
2210 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2211 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2212 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2213 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2214 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2215 __ emit_data64(0xffff3effffffffff, relocInfo::none);
2216 __ emit_data64(0x3b3a393837363534, relocInfo::none);
2217 __ emit_data64(0xffffffffffff3d3c, relocInfo::none);
2218 __ emit_data64(0x06050403020100ff, relocInfo::none);
2219 __ emit_data64(0x0e0d0c0b0a090807, relocInfo::none);
2220 __ emit_data64(0x161514131211100f, relocInfo::none);
2221 __ emit_data64(0x3fffffffff191817, relocInfo::none);
2222 __ emit_data64(0x201f1e1d1c1b1aff, relocInfo::none);
2223 __ emit_data64(0x2827262524232221, relocInfo::none);
2224 __ emit_data64(0x302f2e2d2c2b2a29, relocInfo::none);
2225 __ emit_data64(0xffffffffff333231, relocInfo::none);
2226 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2227 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2228 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2229 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2230 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2231 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2232 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2233 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2234 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2235 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2236 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2237 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2238 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2239 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2240 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2241 __ emit_data64(0xffffffffffffffff, relocInfo::none);
2242
2243 return start;
2244 }
2245
2246
2247 // Code for generating Base64 decoding.
2248 //
2249 // Based on the article (and associated code) from https://arxiv.org/abs/1910.05109.
2250 //
2251 // Intrinsic function prototype in Base64.java:
2252 // private void decodeBlock(byte[] src, int sp, int sl, byte[] dst, int dp, boolean isURL, isMIME) {
2253 address StubGenerator::generate_base64_decodeBlock() {
2254 __ align(CodeEntryAlignment);
2255 StubCodeMark mark(this, "StubRoutines", "implDecode");
2256 address start = __ pc();
2257
2258 __ enter();
2259
2260 // Save callee-saved registers before using them
2261 __ push(r12);
2262 __ push(r13);
2263 __ push(r14);
2264 __ push(r15);
2265 __ push(rbx);
2266
2267 // arguments
2268 const Register source = c_rarg0; // Source Array
2269 const Register start_offset = c_rarg1; // start offset
2270 const Register end_offset = c_rarg2; // end offset
2271 const Register dest = c_rarg3; // destination array
2272 const Register isMIME = rbx;
2273
2274 #ifndef _WIN64
2275 const Register dp = c_rarg4; // Position for writing to dest array
2276 const Register isURL = c_rarg5;// Base64 or URL character set
2277 __ movl(isMIME, Address(rbp, 2 * wordSize));
2278 #else
2279 const Address dp_mem(rbp, 6 * wordSize); // length is on stack on Win64
2280 const Address isURL_mem(rbp, 7 * wordSize);
2281 const Register isURL = r10; // pick the volatile windows register
2282 const Register dp = r12;
2283 __ movl(dp, dp_mem);
2284 __ movl(isURL, isURL_mem);
2285 __ movl(isMIME, Address(rbp, 8 * wordSize));
2286 #endif
2287
2288 const XMMRegister lookup_lo = xmm5;
2289 const XMMRegister lookup_hi = xmm6;
2290 const XMMRegister errorvec = xmm7;
2291 const XMMRegister pack16_op = xmm9;
2292 const XMMRegister pack32_op = xmm8;
2293 const XMMRegister input0 = xmm3;
2294 const XMMRegister input1 = xmm20;
2295 const XMMRegister input2 = xmm21;
2296 const XMMRegister input3 = xmm19;
2297 const XMMRegister join01 = xmm12;
2298 const XMMRegister join12 = xmm11;
2299 const XMMRegister join23 = xmm10;
2300 const XMMRegister translated0 = xmm2;
2301 const XMMRegister translated1 = xmm1;
2302 const XMMRegister translated2 = xmm0;
2303 const XMMRegister translated3 = xmm4;
2304
2305 const XMMRegister merged0 = xmm2;
2306 const XMMRegister merged1 = xmm1;
2307 const XMMRegister merged2 = xmm0;
2308 const XMMRegister merged3 = xmm4;
2309 const XMMRegister merge_ab_bc0 = xmm2;
2310 const XMMRegister merge_ab_bc1 = xmm1;
2311 const XMMRegister merge_ab_bc2 = xmm0;
2312 const XMMRegister merge_ab_bc3 = xmm4;
2313
2314 const XMMRegister pack24bits = xmm4;
2315
2316 const Register length = r14;
2317 const Register output_size = r13;
2318 const Register output_mask = r15;
2319 const KRegister input_mask = k1;
2320
2321 const XMMRegister input_initial_valid_b64 = xmm0;
2322 const XMMRegister tmp = xmm10;
2323 const XMMRegister mask = xmm0;
2324 const XMMRegister invalid_b64 = xmm1;
2325
2326 Label L_process256, L_process64, L_process64Loop, L_exit, L_processdata, L_loadURL;
2327 Label L_continue, L_finalBit, L_padding, L_donePadding, L_bruteForce;
2328 Label L_forceLoop, L_bottomLoop, L_checkMIME, L_exit_no_vzero;
2329
2330 // calculate length from offsets
2331 __ movl(length, end_offset);
2332 __ subl(length, start_offset);
2333 __ push(dest); // Save for return value calc
2334
2335 // If AVX512 VBMI not supported, just compile non-AVX code
2336 if(VM_Version::supports_avx512_vbmi() &&
2337 VM_Version::supports_avx512bw()) {
2338 __ cmpl(length, 128); // 128-bytes is break-even for AVX-512
2339 __ jcc(Assembler::lessEqual, L_bruteForce);
2340
2341 __ cmpl(isMIME, 0);
2342 __ jcc(Assembler::notEqual, L_bruteForce);
2343
2344 // Load lookup tables based on isURL
2345 __ cmpl(isURL, 0);
2346 __ jcc(Assembler::notZero, L_loadURL);
2347
2348 __ evmovdquq(lookup_lo, ExternalAddress(StubRoutines::x86::base64_vbmi_lookup_lo_addr()), Assembler::AVX_512bit, r13);
2349 __ evmovdquq(lookup_hi, ExternalAddress(StubRoutines::x86::base64_vbmi_lookup_hi_addr()), Assembler::AVX_512bit, r13);
2350
2351 __ BIND(L_continue);
2352
2353 __ movl(r15, 0x01400140);
2354 __ evpbroadcastd(pack16_op, r15, Assembler::AVX_512bit);
2355
2356 __ movl(r15, 0x00011000);
2357 __ evpbroadcastd(pack32_op, r15, Assembler::AVX_512bit);
2358
2359 __ cmpl(length, 0xff);
2360 __ jcc(Assembler::lessEqual, L_process64);
2361
2362 // load masks required for decoding data
2363 __ BIND(L_processdata);
2364 __ evmovdquq(join01, ExternalAddress(StubRoutines::x86::base64_vbmi_join_0_1_addr()), Assembler::AVX_512bit,r13);
2365 __ evmovdquq(join12, ExternalAddress(StubRoutines::x86::base64_vbmi_join_1_2_addr()), Assembler::AVX_512bit, r13);
2366 __ evmovdquq(join23, ExternalAddress(StubRoutines::x86::base64_vbmi_join_2_3_addr()), Assembler::AVX_512bit, r13);
2367
2368 __ align32();
2369 __ BIND(L_process256);
2370 // Grab input data
2371 __ evmovdquq(input0, Address(source, start_offset, Address::times_1, 0x00), Assembler::AVX_512bit);
2372 __ evmovdquq(input1, Address(source, start_offset, Address::times_1, 0x40), Assembler::AVX_512bit);
2373 __ evmovdquq(input2, Address(source, start_offset, Address::times_1, 0x80), Assembler::AVX_512bit);
2374 __ evmovdquq(input3, Address(source, start_offset, Address::times_1, 0xc0), Assembler::AVX_512bit);
2375
2376 // Copy the low part of the lookup table into the destination of the permutation
2377 __ evmovdquq(translated0, lookup_lo, Assembler::AVX_512bit);
2378 __ evmovdquq(translated1, lookup_lo, Assembler::AVX_512bit);
2379 __ evmovdquq(translated2, lookup_lo, Assembler::AVX_512bit);
2380 __ evmovdquq(translated3, lookup_lo, Assembler::AVX_512bit);
2381
2382 // Translate the base64 input into "decoded" bytes
2383 __ evpermt2b(translated0, input0, lookup_hi, Assembler::AVX_512bit);
2384 __ evpermt2b(translated1, input1, lookup_hi, Assembler::AVX_512bit);
2385 __ evpermt2b(translated2, input2, lookup_hi, Assembler::AVX_512bit);
2386 __ evpermt2b(translated3, input3, lookup_hi, Assembler::AVX_512bit);
2387
2388 // OR all of the translations together to check for errors (high-order bit of byte set)
2389 __ vpternlogd(input0, 0xfe, input1, input2, Assembler::AVX_512bit);
2390
2391 __ vpternlogd(input3, 0xfe, translated0, translated1, Assembler::AVX_512bit);
2392 __ vpternlogd(input0, 0xfe, translated2, translated3, Assembler::AVX_512bit);
2393 __ vpor(errorvec, input3, input0, Assembler::AVX_512bit);
2394
2395 // Check if there was an error - if so, try 64-byte chunks
2396 __ evpmovb2m(k3, errorvec, Assembler::AVX_512bit);
2397 __ kortestql(k3, k3);
2398 __ jcc(Assembler::notZero, L_process64);
2399
2400 // The merging and shuffling happens here
2401 // We multiply each byte pair [00dddddd | 00cccccc | 00bbbbbb | 00aaaaaa]
2402 // Multiply [00cccccc] by 2^6 added to [00dddddd] to get [0000cccc | ccdddddd]
2403 // The pack16_op is a vector of 0x01400140, so multiply D by 1 and C by 0x40
2404 __ vpmaddubsw(merge_ab_bc0, translated0, pack16_op, Assembler::AVX_512bit);
2405 __ vpmaddubsw(merge_ab_bc1, translated1, pack16_op, Assembler::AVX_512bit);
2406 __ vpmaddubsw(merge_ab_bc2, translated2, pack16_op, Assembler::AVX_512bit);
2407 __ vpmaddubsw(merge_ab_bc3, translated3, pack16_op, Assembler::AVX_512bit);
2408
2409 // Now do the same with packed 16-bit values.
2410 // We start with [0000cccc | ccdddddd | 0000aaaa | aabbbbbb]
2411 // pack32_op is 0x00011000 (2^12, 1), so this multiplies [0000aaaa | aabbbbbb] by 2^12
2412 // and adds [0000cccc | ccdddddd] to yield [00000000 | aaaaaabb | bbbbcccc | ccdddddd]
2413 __ vpmaddwd(merged0, merge_ab_bc0, pack32_op, Assembler::AVX_512bit);
2414 __ vpmaddwd(merged1, merge_ab_bc1, pack32_op, Assembler::AVX_512bit);
2415 __ vpmaddwd(merged2, merge_ab_bc2, pack32_op, Assembler::AVX_512bit);
2416 __ vpmaddwd(merged3, merge_ab_bc3, pack32_op, Assembler::AVX_512bit);
2417
2418 // The join vectors specify which byte from which vector goes into the outputs
2419 // One of every 4 bytes in the extended vector is zero, so we pack them into their
2420 // final positions in the register for storing (256 bytes in, 192 bytes out)
2421 __ evpermt2b(merged0, join01, merged1, Assembler::AVX_512bit);
2422 __ evpermt2b(merged1, join12, merged2, Assembler::AVX_512bit);
2423 __ evpermt2b(merged2, join23, merged3, Assembler::AVX_512bit);
2424
2425 // Store result
2426 __ evmovdquq(Address(dest, dp, Address::times_1, 0x00), merged0, Assembler::AVX_512bit);
2427 __ evmovdquq(Address(dest, dp, Address::times_1, 0x40), merged1, Assembler::AVX_512bit);
2428 __ evmovdquq(Address(dest, dp, Address::times_1, 0x80), merged2, Assembler::AVX_512bit);
2429
2430 __ addptr(source, 0x100);
2431 __ addptr(dest, 0xc0);
2432 __ subl(length, 0x100);
2433 __ cmpl(length, 64 * 4);
2434 __ jcc(Assembler::greaterEqual, L_process256);
2435
2436 // At this point, we've decoded 64 * 4 * n bytes.
2437 // The remaining length will be <= 64 * 4 - 1.
2438 // UNLESS there was an error decoding the first 256-byte chunk. In this
2439 // case, the length will be arbitrarily long.
2440 //
2441 // Note that this will be the path for MIME-encoded strings.
2442
2443 __ BIND(L_process64);
2444
2445 __ evmovdquq(pack24bits, ExternalAddress(StubRoutines::x86::base64_vbmi_pack_vec_addr()), Assembler::AVX_512bit, r13);
2446
2447 __ cmpl(length, 63);
2448 __ jcc(Assembler::lessEqual, L_finalBit);
2449
2450 __ mov64(rax, 0x0000ffffffffffff);
2451 __ kmovql(k2, rax);
2452
2453 __ align32();
2454 __ BIND(L_process64Loop);
2455
2456 // Handle first 64-byte block
2457
2458 __ evmovdquq(input0, Address(source, start_offset), Assembler::AVX_512bit);
2459 __ evmovdquq(translated0, lookup_lo, Assembler::AVX_512bit);
2460 __ evpermt2b(translated0, input0, lookup_hi, Assembler::AVX_512bit);
2461
2462 __ vpor(errorvec, translated0, input0, Assembler::AVX_512bit);
2463
2464 // Check for error and bomb out before updating dest
2465 __ evpmovb2m(k3, errorvec, Assembler::AVX_512bit);
2466 __ kortestql(k3, k3);
2467 __ jcc(Assembler::notZero, L_exit);
2468
2469 // Pack output register, selecting correct byte ordering
2470 __ vpmaddubsw(merge_ab_bc0, translated0, pack16_op, Assembler::AVX_512bit);
2471 __ vpmaddwd(merged0, merge_ab_bc0, pack32_op, Assembler::AVX_512bit);
2472 __ vpermb(merged0, pack24bits, merged0, Assembler::AVX_512bit);
2473
2474 __ evmovdqub(Address(dest, dp), k2, merged0, true, Assembler::AVX_512bit);
2475
2476 __ subl(length, 64);
2477 __ addptr(source, 64);
2478 __ addptr(dest, 48);
2479
2480 __ cmpl(length, 64);
2481 __ jcc(Assembler::greaterEqual, L_process64Loop);
2482
2483 __ cmpl(length, 0);
2484 __ jcc(Assembler::lessEqual, L_exit);
2485
2486 __ BIND(L_finalBit);
2487 // Now have 1 to 63 bytes left to decode
2488
2489 // I was going to let Java take care of the final fragment
2490 // however it will repeatedly call this routine for every 4 bytes
2491 // of input data, so handle the rest here.
2492 __ movq(rax, -1);
2493 __ bzhiq(rax, rax, length); // Input mask in rax
2494
2495 __ movl(output_size, length);
2496 __ shrl(output_size, 2); // Find (len / 4) * 3 (output length)
2497 __ lea(output_size, Address(output_size, output_size, Address::times_2, 0));
2498 // output_size in r13
2499
2500 // Strip pad characters, if any, and adjust length and mask
2501 __ cmpb(Address(source, length, Address::times_1, -1), '=');
2502 __ jcc(Assembler::equal, L_padding);
2503
2504 __ BIND(L_donePadding);
2505
2506 // Output size is (64 - output_size), output mask is (all 1s >> output_size).
2507 __ kmovql(input_mask, rax);
2508 __ movq(output_mask, -1);
2509 __ bzhiq(output_mask, output_mask, output_size);
2510
2511 // Load initial input with all valid base64 characters. Will be used
2512 // in merging source bytes to avoid masking when determining if an error occurred.
2513 __ movl(rax, 0x61616161);
2514 __ evpbroadcastd(input_initial_valid_b64, rax, Assembler::AVX_512bit);
2515
2516 // A register containing all invalid base64 decoded values
2517 __ movl(rax, 0x80808080);
2518 __ evpbroadcastd(invalid_b64, rax, Assembler::AVX_512bit);
2519
2520 // input_mask is in k1
2521 // output_size is in r13
2522 // output_mask is in r15
2523 // zmm0 - free
2524 // zmm1 - 0x00011000
2525 // zmm2 - 0x01400140
2526 // zmm3 - errorvec
2527 // zmm4 - pack vector
2528 // zmm5 - lookup_lo
2529 // zmm6 - lookup_hi
2530 // zmm7 - errorvec
2531 // zmm8 - 0x61616161
2532 // zmm9 - 0x80808080
2533
2534 // Load only the bytes from source, merging into our "fully-valid" register
2535 __ evmovdqub(input_initial_valid_b64, input_mask, Address(source, start_offset, Address::times_1, 0x0), true, Assembler::AVX_512bit);
2536
2537 // Decode all bytes within our merged input
2538 __ evmovdquq(tmp, lookup_lo, Assembler::AVX_512bit);
2539 __ evpermt2b(tmp, input_initial_valid_b64, lookup_hi, Assembler::AVX_512bit);
2540 __ evporq(mask, tmp, input_initial_valid_b64, Assembler::AVX_512bit);
2541
2542 // Check for error. Compare (decoded | initial) to all invalid.
2543 // If any bytes have their high-order bit set, then we have an error.
2544 __ evptestmb(k2, mask, invalid_b64, Assembler::AVX_512bit);
2545 __ kortestql(k2, k2);
2546
2547 // If we have an error, use the brute force loop to decode what we can (4-byte chunks).
2548 __ jcc(Assembler::notZero, L_bruteForce);
2549
2550 // Shuffle output bytes
2551 __ vpmaddubsw(tmp, tmp, pack16_op, Assembler::AVX_512bit);
2552 __ vpmaddwd(tmp, tmp, pack32_op, Assembler::AVX_512bit);
2553
2554 __ vpermb(tmp, pack24bits, tmp, Assembler::AVX_512bit);
2555 __ kmovql(k1, output_mask);
2556 __ evmovdqub(Address(dest, dp), k1, tmp, true, Assembler::AVX_512bit);
2557
2558 __ addptr(dest, output_size);
2559
2560 __ BIND(L_exit);
2561 __ vzeroupper();
2562 __ pop(rax); // Get original dest value
2563 __ subptr(dest, rax); // Number of bytes converted
2564 __ movptr(rax, dest);
2565 __ pop(rbx);
2566 __ pop(r15);
2567 __ pop(r14);
2568 __ pop(r13);
2569 __ pop(r12);
2570 __ leave();
2571 __ ret(0);
2572
2573 __ BIND(L_loadURL);
2574 __ evmovdquq(lookup_lo, ExternalAddress(StubRoutines::x86::base64_vbmi_lookup_lo_url_addr()), Assembler::AVX_512bit, r13);
2575 __ evmovdquq(lookup_hi, ExternalAddress(StubRoutines::x86::base64_vbmi_lookup_hi_url_addr()), Assembler::AVX_512bit, r13);
2576 __ jmp(L_continue);
2577
2578 __ BIND(L_padding);
2579 __ decrementq(output_size, 1);
2580 __ shrq(rax, 1);
2581
2582 __ cmpb(Address(source, length, Address::times_1, -2), '=');
2583 __ jcc(Assembler::notEqual, L_donePadding);
2584
2585 __ decrementq(output_size, 1);
2586 __ shrq(rax, 1);
2587 __ jmp(L_donePadding);
2588
2589 __ align32();
2590 __ BIND(L_bruteForce);
2591 } // End of if(avx512_vbmi)
2592
2593 // Use non-AVX code to decode 4-byte chunks into 3 bytes of output
2594
2595 // Register state (Linux):
2596 // r12-15 - saved on stack
2597 // rdi - src
2598 // rsi - sp
2599 // rdx - sl
2600 // rcx - dst
2601 // r8 - dp
2602 // r9 - isURL
2603
2604 // Register state (Windows):
2605 // r12-15 - saved on stack
2606 // rcx - src
2607 // rdx - sp
2608 // r8 - sl
2609 // r9 - dst
2610 // r12 - dp
2611 // r10 - isURL
2612
2613 // Registers (common):
2614 // length (r14) - bytes in src
2615
2616 const Register decode_table = r11;
2617 const Register out_byte_count = rbx;
2618 const Register byte1 = r13;
2619 const Register byte2 = r15;
2620 const Register byte3 = WIN64_ONLY(r8) NOT_WIN64(rdx);
2621 const Register byte4 = WIN64_ONLY(r10) NOT_WIN64(r9);
2622
2623 __ shrl(length, 2); // Multiple of 4 bytes only - length is # 4-byte chunks
2624 __ cmpl(length, 0);
2625 __ jcc(Assembler::lessEqual, L_exit_no_vzero);
2626
2627 __ shll(isURL, 8); // index into decode table based on isURL
2628 __ lea(decode_table, ExternalAddress(StubRoutines::x86::base64_decoding_table_addr()));
2629 __ addptr(decode_table, isURL);
2630
2631 __ jmp(L_bottomLoop);
2632
2633 __ align32();
2634 __ BIND(L_forceLoop);
2635 __ shll(byte1, 18);
2636 __ shll(byte2, 12);
2637 __ shll(byte3, 6);
2638 __ orl(byte1, byte2);
2639 __ orl(byte1, byte3);
2640 __ orl(byte1, byte4);
2641
2642 __ addptr(source, 4);
2643
2644 __ movb(Address(dest, dp, Address::times_1, 2), byte1);
2645 __ shrl(byte1, 8);
2646 __ movb(Address(dest, dp, Address::times_1, 1), byte1);
2647 __ shrl(byte1, 8);
2648 __ movb(Address(dest, dp, Address::times_1, 0), byte1);
2649
2650 __ addptr(dest, 3);
2651 __ decrementl(length, 1);
2652 __ jcc(Assembler::zero, L_exit_no_vzero);
2653
2654 __ BIND(L_bottomLoop);
2655 __ load_unsigned_byte(byte1, Address(source, start_offset, Address::times_1, 0x00));
2656 __ load_unsigned_byte(byte2, Address(source, start_offset, Address::times_1, 0x01));
2657 __ load_signed_byte(byte1, Address(decode_table, byte1));
2658 __ load_signed_byte(byte2, Address(decode_table, byte2));
2659 __ load_unsigned_byte(byte3, Address(source, start_offset, Address::times_1, 0x02));
2660 __ load_unsigned_byte(byte4, Address(source, start_offset, Address::times_1, 0x03));
2661 __ load_signed_byte(byte3, Address(decode_table, byte3));
2662 __ load_signed_byte(byte4, Address(decode_table, byte4));
2663
2664 __ mov(rax, byte1);
2665 __ orl(rax, byte2);
2666 __ orl(rax, byte3);
2667 __ orl(rax, byte4);
2668 __ jcc(Assembler::positive, L_forceLoop);
2669
2670 __ BIND(L_exit_no_vzero);
2671 __ pop(rax); // Get original dest value
2672 __ subptr(dest, rax); // Number of bytes converted
2673 __ movptr(rax, dest);
2674 __ pop(rbx);
2675 __ pop(r15);
2676 __ pop(r14);
2677 __ pop(r13);
2678 __ pop(r12);
2679 __ leave();
2680 __ ret(0);
2681
2682 return start;
2683 }
2684
2685
2686 /**
2687 * Arguments:
2688 *
2689 * Inputs:
2690 * c_rarg0 - int crc
2691 * c_rarg1 - byte* buf
2692 * c_rarg2 - int length
2693 *
2694 * Output:
2695 * rax - int crc result
2696 */
2697 address StubGenerator::generate_updateBytesCRC32() {
2698 assert(UseCRC32Intrinsics, "need AVX and CLMUL instructions");
2699
2700 __ align(CodeEntryAlignment);
2701 StubCodeMark mark(this, "StubRoutines", "updateBytesCRC32");
2702
2703 address start = __ pc();
2704
2705 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
2706 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
2707 // rscratch1: r10
2708 const Register crc = c_rarg0; // crc
2709 const Register buf = c_rarg1; // source java byte array address
2710 const Register len = c_rarg2; // length
2711 const Register table = c_rarg3; // crc_table address (reuse register)
2712 const Register tmp1 = r11;
2713 const Register tmp2 = r10;
2714 assert_different_registers(crc, buf, len, table, tmp1, tmp2, rax);
2715
2716 BLOCK_COMMENT("Entry:");
2717 __ enter(); // required for proper stackwalking of RuntimeStub frame
2718
2719 if (VM_Version::supports_sse4_1() && VM_Version::supports_avx512_vpclmulqdq() &&
2720 VM_Version::supports_avx512bw() &&
2721 VM_Version::supports_avx512vl()) {
2722 // The constants used in the CRC32 algorithm requires the 1's compliment of the initial crc value.
2723 // However, the constant table for CRC32-C assumes the original crc value. Account for this
2724 // difference before calling and after returning.
2725 __ lea(table, ExternalAddress(StubRoutines::x86::crc_table_avx512_addr()));
2726 __ notl(crc);
2727 __ kernel_crc32_avx512(crc, buf, len, table, tmp1, tmp2);
2728 __ notl(crc);
2729 } else {
2730 __ kernel_crc32(crc, buf, len, table, tmp1);
2731 }
2732
2733 __ movl(rax, crc);
2734 __ vzeroupper();
2735 __ leave(); // required for proper stackwalking of RuntimeStub frame
2736 __ ret(0);
2737
2738 return start;
2739 }
2740
2741 /**
2742 * Arguments:
2743 *
2744 * Inputs:
2745 * c_rarg0 - int crc
2746 * c_rarg1 - byte* buf
2747 * c_rarg2 - long length
2748 * c_rarg3 - table_start - optional (present only when doing a library_call,
2749 * not used by x86 algorithm)
2750 *
2751 * Output:
2752 * rax - int crc result
2753 */
2754 address StubGenerator::generate_updateBytesCRC32C(bool is_pclmulqdq_supported) {
2755 assert(UseCRC32CIntrinsics, "need SSE4_2");
2756 __ align(CodeEntryAlignment);
2757 StubCodeMark mark(this, "StubRoutines", "updateBytesCRC32C");
2758 address start = __ pc();
2759
2760 //reg.arg int#0 int#1 int#2 int#3 int#4 int#5 float regs
2761 //Windows RCX RDX R8 R9 none none XMM0..XMM3
2762 //Lin / Sol RDI RSI RDX RCX R8 R9 XMM0..XMM7
2763 const Register crc = c_rarg0; // crc
2764 const Register buf = c_rarg1; // source java byte array address
2765 const Register len = c_rarg2; // length
2766 const Register a = rax;
2767 const Register j = r9;
2768 const Register k = r10;
2769 const Register l = r11;
2770 #ifdef _WIN64
2771 const Register y = rdi;
2772 const Register z = rsi;
2773 #else
2774 const Register y = rcx;
2775 const Register z = r8;
2776 #endif
2777 assert_different_registers(crc, buf, len, a, j, k, l, y, z);
2778
2779 BLOCK_COMMENT("Entry:");
2780 __ enter(); // required for proper stackwalking of RuntimeStub frame
2781 if (VM_Version::supports_sse4_1() && VM_Version::supports_avx512_vpclmulqdq() &&
2782 VM_Version::supports_avx512bw() &&
2783 VM_Version::supports_avx512vl()) {
2784 __ lea(j, ExternalAddress(StubRoutines::x86::crc32c_table_avx512_addr()));
2785 __ kernel_crc32_avx512(crc, buf, len, j, l, k);
2786 } else {
2787 #ifdef _WIN64
2788 __ push(y);
2789 __ push(z);
2790 #endif
2791 __ crc32c_ipl_alg2_alt2(crc, buf, len,
2792 a, j, k,
2793 l, y, z,
2794 c_farg0, c_farg1, c_farg2,
2795 is_pclmulqdq_supported);
2796 #ifdef _WIN64
2797 __ pop(z);
2798 __ pop(y);
2799 #endif
2800 }
2801 __ movl(rax, crc);
2802 __ vzeroupper();
2803 __ leave(); // required for proper stackwalking of RuntimeStub frame
2804 __ ret(0);
2805
2806 return start;
2807 }
2808
2809
2810 /**
2811 * Arguments:
2812 *
2813 * Input:
2814 * c_rarg0 - x address
2815 * c_rarg1 - x length
2816 * c_rarg2 - y address
2817 * c_rarg3 - y length
2818 * not Win64
2819 * c_rarg4 - z address
2820 * c_rarg5 - z length
2821 * Win64
2822 * rsp+40 - z address
2823 * rsp+48 - z length
2824 */
2825 address StubGenerator::generate_multiplyToLen() {
2826 __ align(CodeEntryAlignment);
2827 StubCodeMark mark(this, "StubRoutines", "multiplyToLen");
2828 address start = __ pc();
2829
2830 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
2831 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
2832 const Register x = rdi;
2833 const Register xlen = rax;
2834 const Register y = rsi;
2835 const Register ylen = rcx;
2836 const Register z = r8;
2837 const Register zlen = r11;
2838
2839 // Next registers will be saved on stack in multiply_to_len().
2840 const Register tmp1 = r12;
2841 const Register tmp2 = r13;
2842 const Register tmp3 = r14;
2843 const Register tmp4 = r15;
2844 const Register tmp5 = rbx;
2845
2846 BLOCK_COMMENT("Entry:");
2847 __ enter(); // required for proper stackwalking of RuntimeStub frame
2848
2849 #ifndef _WIN64
2850 __ movptr(zlen, r9); // Save r9 in r11 - zlen
2851 #endif
2852 setup_arg_regs(4); // x => rdi, xlen => rsi, y => rdx
2853 // ylen => rcx, z => r8, zlen => r11
2854 // r9 and r10 may be used to save non-volatile registers
2855 #ifdef _WIN64
2856 // last 2 arguments (#4, #5) are on stack on Win64
2857 __ movptr(z, Address(rsp, 6 * wordSize));
2858 __ movptr(zlen, Address(rsp, 7 * wordSize));
2859 #endif
2860
2861 __ movptr(xlen, rsi);
2862 __ movptr(y, rdx);
2863 __ multiply_to_len(x, xlen, y, ylen, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5);
2864
2865 restore_arg_regs();
2866
2867 __ leave(); // required for proper stackwalking of RuntimeStub frame
2868 __ ret(0);
2869
2870 return start;
2871 }
2872
2873 /**
2874 * Arguments:
2875 *
2876 * Input:
2877 * c_rarg0 - obja address
2878 * c_rarg1 - objb address
2879 * c_rarg3 - length length
2880 * c_rarg4 - scale log2_array_indxscale
2881 *
2882 * Output:
2883 * rax - int >= mismatched index, < 0 bitwise complement of tail
2884 */
2885 address StubGenerator::generate_vectorizedMismatch() {
2886 __ align(CodeEntryAlignment);
2887 StubCodeMark mark(this, "StubRoutines", "vectorizedMismatch");
2888 address start = __ pc();
2889
2890 BLOCK_COMMENT("Entry:");
2891 __ enter();
2892
2893 #ifdef _WIN64 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
2894 const Register scale = c_rarg0; //rcx, will exchange with r9
2895 const Register objb = c_rarg1; //rdx
2896 const Register length = c_rarg2; //r8
2897 const Register obja = c_rarg3; //r9
2898 __ xchgq(obja, scale); //now obja and scale contains the correct contents
2899
2900 const Register tmp1 = r10;
2901 const Register tmp2 = r11;
2902 #endif
2903 #ifndef _WIN64 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
2904 const Register obja = c_rarg0; //U:rdi
2905 const Register objb = c_rarg1; //U:rsi
2906 const Register length = c_rarg2; //U:rdx
2907 const Register scale = c_rarg3; //U:rcx
2908 const Register tmp1 = r8;
2909 const Register tmp2 = r9;
2910 #endif
2911 const Register result = rax; //return value
2912 const XMMRegister vec0 = xmm0;
2913 const XMMRegister vec1 = xmm1;
2914 const XMMRegister vec2 = xmm2;
2915
2916 __ vectorized_mismatch(obja, objb, length, scale, result, tmp1, tmp2, vec0, vec1, vec2);
2917
2918 __ vzeroupper();
2919 __ leave();
2920 __ ret(0);
2921
2922 return start;
2923 }
2924
2925 /**
2926 * Arguments:
2927 *
2928 // Input:
2929 // c_rarg0 - x address
2930 // c_rarg1 - x length
2931 // c_rarg2 - z address
2932 // c_rarg3 - z length
2933 *
2934 */
2935 address StubGenerator::generate_squareToLen() {
2936
2937 __ align(CodeEntryAlignment);
2938 StubCodeMark mark(this, "StubRoutines", "squareToLen");
2939 address start = __ pc();
2940
2941 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
2942 // Unix: rdi, rsi, rdx, rcx (c_rarg0, c_rarg1, ...)
2943 const Register x = rdi;
2944 const Register len = rsi;
2945 const Register z = r8;
2946 const Register zlen = rcx;
2947
2948 const Register tmp1 = r12;
2949 const Register tmp2 = r13;
2950 const Register tmp3 = r14;
2951 const Register tmp4 = r15;
2952 const Register tmp5 = rbx;
2953
2954 BLOCK_COMMENT("Entry:");
2955 __ enter(); // required for proper stackwalking of RuntimeStub frame
2956
2957 setup_arg_regs(4); // x => rdi, len => rsi, z => rdx
2958 // zlen => rcx
2959 // r9 and r10 may be used to save non-volatile registers
2960 __ movptr(r8, rdx);
2961 __ square_to_len(x, len, z, zlen, tmp1, tmp2, tmp3, tmp4, tmp5, rdx, rax);
2962
2963 restore_arg_regs();
2964
2965 __ leave(); // required for proper stackwalking of RuntimeStub frame
2966 __ ret(0);
2967
2968 return start;
2969 }
2970
2971 address StubGenerator::generate_method_entry_barrier() {
2972 __ align(CodeEntryAlignment);
2973 StubCodeMark mark(this, "StubRoutines", "nmethod_entry_barrier");
2974 address start = __ pc();
2975
2976 Label deoptimize_label;
2977
2978 __ push(-1); // cookie, this is used for writing the new rsp when deoptimizing
2979
2980 BLOCK_COMMENT("Entry:");
2981 __ enter(); // save rbp
2982
2983 // save c_rarg0, because we want to use that value.
2984 // We could do without it but then we depend on the number of slots used by pusha
2985 __ push(c_rarg0);
2986
2987 __ lea(c_rarg0, Address(rsp, wordSize * 3)); // 1 for cookie, 1 for rbp, 1 for c_rarg0 - this should be the return address
2988
2989 __ pusha();
2990
2991 // The method may have floats as arguments, and we must spill them before calling
2992 // the VM runtime.
2993 assert(Argument::n_float_register_parameters_j == 8, "Assumption");
2994 const int xmm_size = wordSize * 2;
2995 const int xmm_spill_size = xmm_size * Argument::n_float_register_parameters_j;
2996 __ subptr(rsp, xmm_spill_size);
2997 __ movdqu(Address(rsp, xmm_size * 7), xmm7);
2998 __ movdqu(Address(rsp, xmm_size * 6), xmm6);
2999 __ movdqu(Address(rsp, xmm_size * 5), xmm5);
3000 __ movdqu(Address(rsp, xmm_size * 4), xmm4);
3001 __ movdqu(Address(rsp, xmm_size * 3), xmm3);
3002 __ movdqu(Address(rsp, xmm_size * 2), xmm2);
3003 __ movdqu(Address(rsp, xmm_size * 1), xmm1);
3004 __ movdqu(Address(rsp, xmm_size * 0), xmm0);
3005
3006 __ call_VM_leaf(CAST_FROM_FN_PTR(address, static_cast<int (*)(address*)>(BarrierSetNMethod::nmethod_stub_entry_barrier)), 1);
3007
3008 __ movdqu(xmm0, Address(rsp, xmm_size * 0));
3009 __ movdqu(xmm1, Address(rsp, xmm_size * 1));
3010 __ movdqu(xmm2, Address(rsp, xmm_size * 2));
3011 __ movdqu(xmm3, Address(rsp, xmm_size * 3));
3012 __ movdqu(xmm4, Address(rsp, xmm_size * 4));
3013 __ movdqu(xmm5, Address(rsp, xmm_size * 5));
3014 __ movdqu(xmm6, Address(rsp, xmm_size * 6));
3015 __ movdqu(xmm7, Address(rsp, xmm_size * 7));
3016 __ addptr(rsp, xmm_spill_size);
3017
3018 __ cmpl(rax, 1); // 1 means deoptimize
3019 __ jcc(Assembler::equal, deoptimize_label);
3020
3021 __ popa();
3022 __ pop(c_rarg0);
3023
3024 __ leave();
3025
3026 __ addptr(rsp, 1 * wordSize); // cookie
3027 __ ret(0);
3028
3029
3030 __ BIND(deoptimize_label);
3031
3032 __ popa();
3033 __ pop(c_rarg0);
3034
3035 __ leave();
3036
3037 // this can be taken out, but is good for verification purposes. getting a SIGSEGV
3038 // here while still having a correct stack is valuable
3039 __ testptr(rsp, Address(rsp, 0));
3040
3041 __ movptr(rsp, Address(rsp, 0)); // new rsp was written in the barrier
3042 __ jmp(Address(rsp, -1 * wordSize)); // jmp target should be callers verified_entry_point
3043
3044 return start;
3045 }
3046
3047 /**
3048 * Arguments:
3049 *
3050 * Input:
3051 * c_rarg0 - out address
3052 * c_rarg1 - in address
3053 * c_rarg2 - offset
3054 * c_rarg3 - len
3055 * not Win64
3056 * c_rarg4 - k
3057 * Win64
3058 * rsp+40 - k
3059 */
3060 address StubGenerator::generate_mulAdd() {
3061 __ align(CodeEntryAlignment);
3062 StubCodeMark mark(this, "StubRoutines", "mulAdd");
3063 address start = __ pc();
3064
3065 // Win64: rcx, rdx, r8, r9 (c_rarg0, c_rarg1, ...)
3066 // Unix: rdi, rsi, rdx, rcx, r8, r9 (c_rarg0, c_rarg1, ...)
3067 const Register out = rdi;
3068 const Register in = rsi;
3069 const Register offset = r11;
3070 const Register len = rcx;
3071 const Register k = r8;
3072
3073 // Next registers will be saved on stack in mul_add().
3074 const Register tmp1 = r12;
3075 const Register tmp2 = r13;
3076 const Register tmp3 = r14;
3077 const Register tmp4 = r15;
3078 const Register tmp5 = rbx;
3079
3080 BLOCK_COMMENT("Entry:");
3081 __ enter(); // required for proper stackwalking of RuntimeStub frame
3082
3083 setup_arg_regs(4); // out => rdi, in => rsi, offset => rdx
3084 // len => rcx, k => r8
3085 // r9 and r10 may be used to save non-volatile registers
3086 #ifdef _WIN64
3087 // last argument is on stack on Win64
3088 __ movl(k, Address(rsp, 6 * wordSize));
3089 #endif
3090 __ movptr(r11, rdx); // move offset in rdx to offset(r11)
3091 __ mul_add(out, in, offset, len, k, tmp1, tmp2, tmp3, tmp4, tmp5, rdx, rax);
3092
3093 restore_arg_regs();
3094
3095 __ leave(); // required for proper stackwalking of RuntimeStub frame
3096 __ ret(0);
3097
3098 return start;
3099 }
3100
3101 address StubGenerator::generate_bigIntegerRightShift() {
3102 __ align(CodeEntryAlignment);
3103 StubCodeMark mark(this, "StubRoutines", "bigIntegerRightShiftWorker");
3104 address start = __ pc();
3105
3106 Label Shift512Loop, ShiftTwo, ShiftTwoLoop, ShiftOne, Exit;
3107 // For Unix, the arguments are as follows: rdi, rsi, rdx, rcx, r8.
3108 const Register newArr = rdi;
3109 const Register oldArr = rsi;
3110 const Register newIdx = rdx;
3111 const Register shiftCount = rcx; // It was intentional to have shiftCount in rcx since it is used implicitly for shift.
3112 const Register totalNumIter = r8;
3113
3114 // For windows, we use r9 and r10 as temps to save rdi and rsi. Thus we cannot allocate them for our temps.
3115 // For everything else, we prefer using r9 and r10 since we do not have to save them before use.
3116 const Register tmp1 = r11; // Caller save.
3117 const Register tmp2 = rax; // Caller save.
3118 const Register tmp3 = WIN64_ONLY(r12) NOT_WIN64(r9); // Windows: Callee save. Linux: Caller save.
3119 const Register tmp4 = WIN64_ONLY(r13) NOT_WIN64(r10); // Windows: Callee save. Linux: Caller save.
3120 const Register tmp5 = r14; // Callee save.
3121 const Register tmp6 = r15;
3122
3123 const XMMRegister x0 = xmm0;
3124 const XMMRegister x1 = xmm1;
3125 const XMMRegister x2 = xmm2;
3126
3127 BLOCK_COMMENT("Entry:");
3128 __ enter(); // required for proper stackwalking of RuntimeStub frame
3129
3130 #ifdef _WIN64
3131 setup_arg_regs(4);
3132 // For windows, since last argument is on stack, we need to move it to the appropriate register.
3133 __ movl(totalNumIter, Address(rsp, 6 * wordSize));
3134 // Save callee save registers.
3135 __ push(tmp3);
3136 __ push(tmp4);
3137 #endif
3138 __ push(tmp5);
3139
3140 // Rename temps used throughout the code.
3141 const Register idx = tmp1;
3142 const Register nIdx = tmp2;
3143
3144 __ xorl(idx, idx);
3145
3146 // Start right shift from end of the array.
3147 // For example, if #iteration = 4 and newIdx = 1
3148 // then dest[4] = src[4] >> shiftCount | src[3] <<< (shiftCount - 32)
3149 // if #iteration = 4 and newIdx = 0
3150 // then dest[3] = src[4] >> shiftCount | src[3] <<< (shiftCount - 32)
3151 __ movl(idx, totalNumIter);
3152 __ movl(nIdx, idx);
3153 __ addl(nIdx, newIdx);
3154
3155 // If vectorization is enabled, check if the number of iterations is at least 64
3156 // If not, then go to ShifTwo processing 2 iterations
3157 if (VM_Version::supports_avx512_vbmi2()) {
3158 __ cmpptr(totalNumIter, (AVX3Threshold/64));
3159 __ jcc(Assembler::less, ShiftTwo);
3160
3161 if (AVX3Threshold < 16 * 64) {
3162 __ cmpl(totalNumIter, 16);
3163 __ jcc(Assembler::less, ShiftTwo);
3164 }
3165 __ evpbroadcastd(x0, shiftCount, Assembler::AVX_512bit);
3166 __ subl(idx, 16);
3167 __ subl(nIdx, 16);
3168 __ BIND(Shift512Loop);
3169 __ evmovdqul(x2, Address(oldArr, idx, Address::times_4, 4), Assembler::AVX_512bit);
3170 __ evmovdqul(x1, Address(oldArr, idx, Address::times_4), Assembler::AVX_512bit);
3171 __ vpshrdvd(x2, x1, x0, Assembler::AVX_512bit);
3172 __ evmovdqul(Address(newArr, nIdx, Address::times_4), x2, Assembler::AVX_512bit);
3173 __ subl(nIdx, 16);
3174 __ subl(idx, 16);
3175 __ jcc(Assembler::greaterEqual, Shift512Loop);
3176 __ addl(idx, 16);
3177 __ addl(nIdx, 16);
3178 }
3179 __ BIND(ShiftTwo);
3180 __ cmpl(idx, 2);
3181 __ jcc(Assembler::less, ShiftOne);
3182 __ subl(idx, 2);
3183 __ subl(nIdx, 2);
3184 __ BIND(ShiftTwoLoop);
3185 __ movl(tmp5, Address(oldArr, idx, Address::times_4, 8));
3186 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 4));
3187 __ movl(tmp3, Address(oldArr, idx, Address::times_4));
3188 __ shrdl(tmp5, tmp4);
3189 __ shrdl(tmp4, tmp3);
3190 __ movl(Address(newArr, nIdx, Address::times_4, 4), tmp5);
3191 __ movl(Address(newArr, nIdx, Address::times_4), tmp4);
3192 __ subl(nIdx, 2);
3193 __ subl(idx, 2);
3194 __ jcc(Assembler::greaterEqual, ShiftTwoLoop);
3195 __ addl(idx, 2);
3196 __ addl(nIdx, 2);
3197
3198 // Do the last iteration
3199 __ BIND(ShiftOne);
3200 __ cmpl(idx, 1);
3201 __ jcc(Assembler::less, Exit);
3202 __ subl(idx, 1);
3203 __ subl(nIdx, 1);
3204 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 4));
3205 __ movl(tmp3, Address(oldArr, idx, Address::times_4));
3206 __ shrdl(tmp4, tmp3);
3207 __ movl(Address(newArr, nIdx, Address::times_4), tmp4);
3208 __ BIND(Exit);
3209 __ vzeroupper();
3210 // Restore callee save registers.
3211 __ pop(tmp5);
3212 #ifdef _WIN64
3213 __ pop(tmp4);
3214 __ pop(tmp3);
3215 restore_arg_regs();
3216 #endif
3217 __ leave(); // required for proper stackwalking of RuntimeStub frame
3218 __ ret(0);
3219
3220 return start;
3221 }
3222
3223 /**
3224 * Arguments:
3225 *
3226 * Input:
3227 * c_rarg0 - newArr address
3228 * c_rarg1 - oldArr address
3229 * c_rarg2 - newIdx
3230 * c_rarg3 - shiftCount
3231 * not Win64
3232 * c_rarg4 - numIter
3233 * Win64
3234 * rsp40 - numIter
3235 */
3236 address StubGenerator::generate_bigIntegerLeftShift() {
3237 __ align(CodeEntryAlignment);
3238 StubCodeMark mark(this, "StubRoutines", "bigIntegerLeftShiftWorker");
3239 address start = __ pc();
3240
3241 Label Shift512Loop, ShiftTwo, ShiftTwoLoop, ShiftOne, Exit;
3242 // For Unix, the arguments are as follows: rdi, rsi, rdx, rcx, r8.
3243 const Register newArr = rdi;
3244 const Register oldArr = rsi;
3245 const Register newIdx = rdx;
3246 const Register shiftCount = rcx; // It was intentional to have shiftCount in rcx since it is used implicitly for shift.
3247 const Register totalNumIter = r8;
3248 // For windows, we use r9 and r10 as temps to save rdi and rsi. Thus we cannot allocate them for our temps.
3249 // For everything else, we prefer using r9 and r10 since we do not have to save them before use.
3250 const Register tmp1 = r11; // Caller save.
3251 const Register tmp2 = rax; // Caller save.
3252 const Register tmp3 = WIN64_ONLY(r12) NOT_WIN64(r9); // Windows: Callee save. Linux: Caller save.
3253 const Register tmp4 = WIN64_ONLY(r13) NOT_WIN64(r10); // Windows: Callee save. Linux: Caller save.
3254 const Register tmp5 = r14; // Callee save.
3255
3256 const XMMRegister x0 = xmm0;
3257 const XMMRegister x1 = xmm1;
3258 const XMMRegister x2 = xmm2;
3259 BLOCK_COMMENT("Entry:");
3260 __ enter(); // required for proper stackwalking of RuntimeStub frame
3261
3262 #ifdef _WIN64
3263 setup_arg_regs(4);
3264 // For windows, since last argument is on stack, we need to move it to the appropriate register.
3265 __ movl(totalNumIter, Address(rsp, 6 * wordSize));
3266 // Save callee save registers.
3267 __ push(tmp3);
3268 __ push(tmp4);
3269 #endif
3270 __ push(tmp5);
3271
3272 // Rename temps used throughout the code
3273 const Register idx = tmp1;
3274 const Register numIterTmp = tmp2;
3275
3276 // Start idx from zero.
3277 __ xorl(idx, idx);
3278 // Compute interior pointer for new array. We do this so that we can use same index for both old and new arrays.
3279 __ lea(newArr, Address(newArr, newIdx, Address::times_4));
3280 __ movl(numIterTmp, totalNumIter);
3281
3282 // If vectorization is enabled, check if the number of iterations is at least 64
3283 // If not, then go to ShiftTwo shifting two numbers at a time
3284 if (VM_Version::supports_avx512_vbmi2()) {
3285 __ cmpl(totalNumIter, (AVX3Threshold/64));
3286 __ jcc(Assembler::less, ShiftTwo);
3287
3288 if (AVX3Threshold < 16 * 64) {
3289 __ cmpl(totalNumIter, 16);
3290 __ jcc(Assembler::less, ShiftTwo);
3291 }
3292 __ evpbroadcastd(x0, shiftCount, Assembler::AVX_512bit);
3293 __ subl(numIterTmp, 16);
3294 __ BIND(Shift512Loop);
3295 __ evmovdqul(x1, Address(oldArr, idx, Address::times_4), Assembler::AVX_512bit);
3296 __ evmovdqul(x2, Address(oldArr, idx, Address::times_4, 0x4), Assembler::AVX_512bit);
3297 __ vpshldvd(x1, x2, x0, Assembler::AVX_512bit);
3298 __ evmovdqul(Address(newArr, idx, Address::times_4), x1, Assembler::AVX_512bit);
3299 __ addl(idx, 16);
3300 __ subl(numIterTmp, 16);
3301 __ jcc(Assembler::greaterEqual, Shift512Loop);
3302 __ addl(numIterTmp, 16);
3303 }
3304 __ BIND(ShiftTwo);
3305 __ cmpl(totalNumIter, 1);
3306 __ jcc(Assembler::less, Exit);
3307 __ movl(tmp3, Address(oldArr, idx, Address::times_4));
3308 __ subl(numIterTmp, 2);
3309 __ jcc(Assembler::less, ShiftOne);
3310
3311 __ BIND(ShiftTwoLoop);
3312 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 0x4));
3313 __ movl(tmp5, Address(oldArr, idx, Address::times_4, 0x8));
3314 __ shldl(tmp3, tmp4);
3315 __ shldl(tmp4, tmp5);
3316 __ movl(Address(newArr, idx, Address::times_4), tmp3);
3317 __ movl(Address(newArr, idx, Address::times_4, 0x4), tmp4);
3318 __ movl(tmp3, tmp5);
3319 __ addl(idx, 2);
3320 __ subl(numIterTmp, 2);
3321 __ jcc(Assembler::greaterEqual, ShiftTwoLoop);
3322
3323 // Do the last iteration
3324 __ BIND(ShiftOne);
3325 __ addl(numIterTmp, 2);
3326 __ cmpl(numIterTmp, 1);
3327 __ jcc(Assembler::less, Exit);
3328 __ movl(tmp4, Address(oldArr, idx, Address::times_4, 0x4));
3329 __ shldl(tmp3, tmp4);
3330 __ movl(Address(newArr, idx, Address::times_4), tmp3);
3331
3332 __ BIND(Exit);
3333 __ vzeroupper();
3334 // Restore callee save registers.
3335 __ pop(tmp5);
3336 #ifdef _WIN64
3337 __ pop(tmp4);
3338 __ pop(tmp3);
3339 restore_arg_regs();
3340 #endif
3341 __ leave(); // required for proper stackwalking of RuntimeStub frame
3342 __ ret(0);
3343
3344 return start;
3345 }
3346
3347 void StubGenerator::generate_libm_stubs() {
3348 if (UseLibmIntrinsic && InlineIntrinsics) {
3349 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dsin)) {
3350 StubRoutines::_dsin = generate_libmSin(); // from stubGenerator_x86_64_sin.cpp
3351 }
3352 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dcos)) {
3353 StubRoutines::_dcos = generate_libmCos(); // from stubGenerator_x86_64_cos.cpp
3354 }
3355 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dtan)) {
3356 StubRoutines::_dtan = generate_libmTan(); // from stubGenerator_x86_64_tan.cpp
3357 }
3358 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dexp)) {
3359 StubRoutines::_dexp = generate_libmExp(); // from stubGenerator_x86_64_exp.cpp
3360 }
3361 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dpow)) {
3362 StubRoutines::_dpow = generate_libmPow(); // from stubGenerator_x86_64_pow.cpp
3363 }
3364 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dlog)) {
3365 StubRoutines::_dlog = generate_libmLog(); // from stubGenerator_x86_64_log.cpp
3366 }
3367 if (vmIntrinsics::is_intrinsic_available(vmIntrinsics::_dlog10)) {
3368 StubRoutines::_dlog10 = generate_libmLog10(); // from stubGenerator_x86_64_log.cpp
3369 }
3370 }
3371 }
3372
3373
3374 address StubGenerator::generate_cont_thaw(const char* label, Continuation::thaw_kind kind) {
3375 if (!Continuations::enabled()) return nullptr;
3376
3377 bool return_barrier = Continuation::is_thaw_return_barrier(kind);
3378 bool return_barrier_exception = Continuation::is_thaw_return_barrier_exception(kind);
3379
3380 StubCodeMark mark(this, "StubRoutines", label);
3381 address start = __ pc();
3382
3383 // TODO: Handle Valhalla return types. May require generating different return barriers.
3384
3385 if (!return_barrier) {
3386 // Pop return address. If we don't do this, we get a drift,
3387 // where the bottom-most frozen frame continuously grows.
3388 __ pop(c_rarg3);
3389 } else {
3390 __ movptr(rsp, Address(r15_thread, JavaThread::cont_entry_offset()));
3391 }
3392
3393 #ifdef ASSERT
3394 {
3395 Label L_good_sp;
3396 __ cmpptr(rsp, Address(r15_thread, JavaThread::cont_entry_offset()));
3397 __ jcc(Assembler::equal, L_good_sp);
3398 __ stop("Incorrect rsp at thaw entry");
3399 __ BIND(L_good_sp);
3400 }
3401 #endif // ASSERT
3402
3403 if (return_barrier) {
3404 // Preserve possible return value from a method returning to the return barrier.
3405 __ push(rax);
3406 __ push_d(xmm0);
3407 }
3408
3409 __ movptr(c_rarg0, r15_thread);
3410 __ movptr(c_rarg1, (return_barrier ? 1 : 0));
3411 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Continuation::prepare_thaw), 2);
3412 __ movptr(rbx, rax);
3413
3414 if (return_barrier) {
3415 // Restore return value from a method returning to the return barrier.
3416 // No safepoint in the call to thaw, so even an oop return value should be OK.
3417 __ pop_d(xmm0);
3418 __ pop(rax);
3419 }
3420
3421 #ifdef ASSERT
3422 {
3423 Label L_good_sp;
3424 __ cmpptr(rsp, Address(r15_thread, JavaThread::cont_entry_offset()));
3425 __ jcc(Assembler::equal, L_good_sp);
3426 __ stop("Incorrect rsp after prepare thaw");
3427 __ BIND(L_good_sp);
3428 }
3429 #endif // ASSERT
3430
3431 // rbx contains the size of the frames to thaw, 0 if overflow or no more frames
3432 Label L_thaw_success;
3433 __ testptr(rbx, rbx);
3434 __ jccb(Assembler::notZero, L_thaw_success);
3435 __ jump(ExternalAddress(StubRoutines::throw_StackOverflowError_entry()));
3436 __ bind(L_thaw_success);
3437
3438 // Make room for the thawed frames and align the stack.
3439 __ subptr(rsp, rbx);
3440 __ andptr(rsp, -StackAlignmentInBytes);
3441
3442 if (return_barrier) {
3443 // Preserve possible return value from a method returning to the return barrier. (Again.)
3444 __ push(rax);
3445 __ push_d(xmm0);
3446 }
3447
3448 // If we want, we can templatize thaw by kind, and have three different entries.
3449 __ movptr(c_rarg0, r15_thread);
3450 __ movptr(c_rarg1, kind);
3451 __ call_VM_leaf(Continuation::thaw_entry(), 2);
3452 __ movptr(rbx, rax);
3453
3454 if (return_barrier) {
3455 // Restore return value from a method returning to the return barrier. (Again.)
3456 // No safepoint in the call to thaw, so even an oop return value should be OK.
3457 __ pop_d(xmm0);
3458 __ pop(rax);
3459 } else {
3460 // Return 0 (success) from doYield.
3461 __ xorptr(rax, rax);
3462 }
3463
3464 // After thawing, rbx is the SP of the yielding frame.
3465 // Move there, and then to saved RBP slot.
3466 __ movptr(rsp, rbx);
3467 __ subptr(rsp, 2*wordSize);
3468
3469 if (return_barrier_exception) {
3470 __ movptr(c_rarg0, r15_thread);
3471 __ movptr(c_rarg1, Address(rsp, wordSize)); // return address
3472
3473 // rax still holds the original exception oop, save it before the call
3474 __ push(rax);
3475
3476 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), 2);
3477 __ movptr(rbx, rax);
3478
3479 // Continue at exception handler:
3480 // rax: exception oop
3481 // rbx: exception handler
3482 // rdx: exception pc
3483 __ pop(rax);
3484 __ verify_oop(rax);
3485 __ pop(rbp); // pop out RBP here too
3486 __ pop(rdx);
3487 __ jmp(rbx);
3488 } else {
3489 // We are "returning" into the topmost thawed frame; see Thaw::push_return_frame
3490 __ pop(rbp);
3491 __ ret(0);
3492 }
3493
3494 return start;
3495 }
3496
3497 address StubGenerator::generate_cont_thaw() {
3498 return generate_cont_thaw("Cont thaw", Continuation::thaw_top);
3499 }
3500
3501 // TODO: will probably need multiple return barriers depending on return type
3502
3503 address StubGenerator::generate_cont_returnBarrier() {
3504 return generate_cont_thaw("Cont thaw return barrier", Continuation::thaw_return_barrier);
3505 }
3506
3507 address StubGenerator::generate_cont_returnBarrier_exception() {
3508 return generate_cont_thaw("Cont thaw return barrier exception", Continuation::thaw_return_barrier_exception);
3509 }
3510
3511 #if INCLUDE_JFR
3512
3513 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
3514 // It returns a jobject handle to the event writer.
3515 // The handle is dereferenced and the return value is the event writer oop.
3516 RuntimeStub* StubGenerator::generate_jfr_write_checkpoint() {
3517 enum layout {
3518 rbp_off,
3519 rbpH_off,
3520 return_off,
3521 return_off2,
3522 framesize // inclusive of return address
3523 };
3524
3525 CodeBuffer code("jfr_write_checkpoint", 512, 64);
3526 MacroAssembler* _masm = new MacroAssembler(&code);
3527 address start = __ pc();
3528
3529 __ enter();
3530 address the_pc = __ pc();
3531
3532 int frame_complete = the_pc - start;
3533
3534 __ set_last_Java_frame(rsp, rbp, the_pc, rscratch1);
3535 __ movptr(c_rarg0, r15_thread);
3536 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), 1);
3537 __ reset_last_Java_frame(true);
3538
3539 // rax is jobject handle result, unpack and process it through a barrier.
3540 Label L_null_jobject;
3541 __ testptr(rax, rax);
3542 __ jcc(Assembler::zero, L_null_jobject);
3543
3544 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
3545 bs->load_at(_masm, ACCESS_READ | IN_NATIVE, T_OBJECT, rax, Address(rax, 0), c_rarg0, r15_thread);
3546
3547 __ bind(L_null_jobject);
3548
3549 __ leave();
3550 __ ret(0);
3551
3552 OopMapSet* oop_maps = new OopMapSet();
3553 OopMap* map = new OopMap(framesize, 1);
3554 oop_maps->add_gc_map(frame_complete, map);
3555
3556 RuntimeStub* stub =
3557 RuntimeStub::new_runtime_stub(code.name(),
3558 &code,
3559 frame_complete,
3560 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3561 oop_maps,
3562 false);
3563 return stub;
3564 }
3565
3566 #endif // INCLUDE_JFR
3567
3568 // Continuation point for throwing of implicit exceptions that are
3569 // not handled in the current activation. Fabricates an exception
3570 // oop and initiates normal exception dispatching in this
3571 // frame. Since we need to preserve callee-saved values (currently
3572 // only for C2, but done for C1 as well) we need a callee-saved oop
3573 // map and therefore have to make these stubs into RuntimeStubs
3574 // rather than BufferBlobs. If the compiler needs all registers to
3575 // be preserved between the fault point and the exception handler
3576 // then it must assume responsibility for that in
3577 // AbstractCompiler::continuation_for_implicit_null_exception or
3578 // continuation_for_implicit_division_by_zero_exception. All other
3579 // implicit exceptions (e.g., NullPointerException or
3580 // AbstractMethodError on entry) are either at call sites or
3581 // otherwise assume that stack unwinding will be initiated, so
3582 // caller saved registers were assumed volatile in the compiler.
3583 address StubGenerator::generate_throw_exception(const char* name,
3584 address runtime_entry,
3585 Register arg1,
3586 Register arg2) {
3587 // Information about frame layout at time of blocking runtime call.
3588 // Note that we only have to preserve callee-saved registers since
3589 // the compilers are responsible for supplying a continuation point
3590 // if they expect all registers to be preserved.
3591 enum layout {
3592 rbp_off = frame::arg_reg_save_area_bytes/BytesPerInt,
3593 rbp_off2,
3594 return_off,
3595 return_off2,
3596 framesize // inclusive of return address
3597 };
3598
3599 int insts_size = 512;
3600 int locs_size = 64;
3601
3602 CodeBuffer code(name, insts_size, locs_size);
3603 OopMapSet* oop_maps = new OopMapSet();
3604 MacroAssembler* _masm = new MacroAssembler(&code);
3605
3606 address start = __ pc();
3607
3608 // This is an inlined and slightly modified version of call_VM
3609 // which has the ability to fetch the return PC out of
3610 // thread-local storage and also sets up last_Java_sp slightly
3611 // differently than the real call_VM
3612
3613 __ enter(); // required for proper stackwalking of RuntimeStub frame
3614
3615 assert(is_even(framesize/2), "sp not 16-byte aligned");
3616
3617 // return address and rbp are already in place
3618 __ subptr(rsp, (framesize-4) << LogBytesPerInt); // prolog
3619
3620 int frame_complete = __ pc() - start;
3621
3622 // Set up last_Java_sp and last_Java_fp
3623 address the_pc = __ pc();
3624 __ set_last_Java_frame(rsp, rbp, the_pc, rscratch1);
3625 __ andptr(rsp, -(StackAlignmentInBytes)); // Align stack
3626
3627 // Call runtime
3628 if (arg1 != noreg) {
3629 assert(arg2 != c_rarg1, "clobbered");
3630 __ movptr(c_rarg1, arg1);
3631 }
3632 if (arg2 != noreg) {
3633 __ movptr(c_rarg2, arg2);
3634 }
3635 __ movptr(c_rarg0, r15_thread);
3636 BLOCK_COMMENT("call runtime_entry");
3637 __ call(RuntimeAddress(runtime_entry));
3638
3639 // Generate oop map
3640 OopMap* map = new OopMap(framesize, 0);
3641
3642 oop_maps->add_gc_map(the_pc - start, map);
3643
3644 __ reset_last_Java_frame(true);
3645
3646 __ leave(); // required for proper stackwalking of RuntimeStub frame
3647
3648 // check for pending exceptions
3649 #ifdef ASSERT
3650 Label L;
3651 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), NULL_WORD);
3652 __ jcc(Assembler::notEqual, L);
3653 __ should_not_reach_here();
3654 __ bind(L);
3655 #endif // ASSERT
3656 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3657
3658
3659 // codeBlob framesize is in words (not VMRegImpl::slot_size)
3660 RuntimeStub* stub =
3661 RuntimeStub::new_runtime_stub(name,
3662 &code,
3663 frame_complete,
3664 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3665 oop_maps, false);
3666 return stub->entry_point();
3667 }
3668
3669 void StubGenerator::create_control_words() {
3670 // Round to nearest, 64-bit mode, exceptions masked
3671 StubRoutines::x86::_mxcsr_std = 0x1F80;
3672 }
3673
3674 // Initialization
3675 void StubGenerator::generate_initial() {
3676 // Generates all stubs and initializes the entry points
3677
3678 // This platform-specific settings are needed by generate_call_stub()
3679 create_control_words();
3680
3681 // entry points that exist in all platforms Note: This is code
3682 // that could be shared among different platforms - however the
3683 // benefit seems to be smaller than the disadvantage of having a
3684 // much more complicated generator structure. See also comment in
3685 // stubRoutines.hpp.
3686
3687 StubRoutines::_forward_exception_entry = generate_forward_exception();
3688
3689 // Generate these first because they are called from other stubs
3690 if (InlineTypeReturnedAsFields) {
3691 StubRoutines::_load_inline_type_fields_in_regs =
3692 generate_return_value_stub(CAST_FROM_FN_PTR(address, SharedRuntime::load_inline_type_fields_in_regs),
3693 "load_inline_type_fields_in_regs", false);
3694 StubRoutines::_store_inline_type_fields_to_buf =
3695 generate_return_value_stub(CAST_FROM_FN_PTR(address, SharedRuntime::store_inline_type_fields_to_buf),
3696 "store_inline_type_fields_to_buf", true);
3697 }
3698
3699 StubRoutines::_call_stub_entry =
3700 generate_call_stub(StubRoutines::_call_stub_return_address);
3701
3702 // is referenced by megamorphic call
3703 StubRoutines::_catch_exception_entry = generate_catch_exception();
3704
3705 // atomic calls
3706 StubRoutines::_fence_entry = generate_orderaccess_fence();
3707
3708 // platform dependent
3709 StubRoutines::x86::_get_previous_sp_entry = generate_get_previous_sp();
3710
3711 StubRoutines::x86::_verify_mxcsr_entry = generate_verify_mxcsr();
3712
3713 StubRoutines::x86::_f2i_fixup = generate_f2i_fixup();
3714 StubRoutines::x86::_f2l_fixup = generate_f2l_fixup();
3715 StubRoutines::x86::_d2i_fixup = generate_d2i_fixup();
3716 StubRoutines::x86::_d2l_fixup = generate_d2l_fixup();
3717
3718 StubRoutines::x86::_float_sign_mask = generate_fp_mask("float_sign_mask", 0x7FFFFFFF7FFFFFFF);
3719 StubRoutines::x86::_float_sign_flip = generate_fp_mask("float_sign_flip", 0x8000000080000000);
3720 StubRoutines::x86::_double_sign_mask = generate_fp_mask("double_sign_mask", 0x7FFFFFFFFFFFFFFF);
3721 StubRoutines::x86::_double_sign_flip = generate_fp_mask("double_sign_flip", 0x8000000000000000);
3722
3723 // Build this early so it's available for the interpreter.
3724 StubRoutines::_throw_StackOverflowError_entry =
3725 generate_throw_exception("StackOverflowError throw_exception",
3726 CAST_FROM_FN_PTR(address,
3727 SharedRuntime::
3728 throw_StackOverflowError));
3729 StubRoutines::_throw_delayed_StackOverflowError_entry =
3730 generate_throw_exception("delayed StackOverflowError throw_exception",
3731 CAST_FROM_FN_PTR(address,
3732 SharedRuntime::
3733 throw_delayed_StackOverflowError));
3734 if (UseCRC32Intrinsics) {
3735 // set table address before stub generation which use it
3736 StubRoutines::_crc_table_adr = (address)StubRoutines::x86::_crc_table;
3737 StubRoutines::_updateBytesCRC32 = generate_updateBytesCRC32();
3738 }
3739
3740 if (UsePolyIntrinsics) {
3741 StubRoutines::_poly1305_processBlocks = generate_poly1305_processBlocks();
3742 }
3743
3744 if (UseCRC32CIntrinsics) {
3745 bool supports_clmul = VM_Version::supports_clmul();
3746 StubRoutines::x86::generate_CRC32C_table(supports_clmul);
3747 StubRoutines::_crc32c_table_addr = (address)StubRoutines::x86::_crc32c_table;
3748 StubRoutines::_updateBytesCRC32C = generate_updateBytesCRC32C(supports_clmul);
3749 }
3750
3751 if (UseAdler32Intrinsics) {
3752 StubRoutines::_updateBytesAdler32 = generate_updateBytesAdler32();
3753 }
3754
3755 generate_libm_stubs();
3756 }
3757
3758 // Call here from the interpreter or compiled code to either load
3759 // multiple returned values from the inline type instance being
3760 // returned to registers or to store returned values to a newly
3761 // allocated inline type instance.
3762 // Register is a class, but it would be assigned numerical value.
3763 // "0" is assigned for xmm0. Thus we need to ignore -Wnonnull.
3764 PRAGMA_DIAG_PUSH
3765 PRAGMA_NONNULL_IGNORED
3766 address StubGenerator::generate_return_value_stub(address destination, const char* name, bool has_res) {
3767 // We need to save all registers the calling convention may use so
3768 // the runtime calls read or update those registers. This needs to
3769 // be in sync with SharedRuntime::java_return_convention().
3770 enum layout {
3771 pad_off = frame::arg_reg_save_area_bytes/BytesPerInt, pad_off_2,
3772 rax_off, rax_off_2,
3773 j_rarg5_off, j_rarg5_2,
3774 j_rarg4_off, j_rarg4_2,
3775 j_rarg3_off, j_rarg3_2,
3776 j_rarg2_off, j_rarg2_2,
3777 j_rarg1_off, j_rarg1_2,
3778 j_rarg0_off, j_rarg0_2,
3779 j_farg0_off, j_farg0_2,
3780 j_farg1_off, j_farg1_2,
3781 j_farg2_off, j_farg2_2,
3782 j_farg3_off, j_farg3_2,
3783 j_farg4_off, j_farg4_2,
3784 j_farg5_off, j_farg5_2,
3785 j_farg6_off, j_farg6_2,
3786 j_farg7_off, j_farg7_2,
3787 rbp_off, rbp_off_2,
3788 return_off, return_off_2,
3789
3790 framesize
3791 };
3792
3793 CodeBuffer buffer(name, 1000, 512);
3794 MacroAssembler* _masm = new MacroAssembler(&buffer);
3795
3796 int frame_size_in_bytes = align_up(framesize*BytesPerInt, 16);
3797 assert(frame_size_in_bytes == framesize*BytesPerInt, "misaligned");
3798 int frame_size_in_slots = frame_size_in_bytes / BytesPerInt;
3799 int frame_size_in_words = frame_size_in_bytes / wordSize;
3800
3801 OopMapSet *oop_maps = new OopMapSet();
3802 OopMap* map = new OopMap(frame_size_in_slots, 0);
3803
3804 map->set_callee_saved(VMRegImpl::stack2reg(rax_off), rax->as_VMReg());
3805 map->set_callee_saved(VMRegImpl::stack2reg(j_rarg5_off), j_rarg5->as_VMReg());
3806 map->set_callee_saved(VMRegImpl::stack2reg(j_rarg4_off), j_rarg4->as_VMReg());
3807 map->set_callee_saved(VMRegImpl::stack2reg(j_rarg3_off), j_rarg3->as_VMReg());
3808 map->set_callee_saved(VMRegImpl::stack2reg(j_rarg2_off), j_rarg2->as_VMReg());
3809 map->set_callee_saved(VMRegImpl::stack2reg(j_rarg1_off), j_rarg1->as_VMReg());
3810 map->set_callee_saved(VMRegImpl::stack2reg(j_rarg0_off), j_rarg0->as_VMReg());
3811 map->set_callee_saved(VMRegImpl::stack2reg(j_farg0_off), j_farg0->as_VMReg());
3812 map->set_callee_saved(VMRegImpl::stack2reg(j_farg1_off), j_farg1->as_VMReg());
3813 map->set_callee_saved(VMRegImpl::stack2reg(j_farg2_off), j_farg2->as_VMReg());
3814 map->set_callee_saved(VMRegImpl::stack2reg(j_farg3_off), j_farg3->as_VMReg());
3815 map->set_callee_saved(VMRegImpl::stack2reg(j_farg4_off), j_farg4->as_VMReg());
3816 map->set_callee_saved(VMRegImpl::stack2reg(j_farg5_off), j_farg5->as_VMReg());
3817 map->set_callee_saved(VMRegImpl::stack2reg(j_farg6_off), j_farg6->as_VMReg());
3818 map->set_callee_saved(VMRegImpl::stack2reg(j_farg7_off), j_farg7->as_VMReg());
3819
3820 int start = __ offset();
3821
3822 __ subptr(rsp, frame_size_in_bytes - 8 /* return address*/);
3823
3824 __ movptr(Address(rsp, rbp_off * BytesPerInt), rbp);
3825 __ movdbl(Address(rsp, j_farg7_off * BytesPerInt), j_farg7);
3826 __ movdbl(Address(rsp, j_farg6_off * BytesPerInt), j_farg6);
3827 __ movdbl(Address(rsp, j_farg5_off * BytesPerInt), j_farg5);
3828 __ movdbl(Address(rsp, j_farg4_off * BytesPerInt), j_farg4);
3829 __ movdbl(Address(rsp, j_farg3_off * BytesPerInt), j_farg3);
3830 __ movdbl(Address(rsp, j_farg2_off * BytesPerInt), j_farg2);
3831 __ movdbl(Address(rsp, j_farg1_off * BytesPerInt), j_farg1);
3832 __ movdbl(Address(rsp, j_farg0_off * BytesPerInt), j_farg0);
3833
3834 __ movptr(Address(rsp, j_rarg0_off * BytesPerInt), j_rarg0);
3835 __ movptr(Address(rsp, j_rarg1_off * BytesPerInt), j_rarg1);
3836 __ movptr(Address(rsp, j_rarg2_off * BytesPerInt), j_rarg2);
3837 __ movptr(Address(rsp, j_rarg3_off * BytesPerInt), j_rarg3);
3838 __ movptr(Address(rsp, j_rarg4_off * BytesPerInt), j_rarg4);
3839 __ movptr(Address(rsp, j_rarg5_off * BytesPerInt), j_rarg5);
3840 __ movptr(Address(rsp, rax_off * BytesPerInt), rax);
3841
3842 int frame_complete = __ offset();
3843
3844 __ set_last_Java_frame(noreg, noreg, NULL, rscratch1);
3845
3846 __ mov(c_rarg0, r15_thread);
3847 __ mov(c_rarg1, rax);
3848
3849 __ call(RuntimeAddress(destination));
3850
3851 // Set an oopmap for the call site.
3852
3853 oop_maps->add_gc_map( __ offset() - start, map);
3854
3855 // clear last_Java_sp
3856 __ reset_last_Java_frame(false);
3857
3858 __ movptr(rbp, Address(rsp, rbp_off * BytesPerInt));
3859 __ movdbl(j_farg7, Address(rsp, j_farg7_off * BytesPerInt));
3860 __ movdbl(j_farg6, Address(rsp, j_farg6_off * BytesPerInt));
3861 __ movdbl(j_farg5, Address(rsp, j_farg5_off * BytesPerInt));
3862 __ movdbl(j_farg4, Address(rsp, j_farg4_off * BytesPerInt));
3863 __ movdbl(j_farg3, Address(rsp, j_farg3_off * BytesPerInt));
3864 __ movdbl(j_farg2, Address(rsp, j_farg2_off * BytesPerInt));
3865 __ movdbl(j_farg1, Address(rsp, j_farg1_off * BytesPerInt));
3866 __ movdbl(j_farg0, Address(rsp, j_farg0_off * BytesPerInt));
3867
3868 __ movptr(j_rarg0, Address(rsp, j_rarg0_off * BytesPerInt));
3869 __ movptr(j_rarg1, Address(rsp, j_rarg1_off * BytesPerInt));
3870 __ movptr(j_rarg2, Address(rsp, j_rarg2_off * BytesPerInt));
3871 __ movptr(j_rarg3, Address(rsp, j_rarg3_off * BytesPerInt));
3872 __ movptr(j_rarg4, Address(rsp, j_rarg4_off * BytesPerInt));
3873 __ movptr(j_rarg5, Address(rsp, j_rarg5_off * BytesPerInt));
3874 __ movptr(rax, Address(rsp, rax_off * BytesPerInt));
3875
3876 __ addptr(rsp, frame_size_in_bytes-8);
3877
3878 // check for pending exceptions
3879 Label pending;
3880 __ cmpptr(Address(r15_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
3881 __ jcc(Assembler::notEqual, pending);
3882
3883 if (has_res) {
3884 __ get_vm_result(rax, r15_thread);
3885 }
3886
3887 __ ret(0);
3888
3889 __ bind(pending);
3890
3891 __ movptr(rax, Address(r15_thread, Thread::pending_exception_offset()));
3892 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3893
3894 // -------------
3895 // make sure all code is generated
3896 _masm->flush();
3897
3898 RuntimeStub* stub = RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, false);
3899 return stub->entry_point();
3900 }
3901
3902 void StubGenerator::generate_phase1() {
3903 // Continuation stubs:
3904 StubRoutines::_cont_thaw = generate_cont_thaw();
3905 StubRoutines::_cont_returnBarrier = generate_cont_returnBarrier();
3906 StubRoutines::_cont_returnBarrierExc = generate_cont_returnBarrier_exception();
3907
3908 JFR_ONLY(StubRoutines::_jfr_write_checkpoint_stub = generate_jfr_write_checkpoint();)
3909 JFR_ONLY(StubRoutines::_jfr_write_checkpoint = StubRoutines::_jfr_write_checkpoint_stub->entry_point();)
3910 }
3911
3912 void StubGenerator::generate_all() {
3913 // Generates all stubs and initializes the entry points
3914
3915 // These entry points require SharedInfo::stack0 to be set up in
3916 // non-core builds and need to be relocatable, so they each
3917 // fabricate a RuntimeStub internally.
3918 StubRoutines::_throw_AbstractMethodError_entry =
3919 generate_throw_exception("AbstractMethodError throw_exception",
3920 CAST_FROM_FN_PTR(address,
3921 SharedRuntime::
3922 throw_AbstractMethodError));
3923
3924 StubRoutines::_throw_IncompatibleClassChangeError_entry =
3925 generate_throw_exception("IncompatibleClassChangeError throw_exception",
3926 CAST_FROM_FN_PTR(address,
3927 SharedRuntime::
3928 throw_IncompatibleClassChangeError));
3929
3930 StubRoutines::_throw_NullPointerException_at_call_entry =
3931 generate_throw_exception("NullPointerException at call throw_exception",
3932 CAST_FROM_FN_PTR(address,
3933 SharedRuntime::
3934 throw_NullPointerException_at_call));
3935
3936 // entry points that are platform specific
3937 StubRoutines::x86::_vector_float_sign_mask = generate_vector_mask("vector_float_sign_mask", 0x7FFFFFFF7FFFFFFF);
3938 StubRoutines::x86::_vector_float_sign_flip = generate_vector_mask("vector_float_sign_flip", 0x8000000080000000);
3939 StubRoutines::x86::_vector_double_sign_mask = generate_vector_mask("vector_double_sign_mask", 0x7FFFFFFFFFFFFFFF);
3940 StubRoutines::x86::_vector_double_sign_flip = generate_vector_mask("vector_double_sign_flip", 0x8000000000000000);
3941 StubRoutines::x86::_vector_all_bits_set = generate_vector_mask("vector_all_bits_set", 0xFFFFFFFFFFFFFFFF);
3942 StubRoutines::x86::_vector_int_mask_cmp_bits = generate_vector_mask("vector_int_mask_cmp_bits", 0x0000000100000001);
3943 StubRoutines::x86::_vector_short_to_byte_mask = generate_vector_mask("vector_short_to_byte_mask", 0x00ff00ff00ff00ff);
3944 StubRoutines::x86::_vector_byte_perm_mask = generate_vector_byte_perm_mask("vector_byte_perm_mask");
3945 StubRoutines::x86::_vector_int_to_byte_mask = generate_vector_mask("vector_int_to_byte_mask", 0x000000ff000000ff);
3946 StubRoutines::x86::_vector_int_to_short_mask = generate_vector_mask("vector_int_to_short_mask", 0x0000ffff0000ffff);
3947 StubRoutines::x86::_vector_32_bit_mask = generate_vector_custom_i32("vector_32_bit_mask", Assembler::AVX_512bit,
3948 0xFFFFFFFF, 0, 0, 0);
3949 StubRoutines::x86::_vector_64_bit_mask = generate_vector_custom_i32("vector_64_bit_mask", Assembler::AVX_512bit,
3950 0xFFFFFFFF, 0xFFFFFFFF, 0, 0);
3951 StubRoutines::x86::_vector_int_shuffle_mask = generate_vector_mask("vector_int_shuffle_mask", 0x0302010003020100);
3952 StubRoutines::x86::_vector_byte_shuffle_mask = generate_vector_byte_shuffle_mask("vector_byte_shuffle_mask");
3953 StubRoutines::x86::_vector_short_shuffle_mask = generate_vector_mask("vector_short_shuffle_mask", 0x0100010001000100);
3954 StubRoutines::x86::_vector_long_shuffle_mask = generate_vector_mask("vector_long_shuffle_mask", 0x0000000100000000);
3955 StubRoutines::x86::_vector_long_sign_mask = generate_vector_mask("vector_long_sign_mask", 0x8000000000000000);
3956 StubRoutines::x86::_vector_iota_indices = generate_iota_indices("iota_indices");
3957 StubRoutines::x86::_vector_count_leading_zeros_lut = generate_count_leading_zeros_lut("count_leading_zeros_lut");
3958 StubRoutines::x86::_vector_reverse_bit_lut = generate_vector_reverse_bit_lut("reverse_bit_lut");
3959 StubRoutines::x86::_vector_reverse_byte_perm_mask_long = generate_vector_reverse_byte_perm_mask_long("perm_mask_long");
3960 StubRoutines::x86::_vector_reverse_byte_perm_mask_int = generate_vector_reverse_byte_perm_mask_int("perm_mask_int");
3961 StubRoutines::x86::_vector_reverse_byte_perm_mask_short = generate_vector_reverse_byte_perm_mask_short("perm_mask_short");
3962
3963 if (VM_Version::supports_avx2() && !VM_Version::supports_avx512_vpopcntdq()) {
3964 // lut implementation influenced by counting 1s algorithm from section 5-1 of Hackers' Delight.
3965 StubRoutines::x86::_vector_popcount_lut = generate_popcount_avx_lut("popcount_lut");
3966 }
3967
3968 // support for verify_oop (must happen after universe_init)
3969 if (VerifyOops) {
3970 StubRoutines::_verify_oop_subroutine_entry = generate_verify_oop();
3971 }
3972
3973 // data cache line writeback
3974 StubRoutines::_data_cache_writeback = generate_data_cache_writeback();
3975 StubRoutines::_data_cache_writeback_sync = generate_data_cache_writeback_sync();
3976
3977 // arraycopy stubs used by compilers
3978 generate_arraycopy_stubs();
3979
3980 generate_aes_stubs();
3981
3982 generate_ghash_stubs();
3983
3984 generate_chacha_stubs();
3985
3986 if (UseMD5Intrinsics) {
3987 StubRoutines::_md5_implCompress = generate_md5_implCompress(false, "md5_implCompress");
3988 StubRoutines::_md5_implCompressMB = generate_md5_implCompress(true, "md5_implCompressMB");
3989 }
3990 if (UseSHA1Intrinsics) {
3991 StubRoutines::x86::_upper_word_mask_addr = generate_upper_word_mask();
3992 StubRoutines::x86::_shuffle_byte_flip_mask_addr = generate_shuffle_byte_flip_mask();
3993 StubRoutines::_sha1_implCompress = generate_sha1_implCompress(false, "sha1_implCompress");
3994 StubRoutines::_sha1_implCompressMB = generate_sha1_implCompress(true, "sha1_implCompressMB");
3995 }
3996 if (UseSHA256Intrinsics) {
3997 StubRoutines::x86::_k256_adr = (address)StubRoutines::x86::_k256;
3998 char* dst = (char*)StubRoutines::x86::_k256_W;
3999 char* src = (char*)StubRoutines::x86::_k256;
4000 for (int ii = 0; ii < 16; ++ii) {
4001 memcpy(dst + 32 * ii, src + 16 * ii, 16);
4002 memcpy(dst + 32 * ii + 16, src + 16 * ii, 16);
4003 }
4004 StubRoutines::x86::_k256_W_adr = (address)StubRoutines::x86::_k256_W;
4005 StubRoutines::x86::_pshuffle_byte_flip_mask_addr = generate_pshuffle_byte_flip_mask();
4006 StubRoutines::_sha256_implCompress = generate_sha256_implCompress(false, "sha256_implCompress");
4007 StubRoutines::_sha256_implCompressMB = generate_sha256_implCompress(true, "sha256_implCompressMB");
4008 }
4009 if (UseSHA512Intrinsics) {
4010 StubRoutines::x86::_k512_W_addr = (address)StubRoutines::x86::_k512_W;
4011 StubRoutines::x86::_pshuffle_byte_flip_mask_addr_sha512 = generate_pshuffle_byte_flip_mask_sha512();
4012 StubRoutines::_sha512_implCompress = generate_sha512_implCompress(false, "sha512_implCompress");
4013 StubRoutines::_sha512_implCompressMB = generate_sha512_implCompress(true, "sha512_implCompressMB");
4014 }
4015
4016 if (UseBASE64Intrinsics) {
4017 if(VM_Version::supports_avx2() &&
4018 VM_Version::supports_avx512bw() &&
4019 VM_Version::supports_avx512vl()) {
4020 StubRoutines::x86::_avx2_shuffle_base64 = base64_avx2_shuffle_addr();
4021 StubRoutines::x86::_avx2_input_mask_base64 = base64_avx2_input_mask_addr();
4022 StubRoutines::x86::_avx2_lut_base64 = base64_avx2_lut_addr();
4023 }
4024 StubRoutines::x86::_encoding_table_base64 = base64_encoding_table_addr();
4025 if (VM_Version::supports_avx512_vbmi()) {
4026 StubRoutines::x86::_shuffle_base64 = base64_shuffle_addr();
4027 StubRoutines::x86::_lookup_lo_base64 = base64_vbmi_lookup_lo_addr();
4028 StubRoutines::x86::_lookup_hi_base64 = base64_vbmi_lookup_hi_addr();
4029 StubRoutines::x86::_lookup_lo_base64url = base64_vbmi_lookup_lo_url_addr();
4030 StubRoutines::x86::_lookup_hi_base64url = base64_vbmi_lookup_hi_url_addr();
4031 StubRoutines::x86::_pack_vec_base64 = base64_vbmi_pack_vec_addr();
4032 StubRoutines::x86::_join_0_1_base64 = base64_vbmi_join_0_1_addr();
4033 StubRoutines::x86::_join_1_2_base64 = base64_vbmi_join_1_2_addr();
4034 StubRoutines::x86::_join_2_3_base64 = base64_vbmi_join_2_3_addr();
4035 }
4036 StubRoutines::x86::_decoding_table_base64 = base64_decoding_table_addr();
4037 StubRoutines::_base64_encodeBlock = generate_base64_encodeBlock();
4038 StubRoutines::_base64_decodeBlock = generate_base64_decodeBlock();
4039 }
4040
4041 BarrierSetNMethod* bs_nm = BarrierSet::barrier_set()->barrier_set_nmethod();
4042 if (bs_nm != NULL) {
4043 StubRoutines::x86::_method_entry_barrier = generate_method_entry_barrier();
4044 }
4045 #ifdef COMPILER2
4046 if (UseMultiplyToLenIntrinsic) {
4047 StubRoutines::_multiplyToLen = generate_multiplyToLen();
4048 }
4049 if (UseSquareToLenIntrinsic) {
4050 StubRoutines::_squareToLen = generate_squareToLen();
4051 }
4052 if (UseMulAddIntrinsic) {
4053 StubRoutines::_mulAdd = generate_mulAdd();
4054 }
4055 if (VM_Version::supports_avx512_vbmi2()) {
4056 StubRoutines::_bigIntegerRightShiftWorker = generate_bigIntegerRightShift();
4057 StubRoutines::_bigIntegerLeftShiftWorker = generate_bigIntegerLeftShift();
4058 }
4059 if (UseMontgomeryMultiplyIntrinsic) {
4060 StubRoutines::_montgomeryMultiply
4061 = CAST_FROM_FN_PTR(address, SharedRuntime::montgomery_multiply);
4062 }
4063 if (UseMontgomerySquareIntrinsic) {
4064 StubRoutines::_montgomerySquare
4065 = CAST_FROM_FN_PTR(address, SharedRuntime::montgomery_square);
4066 }
4067
4068 // Get svml stub routine addresses
4069 void *libjsvml = NULL;
4070 char ebuf[1024];
4071 char dll_name[JVM_MAXPATHLEN];
4072 if (os::dll_locate_lib(dll_name, sizeof(dll_name), Arguments::get_dll_dir(), "jsvml")) {
4073 libjsvml = os::dll_load(dll_name, ebuf, sizeof ebuf);
4074 }
4075 if (libjsvml != NULL) {
4076 // SVML method naming convention
4077 // All the methods are named as __jsvml_op<T><N>_ha_<VV>
4078 // Where:
4079 // ha stands for high accuracy
4080 // <T> is optional to indicate float/double
4081 // Set to f for vector float operation
4082 // Omitted for vector double operation
4083 // <N> is the number of elements in the vector
4084 // 1, 2, 4, 8, 16
4085 // e.g. 128 bit float vector has 4 float elements
4086 // <VV> indicates the avx/sse level:
4087 // z0 is AVX512, l9 is AVX2, e9 is AVX1 and ex is for SSE2
4088 // e.g. __jsvml_expf16_ha_z0 is the method for computing 16 element vector float exp using AVX 512 insns
4089 // __jsvml_exp8_ha_z0 is the method for computing 8 element vector double exp using AVX 512 insns
4090
4091 log_info(library)("Loaded library %s, handle " INTPTR_FORMAT, JNI_LIB_PREFIX "jsvml" JNI_LIB_SUFFIX, p2i(libjsvml));
4092 if (UseAVX > 2) {
4093 for (int op = 0; op < VectorSupport::NUM_SVML_OP; op++) {
4094 int vop = VectorSupport::VECTOR_OP_SVML_START + op;
4095 if ((!VM_Version::supports_avx512dq()) &&
4096 (vop == VectorSupport::VECTOR_OP_LOG || vop == VectorSupport::VECTOR_OP_LOG10 || vop == VectorSupport::VECTOR_OP_POW)) {
4097 continue;
4098 }
4099 snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf16_ha_z0", VectorSupport::svmlname[op]);
4100 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_512][op] = (address)os::dll_lookup(libjsvml, ebuf);
4101
4102 snprintf(ebuf, sizeof(ebuf), "__jsvml_%s8_ha_z0", VectorSupport::svmlname[op]);
4103 StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_512][op] = (address)os::dll_lookup(libjsvml, ebuf);
4104 }
4105 }
4106 const char* avx_sse_str = (UseAVX >= 2) ? "l9" : ((UseAVX == 1) ? "e9" : "ex");
4107 for (int op = 0; op < VectorSupport::NUM_SVML_OP; op++) {
4108 int vop = VectorSupport::VECTOR_OP_SVML_START + op;
4109 if (vop == VectorSupport::VECTOR_OP_POW) {
4110 continue;
4111 }
4112 snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf4_ha_%s", VectorSupport::svmlname[op], avx_sse_str);
4113 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_64][op] = (address)os::dll_lookup(libjsvml, ebuf);
4114
4115 snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf4_ha_%s", VectorSupport::svmlname[op], avx_sse_str);
4116 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_128][op] = (address)os::dll_lookup(libjsvml, ebuf);
4117
4118 snprintf(ebuf, sizeof(ebuf), "__jsvml_%sf8_ha_%s", VectorSupport::svmlname[op], avx_sse_str);
4119 StubRoutines::_vector_f_math[VectorSupport::VEC_SIZE_256][op] = (address)os::dll_lookup(libjsvml, ebuf);
4120
4121 snprintf(ebuf, sizeof(ebuf), "__jsvml_%s1_ha_%s", VectorSupport::svmlname[op], avx_sse_str);
4122 StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_64][op] = (address)os::dll_lookup(libjsvml, ebuf);
4123
4124 snprintf(ebuf, sizeof(ebuf), "__jsvml_%s2_ha_%s", VectorSupport::svmlname[op], avx_sse_str);
4125 StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_128][op] = (address)os::dll_lookup(libjsvml, ebuf);
4126
4127 snprintf(ebuf, sizeof(ebuf), "__jsvml_%s4_ha_%s", VectorSupport::svmlname[op], avx_sse_str);
4128 StubRoutines::_vector_d_math[VectorSupport::VEC_SIZE_256][op] = (address)os::dll_lookup(libjsvml, ebuf);
4129 }
4130 }
4131 #endif // COMPILER2
4132
4133 if (UseVectorizedMismatchIntrinsic) {
4134 StubRoutines::_vectorizedMismatch = generate_vectorizedMismatch();
4135 }
4136 }
4137
4138 void StubGenerator_generate(CodeBuffer* code, int phase) {
4139 if (UnsafeCopyMemory::_table == NULL) {
4140 UnsafeCopyMemory::create_table(16);
4141 }
4142 StubGenerator g(code, phase);
4143 }
4144
4145 #undef __