1 /*
2 * Copyright (c) 2003, 2026, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2014, 2020, Red Hat Inc. All rights reserved.
4 * Copyright (c) 2020, 2023, Huawei Technologies Co., Ltd. All rights reserved.
5 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
6 *
7 * This code is free software; you can redistribute it and/or modify it
8 * under the terms of the GNU General Public License version 2 only, as
9 * published by the Free Software Foundation.
10 *
11 * This code is distributed in the hope that it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
14 * version 2 for more details (a copy is included in the LICENSE file that
15 * accompanied this code).
16 *
17 * You should have received a copy of the GNU General Public License version
18 * 2 along with this work; if not, write to the Free Software Foundation,
19 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
20 *
21 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
22 * or visit www.oracle.com if you need additional information or have any
23 * questions.
24 *
25 */
26
27 #include "asm/macroAssembler.hpp"
28 #include "asm/macroAssembler.inline.hpp"
29 #include "code/compiledIC.hpp"
30 #include "code/debugInfoRec.hpp"
31 #include "code/vtableStubs.hpp"
32 #include "compiler/oopMap.hpp"
33 #include "gc/shared/barrierSetAssembler.hpp"
34 #include "interpreter/interp_masm.hpp"
35 #include "interpreter/interpreter.hpp"
36 #include "logging/log.hpp"
37 #include "memory/resourceArea.hpp"
38 #include "nativeInst_riscv.hpp"
39 #include "oops/klass.inline.hpp"
40 #include "oops/method.inline.hpp"
41 #include "prims/methodHandles.hpp"
42 #include "runtime/continuation.hpp"
43 #include "runtime/continuationEntry.inline.hpp"
44 #include "runtime/globals.hpp"
45 #include "runtime/jniHandles.hpp"
46 #include "runtime/safepointMechanism.hpp"
47 #include "runtime/sharedRuntime.hpp"
48 #include "runtime/signature.hpp"
49 #include "runtime/stubRoutines.hpp"
50 #include "runtime/timerTrace.hpp"
51 #include "runtime/vframeArray.hpp"
52 #include "utilities/align.hpp"
53 #include "utilities/formatBuffer.hpp"
54 #include "vmreg_riscv.inline.hpp"
55 #ifdef COMPILER1
56 #include "c1/c1_Runtime1.hpp"
57 #endif
58 #ifdef COMPILER2
59 #include "adfiles/ad_riscv.hpp"
60 #include "opto/runtime.hpp"
61 #endif
62
63 #define __ masm->
64
65 #ifdef PRODUCT
66 #define BLOCK_COMMENT(str) /* nothing */
67 #else
68 #define BLOCK_COMMENT(str) __ block_comment(str)
69 #endif
70
71 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size;
72
73 class RegisterSaver {
74 const bool _save_vectors;
75 public:
76 RegisterSaver(bool save_vectors) : _save_vectors(UseRVV && save_vectors) {}
77 ~RegisterSaver() {}
78 OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words);
79 void restore_live_registers(MacroAssembler* masm);
80
81 // Offsets into the register save area
82 // Used by deoptimization when it is managing result register
83 // values on its own
84 // gregs:28, float_register:32; except: x1(ra) & x2(sp) & gp(x3) & tp(x4)
85 // |---v0---|<---SP
86 // |---v1---|save vectors only in generate_handler_blob
87 // |-- .. --|
88 // |---v31--|-----
89 // |---f0---|
90 // |---f1---|
91 // | .. |
92 // |---f31--|
93 // |---reserved slot for stack alignment---|
94 // |---x5---|
95 // | x6 |
96 // |---.. --|
97 // |---x31--|
98 // |---fp---|
99 // |---ra---|
100 int v0_offset_in_bytes(void) { return 0; }
101 int f0_offset_in_bytes(void) {
102 int f0_offset = 0;
103 #ifdef COMPILER2
104 if (_save_vectors) {
105 f0_offset += Matcher::scalable_vector_reg_size(T_INT) * VectorRegister::number_of_registers *
106 BytesPerInt;
107 }
108 #endif
109 return f0_offset;
110 }
111 int reserved_slot_offset_in_bytes(void) {
112 return f0_offset_in_bytes() +
113 FloatRegister::max_slots_per_register *
114 FloatRegister::number_of_registers *
115 BytesPerInt;
116 }
117
118 int reg_offset_in_bytes(Register r) {
119 assert (r->encoding() > 4, "ra, sp, gp and tp not saved");
120 return reserved_slot_offset_in_bytes() + (r->encoding() - 4 /* x1, x2, x3, x4 */) * wordSize;
121 }
122
123 int freg_offset_in_bytes(FloatRegister f) {
124 return f0_offset_in_bytes() + f->encoding() * wordSize;
125 }
126
127 int ra_offset_in_bytes(void) {
128 return reserved_slot_offset_in_bytes() +
129 (Register::number_of_registers - 3) *
130 Register::max_slots_per_register *
131 BytesPerInt;
132 }
133 };
134
135 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) {
136 int vector_size_in_bytes = 0;
137 int vector_size_in_slots = 0;
138 #ifdef COMPILER2
139 if (_save_vectors) {
140 vector_size_in_bytes += Matcher::scalable_vector_reg_size(T_BYTE);
141 vector_size_in_slots += Matcher::scalable_vector_reg_size(T_INT);
142 }
143 #endif
144
145 int frame_size_in_bytes = align_up(additional_frame_words * wordSize + ra_offset_in_bytes() + wordSize, 16);
146 // OopMap frame size is in compiler stack slots (jint's) not bytes or words
147 int frame_size_in_slots = frame_size_in_bytes / BytesPerInt;
148 // The caller will allocate additional_frame_words
149 int additional_frame_slots = additional_frame_words * wordSize / BytesPerInt;
150 // CodeBlob frame size is in words.
151 int frame_size_in_words = frame_size_in_bytes / wordSize;
152 *total_frame_words = frame_size_in_words;
153
154 // Save Integer, Float and Vector registers.
155 __ enter();
156 __ push_CPU_state(_save_vectors, vector_size_in_bytes);
157
158 // Set an oopmap for the call site. This oopmap will map all
159 // oop-registers and debug-info registers as callee-saved. This
160 // will allow deoptimization at this safepoint to find all possible
161 // debug-info recordings, as well as let GC find all oops.
162
163 OopMapSet *oop_maps = new OopMapSet();
164 OopMap* oop_map = new OopMap(frame_size_in_slots, 0);
165 assert_cond(oop_maps != nullptr && oop_map != nullptr);
166
167 int sp_offset_in_slots = 0;
168 int step_in_slots = 0;
169 if (_save_vectors) {
170 step_in_slots = vector_size_in_slots;
171 for (int i = 0; i < VectorRegister::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
172 VectorRegister r = as_VectorRegister(i);
173 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots), r->as_VMReg());
174 }
175 }
176
177 step_in_slots = FloatRegister::max_slots_per_register;
178 for (int i = 0; i < FloatRegister::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
179 FloatRegister r = as_FloatRegister(i);
180 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots), r->as_VMReg());
181 }
182
183 step_in_slots = Register::max_slots_per_register;
184 // skip the slot reserved for alignment, see MacroAssembler::push_reg;
185 // also skip x5 ~ x6 on the stack because they are caller-saved registers.
186 sp_offset_in_slots += Register::max_slots_per_register * 3;
187 // besides, we ignore x0 ~ x4 because push_CPU_state won't push them on the stack.
188 for (int i = 7; i < Register::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
189 Register r = as_Register(i);
190 if (r != xthread) {
191 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots + additional_frame_slots), r->as_VMReg());
192 }
193 }
194
195 return oop_map;
196 }
197
198 void RegisterSaver::restore_live_registers(MacroAssembler* masm) {
199 #ifdef COMPILER2
200 __ pop_CPU_state(_save_vectors, Matcher::scalable_vector_reg_size(T_BYTE));
201 #else
202 assert(!_save_vectors, "vectors are generated only by C2");
203 __ pop_CPU_state(_save_vectors);
204 #endif // COMPILER2
205 __ leave();
206 }
207
208 // Is vector's size (in bytes) bigger than a size saved by default?
209 // riscv does not ovlerlay the floating-point registers on vector registers like aarch64.
210 bool SharedRuntime::is_wide_vector(int size) {
211 return UseRVV && size > 0;
212 }
213
214 // ---------------------------------------------------------------------------
215 // Read the array of BasicTypes from a signature, and compute where the
216 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
217 // quantities. Values less than VMRegImpl::stack0 are registers, those above
218 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
219 // as framesizes are fixed.
220 // VMRegImpl::stack0 refers to the first slot 0(sp).
221 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher.
222 // Register up to Register::number_of_registers) are the 64-bit
223 // integer registers.
224
225 // Note: the INPUTS in sig_bt are in units of Java argument words,
226 // which are 64-bit. The OUTPUTS are in 32-bit units.
227
228 // The Java calling convention is a "shifted" version of the C ABI.
229 // By skipping the first C ABI register we can call non-static jni
230 // methods with small numbers of arguments without having to shuffle
231 // the arguments at all. Since we control the java ABI we ought to at
232 // least get some advantage out of it.
233
234 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
235 VMRegPair *regs,
236 int total_args_passed) {
237 // Create the mapping between argument positions and
238 // registers.
239 static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = {
240 j_rarg0, j_rarg1, j_rarg2, j_rarg3,
241 j_rarg4, j_rarg5, j_rarg6, j_rarg7
242 };
243 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_j] = {
244 j_farg0, j_farg1, j_farg2, j_farg3,
245 j_farg4, j_farg5, j_farg6, j_farg7
246 };
247
248 uint int_args = 0;
249 uint fp_args = 0;
250 uint stk_args = 0;
251
252 for (int i = 0; i < total_args_passed; i++) {
253 switch (sig_bt[i]) {
254 case T_BOOLEAN: // fall through
255 case T_CHAR: // fall through
256 case T_BYTE: // fall through
257 case T_SHORT: // fall through
258 case T_INT:
259 if (int_args < Argument::n_int_register_parameters_j) {
260 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
261 } else {
262 stk_args = align_up(stk_args, 2);
263 regs[i].set1(VMRegImpl::stack2reg(stk_args));
264 stk_args += 1;
265 }
266 break;
267 case T_VOID:
268 // halves of T_LONG or T_DOUBLE
269 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
270 regs[i].set_bad();
271 break;
272 case T_LONG: // fall through
273 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
274 case T_OBJECT: // fall through
275 case T_ARRAY: // fall through
276 case T_ADDRESS:
277 if (int_args < Argument::n_int_register_parameters_j) {
278 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
279 } else {
280 stk_args = align_up(stk_args, 2);
281 regs[i].set2(VMRegImpl::stack2reg(stk_args));
282 stk_args += 2;
283 }
284 break;
285 case T_FLOAT:
286 if (fp_args < Argument::n_float_register_parameters_j) {
287 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
288 } else {
289 stk_args = align_up(stk_args, 2);
290 regs[i].set1(VMRegImpl::stack2reg(stk_args));
291 stk_args += 1;
292 }
293 break;
294 case T_DOUBLE:
295 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
296 if (fp_args < Argument::n_float_register_parameters_j) {
297 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
298 } else {
299 stk_args = align_up(stk_args, 2);
300 regs[i].set2(VMRegImpl::stack2reg(stk_args));
301 stk_args += 2;
302 }
303 break;
304 default:
305 ShouldNotReachHere();
306 }
307 }
308
309 return stk_args;
310 }
311
312 // Patch the callers callsite with entry to compiled code if it exists.
313 static void patch_callers_callsite(MacroAssembler *masm) {
314 Label L;
315 __ ld(t0, Address(xmethod, in_bytes(Method::code_offset())));
316 __ beqz(t0, L);
317
318 __ enter();
319 __ push_CPU_state();
320
321 // VM needs caller's callsite
322 // VM needs target method
323 // This needs to be a long call since we will relocate this adapter to
324 // the codeBuffer and it may not reach
325
326 #ifndef PRODUCT
327 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
328 #endif
329
330 __ mv(c_rarg0, xmethod);
331 __ mv(c_rarg1, ra);
332 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite));
333
334 __ pop_CPU_state();
335 // restore sp
336 __ leave();
337 __ bind(L);
338 }
339
340 static void gen_c2i_adapter(MacroAssembler *masm,
341 int comp_args_on_stack,
342 const GrowableArray<SigEntry>* sig,
343 const VMRegPair *regs,
344 Label& skip_fixup) {
345 // Before we get into the guts of the C2I adapter, see if we should be here
346 // at all. We've come from compiled code and are attempting to jump to the
347 // interpreter, which means the caller made a static call to get here
348 // (vcalls always get a compiled target if there is one). Check for a
349 // compiled target. If there is one, we need to patch the caller's call.
350 patch_callers_callsite(masm);
351
352 __ bind(skip_fixup);
353
354 int words_pushed = 0;
355
356 // Since all args are passed on the stack, total_args_passed *
357 // Interpreter::stackElementSize is the space we need.
358
359 int total_args_passed = sig->length();
360 int extraspace = total_args_passed * Interpreter::stackElementSize;
361
362 __ mv(x19_sender_sp, sp);
363
364 // stack is aligned, keep it that way
365 extraspace = align_up(extraspace, 2 * wordSize);
366
367 if (extraspace) {
368 __ sub(sp, sp, extraspace);
369 }
370
371 // Now write the args into the outgoing interpreter space
372 for (int i = 0; i < total_args_passed; i++) {
373 BasicType bt = sig->at(i)._bt;
374 if (bt == T_VOID) {
375 assert(i > 0 && (sig->at(i - 1)._bt == T_LONG || sig->at(i - 1)._bt == T_DOUBLE), "missing half");
376 continue;
377 }
378
379 // offset to start parameters
380 int st_off = (total_args_passed - i - 1) * Interpreter::stackElementSize;
381 int next_off = st_off - Interpreter::stackElementSize;
382
383 // Say 4 args:
384 // i st_off
385 // 0 32 T_LONG
386 // 1 24 T_VOID
387 // 2 16 T_OBJECT
388 // 3 8 T_BOOL
389 // - 0 return address
390 //
391 // However to make thing extra confusing. Because we can fit a Java long/double in
392 // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter
393 // leaves one slot empty and only stores to a single slot. In this case the
394 // slot that is occupied is the T_VOID slot. See I said it was confusing.
395
396 VMReg r_1 = regs[i].first();
397 VMReg r_2 = regs[i].second();
398 if (!r_1->is_valid()) {
399 assert(!r_2->is_valid(), "");
400 continue;
401 }
402 if (r_1->is_stack()) {
403 // memory to memory use t0
404 int ld_off = (r_1->reg2stack() * VMRegImpl::stack_slot_size
405 + extraspace
406 + words_pushed * wordSize);
407 if (!r_2->is_valid()) {
408 __ lwu(t0, Address(sp, ld_off));
409 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
410 } else {
411 __ ld(t0, Address(sp, ld_off), /*temp register*/esp);
412
413 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
414 // T_DOUBLE and T_LONG use two slots in the interpreter
415 if (bt == T_LONG || bt == T_DOUBLE) {
416 // ld_off == LSW, ld_off+wordSize == MSW
417 // st_off == MSW, next_off == LSW
418 __ sd(t0, Address(sp, next_off), /*temp register*/esp);
419 #ifdef ASSERT
420 // Overwrite the unused slot with known junk
421 __ mv(t0, 0xdeadffffdeadaaaaul);
422 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
423 #endif /* ASSERT */
424 } else {
425 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
426 }
427 }
428 } else if (r_1->is_Register()) {
429 Register r = r_1->as_Register();
430 if (!r_2->is_valid()) {
431 // must be only an int (or less ) so move only 32bits to slot
432 __ sd(r, Address(sp, st_off));
433 } else {
434 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
435 // T_DOUBLE and T_LONG use two slots in the interpreter
436 if (bt == T_LONG || bt == T_DOUBLE) {
437 // long/double in gpr
438 #ifdef ASSERT
439 // Overwrite the unused slot with known junk
440 __ mv(t0, 0xdeadffffdeadaaabul);
441 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
442 #endif /* ASSERT */
443 __ sd(r, Address(sp, next_off));
444 } else {
445 __ sd(r, Address(sp, st_off));
446 }
447 }
448 } else {
449 assert(r_1->is_FloatRegister(), "");
450 if (!r_2->is_valid()) {
451 // only a float use just part of the slot
452 __ fsw(r_1->as_FloatRegister(), Address(sp, st_off));
453 } else {
454 #ifdef ASSERT
455 // Overwrite the unused slot with known junk
456 __ mv(t0, 0xdeadffffdeadaaacul);
457 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
458 #endif /* ASSERT */
459 __ fsd(r_1->as_FloatRegister(), Address(sp, next_off));
460 }
461 }
462 }
463
464 __ mv(esp, sp); // Interp expects args on caller's expression stack
465
466 __ ld(t1, Address(xmethod, in_bytes(Method::interpreter_entry_offset())));
467 __ jr(t1);
468 }
469
470 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
471 int comp_args_on_stack,
472 const GrowableArray<SigEntry>* sig,
473 const VMRegPair *regs) {
474 // Note: x19_sender_sp contains the senderSP on entry. We must
475 // preserve it since we may do a i2c -> c2i transition if we lose a
476 // race where compiled code goes non-entrant while we get args
477 // ready.
478
479 // Cut-out for having no stack args.
480 int comp_words_on_stack = align_up(comp_args_on_stack * VMRegImpl::stack_slot_size, wordSize) >> LogBytesPerWord;
481 if (comp_args_on_stack != 0) {
482 __ sub(t0, sp, comp_words_on_stack * wordSize);
483 __ andi(sp, t0, -16);
484 }
485
486 // Will jump to the compiled code just as if compiled code was doing it.
487 // Pre-load the register-jump target early, to schedule it better.
488 __ ld(t1, Address(xmethod, in_bytes(Method::from_compiled_offset())));
489
490 // Now generate the shuffle code.
491 int total_args_passed = sig->length();
492 for (int i = 0; i < total_args_passed; i++) {
493 BasicType bt = sig->at(i)._bt;
494 if (bt == T_VOID) {
495 assert(i > 0 && (sig->at(i - 1)._bt == T_LONG || sig->at(i - 1)._bt == T_DOUBLE), "missing half");
496 continue;
497 }
498
499 // Pick up 0, 1 or 2 words from SP+offset.
500
501 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
502 "scrambled load targets?");
503 // Load in argument order going down.
504 int ld_off = (total_args_passed - i - 1) * Interpreter::stackElementSize;
505 // Point to interpreter value (vs. tag)
506 int next_off = ld_off - Interpreter::stackElementSize;
507
508 VMReg r_1 = regs[i].first();
509 VMReg r_2 = regs[i].second();
510 if (!r_1->is_valid()) {
511 assert(!r_2->is_valid(), "");
512 continue;
513 }
514 if (r_1->is_stack()) {
515 // Convert stack slot to an SP offset (+ wordSize to account for return address )
516 int st_off = regs[i].first()->reg2stack() * VMRegImpl::stack_slot_size;
517 if (!r_2->is_valid()) {
518 __ lw(t0, Address(esp, ld_off));
519 __ sd(t0, Address(sp, st_off), /*temp register*/t2);
520 } else {
521 //
522 // We are using two optoregs. This can be either T_OBJECT,
523 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
524 // two slots but only uses one for thr T_LONG or T_DOUBLE case
525 // So we must adjust where to pick up the data to match the
526 // interpreter.
527 //
528 // Interpreter local[n] == MSW, local[n+1] == LSW however locals
529 // are accessed as negative so LSW is at LOW address
530
531 // ld_off is MSW so get LSW
532 const int offset = (bt == T_LONG || bt == T_DOUBLE) ?
533 next_off : ld_off;
534 __ ld(t0, Address(esp, offset));
535 // st_off is LSW (i.e. reg.first())
536 __ sd(t0, Address(sp, st_off), /*temp register*/t2);
537 }
538 } else if (r_1->is_Register()) { // Register argument
539 Register r = r_1->as_Register();
540 if (r_2->is_valid()) {
541 //
542 // We are using two VMRegs. This can be either T_OBJECT,
543 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
544 // two slots but only uses one for thr T_LONG or T_DOUBLE case
545 // So we must adjust where to pick up the data to match the
546 // interpreter.
547
548 const int offset = (bt == T_LONG || bt == T_DOUBLE) ?
549 next_off : ld_off;
550
551 // this can be a misaligned move
552 __ ld(r, Address(esp, offset));
553 } else {
554 // sign extend and use a full word?
555 __ lw(r, Address(esp, ld_off));
556 }
557 } else {
558 if (!r_2->is_valid()) {
559 __ flw(r_1->as_FloatRegister(), Address(esp, ld_off));
560 } else {
561 __ fld(r_1->as_FloatRegister(), Address(esp, next_off));
562 }
563 }
564 }
565
566 __ push_cont_fastpath(xthread); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about
567
568 // 6243940 We might end up in handle_wrong_method if
569 // the callee is deoptimized as we race thru here. If that
570 // happens we don't want to take a safepoint because the
571 // caller frame will look interpreted and arguments are now
572 // "compiled" so it is much better to make this transition
573 // invisible to the stack walking code. Unfortunately if
574 // we try and find the callee by normal means a safepoint
575 // is possible. So we stash the desired callee in the thread
576 // and the vm will find there should this case occur.
577
578 __ sd(xmethod, Address(xthread, JavaThread::callee_target_offset()));
579
580 __ jr(t1);
581 }
582
583 // ---------------------------------------------------------------
584
585 void SharedRuntime::generate_i2c2i_adapters(MacroAssembler* masm,
586 int comp_args_on_stack,
587 const GrowableArray<SigEntry>* sig,
588 const VMRegPair* regs,
589 const GrowableArray<SigEntry>* sig_cc,
590 const VMRegPair* regs_cc,
591 const GrowableArray<SigEntry>* sig_cc_ro,
592 const VMRegPair* regs_cc_ro,
593 address entry_address[AdapterBlob::ENTRY_COUNT],
594 AdapterBlob*& new_adapter,
595 bool allocate_code_blob) {
596 entry_address[AdapterBlob::I2C] = __ pc();
597 gen_i2c_adapter(masm, comp_args_on_stack, sig, regs);
598
599 entry_address[AdapterBlob::C2I_Unverified] = __ pc();
600 Label skip_fixup;
601
602 const Register receiver = j_rarg0;
603 const Register data = t0;
604
605 // -------------------------------------------------------------------------
606 // Generate a C2I adapter. On entry we know xmethod holds the Method* during calls
607 // to the interpreter. The args start out packed in the compiled layout. They
608 // need to be unpacked into the interpreter layout. This will almost always
609 // require some stack space. We grow the current (compiled) stack, then repack
610 // the args. We finally end in a jump to the generic interpreter entry point.
611 // On exit from the interpreter, the interpreter will restore our SP (lest the
612 // compiled code, which relies solely on SP and not FP, get sick).
613
614 {
615 __ block_comment("c2i_unverified_entry {");
616
617 __ ic_check();
618 __ ld(xmethod, Address(data, CompiledICData::speculated_method_offset()));
619
620 __ ld(t0, Address(xmethod, in_bytes(Method::code_offset())));
621 __ beqz(t0, skip_fixup);
622 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
623 __ block_comment("} c2i_unverified_entry");
624 }
625
626 entry_address[AdapterBlob::C2I] = __ pc();
627
628 // Class initialization barrier for static methods
629 entry_address[AdapterBlob::C2I_No_Clinit_Check] = nullptr;
630 assert(VM_Version::supports_fast_class_init_checks(), "sanity");
631 Label L_skip_barrier;
632
633 // Bypass the barrier for non-static methods
634 __ load_unsigned_short(t0, Address(xmethod, Method::access_flags_offset()));
635 __ test_bit(t1, t0, exact_log2(JVM_ACC_STATIC));
636 __ beqz(t1, L_skip_barrier); // non-static
637
638 __ load_method_holder(t1, xmethod);
639 __ clinit_barrier(t1, t0, &L_skip_barrier);
640 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
641
642 __ bind(L_skip_barrier);
643 entry_address[AdapterBlob::C2I_No_Clinit_Check] = __ pc();
644
645 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
646 bs->c2i_entry_barrier(masm);
647
648 gen_c2i_adapter(masm, comp_args_on_stack, sig, regs, skip_fixup);
649 return;
650 }
651
652 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
653 uint num_bits,
654 uint total_args_passed) {
655 assert(total_args_passed <= Argument::n_vector_register_parameters_c, "unsupported");
656 assert(num_bits >= 64 && num_bits <= 2048 && is_power_of_2(num_bits), "unsupported");
657
658 // check more info at https://github.com/riscv-non-isa/riscv-elf-psabi-doc/blob/master/riscv-cc.adoc
659 static const VectorRegister VEC_ArgReg[Argument::n_vector_register_parameters_c] = {
660 v8, v9, v10, v11, v12, v13, v14, v15,
661 v16, v17, v18, v19, v20, v21, v22, v23
662 };
663
664 const int next_reg_val = 3;
665 for (uint i = 0; i < total_args_passed; i++) {
666 VMReg vmreg = VEC_ArgReg[i]->as_VMReg();
667 regs[i].set_pair(vmreg->next(next_reg_val), vmreg);
668 }
669 return 0;
670 }
671
672 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
673 VMRegPair *regs,
674 int total_args_passed) {
675
676 // We return the amount of VMRegImpl stack slots we need to reserve for all
677 // the arguments NOT counting out_preserve_stack_slots.
678
679 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
680 c_rarg0, c_rarg1, c_rarg2, c_rarg3,
681 c_rarg4, c_rarg5, c_rarg6, c_rarg7
682 };
683 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
684 c_farg0, c_farg1, c_farg2, c_farg3,
685 c_farg4, c_farg5, c_farg6, c_farg7
686 };
687
688 uint int_args = 0;
689 uint fp_args = 0;
690 uint stk_args = 0; // inc by 2 each time
691
692 for (int i = 0; i < total_args_passed; i++) {
693 switch (sig_bt[i]) {
694 case T_BOOLEAN: // fall through
695 case T_CHAR: // fall through
696 case T_BYTE: // fall through
697 case T_SHORT: // fall through
698 case T_INT:
699 if (int_args < Argument::n_int_register_parameters_c) {
700 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
701 } else {
702 regs[i].set1(VMRegImpl::stack2reg(stk_args));
703 stk_args += 2;
704 }
705 break;
706 case T_LONG: // fall through
707 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
708 case T_OBJECT: // fall through
709 case T_ARRAY: // fall through
710 case T_ADDRESS: // fall through
711 case T_METADATA:
712 if (int_args < Argument::n_int_register_parameters_c) {
713 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
714 } else {
715 regs[i].set2(VMRegImpl::stack2reg(stk_args));
716 stk_args += 2;
717 }
718 break;
719 case T_FLOAT:
720 if (fp_args < Argument::n_float_register_parameters_c) {
721 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
722 } else if (int_args < Argument::n_int_register_parameters_c) {
723 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
724 } else {
725 regs[i].set1(VMRegImpl::stack2reg(stk_args));
726 stk_args += 2;
727 }
728 break;
729 case T_DOUBLE:
730 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
731 if (fp_args < Argument::n_float_register_parameters_c) {
732 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
733 } else if (int_args < Argument::n_int_register_parameters_c) {
734 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
735 } else {
736 regs[i].set2(VMRegImpl::stack2reg(stk_args));
737 stk_args += 2;
738 }
739 break;
740 case T_VOID: // Halves of longs and doubles
741 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
742 regs[i].set_bad();
743 break;
744 default:
745 ShouldNotReachHere();
746 }
747 }
748
749 return stk_args;
750 }
751
752 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
753 // We always ignore the frame_slots arg and just use the space just below frame pointer
754 // which by this time is free to use
755 switch (ret_type) {
756 case T_FLOAT:
757 __ fsw(f10, Address(fp, -3 * wordSize));
758 break;
759 case T_DOUBLE:
760 __ fsd(f10, Address(fp, -3 * wordSize));
761 break;
762 case T_VOID: break;
763 default: {
764 __ sd(x10, Address(fp, -3 * wordSize));
765 }
766 }
767 }
768
769 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
770 // We always ignore the frame_slots arg and just use the space just below frame pointer
771 // which by this time is free to use
772 switch (ret_type) {
773 case T_FLOAT:
774 __ flw(f10, Address(fp, -3 * wordSize));
775 break;
776 case T_DOUBLE:
777 __ fld(f10, Address(fp, -3 * wordSize));
778 break;
779 case T_VOID: break;
780 default: {
781 __ ld(x10, Address(fp, -3 * wordSize));
782 }
783 }
784 }
785
786 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
787 RegSet x;
788 for ( int i = first_arg ; i < arg_count ; i++ ) {
789 if (args[i].first()->is_Register()) {
790 x = x + args[i].first()->as_Register();
791 } else if (args[i].first()->is_FloatRegister()) {
792 __ subi(sp, sp, 2 * wordSize);
793 __ fsd(args[i].first()->as_FloatRegister(), Address(sp, 0));
794 }
795 }
796 __ push_reg(x, sp);
797 }
798
799 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
800 RegSet x;
801 for ( int i = first_arg ; i < arg_count ; i++ ) {
802 if (args[i].first()->is_Register()) {
803 x = x + args[i].first()->as_Register();
804 } else {
805 ;
806 }
807 }
808 __ pop_reg(x, sp);
809 for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) {
810 if (args[i].first()->is_Register()) {
811 ;
812 } else if (args[i].first()->is_FloatRegister()) {
813 __ fld(args[i].first()->as_FloatRegister(), Address(sp, 0));
814 __ addi(sp, sp, 2 * wordSize);
815 }
816 }
817 }
818
819 static void verify_oop_args(MacroAssembler* masm,
820 const methodHandle& method,
821 const BasicType* sig_bt,
822 const VMRegPair* regs) {
823 const Register temp_reg = x9; // not part of any compiled calling seq
824 if (VerifyOops) {
825 for (int i = 0; i < method->size_of_parameters(); i++) {
826 if (sig_bt[i] == T_OBJECT ||
827 sig_bt[i] == T_ARRAY) {
828 VMReg r = regs[i].first();
829 assert(r->is_valid(), "bad oop arg");
830 if (r->is_stack()) {
831 __ ld(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
832 __ verify_oop(temp_reg);
833 } else {
834 __ verify_oop(r->as_Register());
835 }
836 }
837 }
838 }
839 }
840
841 // on exit, sp points to the ContinuationEntry
842 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& stack_slots) {
843 assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
844 assert(in_bytes(ContinuationEntry::cont_offset()) % VMRegImpl::stack_slot_size == 0, "");
845 assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
846
847 stack_slots += (int)ContinuationEntry::size() / wordSize;
848 __ sub(sp, sp, (int)ContinuationEntry::size()); // place Continuation metadata
849
850 OopMap* map = new OopMap(((int)ContinuationEntry::size() + wordSize) / VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
851
852 __ ld(t0, Address(xthread, JavaThread::cont_entry_offset()));
853 __ sd(t0, Address(sp, ContinuationEntry::parent_offset()));
854 __ sd(sp, Address(xthread, JavaThread::cont_entry_offset()));
855
856 return map;
857 }
858
859 // on entry c_rarg1 points to the continuation
860 // sp points to ContinuationEntry
861 // c_rarg3 -- isVirtualThread
862 static void fill_continuation_entry(MacroAssembler* masm) {
863 #ifdef ASSERT
864 __ mv(t0, ContinuationEntry::cookie_value());
865 __ sw(t0, Address(sp, ContinuationEntry::cookie_offset()));
866 #endif
867
868 __ sd(c_rarg1, Address(sp, ContinuationEntry::cont_offset()));
869 __ sw(c_rarg3, Address(sp, ContinuationEntry::flags_offset()));
870 __ sd(zr, Address(sp, ContinuationEntry::chunk_offset()));
871 __ sw(zr, Address(sp, ContinuationEntry::argsize_offset()));
872 __ sw(zr, Address(sp, ContinuationEntry::pin_count_offset()));
873
874 __ ld(t0, Address(xthread, JavaThread::cont_fastpath_offset()));
875 __ sd(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
876
877 __ sd(zr, Address(xthread, JavaThread::cont_fastpath_offset()));
878 }
879
880 // on entry, sp points to the ContinuationEntry
881 // on exit, fp points to the spilled fp + 2 * wordSize in the entry frame
882 static void continuation_enter_cleanup(MacroAssembler* masm) {
883 #ifndef PRODUCT
884 Label OK;
885 __ ld(t0, Address(xthread, JavaThread::cont_entry_offset()));
886 __ beq(sp, t0, OK);
887 __ stop("incorrect sp");
888 __ bind(OK);
889 #endif
890
891 __ ld(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
892 __ sd(t0, Address(xthread, JavaThread::cont_fastpath_offset()));
893 __ ld(t0, Address(sp, ContinuationEntry::parent_offset()));
894 __ sd(t0, Address(xthread, JavaThread::cont_entry_offset()));
895 __ add(fp, sp, (int)ContinuationEntry::size() + 2 * wordSize /* 2 extra words to match up with leave() */);
896 }
897
898 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
899 // On entry: c_rarg1 -- the continuation object
900 // c_rarg2 -- isContinue
901 // c_rarg3 -- isVirtualThread
902 static void gen_continuation_enter(MacroAssembler* masm,
903 const methodHandle& method,
904 const BasicType* sig_bt,
905 const VMRegPair* regs,
906 int& exception_offset,
907 OopMapSet*oop_maps,
908 int& frame_complete,
909 int& stack_slots,
910 int& interpreted_entry_offset,
911 int& compiled_entry_offset) {
912 // verify_oop_args(masm, method, sig_bt, regs);
913 Address resolve(SharedRuntime::get_resolve_static_call_stub(), relocInfo::static_call_type);
914
915 address start = __ pc();
916
917 Label call_thaw, exit;
918
919 // i2i entry used at interp_only_mode only
920 interpreted_entry_offset = __ pc() - start;
921 {
922 #ifdef ASSERT
923 Label is_interp_only;
924 __ lw(t0, Address(xthread, JavaThread::interp_only_mode_offset()));
925 __ bnez(t0, is_interp_only);
926 __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
927 __ bind(is_interp_only);
928 #endif
929
930 // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
931 __ ld(c_rarg1, Address(esp, Interpreter::stackElementSize * 2));
932 __ ld(c_rarg2, Address(esp, Interpreter::stackElementSize * 1));
933 __ ld(c_rarg3, Address(esp, Interpreter::stackElementSize * 0));
934 __ push_cont_fastpath(xthread);
935
936 __ enter();
937 stack_slots = 2; // will be adjusted in setup
938 OopMap* map = continuation_enter_setup(masm, stack_slots);
939 // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
940 // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
941
942 fill_continuation_entry(masm);
943
944 __ bnez(c_rarg2, call_thaw);
945
946 address call_pc;
947 {
948 Assembler::IncompressibleScope scope(masm);
949 // Make sure the call is patchable
950 __ align(NativeInstruction::instruction_size);
951
952 call_pc = __ reloc_call(resolve);
953 if (call_pc == nullptr) {
954 fatal("CodeCache is full at gen_continuation_enter");
955 }
956
957 oop_maps->add_gc_map(__ pc() - start, map);
958 __ post_call_nop();
959 }
960 __ j(exit);
961
962 address stub = CompiledDirectCall::emit_to_interp_stub(masm, call_pc);
963 if (stub == nullptr) {
964 fatal("CodeCache is full at gen_continuation_enter");
965 }
966 }
967
968 // compiled entry
969 __ align(CodeEntryAlignment);
970 compiled_entry_offset = __ pc() - start;
971
972 __ enter();
973 stack_slots = 2; // will be adjusted in setup
974 OopMap* map = continuation_enter_setup(masm, stack_slots);
975 frame_complete = __ pc() - start;
976
977 fill_continuation_entry(masm);
978
979 __ bnez(c_rarg2, call_thaw);
980
981 address call_pc;
982 {
983 Assembler::IncompressibleScope scope(masm);
984 // Make sure the call is patchable
985 __ align(NativeInstruction::instruction_size);
986
987 call_pc = __ reloc_call(resolve);
988 if (call_pc == nullptr) {
989 fatal("CodeCache is full at gen_continuation_enter");
990 }
991
992 oop_maps->add_gc_map(__ pc() - start, map);
993 __ post_call_nop();
994 }
995
996 __ j(exit);
997
998 __ bind(call_thaw);
999
1000 // Post call nops must be natural aligned due to cmodx rules.
1001 {
1002 Assembler::IncompressibleScope scope(masm);
1003 __ align(NativeInstruction::instruction_size);
1004
1005 ContinuationEntry::_thaw_call_pc_offset = __ pc() - start;
1006 __ rt_call(CAST_FROM_FN_PTR(address, StubRoutines::cont_thaw()));
1007 oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1008 ContinuationEntry::_return_pc_offset = __ pc() - start;
1009 __ post_call_nop();
1010 }
1011
1012 __ bind(exit);
1013 ContinuationEntry::_cleanup_offset = __ pc() - start;
1014 continuation_enter_cleanup(masm);
1015 __ leave();
1016 __ ret();
1017
1018 // exception handling
1019 exception_offset = __ pc() - start;
1020 {
1021 __ mv(x9, x10); // save return value contaning the exception oop in callee-saved x9
1022
1023 continuation_enter_cleanup(masm);
1024
1025 __ ld(c_rarg1, Address(fp, -1 * wordSize)); // return address
1026 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), xthread, c_rarg1);
1027
1028 // see OptoRuntime::generate_exception_blob: x10 -- exception oop, x13 -- exception pc
1029
1030 __ mv(x11, x10); // the exception handler
1031 __ mv(x10, x9); // restore return value contaning the exception oop
1032 __ verify_oop(x10);
1033
1034 __ leave();
1035 __ mv(x13, ra);
1036 __ jr(x11); // the exception handler
1037 }
1038
1039 address stub = CompiledDirectCall::emit_to_interp_stub(masm, call_pc);
1040 if (stub == nullptr) {
1041 fatal("CodeCache is full at gen_continuation_enter");
1042 }
1043 }
1044
1045 static void gen_continuation_yield(MacroAssembler* masm,
1046 const methodHandle& method,
1047 const BasicType* sig_bt,
1048 const VMRegPair* regs,
1049 OopMapSet* oop_maps,
1050 int& frame_complete,
1051 int& stack_slots,
1052 int& compiled_entry_offset) {
1053 enum layout {
1054 fp_off,
1055 fp_off2,
1056 return_off,
1057 return_off2,
1058 framesize // inclusive of return address
1059 };
1060 // assert(is_even(framesize/2), "sp not 16-byte aligned");
1061
1062 stack_slots = framesize / VMRegImpl::slots_per_word;
1063 assert(stack_slots == 2, "recheck layout");
1064
1065 address start = __ pc();
1066
1067 compiled_entry_offset = __ pc() - start;
1068 __ enter();
1069
1070 __ mv(c_rarg1, sp);
1071
1072 // Post call nops must be natural aligned due to cmodx rules.
1073 __ align(NativeInstruction::instruction_size);
1074
1075 frame_complete = __ pc() - start;
1076 address the_pc = __ pc();
1077
1078 {
1079 Assembler::IncompressibleScope scope(masm);
1080 __ post_call_nop(); // this must be exactly after the pc value that is pushed into the frame info, we use this nop for fast CodeBlob lookup
1081 }
1082
1083 __ mv(c_rarg0, xthread);
1084 __ set_last_Java_frame(sp, fp, the_pc, t0);
1085 __ call_VM_leaf(Continuation::freeze_entry(), 2);
1086 __ reset_last_Java_frame(true);
1087
1088 Label pinned;
1089
1090 __ bnez(x10, pinned);
1091
1092 // We've succeeded, set sp to the ContinuationEntry
1093 __ ld(sp, Address(xthread, JavaThread::cont_entry_offset()));
1094 continuation_enter_cleanup(masm);
1095
1096 __ bind(pinned); // pinned -- return to caller
1097
1098 // handle pending exception thrown by freeze
1099 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1100 Label ok;
1101 __ beqz(t0, ok);
1102 __ leave();
1103 __ j(RuntimeAddress(StubRoutines::forward_exception_entry()));
1104 __ bind(ok);
1105
1106 __ leave();
1107 __ ret();
1108
1109 OopMap* map = new OopMap(framesize, 1);
1110 oop_maps->add_gc_map(the_pc - start, map);
1111 }
1112
1113 void SharedRuntime::continuation_enter_cleanup(MacroAssembler* masm) {
1114 ::continuation_enter_cleanup(masm);
1115 }
1116
1117 static void gen_special_dispatch(MacroAssembler* masm,
1118 const methodHandle& method,
1119 const BasicType* sig_bt,
1120 const VMRegPair* regs) {
1121 verify_oop_args(masm, method, sig_bt, regs);
1122 vmIntrinsics::ID iid = method->intrinsic_id();
1123
1124 // Now write the args into the outgoing interpreter space
1125 bool has_receiver = false;
1126 Register receiver_reg = noreg;
1127 int member_arg_pos = -1;
1128 Register member_reg = noreg;
1129 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1130 if (ref_kind != 0) {
1131 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1132 member_reg = x9; // known to be free at this point
1133 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1134 } else if (iid == vmIntrinsics::_invokeBasic) {
1135 has_receiver = true;
1136 } else if (iid == vmIntrinsics::_linkToNative) {
1137 member_arg_pos = method->size_of_parameters() - 1; // trailing NativeEntryPoint argument
1138 member_reg = x9; // known to be free at this point
1139 } else {
1140 fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1141 }
1142
1143 if (member_reg != noreg) {
1144 // Load the member_arg into register, if necessary.
1145 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1146 VMReg r = regs[member_arg_pos].first();
1147 if (r->is_stack()) {
1148 __ ld(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1149 } else {
1150 // no data motion is needed
1151 member_reg = r->as_Register();
1152 }
1153 }
1154
1155 if (has_receiver) {
1156 // Make sure the receiver is loaded into a register.
1157 assert(method->size_of_parameters() > 0, "oob");
1158 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1159 VMReg r = regs[0].first();
1160 assert(r->is_valid(), "bad receiver arg");
1161 if (r->is_stack()) {
1162 // Porting note: This assumes that compiled calling conventions always
1163 // pass the receiver oop in a register. If this is not true on some
1164 // platform, pick a temp and load the receiver from stack.
1165 fatal("receiver always in a register");
1166 receiver_reg = x12; // known to be free at this point
1167 __ ld(receiver_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1168 } else {
1169 // no data motion is needed
1170 receiver_reg = r->as_Register();
1171 }
1172 }
1173
1174 // Figure out which address we are really jumping to:
1175 MethodHandles::generate_method_handle_dispatch(masm, iid,
1176 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1177 }
1178
1179 // ---------------------------------------------------------------------------
1180 // Generate a native wrapper for a given method. The method takes arguments
1181 // in the Java compiled code convention, marshals them to the native
1182 // convention (handlizes oops, etc), transitions to native, makes the call,
1183 // returns to java state (possibly blocking), unhandlizes any result and
1184 // returns.
1185 //
1186 // Critical native functions are a shorthand for the use of
1187 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1188 // functions. The wrapper is expected to unpack the arguments before
1189 // passing them to the callee and perform checks before and after the
1190 // native call to ensure that they GCLocker
1191 // lock_critical/unlock_critical semantics are followed. Some other
1192 // parts of JNI setup are skipped like the tear down of the JNI handle
1193 // block and the check for pending exceptions it's impossible for them
1194 // to be thrown.
1195 //
1196 // They are roughly structured like this:
1197 // if (GCLocker::needs_gc()) SharedRuntime::block_for_jni_critical()
1198 // tranistion to thread_in_native
1199 // unpack array arguments and call native entry point
1200 // check for safepoint in progress
1201 // check if any thread suspend flags are set
1202 // call into JVM and possible unlock the JNI critical
1203 // if a GC was suppressed while in the critical native.
1204 // transition back to thread_in_Java
1205 // return to caller
1206 //
1207 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1208 const methodHandle& method,
1209 int compile_id,
1210 BasicType* in_sig_bt,
1211 VMRegPair* in_regs,
1212 BasicType ret_type) {
1213 if (method->is_continuation_native_intrinsic()) {
1214 int exception_offset = -1;
1215 OopMapSet* oop_maps = new OopMapSet();
1216 int frame_complete = -1;
1217 int stack_slots = -1;
1218 int interpreted_entry_offset = -1;
1219 int vep_offset = -1;
1220 if (method->is_continuation_enter_intrinsic()) {
1221 gen_continuation_enter(masm,
1222 method,
1223 in_sig_bt,
1224 in_regs,
1225 exception_offset,
1226 oop_maps,
1227 frame_complete,
1228 stack_slots,
1229 interpreted_entry_offset,
1230 vep_offset);
1231 } else if (method->is_continuation_yield_intrinsic()) {
1232 gen_continuation_yield(masm,
1233 method,
1234 in_sig_bt,
1235 in_regs,
1236 oop_maps,
1237 frame_complete,
1238 stack_slots,
1239 vep_offset);
1240 } else {
1241 guarantee(false, "Unknown Continuation native intrinsic");
1242 }
1243
1244 #ifdef ASSERT
1245 if (method->is_continuation_enter_intrinsic()) {
1246 assert(interpreted_entry_offset != -1, "Must be set");
1247 assert(exception_offset != -1, "Must be set");
1248 } else {
1249 assert(interpreted_entry_offset == -1, "Must be unset");
1250 assert(exception_offset == -1, "Must be unset");
1251 }
1252 assert(frame_complete != -1, "Must be set");
1253 assert(stack_slots != -1, "Must be set");
1254 assert(vep_offset != -1, "Must be set");
1255 #endif
1256
1257 __ flush();
1258 nmethod* nm = nmethod::new_native_nmethod(method,
1259 compile_id,
1260 masm->code(),
1261 vep_offset,
1262 frame_complete,
1263 stack_slots,
1264 in_ByteSize(-1),
1265 in_ByteSize(-1),
1266 oop_maps,
1267 exception_offset);
1268 if (nm == nullptr) return nm;
1269 if (method->is_continuation_enter_intrinsic()) {
1270 ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
1271 } else if (method->is_continuation_yield_intrinsic()) {
1272 _cont_doYield_stub = nm;
1273 } else {
1274 guarantee(false, "Unknown Continuation native intrinsic");
1275 }
1276 return nm;
1277 }
1278
1279 if (method->is_method_handle_intrinsic()) {
1280 vmIntrinsics::ID iid = method->intrinsic_id();
1281 intptr_t start = (intptr_t)__ pc();
1282 int vep_offset = ((intptr_t)__ pc()) - start;
1283
1284 // First instruction must be a nop as it may need to be patched on deoptimisation
1285 {
1286 Assembler::IncompressibleScope scope(masm); // keep the nop as 4 bytes for patching.
1287 MacroAssembler::assert_alignment(__ pc());
1288 __ nop(); // 4 bytes
1289 }
1290 gen_special_dispatch(masm,
1291 method,
1292 in_sig_bt,
1293 in_regs);
1294 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
1295 __ flush();
1296 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
1297 return nmethod::new_native_nmethod(method,
1298 compile_id,
1299 masm->code(),
1300 vep_offset,
1301 frame_complete,
1302 stack_slots / VMRegImpl::slots_per_word,
1303 in_ByteSize(-1),
1304 in_ByteSize(-1),
1305 (OopMapSet*)nullptr);
1306 }
1307 address native_func = method->native_function();
1308 assert(native_func != nullptr, "must have function");
1309
1310 // An OopMap for lock (and class if static)
1311 OopMapSet *oop_maps = new OopMapSet();
1312 assert_cond(oop_maps != nullptr);
1313 intptr_t start = (intptr_t)__ pc();
1314
1315 // We have received a description of where all the java arg are located
1316 // on entry to the wrapper. We need to convert these args to where
1317 // the jni function will expect them. To figure out where they go
1318 // we convert the java signature to a C signature by inserting
1319 // the hidden arguments as arg[0] and possibly arg[1] (static method)
1320
1321 const int total_in_args = method->size_of_parameters();
1322 int total_c_args = total_in_args + (method->is_static() ? 2 : 1);
1323
1324 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1325 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1326
1327 int argc = 0;
1328 out_sig_bt[argc++] = T_ADDRESS;
1329 if (method->is_static()) {
1330 out_sig_bt[argc++] = T_OBJECT;
1331 }
1332
1333 for (int i = 0; i < total_in_args ; i++) {
1334 out_sig_bt[argc++] = in_sig_bt[i];
1335 }
1336
1337 // Now figure out where the args must be stored and how much stack space
1338 // they require.
1339 int out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
1340
1341 // Compute framesize for the wrapper. We need to handlize all oops in
1342 // incoming registers
1343
1344 // Calculate the total number of stack slots we will need.
1345
1346 // First count the abi requirement plus all of the outgoing args
1347 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1348
1349 // Now the space for the inbound oop handle area
1350 int total_save_slots = 8 * VMRegImpl::slots_per_word; // 8 arguments passed in registers
1351
1352 int oop_handle_offset = stack_slots;
1353 stack_slots += total_save_slots;
1354
1355 // Now any space we need for handlizing a klass if static method
1356
1357 int klass_slot_offset = 0;
1358 int klass_offset = -1;
1359 int lock_slot_offset = 0;
1360 bool is_static = false;
1361
1362 if (method->is_static()) {
1363 klass_slot_offset = stack_slots;
1364 stack_slots += VMRegImpl::slots_per_word;
1365 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1366 is_static = true;
1367 }
1368
1369 // Plus a lock if needed
1370
1371 if (method->is_synchronized()) {
1372 lock_slot_offset = stack_slots;
1373 stack_slots += VMRegImpl::slots_per_word;
1374 }
1375
1376 // Now a place (+2) to save return values or temp during shuffling
1377 // + 4 for return address (which we own) and saved fp
1378 stack_slots += 6;
1379
1380 // Ok The space we have allocated will look like:
1381 //
1382 //
1383 // FP-> | |
1384 // | 2 slots (ra) |
1385 // | 2 slots (fp) |
1386 // |---------------------|
1387 // | 2 slots for moves |
1388 // |---------------------|
1389 // | lock box (if sync) |
1390 // |---------------------| <- lock_slot_offset
1391 // | klass (if static) |
1392 // |---------------------| <- klass_slot_offset
1393 // | oopHandle area |
1394 // |---------------------| <- oop_handle_offset (8 java arg registers)
1395 // | outbound memory |
1396 // | based arguments |
1397 // | |
1398 // |---------------------|
1399 // | |
1400 // SP-> | out_preserved_slots |
1401 //
1402 //
1403
1404
1405 // Now compute actual number of stack words we need rounding to make
1406 // stack properly aligned.
1407 stack_slots = align_up(stack_slots, StackAlignmentInSlots);
1408
1409 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1410
1411 // First thing make an ic check to see if we should even be here
1412
1413 // We are free to use all registers as temps without saving them and
1414 // restoring them except fp. fp is the only callee save register
1415 // as far as the interpreter and the compiler(s) are concerned.
1416
1417 const Register receiver = j_rarg0;
1418
1419 __ verify_oop(receiver);
1420 assert_different_registers(receiver, t0, t1);
1421
1422 __ ic_check();
1423
1424 int vep_offset = ((intptr_t)__ pc()) - start;
1425
1426 // If we have to make this method not-entrant we'll overwrite its
1427 // first instruction with a jump.
1428 {
1429 Assembler::IncompressibleScope scope(masm); // keep the nop as 4 bytes for patching.
1430 MacroAssembler::assert_alignment(__ pc());
1431 __ nop(); // 4 bytes
1432 }
1433
1434 if (method->needs_clinit_barrier()) {
1435 assert(VM_Version::supports_fast_class_init_checks(), "sanity");
1436 Label L_skip_barrier;
1437 __ mov_metadata(t1, method->method_holder()); // InstanceKlass*
1438 __ clinit_barrier(t1, t0, &L_skip_barrier);
1439 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
1440
1441 __ bind(L_skip_barrier);
1442 }
1443
1444 // Generate stack overflow check
1445 __ bang_stack_with_offset(checked_cast<int>(StackOverflow::stack_shadow_zone_size()));
1446
1447 // Generate a new frame for the wrapper.
1448 __ enter();
1449 // -2 because return address is already present and so is saved fp
1450 __ sub(sp, sp, stack_size - 2 * wordSize);
1451
1452 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1453 assert_cond(bs != nullptr);
1454 bs->nmethod_entry_barrier(masm, nullptr /* slow_path */, nullptr /* continuation */, nullptr /* guard */);
1455
1456 // Frame is now completed as far as size and linkage.
1457 int frame_complete = ((intptr_t)__ pc()) - start;
1458
1459 // We use x18 as the oop handle for the receiver/klass
1460 // It is callee save so it survives the call to native
1461
1462 const Register oop_handle_reg = x18;
1463
1464 //
1465 // We immediately shuffle the arguments so that any vm call we have to
1466 // make from here on out (sync slow path, jvmti, etc.) we will have
1467 // captured the oops from our caller and have a valid oopMap for
1468 // them.
1469
1470 // -----------------
1471 // The Grand Shuffle
1472
1473 // The Java calling convention is either equal (linux) or denser (win64) than the
1474 // c calling convention. However the because of the jni_env argument the c calling
1475 // convention always has at least one more (and two for static) arguments than Java.
1476 // Therefore if we move the args from java -> c backwards then we will never have
1477 // a register->register conflict and we don't have to build a dependency graph
1478 // and figure out how to break any cycles.
1479 //
1480
1481 // Record esp-based slot for receiver on stack for non-static methods
1482 int receiver_offset = -1;
1483
1484 // This is a trick. We double the stack slots so we can claim
1485 // the oops in the caller's frame. Since we are sure to have
1486 // more args than the caller doubling is enough to make
1487 // sure we can capture all the incoming oop args from the
1488 // caller.
1489 //
1490 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1491 assert_cond(map != nullptr);
1492
1493 int float_args = 0;
1494 int int_args = 0;
1495
1496 #ifdef ASSERT
1497 bool reg_destroyed[Register::number_of_registers];
1498 bool freg_destroyed[FloatRegister::number_of_registers];
1499 for ( int r = 0 ; r < Register::number_of_registers ; r++ ) {
1500 reg_destroyed[r] = false;
1501 }
1502 for ( int f = 0 ; f < FloatRegister::number_of_registers ; f++ ) {
1503 freg_destroyed[f] = false;
1504 }
1505
1506 #endif /* ASSERT */
1507
1508 // For JNI natives the incoming and outgoing registers are offset upwards.
1509 GrowableArray<int> arg_order(2 * total_in_args);
1510
1511 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) {
1512 arg_order.push(i);
1513 arg_order.push(c_arg);
1514 }
1515
1516 for (int ai = 0; ai < arg_order.length(); ai += 2) {
1517 int i = arg_order.at(ai);
1518 int c_arg = arg_order.at(ai + 1);
1519 __ block_comment(err_msg("mv %d -> %d", i, c_arg));
1520 assert(c_arg != -1 && i != -1, "wrong order");
1521 #ifdef ASSERT
1522 if (in_regs[i].first()->is_Register()) {
1523 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!");
1524 } else if (in_regs[i].first()->is_FloatRegister()) {
1525 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!");
1526 }
1527 if (out_regs[c_arg].first()->is_Register()) {
1528 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
1529 } else if (out_regs[c_arg].first()->is_FloatRegister()) {
1530 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true;
1531 }
1532 #endif /* ASSERT */
1533 switch (in_sig_bt[i]) {
1534 case T_ARRAY:
1535 case T_OBJECT:
1536 __ object_move(map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
1537 ((i == 0) && (!is_static)),
1538 &receiver_offset);
1539 int_args++;
1540 break;
1541 case T_VOID:
1542 break;
1543
1544 case T_FLOAT:
1545 __ float_move(in_regs[i], out_regs[c_arg]);
1546 float_args++;
1547 break;
1548
1549 case T_DOUBLE:
1550 assert( i + 1 < total_in_args &&
1551 in_sig_bt[i + 1] == T_VOID &&
1552 out_sig_bt[c_arg + 1] == T_VOID, "bad arg list");
1553 __ double_move(in_regs[i], out_regs[c_arg]);
1554 float_args++;
1555 break;
1556
1557 case T_LONG :
1558 __ long_move(in_regs[i], out_regs[c_arg]);
1559 int_args++;
1560 break;
1561
1562 case T_ADDRESS:
1563 assert(false, "found T_ADDRESS in java args");
1564 break;
1565
1566 default:
1567 __ move32_64(in_regs[i], out_regs[c_arg]);
1568 int_args++;
1569 }
1570 }
1571
1572 // point c_arg at the first arg that is already loaded in case we
1573 // need to spill before we call out
1574 int c_arg = total_c_args - total_in_args;
1575
1576 // Pre-load a static method's oop into c_rarg1.
1577 if (method->is_static()) {
1578
1579 // load oop into a register
1580 __ movoop(c_rarg1,
1581 JNIHandles::make_local(method->method_holder()->java_mirror()));
1582
1583 // Now handlize the static class mirror it's known not-null.
1584 __ sd(c_rarg1, Address(sp, klass_offset));
1585 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
1586
1587 // Now get the handle
1588 __ la(c_rarg1, Address(sp, klass_offset));
1589 // and protect the arg if we must spill
1590 c_arg--;
1591 }
1592
1593 // Change state to native (we save the return address in the thread, since it might not
1594 // be pushed on the stack when we do a stack traversal). It is enough that the pc()
1595 // points into the right code segment. It does not have to be the correct return pc.
1596 // We use the same pc/oopMap repeatedly when we call out.
1597
1598 Label native_return;
1599 if (method->is_object_wait0()) {
1600 // For convenience we use the pc we want to resume to in case of preemption on Object.wait.
1601 __ set_last_Java_frame(sp, noreg, native_return, t0);
1602 } else {
1603 intptr_t the_pc = (intptr_t) __ pc();
1604 oop_maps->add_gc_map(the_pc - start, map);
1605
1606 __ set_last_Java_frame(sp, noreg, __ pc(), t0);
1607 }
1608
1609 Label dtrace_method_entry, dtrace_method_entry_done;
1610 if (DTraceMethodProbes) {
1611 __ j(dtrace_method_entry);
1612 __ bind(dtrace_method_entry_done);
1613 }
1614
1615 // RedefineClasses() tracing support for obsolete method entry
1616 if (log_is_enabled(Trace, redefine, class, obsolete)) {
1617 // protect the args we've loaded
1618 save_args(masm, total_c_args, c_arg, out_regs);
1619 __ mov_metadata(c_rarg1, method());
1620 __ call_VM_leaf(
1621 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
1622 xthread, c_rarg1);
1623 restore_args(masm, total_c_args, c_arg, out_regs);
1624 }
1625
1626 // Lock a synchronized method
1627
1628 // Register definitions used by locking and unlocking
1629
1630 const Register swap_reg = x10;
1631 const Register obj_reg = x9; // Will contain the oop
1632 const Register lock_reg = x30; // Address of compiler lock object (BasicLock)
1633 const Register old_hdr = x30; // value of old header at unlock time
1634 const Register lock_tmp = x31; // Temporary used by fast_lock/unlock
1635 const Register tmp = ra;
1636
1637 Label slow_path_lock;
1638 Label lock_done;
1639
1640 if (method->is_synchronized()) {
1641 // Get the handle (the 2nd argument)
1642 __ mv(oop_handle_reg, c_rarg1);
1643
1644 // Get address of the box
1645
1646 __ la(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1647
1648 // Load the oop from the handle
1649 __ ld(obj_reg, Address(oop_handle_reg, 0));
1650
1651 __ fast_lock(lock_reg, obj_reg, swap_reg, tmp, lock_tmp, slow_path_lock);
1652
1653 // Slow path will re-enter here
1654 __ bind(lock_done);
1655 }
1656
1657
1658 // Finally just about ready to make the JNI call
1659
1660 // get JNIEnv* which is first argument to native
1661 __ la(c_rarg0, Address(xthread, in_bytes(JavaThread::jni_environment_offset())));
1662
1663 // Now set thread in native
1664 __ la(t1, Address(xthread, JavaThread::thread_state_offset()));
1665 __ mv(t0, _thread_in_native);
1666 __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore);
1667 __ sw(t0, Address(t1));
1668
1669 // Clobbers t1
1670 __ rt_call(native_func);
1671
1672 // Verify or restore cpu control state after JNI call
1673 __ restore_cpu_control_state_after_jni(t0);
1674
1675 // Unpack native results.
1676 if (ret_type != T_OBJECT && ret_type != T_ARRAY) {
1677 __ cast_primitive_type(ret_type, x10);
1678 }
1679
1680 Label safepoint_in_progress, safepoint_in_progress_done;
1681
1682 // Switch thread to "native transition" state before reading the synchronization state.
1683 // This additional state is necessary because reading and testing the synchronization
1684 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1685 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1686 // VM thread changes sync state to synchronizing and suspends threads for GC.
1687 // Thread A is resumed to finish this native method, but doesn't block here since it
1688 // didn't see any synchronization is progress, and escapes.
1689 __ mv(t0, _thread_in_native_trans);
1690
1691 __ sw(t0, Address(xthread, JavaThread::thread_state_offset()));
1692
1693 // Force this write out before the read below
1694 if (!UseSystemMemoryBarrier) {
1695 __ membar(MacroAssembler::AnyAny);
1696 }
1697
1698 // check for safepoint operation in progress and/or pending suspend requests
1699 {
1700 __ safepoint_poll(safepoint_in_progress, true /* at_return */, false /* in_nmethod */);
1701 __ lwu(t0, Address(xthread, JavaThread::suspend_flags_offset()));
1702 __ bnez(t0, safepoint_in_progress);
1703 __ bind(safepoint_in_progress_done);
1704 }
1705
1706 // change thread state
1707 __ la(t1, Address(xthread, JavaThread::thread_state_offset()));
1708 __ mv(t0, _thread_in_Java);
1709 __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore);
1710 __ sw(t0, Address(t1));
1711
1712 if (method->is_object_wait0()) {
1713 // Check preemption for Object.wait()
1714 __ ld(t1, Address(xthread, JavaThread::preempt_alternate_return_offset()));
1715 __ beqz(t1, native_return);
1716 __ sd(zr, Address(xthread, JavaThread::preempt_alternate_return_offset()));
1717 __ jr(t1);
1718 __ bind(native_return);
1719
1720 intptr_t the_pc = (intptr_t) __ pc();
1721 oop_maps->add_gc_map(the_pc - start, map);
1722 }
1723
1724 Label reguard;
1725 Label reguard_done;
1726 __ lbu(t0, Address(xthread, JavaThread::stack_guard_state_offset()));
1727 __ mv(t1, StackOverflow::stack_guard_yellow_reserved_disabled);
1728 __ beq(t0, t1, reguard);
1729 __ bind(reguard_done);
1730
1731 // native result if any is live
1732
1733 // Unlock
1734 Label unlock_done;
1735 Label slow_path_unlock;
1736 if (method->is_synchronized()) {
1737
1738 // Get locked oop from the handle we passed to jni
1739 __ ld(obj_reg, Address(oop_handle_reg, 0));
1740
1741 // Must save x10 if if it is live now because cmpxchg must use it
1742 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1743 save_native_result(masm, ret_type, stack_slots);
1744 }
1745
1746 __ fast_unlock(obj_reg, old_hdr, swap_reg, lock_tmp, slow_path_unlock);
1747
1748 // slow path re-enters here
1749 __ bind(unlock_done);
1750 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1751 restore_native_result(masm, ret_type, stack_slots);
1752 }
1753 }
1754
1755 Label dtrace_method_exit, dtrace_method_exit_done;
1756 if (DTraceMethodProbes) {
1757 __ j(dtrace_method_exit);
1758 __ bind(dtrace_method_exit_done);
1759 }
1760
1761 __ reset_last_Java_frame(false);
1762
1763 // Unbox oop result, e.g. JNIHandles::resolve result.
1764 if (is_reference_type(ret_type)) {
1765 __ resolve_jobject(x10, x11, x12);
1766 }
1767
1768 if (CheckJNICalls) {
1769 // clear_pending_jni_exception_check
1770 __ sd(zr, Address(xthread, JavaThread::pending_jni_exception_check_fn_offset()));
1771 }
1772
1773 // reset handle block
1774 __ ld(x12, Address(xthread, JavaThread::active_handles_offset()));
1775 __ sd(zr, Address(x12, JNIHandleBlock::top_offset()));
1776
1777 __ leave();
1778
1779 #if INCLUDE_JFR
1780 // We need to do a poll test after unwind in case the sampler
1781 // managed to sample the native frame after returning to Java.
1782 Label L_return;
1783 __ ld(t0, Address(xthread, JavaThread::polling_word_offset()));
1784 address poll_test_pc = __ pc();
1785 __ relocate(relocInfo::poll_return_type);
1786 __ test_bit(t0, t0, log2i_exact(SafepointMechanism::poll_bit()));
1787 __ beqz(t0, L_return);
1788 assert(SharedRuntime::polling_page_return_handler_blob() != nullptr,
1789 "polling page return stub not created yet");
1790 address stub = SharedRuntime::polling_page_return_handler_blob()->entry_point();
1791 __ la(t0, InternalAddress(poll_test_pc));
1792 __ sd(t0, Address(xthread, JavaThread::saved_exception_pc_offset()));
1793 __ far_jump(RuntimeAddress(stub));
1794 __ bind(L_return);
1795 #endif // INCLUDE_JFR
1796
1797 // Any exception pending?
1798 Label exception_pending;
1799 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1800 __ bnez(t0, exception_pending);
1801
1802 // We're done
1803 __ ret();
1804
1805 // Unexpected paths are out of line and go here
1806
1807 // forward the exception
1808 __ bind(exception_pending);
1809
1810 // and forward the exception
1811 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
1812
1813 // Slow path locking & unlocking
1814 if (method->is_synchronized()) {
1815
1816 __ block_comment("Slow path lock {");
1817 __ bind(slow_path_lock);
1818
1819 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM
1820 // args are (oop obj, BasicLock* lock, JavaThread* thread)
1821
1822 // protect the args we've loaded
1823 save_args(masm, total_c_args, c_arg, out_regs);
1824
1825 __ mv(c_rarg0, obj_reg);
1826 __ mv(c_rarg1, lock_reg);
1827 __ mv(c_rarg2, xthread);
1828
1829 // Not a leaf but we have last_Java_frame setup as we want.
1830 // We don't want to unmount in case of contention since that would complicate preserving
1831 // the arguments that had already been marshalled into the native convention. So we force
1832 // the freeze slow path to find this native wrapper frame (see recurse_freeze_native_frame())
1833 // and pin the vthread. Otherwise the fast path won't find it since we don't walk the stack.
1834 __ push_cont_fastpath();
1835 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3);
1836 __ pop_cont_fastpath();
1837 restore_args(masm, total_c_args, c_arg, out_regs);
1838
1839 #ifdef ASSERT
1840 { Label L;
1841 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1842 __ beqz(t0, L);
1843 __ stop("no pending exception allowed on exit from monitorenter");
1844 __ bind(L);
1845 }
1846 #endif
1847 __ j(lock_done);
1848
1849 __ block_comment("} Slow path lock");
1850
1851 __ block_comment("Slow path unlock {");
1852 __ bind(slow_path_unlock);
1853
1854 if (ret_type == T_FLOAT || ret_type == T_DOUBLE) {
1855 save_native_result(masm, ret_type, stack_slots);
1856 }
1857
1858 __ mv(c_rarg2, xthread);
1859 __ la(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1860 __ mv(c_rarg0, obj_reg);
1861
1862 // Save pending exception around call to VM (which contains an EXCEPTION_MARK)
1863 // NOTE that obj_reg == x9 currently
1864 __ ld(x9, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1865 __ sd(zr, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1866
1867 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C));
1868
1869 #ifdef ASSERT
1870 {
1871 Label L;
1872 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1873 __ beqz(t0, L);
1874 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C");
1875 __ bind(L);
1876 }
1877 #endif /* ASSERT */
1878
1879 __ sd(x9, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1880
1881 if (ret_type == T_FLOAT || ret_type == T_DOUBLE) {
1882 restore_native_result(masm, ret_type, stack_slots);
1883 }
1884 __ j(unlock_done);
1885
1886 __ block_comment("} Slow path unlock");
1887
1888 } // synchronized
1889
1890 // SLOW PATH Reguard the stack if needed
1891
1892 __ bind(reguard);
1893 save_native_result(masm, ret_type, stack_slots);
1894 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
1895 restore_native_result(masm, ret_type, stack_slots);
1896 // and continue
1897 __ j(reguard_done);
1898
1899 // SLOW PATH safepoint
1900 {
1901 __ block_comment("safepoint {");
1902 __ bind(safepoint_in_progress);
1903
1904 // Don't use call_VM as it will see a possible pending exception and forward it
1905 // and never return here preventing us from clearing _last_native_pc down below.
1906 //
1907 save_native_result(masm, ret_type, stack_slots);
1908 __ mv(c_rarg0, xthread);
1909 #ifndef PRODUCT
1910 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
1911 #endif
1912 __ rt_call(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans));
1913
1914 // Restore any method result value
1915 restore_native_result(masm, ret_type, stack_slots);
1916
1917 __ j(safepoint_in_progress_done);
1918 __ block_comment("} safepoint");
1919 }
1920
1921 // SLOW PATH dtrace support
1922 if (DTraceMethodProbes) {
1923 {
1924 __ block_comment("dtrace entry {");
1925 __ bind(dtrace_method_entry);
1926
1927 // We have all of the arguments setup at this point. We must not touch any register
1928 // argument registers at this point (what if we save/restore them there are no oop?
1929
1930 save_args(masm, total_c_args, c_arg, out_regs);
1931 __ mov_metadata(c_rarg1, method());
1932 __ call_VM_leaf(
1933 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
1934 xthread, c_rarg1);
1935 restore_args(masm, total_c_args, c_arg, out_regs);
1936 __ j(dtrace_method_entry_done);
1937 __ block_comment("} dtrace entry");
1938 }
1939
1940 {
1941 __ block_comment("dtrace exit {");
1942 __ bind(dtrace_method_exit);
1943 save_native_result(masm, ret_type, stack_slots);
1944 __ mov_metadata(c_rarg1, method());
1945 __ call_VM_leaf(
1946 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
1947 xthread, c_rarg1);
1948 restore_native_result(masm, ret_type, stack_slots);
1949 __ j(dtrace_method_exit_done);
1950 __ block_comment("} dtrace exit");
1951 }
1952 }
1953
1954 __ flush();
1955
1956 nmethod *nm = nmethod::new_native_nmethod(method,
1957 compile_id,
1958 masm->code(),
1959 vep_offset,
1960 frame_complete,
1961 stack_slots / VMRegImpl::slots_per_word,
1962 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
1963 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size),
1964 oop_maps);
1965 assert(nm != nullptr, "create native nmethod fail!");
1966 return nm;
1967 }
1968
1969 // this function returns the adjust size (in number of words) to a c2i adapter
1970 // activation for use during deoptimization
1971 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
1972 assert(callee_locals >= callee_parameters,
1973 "test and remove; got more parms than locals");
1974 if (callee_locals < callee_parameters) {
1975 return 0; // No adjustment for negative locals
1976 }
1977 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords;
1978 // diff is counted in stack words
1979 return align_up(diff, 2);
1980 }
1981
1982 //------------------------------generate_deopt_blob----------------------------
1983 void SharedRuntime::generate_deopt_blob() {
1984 // Allocate space for the code
1985 ResourceMark rm;
1986 // Setup code generation tools
1987 int pad = 0;
1988 const char* name = SharedRuntime::stub_name(StubId::shared_deopt_id);
1989 CodeBuffer buffer(name, 2048 + pad, 1024);
1990 MacroAssembler* masm = new MacroAssembler(&buffer);
1991 int frame_size_in_words = -1;
1992 OopMap* map = nullptr;
1993 OopMapSet *oop_maps = new OopMapSet();
1994 assert_cond(masm != nullptr && oop_maps != nullptr);
1995 RegisterSaver reg_saver(COMPILER2_PRESENT(true) NOT_COMPILER2(false));
1996
1997 // -------------
1998 // This code enters when returning to a de-optimized nmethod. A return
1999 // address has been pushed on the stack, and return values are in
2000 // registers.
2001 // If we are doing a normal deopt then we were called from the patched
2002 // nmethod from the point we returned to the nmethod. So the return
2003 // address on the stack is wrong by NativeCall::instruction_size
2004 // We will adjust the value so it looks like we have the original return
2005 // address on the stack (like when we eagerly deoptimized).
2006 // In the case of an exception pending when deoptimizing, we enter
2007 // with a return address on the stack that points after the call we patched
2008 // into the exception handler. We have the following register state from,
2009 // e.g., the forward exception stub (see stubGenerator_riscv.cpp).
2010 // x10: exception oop
2011 // x9: exception handler
2012 // x13: throwing pc
2013 // So in this case we simply jam x13 into the useless return address and
2014 // the stack looks just like we want.
2015 //
2016 // At this point we need to de-opt. We save the argument return
2017 // registers. We call the first C routine, fetch_unroll_info(). This
2018 // routine captures the return values and returns a structure which
2019 // describes the current frame size and the sizes of all replacement frames.
2020 // The current frame is compiled code and may contain many inlined
2021 // functions, each with their own JVM state. We pop the current frame, then
2022 // push all the new frames. Then we call the C routine unpack_frames() to
2023 // populate these frames. Finally unpack_frames() returns us the new target
2024 // address. Notice that callee-save registers are BLOWN here; they have
2025 // already been captured in the vframeArray at the time the return PC was
2026 // patched.
2027 address start = __ pc();
2028 Label cont;
2029
2030 // Prolog for non exception case!
2031
2032 // Save everything in sight.
2033 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2034
2035 // Normal deoptimization. Save exec mode for unpack_frames.
2036 __ mv(xcpool, Deoptimization::Unpack_deopt); // callee-saved
2037 __ j(cont);
2038
2039 int reexecute_offset = __ pc() - start;
2040 // Reexecute case
2041 // return address is the pc describes what bci to do re-execute at
2042
2043 // No need to update map as each call to save_live_registers will produce identical oopmap
2044 (void) reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2045
2046 __ mv(xcpool, Deoptimization::Unpack_reexecute); // callee-saved
2047 __ j(cont);
2048
2049 int exception_offset = __ pc() - start;
2050
2051 // Prolog for exception case
2052
2053 // all registers are dead at this entry point, except for x10, and
2054 // x13 which contain the exception oop and exception pc
2055 // respectively. Set them in TLS and fall thru to the
2056 // unpack_with_exception_in_tls entry point.
2057
2058 __ sd(x13, Address(xthread, JavaThread::exception_pc_offset()));
2059 __ sd(x10, Address(xthread, JavaThread::exception_oop_offset()));
2060
2061 int exception_in_tls_offset = __ pc() - start;
2062
2063 // new implementation because exception oop is now passed in JavaThread
2064
2065 // Prolog for exception case
2066 // All registers must be preserved because they might be used by LinearScan
2067 // Exceptiop oop and throwing PC are passed in JavaThread
2068 // tos: stack at point of call to method that threw the exception (i.e. only
2069 // args are on the stack, no return address)
2070
2071 // The return address pushed by save_live_registers will be patched
2072 // later with the throwing pc. The correct value is not available
2073 // now because loading it from memory would destroy registers.
2074
2075 // NB: The SP at this point must be the SP of the method that is
2076 // being deoptimized. Deoptimization assumes that the frame created
2077 // here by save_live_registers is immediately below the method's SP.
2078 // This is a somewhat fragile mechanism.
2079
2080 // Save everything in sight.
2081 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2082
2083 // Now it is safe to overwrite any register
2084
2085 // Deopt during an exception. Save exec mode for unpack_frames.
2086 __ mv(xcpool, Deoptimization::Unpack_exception); // callee-saved
2087
2088 // load throwing pc from JavaThread and patch it as the return address
2089 // of the current frame. Then clear the field in JavaThread
2090
2091 __ ld(x13, Address(xthread, JavaThread::exception_pc_offset()));
2092 __ sd(x13, Address(fp, frame::return_addr_offset * wordSize));
2093 __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2094
2095 #ifdef ASSERT
2096 // verify that there is really an exception oop in JavaThread
2097 __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2098 __ verify_oop(x10);
2099
2100 // verify that there is no pending exception
2101 Label no_pending_exception;
2102 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2103 __ beqz(t0, no_pending_exception);
2104 __ stop("must not have pending exception here");
2105 __ bind(no_pending_exception);
2106 #endif
2107
2108 __ bind(cont);
2109
2110 // Call C code. Need thread and this frame, but NOT official VM entry
2111 // crud. We cannot block on this call, no GC can happen.
2112 //
2113 // UnrollBlock* fetch_unroll_info(JavaThread* thread)
2114
2115 // fetch_unroll_info needs to call last_java_frame().
2116
2117 Label retaddr;
2118 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2119 #ifdef ASSERT
2120 {
2121 Label L;
2122 __ ld(t0, Address(xthread,
2123 JavaThread::last_Java_fp_offset()));
2124 __ beqz(t0, L);
2125 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared");
2126 __ bind(L);
2127 }
2128 #endif // ASSERT
2129 __ mv(c_rarg0, xthread);
2130 __ mv(c_rarg1, xcpool);
2131 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info));
2132 __ bind(retaddr);
2133
2134 // Need to have an oopmap that tells fetch_unroll_info where to
2135 // find any register it might need.
2136 oop_maps->add_gc_map(__ pc() - start, map);
2137
2138 __ reset_last_Java_frame(false);
2139
2140 // Load UnrollBlock* into x15
2141 __ mv(x15, x10);
2142
2143 __ lwu(xcpool, Address(x15, Deoptimization::UnrollBlock::unpack_kind_offset()));
2144 Label noException;
2145 __ mv(t0, Deoptimization::Unpack_exception);
2146 __ bne(xcpool, t0, noException); // Was exception pending?
2147 __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2148 __ ld(x13, Address(xthread, JavaThread::exception_pc_offset()));
2149 __ sd(zr, Address(xthread, JavaThread::exception_oop_offset()));
2150 __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2151
2152 __ verify_oop(x10);
2153
2154 // Overwrite the result registers with the exception results.
2155 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2156
2157 __ bind(noException);
2158
2159 // Only register save data is on the stack.
2160 // Now restore the result registers. Everything else is either dead
2161 // or captured in the vframeArray.
2162
2163 // Restore fp result register
2164 __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2165 // Restore integer result register
2166 __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2167
2168 // Pop all of the register save area off the stack
2169 __ add(sp, sp, frame_size_in_words * wordSize);
2170
2171 // All of the register save area has been popped of the stack. Only the
2172 // return address remains.
2173
2174 // Pop all the frames we must move/replace.
2175 //
2176 // Frame picture (youngest to oldest)
2177 // 1: self-frame (no frame link)
2178 // 2: deopting frame (no frame link)
2179 // 3: caller of deopting frame (could be compiled/interpreted).
2180 //
2181 // Note: by leaving the return address of self-frame on the stack
2182 // and using the size of frame 2 to adjust the stack
2183 // when we are done the return to frame 3 will still be on the stack.
2184
2185 // Pop deoptimized frame
2186 __ lwu(x12, Address(x15, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset()));
2187 __ subi(x12, x12, 2 * wordSize);
2188 __ add(sp, sp, x12);
2189 __ ld(fp, Address(sp, 0));
2190 __ ld(ra, Address(sp, wordSize));
2191 __ addi(sp, sp, 2 * wordSize);
2192 // RA should now be the return address to the caller (3)
2193
2194 #ifdef ASSERT
2195 // Compilers generate code that bang the stack by as much as the
2196 // interpreter would need. So this stack banging should never
2197 // trigger a fault. Verify that it does not on non product builds.
2198 __ lwu(x9, Address(x15, Deoptimization::UnrollBlock::total_frame_sizes_offset()));
2199 __ bang_stack_size(x9, x12);
2200 #endif
2201 // Load address of array of frame pcs into x12
2202 __ ld(x12, Address(x15, Deoptimization::UnrollBlock::frame_pcs_offset()));
2203
2204 // Load address of array of frame sizes into x14
2205 __ ld(x14, Address(x15, Deoptimization::UnrollBlock::frame_sizes_offset()));
2206
2207 // Load counter into x13
2208 __ lwu(x13, Address(x15, Deoptimization::UnrollBlock::number_of_frames_offset()));
2209
2210 // Now adjust the caller's stack to make up for the extra locals
2211 // but record the original sp so that we can save it in the skeletal interpreter
2212 // frame and the stack walking of interpreter_sender will get the unextended sp
2213 // value and not the "real" sp value.
2214
2215 const Register sender_sp = x16;
2216
2217 __ mv(sender_sp, sp);
2218 __ lwu(x9, Address(x15,
2219 Deoptimization::UnrollBlock::
2220 caller_adjustment_offset()));
2221 __ sub(sp, sp, x9);
2222
2223 // Push interpreter frames in a loop
2224 __ mv(t0, 0xDEADDEAD); // Make a recognizable pattern
2225 __ mv(t1, t0);
2226 Label loop;
2227 __ bind(loop);
2228 __ ld(x9, Address(x14, 0)); // Load frame size
2229 __ addi(x14, x14, wordSize);
2230 __ subi(x9, x9, 2 * wordSize); // We'll push pc and fp by hand
2231 __ ld(ra, Address(x12, 0)); // Load pc
2232 __ addi(x12, x12, wordSize);
2233 __ enter(); // Save old & set new fp
2234 __ sub(sp, sp, x9); // Prolog
2235 // This value is corrected by layout_activation_impl
2236 __ sd(zr, Address(fp, frame::interpreter_frame_last_sp_offset * wordSize));
2237 __ sd(sender_sp, Address(fp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable
2238 __ mv(sender_sp, sp); // Pass sender_sp to next frame
2239 __ subi(x13, x13, 1); // Decrement counter
2240 __ bnez(x13, loop);
2241
2242 // Re-push self-frame
2243 __ ld(ra, Address(x12));
2244 __ enter();
2245
2246 // Allocate a full sized register save area. We subtract 2 because
2247 // enter() just pushed 2 words
2248 __ sub(sp, sp, (frame_size_in_words - 2) * wordSize);
2249
2250 // Restore frame locals after moving the frame
2251 __ fsd(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2252 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2253
2254 // Call C code. Need thread but NOT official VM entry
2255 // crud. We cannot block on this call, no GC can happen. Call should
2256 // restore return values to their stack-slots with the new SP.
2257 //
2258 // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)
2259
2260 // Use fp because the frames look interpreted now
2261 // Don't need the precise return PC here, just precise enough to point into this code blob.
2262 address the_pc = __ pc();
2263 __ set_last_Java_frame(sp, fp, the_pc, t0);
2264
2265 __ mv(c_rarg0, xthread);
2266 __ mv(c_rarg1, xcpool); // second arg: exec_mode
2267 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames));
2268
2269 // Set an oopmap for the call site
2270 // Use the same PC we used for the last java frame
2271 oop_maps->add_gc_map(the_pc - start,
2272 new OopMap(frame_size_in_words, 0));
2273
2274 // Clear fp AND pc
2275 __ reset_last_Java_frame(true);
2276
2277 // Collect return values
2278 __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2279 __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2280
2281 // Pop self-frame.
2282 __ leave(); // Epilog
2283
2284 // Jump to interpreter
2285 __ ret();
2286
2287 // Make sure all code is generated
2288 masm->flush();
2289
2290 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words);
2291 assert(_deopt_blob != nullptr, "create deoptimization blob fail!");
2292 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
2293 }
2294
2295 // Number of stack slots between incoming argument block and the start of
2296 // a new frame. The PROLOG must add this many slots to the stack. The
2297 // EPILOG must remove this many slots.
2298 // RISCV needs two words for RA (return address) and FP (frame pointer).
2299 uint SharedRuntime::in_preserve_stack_slots() {
2300 return 2 * VMRegImpl::slots_per_word + (VerifyStackAtCalls ? 0 : 2) ;
2301 }
2302
2303 uint SharedRuntime::out_preserve_stack_slots() {
2304 return 0;
2305 }
2306
2307 VMReg SharedRuntime::thread_register() {
2308 return xthread->as_VMReg();
2309 }
2310
2311 //------------------------------generate_handler_blob------
2312 //
2313 // Generate a special Compile2Runtime blob that saves all registers,
2314 // and setup oopmap.
2315 //
2316 SafepointBlob* SharedRuntime::generate_handler_blob(StubId id, address call_ptr) {
2317 assert(is_polling_page_id(id), "expected a polling page stub id");
2318
2319 ResourceMark rm;
2320 OopMapSet *oop_maps = new OopMapSet();
2321 assert_cond(oop_maps != nullptr);
2322 OopMap* map = nullptr;
2323
2324 // Allocate space for the code. Setup code generation tools.
2325 const char* name = SharedRuntime::stub_name(id);
2326 CodeBuffer buffer(name, 2048, 1024);
2327 MacroAssembler* masm = new MacroAssembler(&buffer);
2328 assert_cond(masm != nullptr);
2329
2330 address start = __ pc();
2331 address call_pc = nullptr;
2332 int frame_size_in_words = -1;
2333 bool cause_return = (id == StubId::shared_polling_page_return_handler_id);
2334 RegisterSaver reg_saver(id == StubId::shared_polling_page_vectors_safepoint_handler_id /* save_vectors */);
2335
2336 // Save Integer and Float registers.
2337 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2338
2339 // The following is basically a call_VM. However, we need the precise
2340 // address of the call in order to generate an oopmap. Hence, we do all the
2341 // work ourselves.
2342
2343 Label retaddr;
2344 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2345
2346 // The return address must always be correct so that frame constructor never
2347 // sees an invalid pc.
2348
2349 if (!cause_return) {
2350 // overwrite the return address pushed by save_live_registers
2351 // Additionally, x18 is a callee-saved register so we can look at
2352 // it later to determine if someone changed the return address for
2353 // us!
2354 __ ld(x18, Address(xthread, JavaThread::saved_exception_pc_offset()));
2355 __ sd(x18, Address(fp, frame::return_addr_offset * wordSize));
2356 }
2357
2358 // Do the call
2359 __ mv(c_rarg0, xthread);
2360 __ rt_call(call_ptr);
2361 __ bind(retaddr);
2362
2363 // Set an oopmap for the call site. This oopmap will map all
2364 // oop-registers and debug-info registers as callee-saved. This
2365 // will allow deoptimization at this safepoint to find all possible
2366 // debug-info recordings, as well as let GC find all oops.
2367
2368 oop_maps->add_gc_map( __ pc() - start, map);
2369
2370 Label noException;
2371
2372 __ reset_last_Java_frame(false);
2373
2374 __ membar(MacroAssembler::LoadLoad | MacroAssembler::LoadStore);
2375
2376 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2377 __ beqz(t0, noException);
2378
2379 // Exception pending
2380
2381 reg_saver.restore_live_registers(masm);
2382
2383 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2384
2385 // No exception case
2386 __ bind(noException);
2387
2388 Label no_adjust, bail;
2389 if (!cause_return) {
2390 // If our stashed return pc was modified by the runtime we avoid touching it
2391 __ ld(t0, Address(fp, frame::return_addr_offset * wordSize));
2392 __ bne(x18, t0, no_adjust);
2393
2394 #ifdef ASSERT
2395 // Verify the correct encoding of the poll we're about to skip.
2396 // See NativeInstruction::is_lwu_to_zr()
2397 __ lwu(t0, Address(x18));
2398 __ andi(t1, t0, 0b1111111);
2399 __ mv(t2, 0b0000011);
2400 __ bne(t1, t2, bail); // 0-6:0b0000011
2401 __ srli(t1, t0, 7);
2402 __ andi(t1, t1, 0b11111);
2403 __ bnez(t1, bail); // 7-11:0b00000
2404 __ srli(t1, t0, 12);
2405 __ andi(t1, t1, 0b111);
2406 __ mv(t2, 0b110);
2407 __ bne(t1, t2, bail); // 12-14:0b110
2408 #endif
2409
2410 // Adjust return pc forward to step over the safepoint poll instruction
2411 __ addi(x18, x18, NativeInstruction::instruction_size);
2412 __ sd(x18, Address(fp, frame::return_addr_offset * wordSize));
2413 }
2414
2415 __ bind(no_adjust);
2416 // Normal exit, restore registers and exit.
2417
2418 reg_saver.restore_live_registers(masm);
2419 __ ret();
2420
2421 #ifdef ASSERT
2422 __ bind(bail);
2423 __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected");
2424 #endif
2425
2426 // Make sure all code is generated
2427 masm->flush();
2428
2429 // Fill-out other meta info
2430 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words);
2431 }
2432
2433 //
2434 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
2435 //
2436 // Generate a stub that calls into vm to find out the proper destination
2437 // of a java call. All the argument registers are live at this point
2438 // but since this is generic code we don't know what they are and the caller
2439 // must do any gc of the args.
2440 //
2441 RuntimeStub* SharedRuntime::generate_resolve_blob(StubId id, address destination) {
2442 assert(StubRoutines::forward_exception_entry() != nullptr, "must be generated before");
2443 assert(is_resolve_id(id), "expected a resolve stub id");
2444
2445 // allocate space for the code
2446 ResourceMark rm;
2447
2448 const char* name = SharedRuntime::stub_name(id);
2449 CodeBuffer buffer(name, 1000, 512);
2450 MacroAssembler* masm = new MacroAssembler(&buffer);
2451 assert_cond(masm != nullptr);
2452
2453 int frame_size_in_words = -1;
2454 RegisterSaver reg_saver(false /* save_vectors */);
2455
2456 OopMapSet *oop_maps = new OopMapSet();
2457 assert_cond(oop_maps != nullptr);
2458 OopMap* map = nullptr;
2459
2460 int start = __ offset();
2461
2462 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2463
2464 int frame_complete = __ offset();
2465
2466 {
2467 Label retaddr;
2468 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2469
2470 __ mv(c_rarg0, xthread);
2471 __ rt_call(destination);
2472 __ bind(retaddr);
2473 }
2474
2475 // Set an oopmap for the call site.
2476 // We need this not only for callee-saved registers, but also for volatile
2477 // registers that the compiler might be keeping live across a safepoint.
2478
2479 oop_maps->add_gc_map( __ offset() - start, map);
2480
2481 // x10 contains the address we are going to jump to assuming no exception got installed
2482
2483 // clear last_Java_sp
2484 __ reset_last_Java_frame(false);
2485 // check for pending exceptions
2486 Label pending;
2487 __ ld(t1, Address(xthread, Thread::pending_exception_offset()));
2488 __ bnez(t1, pending);
2489
2490 // get the returned Method*
2491 __ get_vm_result_metadata(xmethod, xthread);
2492 __ sd(xmethod, Address(sp, reg_saver.reg_offset_in_bytes(xmethod)));
2493
2494 // x10 is where we want to jump, overwrite t1 which is saved and temporary
2495 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(t1)));
2496 reg_saver.restore_live_registers(masm);
2497
2498 // We are back to the original state on entry and ready to go.
2499 __ jr(t1);
2500
2501 // Pending exception after the safepoint
2502
2503 __ bind(pending);
2504
2505 reg_saver.restore_live_registers(masm);
2506
2507 // exception pending => remove activation and forward to exception handler
2508
2509 __ sd(zr, Address(xthread, JavaThread::vm_result_oop_offset()));
2510
2511 __ ld(x10, Address(xthread, Thread::pending_exception_offset()));
2512 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2513
2514 // -------------
2515 // make sure all code is generated
2516 masm->flush();
2517
2518 // return the blob
2519 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true);
2520 }
2521
2522 // Continuation point for throwing of implicit exceptions that are
2523 // not handled in the current activation. Fabricates an exception
2524 // oop and initiates normal exception dispatching in this
2525 // frame. Since we need to preserve callee-saved values (currently
2526 // only for C2, but done for C1 as well) we need a callee-saved oop
2527 // map and therefore have to make these stubs into RuntimeStubs
2528 // rather than BufferBlobs. If the compiler needs all registers to
2529 // be preserved between the fault point and the exception handler
2530 // then it must assume responsibility for that in
2531 // AbstractCompiler::continuation_for_implicit_null_exception or
2532 // continuation_for_implicit_division_by_zero_exception. All other
2533 // implicit exceptions (e.g., NullPointerException or
2534 // AbstractMethodError on entry) are either at call sites or
2535 // otherwise assume that stack unwinding will be initiated, so
2536 // caller saved registers were assumed volatile in the compiler.
2537
2538 RuntimeStub* SharedRuntime::generate_throw_exception(StubId id, address runtime_entry) {
2539 assert(is_throw_id(id), "expected a throw stub id");
2540
2541 const char* name = SharedRuntime::stub_name(id);
2542
2543 // Information about frame layout at time of blocking runtime call.
2544 // Note that we only have to preserve callee-saved registers since
2545 // the compilers are responsible for supplying a continuation point
2546 // if they expect all registers to be preserved.
2547 // n.b. riscv asserts that frame::arg_reg_save_area_bytes == 0
2548 assert_cond(runtime_entry != nullptr);
2549 enum layout {
2550 fp_off = 0,
2551 fp_off2,
2552 return_off,
2553 return_off2,
2554 framesize // inclusive of return address
2555 };
2556
2557 const int insts_size = 1024;
2558 const int locs_size = 64;
2559
2560 ResourceMark rm;
2561 const char* timer_msg = "SharedRuntime generate_throw_exception";
2562 TraceTime timer(timer_msg, TRACETIME_LOG(Info, startuptime));
2563
2564 CodeBuffer code(name, insts_size, locs_size);
2565 OopMapSet* oop_maps = new OopMapSet();
2566 MacroAssembler* masm = new MacroAssembler(&code);
2567 assert_cond(oop_maps != nullptr && masm != nullptr);
2568
2569 address start = __ pc();
2570
2571 // This is an inlined and slightly modified version of call_VM
2572 // which has the ability to fetch the return PC out of
2573 // thread-local storage and also sets up last_Java_sp slightly
2574 // differently than the real call_VM
2575
2576 __ enter(); // Save FP and RA before call
2577
2578 assert(is_even(framesize / 2), "sp not 16-byte aligned");
2579
2580 // ra and fp are already in place
2581 __ subi(sp, fp, (unsigned)framesize << LogBytesPerInt); // prolog
2582
2583 int frame_complete = __ pc() - start;
2584
2585 // Set up last_Java_sp and last_Java_fp
2586 address the_pc = __ pc();
2587 __ set_last_Java_frame(sp, fp, the_pc, t0);
2588
2589 // Call runtime
2590 __ mv(c_rarg0, xthread);
2591 BLOCK_COMMENT("call runtime_entry");
2592 __ rt_call(runtime_entry);
2593
2594 // Generate oop map
2595 OopMap* map = new OopMap(framesize, 0);
2596 assert_cond(map != nullptr);
2597
2598 oop_maps->add_gc_map(the_pc - start, map);
2599
2600 __ reset_last_Java_frame(true);
2601
2602 __ leave();
2603
2604 // check for pending exceptions
2605 #ifdef ASSERT
2606 Label L;
2607 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2608 __ bnez(t0, L);
2609 __ should_not_reach_here();
2610 __ bind(L);
2611 #endif // ASSERT
2612 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2613
2614 // codeBlob framesize is in words (not VMRegImpl::slot_size)
2615 RuntimeStub* stub =
2616 RuntimeStub::new_runtime_stub(name,
2617 &code,
2618 frame_complete,
2619 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2620 oop_maps, false);
2621 assert(stub != nullptr, "create runtime stub fail!");
2622 return stub;
2623 }
2624
2625 #if INCLUDE_JFR
2626
2627 static void jfr_prologue(address the_pc, MacroAssembler* masm, Register thread) {
2628 __ set_last_Java_frame(sp, fp, the_pc, t0);
2629 __ mv(c_rarg0, thread);
2630 }
2631
2632 static void jfr_epilogue(MacroAssembler* masm) {
2633 __ reset_last_Java_frame(true);
2634 }
2635 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
2636 // It returns a jobject handle to the event writer.
2637 // The handle is dereferenced and the return value is the event writer oop.
2638 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
2639 enum layout {
2640 fp_off,
2641 fp_off2,
2642 return_off,
2643 return_off2,
2644 framesize // inclusive of return address
2645 };
2646
2647 int insts_size = 1024;
2648 int locs_size = 64;
2649 const char* name = SharedRuntime::stub_name(StubId::shared_jfr_write_checkpoint_id);
2650 CodeBuffer code(name, insts_size, locs_size);
2651 OopMapSet* oop_maps = new OopMapSet();
2652 MacroAssembler* masm = new MacroAssembler(&code);
2653
2654 address start = __ pc();
2655 __ enter();
2656 int frame_complete = __ pc() - start;
2657 address the_pc = __ pc();
2658 jfr_prologue(the_pc, masm, xthread);
2659 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), 1);
2660
2661 jfr_epilogue(masm);
2662 __ resolve_global_jobject(x10, t0, t1);
2663 __ leave();
2664 __ ret();
2665
2666 OopMap* map = new OopMap(framesize, 1);
2667 oop_maps->add_gc_map(the_pc - start, map);
2668
2669 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
2670 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
2671 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2672 oop_maps, false);
2673 return stub;
2674 }
2675
2676 // For c2: call to return a leased buffer.
2677 RuntimeStub* SharedRuntime::generate_jfr_return_lease() {
2678 enum layout {
2679 fp_off,
2680 fp_off2,
2681 return_off,
2682 return_off2,
2683 framesize // inclusive of return address
2684 };
2685
2686 int insts_size = 1024;
2687 int locs_size = 64;
2688 const char* name = SharedRuntime::stub_name(StubId::shared_jfr_return_lease_id);
2689 CodeBuffer code(name, insts_size, locs_size);
2690 OopMapSet* oop_maps = new OopMapSet();
2691 MacroAssembler* masm = new MacroAssembler(&code);
2692
2693 address start = __ pc();
2694 __ enter();
2695 int frame_complete = __ pc() - start;
2696 address the_pc = __ pc();
2697 jfr_prologue(the_pc, masm, xthread);
2698 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::return_lease), 1);
2699
2700 jfr_epilogue(masm);
2701 __ leave();
2702 __ ret();
2703
2704 OopMap* map = new OopMap(framesize, 1);
2705 oop_maps->add_gc_map(the_pc - start, map);
2706
2707 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
2708 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
2709 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2710 oop_maps, false);
2711 return stub;
2712 }
2713
2714 #endif // INCLUDE_JFR
2715
2716 const uint SharedRuntime::java_return_convention_max_int = Argument::n_int_register_parameters_j;
2717 const uint SharedRuntime::java_return_convention_max_float = Argument::n_float_register_parameters_j;
2718
2719 int SharedRuntime::java_return_convention(const BasicType *sig_bt, VMRegPair *regs, int total_args_passed) {
2720 Unimplemented();
2721 return 0;
2722 }
2723
2724 BufferedInlineTypeBlob* SharedRuntime::generate_buffered_inline_type_adapter(const InlineKlass* vk) {
2725 Unimplemented();
2726 return nullptr;
2727 }