1 /*
2 * Copyright (c) 2003, 2025, 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 #if INCLUDE_JVMCI
63 #include "jvmci/jvmciJavaClasses.hpp"
64 #endif
65
66 #define __ masm->
67
68 #ifdef PRODUCT
69 #define BLOCK_COMMENT(str) /* nothing */
70 #else
71 #define BLOCK_COMMENT(str) __ block_comment(str)
72 #endif
73
74 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size;
75
76 class RegisterSaver {
77 const bool _save_vectors;
78 public:
79 RegisterSaver(bool save_vectors) : _save_vectors(UseRVV && save_vectors) {}
80 ~RegisterSaver() {}
81 OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words);
82 void restore_live_registers(MacroAssembler* masm);
83
84 // Offsets into the register save area
85 // Used by deoptimization when it is managing result register
86 // values on its own
87 // gregs:28, float_register:32; except: x1(ra) & x2(sp) & gp(x3) & tp(x4)
88 // |---v0---|<---SP
89 // |---v1---|save vectors only in generate_handler_blob
90 // |-- .. --|
91 // |---v31--|-----
92 // |---f0---|
93 // |---f1---|
94 // | .. |
95 // |---f31--|
96 // |---reserved slot for stack alignment---|
97 // |---x5---|
98 // | x6 |
99 // |---.. --|
100 // |---x31--|
101 // |---fp---|
102 // |---ra---|
103 int v0_offset_in_bytes(void) { return 0; }
104 int f0_offset_in_bytes(void) {
105 int f0_offset = 0;
106 #ifdef COMPILER2
107 if (_save_vectors) {
108 f0_offset += Matcher::scalable_vector_reg_size(T_INT) * VectorRegister::number_of_registers *
109 BytesPerInt;
110 }
111 #endif
112 return f0_offset;
113 }
114 int reserved_slot_offset_in_bytes(void) {
115 return f0_offset_in_bytes() +
116 FloatRegister::max_slots_per_register *
117 FloatRegister::number_of_registers *
118 BytesPerInt;
119 }
120
121 int reg_offset_in_bytes(Register r) {
122 assert (r->encoding() > 4, "ra, sp, gp and tp not saved");
123 return reserved_slot_offset_in_bytes() + (r->encoding() - 4 /* x1, x2, x3, x4 */) * wordSize;
124 }
125
126 int freg_offset_in_bytes(FloatRegister f) {
127 return f0_offset_in_bytes() + f->encoding() * wordSize;
128 }
129
130 int ra_offset_in_bytes(void) {
131 return reserved_slot_offset_in_bytes() +
132 (Register::number_of_registers - 3) *
133 Register::max_slots_per_register *
134 BytesPerInt;
135 }
136 };
137
138 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) {
139 int vector_size_in_bytes = 0;
140 int vector_size_in_slots = 0;
141 #ifdef COMPILER2
142 if (_save_vectors) {
143 vector_size_in_bytes += Matcher::scalable_vector_reg_size(T_BYTE);
144 vector_size_in_slots += Matcher::scalable_vector_reg_size(T_INT);
145 }
146 #endif
147
148 int frame_size_in_bytes = align_up(additional_frame_words * wordSize + ra_offset_in_bytes() + wordSize, 16);
149 // OopMap frame size is in compiler stack slots (jint's) not bytes or words
150 int frame_size_in_slots = frame_size_in_bytes / BytesPerInt;
151 // The caller will allocate additional_frame_words
152 int additional_frame_slots = additional_frame_words * wordSize / BytesPerInt;
153 // CodeBlob frame size is in words.
154 int frame_size_in_words = frame_size_in_bytes / wordSize;
155 *total_frame_words = frame_size_in_words;
156
157 // Save Integer, Float and Vector registers.
158 __ enter();
159 __ push_CPU_state(_save_vectors, vector_size_in_bytes);
160
161 // Set an oopmap for the call site. This oopmap will map all
162 // oop-registers and debug-info registers as callee-saved. This
163 // will allow deoptimization at this safepoint to find all possible
164 // debug-info recordings, as well as let GC find all oops.
165
166 OopMapSet *oop_maps = new OopMapSet();
167 OopMap* oop_map = new OopMap(frame_size_in_slots, 0);
168 assert_cond(oop_maps != nullptr && oop_map != nullptr);
169
170 int sp_offset_in_slots = 0;
171 int step_in_slots = 0;
172 if (_save_vectors) {
173 step_in_slots = vector_size_in_slots;
174 for (int i = 0; i < VectorRegister::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
175 VectorRegister r = as_VectorRegister(i);
176 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots), r->as_VMReg());
177 }
178 }
179
180 step_in_slots = FloatRegister::max_slots_per_register;
181 for (int i = 0; i < FloatRegister::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
182 FloatRegister r = as_FloatRegister(i);
183 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots), r->as_VMReg());
184 }
185
186 step_in_slots = Register::max_slots_per_register;
187 // skip the slot reserved for alignment, see MacroAssembler::push_reg;
188 // also skip x5 ~ x6 on the stack because they are caller-saved registers.
189 sp_offset_in_slots += Register::max_slots_per_register * 3;
190 // besides, we ignore x0 ~ x4 because push_CPU_state won't push them on the stack.
191 for (int i = 7; i < Register::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
192 Register r = as_Register(i);
193 if (r != xthread) {
194 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots + additional_frame_slots), r->as_VMReg());
195 }
196 }
197
198 return oop_map;
199 }
200
201 void RegisterSaver::restore_live_registers(MacroAssembler* masm) {
202 #ifdef COMPILER2
203 __ pop_CPU_state(_save_vectors, Matcher::scalable_vector_reg_size(T_BYTE));
204 #else
205 #if !INCLUDE_JVMCI
206 assert(!_save_vectors, "vectors are generated only by C2 and JVMCI");
207 #endif
208 __ pop_CPU_state(_save_vectors);
209 #endif
210 __ leave();
211 }
212
213 // Is vector's size (in bytes) bigger than a size saved by default?
214 // riscv does not ovlerlay the floating-point registers on vector registers like aarch64.
215 bool SharedRuntime::is_wide_vector(int size) {
216 return UseRVV;
217 }
218
219 // ---------------------------------------------------------------------------
220 // Read the array of BasicTypes from a signature, and compute where the
221 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
222 // quantities. Values less than VMRegImpl::stack0 are registers, those above
223 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
224 // as framesizes are fixed.
225 // VMRegImpl::stack0 refers to the first slot 0(sp).
226 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher.
227 // Register up to Register::number_of_registers) are the 64-bit
228 // integer registers.
229
230 // Note: the INPUTS in sig_bt are in units of Java argument words,
231 // which are 64-bit. The OUTPUTS are in 32-bit units.
232
233 // The Java calling convention is a "shifted" version of the C ABI.
234 // By skipping the first C ABI register we can call non-static jni
235 // methods with small numbers of arguments without having to shuffle
236 // the arguments at all. Since we control the java ABI we ought to at
237 // least get some advantage out of it.
238
239 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
240 VMRegPair *regs,
241 int total_args_passed) {
242 // Create the mapping between argument positions and
243 // registers.
244 static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = {
245 j_rarg0, j_rarg1, j_rarg2, j_rarg3,
246 j_rarg4, j_rarg5, j_rarg6, j_rarg7
247 };
248 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_j] = {
249 j_farg0, j_farg1, j_farg2, j_farg3,
250 j_farg4, j_farg5, j_farg6, j_farg7
251 };
252
253 uint int_args = 0;
254 uint fp_args = 0;
255 uint stk_args = 0;
256
257 for (int i = 0; i < total_args_passed; i++) {
258 switch (sig_bt[i]) {
259 case T_BOOLEAN: // fall through
260 case T_CHAR: // fall through
261 case T_BYTE: // fall through
262 case T_SHORT: // fall through
263 case T_INT:
264 if (int_args < Argument::n_int_register_parameters_j) {
265 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
266 } else {
267 stk_args = align_up(stk_args, 2);
268 regs[i].set1(VMRegImpl::stack2reg(stk_args));
269 stk_args += 1;
270 }
271 break;
272 case T_VOID:
273 // halves of T_LONG or T_DOUBLE
274 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
275 regs[i].set_bad();
276 break;
277 case T_LONG: // fall through
278 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
279 case T_OBJECT: // fall through
280 case T_ARRAY: // fall through
281 case T_ADDRESS:
282 if (int_args < Argument::n_int_register_parameters_j) {
283 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
284 } else {
285 stk_args = align_up(stk_args, 2);
286 regs[i].set2(VMRegImpl::stack2reg(stk_args));
287 stk_args += 2;
288 }
289 break;
290 case T_FLOAT:
291 if (fp_args < Argument::n_float_register_parameters_j) {
292 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
293 } else {
294 stk_args = align_up(stk_args, 2);
295 regs[i].set1(VMRegImpl::stack2reg(stk_args));
296 stk_args += 1;
297 }
298 break;
299 case T_DOUBLE:
300 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
301 if (fp_args < Argument::n_float_register_parameters_j) {
302 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
303 } else {
304 stk_args = align_up(stk_args, 2);
305 regs[i].set2(VMRegImpl::stack2reg(stk_args));
306 stk_args += 2;
307 }
308 break;
309 default:
310 ShouldNotReachHere();
311 }
312 }
313
314 return stk_args;
315 }
316
317 // Patch the callers callsite with entry to compiled code if it exists.
318 static void patch_callers_callsite(MacroAssembler *masm) {
319 Label L;
320 __ ld(t0, Address(xmethod, in_bytes(Method::code_offset())));
321 __ beqz(t0, L);
322
323 __ enter();
324 __ push_CPU_state();
325
326 // VM needs caller's callsite
327 // VM needs target method
328 // This needs to be a long call since we will relocate this adapter to
329 // the codeBuffer and it may not reach
330
331 #ifndef PRODUCT
332 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
333 #endif
334
335 __ mv(c_rarg0, xmethod);
336 __ mv(c_rarg1, ra);
337 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite));
338
339 __ pop_CPU_state();
340 // restore sp
341 __ leave();
342 __ bind(L);
343 }
344
345 static void gen_c2i_adapter(MacroAssembler *masm,
346 int total_args_passed,
347 int comp_args_on_stack,
348 const BasicType *sig_bt,
349 const VMRegPair *regs,
350 Label& skip_fixup) {
351 // Before we get into the guts of the C2I adapter, see if we should be here
352 // at all. We've come from compiled code and are attempting to jump to the
353 // interpreter, which means the caller made a static call to get here
354 // (vcalls always get a compiled target if there is one). Check for a
355 // compiled target. If there is one, we need to patch the caller's call.
356 patch_callers_callsite(masm);
357
358 __ bind(skip_fixup);
359
360 int words_pushed = 0;
361
362 // Since all args are passed on the stack, total_args_passed *
363 // Interpreter::stackElementSize is the space we need.
364
365 int extraspace = total_args_passed * Interpreter::stackElementSize;
366
367 __ mv(x19_sender_sp, sp);
368
369 // stack is aligned, keep it that way
370 extraspace = align_up(extraspace, 2 * wordSize);
371
372 if (extraspace) {
373 __ sub(sp, sp, extraspace);
374 }
375
376 // Now write the args into the outgoing interpreter space
377 for (int i = 0; i < total_args_passed; i++) {
378 if (sig_bt[i] == T_VOID) {
379 assert(i > 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "missing half");
380 continue;
381 }
382
383 // offset to start parameters
384 int st_off = (total_args_passed - i - 1) * Interpreter::stackElementSize;
385 int next_off = st_off - Interpreter::stackElementSize;
386
387 // Say 4 args:
388 // i st_off
389 // 0 32 T_LONG
390 // 1 24 T_VOID
391 // 2 16 T_OBJECT
392 // 3 8 T_BOOL
393 // - 0 return address
394 //
395 // However to make thing extra confusing. Because we can fit a Java long/double in
396 // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter
397 // leaves one slot empty and only stores to a single slot. In this case the
398 // slot that is occupied is the T_VOID slot. See I said it was confusing.
399
400 VMReg r_1 = regs[i].first();
401 VMReg r_2 = regs[i].second();
402 if (!r_1->is_valid()) {
403 assert(!r_2->is_valid(), "");
404 continue;
405 }
406 if (r_1->is_stack()) {
407 // memory to memory use t0
408 int ld_off = (r_1->reg2stack() * VMRegImpl::stack_slot_size
409 + extraspace
410 + words_pushed * wordSize);
411 if (!r_2->is_valid()) {
412 __ lwu(t0, Address(sp, ld_off));
413 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
414 } else {
415 __ ld(t0, Address(sp, ld_off), /*temp register*/esp);
416
417 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
418 // T_DOUBLE and T_LONG use two slots in the interpreter
419 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
420 // ld_off == LSW, ld_off+wordSize == MSW
421 // st_off == MSW, next_off == LSW
422 __ sd(t0, Address(sp, next_off), /*temp register*/esp);
423 #ifdef ASSERT
424 // Overwrite the unused slot with known junk
425 __ mv(t0, 0xdeadffffdeadaaaaul);
426 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
427 #endif /* ASSERT */
428 } else {
429 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
430 }
431 }
432 } else if (r_1->is_Register()) {
433 Register r = r_1->as_Register();
434 if (!r_2->is_valid()) {
435 // must be only an int (or less ) so move only 32bits to slot
436 __ sd(r, Address(sp, st_off));
437 } else {
438 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
439 // T_DOUBLE and T_LONG use two slots in the interpreter
440 if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
441 // long/double in gpr
442 #ifdef ASSERT
443 // Overwrite the unused slot with known junk
444 __ mv(t0, 0xdeadffffdeadaaabul);
445 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
446 #endif /* ASSERT */
447 __ sd(r, Address(sp, next_off));
448 } else {
449 __ sd(r, Address(sp, st_off));
450 }
451 }
452 } else {
453 assert(r_1->is_FloatRegister(), "");
454 if (!r_2->is_valid()) {
455 // only a float use just part of the slot
456 __ fsw(r_1->as_FloatRegister(), Address(sp, st_off));
457 } else {
458 #ifdef ASSERT
459 // Overwrite the unused slot with known junk
460 __ mv(t0, 0xdeadffffdeadaaacul);
461 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
462 #endif /* ASSERT */
463 __ fsd(r_1->as_FloatRegister(), Address(sp, next_off));
464 }
465 }
466 }
467
468 __ mv(esp, sp); // Interp expects args on caller's expression stack
469
470 __ ld(t1, Address(xmethod, in_bytes(Method::interpreter_entry_offset())));
471 __ jr(t1);
472 }
473
474 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
475 int total_args_passed,
476 int comp_args_on_stack,
477 const BasicType *sig_bt,
478 const VMRegPair *regs) {
479 // Note: x19_sender_sp contains the senderSP on entry. We must
480 // preserve it since we may do a i2c -> c2i transition if we lose a
481 // race where compiled code goes non-entrant while we get args
482 // ready.
483
484 // Cut-out for having no stack args.
485 int comp_words_on_stack = align_up(comp_args_on_stack * VMRegImpl::stack_slot_size, wordSize) >> LogBytesPerWord;
486 if (comp_args_on_stack != 0) {
487 __ sub(t0, sp, comp_words_on_stack * wordSize);
488 __ andi(sp, t0, -16);
489 }
490
491 // Will jump to the compiled code just as if compiled code was doing it.
492 // Pre-load the register-jump target early, to schedule it better.
493 __ ld(t1, Address(xmethod, in_bytes(Method::from_compiled_offset())));
494
495 #if INCLUDE_JVMCI
496 if (EnableJVMCI) {
497 // check if this call should be routed towards a specific entry point
498 __ ld(t0, Address(xthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
499 Label no_alternative_target;
500 __ beqz(t0, no_alternative_target);
501 __ mv(t1, t0);
502 __ sd(zr, Address(xthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
503 __ bind(no_alternative_target);
504 }
505 #endif // INCLUDE_JVMCI
506
507 // Now generate the shuffle code.
508 for (int i = 0; i < total_args_passed; i++) {
509 if (sig_bt[i] == T_VOID) {
510 assert(i > 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "missing half");
511 continue;
512 }
513
514 // Pick up 0, 1 or 2 words from SP+offset.
515
516 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
517 "scrambled load targets?");
518 // Load in argument order going down.
519 int ld_off = (total_args_passed - i - 1) * Interpreter::stackElementSize;
520 // Point to interpreter value (vs. tag)
521 int next_off = ld_off - Interpreter::stackElementSize;
522
523 VMReg r_1 = regs[i].first();
524 VMReg r_2 = regs[i].second();
525 if (!r_1->is_valid()) {
526 assert(!r_2->is_valid(), "");
527 continue;
528 }
529 if (r_1->is_stack()) {
530 // Convert stack slot to an SP offset (+ wordSize to account for return address )
531 int st_off = regs[i].first()->reg2stack() * VMRegImpl::stack_slot_size;
532 if (!r_2->is_valid()) {
533 __ lw(t0, Address(esp, ld_off));
534 __ sd(t0, Address(sp, st_off), /*temp register*/t2);
535 } else {
536 //
537 // We are using two optoregs. This can be either T_OBJECT,
538 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
539 // two slots but only uses one for thr T_LONG or T_DOUBLE case
540 // So we must adjust where to pick up the data to match the
541 // interpreter.
542 //
543 // Interpreter local[n] == MSW, local[n+1] == LSW however locals
544 // are accessed as negative so LSW is at LOW address
545
546 // ld_off is MSW so get LSW
547 const int offset = (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) ?
548 next_off : ld_off;
549 __ ld(t0, Address(esp, offset));
550 // st_off is LSW (i.e. reg.first())
551 __ sd(t0, Address(sp, st_off), /*temp register*/t2);
552 }
553 } else if (r_1->is_Register()) { // Register argument
554 Register r = r_1->as_Register();
555 if (r_2->is_valid()) {
556 //
557 // We are using two VMRegs. This can be either T_OBJECT,
558 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
559 // two slots but only uses one for thr T_LONG or T_DOUBLE case
560 // So we must adjust where to pick up the data to match the
561 // interpreter.
562
563 const int offset = (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) ?
564 next_off : ld_off;
565
566 // this can be a misaligned move
567 __ ld(r, Address(esp, offset));
568 } else {
569 // sign extend and use a full word?
570 __ lw(r, Address(esp, ld_off));
571 }
572 } else {
573 if (!r_2->is_valid()) {
574 __ flw(r_1->as_FloatRegister(), Address(esp, ld_off));
575 } else {
576 __ fld(r_1->as_FloatRegister(), Address(esp, next_off));
577 }
578 }
579 }
580
581 __ push_cont_fastpath(xthread); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about
582
583 // 6243940 We might end up in handle_wrong_method if
584 // the callee is deoptimized as we race thru here. If that
585 // happens we don't want to take a safepoint because the
586 // caller frame will look interpreted and arguments are now
587 // "compiled" so it is much better to make this transition
588 // invisible to the stack walking code. Unfortunately if
589 // we try and find the callee by normal means a safepoint
590 // is possible. So we stash the desired callee in the thread
591 // and the vm will find there should this case occur.
592
593 __ sd(xmethod, Address(xthread, JavaThread::callee_target_offset()));
594
595 __ jr(t1);
596 }
597
598 // ---------------------------------------------------------------
599 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
600 int total_args_passed,
601 int comp_args_on_stack,
602 const BasicType *sig_bt,
603 const VMRegPair *regs,
604 AdapterFingerPrint* fingerprint) {
605 address i2c_entry = __ pc();
606 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
607
608 address c2i_unverified_entry = __ pc();
609 Label skip_fixup;
610
611 const Register receiver = j_rarg0;
612 const Register data = t0;
613
614 // -------------------------------------------------------------------------
615 // Generate a C2I adapter. On entry we know xmethod holds the Method* during calls
616 // to the interpreter. The args start out packed in the compiled layout. They
617 // need to be unpacked into the interpreter layout. This will almost always
618 // require some stack space. We grow the current (compiled) stack, then repack
619 // the args. We finally end in a jump to the generic interpreter entry point.
620 // On exit from the interpreter, the interpreter will restore our SP (lest the
621 // compiled code, which relies solely on SP and not FP, get sick).
622
623 {
624 __ block_comment("c2i_unverified_entry {");
625
626 __ ic_check();
627 __ ld(xmethod, Address(data, CompiledICData::speculated_method_offset()));
628
629 __ ld(t0, Address(xmethod, in_bytes(Method::code_offset())));
630 __ beqz(t0, skip_fixup);
631 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
632 __ block_comment("} c2i_unverified_entry");
633 }
634
635 address c2i_entry = __ pc();
636
637 // Class initialization barrier for static methods
638 address c2i_no_clinit_check_entry = nullptr;
639 if (VM_Version::supports_fast_class_init_checks()) {
640 Label L_skip_barrier;
641
642 { // Bypass the barrier for non-static methods
643 __ load_unsigned_short(t0, Address(xmethod, Method::access_flags_offset()));
644 __ test_bit(t1, t0, exact_log2(JVM_ACC_STATIC));
645 __ beqz(t1, L_skip_barrier); // non-static
646 }
647
648 __ load_method_holder(t1, xmethod);
649 __ clinit_barrier(t1, t0, &L_skip_barrier);
650 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
651
652 __ bind(L_skip_barrier);
653 c2i_no_clinit_check_entry = __ pc();
654 }
655
656 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
657 bs->c2i_entry_barrier(masm);
658
659 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup);
660
661 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry, c2i_no_clinit_check_entry);
662 }
663
664 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
665 uint num_bits,
666 uint total_args_passed) {
667 assert(total_args_passed <= Argument::n_vector_register_parameters_c, "unsupported");
668 assert(num_bits >= 64 && num_bits <= 2048 && is_power_of_2(num_bits), "unsupported");
669
670 // check more info at https://github.com/riscv-non-isa/riscv-elf-psabi-doc/blob/master/riscv-cc.adoc
671 static const VectorRegister VEC_ArgReg[Argument::n_vector_register_parameters_c] = {
672 v8, v9, v10, v11, v12, v13, v14, v15,
673 v16, v17, v18, v19, v20, v21, v22, v23
674 };
675
676 const int next_reg_val = 3;
677 for (uint i = 0; i < total_args_passed; i++) {
678 VMReg vmreg = VEC_ArgReg[i]->as_VMReg();
679 regs[i].set_pair(vmreg->next(next_reg_val), vmreg);
680 }
681 return 0;
682 }
683
684 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
685 VMRegPair *regs,
686 int total_args_passed) {
687
688 // We return the amount of VMRegImpl stack slots we need to reserve for all
689 // the arguments NOT counting out_preserve_stack_slots.
690
691 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
692 c_rarg0, c_rarg1, c_rarg2, c_rarg3,
693 c_rarg4, c_rarg5, c_rarg6, c_rarg7
694 };
695 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
696 c_farg0, c_farg1, c_farg2, c_farg3,
697 c_farg4, c_farg5, c_farg6, c_farg7
698 };
699
700 uint int_args = 0;
701 uint fp_args = 0;
702 uint stk_args = 0; // inc by 2 each time
703
704 for (int i = 0; i < total_args_passed; i++) {
705 switch (sig_bt[i]) {
706 case T_BOOLEAN: // fall through
707 case T_CHAR: // fall through
708 case T_BYTE: // fall through
709 case T_SHORT: // fall through
710 case T_INT:
711 if (int_args < Argument::n_int_register_parameters_c) {
712 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
713 } else {
714 regs[i].set1(VMRegImpl::stack2reg(stk_args));
715 stk_args += 2;
716 }
717 break;
718 case T_LONG: // fall through
719 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
720 case T_OBJECT: // fall through
721 case T_ARRAY: // fall through
722 case T_ADDRESS: // fall through
723 case T_METADATA:
724 if (int_args < Argument::n_int_register_parameters_c) {
725 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
726 } else {
727 regs[i].set2(VMRegImpl::stack2reg(stk_args));
728 stk_args += 2;
729 }
730 break;
731 case T_FLOAT:
732 if (fp_args < Argument::n_float_register_parameters_c) {
733 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
734 } else if (int_args < Argument::n_int_register_parameters_c) {
735 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
736 } else {
737 regs[i].set1(VMRegImpl::stack2reg(stk_args));
738 stk_args += 2;
739 }
740 break;
741 case T_DOUBLE:
742 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
743 if (fp_args < Argument::n_float_register_parameters_c) {
744 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
745 } else if (int_args < Argument::n_int_register_parameters_c) {
746 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
747 } else {
748 regs[i].set2(VMRegImpl::stack2reg(stk_args));
749 stk_args += 2;
750 }
751 break;
752 case T_VOID: // Halves of longs and doubles
753 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
754 regs[i].set_bad();
755 break;
756 default:
757 ShouldNotReachHere();
758 }
759 }
760
761 return stk_args;
762 }
763
764 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
765 // We always ignore the frame_slots arg and just use the space just below frame pointer
766 // which by this time is free to use
767 switch (ret_type) {
768 case T_FLOAT:
769 __ fsw(f10, Address(fp, -3 * wordSize));
770 break;
771 case T_DOUBLE:
772 __ fsd(f10, Address(fp, -3 * wordSize));
773 break;
774 case T_VOID: break;
775 default: {
776 __ sd(x10, Address(fp, -3 * wordSize));
777 }
778 }
779 }
780
781 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
782 // We always ignore the frame_slots arg and just use the space just below frame pointer
783 // which by this time is free to use
784 switch (ret_type) {
785 case T_FLOAT:
786 __ flw(f10, Address(fp, -3 * wordSize));
787 break;
788 case T_DOUBLE:
789 __ fld(f10, Address(fp, -3 * wordSize));
790 break;
791 case T_VOID: break;
792 default: {
793 __ ld(x10, Address(fp, -3 * wordSize));
794 }
795 }
796 }
797
798 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
799 RegSet x;
800 for ( int i = first_arg ; i < arg_count ; i++ ) {
801 if (args[i].first()->is_Register()) {
802 x = x + args[i].first()->as_Register();
803 } else if (args[i].first()->is_FloatRegister()) {
804 __ subi(sp, sp, 2 * wordSize);
805 __ fsd(args[i].first()->as_FloatRegister(), Address(sp, 0));
806 }
807 }
808 __ push_reg(x, sp);
809 }
810
811 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
812 RegSet x;
813 for ( int i = first_arg ; i < arg_count ; i++ ) {
814 if (args[i].first()->is_Register()) {
815 x = x + args[i].first()->as_Register();
816 } else {
817 ;
818 }
819 }
820 __ pop_reg(x, sp);
821 for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) {
822 if (args[i].first()->is_Register()) {
823 ;
824 } else if (args[i].first()->is_FloatRegister()) {
825 __ fld(args[i].first()->as_FloatRegister(), Address(sp, 0));
826 __ addi(sp, sp, 2 * wordSize);
827 }
828 }
829 }
830
831 static void verify_oop_args(MacroAssembler* masm,
832 const methodHandle& method,
833 const BasicType* sig_bt,
834 const VMRegPair* regs) {
835 const Register temp_reg = x9; // not part of any compiled calling seq
836 if (VerifyOops) {
837 for (int i = 0; i < method->size_of_parameters(); i++) {
838 if (sig_bt[i] == T_OBJECT ||
839 sig_bt[i] == T_ARRAY) {
840 VMReg r = regs[i].first();
841 assert(r->is_valid(), "bad oop arg");
842 if (r->is_stack()) {
843 __ ld(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
844 __ verify_oop(temp_reg);
845 } else {
846 __ verify_oop(r->as_Register());
847 }
848 }
849 }
850 }
851 }
852
853 // on exit, sp points to the ContinuationEntry
854 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& stack_slots) {
855 assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
856 assert(in_bytes(ContinuationEntry::cont_offset()) % VMRegImpl::stack_slot_size == 0, "");
857 assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
858
859 stack_slots += (int)ContinuationEntry::size() / wordSize;
860 __ sub(sp, sp, (int)ContinuationEntry::size()); // place Continuation metadata
861
862 OopMap* map = new OopMap(((int)ContinuationEntry::size() + wordSize) / VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
863
864 __ ld(t0, Address(xthread, JavaThread::cont_entry_offset()));
865 __ sd(t0, Address(sp, ContinuationEntry::parent_offset()));
866 __ sd(sp, Address(xthread, JavaThread::cont_entry_offset()));
867
868 return map;
869 }
870
871 // on entry c_rarg1 points to the continuation
872 // sp points to ContinuationEntry
873 // c_rarg3 -- isVirtualThread
874 static void fill_continuation_entry(MacroAssembler* masm) {
875 #ifdef ASSERT
876 __ mv(t0, ContinuationEntry::cookie_value());
877 __ sw(t0, Address(sp, ContinuationEntry::cookie_offset()));
878 #endif
879
880 __ sd(c_rarg1, Address(sp, ContinuationEntry::cont_offset()));
881 __ sw(c_rarg3, Address(sp, ContinuationEntry::flags_offset()));
882 __ sd(zr, Address(sp, ContinuationEntry::chunk_offset()));
883 __ sw(zr, Address(sp, ContinuationEntry::argsize_offset()));
884 __ sw(zr, Address(sp, ContinuationEntry::pin_count_offset()));
885
886 __ ld(t0, Address(xthread, JavaThread::cont_fastpath_offset()));
887 __ sd(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
888 __ ld(t0, Address(xthread, JavaThread::held_monitor_count_offset()));
889 __ sd(t0, Address(sp, ContinuationEntry::parent_held_monitor_count_offset()));
890
891 __ sd(zr, Address(xthread, JavaThread::cont_fastpath_offset()));
892 __ sd(zr, Address(xthread, JavaThread::held_monitor_count_offset()));
893 }
894
895 // on entry, sp points to the ContinuationEntry
896 // on exit, fp points to the spilled fp + 2 * wordSize in the entry frame
897 static void continuation_enter_cleanup(MacroAssembler* masm) {
898 #ifndef PRODUCT
899 Label OK;
900 __ ld(t0, Address(xthread, JavaThread::cont_entry_offset()));
901 __ beq(sp, t0, OK);
902 __ stop("incorrect sp");
903 __ bind(OK);
904 #endif
905
906 __ ld(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
907 __ sd(t0, Address(xthread, JavaThread::cont_fastpath_offset()));
908
909 if (CheckJNICalls) {
910 // Check if this is a virtual thread continuation
911 Label L_skip_vthread_code;
912 __ lwu(t0, Address(sp, ContinuationEntry::flags_offset()));
913 __ beqz(t0, L_skip_vthread_code);
914
915 // If the held monitor count is > 0 and this vthread is terminating then
916 // it failed to release a JNI monitor. So we issue the same log message
917 // that JavaThread::exit does.
918 __ ld(t0, Address(xthread, JavaThread::jni_monitor_count_offset()));
919 __ beqz(t0, L_skip_vthread_code);
920
921 // Save return value potentially containing the exception oop in callee-saved x9
922 __ mv(x9, x10);
923 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::log_jni_monitor_still_held));
924 // Restore potential return value
925 __ mv(x10, x9);
926
927 // For vthreads we have to explicitly zero the JNI monitor count of the carrier
928 // on termination. The held count is implicitly zeroed below when we restore from
929 // the parent held count (which has to be zero).
930 __ sd(zr, Address(xthread, JavaThread::jni_monitor_count_offset()));
931
932 __ bind(L_skip_vthread_code);
933 }
934 #ifdef ASSERT
935 else {
936 // Check if this is a virtual thread continuation
937 Label L_skip_vthread_code;
938 __ lwu(t0, Address(sp, ContinuationEntry::flags_offset()));
939 __ beqz(t0, L_skip_vthread_code);
940
941 // See comment just above. If not checking JNI calls the JNI count is only
942 // needed for assertion checking.
943 __ sd(zr, Address(xthread, JavaThread::jni_monitor_count_offset()));
944
945 __ bind(L_skip_vthread_code);
946 }
947 #endif
948
949 __ ld(t0, Address(sp, ContinuationEntry::parent_held_monitor_count_offset()));
950 __ sd(t0, Address(xthread, JavaThread::held_monitor_count_offset()));
951
952 __ ld(t0, Address(sp, ContinuationEntry::parent_offset()));
953 __ sd(t0, Address(xthread, JavaThread::cont_entry_offset()));
954 __ add(fp, sp, (int)ContinuationEntry::size() + 2 * wordSize /* 2 extra words to match up with leave() */);
955 }
956
957 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
958 // On entry: c_rarg1 -- the continuation object
959 // c_rarg2 -- isContinue
960 // c_rarg3 -- isVirtualThread
961 static void gen_continuation_enter(MacroAssembler* masm,
962 const methodHandle& method,
963 const BasicType* sig_bt,
964 const VMRegPair* regs,
965 int& exception_offset,
966 OopMapSet*oop_maps,
967 int& frame_complete,
968 int& stack_slots,
969 int& interpreted_entry_offset,
970 int& compiled_entry_offset) {
971 // verify_oop_args(masm, method, sig_bt, regs);
972 Address resolve(SharedRuntime::get_resolve_static_call_stub(), relocInfo::static_call_type);
973
974 address start = __ pc();
975
976 Label call_thaw, exit;
977
978 // i2i entry used at interp_only_mode only
979 interpreted_entry_offset = __ pc() - start;
980 {
981 #ifdef ASSERT
982 Label is_interp_only;
983 __ lw(t0, Address(xthread, JavaThread::interp_only_mode_offset()));
984 __ bnez(t0, is_interp_only);
985 __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
986 __ bind(is_interp_only);
987 #endif
988
989 // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
990 __ ld(c_rarg1, Address(esp, Interpreter::stackElementSize * 2));
991 __ ld(c_rarg2, Address(esp, Interpreter::stackElementSize * 1));
992 __ ld(c_rarg3, Address(esp, Interpreter::stackElementSize * 0));
993 __ push_cont_fastpath(xthread);
994
995 __ enter();
996 stack_slots = 2; // will be adjusted in setup
997 OopMap* map = continuation_enter_setup(masm, stack_slots);
998 // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
999 // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
1000
1001 fill_continuation_entry(masm);
1002
1003 __ bnez(c_rarg2, call_thaw);
1004
1005 // Make sure the call is patchable
1006 __ align(NativeInstruction::instruction_size);
1007
1008 const address tr_call = __ reloc_call(resolve);
1009 if (tr_call == nullptr) {
1010 fatal("CodeCache is full at gen_continuation_enter");
1011 }
1012
1013 oop_maps->add_gc_map(__ pc() - start, map);
1014 __ post_call_nop();
1015
1016 __ j(exit);
1017
1018 address stub = CompiledDirectCall::emit_to_interp_stub(masm, tr_call);
1019 if (stub == nullptr) {
1020 fatal("CodeCache is full at gen_continuation_enter");
1021 }
1022 }
1023
1024 // compiled entry
1025 __ align(CodeEntryAlignment);
1026 compiled_entry_offset = __ pc() - start;
1027
1028 __ enter();
1029 stack_slots = 2; // will be adjusted in setup
1030 OopMap* map = continuation_enter_setup(masm, stack_slots);
1031 frame_complete = __ pc() - start;
1032
1033 fill_continuation_entry(masm);
1034
1035 __ bnez(c_rarg2, call_thaw);
1036
1037 // Make sure the call is patchable
1038 __ align(NativeInstruction::instruction_size);
1039
1040 const address tr_call = __ reloc_call(resolve);
1041 if (tr_call == nullptr) {
1042 fatal("CodeCache is full at gen_continuation_enter");
1043 }
1044
1045 oop_maps->add_gc_map(__ pc() - start, map);
1046 __ post_call_nop();
1047
1048 __ j(exit);
1049
1050 __ bind(call_thaw);
1051
1052 ContinuationEntry::_thaw_call_pc_offset = __ pc() - start;
1053 __ rt_call(CAST_FROM_FN_PTR(address, StubRoutines::cont_thaw()));
1054 oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1055 ContinuationEntry::_return_pc_offset = __ pc() - start;
1056 __ post_call_nop();
1057
1058 __ bind(exit);
1059 ContinuationEntry::_cleanup_offset = __ pc() - start;
1060 continuation_enter_cleanup(masm);
1061 __ leave();
1062 __ ret();
1063
1064 // exception handling
1065 exception_offset = __ pc() - start;
1066 {
1067 __ mv(x9, x10); // save return value contaning the exception oop in callee-saved x9
1068
1069 continuation_enter_cleanup(masm);
1070
1071 __ ld(c_rarg1, Address(fp, -1 * wordSize)); // return address
1072 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), xthread, c_rarg1);
1073
1074 // see OptoRuntime::generate_exception_blob: x10 -- exception oop, x13 -- exception pc
1075
1076 __ mv(x11, x10); // the exception handler
1077 __ mv(x10, x9); // restore return value contaning the exception oop
1078 __ verify_oop(x10);
1079
1080 __ leave();
1081 __ mv(x13, ra);
1082 __ jr(x11); // the exception handler
1083 }
1084
1085 address stub = CompiledDirectCall::emit_to_interp_stub(masm, tr_call);
1086 if (stub == nullptr) {
1087 fatal("CodeCache is full at gen_continuation_enter");
1088 }
1089 }
1090
1091 static void gen_continuation_yield(MacroAssembler* masm,
1092 const methodHandle& method,
1093 const BasicType* sig_bt,
1094 const VMRegPair* regs,
1095 OopMapSet* oop_maps,
1096 int& frame_complete,
1097 int& stack_slots,
1098 int& compiled_entry_offset) {
1099 enum layout {
1100 fp_off,
1101 fp_off2,
1102 return_off,
1103 return_off2,
1104 framesize // inclusive of return address
1105 };
1106 // assert(is_even(framesize/2), "sp not 16-byte aligned");
1107
1108 stack_slots = framesize / VMRegImpl::slots_per_word;
1109 assert(stack_slots == 2, "recheck layout");
1110
1111 address start = __ pc();
1112
1113 compiled_entry_offset = __ pc() - start;
1114 __ enter();
1115
1116 __ mv(c_rarg1, sp);
1117
1118 frame_complete = __ pc() - start;
1119 address the_pc = __ pc();
1120
1121 __ 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
1122
1123 __ mv(c_rarg0, xthread);
1124 __ set_last_Java_frame(sp, fp, the_pc, t0);
1125 __ call_VM_leaf(Continuation::freeze_entry(), 2);
1126 __ reset_last_Java_frame(true);
1127
1128 Label pinned;
1129
1130 __ bnez(x10, pinned);
1131
1132 // We've succeeded, set sp to the ContinuationEntry
1133 __ ld(sp, Address(xthread, JavaThread::cont_entry_offset()));
1134 continuation_enter_cleanup(masm);
1135
1136 __ bind(pinned); // pinned -- return to caller
1137
1138 // handle pending exception thrown by freeze
1139 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1140 Label ok;
1141 __ beqz(t0, ok);
1142 __ leave();
1143 __ j(RuntimeAddress(StubRoutines::forward_exception_entry()));
1144 __ bind(ok);
1145
1146 __ leave();
1147 __ ret();
1148
1149 OopMap* map = new OopMap(framesize, 1);
1150 oop_maps->add_gc_map(the_pc - start, map);
1151 }
1152
1153 void SharedRuntime::continuation_enter_cleanup(MacroAssembler* masm) {
1154 ::continuation_enter_cleanup(masm);
1155 }
1156
1157 static void gen_special_dispatch(MacroAssembler* masm,
1158 const methodHandle& method,
1159 const BasicType* sig_bt,
1160 const VMRegPair* regs) {
1161 verify_oop_args(masm, method, sig_bt, regs);
1162 vmIntrinsics::ID iid = method->intrinsic_id();
1163
1164 // Now write the args into the outgoing interpreter space
1165 bool has_receiver = false;
1166 Register receiver_reg = noreg;
1167 int member_arg_pos = -1;
1168 Register member_reg = noreg;
1169 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1170 if (ref_kind != 0) {
1171 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1172 member_reg = x9; // known to be free at this point
1173 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1174 } else if (iid == vmIntrinsics::_invokeBasic) {
1175 has_receiver = true;
1176 } else if (iid == vmIntrinsics::_linkToNative) {
1177 member_arg_pos = method->size_of_parameters() - 1; // trailing NativeEntryPoint argument
1178 member_reg = x9; // known to be free at this point
1179 } else {
1180 fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1181 }
1182
1183 if (member_reg != noreg) {
1184 // Load the member_arg into register, if necessary.
1185 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1186 VMReg r = regs[member_arg_pos].first();
1187 if (r->is_stack()) {
1188 __ ld(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1189 } else {
1190 // no data motion is needed
1191 member_reg = r->as_Register();
1192 }
1193 }
1194
1195 if (has_receiver) {
1196 // Make sure the receiver is loaded into a register.
1197 assert(method->size_of_parameters() > 0, "oob");
1198 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1199 VMReg r = regs[0].first();
1200 assert(r->is_valid(), "bad receiver arg");
1201 if (r->is_stack()) {
1202 // Porting note: This assumes that compiled calling conventions always
1203 // pass the receiver oop in a register. If this is not true on some
1204 // platform, pick a temp and load the receiver from stack.
1205 fatal("receiver always in a register");
1206 receiver_reg = x12; // known to be free at this point
1207 __ ld(receiver_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1208 } else {
1209 // no data motion is needed
1210 receiver_reg = r->as_Register();
1211 }
1212 }
1213
1214 // Figure out which address we are really jumping to:
1215 MethodHandles::generate_method_handle_dispatch(masm, iid,
1216 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1217 }
1218
1219 // ---------------------------------------------------------------------------
1220 // Generate a native wrapper for a given method. The method takes arguments
1221 // in the Java compiled code convention, marshals them to the native
1222 // convention (handlizes oops, etc), transitions to native, makes the call,
1223 // returns to java state (possibly blocking), unhandlizes any result and
1224 // returns.
1225 //
1226 // Critical native functions are a shorthand for the use of
1227 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1228 // functions. The wrapper is expected to unpack the arguments before
1229 // passing them to the callee and perform checks before and after the
1230 // native call to ensure that they GCLocker
1231 // lock_critical/unlock_critical semantics are followed. Some other
1232 // parts of JNI setup are skipped like the tear down of the JNI handle
1233 // block and the check for pending exceptions it's impossible for them
1234 // to be thrown.
1235 //
1236 // They are roughly structured like this:
1237 // if (GCLocker::needs_gc()) SharedRuntime::block_for_jni_critical()
1238 // tranistion to thread_in_native
1239 // unpack array arguments and call native entry point
1240 // check for safepoint in progress
1241 // check if any thread suspend flags are set
1242 // call into JVM and possible unlock the JNI critical
1243 // if a GC was suppressed while in the critical native.
1244 // transition back to thread_in_Java
1245 // return to caller
1246 //
1247 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1248 const methodHandle& method,
1249 int compile_id,
1250 BasicType* in_sig_bt,
1251 VMRegPair* in_regs,
1252 BasicType ret_type) {
1253 if (method->is_continuation_native_intrinsic()) {
1254 int exception_offset = -1;
1255 OopMapSet* oop_maps = new OopMapSet();
1256 int frame_complete = -1;
1257 int stack_slots = -1;
1258 int interpreted_entry_offset = -1;
1259 int vep_offset = -1;
1260 if (method->is_continuation_enter_intrinsic()) {
1261 gen_continuation_enter(masm,
1262 method,
1263 in_sig_bt,
1264 in_regs,
1265 exception_offset,
1266 oop_maps,
1267 frame_complete,
1268 stack_slots,
1269 interpreted_entry_offset,
1270 vep_offset);
1271 } else if (method->is_continuation_yield_intrinsic()) {
1272 gen_continuation_yield(masm,
1273 method,
1274 in_sig_bt,
1275 in_regs,
1276 oop_maps,
1277 frame_complete,
1278 stack_slots,
1279 vep_offset);
1280 } else {
1281 guarantee(false, "Unknown Continuation native intrinsic");
1282 }
1283
1284 #ifdef ASSERT
1285 if (method->is_continuation_enter_intrinsic()) {
1286 assert(interpreted_entry_offset != -1, "Must be set");
1287 assert(exception_offset != -1, "Must be set");
1288 } else {
1289 assert(interpreted_entry_offset == -1, "Must be unset");
1290 assert(exception_offset == -1, "Must be unset");
1291 }
1292 assert(frame_complete != -1, "Must be set");
1293 assert(stack_slots != -1, "Must be set");
1294 assert(vep_offset != -1, "Must be set");
1295 #endif
1296
1297 __ flush();
1298 nmethod* nm = nmethod::new_native_nmethod(method,
1299 compile_id,
1300 masm->code(),
1301 vep_offset,
1302 frame_complete,
1303 stack_slots,
1304 in_ByteSize(-1),
1305 in_ByteSize(-1),
1306 oop_maps,
1307 exception_offset);
1308 if (nm == nullptr) return nm;
1309 if (method->is_continuation_enter_intrinsic()) {
1310 ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
1311 } else if (method->is_continuation_yield_intrinsic()) {
1312 _cont_doYield_stub = nm;
1313 } else {
1314 guarantee(false, "Unknown Continuation native intrinsic");
1315 }
1316 return nm;
1317 }
1318
1319 if (method->is_method_handle_intrinsic()) {
1320 vmIntrinsics::ID iid = method->intrinsic_id();
1321 intptr_t start = (intptr_t)__ pc();
1322 int vep_offset = ((intptr_t)__ pc()) - start;
1323
1324 // First instruction must be a nop as it may need to be patched on deoptimisation
1325 {
1326 Assembler::IncompressibleRegion ir(masm); // keep the nop as 4 bytes for patching.
1327 MacroAssembler::assert_alignment(__ pc());
1328 __ nop(); // 4 bytes
1329 }
1330 gen_special_dispatch(masm,
1331 method,
1332 in_sig_bt,
1333 in_regs);
1334 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
1335 __ flush();
1336 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
1337 return nmethod::new_native_nmethod(method,
1338 compile_id,
1339 masm->code(),
1340 vep_offset,
1341 frame_complete,
1342 stack_slots / VMRegImpl::slots_per_word,
1343 in_ByteSize(-1),
1344 in_ByteSize(-1),
1345 (OopMapSet*)nullptr);
1346 }
1347 address native_func = method->native_function();
1348 assert(native_func != nullptr, "must have function");
1349
1350 // An OopMap for lock (and class if static)
1351 OopMapSet *oop_maps = new OopMapSet();
1352 assert_cond(oop_maps != nullptr);
1353 intptr_t start = (intptr_t)__ pc();
1354
1355 // We have received a description of where all the java arg are located
1356 // on entry to the wrapper. We need to convert these args to where
1357 // the jni function will expect them. To figure out where they go
1358 // we convert the java signature to a C signature by inserting
1359 // the hidden arguments as arg[0] and possibly arg[1] (static method)
1360
1361 const int total_in_args = method->size_of_parameters();
1362 int total_c_args = total_in_args + (method->is_static() ? 2 : 1);
1363
1364 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1365 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1366
1367 int argc = 0;
1368 out_sig_bt[argc++] = T_ADDRESS;
1369 if (method->is_static()) {
1370 out_sig_bt[argc++] = T_OBJECT;
1371 }
1372
1373 for (int i = 0; i < total_in_args ; i++) {
1374 out_sig_bt[argc++] = in_sig_bt[i];
1375 }
1376
1377 // Now figure out where the args must be stored and how much stack space
1378 // they require.
1379 int out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
1380
1381 // Compute framesize for the wrapper. We need to handlize all oops in
1382 // incoming registers
1383
1384 // Calculate the total number of stack slots we will need.
1385
1386 // First count the abi requirement plus all of the outgoing args
1387 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1388
1389 // Now the space for the inbound oop handle area
1390 int total_save_slots = 8 * VMRegImpl::slots_per_word; // 8 arguments passed in registers
1391
1392 int oop_handle_offset = stack_slots;
1393 stack_slots += total_save_slots;
1394
1395 // Now any space we need for handlizing a klass if static method
1396
1397 int klass_slot_offset = 0;
1398 int klass_offset = -1;
1399 int lock_slot_offset = 0;
1400 bool is_static = false;
1401
1402 if (method->is_static()) {
1403 klass_slot_offset = stack_slots;
1404 stack_slots += VMRegImpl::slots_per_word;
1405 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1406 is_static = true;
1407 }
1408
1409 // Plus a lock if needed
1410
1411 if (method->is_synchronized()) {
1412 lock_slot_offset = stack_slots;
1413 stack_slots += VMRegImpl::slots_per_word;
1414 }
1415
1416 // Now a place (+2) to save return values or temp during shuffling
1417 // + 4 for return address (which we own) and saved fp
1418 stack_slots += 6;
1419
1420 // Ok The space we have allocated will look like:
1421 //
1422 //
1423 // FP-> | |
1424 // | 2 slots (ra) |
1425 // | 2 slots (fp) |
1426 // |---------------------|
1427 // | 2 slots for moves |
1428 // |---------------------|
1429 // | lock box (if sync) |
1430 // |---------------------| <- lock_slot_offset
1431 // | klass (if static) |
1432 // |---------------------| <- klass_slot_offset
1433 // | oopHandle area |
1434 // |---------------------| <- oop_handle_offset (8 java arg registers)
1435 // | outbound memory |
1436 // | based arguments |
1437 // | |
1438 // |---------------------|
1439 // | |
1440 // SP-> | out_preserved_slots |
1441 //
1442 //
1443
1444
1445 // Now compute actual number of stack words we need rounding to make
1446 // stack properly aligned.
1447 stack_slots = align_up(stack_slots, StackAlignmentInSlots);
1448
1449 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1450
1451 // First thing make an ic check to see if we should even be here
1452
1453 // We are free to use all registers as temps without saving them and
1454 // restoring them except fp. fp is the only callee save register
1455 // as far as the interpreter and the compiler(s) are concerned.
1456
1457 const Register receiver = j_rarg0;
1458
1459 __ verify_oop(receiver);
1460 assert_different_registers(receiver, t0, t1);
1461
1462 __ ic_check();
1463
1464 int vep_offset = ((intptr_t)__ pc()) - start;
1465
1466 // If we have to make this method not-entrant we'll overwrite its
1467 // first instruction with a jump.
1468 {
1469 Assembler::IncompressibleRegion ir(masm); // keep the nop as 4 bytes for patching.
1470 MacroAssembler::assert_alignment(__ pc());
1471 __ nop(); // 4 bytes
1472 }
1473
1474 if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) {
1475 Label L_skip_barrier;
1476 __ mov_metadata(t1, method->method_holder()); // InstanceKlass*
1477 __ clinit_barrier(t1, t0, &L_skip_barrier);
1478 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
1479
1480 __ bind(L_skip_barrier);
1481 }
1482
1483 // Generate stack overflow check
1484 __ bang_stack_with_offset(checked_cast<int>(StackOverflow::stack_shadow_zone_size()));
1485
1486 // Generate a new frame for the wrapper.
1487 __ enter();
1488 // -2 because return address is already present and so is saved fp
1489 __ sub(sp, sp, stack_size - 2 * wordSize);
1490
1491 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1492 assert_cond(bs != nullptr);
1493 bs->nmethod_entry_barrier(masm, nullptr /* slow_path */, nullptr /* continuation */, nullptr /* guard */);
1494
1495 // Frame is now completed as far as size and linkage.
1496 int frame_complete = ((intptr_t)__ pc()) - start;
1497
1498 // We use x18 as the oop handle for the receiver/klass
1499 // It is callee save so it survives the call to native
1500
1501 const Register oop_handle_reg = x18;
1502
1503 //
1504 // We immediately shuffle the arguments so that any vm call we have to
1505 // make from here on out (sync slow path, jvmti, etc.) we will have
1506 // captured the oops from our caller and have a valid oopMap for
1507 // them.
1508
1509 // -----------------
1510 // The Grand Shuffle
1511
1512 // The Java calling convention is either equal (linux) or denser (win64) than the
1513 // c calling convention. However the because of the jni_env argument the c calling
1514 // convention always has at least one more (and two for static) arguments than Java.
1515 // Therefore if we move the args from java -> c backwards then we will never have
1516 // a register->register conflict and we don't have to build a dependency graph
1517 // and figure out how to break any cycles.
1518 //
1519
1520 // Record esp-based slot for receiver on stack for non-static methods
1521 int receiver_offset = -1;
1522
1523 // This is a trick. We double the stack slots so we can claim
1524 // the oops in the caller's frame. Since we are sure to have
1525 // more args than the caller doubling is enough to make
1526 // sure we can capture all the incoming oop args from the
1527 // caller.
1528 //
1529 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1530 assert_cond(map != nullptr);
1531
1532 int float_args = 0;
1533 int int_args = 0;
1534
1535 #ifdef ASSERT
1536 bool reg_destroyed[Register::number_of_registers];
1537 bool freg_destroyed[FloatRegister::number_of_registers];
1538 for ( int r = 0 ; r < Register::number_of_registers ; r++ ) {
1539 reg_destroyed[r] = false;
1540 }
1541 for ( int f = 0 ; f < FloatRegister::number_of_registers ; f++ ) {
1542 freg_destroyed[f] = false;
1543 }
1544
1545 #endif /* ASSERT */
1546
1547 // For JNI natives the incoming and outgoing registers are offset upwards.
1548 GrowableArray<int> arg_order(2 * total_in_args);
1549
1550 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) {
1551 arg_order.push(i);
1552 arg_order.push(c_arg);
1553 }
1554
1555 for (int ai = 0; ai < arg_order.length(); ai += 2) {
1556 int i = arg_order.at(ai);
1557 int c_arg = arg_order.at(ai + 1);
1558 __ block_comment(err_msg("mv %d -> %d", i, c_arg));
1559 assert(c_arg != -1 && i != -1, "wrong order");
1560 #ifdef ASSERT
1561 if (in_regs[i].first()->is_Register()) {
1562 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!");
1563 } else if (in_regs[i].first()->is_FloatRegister()) {
1564 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!");
1565 }
1566 if (out_regs[c_arg].first()->is_Register()) {
1567 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
1568 } else if (out_regs[c_arg].first()->is_FloatRegister()) {
1569 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true;
1570 }
1571 #endif /* ASSERT */
1572 switch (in_sig_bt[i]) {
1573 case T_ARRAY:
1574 case T_OBJECT:
1575 __ object_move(map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
1576 ((i == 0) && (!is_static)),
1577 &receiver_offset);
1578 int_args++;
1579 break;
1580 case T_VOID:
1581 break;
1582
1583 case T_FLOAT:
1584 __ float_move(in_regs[i], out_regs[c_arg]);
1585 float_args++;
1586 break;
1587
1588 case T_DOUBLE:
1589 assert( i + 1 < total_in_args &&
1590 in_sig_bt[i + 1] == T_VOID &&
1591 out_sig_bt[c_arg + 1] == T_VOID, "bad arg list");
1592 __ double_move(in_regs[i], out_regs[c_arg]);
1593 float_args++;
1594 break;
1595
1596 case T_LONG :
1597 __ long_move(in_regs[i], out_regs[c_arg]);
1598 int_args++;
1599 break;
1600
1601 case T_ADDRESS:
1602 assert(false, "found T_ADDRESS in java args");
1603 break;
1604
1605 default:
1606 __ move32_64(in_regs[i], out_regs[c_arg]);
1607 int_args++;
1608 }
1609 }
1610
1611 // point c_arg at the first arg that is already loaded in case we
1612 // need to spill before we call out
1613 int c_arg = total_c_args - total_in_args;
1614
1615 // Pre-load a static method's oop into c_rarg1.
1616 if (method->is_static()) {
1617
1618 // load oop into a register
1619 __ movoop(c_rarg1,
1620 JNIHandles::make_local(method->method_holder()->java_mirror()));
1621
1622 // Now handlize the static class mirror it's known not-null.
1623 __ sd(c_rarg1, Address(sp, klass_offset));
1624 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
1625
1626 // Now get the handle
1627 __ la(c_rarg1, Address(sp, klass_offset));
1628 // and protect the arg if we must spill
1629 c_arg--;
1630 }
1631
1632 // Change state to native (we save the return address in the thread, since it might not
1633 // be pushed on the stack when we do a stack traversal). It is enough that the pc()
1634 // points into the right code segment. It does not have to be the correct return pc.
1635 // We use the same pc/oopMap repeatedly when we call out.
1636
1637 Label native_return;
1638 if (LockingMode != LM_LEGACY && method->is_object_wait0()) {
1639 // For convenience we use the pc we want to resume to in case of preemption on Object.wait.
1640 __ set_last_Java_frame(sp, noreg, native_return, t0);
1641 } else {
1642 intptr_t the_pc = (intptr_t) __ pc();
1643 oop_maps->add_gc_map(the_pc - start, map);
1644
1645 __ set_last_Java_frame(sp, noreg, __ pc(), t0);
1646 }
1647
1648 Label dtrace_method_entry, dtrace_method_entry_done;
1649 if (DTraceMethodProbes) {
1650 __ j(dtrace_method_entry);
1651 __ bind(dtrace_method_entry_done);
1652 }
1653
1654 // RedefineClasses() tracing support for obsolete method entry
1655 if (log_is_enabled(Trace, redefine, class, obsolete)) {
1656 // protect the args we've loaded
1657 save_args(masm, total_c_args, c_arg, out_regs);
1658 __ mov_metadata(c_rarg1, method());
1659 __ call_VM_leaf(
1660 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
1661 xthread, c_rarg1);
1662 restore_args(masm, total_c_args, c_arg, out_regs);
1663 }
1664
1665 // Lock a synchronized method
1666
1667 // Register definitions used by locking and unlocking
1668
1669 const Register swap_reg = x10;
1670 const Register obj_reg = x9; // Will contain the oop
1671 const Register lock_reg = x30; // Address of compiler lock object (BasicLock)
1672 const Register old_hdr = x30; // value of old header at unlock time
1673 const Register lock_tmp = x31; // Temporary used by lightweight_lock/unlock
1674 const Register tmp = ra;
1675
1676 Label slow_path_lock;
1677 Label lock_done;
1678
1679 if (method->is_synchronized()) {
1680 Label count;
1681
1682 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes();
1683
1684 // Get the handle (the 2nd argument)
1685 __ mv(oop_handle_reg, c_rarg1);
1686
1687 // Get address of the box
1688
1689 __ la(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1690
1691 // Load the oop from the handle
1692 __ ld(obj_reg, Address(oop_handle_reg, 0));
1693
1694 if (LockingMode == LM_MONITOR) {
1695 __ j(slow_path_lock);
1696 } else if (LockingMode == LM_LEGACY) {
1697 // Load (object->mark() | 1) into swap_reg % x10
1698 __ ld(t0, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
1699 __ ori(swap_reg, t0, 1);
1700
1701 // Save (object->mark() | 1) into BasicLock's displaced header
1702 __ sd(swap_reg, Address(lock_reg, mark_word_offset));
1703
1704 // src -> dest if dest == x10 else x10 <- dest
1705 __ cmpxchg_obj_header(x10, lock_reg, obj_reg, lock_tmp, count, /*fallthrough*/nullptr);
1706
1707 // Test if the oopMark is an obvious stack pointer, i.e.,
1708 // 1) (mark & 3) == 0, and
1709 // 2) sp <= mark < mark + os::pagesize()
1710 // These 3 tests can be done by evaluating the following
1711 // expression: ((mark - sp) & (3 - os::vm_page_size())),
1712 // assuming both stack pointer and pagesize have their
1713 // least significant 2 bits clear.
1714 // NOTE: the oopMark is in swap_reg % 10 as the result of cmpxchg
1715
1716 __ sub(swap_reg, swap_reg, sp);
1717 __ mv(t0, 3 - (int)os::vm_page_size());
1718 __ andr(swap_reg, swap_reg, t0);
1719
1720 // Save the test result, for recursive case, the result is zero
1721 __ sd(swap_reg, Address(lock_reg, mark_word_offset));
1722 __ bnez(swap_reg, slow_path_lock);
1723
1724 __ bind(count);
1725 __ inc_held_monitor_count(t0);
1726 } else {
1727 assert(LockingMode == LM_LIGHTWEIGHT, "must be");
1728 __ lightweight_lock(lock_reg, obj_reg, swap_reg, tmp, lock_tmp, slow_path_lock);
1729 }
1730
1731 // Slow path will re-enter here
1732 __ bind(lock_done);
1733 }
1734
1735
1736 // Finally just about ready to make the JNI call
1737
1738 // get JNIEnv* which is first argument to native
1739 __ la(c_rarg0, Address(xthread, in_bytes(JavaThread::jni_environment_offset())));
1740
1741 // Now set thread in native
1742 __ la(t1, Address(xthread, JavaThread::thread_state_offset()));
1743 __ mv(t0, _thread_in_native);
1744 __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore);
1745 __ sw(t0, Address(t1));
1746
1747 // Clobbers t1
1748 __ rt_call(native_func);
1749
1750 // Verify or restore cpu control state after JNI call
1751 __ restore_cpu_control_state_after_jni(t0);
1752
1753 // Unpack native results.
1754 if (ret_type != T_OBJECT && ret_type != T_ARRAY) {
1755 __ cast_primitive_type(ret_type, x10);
1756 }
1757
1758 Label safepoint_in_progress, safepoint_in_progress_done;
1759
1760 // Switch thread to "native transition" state before reading the synchronization state.
1761 // This additional state is necessary because reading and testing the synchronization
1762 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1763 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1764 // VM thread changes sync state to synchronizing and suspends threads for GC.
1765 // Thread A is resumed to finish this native method, but doesn't block here since it
1766 // didn't see any synchronization is progress, and escapes.
1767 __ mv(t0, _thread_in_native_trans);
1768
1769 __ sw(t0, Address(xthread, JavaThread::thread_state_offset()));
1770
1771 // Force this write out before the read below
1772 if (!UseSystemMemoryBarrier) {
1773 __ membar(MacroAssembler::AnyAny);
1774 }
1775
1776 // check for safepoint operation in progress and/or pending suspend requests
1777 {
1778 // We need an acquire here to ensure that any subsequent load of the
1779 // global SafepointSynchronize::_state flag is ordered after this load
1780 // of the thread-local polling word. We don't want this poll to
1781 // return false (i.e. not safepointing) and a later poll of the global
1782 // SafepointSynchronize::_state spuriously to return true.
1783 // This is to avoid a race when we're in a native->Java transition
1784 // racing the code which wakes up from a safepoint.
1785
1786 __ safepoint_poll(safepoint_in_progress, true /* at_return */, true /* acquire */, false /* in_nmethod */);
1787 __ lwu(t0, Address(xthread, JavaThread::suspend_flags_offset()));
1788 __ bnez(t0, safepoint_in_progress);
1789 __ bind(safepoint_in_progress_done);
1790 }
1791
1792 // change thread state
1793 __ la(t1, Address(xthread, JavaThread::thread_state_offset()));
1794 __ mv(t0, _thread_in_Java);
1795 __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore);
1796 __ sw(t0, Address(t1));
1797
1798 if (LockingMode != LM_LEGACY && method->is_object_wait0()) {
1799 // Check preemption for Object.wait()
1800 __ ld(t1, Address(xthread, JavaThread::preempt_alternate_return_offset()));
1801 __ beqz(t1, native_return);
1802 __ sd(zr, Address(xthread, JavaThread::preempt_alternate_return_offset()));
1803 __ jr(t1);
1804 __ bind(native_return);
1805
1806 intptr_t the_pc = (intptr_t) __ pc();
1807 oop_maps->add_gc_map(the_pc - start, map);
1808 }
1809
1810 Label reguard;
1811 Label reguard_done;
1812 __ lbu(t0, Address(xthread, JavaThread::stack_guard_state_offset()));
1813 __ mv(t1, StackOverflow::stack_guard_yellow_reserved_disabled);
1814 __ beq(t0, t1, reguard);
1815 __ bind(reguard_done);
1816
1817 // native result if any is live
1818
1819 // Unlock
1820 Label unlock_done;
1821 Label slow_path_unlock;
1822 if (method->is_synchronized()) {
1823
1824 // Get locked oop from the handle we passed to jni
1825 __ ld(obj_reg, Address(oop_handle_reg, 0));
1826
1827 Label done, not_recursive;
1828
1829 if (LockingMode == LM_LEGACY) {
1830 // Simple recursive lock?
1831 __ ld(t0, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1832 __ bnez(t0, not_recursive);
1833 __ dec_held_monitor_count(t0);
1834 __ j(done);
1835 }
1836
1837 __ bind(not_recursive);
1838
1839 // Must save x10 if if it is live now because cmpxchg must use it
1840 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1841 save_native_result(masm, ret_type, stack_slots);
1842 }
1843
1844 if (LockingMode == LM_MONITOR) {
1845 __ j(slow_path_unlock);
1846 } else if (LockingMode == LM_LEGACY) {
1847 // get address of the stack lock
1848 __ la(x10, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1849 // get old displaced header
1850 __ ld(old_hdr, Address(x10, 0));
1851
1852 // Atomic swap old header if oop still contains the stack lock
1853 Label count;
1854 __ cmpxchg_obj_header(x10, old_hdr, obj_reg, lock_tmp, count, &slow_path_unlock);
1855 __ bind(count);
1856 __ dec_held_monitor_count(t0);
1857 } else {
1858 assert(LockingMode == LM_LIGHTWEIGHT, "");
1859 __ lightweight_unlock(obj_reg, old_hdr, swap_reg, lock_tmp, slow_path_unlock);
1860 }
1861
1862 // slow path re-enters here
1863 __ bind(unlock_done);
1864 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1865 restore_native_result(masm, ret_type, stack_slots);
1866 }
1867
1868 __ bind(done);
1869 }
1870
1871 Label dtrace_method_exit, dtrace_method_exit_done;
1872 if (DTraceMethodProbes) {
1873 __ j(dtrace_method_exit);
1874 __ bind(dtrace_method_exit_done);
1875 }
1876
1877 __ reset_last_Java_frame(false);
1878
1879 // Unbox oop result, e.g. JNIHandles::resolve result.
1880 if (is_reference_type(ret_type)) {
1881 __ resolve_jobject(x10, x11, x12);
1882 }
1883
1884 if (CheckJNICalls) {
1885 // clear_pending_jni_exception_check
1886 __ sd(zr, Address(xthread, JavaThread::pending_jni_exception_check_fn_offset()));
1887 }
1888
1889 // reset handle block
1890 __ ld(x12, Address(xthread, JavaThread::active_handles_offset()));
1891 __ sd(zr, Address(x12, JNIHandleBlock::top_offset()));
1892
1893 __ leave();
1894
1895 // Any exception pending?
1896 Label exception_pending;
1897 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1898 __ bnez(t0, exception_pending);
1899
1900 // We're done
1901 __ ret();
1902
1903 // Unexpected paths are out of line and go here
1904
1905 // forward the exception
1906 __ bind(exception_pending);
1907
1908 // and forward the exception
1909 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
1910
1911 // Slow path locking & unlocking
1912 if (method->is_synchronized()) {
1913
1914 __ block_comment("Slow path lock {");
1915 __ bind(slow_path_lock);
1916
1917 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM
1918 // args are (oop obj, BasicLock* lock, JavaThread* thread)
1919
1920 // protect the args we've loaded
1921 save_args(masm, total_c_args, c_arg, out_regs);
1922
1923 __ mv(c_rarg0, obj_reg);
1924 __ mv(c_rarg1, lock_reg);
1925 __ mv(c_rarg2, xthread);
1926
1927 // Not a leaf but we have last_Java_frame setup as we want.
1928 // We don't want to unmount in case of contention since that would complicate preserving
1929 // the arguments that had already been marshalled into the native convention. So we force
1930 // the freeze slow path to find this native wrapper frame (see recurse_freeze_native_frame())
1931 // and pin the vthread. Otherwise the fast path won't find it since we don't walk the stack.
1932 __ push_cont_fastpath();
1933 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3);
1934 __ pop_cont_fastpath();
1935 restore_args(masm, total_c_args, c_arg, out_regs);
1936
1937 #ifdef ASSERT
1938 { Label L;
1939 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1940 __ beqz(t0, L);
1941 __ stop("no pending exception allowed on exit from monitorenter");
1942 __ bind(L);
1943 }
1944 #endif
1945 __ j(lock_done);
1946
1947 __ block_comment("} Slow path lock");
1948
1949 __ block_comment("Slow path unlock {");
1950 __ bind(slow_path_unlock);
1951
1952 if (ret_type == T_FLOAT || ret_type == T_DOUBLE) {
1953 save_native_result(masm, ret_type, stack_slots);
1954 }
1955
1956 __ mv(c_rarg2, xthread);
1957 __ la(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1958 __ mv(c_rarg0, obj_reg);
1959
1960 // Save pending exception around call to VM (which contains an EXCEPTION_MARK)
1961 // NOTE that obj_reg == x9 currently
1962 __ ld(x9, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1963 __ sd(zr, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1964
1965 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C));
1966
1967 #ifdef ASSERT
1968 {
1969 Label L;
1970 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1971 __ beqz(t0, L);
1972 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C");
1973 __ bind(L);
1974 }
1975 #endif /* ASSERT */
1976
1977 __ sd(x9, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1978
1979 if (ret_type == T_FLOAT || ret_type == T_DOUBLE) {
1980 restore_native_result(masm, ret_type, stack_slots);
1981 }
1982 __ j(unlock_done);
1983
1984 __ block_comment("} Slow path unlock");
1985
1986 } // synchronized
1987
1988 // SLOW PATH Reguard the stack if needed
1989
1990 __ bind(reguard);
1991 save_native_result(masm, ret_type, stack_slots);
1992 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
1993 restore_native_result(masm, ret_type, stack_slots);
1994 // and continue
1995 __ j(reguard_done);
1996
1997 // SLOW PATH safepoint
1998 {
1999 __ block_comment("safepoint {");
2000 __ bind(safepoint_in_progress);
2001
2002 // Don't use call_VM as it will see a possible pending exception and forward it
2003 // and never return here preventing us from clearing _last_native_pc down below.
2004 //
2005 save_native_result(masm, ret_type, stack_slots);
2006 __ mv(c_rarg0, xthread);
2007 #ifndef PRODUCT
2008 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
2009 #endif
2010 __ rt_call(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans));
2011
2012 // Restore any method result value
2013 restore_native_result(masm, ret_type, stack_slots);
2014
2015 __ j(safepoint_in_progress_done);
2016 __ block_comment("} safepoint");
2017 }
2018
2019 // SLOW PATH dtrace support
2020 if (DTraceMethodProbes) {
2021 {
2022 __ block_comment("dtrace entry {");
2023 __ bind(dtrace_method_entry);
2024
2025 // We have all of the arguments setup at this point. We must not touch any register
2026 // argument registers at this point (what if we save/restore them there are no oop?
2027
2028 save_args(masm, total_c_args, c_arg, out_regs);
2029 __ mov_metadata(c_rarg1, method());
2030 __ call_VM_leaf(
2031 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
2032 xthread, c_rarg1);
2033 restore_args(masm, total_c_args, c_arg, out_regs);
2034 __ j(dtrace_method_entry_done);
2035 __ block_comment("} dtrace entry");
2036 }
2037
2038 {
2039 __ block_comment("dtrace exit {");
2040 __ bind(dtrace_method_exit);
2041 save_native_result(masm, ret_type, stack_slots);
2042 __ mov_metadata(c_rarg1, method());
2043 __ call_VM_leaf(
2044 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
2045 xthread, c_rarg1);
2046 restore_native_result(masm, ret_type, stack_slots);
2047 __ j(dtrace_method_exit_done);
2048 __ block_comment("} dtrace exit");
2049 }
2050 }
2051
2052 __ flush();
2053
2054 nmethod *nm = nmethod::new_native_nmethod(method,
2055 compile_id,
2056 masm->code(),
2057 vep_offset,
2058 frame_complete,
2059 stack_slots / VMRegImpl::slots_per_word,
2060 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2061 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size),
2062 oop_maps);
2063 assert(nm != nullptr, "create native nmethod fail!");
2064 return nm;
2065 }
2066
2067 // this function returns the adjust size (in number of words) to a c2i adapter
2068 // activation for use during deoptimization
2069 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2070 assert(callee_locals >= callee_parameters,
2071 "test and remove; got more parms than locals");
2072 if (callee_locals < callee_parameters) {
2073 return 0; // No adjustment for negative locals
2074 }
2075 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords;
2076 // diff is counted in stack words
2077 return align_up(diff, 2);
2078 }
2079
2080 //------------------------------generate_deopt_blob----------------------------
2081 void SharedRuntime::generate_deopt_blob() {
2082 // Allocate space for the code
2083 ResourceMark rm;
2084 // Setup code generation tools
2085 int pad = 0;
2086 #if INCLUDE_JVMCI
2087 if (EnableJVMCI) {
2088 pad += 512; // Increase the buffer size when compiling for JVMCI
2089 }
2090 #endif
2091 const char* name = SharedRuntime::stub_name(SharedStubId::deopt_id);
2092 CodeBuffer buffer(name, 2048 + pad, 1024);
2093 MacroAssembler* masm = new MacroAssembler(&buffer);
2094 int frame_size_in_words = -1;
2095 OopMap* map = nullptr;
2096 OopMapSet *oop_maps = new OopMapSet();
2097 assert_cond(masm != nullptr && oop_maps != nullptr);
2098 RegisterSaver reg_saver(COMPILER2_OR_JVMCI != 0);
2099
2100 // -------------
2101 // This code enters when returning to a de-optimized nmethod. A return
2102 // address has been pushed on the stack, and return values are in
2103 // registers.
2104 // If we are doing a normal deopt then we were called from the patched
2105 // nmethod from the point we returned to the nmethod. So the return
2106 // address on the stack is wrong by NativeCall::instruction_size
2107 // We will adjust the value so it looks like we have the original return
2108 // address on the stack (like when we eagerly deoptimized).
2109 // In the case of an exception pending when deoptimizing, we enter
2110 // with a return address on the stack that points after the call we patched
2111 // into the exception handler. We have the following register state from,
2112 // e.g., the forward exception stub (see stubGenerator_riscv.cpp).
2113 // x10: exception oop
2114 // x9: exception handler
2115 // x13: throwing pc
2116 // So in this case we simply jam x13 into the useless return address and
2117 // the stack looks just like we want.
2118 //
2119 // At this point we need to de-opt. We save the argument return
2120 // registers. We call the first C routine, fetch_unroll_info(). This
2121 // routine captures the return values and returns a structure which
2122 // describes the current frame size and the sizes of all replacement frames.
2123 // The current frame is compiled code and may contain many inlined
2124 // functions, each with their own JVM state. We pop the current frame, then
2125 // push all the new frames. Then we call the C routine unpack_frames() to
2126 // populate these frames. Finally unpack_frames() returns us the new target
2127 // address. Notice that callee-save registers are BLOWN here; they have
2128 // already been captured in the vframeArray at the time the return PC was
2129 // patched.
2130 address start = __ pc();
2131 Label cont;
2132
2133 // Prolog for non exception case!
2134
2135 // Save everything in sight.
2136 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2137
2138 // Normal deoptimization. Save exec mode for unpack_frames.
2139 __ mv(xcpool, Deoptimization::Unpack_deopt); // callee-saved
2140 __ j(cont);
2141
2142 int reexecute_offset = __ pc() - start;
2143 #if INCLUDE_JVMCI && !defined(COMPILER1)
2144 if (UseJVMCICompiler) {
2145 // JVMCI does not use this kind of deoptimization
2146 __ should_not_reach_here();
2147 }
2148 #endif
2149
2150 // Reexecute case
2151 // return address is the pc describes what bci to do re-execute at
2152
2153 // No need to update map as each call to save_live_registers will produce identical oopmap
2154 (void) reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2155
2156 __ mv(xcpool, Deoptimization::Unpack_reexecute); // callee-saved
2157 __ j(cont);
2158
2159 #if INCLUDE_JVMCI
2160 Label after_fetch_unroll_info_call;
2161 int implicit_exception_uncommon_trap_offset = 0;
2162 int uncommon_trap_offset = 0;
2163
2164 if (EnableJVMCI) {
2165 implicit_exception_uncommon_trap_offset = __ pc() - start;
2166
2167 __ ld(ra, Address(xthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2168 __ sd(zr, Address(xthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2169
2170 uncommon_trap_offset = __ pc() - start;
2171
2172 // Save everything in sight.
2173 reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2174 // fetch_unroll_info needs to call last_java_frame()
2175 Label retaddr;
2176 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2177
2178 __ lw(c_rarg1, Address(xthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2179 __ mv(t0, -1);
2180 __ sw(t0, Address(xthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2181
2182 __ mv(xcpool, Deoptimization::Unpack_reexecute);
2183 __ mv(c_rarg0, xthread);
2184 __ orrw(c_rarg2, zr, xcpool); // exec mode
2185 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap));
2186 __ bind(retaddr);
2187 oop_maps->add_gc_map( __ pc()-start, map->deep_copy());
2188
2189 __ reset_last_Java_frame(false);
2190
2191 __ j(after_fetch_unroll_info_call);
2192 } // EnableJVMCI
2193 #endif // INCLUDE_JVMCI
2194
2195 int exception_offset = __ pc() - start;
2196
2197 // Prolog for exception case
2198
2199 // all registers are dead at this entry point, except for x10, and
2200 // x13 which contain the exception oop and exception pc
2201 // respectively. Set them in TLS and fall thru to the
2202 // unpack_with_exception_in_tls entry point.
2203
2204 __ sd(x13, Address(xthread, JavaThread::exception_pc_offset()));
2205 __ sd(x10, Address(xthread, JavaThread::exception_oop_offset()));
2206
2207 int exception_in_tls_offset = __ pc() - start;
2208
2209 // new implementation because exception oop is now passed in JavaThread
2210
2211 // Prolog for exception case
2212 // All registers must be preserved because they might be used by LinearScan
2213 // Exceptiop oop and throwing PC are passed in JavaThread
2214 // tos: stack at point of call to method that threw the exception (i.e. only
2215 // args are on the stack, no return address)
2216
2217 // The return address pushed by save_live_registers will be patched
2218 // later with the throwing pc. The correct value is not available
2219 // now because loading it from memory would destroy registers.
2220
2221 // NB: The SP at this point must be the SP of the method that is
2222 // being deoptimized. Deoptimization assumes that the frame created
2223 // here by save_live_registers is immediately below the method's SP.
2224 // This is a somewhat fragile mechanism.
2225
2226 // Save everything in sight.
2227 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2228
2229 // Now it is safe to overwrite any register
2230
2231 // Deopt during an exception. Save exec mode for unpack_frames.
2232 __ mv(xcpool, Deoptimization::Unpack_exception); // callee-saved
2233
2234 // load throwing pc from JavaThread and patch it as the return address
2235 // of the current frame. Then clear the field in JavaThread
2236
2237 __ ld(x13, Address(xthread, JavaThread::exception_pc_offset()));
2238 __ sd(x13, Address(fp, frame::return_addr_offset * wordSize));
2239 __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2240
2241 #ifdef ASSERT
2242 // verify that there is really an exception oop in JavaThread
2243 __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2244 __ verify_oop(x10);
2245
2246 // verify that there is no pending exception
2247 Label no_pending_exception;
2248 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2249 __ beqz(t0, no_pending_exception);
2250 __ stop("must not have pending exception here");
2251 __ bind(no_pending_exception);
2252 #endif
2253
2254 __ bind(cont);
2255
2256 // Call C code. Need thread and this frame, but NOT official VM entry
2257 // crud. We cannot block on this call, no GC can happen.
2258 //
2259 // UnrollBlock* fetch_unroll_info(JavaThread* thread)
2260
2261 // fetch_unroll_info needs to call last_java_frame().
2262
2263 Label retaddr;
2264 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2265 #ifdef ASSERT
2266 {
2267 Label L;
2268 __ ld(t0, Address(xthread,
2269 JavaThread::last_Java_fp_offset()));
2270 __ beqz(t0, L);
2271 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared");
2272 __ bind(L);
2273 }
2274 #endif // ASSERT
2275 __ mv(c_rarg0, xthread);
2276 __ mv(c_rarg1, xcpool);
2277 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info));
2278 __ bind(retaddr);
2279
2280 // Need to have an oopmap that tells fetch_unroll_info where to
2281 // find any register it might need.
2282 oop_maps->add_gc_map(__ pc() - start, map);
2283
2284 __ reset_last_Java_frame(false);
2285
2286 #if INCLUDE_JVMCI
2287 if (EnableJVMCI) {
2288 __ bind(after_fetch_unroll_info_call);
2289 }
2290 #endif
2291
2292 // Load UnrollBlock* into x15
2293 __ mv(x15, x10);
2294
2295 __ lwu(xcpool, Address(x15, Deoptimization::UnrollBlock::unpack_kind_offset()));
2296 Label noException;
2297 __ mv(t0, Deoptimization::Unpack_exception);
2298 __ bne(xcpool, t0, noException); // Was exception pending?
2299 __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2300 __ ld(x13, Address(xthread, JavaThread::exception_pc_offset()));
2301 __ sd(zr, Address(xthread, JavaThread::exception_oop_offset()));
2302 __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2303
2304 __ verify_oop(x10);
2305
2306 // Overwrite the result registers with the exception results.
2307 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2308
2309 __ bind(noException);
2310
2311 // Only register save data is on the stack.
2312 // Now restore the result registers. Everything else is either dead
2313 // or captured in the vframeArray.
2314
2315 // Restore fp result register
2316 __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2317 // Restore integer result register
2318 __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2319
2320 // Pop all of the register save area off the stack
2321 __ add(sp, sp, frame_size_in_words * wordSize);
2322
2323 // All of the register save area has been popped of the stack. Only the
2324 // return address remains.
2325
2326 // Pop all the frames we must move/replace.
2327 //
2328 // Frame picture (youngest to oldest)
2329 // 1: self-frame (no frame link)
2330 // 2: deopting frame (no frame link)
2331 // 3: caller of deopting frame (could be compiled/interpreted).
2332 //
2333 // Note: by leaving the return address of self-frame on the stack
2334 // and using the size of frame 2 to adjust the stack
2335 // when we are done the return to frame 3 will still be on the stack.
2336
2337 // Pop deoptimized frame
2338 __ lwu(x12, Address(x15, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset()));
2339 __ subi(x12, x12, 2 * wordSize);
2340 __ add(sp, sp, x12);
2341 __ ld(fp, Address(sp, 0));
2342 __ ld(ra, Address(sp, wordSize));
2343 __ addi(sp, sp, 2 * wordSize);
2344 // RA should now be the return address to the caller (3)
2345
2346 #ifdef ASSERT
2347 // Compilers generate code that bang the stack by as much as the
2348 // interpreter would need. So this stack banging should never
2349 // trigger a fault. Verify that it does not on non product builds.
2350 __ lwu(x9, Address(x15, Deoptimization::UnrollBlock::total_frame_sizes_offset()));
2351 __ bang_stack_size(x9, x12);
2352 #endif
2353 // Load address of array of frame pcs into x12
2354 __ ld(x12, Address(x15, Deoptimization::UnrollBlock::frame_pcs_offset()));
2355
2356 // Load address of array of frame sizes into x14
2357 __ ld(x14, Address(x15, Deoptimization::UnrollBlock::frame_sizes_offset()));
2358
2359 // Load counter into x13
2360 __ lwu(x13, Address(x15, Deoptimization::UnrollBlock::number_of_frames_offset()));
2361
2362 // Now adjust the caller's stack to make up for the extra locals
2363 // but record the original sp so that we can save it in the skeletal interpreter
2364 // frame and the stack walking of interpreter_sender will get the unextended sp
2365 // value and not the "real" sp value.
2366
2367 const Register sender_sp = x16;
2368
2369 __ mv(sender_sp, sp);
2370 __ lwu(x9, Address(x15,
2371 Deoptimization::UnrollBlock::
2372 caller_adjustment_offset()));
2373 __ sub(sp, sp, x9);
2374
2375 // Push interpreter frames in a loop
2376 __ mv(t0, 0xDEADDEAD); // Make a recognizable pattern
2377 __ mv(t1, t0);
2378 Label loop;
2379 __ bind(loop);
2380 __ ld(x9, Address(x14, 0)); // Load frame size
2381 __ addi(x14, x14, wordSize);
2382 __ subi(x9, x9, 2 * wordSize); // We'll push pc and fp by hand
2383 __ ld(ra, Address(x12, 0)); // Load pc
2384 __ addi(x12, x12, wordSize);
2385 __ enter(); // Save old & set new fp
2386 __ sub(sp, sp, x9); // Prolog
2387 // This value is corrected by layout_activation_impl
2388 __ sd(zr, Address(fp, frame::interpreter_frame_last_sp_offset * wordSize));
2389 __ sd(sender_sp, Address(fp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable
2390 __ mv(sender_sp, sp); // Pass sender_sp to next frame
2391 __ subi(x13, x13, 1); // Decrement counter
2392 __ bnez(x13, loop);
2393
2394 // Re-push self-frame
2395 __ ld(ra, Address(x12));
2396 __ enter();
2397
2398 // Allocate a full sized register save area. We subtract 2 because
2399 // enter() just pushed 2 words
2400 __ sub(sp, sp, (frame_size_in_words - 2) * wordSize);
2401
2402 // Restore frame locals after moving the frame
2403 __ fsd(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2404 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2405
2406 // Call C code. Need thread but NOT official VM entry
2407 // crud. We cannot block on this call, no GC can happen. Call should
2408 // restore return values to their stack-slots with the new SP.
2409 //
2410 // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)
2411
2412 // Use fp because the frames look interpreted now
2413 // Don't need the precise return PC here, just precise enough to point into this code blob.
2414 address the_pc = __ pc();
2415 __ set_last_Java_frame(sp, fp, the_pc, t0);
2416
2417 __ mv(c_rarg0, xthread);
2418 __ mv(c_rarg1, xcpool); // second arg: exec_mode
2419 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames));
2420
2421 // Set an oopmap for the call site
2422 // Use the same PC we used for the last java frame
2423 oop_maps->add_gc_map(the_pc - start,
2424 new OopMap(frame_size_in_words, 0));
2425
2426 // Clear fp AND pc
2427 __ reset_last_Java_frame(true);
2428
2429 // Collect return values
2430 __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2431 __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2432
2433 // Pop self-frame.
2434 __ leave(); // Epilog
2435
2436 // Jump to interpreter
2437 __ ret();
2438
2439 // Make sure all code is generated
2440 masm->flush();
2441
2442 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words);
2443 assert(_deopt_blob != nullptr, "create deoptimization blob fail!");
2444 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
2445 #if INCLUDE_JVMCI
2446 if (EnableJVMCI) {
2447 _deopt_blob->set_uncommon_trap_offset(uncommon_trap_offset);
2448 _deopt_blob->set_implicit_exception_uncommon_trap_offset(implicit_exception_uncommon_trap_offset);
2449 }
2450 #endif
2451 }
2452
2453 // Number of stack slots between incoming argument block and the start of
2454 // a new frame. The PROLOG must add this many slots to the stack. The
2455 // EPILOG must remove this many slots.
2456 // RISCV needs two words for RA (return address) and FP (frame pointer).
2457 uint SharedRuntime::in_preserve_stack_slots() {
2458 return 2 * VMRegImpl::slots_per_word;
2459 }
2460
2461 uint SharedRuntime::out_preserve_stack_slots() {
2462 return 0;
2463 }
2464
2465 VMReg SharedRuntime::thread_register() {
2466 return xthread->as_VMReg();
2467 }
2468
2469 //------------------------------generate_handler_blob------
2470 //
2471 // Generate a special Compile2Runtime blob that saves all registers,
2472 // and setup oopmap.
2473 //
2474 SafepointBlob* SharedRuntime::generate_handler_blob(SharedStubId id, address call_ptr) {
2475 assert(is_polling_page_id(id), "expected a polling page stub id");
2476
2477 ResourceMark rm;
2478 OopMapSet *oop_maps = new OopMapSet();
2479 assert_cond(oop_maps != nullptr);
2480 OopMap* map = nullptr;
2481
2482 // Allocate space for the code. Setup code generation tools.
2483 const char* name = SharedRuntime::stub_name(id);
2484 CodeBuffer buffer(name, 2048, 1024);
2485 MacroAssembler* masm = new MacroAssembler(&buffer);
2486 assert_cond(masm != nullptr);
2487
2488 address start = __ pc();
2489 address call_pc = nullptr;
2490 int frame_size_in_words = -1;
2491 bool cause_return = (id == SharedStubId::polling_page_return_handler_id);
2492 RegisterSaver reg_saver(id == SharedStubId::polling_page_vectors_safepoint_handler_id /* save_vectors */);
2493
2494 // Save Integer and Float registers.
2495 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2496
2497 // The following is basically a call_VM. However, we need the precise
2498 // address of the call in order to generate an oopmap. Hence, we do all the
2499 // work ourselves.
2500
2501 Label retaddr;
2502 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2503
2504 // The return address must always be correct so that frame constructor never
2505 // sees an invalid pc.
2506
2507 if (!cause_return) {
2508 // overwrite the return address pushed by save_live_registers
2509 // Additionally, x18 is a callee-saved register so we can look at
2510 // it later to determine if someone changed the return address for
2511 // us!
2512 __ ld(x18, Address(xthread, JavaThread::saved_exception_pc_offset()));
2513 __ sd(x18, Address(fp, frame::return_addr_offset * wordSize));
2514 }
2515
2516 // Do the call
2517 __ mv(c_rarg0, xthread);
2518 __ rt_call(call_ptr);
2519 __ bind(retaddr);
2520
2521 // Set an oopmap for the call site. This oopmap will map all
2522 // oop-registers and debug-info registers as callee-saved. This
2523 // will allow deoptimization at this safepoint to find all possible
2524 // debug-info recordings, as well as let GC find all oops.
2525
2526 oop_maps->add_gc_map( __ pc() - start, map);
2527
2528 Label noException;
2529
2530 __ reset_last_Java_frame(false);
2531
2532 __ membar(MacroAssembler::LoadLoad | MacroAssembler::LoadStore);
2533
2534 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2535 __ beqz(t0, noException);
2536
2537 // Exception pending
2538
2539 reg_saver.restore_live_registers(masm);
2540
2541 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2542
2543 // No exception case
2544 __ bind(noException);
2545
2546 Label no_adjust, bail;
2547 if (!cause_return) {
2548 // If our stashed return pc was modified by the runtime we avoid touching it
2549 __ ld(t0, Address(fp, frame::return_addr_offset * wordSize));
2550 __ bne(x18, t0, no_adjust);
2551
2552 #ifdef ASSERT
2553 // Verify the correct encoding of the poll we're about to skip.
2554 // See NativeInstruction::is_lwu_to_zr()
2555 __ lwu(t0, Address(x18));
2556 __ andi(t1, t0, 0b1111111);
2557 __ mv(t2, 0b0000011);
2558 __ bne(t1, t2, bail); // 0-6:0b0000011
2559 __ srli(t1, t0, 7);
2560 __ andi(t1, t1, 0b11111);
2561 __ bnez(t1, bail); // 7-11:0b00000
2562 __ srli(t1, t0, 12);
2563 __ andi(t1, t1, 0b111);
2564 __ mv(t2, 0b110);
2565 __ bne(t1, t2, bail); // 12-14:0b110
2566 #endif
2567
2568 // Adjust return pc forward to step over the safepoint poll instruction
2569 __ addi(x18, x18, NativeInstruction::instruction_size);
2570 __ sd(x18, Address(fp, frame::return_addr_offset * wordSize));
2571 }
2572
2573 __ bind(no_adjust);
2574 // Normal exit, restore registers and exit.
2575
2576 reg_saver.restore_live_registers(masm);
2577 __ ret();
2578
2579 #ifdef ASSERT
2580 __ bind(bail);
2581 __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected");
2582 #endif
2583
2584 // Make sure all code is generated
2585 masm->flush();
2586
2587 // Fill-out other meta info
2588 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words);
2589 }
2590
2591 //
2592 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
2593 //
2594 // Generate a stub that calls into vm to find out the proper destination
2595 // of a java call. All the argument registers are live at this point
2596 // but since this is generic code we don't know what they are and the caller
2597 // must do any gc of the args.
2598 //
2599 RuntimeStub* SharedRuntime::generate_resolve_blob(SharedStubId id, address destination) {
2600 assert(StubRoutines::forward_exception_entry() != nullptr, "must be generated before");
2601 assert(is_resolve_id(id), "expected a resolve stub id");
2602
2603 // allocate space for the code
2604 ResourceMark rm;
2605
2606 const char* name = SharedRuntime::stub_name(id);
2607 CodeBuffer buffer(name, 1000, 512);
2608 MacroAssembler* masm = new MacroAssembler(&buffer);
2609 assert_cond(masm != nullptr);
2610
2611 int frame_size_in_words = -1;
2612 RegisterSaver reg_saver(false /* save_vectors */);
2613
2614 OopMapSet *oop_maps = new OopMapSet();
2615 assert_cond(oop_maps != nullptr);
2616 OopMap* map = nullptr;
2617
2618 int start = __ offset();
2619
2620 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2621
2622 int frame_complete = __ offset();
2623
2624 {
2625 Label retaddr;
2626 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2627
2628 __ mv(c_rarg0, xthread);
2629 __ rt_call(destination);
2630 __ bind(retaddr);
2631 }
2632
2633 // Set an oopmap for the call site.
2634 // We need this not only for callee-saved registers, but also for volatile
2635 // registers that the compiler might be keeping live across a safepoint.
2636
2637 oop_maps->add_gc_map( __ offset() - start, map);
2638
2639 // x10 contains the address we are going to jump to assuming no exception got installed
2640
2641 // clear last_Java_sp
2642 __ reset_last_Java_frame(false);
2643 // check for pending exceptions
2644 Label pending;
2645 __ ld(t1, Address(xthread, Thread::pending_exception_offset()));
2646 __ bnez(t1, pending);
2647
2648 // get the returned Method*
2649 __ get_vm_result_2(xmethod, xthread);
2650 __ sd(xmethod, Address(sp, reg_saver.reg_offset_in_bytes(xmethod)));
2651
2652 // x10 is where we want to jump, overwrite t1 which is saved and temporary
2653 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(t1)));
2654 reg_saver.restore_live_registers(masm);
2655
2656 // We are back to the original state on entry and ready to go.
2657 __ jr(t1);
2658
2659 // Pending exception after the safepoint
2660
2661 __ bind(pending);
2662
2663 reg_saver.restore_live_registers(masm);
2664
2665 // exception pending => remove activation and forward to exception handler
2666
2667 __ sd(zr, Address(xthread, JavaThread::vm_result_offset()));
2668
2669 __ ld(x10, Address(xthread, Thread::pending_exception_offset()));
2670 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2671
2672 // -------------
2673 // make sure all code is generated
2674 masm->flush();
2675
2676 // return the blob
2677 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true);
2678 }
2679
2680 // Continuation point for throwing of implicit exceptions that are
2681 // not handled in the current activation. Fabricates an exception
2682 // oop and initiates normal exception dispatching in this
2683 // frame. Since we need to preserve callee-saved values (currently
2684 // only for C2, but done for C1 as well) we need a callee-saved oop
2685 // map and therefore have to make these stubs into RuntimeStubs
2686 // rather than BufferBlobs. If the compiler needs all registers to
2687 // be preserved between the fault point and the exception handler
2688 // then it must assume responsibility for that in
2689 // AbstractCompiler::continuation_for_implicit_null_exception or
2690 // continuation_for_implicit_division_by_zero_exception. All other
2691 // implicit exceptions (e.g., NullPointerException or
2692 // AbstractMethodError on entry) are either at call sites or
2693 // otherwise assume that stack unwinding will be initiated, so
2694 // caller saved registers were assumed volatile in the compiler.
2695
2696 RuntimeStub* SharedRuntime::generate_throw_exception(SharedStubId id, address runtime_entry) {
2697 assert(is_throw_id(id), "expected a throw stub id");
2698
2699 const char* name = SharedRuntime::stub_name(id);
2700
2701 // Information about frame layout at time of blocking runtime call.
2702 // Note that we only have to preserve callee-saved registers since
2703 // the compilers are responsible for supplying a continuation point
2704 // if they expect all registers to be preserved.
2705 // n.b. riscv asserts that frame::arg_reg_save_area_bytes == 0
2706 assert_cond(runtime_entry != nullptr);
2707 enum layout {
2708 fp_off = 0,
2709 fp_off2,
2710 return_off,
2711 return_off2,
2712 framesize // inclusive of return address
2713 };
2714
2715 const int insts_size = 1024;
2716 const int locs_size = 64;
2717
2718 ResourceMark rm;
2719 const char* timer_msg = "SharedRuntime generate_throw_exception";
2720 TraceTime timer(timer_msg, TRACETIME_LOG(Info, startuptime));
2721
2722 CodeBuffer code(name, insts_size, locs_size);
2723 OopMapSet* oop_maps = new OopMapSet();
2724 MacroAssembler* masm = new MacroAssembler(&code);
2725 assert_cond(oop_maps != nullptr && masm != nullptr);
2726
2727 address start = __ pc();
2728
2729 // This is an inlined and slightly modified version of call_VM
2730 // which has the ability to fetch the return PC out of
2731 // thread-local storage and also sets up last_Java_sp slightly
2732 // differently than the real call_VM
2733
2734 __ enter(); // Save FP and RA before call
2735
2736 assert(is_even(framesize / 2), "sp not 16-byte aligned");
2737
2738 // ra and fp are already in place
2739 __ subi(sp, fp, (unsigned)framesize << LogBytesPerInt); // prolog
2740
2741 int frame_complete = __ pc() - start;
2742
2743 // Set up last_Java_sp and last_Java_fp
2744 address the_pc = __ pc();
2745 __ set_last_Java_frame(sp, fp, the_pc, t0);
2746
2747 // Call runtime
2748 __ mv(c_rarg0, xthread);
2749 BLOCK_COMMENT("call runtime_entry");
2750 __ rt_call(runtime_entry);
2751
2752 // Generate oop map
2753 OopMap* map = new OopMap(framesize, 0);
2754 assert_cond(map != nullptr);
2755
2756 oop_maps->add_gc_map(the_pc - start, map);
2757
2758 __ reset_last_Java_frame(true);
2759
2760 __ leave();
2761
2762 // check for pending exceptions
2763 #ifdef ASSERT
2764 Label L;
2765 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2766 __ bnez(t0, L);
2767 __ should_not_reach_here();
2768 __ bind(L);
2769 #endif // ASSERT
2770 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2771
2772 // codeBlob framesize is in words (not VMRegImpl::slot_size)
2773 RuntimeStub* stub =
2774 RuntimeStub::new_runtime_stub(name,
2775 &code,
2776 frame_complete,
2777 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2778 oop_maps, false);
2779 assert(stub != nullptr, "create runtime stub fail!");
2780 return stub;
2781 }
2782
2783 #if INCLUDE_JFR
2784
2785 static void jfr_prologue(address the_pc, MacroAssembler* masm, Register thread) {
2786 __ set_last_Java_frame(sp, fp, the_pc, t0);
2787 __ mv(c_rarg0, thread);
2788 }
2789
2790 static void jfr_epilogue(MacroAssembler* masm) {
2791 __ reset_last_Java_frame(true);
2792 }
2793 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
2794 // It returns a jobject handle to the event writer.
2795 // The handle is dereferenced and the return value is the event writer oop.
2796 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
2797 enum layout {
2798 fp_off,
2799 fp_off2,
2800 return_off,
2801 return_off2,
2802 framesize // inclusive of return address
2803 };
2804
2805 int insts_size = 1024;
2806 int locs_size = 64;
2807 const char* name = SharedRuntime::stub_name(SharedStubId::jfr_write_checkpoint_id);
2808 CodeBuffer code(name, insts_size, locs_size);
2809 OopMapSet* oop_maps = new OopMapSet();
2810 MacroAssembler* masm = new MacroAssembler(&code);
2811
2812 address start = __ pc();
2813 __ enter();
2814 int frame_complete = __ pc() - start;
2815 address the_pc = __ pc();
2816 jfr_prologue(the_pc, masm, xthread);
2817 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), 1);
2818
2819 jfr_epilogue(masm);
2820 __ resolve_global_jobject(x10, t0, t1);
2821 __ leave();
2822 __ ret();
2823
2824 OopMap* map = new OopMap(framesize, 1);
2825 oop_maps->add_gc_map(the_pc - start, map);
2826
2827 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
2828 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
2829 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2830 oop_maps, false);
2831 return stub;
2832 }
2833
2834 // For c2: call to return a leased buffer.
2835 RuntimeStub* SharedRuntime::generate_jfr_return_lease() {
2836 enum layout {
2837 fp_off,
2838 fp_off2,
2839 return_off,
2840 return_off2,
2841 framesize // inclusive of return address
2842 };
2843
2844 int insts_size = 1024;
2845 int locs_size = 64;
2846 const char* name = SharedRuntime::stub_name(SharedStubId::jfr_return_lease_id);
2847 CodeBuffer code(name, insts_size, locs_size);
2848 OopMapSet* oop_maps = new OopMapSet();
2849 MacroAssembler* masm = new MacroAssembler(&code);
2850
2851 address start = __ pc();
2852 __ enter();
2853 int frame_complete = __ pc() - start;
2854 address the_pc = __ pc();
2855 jfr_prologue(the_pc, masm, xthread);
2856 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::return_lease), 1);
2857
2858 jfr_epilogue(masm);
2859 __ leave();
2860 __ ret();
2861
2862 OopMap* map = new OopMap(framesize, 1);
2863 oop_maps->add_gc_map(the_pc - start, map);
2864
2865 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
2866 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
2867 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2868 oop_maps, false);
2869 return stub;
2870 }
2871
2872 #endif // INCLUDE_JFR