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
600 void SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
601 int total_args_passed,
602 int comp_args_on_stack,
603 const BasicType *sig_bt,
604 const VMRegPair *regs,
605 AdapterHandlerEntry* handler) {
606 address i2c_entry = __ pc();
607 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
608
609 address c2i_unverified_entry = __ pc();
610 Label skip_fixup;
611
612 const Register receiver = j_rarg0;
613 const Register data = t0;
614
615 // -------------------------------------------------------------------------
616 // Generate a C2I adapter. On entry we know xmethod holds the Method* during calls
617 // to the interpreter. The args start out packed in the compiled layout. They
618 // need to be unpacked into the interpreter layout. This will almost always
619 // require some stack space. We grow the current (compiled) stack, then repack
620 // the args. We finally end in a jump to the generic interpreter entry point.
621 // On exit from the interpreter, the interpreter will restore our SP (lest the
622 // compiled code, which relies solely on SP and not FP, get sick).
623
624 {
625 __ block_comment("c2i_unverified_entry {");
626
627 __ ic_check();
628 __ ld(xmethod, Address(data, CompiledICData::speculated_method_offset()));
629
630 __ ld(t0, Address(xmethod, in_bytes(Method::code_offset())));
631 __ beqz(t0, skip_fixup);
632 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
633 __ block_comment("} c2i_unverified_entry");
634 }
635
636 address c2i_entry = __ pc();
637
638 // Class initialization barrier for static methods
639 address c2i_no_clinit_check_entry = nullptr;
640 if (VM_Version::supports_fast_class_init_checks()) {
641 Label L_skip_barrier;
642
643 { // Bypass the barrier for non-static methods
644 __ load_unsigned_short(t0, Address(xmethod, Method::access_flags_offset()));
645 __ test_bit(t1, t0, exact_log2(JVM_ACC_STATIC));
646 __ beqz(t1, L_skip_barrier); // non-static
647 }
648
649 __ load_method_holder(t1, xmethod);
650 __ clinit_barrier(t1, t0, &L_skip_barrier);
651 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
652
653 __ bind(L_skip_barrier);
654 c2i_no_clinit_check_entry = __ pc();
655 }
656
657 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
658 bs->c2i_entry_barrier(masm);
659
660 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup);
661
662 handler->set_entry_points(i2c_entry, c2i_entry, c2i_unverified_entry, c2i_no_clinit_check_entry);
663 return;
664 }
665
666 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
667 uint num_bits,
668 uint total_args_passed) {
669 assert(total_args_passed <= Argument::n_vector_register_parameters_c, "unsupported");
670 assert(num_bits >= 64 && num_bits <= 2048 && is_power_of_2(num_bits), "unsupported");
671
672 // check more info at https://github.com/riscv-non-isa/riscv-elf-psabi-doc/blob/master/riscv-cc.adoc
673 static const VectorRegister VEC_ArgReg[Argument::n_vector_register_parameters_c] = {
674 v8, v9, v10, v11, v12, v13, v14, v15,
675 v16, v17, v18, v19, v20, v21, v22, v23
676 };
677
678 const int next_reg_val = 3;
679 for (uint i = 0; i < total_args_passed; i++) {
680 VMReg vmreg = VEC_ArgReg[i]->as_VMReg();
681 regs[i].set_pair(vmreg->next(next_reg_val), vmreg);
682 }
683 return 0;
684 }
685
686 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
687 VMRegPair *regs,
688 int total_args_passed) {
689
690 // We return the amount of VMRegImpl stack slots we need to reserve for all
691 // the arguments NOT counting out_preserve_stack_slots.
692
693 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
694 c_rarg0, c_rarg1, c_rarg2, c_rarg3,
695 c_rarg4, c_rarg5, c_rarg6, c_rarg7
696 };
697 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
698 c_farg0, c_farg1, c_farg2, c_farg3,
699 c_farg4, c_farg5, c_farg6, c_farg7
700 };
701
702 uint int_args = 0;
703 uint fp_args = 0;
704 uint stk_args = 0; // inc by 2 each time
705
706 for (int i = 0; i < total_args_passed; i++) {
707 switch (sig_bt[i]) {
708 case T_BOOLEAN: // fall through
709 case T_CHAR: // fall through
710 case T_BYTE: // fall through
711 case T_SHORT: // fall through
712 case T_INT:
713 if (int_args < Argument::n_int_register_parameters_c) {
714 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
715 } else {
716 regs[i].set1(VMRegImpl::stack2reg(stk_args));
717 stk_args += 2;
718 }
719 break;
720 case T_LONG: // fall through
721 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
722 case T_OBJECT: // fall through
723 case T_ARRAY: // fall through
724 case T_ADDRESS: // fall through
725 case T_METADATA:
726 if (int_args < Argument::n_int_register_parameters_c) {
727 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
728 } else {
729 regs[i].set2(VMRegImpl::stack2reg(stk_args));
730 stk_args += 2;
731 }
732 break;
733 case T_FLOAT:
734 if (fp_args < Argument::n_float_register_parameters_c) {
735 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
736 } else if (int_args < Argument::n_int_register_parameters_c) {
737 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
738 } else {
739 regs[i].set1(VMRegImpl::stack2reg(stk_args));
740 stk_args += 2;
741 }
742 break;
743 case T_DOUBLE:
744 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
745 if (fp_args < Argument::n_float_register_parameters_c) {
746 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
747 } else if (int_args < Argument::n_int_register_parameters_c) {
748 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
749 } else {
750 regs[i].set2(VMRegImpl::stack2reg(stk_args));
751 stk_args += 2;
752 }
753 break;
754 case T_VOID: // Halves of longs and doubles
755 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
756 regs[i].set_bad();
757 break;
758 default:
759 ShouldNotReachHere();
760 }
761 }
762
763 return stk_args;
764 }
765
766 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
767 // We always ignore the frame_slots arg and just use the space just below frame pointer
768 // which by this time is free to use
769 switch (ret_type) {
770 case T_FLOAT:
771 __ fsw(f10, Address(fp, -3 * wordSize));
772 break;
773 case T_DOUBLE:
774 __ fsd(f10, Address(fp, -3 * wordSize));
775 break;
776 case T_VOID: break;
777 default: {
778 __ sd(x10, Address(fp, -3 * wordSize));
779 }
780 }
781 }
782
783 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
784 // We always ignore the frame_slots arg and just use the space just below frame pointer
785 // which by this time is free to use
786 switch (ret_type) {
787 case T_FLOAT:
788 __ flw(f10, Address(fp, -3 * wordSize));
789 break;
790 case T_DOUBLE:
791 __ fld(f10, Address(fp, -3 * wordSize));
792 break;
793 case T_VOID: break;
794 default: {
795 __ ld(x10, Address(fp, -3 * wordSize));
796 }
797 }
798 }
799
800 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
801 RegSet x;
802 for ( int i = first_arg ; i < arg_count ; i++ ) {
803 if (args[i].first()->is_Register()) {
804 x = x + args[i].first()->as_Register();
805 } else if (args[i].first()->is_FloatRegister()) {
806 __ subi(sp, sp, 2 * wordSize);
807 __ fsd(args[i].first()->as_FloatRegister(), Address(sp, 0));
808 }
809 }
810 __ push_reg(x, sp);
811 }
812
813 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
814 RegSet x;
815 for ( int i = first_arg ; i < arg_count ; i++ ) {
816 if (args[i].first()->is_Register()) {
817 x = x + args[i].first()->as_Register();
818 } else {
819 ;
820 }
821 }
822 __ pop_reg(x, sp);
823 for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) {
824 if (args[i].first()->is_Register()) {
825 ;
826 } else if (args[i].first()->is_FloatRegister()) {
827 __ fld(args[i].first()->as_FloatRegister(), Address(sp, 0));
828 __ addi(sp, sp, 2 * wordSize);
829 }
830 }
831 }
832
833 static void verify_oop_args(MacroAssembler* masm,
834 const methodHandle& method,
835 const BasicType* sig_bt,
836 const VMRegPair* regs) {
837 const Register temp_reg = x9; // not part of any compiled calling seq
838 if (VerifyOops) {
839 for (int i = 0; i < method->size_of_parameters(); i++) {
840 if (sig_bt[i] == T_OBJECT ||
841 sig_bt[i] == T_ARRAY) {
842 VMReg r = regs[i].first();
843 assert(r->is_valid(), "bad oop arg");
844 if (r->is_stack()) {
845 __ ld(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
846 __ verify_oop(temp_reg);
847 } else {
848 __ verify_oop(r->as_Register());
849 }
850 }
851 }
852 }
853 }
854
855 // on exit, sp points to the ContinuationEntry
856 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& stack_slots) {
857 assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
858 assert(in_bytes(ContinuationEntry::cont_offset()) % VMRegImpl::stack_slot_size == 0, "");
859 assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
860
861 stack_slots += (int)ContinuationEntry::size() / wordSize;
862 __ sub(sp, sp, (int)ContinuationEntry::size()); // place Continuation metadata
863
864 OopMap* map = new OopMap(((int)ContinuationEntry::size() + wordSize) / VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
865
866 __ ld(t0, Address(xthread, JavaThread::cont_entry_offset()));
867 __ sd(t0, Address(sp, ContinuationEntry::parent_offset()));
868 __ sd(sp, Address(xthread, JavaThread::cont_entry_offset()));
869
870 return map;
871 }
872
873 // on entry c_rarg1 points to the continuation
874 // sp points to ContinuationEntry
875 // c_rarg3 -- isVirtualThread
876 static void fill_continuation_entry(MacroAssembler* masm) {
877 #ifdef ASSERT
878 __ mv(t0, ContinuationEntry::cookie_value());
879 __ sw(t0, Address(sp, ContinuationEntry::cookie_offset()));
880 #endif
881
882 __ sd(c_rarg1, Address(sp, ContinuationEntry::cont_offset()));
883 __ sw(c_rarg3, Address(sp, ContinuationEntry::flags_offset()));
884 __ sd(zr, Address(sp, ContinuationEntry::chunk_offset()));
885 __ sw(zr, Address(sp, ContinuationEntry::argsize_offset()));
886 __ sw(zr, Address(sp, ContinuationEntry::pin_count_offset()));
887
888 __ ld(t0, Address(xthread, JavaThread::cont_fastpath_offset()));
889 __ sd(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
890 __ ld(t0, Address(xthread, JavaThread::held_monitor_count_offset()));
891 __ sd(t0, Address(sp, ContinuationEntry::parent_held_monitor_count_offset()));
892
893 __ sd(zr, Address(xthread, JavaThread::cont_fastpath_offset()));
894 __ sd(zr, Address(xthread, JavaThread::held_monitor_count_offset()));
895 }
896
897 // on entry, sp points to the ContinuationEntry
898 // on exit, fp points to the spilled fp + 2 * wordSize in the entry frame
899 static void continuation_enter_cleanup(MacroAssembler* masm) {
900 #ifndef PRODUCT
901 Label OK;
902 __ ld(t0, Address(xthread, JavaThread::cont_entry_offset()));
903 __ beq(sp, t0, OK);
904 __ stop("incorrect sp");
905 __ bind(OK);
906 #endif
907
908 __ ld(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
909 __ sd(t0, Address(xthread, JavaThread::cont_fastpath_offset()));
910
911 if (CheckJNICalls) {
912 // Check if this is a virtual thread continuation
913 Label L_skip_vthread_code;
914 __ lwu(t0, Address(sp, ContinuationEntry::flags_offset()));
915 __ beqz(t0, L_skip_vthread_code);
916
917 // If the held monitor count is > 0 and this vthread is terminating then
918 // it failed to release a JNI monitor. So we issue the same log message
919 // that JavaThread::exit does.
920 __ ld(t0, Address(xthread, JavaThread::jni_monitor_count_offset()));
921 __ beqz(t0, L_skip_vthread_code);
922
923 // Save return value potentially containing the exception oop in callee-saved x9
924 __ mv(x9, x10);
925 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::log_jni_monitor_still_held));
926 // Restore potential return value
927 __ mv(x10, x9);
928
929 // For vthreads we have to explicitly zero the JNI monitor count of the carrier
930 // on termination. The held count is implicitly zeroed below when we restore from
931 // the parent held count (which has to be zero).
932 __ sd(zr, Address(xthread, JavaThread::jni_monitor_count_offset()));
933
934 __ bind(L_skip_vthread_code);
935 }
936 #ifdef ASSERT
937 else {
938 // Check if this is a virtual thread continuation
939 Label L_skip_vthread_code;
940 __ lwu(t0, Address(sp, ContinuationEntry::flags_offset()));
941 __ beqz(t0, L_skip_vthread_code);
942
943 // See comment just above. If not checking JNI calls the JNI count is only
944 // needed for assertion checking.
945 __ sd(zr, Address(xthread, JavaThread::jni_monitor_count_offset()));
946
947 __ bind(L_skip_vthread_code);
948 }
949 #endif
950
951 __ ld(t0, Address(sp, ContinuationEntry::parent_held_monitor_count_offset()));
952 __ sd(t0, Address(xthread, JavaThread::held_monitor_count_offset()));
953
954 __ ld(t0, Address(sp, ContinuationEntry::parent_offset()));
955 __ sd(t0, Address(xthread, JavaThread::cont_entry_offset()));
956 __ add(fp, sp, (int)ContinuationEntry::size() + 2 * wordSize /* 2 extra words to match up with leave() */);
957 }
958
959 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
960 // On entry: c_rarg1 -- the continuation object
961 // c_rarg2 -- isContinue
962 // c_rarg3 -- isVirtualThread
963 static void gen_continuation_enter(MacroAssembler* masm,
964 const methodHandle& method,
965 const BasicType* sig_bt,
966 const VMRegPair* regs,
967 int& exception_offset,
968 OopMapSet*oop_maps,
969 int& frame_complete,
970 int& stack_slots,
971 int& interpreted_entry_offset,
972 int& compiled_entry_offset) {
973 // verify_oop_args(masm, method, sig_bt, regs);
974 Address resolve(SharedRuntime::get_resolve_static_call_stub(), relocInfo::static_call_type);
975
976 address start = __ pc();
977
978 Label call_thaw, exit;
979
980 // i2i entry used at interp_only_mode only
981 interpreted_entry_offset = __ pc() - start;
982 {
983 #ifdef ASSERT
984 Label is_interp_only;
985 __ lw(t0, Address(xthread, JavaThread::interp_only_mode_offset()));
986 __ bnez(t0, is_interp_only);
987 __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
988 __ bind(is_interp_only);
989 #endif
990
991 // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
992 __ ld(c_rarg1, Address(esp, Interpreter::stackElementSize * 2));
993 __ ld(c_rarg2, Address(esp, Interpreter::stackElementSize * 1));
994 __ ld(c_rarg3, Address(esp, Interpreter::stackElementSize * 0));
995 __ push_cont_fastpath(xthread);
996
997 __ enter();
998 stack_slots = 2; // will be adjusted in setup
999 OopMap* map = continuation_enter_setup(masm, stack_slots);
1000 // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
1001 // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
1002
1003 fill_continuation_entry(masm);
1004
1005 __ bnez(c_rarg2, call_thaw);
1006
1007 // Make sure the call is patchable
1008 __ align(NativeInstruction::instruction_size);
1009
1010 const address tr_call = __ reloc_call(resolve);
1011 if (tr_call == nullptr) {
1012 fatal("CodeCache is full at gen_continuation_enter");
1013 }
1014
1015 oop_maps->add_gc_map(__ pc() - start, map);
1016 __ post_call_nop();
1017
1018 __ j(exit);
1019
1020 address stub = CompiledDirectCall::emit_to_interp_stub(masm, tr_call);
1021 if (stub == nullptr) {
1022 fatal("CodeCache is full at gen_continuation_enter");
1023 }
1024 }
1025
1026 // compiled entry
1027 __ align(CodeEntryAlignment);
1028 compiled_entry_offset = __ pc() - start;
1029
1030 __ enter();
1031 stack_slots = 2; // will be adjusted in setup
1032 OopMap* map = continuation_enter_setup(masm, stack_slots);
1033 frame_complete = __ pc() - start;
1034
1035 fill_continuation_entry(masm);
1036
1037 __ bnez(c_rarg2, call_thaw);
1038
1039 // Make sure the call is patchable
1040 __ align(NativeInstruction::instruction_size);
1041
1042 const address tr_call = __ reloc_call(resolve);
1043 if (tr_call == nullptr) {
1044 fatal("CodeCache is full at gen_continuation_enter");
1045 }
1046
1047 oop_maps->add_gc_map(__ pc() - start, map);
1048 __ post_call_nop();
1049
1050 __ j(exit);
1051
1052 __ bind(call_thaw);
1053
1054 ContinuationEntry::_thaw_call_pc_offset = __ pc() - start;
1055 __ rt_call(CAST_FROM_FN_PTR(address, StubRoutines::cont_thaw()));
1056 oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1057 ContinuationEntry::_return_pc_offset = __ pc() - start;
1058 __ post_call_nop();
1059
1060 __ bind(exit);
1061 ContinuationEntry::_cleanup_offset = __ pc() - start;
1062 continuation_enter_cleanup(masm);
1063 __ leave();
1064 __ ret();
1065
1066 // exception handling
1067 exception_offset = __ pc() - start;
1068 {
1069 __ mv(x9, x10); // save return value contaning the exception oop in callee-saved x9
1070
1071 continuation_enter_cleanup(masm);
1072
1073 __ ld(c_rarg1, Address(fp, -1 * wordSize)); // return address
1074 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), xthread, c_rarg1);
1075
1076 // see OptoRuntime::generate_exception_blob: x10 -- exception oop, x13 -- exception pc
1077
1078 __ mv(x11, x10); // the exception handler
1079 __ mv(x10, x9); // restore return value contaning the exception oop
1080 __ verify_oop(x10);
1081
1082 __ leave();
1083 __ mv(x13, ra);
1084 __ jr(x11); // the exception handler
1085 }
1086
1087 address stub = CompiledDirectCall::emit_to_interp_stub(masm, tr_call);
1088 if (stub == nullptr) {
1089 fatal("CodeCache is full at gen_continuation_enter");
1090 }
1091 }
1092
1093 static void gen_continuation_yield(MacroAssembler* masm,
1094 const methodHandle& method,
1095 const BasicType* sig_bt,
1096 const VMRegPair* regs,
1097 OopMapSet* oop_maps,
1098 int& frame_complete,
1099 int& stack_slots,
1100 int& compiled_entry_offset) {
1101 enum layout {
1102 fp_off,
1103 fp_off2,
1104 return_off,
1105 return_off2,
1106 framesize // inclusive of return address
1107 };
1108 // assert(is_even(framesize/2), "sp not 16-byte aligned");
1109
1110 stack_slots = framesize / VMRegImpl::slots_per_word;
1111 assert(stack_slots == 2, "recheck layout");
1112
1113 address start = __ pc();
1114
1115 compiled_entry_offset = __ pc() - start;
1116 __ enter();
1117
1118 __ mv(c_rarg1, sp);
1119
1120 frame_complete = __ pc() - start;
1121 address the_pc = __ pc();
1122
1123 __ 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
1124
1125 __ mv(c_rarg0, xthread);
1126 __ set_last_Java_frame(sp, fp, the_pc, t0);
1127 __ call_VM_leaf(Continuation::freeze_entry(), 2);
1128 __ reset_last_Java_frame(true);
1129
1130 Label pinned;
1131
1132 __ bnez(x10, pinned);
1133
1134 // We've succeeded, set sp to the ContinuationEntry
1135 __ ld(sp, Address(xthread, JavaThread::cont_entry_offset()));
1136 continuation_enter_cleanup(masm);
1137
1138 __ bind(pinned); // pinned -- return to caller
1139
1140 // handle pending exception thrown by freeze
1141 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1142 Label ok;
1143 __ beqz(t0, ok);
1144 __ leave();
1145 __ j(RuntimeAddress(StubRoutines::forward_exception_entry()));
1146 __ bind(ok);
1147
1148 __ leave();
1149 __ ret();
1150
1151 OopMap* map = new OopMap(framesize, 1);
1152 oop_maps->add_gc_map(the_pc - start, map);
1153 }
1154
1155 void SharedRuntime::continuation_enter_cleanup(MacroAssembler* masm) {
1156 ::continuation_enter_cleanup(masm);
1157 }
1158
1159 static void gen_special_dispatch(MacroAssembler* masm,
1160 const methodHandle& method,
1161 const BasicType* sig_bt,
1162 const VMRegPair* regs) {
1163 verify_oop_args(masm, method, sig_bt, regs);
1164 vmIntrinsics::ID iid = method->intrinsic_id();
1165
1166 // Now write the args into the outgoing interpreter space
1167 bool has_receiver = false;
1168 Register receiver_reg = noreg;
1169 int member_arg_pos = -1;
1170 Register member_reg = noreg;
1171 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1172 if (ref_kind != 0) {
1173 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1174 member_reg = x9; // known to be free at this point
1175 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1176 } else if (iid == vmIntrinsics::_invokeBasic) {
1177 has_receiver = true;
1178 } else if (iid == vmIntrinsics::_linkToNative) {
1179 member_arg_pos = method->size_of_parameters() - 1; // trailing NativeEntryPoint argument
1180 member_reg = x9; // known to be free at this point
1181 } else {
1182 fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1183 }
1184
1185 if (member_reg != noreg) {
1186 // Load the member_arg into register, if necessary.
1187 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1188 VMReg r = regs[member_arg_pos].first();
1189 if (r->is_stack()) {
1190 __ ld(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1191 } else {
1192 // no data motion is needed
1193 member_reg = r->as_Register();
1194 }
1195 }
1196
1197 if (has_receiver) {
1198 // Make sure the receiver is loaded into a register.
1199 assert(method->size_of_parameters() > 0, "oob");
1200 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1201 VMReg r = regs[0].first();
1202 assert(r->is_valid(), "bad receiver arg");
1203 if (r->is_stack()) {
1204 // Porting note: This assumes that compiled calling conventions always
1205 // pass the receiver oop in a register. If this is not true on some
1206 // platform, pick a temp and load the receiver from stack.
1207 fatal("receiver always in a register");
1208 receiver_reg = x12; // known to be free at this point
1209 __ ld(receiver_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1210 } else {
1211 // no data motion is needed
1212 receiver_reg = r->as_Register();
1213 }
1214 }
1215
1216 // Figure out which address we are really jumping to:
1217 MethodHandles::generate_method_handle_dispatch(masm, iid,
1218 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1219 }
1220
1221 // ---------------------------------------------------------------------------
1222 // Generate a native wrapper for a given method. The method takes arguments
1223 // in the Java compiled code convention, marshals them to the native
1224 // convention (handlizes oops, etc), transitions to native, makes the call,
1225 // returns to java state (possibly blocking), unhandlizes any result and
1226 // returns.
1227 //
1228 // Critical native functions are a shorthand for the use of
1229 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1230 // functions. The wrapper is expected to unpack the arguments before
1231 // passing them to the callee and perform checks before and after the
1232 // native call to ensure that they GCLocker
1233 // lock_critical/unlock_critical semantics are followed. Some other
1234 // parts of JNI setup are skipped like the tear down of the JNI handle
1235 // block and the check for pending exceptions it's impossible for them
1236 // to be thrown.
1237 //
1238 // They are roughly structured like this:
1239 // if (GCLocker::needs_gc()) SharedRuntime::block_for_jni_critical()
1240 // tranistion to thread_in_native
1241 // unpack array arguments and call native entry point
1242 // check for safepoint in progress
1243 // check if any thread suspend flags are set
1244 // call into JVM and possible unlock the JNI critical
1245 // if a GC was suppressed while in the critical native.
1246 // transition back to thread_in_Java
1247 // return to caller
1248 //
1249 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1250 const methodHandle& method,
1251 int compile_id,
1252 BasicType* in_sig_bt,
1253 VMRegPair* in_regs,
1254 BasicType ret_type) {
1255 if (method->is_continuation_native_intrinsic()) {
1256 int exception_offset = -1;
1257 OopMapSet* oop_maps = new OopMapSet();
1258 int frame_complete = -1;
1259 int stack_slots = -1;
1260 int interpreted_entry_offset = -1;
1261 int vep_offset = -1;
1262 if (method->is_continuation_enter_intrinsic()) {
1263 gen_continuation_enter(masm,
1264 method,
1265 in_sig_bt,
1266 in_regs,
1267 exception_offset,
1268 oop_maps,
1269 frame_complete,
1270 stack_slots,
1271 interpreted_entry_offset,
1272 vep_offset);
1273 } else if (method->is_continuation_yield_intrinsic()) {
1274 gen_continuation_yield(masm,
1275 method,
1276 in_sig_bt,
1277 in_regs,
1278 oop_maps,
1279 frame_complete,
1280 stack_slots,
1281 vep_offset);
1282 } else {
1283 guarantee(false, "Unknown Continuation native intrinsic");
1284 }
1285
1286 #ifdef ASSERT
1287 if (method->is_continuation_enter_intrinsic()) {
1288 assert(interpreted_entry_offset != -1, "Must be set");
1289 assert(exception_offset != -1, "Must be set");
1290 } else {
1291 assert(interpreted_entry_offset == -1, "Must be unset");
1292 assert(exception_offset == -1, "Must be unset");
1293 }
1294 assert(frame_complete != -1, "Must be set");
1295 assert(stack_slots != -1, "Must be set");
1296 assert(vep_offset != -1, "Must be set");
1297 #endif
1298
1299 __ flush();
1300 nmethod* nm = nmethod::new_native_nmethod(method,
1301 compile_id,
1302 masm->code(),
1303 vep_offset,
1304 frame_complete,
1305 stack_slots,
1306 in_ByteSize(-1),
1307 in_ByteSize(-1),
1308 oop_maps,
1309 exception_offset);
1310 if (nm == nullptr) return nm;
1311 if (method->is_continuation_enter_intrinsic()) {
1312 ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
1313 } else if (method->is_continuation_yield_intrinsic()) {
1314 _cont_doYield_stub = nm;
1315 } else {
1316 guarantee(false, "Unknown Continuation native intrinsic");
1317 }
1318 return nm;
1319 }
1320
1321 if (method->is_method_handle_intrinsic()) {
1322 vmIntrinsics::ID iid = method->intrinsic_id();
1323 intptr_t start = (intptr_t)__ pc();
1324 int vep_offset = ((intptr_t)__ pc()) - start;
1325
1326 // First instruction must be a nop as it may need to be patched on deoptimisation
1327 {
1328 Assembler::IncompressibleScope scope(masm); // keep the nop as 4 bytes for patching.
1329 MacroAssembler::assert_alignment(__ pc());
1330 __ nop(); // 4 bytes
1331 }
1332 gen_special_dispatch(masm,
1333 method,
1334 in_sig_bt,
1335 in_regs);
1336 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
1337 __ flush();
1338 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
1339 return nmethod::new_native_nmethod(method,
1340 compile_id,
1341 masm->code(),
1342 vep_offset,
1343 frame_complete,
1344 stack_slots / VMRegImpl::slots_per_word,
1345 in_ByteSize(-1),
1346 in_ByteSize(-1),
1347 (OopMapSet*)nullptr);
1348 }
1349 address native_func = method->native_function();
1350 assert(native_func != nullptr, "must have function");
1351
1352 // An OopMap for lock (and class if static)
1353 OopMapSet *oop_maps = new OopMapSet();
1354 assert_cond(oop_maps != nullptr);
1355 intptr_t start = (intptr_t)__ pc();
1356
1357 // We have received a description of where all the java arg are located
1358 // on entry to the wrapper. We need to convert these args to where
1359 // the jni function will expect them. To figure out where they go
1360 // we convert the java signature to a C signature by inserting
1361 // the hidden arguments as arg[0] and possibly arg[1] (static method)
1362
1363 const int total_in_args = method->size_of_parameters();
1364 int total_c_args = total_in_args + (method->is_static() ? 2 : 1);
1365
1366 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1367 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1368
1369 int argc = 0;
1370 out_sig_bt[argc++] = T_ADDRESS;
1371 if (method->is_static()) {
1372 out_sig_bt[argc++] = T_OBJECT;
1373 }
1374
1375 for (int i = 0; i < total_in_args ; i++) {
1376 out_sig_bt[argc++] = in_sig_bt[i];
1377 }
1378
1379 // Now figure out where the args must be stored and how much stack space
1380 // they require.
1381 int out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
1382
1383 // Compute framesize for the wrapper. We need to handlize all oops in
1384 // incoming registers
1385
1386 // Calculate the total number of stack slots we will need.
1387
1388 // First count the abi requirement plus all of the outgoing args
1389 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1390
1391 // Now the space for the inbound oop handle area
1392 int total_save_slots = 8 * VMRegImpl::slots_per_word; // 8 arguments passed in registers
1393
1394 int oop_handle_offset = stack_slots;
1395 stack_slots += total_save_slots;
1396
1397 // Now any space we need for handlizing a klass if static method
1398
1399 int klass_slot_offset = 0;
1400 int klass_offset = -1;
1401 int lock_slot_offset = 0;
1402 bool is_static = false;
1403
1404 if (method->is_static()) {
1405 klass_slot_offset = stack_slots;
1406 stack_slots += VMRegImpl::slots_per_word;
1407 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1408 is_static = true;
1409 }
1410
1411 // Plus a lock if needed
1412
1413 if (method->is_synchronized()) {
1414 lock_slot_offset = stack_slots;
1415 stack_slots += VMRegImpl::slots_per_word;
1416 }
1417
1418 // Now a place (+2) to save return values or temp during shuffling
1419 // + 4 for return address (which we own) and saved fp
1420 stack_slots += 6;
1421
1422 // Ok The space we have allocated will look like:
1423 //
1424 //
1425 // FP-> | |
1426 // | 2 slots (ra) |
1427 // | 2 slots (fp) |
1428 // |---------------------|
1429 // | 2 slots for moves |
1430 // |---------------------|
1431 // | lock box (if sync) |
1432 // |---------------------| <- lock_slot_offset
1433 // | klass (if static) |
1434 // |---------------------| <- klass_slot_offset
1435 // | oopHandle area |
1436 // |---------------------| <- oop_handle_offset (8 java arg registers)
1437 // | outbound memory |
1438 // | based arguments |
1439 // | |
1440 // |---------------------|
1441 // | |
1442 // SP-> | out_preserved_slots |
1443 //
1444 //
1445
1446
1447 // Now compute actual number of stack words we need rounding to make
1448 // stack properly aligned.
1449 stack_slots = align_up(stack_slots, StackAlignmentInSlots);
1450
1451 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1452
1453 // First thing make an ic check to see if we should even be here
1454
1455 // We are free to use all registers as temps without saving them and
1456 // restoring them except fp. fp is the only callee save register
1457 // as far as the interpreter and the compiler(s) are concerned.
1458
1459 const Register receiver = j_rarg0;
1460
1461 __ verify_oop(receiver);
1462 assert_different_registers(receiver, t0, t1);
1463
1464 __ ic_check();
1465
1466 int vep_offset = ((intptr_t)__ pc()) - start;
1467
1468 // If we have to make this method not-entrant we'll overwrite its
1469 // first instruction with a jump.
1470 {
1471 Assembler::IncompressibleScope scope(masm); // keep the nop as 4 bytes for patching.
1472 MacroAssembler::assert_alignment(__ pc());
1473 __ nop(); // 4 bytes
1474 }
1475
1476 if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) {
1477 Label L_skip_barrier;
1478 __ mov_metadata(t1, method->method_holder()); // InstanceKlass*
1479 __ clinit_barrier(t1, t0, &L_skip_barrier);
1480 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
1481
1482 __ bind(L_skip_barrier);
1483 }
1484
1485 // Generate stack overflow check
1486 __ bang_stack_with_offset(checked_cast<int>(StackOverflow::stack_shadow_zone_size()));
1487
1488 // Generate a new frame for the wrapper.
1489 __ enter();
1490 // -2 because return address is already present and so is saved fp
1491 __ sub(sp, sp, stack_size - 2 * wordSize);
1492
1493 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1494 assert_cond(bs != nullptr);
1495 bs->nmethod_entry_barrier(masm, nullptr /* slow_path */, nullptr /* continuation */, nullptr /* guard */);
1496
1497 // Frame is now completed as far as size and linkage.
1498 int frame_complete = ((intptr_t)__ pc()) - start;
1499
1500 // We use x18 as the oop handle for the receiver/klass
1501 // It is callee save so it survives the call to native
1502
1503 const Register oop_handle_reg = x18;
1504
1505 //
1506 // We immediately shuffle the arguments so that any vm call we have to
1507 // make from here on out (sync slow path, jvmti, etc.) we will have
1508 // captured the oops from our caller and have a valid oopMap for
1509 // them.
1510
1511 // -----------------
1512 // The Grand Shuffle
1513
1514 // The Java calling convention is either equal (linux) or denser (win64) than the
1515 // c calling convention. However the because of the jni_env argument the c calling
1516 // convention always has at least one more (and two for static) arguments than Java.
1517 // Therefore if we move the args from java -> c backwards then we will never have
1518 // a register->register conflict and we don't have to build a dependency graph
1519 // and figure out how to break any cycles.
1520 //
1521
1522 // Record esp-based slot for receiver on stack for non-static methods
1523 int receiver_offset = -1;
1524
1525 // This is a trick. We double the stack slots so we can claim
1526 // the oops in the caller's frame. Since we are sure to have
1527 // more args than the caller doubling is enough to make
1528 // sure we can capture all the incoming oop args from the
1529 // caller.
1530 //
1531 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1532 assert_cond(map != nullptr);
1533
1534 int float_args = 0;
1535 int int_args = 0;
1536
1537 #ifdef ASSERT
1538 bool reg_destroyed[Register::number_of_registers];
1539 bool freg_destroyed[FloatRegister::number_of_registers];
1540 for ( int r = 0 ; r < Register::number_of_registers ; r++ ) {
1541 reg_destroyed[r] = false;
1542 }
1543 for ( int f = 0 ; f < FloatRegister::number_of_registers ; f++ ) {
1544 freg_destroyed[f] = false;
1545 }
1546
1547 #endif /* ASSERT */
1548
1549 // For JNI natives the incoming and outgoing registers are offset upwards.
1550 GrowableArray<int> arg_order(2 * total_in_args);
1551
1552 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) {
1553 arg_order.push(i);
1554 arg_order.push(c_arg);
1555 }
1556
1557 for (int ai = 0; ai < arg_order.length(); ai += 2) {
1558 int i = arg_order.at(ai);
1559 int c_arg = arg_order.at(ai + 1);
1560 __ block_comment(err_msg("mv %d -> %d", i, c_arg));
1561 assert(c_arg != -1 && i != -1, "wrong order");
1562 #ifdef ASSERT
1563 if (in_regs[i].first()->is_Register()) {
1564 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!");
1565 } else if (in_regs[i].first()->is_FloatRegister()) {
1566 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!");
1567 }
1568 if (out_regs[c_arg].first()->is_Register()) {
1569 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
1570 } else if (out_regs[c_arg].first()->is_FloatRegister()) {
1571 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true;
1572 }
1573 #endif /* ASSERT */
1574 switch (in_sig_bt[i]) {
1575 case T_ARRAY:
1576 case T_OBJECT:
1577 __ object_move(map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
1578 ((i == 0) && (!is_static)),
1579 &receiver_offset);
1580 int_args++;
1581 break;
1582 case T_VOID:
1583 break;
1584
1585 case T_FLOAT:
1586 __ float_move(in_regs[i], out_regs[c_arg]);
1587 float_args++;
1588 break;
1589
1590 case T_DOUBLE:
1591 assert( i + 1 < total_in_args &&
1592 in_sig_bt[i + 1] == T_VOID &&
1593 out_sig_bt[c_arg + 1] == T_VOID, "bad arg list");
1594 __ double_move(in_regs[i], out_regs[c_arg]);
1595 float_args++;
1596 break;
1597
1598 case T_LONG :
1599 __ long_move(in_regs[i], out_regs[c_arg]);
1600 int_args++;
1601 break;
1602
1603 case T_ADDRESS:
1604 assert(false, "found T_ADDRESS in java args");
1605 break;
1606
1607 default:
1608 __ move32_64(in_regs[i], out_regs[c_arg]);
1609 int_args++;
1610 }
1611 }
1612
1613 // point c_arg at the first arg that is already loaded in case we
1614 // need to spill before we call out
1615 int c_arg = total_c_args - total_in_args;
1616
1617 // Pre-load a static method's oop into c_rarg1.
1618 if (method->is_static()) {
1619
1620 // load oop into a register
1621 __ movoop(c_rarg1,
1622 JNIHandles::make_local(method->method_holder()->java_mirror()));
1623
1624 // Now handlize the static class mirror it's known not-null.
1625 __ sd(c_rarg1, Address(sp, klass_offset));
1626 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
1627
1628 // Now get the handle
1629 __ la(c_rarg1, Address(sp, klass_offset));
1630 // and protect the arg if we must spill
1631 c_arg--;
1632 }
1633
1634 // Change state to native (we save the return address in the thread, since it might not
1635 // be pushed on the stack when we do a stack traversal). It is enough that the pc()
1636 // points into the right code segment. It does not have to be the correct return pc.
1637 // We use the same pc/oopMap repeatedly when we call out.
1638
1639 Label native_return;
1640 if (LockingMode != LM_LEGACY && method->is_object_wait0()) {
1641 // For convenience we use the pc we want to resume to in case of preemption on Object.wait.
1642 __ set_last_Java_frame(sp, noreg, native_return, t0);
1643 } else {
1644 intptr_t the_pc = (intptr_t) __ pc();
1645 oop_maps->add_gc_map(the_pc - start, map);
1646
1647 __ set_last_Java_frame(sp, noreg, __ pc(), t0);
1648 }
1649
1650 Label dtrace_method_entry, dtrace_method_entry_done;
1651 if (DTraceMethodProbes) {
1652 __ j(dtrace_method_entry);
1653 __ bind(dtrace_method_entry_done);
1654 }
1655
1656 // RedefineClasses() tracing support for obsolete method entry
1657 if (log_is_enabled(Trace, redefine, class, obsolete)) {
1658 // protect the args we've loaded
1659 save_args(masm, total_c_args, c_arg, out_regs);
1660 __ mov_metadata(c_rarg1, method());
1661 __ call_VM_leaf(
1662 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
1663 xthread, c_rarg1);
1664 restore_args(masm, total_c_args, c_arg, out_regs);
1665 }
1666
1667 // Lock a synchronized method
1668
1669 // Register definitions used by locking and unlocking
1670
1671 const Register swap_reg = x10;
1672 const Register obj_reg = x9; // Will contain the oop
1673 const Register lock_reg = x30; // Address of compiler lock object (BasicLock)
1674 const Register old_hdr = x30; // value of old header at unlock time
1675 const Register lock_tmp = x31; // Temporary used by lightweight_lock/unlock
1676 const Register tmp = ra;
1677
1678 Label slow_path_lock;
1679 Label lock_done;
1680
1681 if (method->is_synchronized()) {
1682 Label count;
1683
1684 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes();
1685
1686 // Get the handle (the 2nd argument)
1687 __ mv(oop_handle_reg, c_rarg1);
1688
1689 // Get address of the box
1690
1691 __ la(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1692
1693 // Load the oop from the handle
1694 __ ld(obj_reg, Address(oop_handle_reg, 0));
1695
1696 if (LockingMode == LM_MONITOR) {
1697 __ j(slow_path_lock);
1698 } else if (LockingMode == LM_LEGACY) {
1699 // Load (object->mark() | 1) into swap_reg % x10
1700 __ ld(t0, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
1701 __ ori(swap_reg, t0, 1);
1702
1703 // Save (object->mark() | 1) into BasicLock's displaced header
1704 __ sd(swap_reg, Address(lock_reg, mark_word_offset));
1705
1706 // src -> dest if dest == x10 else x10 <- dest
1707 __ cmpxchg_obj_header(x10, lock_reg, obj_reg, lock_tmp, count, /*fallthrough*/nullptr);
1708
1709 // Test if the oopMark is an obvious stack pointer, i.e.,
1710 // 1) (mark & 3) == 0, and
1711 // 2) sp <= mark < mark + os::pagesize()
1712 // These 3 tests can be done by evaluating the following
1713 // expression: ((mark - sp) & (3 - os::vm_page_size())),
1714 // assuming both stack pointer and pagesize have their
1715 // least significant 2 bits clear.
1716 // NOTE: the oopMark is in swap_reg % 10 as the result of cmpxchg
1717
1718 __ sub(swap_reg, swap_reg, sp);
1719 __ mv(t0, 3 - (int)os::vm_page_size());
1720 __ andr(swap_reg, swap_reg, t0);
1721
1722 // Save the test result, for recursive case, the result is zero
1723 __ sd(swap_reg, Address(lock_reg, mark_word_offset));
1724 __ bnez(swap_reg, slow_path_lock);
1725
1726 __ bind(count);
1727 __ inc_held_monitor_count(t0);
1728 } else {
1729 assert(LockingMode == LM_LIGHTWEIGHT, "must be");
1730 __ lightweight_lock(lock_reg, obj_reg, swap_reg, tmp, lock_tmp, slow_path_lock);
1731 }
1732
1733 // Slow path will re-enter here
1734 __ bind(lock_done);
1735 }
1736
1737
1738 // Finally just about ready to make the JNI call
1739
1740 // get JNIEnv* which is first argument to native
1741 __ la(c_rarg0, Address(xthread, in_bytes(JavaThread::jni_environment_offset())));
1742
1743 // Now set thread in native
1744 __ la(t1, Address(xthread, JavaThread::thread_state_offset()));
1745 __ mv(t0, _thread_in_native);
1746 __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore);
1747 __ sw(t0, Address(t1));
1748
1749 // Clobbers t1
1750 __ rt_call(native_func);
1751
1752 // Verify or restore cpu control state after JNI call
1753 __ restore_cpu_control_state_after_jni(t0);
1754
1755 // Unpack native results.
1756 if (ret_type != T_OBJECT && ret_type != T_ARRAY) {
1757 __ cast_primitive_type(ret_type, x10);
1758 }
1759
1760 Label safepoint_in_progress, safepoint_in_progress_done;
1761
1762 // Switch thread to "native transition" state before reading the synchronization state.
1763 // This additional state is necessary because reading and testing the synchronization
1764 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1765 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1766 // VM thread changes sync state to synchronizing and suspends threads for GC.
1767 // Thread A is resumed to finish this native method, but doesn't block here since it
1768 // didn't see any synchronization is progress, and escapes.
1769 __ mv(t0, _thread_in_native_trans);
1770
1771 __ sw(t0, Address(xthread, JavaThread::thread_state_offset()));
1772
1773 // Force this write out before the read below
1774 if (!UseSystemMemoryBarrier) {
1775 __ membar(MacroAssembler::AnyAny);
1776 }
1777
1778 // check for safepoint operation in progress and/or pending suspend requests
1779 {
1780 __ safepoint_poll(safepoint_in_progress, true /* at_return */, false /* in_nmethod */);
1781 __ lwu(t0, Address(xthread, JavaThread::suspend_flags_offset()));
1782 __ bnez(t0, safepoint_in_progress);
1783 __ bind(safepoint_in_progress_done);
1784 }
1785
1786 // change thread state
1787 __ la(t1, Address(xthread, JavaThread::thread_state_offset()));
1788 __ mv(t0, _thread_in_Java);
1789 __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore);
1790 __ sw(t0, Address(t1));
1791
1792 if (LockingMode != LM_LEGACY && method->is_object_wait0()) {
1793 // Check preemption for Object.wait()
1794 __ ld(t1, Address(xthread, JavaThread::preempt_alternate_return_offset()));
1795 __ beqz(t1, native_return);
1796 __ sd(zr, Address(xthread, JavaThread::preempt_alternate_return_offset()));
1797 __ jr(t1);
1798 __ bind(native_return);
1799
1800 intptr_t the_pc = (intptr_t) __ pc();
1801 oop_maps->add_gc_map(the_pc - start, map);
1802 }
1803
1804 Label reguard;
1805 Label reguard_done;
1806 __ lbu(t0, Address(xthread, JavaThread::stack_guard_state_offset()));
1807 __ mv(t1, StackOverflow::stack_guard_yellow_reserved_disabled);
1808 __ beq(t0, t1, reguard);
1809 __ bind(reguard_done);
1810
1811 // native result if any is live
1812
1813 // Unlock
1814 Label unlock_done;
1815 Label slow_path_unlock;
1816 if (method->is_synchronized()) {
1817
1818 // Get locked oop from the handle we passed to jni
1819 __ ld(obj_reg, Address(oop_handle_reg, 0));
1820
1821 Label done, not_recursive;
1822
1823 if (LockingMode == LM_LEGACY) {
1824 // Simple recursive lock?
1825 __ ld(t0, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1826 __ bnez(t0, not_recursive);
1827 __ dec_held_monitor_count(t0);
1828 __ j(done);
1829 }
1830
1831 __ bind(not_recursive);
1832
1833 // Must save x10 if if it is live now because cmpxchg must use it
1834 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1835 save_native_result(masm, ret_type, stack_slots);
1836 }
1837
1838 if (LockingMode == LM_MONITOR) {
1839 __ j(slow_path_unlock);
1840 } else if (LockingMode == LM_LEGACY) {
1841 // get address of the stack lock
1842 __ la(x10, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1843 // get old displaced header
1844 __ ld(old_hdr, Address(x10, 0));
1845
1846 // Atomic swap old header if oop still contains the stack lock
1847 Label count;
1848 __ cmpxchg_obj_header(x10, old_hdr, obj_reg, lock_tmp, count, &slow_path_unlock);
1849 __ bind(count);
1850 __ dec_held_monitor_count(t0);
1851 } else {
1852 assert(LockingMode == LM_LIGHTWEIGHT, "");
1853 __ lightweight_unlock(obj_reg, old_hdr, swap_reg, lock_tmp, slow_path_unlock);
1854 }
1855
1856 // slow path re-enters here
1857 __ bind(unlock_done);
1858 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1859 restore_native_result(masm, ret_type, stack_slots);
1860 }
1861
1862 __ bind(done);
1863 }
1864
1865 Label dtrace_method_exit, dtrace_method_exit_done;
1866 if (DTraceMethodProbes) {
1867 __ j(dtrace_method_exit);
1868 __ bind(dtrace_method_exit_done);
1869 }
1870
1871 __ reset_last_Java_frame(false);
1872
1873 // Unbox oop result, e.g. JNIHandles::resolve result.
1874 if (is_reference_type(ret_type)) {
1875 __ resolve_jobject(x10, x11, x12);
1876 }
1877
1878 if (CheckJNICalls) {
1879 // clear_pending_jni_exception_check
1880 __ sd(zr, Address(xthread, JavaThread::pending_jni_exception_check_fn_offset()));
1881 }
1882
1883 // reset handle block
1884 __ ld(x12, Address(xthread, JavaThread::active_handles_offset()));
1885 __ sd(zr, Address(x12, JNIHandleBlock::top_offset()));
1886
1887 __ leave();
1888
1889 #if INCLUDE_JFR
1890 // We need to do a poll test after unwind in case the sampler
1891 // managed to sample the native frame after returning to Java.
1892 Label L_return;
1893 __ ld(t0, Address(xthread, JavaThread::polling_word_offset()));
1894 address poll_test_pc = __ pc();
1895 __ relocate(relocInfo::poll_return_type);
1896 __ test_bit(t0, t0, log2i_exact(SafepointMechanism::poll_bit()));
1897 __ beqz(t0, L_return);
1898 assert(SharedRuntime::polling_page_return_handler_blob() != nullptr,
1899 "polling page return stub not created yet");
1900 address stub = SharedRuntime::polling_page_return_handler_blob()->entry_point();
1901 __ la(t0, InternalAddress(poll_test_pc));
1902 __ sd(t0, Address(xthread, JavaThread::saved_exception_pc_offset()));
1903 __ far_jump(RuntimeAddress(stub));
1904 __ bind(L_return);
1905 #endif // INCLUDE_JFR
1906
1907 // Any exception pending?
1908 Label exception_pending;
1909 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1910 __ bnez(t0, exception_pending);
1911
1912 // We're done
1913 __ ret();
1914
1915 // Unexpected paths are out of line and go here
1916
1917 // forward the exception
1918 __ bind(exception_pending);
1919
1920 // and forward the exception
1921 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
1922
1923 // Slow path locking & unlocking
1924 if (method->is_synchronized()) {
1925
1926 __ block_comment("Slow path lock {");
1927 __ bind(slow_path_lock);
1928
1929 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM
1930 // args are (oop obj, BasicLock* lock, JavaThread* thread)
1931
1932 // protect the args we've loaded
1933 save_args(masm, total_c_args, c_arg, out_regs);
1934
1935 __ mv(c_rarg0, obj_reg);
1936 __ mv(c_rarg1, lock_reg);
1937 __ mv(c_rarg2, xthread);
1938
1939 // Not a leaf but we have last_Java_frame setup as we want.
1940 // We don't want to unmount in case of contention since that would complicate preserving
1941 // the arguments that had already been marshalled into the native convention. So we force
1942 // the freeze slow path to find this native wrapper frame (see recurse_freeze_native_frame())
1943 // and pin the vthread. Otherwise the fast path won't find it since we don't walk the stack.
1944 __ push_cont_fastpath();
1945 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3);
1946 __ pop_cont_fastpath();
1947 restore_args(masm, total_c_args, c_arg, out_regs);
1948
1949 #ifdef ASSERT
1950 { Label L;
1951 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1952 __ beqz(t0, L);
1953 __ stop("no pending exception allowed on exit from monitorenter");
1954 __ bind(L);
1955 }
1956 #endif
1957 __ j(lock_done);
1958
1959 __ block_comment("} Slow path lock");
1960
1961 __ block_comment("Slow path unlock {");
1962 __ bind(slow_path_unlock);
1963
1964 if (ret_type == T_FLOAT || ret_type == T_DOUBLE) {
1965 save_native_result(masm, ret_type, stack_slots);
1966 }
1967
1968 __ mv(c_rarg2, xthread);
1969 __ la(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1970 __ mv(c_rarg0, obj_reg);
1971
1972 // Save pending exception around call to VM (which contains an EXCEPTION_MARK)
1973 // NOTE that obj_reg == x9 currently
1974 __ ld(x9, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1975 __ sd(zr, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1976
1977 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C));
1978
1979 #ifdef ASSERT
1980 {
1981 Label L;
1982 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1983 __ beqz(t0, L);
1984 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C");
1985 __ bind(L);
1986 }
1987 #endif /* ASSERT */
1988
1989 __ sd(x9, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1990
1991 if (ret_type == T_FLOAT || ret_type == T_DOUBLE) {
1992 restore_native_result(masm, ret_type, stack_slots);
1993 }
1994 __ j(unlock_done);
1995
1996 __ block_comment("} Slow path unlock");
1997
1998 } // synchronized
1999
2000 // SLOW PATH Reguard the stack if needed
2001
2002 __ bind(reguard);
2003 save_native_result(masm, ret_type, stack_slots);
2004 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2005 restore_native_result(masm, ret_type, stack_slots);
2006 // and continue
2007 __ j(reguard_done);
2008
2009 // SLOW PATH safepoint
2010 {
2011 __ block_comment("safepoint {");
2012 __ bind(safepoint_in_progress);
2013
2014 // Don't use call_VM as it will see a possible pending exception and forward it
2015 // and never return here preventing us from clearing _last_native_pc down below.
2016 //
2017 save_native_result(masm, ret_type, stack_slots);
2018 __ mv(c_rarg0, xthread);
2019 #ifndef PRODUCT
2020 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
2021 #endif
2022 __ rt_call(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans));
2023
2024 // Restore any method result value
2025 restore_native_result(masm, ret_type, stack_slots);
2026
2027 __ j(safepoint_in_progress_done);
2028 __ block_comment("} safepoint");
2029 }
2030
2031 // SLOW PATH dtrace support
2032 if (DTraceMethodProbes) {
2033 {
2034 __ block_comment("dtrace entry {");
2035 __ bind(dtrace_method_entry);
2036
2037 // We have all of the arguments setup at this point. We must not touch any register
2038 // argument registers at this point (what if we save/restore them there are no oop?
2039
2040 save_args(masm, total_c_args, c_arg, out_regs);
2041 __ mov_metadata(c_rarg1, method());
2042 __ call_VM_leaf(
2043 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
2044 xthread, c_rarg1);
2045 restore_args(masm, total_c_args, c_arg, out_regs);
2046 __ j(dtrace_method_entry_done);
2047 __ block_comment("} dtrace entry");
2048 }
2049
2050 {
2051 __ block_comment("dtrace exit {");
2052 __ bind(dtrace_method_exit);
2053 save_native_result(masm, ret_type, stack_slots);
2054 __ mov_metadata(c_rarg1, method());
2055 __ call_VM_leaf(
2056 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
2057 xthread, c_rarg1);
2058 restore_native_result(masm, ret_type, stack_slots);
2059 __ j(dtrace_method_exit_done);
2060 __ block_comment("} dtrace exit");
2061 }
2062 }
2063
2064 __ flush();
2065
2066 nmethod *nm = nmethod::new_native_nmethod(method,
2067 compile_id,
2068 masm->code(),
2069 vep_offset,
2070 frame_complete,
2071 stack_slots / VMRegImpl::slots_per_word,
2072 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2073 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size),
2074 oop_maps);
2075 assert(nm != nullptr, "create native nmethod fail!");
2076 return nm;
2077 }
2078
2079 // this function returns the adjust size (in number of words) to a c2i adapter
2080 // activation for use during deoptimization
2081 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2082 assert(callee_locals >= callee_parameters,
2083 "test and remove; got more parms than locals");
2084 if (callee_locals < callee_parameters) {
2085 return 0; // No adjustment for negative locals
2086 }
2087 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords;
2088 // diff is counted in stack words
2089 return align_up(diff, 2);
2090 }
2091
2092 //------------------------------generate_deopt_blob----------------------------
2093 void SharedRuntime::generate_deopt_blob() {
2094 // Allocate space for the code
2095 ResourceMark rm;
2096 // Setup code generation tools
2097 int pad = 0;
2098 #if INCLUDE_JVMCI
2099 if (EnableJVMCI) {
2100 pad += 512; // Increase the buffer size when compiling for JVMCI
2101 }
2102 #endif
2103 const char* name = SharedRuntime::stub_name(StubId::shared_deopt_id);
2104 CodeBuffer buffer(name, 2048 + pad, 1024);
2105 MacroAssembler* masm = new MacroAssembler(&buffer);
2106 int frame_size_in_words = -1;
2107 OopMap* map = nullptr;
2108 OopMapSet *oop_maps = new OopMapSet();
2109 assert_cond(masm != nullptr && oop_maps != nullptr);
2110 RegisterSaver reg_saver(COMPILER2_OR_JVMCI != 0);
2111
2112 // -------------
2113 // This code enters when returning to a de-optimized nmethod. A return
2114 // address has been pushed on the stack, and return values are in
2115 // registers.
2116 // If we are doing a normal deopt then we were called from the patched
2117 // nmethod from the point we returned to the nmethod. So the return
2118 // address on the stack is wrong by NativeCall::instruction_size
2119 // We will adjust the value so it looks like we have the original return
2120 // address on the stack (like when we eagerly deoptimized).
2121 // In the case of an exception pending when deoptimizing, we enter
2122 // with a return address on the stack that points after the call we patched
2123 // into the exception handler. We have the following register state from,
2124 // e.g., the forward exception stub (see stubGenerator_riscv.cpp).
2125 // x10: exception oop
2126 // x9: exception handler
2127 // x13: throwing pc
2128 // So in this case we simply jam x13 into the useless return address and
2129 // the stack looks just like we want.
2130 //
2131 // At this point we need to de-opt. We save the argument return
2132 // registers. We call the first C routine, fetch_unroll_info(). This
2133 // routine captures the return values and returns a structure which
2134 // describes the current frame size and the sizes of all replacement frames.
2135 // The current frame is compiled code and may contain many inlined
2136 // functions, each with their own JVM state. We pop the current frame, then
2137 // push all the new frames. Then we call the C routine unpack_frames() to
2138 // populate these frames. Finally unpack_frames() returns us the new target
2139 // address. Notice that callee-save registers are BLOWN here; they have
2140 // already been captured in the vframeArray at the time the return PC was
2141 // patched.
2142 address start = __ pc();
2143 Label cont;
2144
2145 // Prolog for non exception case!
2146
2147 // Save everything in sight.
2148 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2149
2150 // Normal deoptimization. Save exec mode for unpack_frames.
2151 __ mv(xcpool, Deoptimization::Unpack_deopt); // callee-saved
2152 __ j(cont);
2153
2154 int reexecute_offset = __ pc() - start;
2155 #if INCLUDE_JVMCI && !defined(COMPILER1)
2156 if (UseJVMCICompiler) {
2157 // JVMCI does not use this kind of deoptimization
2158 __ should_not_reach_here();
2159 }
2160 #endif
2161
2162 // Reexecute case
2163 // return address is the pc describes what bci to do re-execute at
2164
2165 // No need to update map as each call to save_live_registers will produce identical oopmap
2166 (void) reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2167
2168 __ mv(xcpool, Deoptimization::Unpack_reexecute); // callee-saved
2169 __ j(cont);
2170
2171 #if INCLUDE_JVMCI
2172 Label after_fetch_unroll_info_call;
2173 int implicit_exception_uncommon_trap_offset = 0;
2174 int uncommon_trap_offset = 0;
2175
2176 if (EnableJVMCI) {
2177 implicit_exception_uncommon_trap_offset = __ pc() - start;
2178
2179 __ ld(ra, Address(xthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2180 __ sd(zr, Address(xthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2181
2182 uncommon_trap_offset = __ pc() - start;
2183
2184 // Save everything in sight.
2185 reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2186 // fetch_unroll_info needs to call last_java_frame()
2187 Label retaddr;
2188 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2189
2190 __ lw(c_rarg1, Address(xthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2191 __ mv(t0, -1);
2192 __ sw(t0, Address(xthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2193
2194 __ mv(xcpool, Deoptimization::Unpack_reexecute);
2195 __ mv(c_rarg0, xthread);
2196 __ orrw(c_rarg2, zr, xcpool); // exec mode
2197 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap));
2198 __ bind(retaddr);
2199 oop_maps->add_gc_map( __ pc()-start, map->deep_copy());
2200
2201 __ reset_last_Java_frame(false);
2202
2203 __ j(after_fetch_unroll_info_call);
2204 } // EnableJVMCI
2205 #endif // INCLUDE_JVMCI
2206
2207 int exception_offset = __ pc() - start;
2208
2209 // Prolog for exception case
2210
2211 // all registers are dead at this entry point, except for x10, and
2212 // x13 which contain the exception oop and exception pc
2213 // respectively. Set them in TLS and fall thru to the
2214 // unpack_with_exception_in_tls entry point.
2215
2216 __ sd(x13, Address(xthread, JavaThread::exception_pc_offset()));
2217 __ sd(x10, Address(xthread, JavaThread::exception_oop_offset()));
2218
2219 int exception_in_tls_offset = __ pc() - start;
2220
2221 // new implementation because exception oop is now passed in JavaThread
2222
2223 // Prolog for exception case
2224 // All registers must be preserved because they might be used by LinearScan
2225 // Exceptiop oop and throwing PC are passed in JavaThread
2226 // tos: stack at point of call to method that threw the exception (i.e. only
2227 // args are on the stack, no return address)
2228
2229 // The return address pushed by save_live_registers will be patched
2230 // later with the throwing pc. The correct value is not available
2231 // now because loading it from memory would destroy registers.
2232
2233 // NB: The SP at this point must be the SP of the method that is
2234 // being deoptimized. Deoptimization assumes that the frame created
2235 // here by save_live_registers is immediately below the method's SP.
2236 // This is a somewhat fragile mechanism.
2237
2238 // Save everything in sight.
2239 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2240
2241 // Now it is safe to overwrite any register
2242
2243 // Deopt during an exception. Save exec mode for unpack_frames.
2244 __ mv(xcpool, Deoptimization::Unpack_exception); // callee-saved
2245
2246 // load throwing pc from JavaThread and patch it as the return address
2247 // of the current frame. Then clear the field in JavaThread
2248
2249 __ ld(x13, Address(xthread, JavaThread::exception_pc_offset()));
2250 __ sd(x13, Address(fp, frame::return_addr_offset * wordSize));
2251 __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2252
2253 #ifdef ASSERT
2254 // verify that there is really an exception oop in JavaThread
2255 __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2256 __ verify_oop(x10);
2257
2258 // verify that there is no pending exception
2259 Label no_pending_exception;
2260 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2261 __ beqz(t0, no_pending_exception);
2262 __ stop("must not have pending exception here");
2263 __ bind(no_pending_exception);
2264 #endif
2265
2266 __ bind(cont);
2267
2268 // Call C code. Need thread and this frame, but NOT official VM entry
2269 // crud. We cannot block on this call, no GC can happen.
2270 //
2271 // UnrollBlock* fetch_unroll_info(JavaThread* thread)
2272
2273 // fetch_unroll_info needs to call last_java_frame().
2274
2275 Label retaddr;
2276 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2277 #ifdef ASSERT
2278 {
2279 Label L;
2280 __ ld(t0, Address(xthread,
2281 JavaThread::last_Java_fp_offset()));
2282 __ beqz(t0, L);
2283 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared");
2284 __ bind(L);
2285 }
2286 #endif // ASSERT
2287 __ mv(c_rarg0, xthread);
2288 __ mv(c_rarg1, xcpool);
2289 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info));
2290 __ bind(retaddr);
2291
2292 // Need to have an oopmap that tells fetch_unroll_info where to
2293 // find any register it might need.
2294 oop_maps->add_gc_map(__ pc() - start, map);
2295
2296 __ reset_last_Java_frame(false);
2297
2298 #if INCLUDE_JVMCI
2299 if (EnableJVMCI) {
2300 __ bind(after_fetch_unroll_info_call);
2301 }
2302 #endif
2303
2304 // Load UnrollBlock* into x15
2305 __ mv(x15, x10);
2306
2307 __ lwu(xcpool, Address(x15, Deoptimization::UnrollBlock::unpack_kind_offset()));
2308 Label noException;
2309 __ mv(t0, Deoptimization::Unpack_exception);
2310 __ bne(xcpool, t0, noException); // Was exception pending?
2311 __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2312 __ ld(x13, Address(xthread, JavaThread::exception_pc_offset()));
2313 __ sd(zr, Address(xthread, JavaThread::exception_oop_offset()));
2314 __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2315
2316 __ verify_oop(x10);
2317
2318 // Overwrite the result registers with the exception results.
2319 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2320
2321 __ bind(noException);
2322
2323 // Only register save data is on the stack.
2324 // Now restore the result registers. Everything else is either dead
2325 // or captured in the vframeArray.
2326
2327 // Restore fp result register
2328 __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2329 // Restore integer result register
2330 __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2331
2332 // Pop all of the register save area off the stack
2333 __ add(sp, sp, frame_size_in_words * wordSize);
2334
2335 // All of the register save area has been popped of the stack. Only the
2336 // return address remains.
2337
2338 // Pop all the frames we must move/replace.
2339 //
2340 // Frame picture (youngest to oldest)
2341 // 1: self-frame (no frame link)
2342 // 2: deopting frame (no frame link)
2343 // 3: caller of deopting frame (could be compiled/interpreted).
2344 //
2345 // Note: by leaving the return address of self-frame on the stack
2346 // and using the size of frame 2 to adjust the stack
2347 // when we are done the return to frame 3 will still be on the stack.
2348
2349 // Pop deoptimized frame
2350 __ lwu(x12, Address(x15, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset()));
2351 __ subi(x12, x12, 2 * wordSize);
2352 __ add(sp, sp, x12);
2353 __ ld(fp, Address(sp, 0));
2354 __ ld(ra, Address(sp, wordSize));
2355 __ addi(sp, sp, 2 * wordSize);
2356 // RA should now be the return address to the caller (3)
2357
2358 #ifdef ASSERT
2359 // Compilers generate code that bang the stack by as much as the
2360 // interpreter would need. So this stack banging should never
2361 // trigger a fault. Verify that it does not on non product builds.
2362 __ lwu(x9, Address(x15, Deoptimization::UnrollBlock::total_frame_sizes_offset()));
2363 __ bang_stack_size(x9, x12);
2364 #endif
2365 // Load address of array of frame pcs into x12
2366 __ ld(x12, Address(x15, Deoptimization::UnrollBlock::frame_pcs_offset()));
2367
2368 // Load address of array of frame sizes into x14
2369 __ ld(x14, Address(x15, Deoptimization::UnrollBlock::frame_sizes_offset()));
2370
2371 // Load counter into x13
2372 __ lwu(x13, Address(x15, Deoptimization::UnrollBlock::number_of_frames_offset()));
2373
2374 // Now adjust the caller's stack to make up for the extra locals
2375 // but record the original sp so that we can save it in the skeletal interpreter
2376 // frame and the stack walking of interpreter_sender will get the unextended sp
2377 // value and not the "real" sp value.
2378
2379 const Register sender_sp = x16;
2380
2381 __ mv(sender_sp, sp);
2382 __ lwu(x9, Address(x15,
2383 Deoptimization::UnrollBlock::
2384 caller_adjustment_offset()));
2385 __ sub(sp, sp, x9);
2386
2387 // Push interpreter frames in a loop
2388 __ mv(t0, 0xDEADDEAD); // Make a recognizable pattern
2389 __ mv(t1, t0);
2390 Label loop;
2391 __ bind(loop);
2392 __ ld(x9, Address(x14, 0)); // Load frame size
2393 __ addi(x14, x14, wordSize);
2394 __ subi(x9, x9, 2 * wordSize); // We'll push pc and fp by hand
2395 __ ld(ra, Address(x12, 0)); // Load pc
2396 __ addi(x12, x12, wordSize);
2397 __ enter(); // Save old & set new fp
2398 __ sub(sp, sp, x9); // Prolog
2399 // This value is corrected by layout_activation_impl
2400 __ sd(zr, Address(fp, frame::interpreter_frame_last_sp_offset * wordSize));
2401 __ sd(sender_sp, Address(fp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable
2402 __ mv(sender_sp, sp); // Pass sender_sp to next frame
2403 __ subi(x13, x13, 1); // Decrement counter
2404 __ bnez(x13, loop);
2405
2406 // Re-push self-frame
2407 __ ld(ra, Address(x12));
2408 __ enter();
2409
2410 // Allocate a full sized register save area. We subtract 2 because
2411 // enter() just pushed 2 words
2412 __ sub(sp, sp, (frame_size_in_words - 2) * wordSize);
2413
2414 // Restore frame locals after moving the frame
2415 __ fsd(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2416 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2417
2418 // Call C code. Need thread but NOT official VM entry
2419 // crud. We cannot block on this call, no GC can happen. Call should
2420 // restore return values to their stack-slots with the new SP.
2421 //
2422 // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)
2423
2424 // Use fp because the frames look interpreted now
2425 // Don't need the precise return PC here, just precise enough to point into this code blob.
2426 address the_pc = __ pc();
2427 __ set_last_Java_frame(sp, fp, the_pc, t0);
2428
2429 __ mv(c_rarg0, xthread);
2430 __ mv(c_rarg1, xcpool); // second arg: exec_mode
2431 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames));
2432
2433 // Set an oopmap for the call site
2434 // Use the same PC we used for the last java frame
2435 oop_maps->add_gc_map(the_pc - start,
2436 new OopMap(frame_size_in_words, 0));
2437
2438 // Clear fp AND pc
2439 __ reset_last_Java_frame(true);
2440
2441 // Collect return values
2442 __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2443 __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2444
2445 // Pop self-frame.
2446 __ leave(); // Epilog
2447
2448 // Jump to interpreter
2449 __ ret();
2450
2451 // Make sure all code is generated
2452 masm->flush();
2453
2454 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words);
2455 assert(_deopt_blob != nullptr, "create deoptimization blob fail!");
2456 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
2457 #if INCLUDE_JVMCI
2458 if (EnableJVMCI) {
2459 _deopt_blob->set_uncommon_trap_offset(uncommon_trap_offset);
2460 _deopt_blob->set_implicit_exception_uncommon_trap_offset(implicit_exception_uncommon_trap_offset);
2461 }
2462 #endif
2463 }
2464
2465 // Number of stack slots between incoming argument block and the start of
2466 // a new frame. The PROLOG must add this many slots to the stack. The
2467 // EPILOG must remove this many slots.
2468 // RISCV needs two words for RA (return address) and FP (frame pointer).
2469 uint SharedRuntime::in_preserve_stack_slots() {
2470 return 2 * VMRegImpl::slots_per_word;
2471 }
2472
2473 uint SharedRuntime::out_preserve_stack_slots() {
2474 return 0;
2475 }
2476
2477 VMReg SharedRuntime::thread_register() {
2478 return xthread->as_VMReg();
2479 }
2480
2481 //------------------------------generate_handler_blob------
2482 //
2483 // Generate a special Compile2Runtime blob that saves all registers,
2484 // and setup oopmap.
2485 //
2486 SafepointBlob* SharedRuntime::generate_handler_blob(StubId id, address call_ptr) {
2487 assert(is_polling_page_id(id), "expected a polling page stub id");
2488
2489 ResourceMark rm;
2490 OopMapSet *oop_maps = new OopMapSet();
2491 assert_cond(oop_maps != nullptr);
2492 OopMap* map = nullptr;
2493
2494 // Allocate space for the code. Setup code generation tools.
2495 const char* name = SharedRuntime::stub_name(id);
2496 CodeBuffer buffer(name, 2048, 1024);
2497 MacroAssembler* masm = new MacroAssembler(&buffer);
2498 assert_cond(masm != nullptr);
2499
2500 address start = __ pc();
2501 address call_pc = nullptr;
2502 int frame_size_in_words = -1;
2503 bool cause_return = (id == StubId::shared_polling_page_return_handler_id);
2504 RegisterSaver reg_saver(id == StubId::shared_polling_page_vectors_safepoint_handler_id /* save_vectors */);
2505
2506 // Save Integer and Float registers.
2507 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2508
2509 // The following is basically a call_VM. However, we need the precise
2510 // address of the call in order to generate an oopmap. Hence, we do all the
2511 // work ourselves.
2512
2513 Label retaddr;
2514 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2515
2516 // The return address must always be correct so that frame constructor never
2517 // sees an invalid pc.
2518
2519 if (!cause_return) {
2520 // overwrite the return address pushed by save_live_registers
2521 // Additionally, x18 is a callee-saved register so we can look at
2522 // it later to determine if someone changed the return address for
2523 // us!
2524 __ ld(x18, Address(xthread, JavaThread::saved_exception_pc_offset()));
2525 __ sd(x18, Address(fp, frame::return_addr_offset * wordSize));
2526 }
2527
2528 // Do the call
2529 __ mv(c_rarg0, xthread);
2530 __ rt_call(call_ptr);
2531 __ bind(retaddr);
2532
2533 // Set an oopmap for the call site. This oopmap will map all
2534 // oop-registers and debug-info registers as callee-saved. This
2535 // will allow deoptimization at this safepoint to find all possible
2536 // debug-info recordings, as well as let GC find all oops.
2537
2538 oop_maps->add_gc_map( __ pc() - start, map);
2539
2540 Label noException;
2541
2542 __ reset_last_Java_frame(false);
2543
2544 __ membar(MacroAssembler::LoadLoad | MacroAssembler::LoadStore);
2545
2546 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2547 __ beqz(t0, noException);
2548
2549 // Exception pending
2550
2551 reg_saver.restore_live_registers(masm);
2552
2553 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2554
2555 // No exception case
2556 __ bind(noException);
2557
2558 Label no_adjust, bail;
2559 if (!cause_return) {
2560 // If our stashed return pc was modified by the runtime we avoid touching it
2561 __ ld(t0, Address(fp, frame::return_addr_offset * wordSize));
2562 __ bne(x18, t0, no_adjust);
2563
2564 #ifdef ASSERT
2565 // Verify the correct encoding of the poll we're about to skip.
2566 // See NativeInstruction::is_lwu_to_zr()
2567 __ lwu(t0, Address(x18));
2568 __ andi(t1, t0, 0b1111111);
2569 __ mv(t2, 0b0000011);
2570 __ bne(t1, t2, bail); // 0-6:0b0000011
2571 __ srli(t1, t0, 7);
2572 __ andi(t1, t1, 0b11111);
2573 __ bnez(t1, bail); // 7-11:0b00000
2574 __ srli(t1, t0, 12);
2575 __ andi(t1, t1, 0b111);
2576 __ mv(t2, 0b110);
2577 __ bne(t1, t2, bail); // 12-14:0b110
2578 #endif
2579
2580 // Adjust return pc forward to step over the safepoint poll instruction
2581 __ addi(x18, x18, NativeInstruction::instruction_size);
2582 __ sd(x18, Address(fp, frame::return_addr_offset * wordSize));
2583 }
2584
2585 __ bind(no_adjust);
2586 // Normal exit, restore registers and exit.
2587
2588 reg_saver.restore_live_registers(masm);
2589 __ ret();
2590
2591 #ifdef ASSERT
2592 __ bind(bail);
2593 __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected");
2594 #endif
2595
2596 // Make sure all code is generated
2597 masm->flush();
2598
2599 // Fill-out other meta info
2600 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words);
2601 }
2602
2603 //
2604 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
2605 //
2606 // Generate a stub that calls into vm to find out the proper destination
2607 // of a java call. All the argument registers are live at this point
2608 // but since this is generic code we don't know what they are and the caller
2609 // must do any gc of the args.
2610 //
2611 RuntimeStub* SharedRuntime::generate_resolve_blob(StubId id, address destination) {
2612 assert(StubRoutines::forward_exception_entry() != nullptr, "must be generated before");
2613 assert(is_resolve_id(id), "expected a resolve stub id");
2614
2615 // allocate space for the code
2616 ResourceMark rm;
2617
2618 const char* name = SharedRuntime::stub_name(id);
2619 CodeBuffer buffer(name, 1000, 512);
2620 MacroAssembler* masm = new MacroAssembler(&buffer);
2621 assert_cond(masm != nullptr);
2622
2623 int frame_size_in_words = -1;
2624 RegisterSaver reg_saver(false /* save_vectors */);
2625
2626 OopMapSet *oop_maps = new OopMapSet();
2627 assert_cond(oop_maps != nullptr);
2628 OopMap* map = nullptr;
2629
2630 int start = __ offset();
2631
2632 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2633
2634 int frame_complete = __ offset();
2635
2636 {
2637 Label retaddr;
2638 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2639
2640 __ mv(c_rarg0, xthread);
2641 __ rt_call(destination);
2642 __ bind(retaddr);
2643 }
2644
2645 // Set an oopmap for the call site.
2646 // We need this not only for callee-saved registers, but also for volatile
2647 // registers that the compiler might be keeping live across a safepoint.
2648
2649 oop_maps->add_gc_map( __ offset() - start, map);
2650
2651 // x10 contains the address we are going to jump to assuming no exception got installed
2652
2653 // clear last_Java_sp
2654 __ reset_last_Java_frame(false);
2655 // check for pending exceptions
2656 Label pending;
2657 __ ld(t1, Address(xthread, Thread::pending_exception_offset()));
2658 __ bnez(t1, pending);
2659
2660 // get the returned Method*
2661 __ get_vm_result_metadata(xmethod, xthread);
2662 __ sd(xmethod, Address(sp, reg_saver.reg_offset_in_bytes(xmethod)));
2663
2664 // x10 is where we want to jump, overwrite t1 which is saved and temporary
2665 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(t1)));
2666 reg_saver.restore_live_registers(masm);
2667
2668 // We are back to the original state on entry and ready to go.
2669 __ jr(t1);
2670
2671 // Pending exception after the safepoint
2672
2673 __ bind(pending);
2674
2675 reg_saver.restore_live_registers(masm);
2676
2677 // exception pending => remove activation and forward to exception handler
2678
2679 __ sd(zr, Address(xthread, JavaThread::vm_result_oop_offset()));
2680
2681 __ ld(x10, Address(xthread, Thread::pending_exception_offset()));
2682 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2683
2684 // -------------
2685 // make sure all code is generated
2686 masm->flush();
2687
2688 // return the blob
2689 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true);
2690 }
2691
2692 // Continuation point for throwing of implicit exceptions that are
2693 // not handled in the current activation. Fabricates an exception
2694 // oop and initiates normal exception dispatching in this
2695 // frame. Since we need to preserve callee-saved values (currently
2696 // only for C2, but done for C1 as well) we need a callee-saved oop
2697 // map and therefore have to make these stubs into RuntimeStubs
2698 // rather than BufferBlobs. If the compiler needs all registers to
2699 // be preserved between the fault point and the exception handler
2700 // then it must assume responsibility for that in
2701 // AbstractCompiler::continuation_for_implicit_null_exception or
2702 // continuation_for_implicit_division_by_zero_exception. All other
2703 // implicit exceptions (e.g., NullPointerException or
2704 // AbstractMethodError on entry) are either at call sites or
2705 // otherwise assume that stack unwinding will be initiated, so
2706 // caller saved registers were assumed volatile in the compiler.
2707
2708 RuntimeStub* SharedRuntime::generate_throw_exception(StubId id, address runtime_entry) {
2709 assert(is_throw_id(id), "expected a throw stub id");
2710
2711 const char* name = SharedRuntime::stub_name(id);
2712
2713 // Information about frame layout at time of blocking runtime call.
2714 // Note that we only have to preserve callee-saved registers since
2715 // the compilers are responsible for supplying a continuation point
2716 // if they expect all registers to be preserved.
2717 // n.b. riscv asserts that frame::arg_reg_save_area_bytes == 0
2718 assert_cond(runtime_entry != nullptr);
2719 enum layout {
2720 fp_off = 0,
2721 fp_off2,
2722 return_off,
2723 return_off2,
2724 framesize // inclusive of return address
2725 };
2726
2727 const int insts_size = 1024;
2728 const int locs_size = 64;
2729
2730 ResourceMark rm;
2731 const char* timer_msg = "SharedRuntime generate_throw_exception";
2732 TraceTime timer(timer_msg, TRACETIME_LOG(Info, startuptime));
2733
2734 CodeBuffer code(name, insts_size, locs_size);
2735 OopMapSet* oop_maps = new OopMapSet();
2736 MacroAssembler* masm = new MacroAssembler(&code);
2737 assert_cond(oop_maps != nullptr && masm != nullptr);
2738
2739 address start = __ pc();
2740
2741 // This is an inlined and slightly modified version of call_VM
2742 // which has the ability to fetch the return PC out of
2743 // thread-local storage and also sets up last_Java_sp slightly
2744 // differently than the real call_VM
2745
2746 __ enter(); // Save FP and RA before call
2747
2748 assert(is_even(framesize / 2), "sp not 16-byte aligned");
2749
2750 // ra and fp are already in place
2751 __ subi(sp, fp, (unsigned)framesize << LogBytesPerInt); // prolog
2752
2753 int frame_complete = __ pc() - start;
2754
2755 // Set up last_Java_sp and last_Java_fp
2756 address the_pc = __ pc();
2757 __ set_last_Java_frame(sp, fp, the_pc, t0);
2758
2759 // Call runtime
2760 __ mv(c_rarg0, xthread);
2761 BLOCK_COMMENT("call runtime_entry");
2762 __ rt_call(runtime_entry);
2763
2764 // Generate oop map
2765 OopMap* map = new OopMap(framesize, 0);
2766 assert_cond(map != nullptr);
2767
2768 oop_maps->add_gc_map(the_pc - start, map);
2769
2770 __ reset_last_Java_frame(true);
2771
2772 __ leave();
2773
2774 // check for pending exceptions
2775 #ifdef ASSERT
2776 Label L;
2777 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2778 __ bnez(t0, L);
2779 __ should_not_reach_here();
2780 __ bind(L);
2781 #endif // ASSERT
2782 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2783
2784 // codeBlob framesize is in words (not VMRegImpl::slot_size)
2785 RuntimeStub* stub =
2786 RuntimeStub::new_runtime_stub(name,
2787 &code,
2788 frame_complete,
2789 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2790 oop_maps, false);
2791 assert(stub != nullptr, "create runtime stub fail!");
2792 return stub;
2793 }
2794
2795 #if INCLUDE_JFR
2796
2797 static void jfr_prologue(address the_pc, MacroAssembler* masm, Register thread) {
2798 __ set_last_Java_frame(sp, fp, the_pc, t0);
2799 __ mv(c_rarg0, thread);
2800 }
2801
2802 static void jfr_epilogue(MacroAssembler* masm) {
2803 __ reset_last_Java_frame(true);
2804 }
2805 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
2806 // It returns a jobject handle to the event writer.
2807 // The handle is dereferenced and the return value is the event writer oop.
2808 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
2809 enum layout {
2810 fp_off,
2811 fp_off2,
2812 return_off,
2813 return_off2,
2814 framesize // inclusive of return address
2815 };
2816
2817 int insts_size = 1024;
2818 int locs_size = 64;
2819 const char* name = SharedRuntime::stub_name(StubId::shared_jfr_write_checkpoint_id);
2820 CodeBuffer code(name, insts_size, locs_size);
2821 OopMapSet* oop_maps = new OopMapSet();
2822 MacroAssembler* masm = new MacroAssembler(&code);
2823
2824 address start = __ pc();
2825 __ enter();
2826 int frame_complete = __ pc() - start;
2827 address the_pc = __ pc();
2828 jfr_prologue(the_pc, masm, xthread);
2829 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), 1);
2830
2831 jfr_epilogue(masm);
2832 __ resolve_global_jobject(x10, t0, t1);
2833 __ leave();
2834 __ ret();
2835
2836 OopMap* map = new OopMap(framesize, 1);
2837 oop_maps->add_gc_map(the_pc - start, map);
2838
2839 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
2840 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
2841 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2842 oop_maps, false);
2843 return stub;
2844 }
2845
2846 // For c2: call to return a leased buffer.
2847 RuntimeStub* SharedRuntime::generate_jfr_return_lease() {
2848 enum layout {
2849 fp_off,
2850 fp_off2,
2851 return_off,
2852 return_off2,
2853 framesize // inclusive of return address
2854 };
2855
2856 int insts_size = 1024;
2857 int locs_size = 64;
2858 const char* name = SharedRuntime::stub_name(StubId::shared_jfr_return_lease_id);
2859 CodeBuffer code(name, insts_size, locs_size);
2860 OopMapSet* oop_maps = new OopMapSet();
2861 MacroAssembler* masm = new MacroAssembler(&code);
2862
2863 address start = __ pc();
2864 __ enter();
2865 int frame_complete = __ pc() - start;
2866 address the_pc = __ pc();
2867 jfr_prologue(the_pc, masm, xthread);
2868 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::return_lease), 1);
2869
2870 jfr_epilogue(masm);
2871 __ leave();
2872 __ ret();
2873
2874 OopMap* map = new OopMap(framesize, 1);
2875 oop_maps->add_gc_map(the_pc - start, map);
2876
2877 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
2878 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
2879 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2880 oop_maps, false);
2881 return stub;
2882 }
2883
2884 #endif // INCLUDE_JFR