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 address entry_address[AdapterBlob::ENTRY_COUNT]) {
606 entry_address[AdapterBlob::I2C] = __ pc();
607 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
608
609 entry_address[AdapterBlob::C2I_Unverified] = __ 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 entry_address[AdapterBlob::C2I] = __ pc();
637
638 // Class initialization barrier for static methods
639 entry_address[AdapterBlob::C2I_No_Clinit_Check] = 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 entry_address[AdapterBlob::C2I_No_Clinit_Check] = __ 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 return;
662 }
663
664 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
665 uint num_bits,
666 uint total_args_passed) {
667 assert(total_args_passed <= Argument::n_vector_register_parameters_c, "unsupported");
668 assert(num_bits >= 64 && num_bits <= 2048 && is_power_of_2(num_bits), "unsupported");
669
670 // check more info at https://github.com/riscv-non-isa/riscv-elf-psabi-doc/blob/master/riscv-cc.adoc
671 static const VectorRegister VEC_ArgReg[Argument::n_vector_register_parameters_c] = {
672 v8, v9, v10, v11, v12, v13, v14, v15,
673 v16, v17, v18, v19, v20, v21, v22, v23
674 };
675
676 const int next_reg_val = 3;
677 for (uint i = 0; i < total_args_passed; i++) {
678 VMReg vmreg = VEC_ArgReg[i]->as_VMReg();
679 regs[i].set_pair(vmreg->next(next_reg_val), vmreg);
680 }
681 return 0;
682 }
683
684 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
685 VMRegPair *regs,
686 int total_args_passed) {
687
688 // We return the amount of VMRegImpl stack slots we need to reserve for all
689 // the arguments NOT counting out_preserve_stack_slots.
690
691 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
692 c_rarg0, c_rarg1, c_rarg2, c_rarg3,
693 c_rarg4, c_rarg5, c_rarg6, c_rarg7
694 };
695 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
696 c_farg0, c_farg1, c_farg2, c_farg3,
697 c_farg4, c_farg5, c_farg6, c_farg7
698 };
699
700 uint int_args = 0;
701 uint fp_args = 0;
702 uint stk_args = 0; // inc by 2 each time
703
704 for (int i = 0; i < total_args_passed; i++) {
705 switch (sig_bt[i]) {
706 case T_BOOLEAN: // fall through
707 case T_CHAR: // fall through
708 case T_BYTE: // fall through
709 case T_SHORT: // fall through
710 case T_INT:
711 if (int_args < Argument::n_int_register_parameters_c) {
712 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
713 } else {
714 regs[i].set1(VMRegImpl::stack2reg(stk_args));
715 stk_args += 2;
716 }
717 break;
718 case T_LONG: // fall through
719 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
720 case T_OBJECT: // fall through
721 case T_ARRAY: // fall through
722 case T_ADDRESS: // fall through
723 case T_METADATA:
724 if (int_args < Argument::n_int_register_parameters_c) {
725 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
726 } else {
727 regs[i].set2(VMRegImpl::stack2reg(stk_args));
728 stk_args += 2;
729 }
730 break;
731 case T_FLOAT:
732 if (fp_args < Argument::n_float_register_parameters_c) {
733 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
734 } else if (int_args < Argument::n_int_register_parameters_c) {
735 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
736 } else {
737 regs[i].set1(VMRegImpl::stack2reg(stk_args));
738 stk_args += 2;
739 }
740 break;
741 case T_DOUBLE:
742 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
743 if (fp_args < Argument::n_float_register_parameters_c) {
744 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
745 } else if (int_args < Argument::n_int_register_parameters_c) {
746 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
747 } else {
748 regs[i].set2(VMRegImpl::stack2reg(stk_args));
749 stk_args += 2;
750 }
751 break;
752 case T_VOID: // Halves of longs and doubles
753 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
754 regs[i].set_bad();
755 break;
756 default:
757 ShouldNotReachHere();
758 }
759 }
760
761 return stk_args;
762 }
763
764 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
765 // We always ignore the frame_slots arg and just use the space just below frame pointer
766 // which by this time is free to use
767 switch (ret_type) {
768 case T_FLOAT:
769 __ fsw(f10, Address(fp, -3 * wordSize));
770 break;
771 case T_DOUBLE:
772 __ fsd(f10, Address(fp, -3 * wordSize));
773 break;
774 case T_VOID: break;
775 default: {
776 __ sd(x10, Address(fp, -3 * wordSize));
777 }
778 }
779 }
780
781 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
782 // We always ignore the frame_slots arg and just use the space just below frame pointer
783 // which by this time is free to use
784 switch (ret_type) {
785 case T_FLOAT:
786 __ flw(f10, Address(fp, -3 * wordSize));
787 break;
788 case T_DOUBLE:
789 __ fld(f10, Address(fp, -3 * wordSize));
790 break;
791 case T_VOID: break;
792 default: {
793 __ ld(x10, Address(fp, -3 * wordSize));
794 }
795 }
796 }
797
798 static void save_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
799 RegSet x;
800 for ( int i = first_arg ; i < arg_count ; i++ ) {
801 if (args[i].first()->is_Register()) {
802 x = x + args[i].first()->as_Register();
803 } else if (args[i].first()->is_FloatRegister()) {
804 __ subi(sp, sp, 2 * wordSize);
805 __ fsd(args[i].first()->as_FloatRegister(), Address(sp, 0));
806 }
807 }
808 __ push_reg(x, sp);
809 }
810
811 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
812 RegSet x;
813 for ( int i = first_arg ; i < arg_count ; i++ ) {
814 if (args[i].first()->is_Register()) {
815 x = x + args[i].first()->as_Register();
816 } else {
817 ;
818 }
819 }
820 __ pop_reg(x, sp);
821 for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) {
822 if (args[i].first()->is_Register()) {
823 ;
824 } else if (args[i].first()->is_FloatRegister()) {
825 __ fld(args[i].first()->as_FloatRegister(), Address(sp, 0));
826 __ addi(sp, sp, 2 * wordSize);
827 }
828 }
829 }
830
831 static void verify_oop_args(MacroAssembler* masm,
832 const methodHandle& method,
833 const BasicType* sig_bt,
834 const VMRegPair* regs) {
835 const Register temp_reg = x9; // not part of any compiled calling seq
836 if (VerifyOops) {
837 for (int i = 0; i < method->size_of_parameters(); i++) {
838 if (sig_bt[i] == T_OBJECT ||
839 sig_bt[i] == T_ARRAY) {
840 VMReg r = regs[i].first();
841 assert(r->is_valid(), "bad oop arg");
842 if (r->is_stack()) {
843 __ ld(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
844 __ verify_oop(temp_reg);
845 } else {
846 __ verify_oop(r->as_Register());
847 }
848 }
849 }
850 }
851 }
852
853 // on exit, sp points to the ContinuationEntry
854 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& stack_slots) {
855 assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
856 assert(in_bytes(ContinuationEntry::cont_offset()) % VMRegImpl::stack_slot_size == 0, "");
857 assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
858
859 stack_slots += (int)ContinuationEntry::size() / wordSize;
860 __ sub(sp, sp, (int)ContinuationEntry::size()); // place Continuation metadata
861
862 OopMap* map = new OopMap(((int)ContinuationEntry::size() + wordSize) / VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
863
864 __ ld(t0, Address(xthread, JavaThread::cont_entry_offset()));
865 __ sd(t0, Address(sp, ContinuationEntry::parent_offset()));
866 __ sd(sp, Address(xthread, JavaThread::cont_entry_offset()));
867
868 return map;
869 }
870
871 // on entry c_rarg1 points to the continuation
872 // sp points to ContinuationEntry
873 // c_rarg3 -- isVirtualThread
874 static void fill_continuation_entry(MacroAssembler* masm) {
875 #ifdef ASSERT
876 __ mv(t0, ContinuationEntry::cookie_value());
877 __ sw(t0, Address(sp, ContinuationEntry::cookie_offset()));
878 #endif
879
880 __ sd(c_rarg1, Address(sp, ContinuationEntry::cont_offset()));
881 __ sw(c_rarg3, Address(sp, ContinuationEntry::flags_offset()));
882 __ sd(zr, Address(sp, ContinuationEntry::chunk_offset()));
883 __ sw(zr, Address(sp, ContinuationEntry::argsize_offset()));
884 __ sw(zr, Address(sp, ContinuationEntry::pin_count_offset()));
885
886 __ ld(t0, Address(xthread, JavaThread::cont_fastpath_offset()));
887 __ sd(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
888 __ ld(t0, Address(xthread, JavaThread::held_monitor_count_offset()));
889 __ sd(t0, Address(sp, ContinuationEntry::parent_held_monitor_count_offset()));
890
891 __ sd(zr, Address(xthread, JavaThread::cont_fastpath_offset()));
892 __ sd(zr, Address(xthread, JavaThread::held_monitor_count_offset()));
893 }
894
895 // on entry, sp points to the ContinuationEntry
896 // on exit, fp points to the spilled fp + 2 * wordSize in the entry frame
897 static void continuation_enter_cleanup(MacroAssembler* masm) {
898 #ifndef PRODUCT
899 Label OK;
900 __ ld(t0, Address(xthread, JavaThread::cont_entry_offset()));
901 __ beq(sp, t0, OK);
902 __ stop("incorrect sp");
903 __ bind(OK);
904 #endif
905
906 __ ld(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
907 __ sd(t0, Address(xthread, JavaThread::cont_fastpath_offset()));
908
909 if (CheckJNICalls) {
910 // Check if this is a virtual thread continuation
911 Label L_skip_vthread_code;
912 __ lwu(t0, Address(sp, ContinuationEntry::flags_offset()));
913 __ beqz(t0, L_skip_vthread_code);
914
915 // If the held monitor count is > 0 and this vthread is terminating then
916 // it failed to release a JNI monitor. So we issue the same log message
917 // that JavaThread::exit does.
918 __ ld(t0, Address(xthread, JavaThread::jni_monitor_count_offset()));
919 __ beqz(t0, L_skip_vthread_code);
920
921 // Save return value potentially containing the exception oop in callee-saved x9
922 __ mv(x9, x10);
923 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::log_jni_monitor_still_held));
924 // Restore potential return value
925 __ mv(x10, x9);
926
927 // For vthreads we have to explicitly zero the JNI monitor count of the carrier
928 // on termination. The held count is implicitly zeroed below when we restore from
929 // the parent held count (which has to be zero).
930 __ sd(zr, Address(xthread, JavaThread::jni_monitor_count_offset()));
931
932 __ bind(L_skip_vthread_code);
933 }
934 #ifdef ASSERT
935 else {
936 // Check if this is a virtual thread continuation
937 Label L_skip_vthread_code;
938 __ lwu(t0, Address(sp, ContinuationEntry::flags_offset()));
939 __ beqz(t0, L_skip_vthread_code);
940
941 // See comment just above. If not checking JNI calls the JNI count is only
942 // needed for assertion checking.
943 __ sd(zr, Address(xthread, JavaThread::jni_monitor_count_offset()));
944
945 __ bind(L_skip_vthread_code);
946 }
947 #endif
948
949 __ ld(t0, Address(sp, ContinuationEntry::parent_held_monitor_count_offset()));
950 __ sd(t0, Address(xthread, JavaThread::held_monitor_count_offset()));
951
952 __ ld(t0, Address(sp, ContinuationEntry::parent_offset()));
953 __ sd(t0, Address(xthread, JavaThread::cont_entry_offset()));
954 __ add(fp, sp, (int)ContinuationEntry::size() + 2 * wordSize /* 2 extra words to match up with leave() */);
955 }
956
957 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
958 // On entry: c_rarg1 -- the continuation object
959 // c_rarg2 -- isContinue
960 // c_rarg3 -- isVirtualThread
961 static void gen_continuation_enter(MacroAssembler* masm,
962 const methodHandle& method,
963 const BasicType* sig_bt,
964 const VMRegPair* regs,
965 int& exception_offset,
966 OopMapSet*oop_maps,
967 int& frame_complete,
968 int& stack_slots,
969 int& interpreted_entry_offset,
970 int& compiled_entry_offset) {
971 // verify_oop_args(masm, method, sig_bt, regs);
972 Address resolve(SharedRuntime::get_resolve_static_call_stub(), relocInfo::static_call_type);
973
974 address start = __ pc();
975
976 Label call_thaw, exit;
977
978 // i2i entry used at interp_only_mode only
979 interpreted_entry_offset = __ pc() - start;
980 {
981 #ifdef ASSERT
982 Label is_interp_only;
983 __ lw(t0, Address(xthread, JavaThread::interp_only_mode_offset()));
984 __ bnez(t0, is_interp_only);
985 __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
986 __ bind(is_interp_only);
987 #endif
988
989 // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
990 __ ld(c_rarg1, Address(esp, Interpreter::stackElementSize * 2));
991 __ ld(c_rarg2, Address(esp, Interpreter::stackElementSize * 1));
992 __ ld(c_rarg3, Address(esp, Interpreter::stackElementSize * 0));
993 __ push_cont_fastpath(xthread);
994
995 __ enter();
996 stack_slots = 2; // will be adjusted in setup
997 OopMap* map = continuation_enter_setup(masm, stack_slots);
998 // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
999 // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
1000
1001 fill_continuation_entry(masm);
1002
1003 __ bnez(c_rarg2, call_thaw);
1004
1005 address call_pc;
1006 {
1007 Assembler::IncompressibleScope scope(masm);
1008 // Make sure the call is patchable
1009 __ align(NativeInstruction::instruction_size);
1010
1011 call_pc = __ reloc_call(resolve);
1012 if (call_pc == nullptr) {
1013 fatal("CodeCache is full at gen_continuation_enter");
1014 }
1015
1016 oop_maps->add_gc_map(__ pc() - start, map);
1017 __ post_call_nop();
1018 }
1019 __ j(exit);
1020
1021 address stub = CompiledDirectCall::emit_to_interp_stub(masm, call_pc);
1022 if (stub == nullptr) {
1023 fatal("CodeCache is full at gen_continuation_enter");
1024 }
1025 }
1026
1027 // compiled entry
1028 __ align(CodeEntryAlignment);
1029 compiled_entry_offset = __ pc() - start;
1030
1031 __ enter();
1032 stack_slots = 2; // will be adjusted in setup
1033 OopMap* map = continuation_enter_setup(masm, stack_slots);
1034 frame_complete = __ pc() - start;
1035
1036 fill_continuation_entry(masm);
1037
1038 __ bnez(c_rarg2, call_thaw);
1039
1040 address call_pc;
1041 {
1042 Assembler::IncompressibleScope scope(masm);
1043 // Make sure the call is patchable
1044 __ align(NativeInstruction::instruction_size);
1045
1046 call_pc = __ reloc_call(resolve);
1047 if (call_pc == nullptr) {
1048 fatal("CodeCache is full at gen_continuation_enter");
1049 }
1050
1051 oop_maps->add_gc_map(__ pc() - start, map);
1052 __ post_call_nop();
1053 }
1054
1055 __ j(exit);
1056
1057 __ bind(call_thaw);
1058
1059 // Post call nops must be natural aligned due to cmodx rules.
1060 {
1061 Assembler::IncompressibleScope scope(masm);
1062 __ align(NativeInstruction::instruction_size);
1063
1064 ContinuationEntry::_thaw_call_pc_offset = __ pc() - start;
1065 __ rt_call(CAST_FROM_FN_PTR(address, StubRoutines::cont_thaw()));
1066 oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1067 ContinuationEntry::_return_pc_offset = __ pc() - start;
1068 __ post_call_nop();
1069 }
1070
1071 __ bind(exit);
1072 ContinuationEntry::_cleanup_offset = __ pc() - start;
1073 continuation_enter_cleanup(masm);
1074 __ leave();
1075 __ ret();
1076
1077 // exception handling
1078 exception_offset = __ pc() - start;
1079 {
1080 __ mv(x9, x10); // save return value contaning the exception oop in callee-saved x9
1081
1082 continuation_enter_cleanup(masm);
1083
1084 __ ld(c_rarg1, Address(fp, -1 * wordSize)); // return address
1085 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), xthread, c_rarg1);
1086
1087 // see OptoRuntime::generate_exception_blob: x10 -- exception oop, x13 -- exception pc
1088
1089 __ mv(x11, x10); // the exception handler
1090 __ mv(x10, x9); // restore return value contaning the exception oop
1091 __ verify_oop(x10);
1092
1093 __ leave();
1094 __ mv(x13, ra);
1095 __ jr(x11); // the exception handler
1096 }
1097
1098 address stub = CompiledDirectCall::emit_to_interp_stub(masm, call_pc);
1099 if (stub == nullptr) {
1100 fatal("CodeCache is full at gen_continuation_enter");
1101 }
1102 }
1103
1104 static void gen_continuation_yield(MacroAssembler* masm,
1105 const methodHandle& method,
1106 const BasicType* sig_bt,
1107 const VMRegPair* regs,
1108 OopMapSet* oop_maps,
1109 int& frame_complete,
1110 int& stack_slots,
1111 int& compiled_entry_offset) {
1112 enum layout {
1113 fp_off,
1114 fp_off2,
1115 return_off,
1116 return_off2,
1117 framesize // inclusive of return address
1118 };
1119 // assert(is_even(framesize/2), "sp not 16-byte aligned");
1120
1121 stack_slots = framesize / VMRegImpl::slots_per_word;
1122 assert(stack_slots == 2, "recheck layout");
1123
1124 address start = __ pc();
1125
1126 compiled_entry_offset = __ pc() - start;
1127 __ enter();
1128
1129 __ mv(c_rarg1, sp);
1130
1131 // Post call nops must be natural aligned due to cmodx rules.
1132 __ align(NativeInstruction::instruction_size);
1133
1134 frame_complete = __ pc() - start;
1135 address the_pc = __ pc();
1136
1137 {
1138 Assembler::IncompressibleScope scope(masm);
1139 __ 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
1140 }
1141
1142 __ mv(c_rarg0, xthread);
1143 __ set_last_Java_frame(sp, fp, the_pc, t0);
1144 __ call_VM_leaf(Continuation::freeze_entry(), 2);
1145 __ reset_last_Java_frame(true);
1146
1147 Label pinned;
1148
1149 __ bnez(x10, pinned);
1150
1151 // We've succeeded, set sp to the ContinuationEntry
1152 __ ld(sp, Address(xthread, JavaThread::cont_entry_offset()));
1153 continuation_enter_cleanup(masm);
1154
1155 __ bind(pinned); // pinned -- return to caller
1156
1157 // handle pending exception thrown by freeze
1158 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1159 Label ok;
1160 __ beqz(t0, ok);
1161 __ leave();
1162 __ j(RuntimeAddress(StubRoutines::forward_exception_entry()));
1163 __ bind(ok);
1164
1165 __ leave();
1166 __ ret();
1167
1168 OopMap* map = new OopMap(framesize, 1);
1169 oop_maps->add_gc_map(the_pc - start, map);
1170 }
1171
1172 void SharedRuntime::continuation_enter_cleanup(MacroAssembler* masm) {
1173 ::continuation_enter_cleanup(masm);
1174 }
1175
1176 static void gen_special_dispatch(MacroAssembler* masm,
1177 const methodHandle& method,
1178 const BasicType* sig_bt,
1179 const VMRegPair* regs) {
1180 verify_oop_args(masm, method, sig_bt, regs);
1181 vmIntrinsics::ID iid = method->intrinsic_id();
1182
1183 // Now write the args into the outgoing interpreter space
1184 bool has_receiver = false;
1185 Register receiver_reg = noreg;
1186 int member_arg_pos = -1;
1187 Register member_reg = noreg;
1188 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1189 if (ref_kind != 0) {
1190 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1191 member_reg = x9; // known to be free at this point
1192 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1193 } else if (iid == vmIntrinsics::_invokeBasic) {
1194 has_receiver = true;
1195 } else if (iid == vmIntrinsics::_linkToNative) {
1196 member_arg_pos = method->size_of_parameters() - 1; // trailing NativeEntryPoint argument
1197 member_reg = x9; // known to be free at this point
1198 } else {
1199 fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1200 }
1201
1202 if (member_reg != noreg) {
1203 // Load the member_arg into register, if necessary.
1204 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1205 VMReg r = regs[member_arg_pos].first();
1206 if (r->is_stack()) {
1207 __ ld(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1208 } else {
1209 // no data motion is needed
1210 member_reg = r->as_Register();
1211 }
1212 }
1213
1214 if (has_receiver) {
1215 // Make sure the receiver is loaded into a register.
1216 assert(method->size_of_parameters() > 0, "oob");
1217 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1218 VMReg r = regs[0].first();
1219 assert(r->is_valid(), "bad receiver arg");
1220 if (r->is_stack()) {
1221 // Porting note: This assumes that compiled calling conventions always
1222 // pass the receiver oop in a register. If this is not true on some
1223 // platform, pick a temp and load the receiver from stack.
1224 fatal("receiver always in a register");
1225 receiver_reg = x12; // known to be free at this point
1226 __ ld(receiver_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1227 } else {
1228 // no data motion is needed
1229 receiver_reg = r->as_Register();
1230 }
1231 }
1232
1233 // Figure out which address we are really jumping to:
1234 MethodHandles::generate_method_handle_dispatch(masm, iid,
1235 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1236 }
1237
1238 // ---------------------------------------------------------------------------
1239 // Generate a native wrapper for a given method. The method takes arguments
1240 // in the Java compiled code convention, marshals them to the native
1241 // convention (handlizes oops, etc), transitions to native, makes the call,
1242 // returns to java state (possibly blocking), unhandlizes any result and
1243 // returns.
1244 //
1245 // Critical native functions are a shorthand for the use of
1246 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1247 // functions. The wrapper is expected to unpack the arguments before
1248 // passing them to the callee and perform checks before and after the
1249 // native call to ensure that they GCLocker
1250 // lock_critical/unlock_critical semantics are followed. Some other
1251 // parts of JNI setup are skipped like the tear down of the JNI handle
1252 // block and the check for pending exceptions it's impossible for them
1253 // to be thrown.
1254 //
1255 // They are roughly structured like this:
1256 // if (GCLocker::needs_gc()) SharedRuntime::block_for_jni_critical()
1257 // tranistion to thread_in_native
1258 // unpack array arguments and call native entry point
1259 // check for safepoint in progress
1260 // check if any thread suspend flags are set
1261 // call into JVM and possible unlock the JNI critical
1262 // if a GC was suppressed while in the critical native.
1263 // transition back to thread_in_Java
1264 // return to caller
1265 //
1266 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1267 const methodHandle& method,
1268 int compile_id,
1269 BasicType* in_sig_bt,
1270 VMRegPair* in_regs,
1271 BasicType ret_type) {
1272 if (method->is_continuation_native_intrinsic()) {
1273 int exception_offset = -1;
1274 OopMapSet* oop_maps = new OopMapSet();
1275 int frame_complete = -1;
1276 int stack_slots = -1;
1277 int interpreted_entry_offset = -1;
1278 int vep_offset = -1;
1279 if (method->is_continuation_enter_intrinsic()) {
1280 gen_continuation_enter(masm,
1281 method,
1282 in_sig_bt,
1283 in_regs,
1284 exception_offset,
1285 oop_maps,
1286 frame_complete,
1287 stack_slots,
1288 interpreted_entry_offset,
1289 vep_offset);
1290 } else if (method->is_continuation_yield_intrinsic()) {
1291 gen_continuation_yield(masm,
1292 method,
1293 in_sig_bt,
1294 in_regs,
1295 oop_maps,
1296 frame_complete,
1297 stack_slots,
1298 vep_offset);
1299 } else {
1300 guarantee(false, "Unknown Continuation native intrinsic");
1301 }
1302
1303 #ifdef ASSERT
1304 if (method->is_continuation_enter_intrinsic()) {
1305 assert(interpreted_entry_offset != -1, "Must be set");
1306 assert(exception_offset != -1, "Must be set");
1307 } else {
1308 assert(interpreted_entry_offset == -1, "Must be unset");
1309 assert(exception_offset == -1, "Must be unset");
1310 }
1311 assert(frame_complete != -1, "Must be set");
1312 assert(stack_slots != -1, "Must be set");
1313 assert(vep_offset != -1, "Must be set");
1314 #endif
1315
1316 __ flush();
1317 nmethod* nm = nmethod::new_native_nmethod(method,
1318 compile_id,
1319 masm->code(),
1320 vep_offset,
1321 frame_complete,
1322 stack_slots,
1323 in_ByteSize(-1),
1324 in_ByteSize(-1),
1325 oop_maps,
1326 exception_offset);
1327 if (nm == nullptr) return nm;
1328 if (method->is_continuation_enter_intrinsic()) {
1329 ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
1330 } else if (method->is_continuation_yield_intrinsic()) {
1331 ContinuationEntry::set_yield_code(nm);
1332 } else {
1333 guarantee(false, "Unknown Continuation native intrinsic");
1334 }
1335 return nm;
1336 }
1337
1338 if (method->is_method_handle_intrinsic()) {
1339 vmIntrinsics::ID iid = method->intrinsic_id();
1340 intptr_t start = (intptr_t)__ pc();
1341 int vep_offset = ((intptr_t)__ pc()) - start;
1342
1343 // First instruction must be a nop as it may need to be patched on deoptimisation
1344 {
1345 Assembler::IncompressibleScope scope(masm); // keep the nop as 4 bytes for patching.
1346 MacroAssembler::assert_alignment(__ pc());
1347 __ nop(); // 4 bytes
1348 }
1349 gen_special_dispatch(masm,
1350 method,
1351 in_sig_bt,
1352 in_regs);
1353 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
1354 __ flush();
1355 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
1356 return nmethod::new_native_nmethod(method,
1357 compile_id,
1358 masm->code(),
1359 vep_offset,
1360 frame_complete,
1361 stack_slots / VMRegImpl::slots_per_word,
1362 in_ByteSize(-1),
1363 in_ByteSize(-1),
1364 (OopMapSet*)nullptr);
1365 }
1366 address native_func = method->native_function();
1367 assert(native_func != nullptr, "must have function");
1368
1369 // An OopMap for lock (and class if static)
1370 OopMapSet *oop_maps = new OopMapSet();
1371 assert_cond(oop_maps != nullptr);
1372 intptr_t start = (intptr_t)__ pc();
1373
1374 // We have received a description of where all the java arg are located
1375 // on entry to the wrapper. We need to convert these args to where
1376 // the jni function will expect them. To figure out where they go
1377 // we convert the java signature to a C signature by inserting
1378 // the hidden arguments as arg[0] and possibly arg[1] (static method)
1379
1380 const int total_in_args = method->size_of_parameters();
1381 int total_c_args = total_in_args + (method->is_static() ? 2 : 1);
1382
1383 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1384 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1385
1386 int argc = 0;
1387 out_sig_bt[argc++] = T_ADDRESS;
1388 if (method->is_static()) {
1389 out_sig_bt[argc++] = T_OBJECT;
1390 }
1391
1392 for (int i = 0; i < total_in_args ; i++) {
1393 out_sig_bt[argc++] = in_sig_bt[i];
1394 }
1395
1396 // Now figure out where the args must be stored and how much stack space
1397 // they require.
1398 int out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
1399
1400 // Compute framesize for the wrapper. We need to handlize all oops in
1401 // incoming registers
1402
1403 // Calculate the total number of stack slots we will need.
1404
1405 // First count the abi requirement plus all of the outgoing args
1406 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1407
1408 // Now the space for the inbound oop handle area
1409 int total_save_slots = 8 * VMRegImpl::slots_per_word; // 8 arguments passed in registers
1410
1411 int oop_handle_offset = stack_slots;
1412 stack_slots += total_save_slots;
1413
1414 // Now any space we need for handlizing a klass if static method
1415
1416 int klass_slot_offset = 0;
1417 int klass_offset = -1;
1418 int lock_slot_offset = 0;
1419 bool is_static = false;
1420
1421 if (method->is_static()) {
1422 klass_slot_offset = stack_slots;
1423 stack_slots += VMRegImpl::slots_per_word;
1424 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1425 is_static = true;
1426 }
1427
1428 // Plus a lock if needed
1429
1430 if (method->is_synchronized()) {
1431 lock_slot_offset = stack_slots;
1432 stack_slots += VMRegImpl::slots_per_word;
1433 }
1434
1435 // Now a place (+2) to save return values or temp during shuffling
1436 // + 4 for return address (which we own) and saved fp
1437 stack_slots += 6;
1438
1439 // Ok The space we have allocated will look like:
1440 //
1441 //
1442 // FP-> | |
1443 // | 2 slots (ra) |
1444 // | 2 slots (fp) |
1445 // |---------------------|
1446 // | 2 slots for moves |
1447 // |---------------------|
1448 // | lock box (if sync) |
1449 // |---------------------| <- lock_slot_offset
1450 // | klass (if static) |
1451 // |---------------------| <- klass_slot_offset
1452 // | oopHandle area |
1453 // |---------------------| <- oop_handle_offset (8 java arg registers)
1454 // | outbound memory |
1455 // | based arguments |
1456 // | |
1457 // |---------------------|
1458 // | |
1459 // SP-> | out_preserved_slots |
1460 //
1461 //
1462
1463
1464 // Now compute actual number of stack words we need rounding to make
1465 // stack properly aligned.
1466 stack_slots = align_up(stack_slots, StackAlignmentInSlots);
1467
1468 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1469
1470 // First thing make an ic check to see if we should even be here
1471
1472 // We are free to use all registers as temps without saving them and
1473 // restoring them except fp. fp is the only callee save register
1474 // as far as the interpreter and the compiler(s) are concerned.
1475
1476 const Register receiver = j_rarg0;
1477
1478 __ verify_oop(receiver);
1479 assert_different_registers(receiver, t0, t1);
1480
1481 __ ic_check();
1482
1483 int vep_offset = ((intptr_t)__ pc()) - start;
1484
1485 // If we have to make this method not-entrant we'll overwrite its
1486 // first instruction with a jump.
1487 {
1488 Assembler::IncompressibleScope scope(masm); // keep the nop as 4 bytes for patching.
1489 MacroAssembler::assert_alignment(__ pc());
1490 __ nop(); // 4 bytes
1491 }
1492
1493 if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) {
1494 Label L_skip_barrier;
1495 __ mov_metadata(t1, method->method_holder()); // InstanceKlass*
1496 __ clinit_barrier(t1, t0, &L_skip_barrier);
1497 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
1498
1499 __ bind(L_skip_barrier);
1500 }
1501
1502 // Generate stack overflow check
1503 __ bang_stack_with_offset(checked_cast<int>(StackOverflow::stack_shadow_zone_size()));
1504
1505 // Generate a new frame for the wrapper.
1506 __ enter();
1507 // -2 because return address is already present and so is saved fp
1508 __ sub(sp, sp, stack_size - 2 * wordSize);
1509
1510 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1511 assert_cond(bs != nullptr);
1512 bs->nmethod_entry_barrier(masm, nullptr /* slow_path */, nullptr /* continuation */, nullptr /* guard */);
1513
1514 // Frame is now completed as far as size and linkage.
1515 int frame_complete = ((intptr_t)__ pc()) - start;
1516
1517 // We use x18 as the oop handle for the receiver/klass
1518 // It is callee save so it survives the call to native
1519
1520 const Register oop_handle_reg = x18;
1521
1522 //
1523 // We immediately shuffle the arguments so that any vm call we have to
1524 // make from here on out (sync slow path, jvmti, etc.) we will have
1525 // captured the oops from our caller and have a valid oopMap for
1526 // them.
1527
1528 // -----------------
1529 // The Grand Shuffle
1530
1531 // The Java calling convention is either equal (linux) or denser (win64) than the
1532 // c calling convention. However the because of the jni_env argument the c calling
1533 // convention always has at least one more (and two for static) arguments than Java.
1534 // Therefore if we move the args from java -> c backwards then we will never have
1535 // a register->register conflict and we don't have to build a dependency graph
1536 // and figure out how to break any cycles.
1537 //
1538
1539 // Record esp-based slot for receiver on stack for non-static methods
1540 int receiver_offset = -1;
1541
1542 // This is a trick. We double the stack slots so we can claim
1543 // the oops in the caller's frame. Since we are sure to have
1544 // more args than the caller doubling is enough to make
1545 // sure we can capture all the incoming oop args from the
1546 // caller.
1547 //
1548 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1549 assert_cond(map != nullptr);
1550
1551 int float_args = 0;
1552 int int_args = 0;
1553
1554 #ifdef ASSERT
1555 bool reg_destroyed[Register::number_of_registers];
1556 bool freg_destroyed[FloatRegister::number_of_registers];
1557 for ( int r = 0 ; r < Register::number_of_registers ; r++ ) {
1558 reg_destroyed[r] = false;
1559 }
1560 for ( int f = 0 ; f < FloatRegister::number_of_registers ; f++ ) {
1561 freg_destroyed[f] = false;
1562 }
1563
1564 #endif /* ASSERT */
1565
1566 // For JNI natives the incoming and outgoing registers are offset upwards.
1567 GrowableArray<int> arg_order(2 * total_in_args);
1568
1569 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) {
1570 arg_order.push(i);
1571 arg_order.push(c_arg);
1572 }
1573
1574 for (int ai = 0; ai < arg_order.length(); ai += 2) {
1575 int i = arg_order.at(ai);
1576 int c_arg = arg_order.at(ai + 1);
1577 __ block_comment(err_msg("mv %d -> %d", i, c_arg));
1578 assert(c_arg != -1 && i != -1, "wrong order");
1579 #ifdef ASSERT
1580 if (in_regs[i].first()->is_Register()) {
1581 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!");
1582 } else if (in_regs[i].first()->is_FloatRegister()) {
1583 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!");
1584 }
1585 if (out_regs[c_arg].first()->is_Register()) {
1586 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
1587 } else if (out_regs[c_arg].first()->is_FloatRegister()) {
1588 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true;
1589 }
1590 #endif /* ASSERT */
1591 switch (in_sig_bt[i]) {
1592 case T_ARRAY:
1593 case T_OBJECT:
1594 __ object_move(map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
1595 ((i == 0) && (!is_static)),
1596 &receiver_offset);
1597 int_args++;
1598 break;
1599 case T_VOID:
1600 break;
1601
1602 case T_FLOAT:
1603 __ float_move(in_regs[i], out_regs[c_arg]);
1604 float_args++;
1605 break;
1606
1607 case T_DOUBLE:
1608 assert( i + 1 < total_in_args &&
1609 in_sig_bt[i + 1] == T_VOID &&
1610 out_sig_bt[c_arg + 1] == T_VOID, "bad arg list");
1611 __ double_move(in_regs[i], out_regs[c_arg]);
1612 float_args++;
1613 break;
1614
1615 case T_LONG :
1616 __ long_move(in_regs[i], out_regs[c_arg]);
1617 int_args++;
1618 break;
1619
1620 case T_ADDRESS:
1621 assert(false, "found T_ADDRESS in java args");
1622 break;
1623
1624 default:
1625 __ move32_64(in_regs[i], out_regs[c_arg]);
1626 int_args++;
1627 }
1628 }
1629
1630 // point c_arg at the first arg that is already loaded in case we
1631 // need to spill before we call out
1632 int c_arg = total_c_args - total_in_args;
1633
1634 // Pre-load a static method's oop into c_rarg1.
1635 if (method->is_static()) {
1636
1637 // load oop into a register
1638 __ movoop(c_rarg1,
1639 JNIHandles::make_local(method->method_holder()->java_mirror()));
1640
1641 // Now handlize the static class mirror it's known not-null.
1642 __ sd(c_rarg1, Address(sp, klass_offset));
1643 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
1644
1645 // Now get the handle
1646 __ la(c_rarg1, Address(sp, klass_offset));
1647 // and protect the arg if we must spill
1648 c_arg--;
1649 }
1650
1651 // Change state to native (we save the return address in the thread, since it might not
1652 // be pushed on the stack when we do a stack traversal). It is enough that the pc()
1653 // points into the right code segment. It does not have to be the correct return pc.
1654 // We use the same pc/oopMap repeatedly when we call out.
1655
1656 Label native_return;
1657 if (method->is_object_wait0()) {
1658 // For convenience we use the pc we want to resume to in case of preemption on Object.wait.
1659 __ set_last_Java_frame(sp, noreg, native_return, t0);
1660 } else {
1661 intptr_t the_pc = (intptr_t) __ pc();
1662 oop_maps->add_gc_map(the_pc - start, map);
1663
1664 __ set_last_Java_frame(sp, noreg, __ pc(), t0);
1665 }
1666
1667 Label dtrace_method_entry, dtrace_method_entry_done;
1668 if (DTraceMethodProbes) {
1669 __ j(dtrace_method_entry);
1670 __ bind(dtrace_method_entry_done);
1671 }
1672
1673 // RedefineClasses() tracing support for obsolete method entry
1674 if (log_is_enabled(Trace, redefine, class, obsolete)) {
1675 // protect the args we've loaded
1676 save_args(masm, total_c_args, c_arg, out_regs);
1677 __ mov_metadata(c_rarg1, method());
1678 __ call_VM_leaf(
1679 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
1680 xthread, c_rarg1);
1681 restore_args(masm, total_c_args, c_arg, out_regs);
1682 }
1683
1684 // Lock a synchronized method
1685
1686 // Register definitions used by locking and unlocking
1687
1688 const Register swap_reg = x10;
1689 const Register obj_reg = x9; // Will contain the oop
1690 const Register lock_reg = x30; // Address of compiler lock object (BasicLock)
1691 const Register old_hdr = x30; // value of old header at unlock time
1692 const Register lock_tmp = x31; // Temporary used by lightweight_lock/unlock
1693 const Register tmp = ra;
1694
1695 Label slow_path_lock;
1696 Label lock_done;
1697
1698 if (method->is_synchronized()) {
1699 // Get the handle (the 2nd argument)
1700 __ mv(oop_handle_reg, c_rarg1);
1701
1702 // Get address of the box
1703
1704 __ la(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1705
1706 // Load the oop from the handle
1707 __ ld(obj_reg, Address(oop_handle_reg, 0));
1708
1709 __ lightweight_lock(lock_reg, obj_reg, swap_reg, tmp, lock_tmp, slow_path_lock);
1710
1711 // Slow path will re-enter here
1712 __ bind(lock_done);
1713 }
1714
1715
1716 // Finally just about ready to make the JNI call
1717
1718 // get JNIEnv* which is first argument to native
1719 __ la(c_rarg0, Address(xthread, in_bytes(JavaThread::jni_environment_offset())));
1720
1721 // Now set thread in native
1722 __ la(t1, Address(xthread, JavaThread::thread_state_offset()));
1723 __ mv(t0, _thread_in_native);
1724 __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore);
1725 __ sw(t0, Address(t1));
1726
1727 // Clobbers t1
1728 __ rt_call(native_func);
1729
1730 // Verify or restore cpu control state after JNI call
1731 __ restore_cpu_control_state_after_jni(t0);
1732
1733 // Unpack native results.
1734 if (ret_type != T_OBJECT && ret_type != T_ARRAY) {
1735 __ cast_primitive_type(ret_type, x10);
1736 }
1737
1738 Label safepoint_in_progress, safepoint_in_progress_done;
1739
1740 // Switch thread to "native transition" state before reading the synchronization state.
1741 // This additional state is necessary because reading and testing the synchronization
1742 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1743 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1744 // VM thread changes sync state to synchronizing and suspends threads for GC.
1745 // Thread A is resumed to finish this native method, but doesn't block here since it
1746 // didn't see any synchronization is progress, and escapes.
1747 __ mv(t0, _thread_in_native_trans);
1748
1749 __ sw(t0, Address(xthread, JavaThread::thread_state_offset()));
1750
1751 // Force this write out before the read below
1752 if (!UseSystemMemoryBarrier) {
1753 __ membar(MacroAssembler::AnyAny);
1754 }
1755
1756 // check for safepoint operation in progress and/or pending suspend requests
1757 {
1758 __ safepoint_poll(safepoint_in_progress, true /* at_return */, false /* in_nmethod */);
1759 __ lwu(t0, Address(xthread, JavaThread::suspend_flags_offset()));
1760 __ bnez(t0, safepoint_in_progress);
1761 __ bind(safepoint_in_progress_done);
1762 }
1763
1764 // change thread state
1765 __ la(t1, Address(xthread, JavaThread::thread_state_offset()));
1766 __ mv(t0, _thread_in_Java);
1767 __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore);
1768 __ sw(t0, Address(t1));
1769
1770 if (method->is_object_wait0()) {
1771 // Check preemption for Object.wait()
1772 __ ld(t1, Address(xthread, JavaThread::preempt_alternate_return_offset()));
1773 __ beqz(t1, native_return);
1774 __ sd(zr, Address(xthread, JavaThread::preempt_alternate_return_offset()));
1775 __ jr(t1);
1776 __ bind(native_return);
1777
1778 intptr_t the_pc = (intptr_t) __ pc();
1779 oop_maps->add_gc_map(the_pc - start, map);
1780 }
1781
1782 Label reguard;
1783 Label reguard_done;
1784 __ lbu(t0, Address(xthread, JavaThread::stack_guard_state_offset()));
1785 __ mv(t1, StackOverflow::stack_guard_yellow_reserved_disabled);
1786 __ beq(t0, t1, reguard);
1787 __ bind(reguard_done);
1788
1789 // native result if any is live
1790
1791 // Unlock
1792 Label unlock_done;
1793 Label slow_path_unlock;
1794 if (method->is_synchronized()) {
1795
1796 // Get locked oop from the handle we passed to jni
1797 __ ld(obj_reg, Address(oop_handle_reg, 0));
1798
1799 // Must save x10 if if it is live now because cmpxchg must use it
1800 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1801 save_native_result(masm, ret_type, stack_slots);
1802 }
1803
1804 __ lightweight_unlock(obj_reg, old_hdr, swap_reg, lock_tmp, slow_path_unlock);
1805
1806 // slow path re-enters here
1807 __ bind(unlock_done);
1808 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1809 restore_native_result(masm, ret_type, stack_slots);
1810 }
1811 }
1812
1813 Label dtrace_method_exit, dtrace_method_exit_done;
1814 if (DTraceMethodProbes) {
1815 __ j(dtrace_method_exit);
1816 __ bind(dtrace_method_exit_done);
1817 }
1818
1819 __ reset_last_Java_frame(false);
1820
1821 // Unbox oop result, e.g. JNIHandles::resolve result.
1822 if (is_reference_type(ret_type)) {
1823 __ resolve_jobject(x10, x11, x12);
1824 }
1825
1826 if (CheckJNICalls) {
1827 // clear_pending_jni_exception_check
1828 __ sd(zr, Address(xthread, JavaThread::pending_jni_exception_check_fn_offset()));
1829 }
1830
1831 // reset handle block
1832 __ ld(x12, Address(xthread, JavaThread::active_handles_offset()));
1833 __ sd(zr, Address(x12, JNIHandleBlock::top_offset()));
1834
1835 __ leave();
1836
1837 #if INCLUDE_JFR
1838 // We need to do a poll test after unwind in case the sampler
1839 // managed to sample the native frame after returning to Java.
1840 Label L_return;
1841 __ ld(t0, Address(xthread, JavaThread::polling_word_offset()));
1842 address poll_test_pc = __ pc();
1843 __ relocate(relocInfo::poll_return_type);
1844 __ test_bit(t0, t0, log2i_exact(SafepointMechanism::poll_bit()));
1845 __ beqz(t0, L_return);
1846 assert(SharedRuntime::polling_page_return_handler_blob() != nullptr,
1847 "polling page return stub not created yet");
1848 address stub = SharedRuntime::polling_page_return_handler_blob()->entry_point();
1849 __ la(t0, InternalAddress(poll_test_pc));
1850 __ sd(t0, Address(xthread, JavaThread::saved_exception_pc_offset()));
1851 __ far_jump(RuntimeAddress(stub));
1852 __ bind(L_return);
1853 #endif // INCLUDE_JFR
1854
1855 // Any exception pending?
1856 Label exception_pending;
1857 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1858 __ bnez(t0, exception_pending);
1859
1860 // We're done
1861 __ ret();
1862
1863 // Unexpected paths are out of line and go here
1864
1865 // forward the exception
1866 __ bind(exception_pending);
1867
1868 // and forward the exception
1869 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
1870
1871 // Slow path locking & unlocking
1872 if (method->is_synchronized()) {
1873
1874 __ block_comment("Slow path lock {");
1875 __ bind(slow_path_lock);
1876
1877 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM
1878 // args are (oop obj, BasicLock* lock, JavaThread* thread)
1879
1880 // protect the args we've loaded
1881 save_args(masm, total_c_args, c_arg, out_regs);
1882
1883 __ mv(c_rarg0, obj_reg);
1884 __ mv(c_rarg1, lock_reg);
1885 __ mv(c_rarg2, xthread);
1886
1887 // Not a leaf but we have last_Java_frame setup as we want.
1888 // We don't want to unmount in case of contention since that would complicate preserving
1889 // the arguments that had already been marshalled into the native convention. So we force
1890 // the freeze slow path to find this native wrapper frame (see recurse_freeze_native_frame())
1891 // and pin the vthread. Otherwise the fast path won't find it since we don't walk the stack.
1892 __ push_cont_fastpath();
1893 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3);
1894 __ pop_cont_fastpath();
1895 restore_args(masm, total_c_args, c_arg, out_regs);
1896
1897 #ifdef ASSERT
1898 { Label L;
1899 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1900 __ beqz(t0, L);
1901 __ stop("no pending exception allowed on exit from monitorenter");
1902 __ bind(L);
1903 }
1904 #endif
1905 __ j(lock_done);
1906
1907 __ block_comment("} Slow path lock");
1908
1909 __ block_comment("Slow path unlock {");
1910 __ bind(slow_path_unlock);
1911
1912 if (ret_type == T_FLOAT || ret_type == T_DOUBLE) {
1913 save_native_result(masm, ret_type, stack_slots);
1914 }
1915
1916 __ mv(c_rarg2, xthread);
1917 __ la(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1918 __ mv(c_rarg0, obj_reg);
1919
1920 // Save pending exception around call to VM (which contains an EXCEPTION_MARK)
1921 // NOTE that obj_reg == x9 currently
1922 __ ld(x9, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1923 __ sd(zr, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1924
1925 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C));
1926
1927 #ifdef ASSERT
1928 {
1929 Label L;
1930 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1931 __ beqz(t0, L);
1932 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C");
1933 __ bind(L);
1934 }
1935 #endif /* ASSERT */
1936
1937 __ sd(x9, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1938
1939 if (ret_type == T_FLOAT || ret_type == T_DOUBLE) {
1940 restore_native_result(masm, ret_type, stack_slots);
1941 }
1942 __ j(unlock_done);
1943
1944 __ block_comment("} Slow path unlock");
1945
1946 } // synchronized
1947
1948 // SLOW PATH Reguard the stack if needed
1949
1950 __ bind(reguard);
1951 save_native_result(masm, ret_type, stack_slots);
1952 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
1953 restore_native_result(masm, ret_type, stack_slots);
1954 // and continue
1955 __ j(reguard_done);
1956
1957 // SLOW PATH safepoint
1958 {
1959 __ block_comment("safepoint {");
1960 __ bind(safepoint_in_progress);
1961
1962 // Don't use call_VM as it will see a possible pending exception and forward it
1963 // and never return here preventing us from clearing _last_native_pc down below.
1964 //
1965 save_native_result(masm, ret_type, stack_slots);
1966 __ mv(c_rarg0, xthread);
1967 #ifndef PRODUCT
1968 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
1969 #endif
1970 __ rt_call(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans));
1971
1972 // Restore any method result value
1973 restore_native_result(masm, ret_type, stack_slots);
1974
1975 __ j(safepoint_in_progress_done);
1976 __ block_comment("} safepoint");
1977 }
1978
1979 // SLOW PATH dtrace support
1980 if (DTraceMethodProbes) {
1981 {
1982 __ block_comment("dtrace entry {");
1983 __ bind(dtrace_method_entry);
1984
1985 // We have all of the arguments setup at this point. We must not touch any register
1986 // argument registers at this point (what if we save/restore them there are no oop?
1987
1988 save_args(masm, total_c_args, c_arg, out_regs);
1989 __ mov_metadata(c_rarg1, method());
1990 __ call_VM_leaf(
1991 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
1992 xthread, c_rarg1);
1993 restore_args(masm, total_c_args, c_arg, out_regs);
1994 __ j(dtrace_method_entry_done);
1995 __ block_comment("} dtrace entry");
1996 }
1997
1998 {
1999 __ block_comment("dtrace exit {");
2000 __ bind(dtrace_method_exit);
2001 save_native_result(masm, ret_type, stack_slots);
2002 __ mov_metadata(c_rarg1, method());
2003 __ call_VM_leaf(
2004 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
2005 xthread, c_rarg1);
2006 restore_native_result(masm, ret_type, stack_slots);
2007 __ j(dtrace_method_exit_done);
2008 __ block_comment("} dtrace exit");
2009 }
2010 }
2011
2012 __ flush();
2013
2014 nmethod *nm = nmethod::new_native_nmethod(method,
2015 compile_id,
2016 masm->code(),
2017 vep_offset,
2018 frame_complete,
2019 stack_slots / VMRegImpl::slots_per_word,
2020 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2021 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size),
2022 oop_maps);
2023 assert(nm != nullptr, "create native nmethod fail!");
2024 return nm;
2025 }
2026
2027 // this function returns the adjust size (in number of words) to a c2i adapter
2028 // activation for use during deoptimization
2029 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2030 assert(callee_locals >= callee_parameters,
2031 "test and remove; got more parms than locals");
2032 if (callee_locals < callee_parameters) {
2033 return 0; // No adjustment for negative locals
2034 }
2035 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords;
2036 // diff is counted in stack words
2037 return align_up(diff, 2);
2038 }
2039
2040 //------------------------------generate_deopt_blob----------------------------
2041 void SharedRuntime::generate_deopt_blob() {
2042 // Allocate space for the code
2043 ResourceMark rm;
2044 // Setup code generation tools
2045 int pad = 0;
2046 #if INCLUDE_JVMCI
2047 if (EnableJVMCI) {
2048 pad += 512; // Increase the buffer size when compiling for JVMCI
2049 }
2050 #endif
2051 const char* name = SharedRuntime::stub_name(StubId::shared_deopt_id);
2052 CodeBuffer buffer(name, 2048 + pad, 1024);
2053 MacroAssembler* masm = new MacroAssembler(&buffer);
2054 int frame_size_in_words = -1;
2055 OopMap* map = nullptr;
2056 OopMapSet *oop_maps = new OopMapSet();
2057 assert_cond(masm != nullptr && oop_maps != nullptr);
2058 RegisterSaver reg_saver(COMPILER2_OR_JVMCI != 0);
2059
2060 // -------------
2061 // This code enters when returning to a de-optimized nmethod. A return
2062 // address has been pushed on the stack, and return values are in
2063 // registers.
2064 // If we are doing a normal deopt then we were called from the patched
2065 // nmethod from the point we returned to the nmethod. So the return
2066 // address on the stack is wrong by NativeCall::instruction_size
2067 // We will adjust the value so it looks like we have the original return
2068 // address on the stack (like when we eagerly deoptimized).
2069 // In the case of an exception pending when deoptimizing, we enter
2070 // with a return address on the stack that points after the call we patched
2071 // into the exception handler. We have the following register state from,
2072 // e.g., the forward exception stub (see stubGenerator_riscv.cpp).
2073 // x10: exception oop
2074 // x9: exception handler
2075 // x13: throwing pc
2076 // So in this case we simply jam x13 into the useless return address and
2077 // the stack looks just like we want.
2078 //
2079 // At this point we need to de-opt. We save the argument return
2080 // registers. We call the first C routine, fetch_unroll_info(). This
2081 // routine captures the return values and returns a structure which
2082 // describes the current frame size and the sizes of all replacement frames.
2083 // The current frame is compiled code and may contain many inlined
2084 // functions, each with their own JVM state. We pop the current frame, then
2085 // push all the new frames. Then we call the C routine unpack_frames() to
2086 // populate these frames. Finally unpack_frames() returns us the new target
2087 // address. Notice that callee-save registers are BLOWN here; they have
2088 // already been captured in the vframeArray at the time the return PC was
2089 // patched.
2090 address start = __ pc();
2091 Label cont;
2092
2093 // Prolog for non exception case!
2094
2095 // Save everything in sight.
2096 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2097
2098 // Normal deoptimization. Save exec mode for unpack_frames.
2099 __ mv(xcpool, Deoptimization::Unpack_deopt); // callee-saved
2100 __ j(cont);
2101
2102 int reexecute_offset = __ pc() - start;
2103 #if INCLUDE_JVMCI && !defined(COMPILER1)
2104 if (UseJVMCICompiler) {
2105 // JVMCI does not use this kind of deoptimization
2106 __ should_not_reach_here();
2107 }
2108 #endif
2109
2110 // Reexecute case
2111 // return address is the pc describes what bci to do re-execute at
2112
2113 // No need to update map as each call to save_live_registers will produce identical oopmap
2114 (void) reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2115
2116 __ mv(xcpool, Deoptimization::Unpack_reexecute); // callee-saved
2117 __ j(cont);
2118
2119 #if INCLUDE_JVMCI
2120 Label after_fetch_unroll_info_call;
2121 int implicit_exception_uncommon_trap_offset = 0;
2122 int uncommon_trap_offset = 0;
2123
2124 if (EnableJVMCI) {
2125 implicit_exception_uncommon_trap_offset = __ pc() - start;
2126
2127 __ ld(ra, Address(xthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2128 __ sd(zr, Address(xthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2129
2130 uncommon_trap_offset = __ pc() - start;
2131
2132 // Save everything in sight.
2133 reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2134 // fetch_unroll_info needs to call last_java_frame()
2135 Label retaddr;
2136 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2137
2138 __ lw(c_rarg1, Address(xthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2139 __ mv(t0, -1);
2140 __ sw(t0, Address(xthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2141
2142 __ mv(xcpool, Deoptimization::Unpack_reexecute);
2143 __ mv(c_rarg0, xthread);
2144 __ orrw(c_rarg2, zr, xcpool); // exec mode
2145 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap));
2146 __ bind(retaddr);
2147 oop_maps->add_gc_map( __ pc()-start, map->deep_copy());
2148
2149 __ reset_last_Java_frame(false);
2150
2151 __ j(after_fetch_unroll_info_call);
2152 } // EnableJVMCI
2153 #endif // INCLUDE_JVMCI
2154
2155 int exception_offset = __ pc() - start;
2156
2157 // Prolog for exception case
2158
2159 // all registers are dead at this entry point, except for x10, and
2160 // x13 which contain the exception oop and exception pc
2161 // respectively. Set them in TLS and fall thru to the
2162 // unpack_with_exception_in_tls entry point.
2163
2164 __ sd(x13, Address(xthread, JavaThread::exception_pc_offset()));
2165 __ sd(x10, Address(xthread, JavaThread::exception_oop_offset()));
2166
2167 int exception_in_tls_offset = __ pc() - start;
2168
2169 // new implementation because exception oop is now passed in JavaThread
2170
2171 // Prolog for exception case
2172 // All registers must be preserved because they might be used by LinearScan
2173 // Exceptiop oop and throwing PC are passed in JavaThread
2174 // tos: stack at point of call to method that threw the exception (i.e. only
2175 // args are on the stack, no return address)
2176
2177 // The return address pushed by save_live_registers will be patched
2178 // later with the throwing pc. The correct value is not available
2179 // now because loading it from memory would destroy registers.
2180
2181 // NB: The SP at this point must be the SP of the method that is
2182 // being deoptimized. Deoptimization assumes that the frame created
2183 // here by save_live_registers is immediately below the method's SP.
2184 // This is a somewhat fragile mechanism.
2185
2186 // Save everything in sight.
2187 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2188
2189 // Now it is safe to overwrite any register
2190
2191 // Deopt during an exception. Save exec mode for unpack_frames.
2192 __ mv(xcpool, Deoptimization::Unpack_exception); // callee-saved
2193
2194 // load throwing pc from JavaThread and patch it as the return address
2195 // of the current frame. Then clear the field in JavaThread
2196
2197 __ ld(x13, Address(xthread, JavaThread::exception_pc_offset()));
2198 __ sd(x13, Address(fp, frame::return_addr_offset * wordSize));
2199 __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2200
2201 #ifdef ASSERT
2202 // verify that there is really an exception oop in JavaThread
2203 __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2204 __ verify_oop(x10);
2205
2206 // verify that there is no pending exception
2207 Label no_pending_exception;
2208 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2209 __ beqz(t0, no_pending_exception);
2210 __ stop("must not have pending exception here");
2211 __ bind(no_pending_exception);
2212 #endif
2213
2214 __ bind(cont);
2215
2216 // Call C code. Need thread and this frame, but NOT official VM entry
2217 // crud. We cannot block on this call, no GC can happen.
2218 //
2219 // UnrollBlock* fetch_unroll_info(JavaThread* thread)
2220
2221 // fetch_unroll_info needs to call last_java_frame().
2222
2223 Label retaddr;
2224 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2225 #ifdef ASSERT
2226 {
2227 Label L;
2228 __ ld(t0, Address(xthread,
2229 JavaThread::last_Java_fp_offset()));
2230 __ beqz(t0, L);
2231 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared");
2232 __ bind(L);
2233 }
2234 #endif // ASSERT
2235 __ mv(c_rarg0, xthread);
2236 __ mv(c_rarg1, xcpool);
2237 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info));
2238 __ bind(retaddr);
2239
2240 // Need to have an oopmap that tells fetch_unroll_info where to
2241 // find any register it might need.
2242 oop_maps->add_gc_map(__ pc() - start, map);
2243
2244 __ reset_last_Java_frame(false);
2245
2246 #if INCLUDE_JVMCI
2247 if (EnableJVMCI) {
2248 __ bind(after_fetch_unroll_info_call);
2249 }
2250 #endif
2251
2252 // Load UnrollBlock* into x15
2253 __ mv(x15, x10);
2254
2255 __ lwu(xcpool, Address(x15, Deoptimization::UnrollBlock::unpack_kind_offset()));
2256 Label noException;
2257 __ mv(t0, Deoptimization::Unpack_exception);
2258 __ bne(xcpool, t0, noException); // Was exception pending?
2259 __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2260 __ ld(x13, Address(xthread, JavaThread::exception_pc_offset()));
2261 __ sd(zr, Address(xthread, JavaThread::exception_oop_offset()));
2262 __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2263
2264 __ verify_oop(x10);
2265
2266 // Overwrite the result registers with the exception results.
2267 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2268
2269 __ bind(noException);
2270
2271 // Only register save data is on the stack.
2272 // Now restore the result registers. Everything else is either dead
2273 // or captured in the vframeArray.
2274
2275 // Restore fp result register
2276 __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2277 // Restore integer result register
2278 __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2279
2280 // Pop all of the register save area off the stack
2281 __ add(sp, sp, frame_size_in_words * wordSize);
2282
2283 // All of the register save area has been popped of the stack. Only the
2284 // return address remains.
2285
2286 // Pop all the frames we must move/replace.
2287 //
2288 // Frame picture (youngest to oldest)
2289 // 1: self-frame (no frame link)
2290 // 2: deopting frame (no frame link)
2291 // 3: caller of deopting frame (could be compiled/interpreted).
2292 //
2293 // Note: by leaving the return address of self-frame on the stack
2294 // and using the size of frame 2 to adjust the stack
2295 // when we are done the return to frame 3 will still be on the stack.
2296
2297 // Pop deoptimized frame
2298 __ lwu(x12, Address(x15, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset()));
2299 __ subi(x12, x12, 2 * wordSize);
2300 __ add(sp, sp, x12);
2301 __ ld(fp, Address(sp, 0));
2302 __ ld(ra, Address(sp, wordSize));
2303 __ addi(sp, sp, 2 * wordSize);
2304 // RA should now be the return address to the caller (3)
2305
2306 #ifdef ASSERT
2307 // Compilers generate code that bang the stack by as much as the
2308 // interpreter would need. So this stack banging should never
2309 // trigger a fault. Verify that it does not on non product builds.
2310 __ lwu(x9, Address(x15, Deoptimization::UnrollBlock::total_frame_sizes_offset()));
2311 __ bang_stack_size(x9, x12);
2312 #endif
2313 // Load address of array of frame pcs into x12
2314 __ ld(x12, Address(x15, Deoptimization::UnrollBlock::frame_pcs_offset()));
2315
2316 // Load address of array of frame sizes into x14
2317 __ ld(x14, Address(x15, Deoptimization::UnrollBlock::frame_sizes_offset()));
2318
2319 // Load counter into x13
2320 __ lwu(x13, Address(x15, Deoptimization::UnrollBlock::number_of_frames_offset()));
2321
2322 // Now adjust the caller's stack to make up for the extra locals
2323 // but record the original sp so that we can save it in the skeletal interpreter
2324 // frame and the stack walking of interpreter_sender will get the unextended sp
2325 // value and not the "real" sp value.
2326
2327 const Register sender_sp = x16;
2328
2329 __ mv(sender_sp, sp);
2330 __ lwu(x9, Address(x15,
2331 Deoptimization::UnrollBlock::
2332 caller_adjustment_offset()));
2333 __ sub(sp, sp, x9);
2334
2335 // Push interpreter frames in a loop
2336 __ mv(t0, 0xDEADDEAD); // Make a recognizable pattern
2337 __ mv(t1, t0);
2338 Label loop;
2339 __ bind(loop);
2340 __ ld(x9, Address(x14, 0)); // Load frame size
2341 __ addi(x14, x14, wordSize);
2342 __ subi(x9, x9, 2 * wordSize); // We'll push pc and fp by hand
2343 __ ld(ra, Address(x12, 0)); // Load pc
2344 __ addi(x12, x12, wordSize);
2345 __ enter(); // Save old & set new fp
2346 __ sub(sp, sp, x9); // Prolog
2347 // This value is corrected by layout_activation_impl
2348 __ sd(zr, Address(fp, frame::interpreter_frame_last_sp_offset * wordSize));
2349 __ sd(sender_sp, Address(fp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable
2350 __ mv(sender_sp, sp); // Pass sender_sp to next frame
2351 __ subi(x13, x13, 1); // Decrement counter
2352 __ bnez(x13, loop);
2353
2354 // Re-push self-frame
2355 __ ld(ra, Address(x12));
2356 __ enter();
2357
2358 // Allocate a full sized register save area. We subtract 2 because
2359 // enter() just pushed 2 words
2360 __ sub(sp, sp, (frame_size_in_words - 2) * wordSize);
2361
2362 // Restore frame locals after moving the frame
2363 __ fsd(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2364 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2365
2366 // Call C code. Need thread but NOT official VM entry
2367 // crud. We cannot block on this call, no GC can happen. Call should
2368 // restore return values to their stack-slots with the new SP.
2369 //
2370 // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)
2371
2372 // Use fp because the frames look interpreted now
2373 // Don't need the precise return PC here, just precise enough to point into this code blob.
2374 address the_pc = __ pc();
2375 __ set_last_Java_frame(sp, fp, the_pc, t0);
2376
2377 __ mv(c_rarg0, xthread);
2378 __ mv(c_rarg1, xcpool); // second arg: exec_mode
2379 __ rt_call(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames));
2380
2381 // Set an oopmap for the call site
2382 // Use the same PC we used for the last java frame
2383 oop_maps->add_gc_map(the_pc - start,
2384 new OopMap(frame_size_in_words, 0));
2385
2386 // Clear fp AND pc
2387 __ reset_last_Java_frame(true);
2388
2389 // Collect return values
2390 __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2391 __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2392
2393 // Pop self-frame.
2394 __ leave(); // Epilog
2395
2396 // Jump to interpreter
2397 __ ret();
2398
2399 // Make sure all code is generated
2400 masm->flush();
2401
2402 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words);
2403 assert(_deopt_blob != nullptr, "create deoptimization blob fail!");
2404 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
2405 #if INCLUDE_JVMCI
2406 if (EnableJVMCI) {
2407 _deopt_blob->set_uncommon_trap_offset(uncommon_trap_offset);
2408 _deopt_blob->set_implicit_exception_uncommon_trap_offset(implicit_exception_uncommon_trap_offset);
2409 }
2410 #endif
2411 }
2412
2413 // Number of stack slots between incoming argument block and the start of
2414 // a new frame. The PROLOG must add this many slots to the stack. The
2415 // EPILOG must remove this many slots.
2416 // RISCV needs two words for RA (return address) and FP (frame pointer).
2417 uint SharedRuntime::in_preserve_stack_slots() {
2418 return 2 * VMRegImpl::slots_per_word;
2419 }
2420
2421 uint SharedRuntime::out_preserve_stack_slots() {
2422 return 0;
2423 }
2424
2425 VMReg SharedRuntime::thread_register() {
2426 return xthread->as_VMReg();
2427 }
2428
2429 //------------------------------generate_handler_blob------
2430 //
2431 // Generate a special Compile2Runtime blob that saves all registers,
2432 // and setup oopmap.
2433 //
2434 SafepointBlob* SharedRuntime::generate_handler_blob(StubId id, address call_ptr) {
2435 assert(is_polling_page_id(id), "expected a polling page stub id");
2436
2437 ResourceMark rm;
2438 OopMapSet *oop_maps = new OopMapSet();
2439 assert_cond(oop_maps != nullptr);
2440 OopMap* map = nullptr;
2441
2442 // Allocate space for the code. Setup code generation tools.
2443 const char* name = SharedRuntime::stub_name(id);
2444 CodeBuffer buffer(name, 2048, 1024);
2445 MacroAssembler* masm = new MacroAssembler(&buffer);
2446 assert_cond(masm != nullptr);
2447
2448 address start = __ pc();
2449 address call_pc = nullptr;
2450 int frame_size_in_words = -1;
2451 bool cause_return = (id == StubId::shared_polling_page_return_handler_id);
2452 RegisterSaver reg_saver(id == StubId::shared_polling_page_vectors_safepoint_handler_id /* save_vectors */);
2453
2454 // Save Integer and Float registers.
2455 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2456
2457 // The following is basically a call_VM. However, we need the precise
2458 // address of the call in order to generate an oopmap. Hence, we do all the
2459 // work ourselves.
2460
2461 Label retaddr;
2462 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2463
2464 // The return address must always be correct so that frame constructor never
2465 // sees an invalid pc.
2466
2467 if (!cause_return) {
2468 // overwrite the return address pushed by save_live_registers
2469 // Additionally, x18 is a callee-saved register so we can look at
2470 // it later to determine if someone changed the return address for
2471 // us!
2472 __ ld(x18, Address(xthread, JavaThread::saved_exception_pc_offset()));
2473 __ sd(x18, Address(fp, frame::return_addr_offset * wordSize));
2474 }
2475
2476 // Do the call
2477 __ mv(c_rarg0, xthread);
2478 __ rt_call(call_ptr);
2479 __ bind(retaddr);
2480
2481 // Set an oopmap for the call site. This oopmap will map all
2482 // oop-registers and debug-info registers as callee-saved. This
2483 // will allow deoptimization at this safepoint to find all possible
2484 // debug-info recordings, as well as let GC find all oops.
2485
2486 oop_maps->add_gc_map( __ pc() - start, map);
2487
2488 Label noException;
2489
2490 __ reset_last_Java_frame(false);
2491
2492 __ membar(MacroAssembler::LoadLoad | MacroAssembler::LoadStore);
2493
2494 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2495 __ beqz(t0, noException);
2496
2497 // Exception pending
2498
2499 reg_saver.restore_live_registers(masm);
2500
2501 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2502
2503 // No exception case
2504 __ bind(noException);
2505
2506 Label no_adjust, bail;
2507 if (!cause_return) {
2508 // If our stashed return pc was modified by the runtime we avoid touching it
2509 __ ld(t0, Address(fp, frame::return_addr_offset * wordSize));
2510 __ bne(x18, t0, no_adjust);
2511
2512 #ifdef ASSERT
2513 // Verify the correct encoding of the poll we're about to skip.
2514 // See NativeInstruction::is_lwu_to_zr()
2515 __ lwu(t0, Address(x18));
2516 __ andi(t1, t0, 0b1111111);
2517 __ mv(t2, 0b0000011);
2518 __ bne(t1, t2, bail); // 0-6:0b0000011
2519 __ srli(t1, t0, 7);
2520 __ andi(t1, t1, 0b11111);
2521 __ bnez(t1, bail); // 7-11:0b00000
2522 __ srli(t1, t0, 12);
2523 __ andi(t1, t1, 0b111);
2524 __ mv(t2, 0b110);
2525 __ bne(t1, t2, bail); // 12-14:0b110
2526 #endif
2527
2528 // Adjust return pc forward to step over the safepoint poll instruction
2529 __ addi(x18, x18, NativeInstruction::instruction_size);
2530 __ sd(x18, Address(fp, frame::return_addr_offset * wordSize));
2531 }
2532
2533 __ bind(no_adjust);
2534 // Normal exit, restore registers and exit.
2535
2536 reg_saver.restore_live_registers(masm);
2537 __ ret();
2538
2539 #ifdef ASSERT
2540 __ bind(bail);
2541 __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected");
2542 #endif
2543
2544 // Make sure all code is generated
2545 masm->flush();
2546
2547 // Fill-out other meta info
2548 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words);
2549 }
2550
2551 //
2552 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
2553 //
2554 // Generate a stub that calls into vm to find out the proper destination
2555 // of a java call. All the argument registers are live at this point
2556 // but since this is generic code we don't know what they are and the caller
2557 // must do any gc of the args.
2558 //
2559 RuntimeStub* SharedRuntime::generate_resolve_blob(StubId id, address destination) {
2560 assert(StubRoutines::forward_exception_entry() != nullptr, "must be generated before");
2561 assert(is_resolve_id(id), "expected a resolve stub id");
2562
2563 // allocate space for the code
2564 ResourceMark rm;
2565
2566 const char* name = SharedRuntime::stub_name(id);
2567 CodeBuffer buffer(name, 1000, 512);
2568 MacroAssembler* masm = new MacroAssembler(&buffer);
2569 assert_cond(masm != nullptr);
2570
2571 int frame_size_in_words = -1;
2572 RegisterSaver reg_saver(false /* save_vectors */);
2573
2574 OopMapSet *oop_maps = new OopMapSet();
2575 assert_cond(oop_maps != nullptr);
2576 OopMap* map = nullptr;
2577
2578 int start = __ offset();
2579
2580 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2581
2582 int frame_complete = __ offset();
2583
2584 {
2585 Label retaddr;
2586 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2587
2588 __ mv(c_rarg0, xthread);
2589 __ rt_call(destination);
2590 __ bind(retaddr);
2591 }
2592
2593 // Set an oopmap for the call site.
2594 // We need this not only for callee-saved registers, but also for volatile
2595 // registers that the compiler might be keeping live across a safepoint.
2596
2597 oop_maps->add_gc_map( __ offset() - start, map);
2598
2599 // x10 contains the address we are going to jump to assuming no exception got installed
2600
2601 // clear last_Java_sp
2602 __ reset_last_Java_frame(false);
2603 // check for pending exceptions
2604 Label pending;
2605 __ ld(t1, Address(xthread, Thread::pending_exception_offset()));
2606 __ bnez(t1, pending);
2607
2608 // get the returned Method*
2609 __ get_vm_result_metadata(xmethod, xthread);
2610 __ sd(xmethod, Address(sp, reg_saver.reg_offset_in_bytes(xmethod)));
2611
2612 // x10 is where we want to jump, overwrite t1 which is saved and temporary
2613 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(t1)));
2614 reg_saver.restore_live_registers(masm);
2615
2616 // We are back to the original state on entry and ready to go.
2617 __ jr(t1);
2618
2619 // Pending exception after the safepoint
2620
2621 __ bind(pending);
2622
2623 reg_saver.restore_live_registers(masm);
2624
2625 // exception pending => remove activation and forward to exception handler
2626
2627 __ sd(zr, Address(xthread, JavaThread::vm_result_oop_offset()));
2628
2629 __ ld(x10, Address(xthread, Thread::pending_exception_offset()));
2630 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2631
2632 // -------------
2633 // make sure all code is generated
2634 masm->flush();
2635
2636 // return the blob
2637 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true);
2638 }
2639
2640 // Continuation point for throwing of implicit exceptions that are
2641 // not handled in the current activation. Fabricates an exception
2642 // oop and initiates normal exception dispatching in this
2643 // frame. Since we need to preserve callee-saved values (currently
2644 // only for C2, but done for C1 as well) we need a callee-saved oop
2645 // map and therefore have to make these stubs into RuntimeStubs
2646 // rather than BufferBlobs. If the compiler needs all registers to
2647 // be preserved between the fault point and the exception handler
2648 // then it must assume responsibility for that in
2649 // AbstractCompiler::continuation_for_implicit_null_exception or
2650 // continuation_for_implicit_division_by_zero_exception. All other
2651 // implicit exceptions (e.g., NullPointerException or
2652 // AbstractMethodError on entry) are either at call sites or
2653 // otherwise assume that stack unwinding will be initiated, so
2654 // caller saved registers were assumed volatile in the compiler.
2655
2656 RuntimeStub* SharedRuntime::generate_throw_exception(StubId id, address runtime_entry) {
2657 assert(is_throw_id(id), "expected a throw stub id");
2658
2659 const char* name = SharedRuntime::stub_name(id);
2660
2661 // Information about frame layout at time of blocking runtime call.
2662 // Note that we only have to preserve callee-saved registers since
2663 // the compilers are responsible for supplying a continuation point
2664 // if they expect all registers to be preserved.
2665 // n.b. riscv asserts that frame::arg_reg_save_area_bytes == 0
2666 assert_cond(runtime_entry != nullptr);
2667 enum layout {
2668 fp_off = 0,
2669 fp_off2,
2670 return_off,
2671 return_off2,
2672 framesize // inclusive of return address
2673 };
2674
2675 const int insts_size = 1024;
2676 const int locs_size = 64;
2677
2678 ResourceMark rm;
2679 const char* timer_msg = "SharedRuntime generate_throw_exception";
2680 TraceTime timer(timer_msg, TRACETIME_LOG(Info, startuptime));
2681
2682 CodeBuffer code(name, insts_size, locs_size);
2683 OopMapSet* oop_maps = new OopMapSet();
2684 MacroAssembler* masm = new MacroAssembler(&code);
2685 assert_cond(oop_maps != nullptr && masm != nullptr);
2686
2687 address start = __ pc();
2688
2689 // This is an inlined and slightly modified version of call_VM
2690 // which has the ability to fetch the return PC out of
2691 // thread-local storage and also sets up last_Java_sp slightly
2692 // differently than the real call_VM
2693
2694 __ enter(); // Save FP and RA before call
2695
2696 assert(is_even(framesize / 2), "sp not 16-byte aligned");
2697
2698 // ra and fp are already in place
2699 __ subi(sp, fp, (unsigned)framesize << LogBytesPerInt); // prolog
2700
2701 int frame_complete = __ pc() - start;
2702
2703 // Set up last_Java_sp and last_Java_fp
2704 address the_pc = __ pc();
2705 __ set_last_Java_frame(sp, fp, the_pc, t0);
2706
2707 // Call runtime
2708 __ mv(c_rarg0, xthread);
2709 BLOCK_COMMENT("call runtime_entry");
2710 __ rt_call(runtime_entry);
2711
2712 // Generate oop map
2713 OopMap* map = new OopMap(framesize, 0);
2714 assert_cond(map != nullptr);
2715
2716 oop_maps->add_gc_map(the_pc - start, map);
2717
2718 __ reset_last_Java_frame(true);
2719
2720 __ leave();
2721
2722 // check for pending exceptions
2723 #ifdef ASSERT
2724 Label L;
2725 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2726 __ bnez(t0, L);
2727 __ should_not_reach_here();
2728 __ bind(L);
2729 #endif // ASSERT
2730 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2731
2732 // codeBlob framesize is in words (not VMRegImpl::slot_size)
2733 RuntimeStub* stub =
2734 RuntimeStub::new_runtime_stub(name,
2735 &code,
2736 frame_complete,
2737 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2738 oop_maps, false);
2739 assert(stub != nullptr, "create runtime stub fail!");
2740 return stub;
2741 }
2742
2743 #if INCLUDE_JFR
2744
2745 static void jfr_prologue(address the_pc, MacroAssembler* masm, Register thread) {
2746 __ set_last_Java_frame(sp, fp, the_pc, t0);
2747 __ mv(c_rarg0, thread);
2748 }
2749
2750 static void jfr_epilogue(MacroAssembler* masm) {
2751 __ reset_last_Java_frame(true);
2752 }
2753 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
2754 // It returns a jobject handle to the event writer.
2755 // The handle is dereferenced and the return value is the event writer oop.
2756 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
2757 enum layout {
2758 fp_off,
2759 fp_off2,
2760 return_off,
2761 return_off2,
2762 framesize // inclusive of return address
2763 };
2764
2765 int insts_size = 1024;
2766 int locs_size = 64;
2767 const char* name = SharedRuntime::stub_name(StubId::shared_jfr_write_checkpoint_id);
2768 CodeBuffer code(name, insts_size, locs_size);
2769 OopMapSet* oop_maps = new OopMapSet();
2770 MacroAssembler* masm = new MacroAssembler(&code);
2771
2772 address start = __ pc();
2773 __ enter();
2774 int frame_complete = __ pc() - start;
2775 address the_pc = __ pc();
2776 jfr_prologue(the_pc, masm, xthread);
2777 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), 1);
2778
2779 jfr_epilogue(masm);
2780 __ resolve_global_jobject(x10, t0, t1);
2781 __ leave();
2782 __ ret();
2783
2784 OopMap* map = new OopMap(framesize, 1);
2785 oop_maps->add_gc_map(the_pc - start, map);
2786
2787 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
2788 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
2789 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2790 oop_maps, false);
2791 return stub;
2792 }
2793
2794 // For c2: call to return a leased buffer.
2795 RuntimeStub* SharedRuntime::generate_jfr_return_lease() {
2796 enum layout {
2797 fp_off,
2798 fp_off2,
2799 return_off,
2800 return_off2,
2801 framesize // inclusive of return address
2802 };
2803
2804 int insts_size = 1024;
2805 int locs_size = 64;
2806 const char* name = SharedRuntime::stub_name(StubId::shared_jfr_return_lease_id);
2807 CodeBuffer code(name, insts_size, locs_size);
2808 OopMapSet* oop_maps = new OopMapSet();
2809 MacroAssembler* masm = new MacroAssembler(&code);
2810
2811 address start = __ pc();
2812 __ enter();
2813 int frame_complete = __ pc() - start;
2814 address the_pc = __ pc();
2815 jfr_prologue(the_pc, masm, xthread);
2816 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::return_lease), 1);
2817
2818 jfr_epilogue(masm);
2819 __ leave();
2820 __ ret();
2821
2822 OopMap* map = new OopMap(framesize, 1);
2823 oop_maps->add_gc_map(the_pc - start, map);
2824
2825 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
2826 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
2827 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
2828 oop_maps, false);
2829 return stub;
2830 }
2831
2832 #endif // INCLUDE_JFR