1 /*
2 * Copyright (c) 2003, 2023, 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 "precompiled.hpp"
28 #include "asm/macroAssembler.hpp"
29 #include "asm/macroAssembler.inline.hpp"
30 #include "code/compiledIC.hpp"
31 #include "code/debugInfoRec.hpp"
32 #include "code/icBuffer.hpp"
33 #include "code/vtableStubs.hpp"
34 #include "compiler/oopMap.hpp"
35 #include "gc/shared/barrierSetAssembler.hpp"
36 #include "interpreter/interp_masm.hpp"
37 #include "interpreter/interpreter.hpp"
38 #include "logging/log.hpp"
39 #include "memory/resourceArea.hpp"
40 #include "nativeInst_riscv.hpp"
41 #include "oops/compiledICHolder.hpp"
42 #include "oops/klass.inline.hpp"
43 #include "oops/method.inline.hpp"
44 #include "prims/methodHandles.hpp"
45 #include "runtime/continuation.hpp"
46 #include "runtime/continuationEntry.inline.hpp"
47 #include "runtime/globals.hpp"
48 #include "runtime/jniHandles.hpp"
49 #include "runtime/safepointMechanism.hpp"
50 #include "runtime/sharedRuntime.hpp"
51 #include "runtime/signature.hpp"
52 #include "runtime/stubRoutines.hpp"
53 #include "runtime/vframeArray.hpp"
54 #include "utilities/align.hpp"
55 #include "utilities/formatBuffer.hpp"
56 #include "vmreg_riscv.inline.hpp"
57 #ifdef COMPILER1
58 #include "c1/c1_Runtime1.hpp"
59 #endif
60 #ifdef COMPILER2
61 #include "adfiles/ad_riscv.hpp"
62 #include "opto/runtime.hpp"
63 #endif
64 #if INCLUDE_JVMCI
65 #include "jvmci/jvmciJavaClasses.hpp"
66 #endif
67
68 #define __ masm->
69
70 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size;
71
72 class SimpleRuntimeFrame {
73 public:
74
75 // Most of the runtime stubs have this simple frame layout.
76 // This class exists to make the layout shared in one place.
77 // Offsets are for compiler stack slots, which are jints.
78 enum layout {
79 // The frame sender code expects that fp will be in the "natural" place and
80 // will override any oopMap setting for it. We must therefore force the layout
81 // so that it agrees with the frame sender code.
82 // we don't expect any arg reg save area so riscv asserts that
83 // frame::arg_reg_save_area_bytes == 0
84 fp_off = 0, fp_off2,
85 return_off, return_off2,
86 framesize
87 };
88 };
89
90 class RegisterSaver {
91 const bool _save_vectors;
92 public:
93 RegisterSaver(bool save_vectors) : _save_vectors(UseRVV && save_vectors) {}
94 ~RegisterSaver() {}
95 OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words);
96 void restore_live_registers(MacroAssembler* masm);
97
98 // Offsets into the register save area
99 // Used by deoptimization when it is managing result register
100 // values on its own
101 // gregs:28, float_register:32; except: x1(ra) & x2(sp) & gp(x3) & tp(x4)
102 // |---v0---|<---SP
103 // |---v1---|save vectors only in generate_handler_blob
104 // |-- .. --|
105 // |---v31--|-----
106 // |---f0---|
107 // |---f1---|
108 // | .. |
109 // |---f31--|
110 // |---reserved slot for stack alignment---|
111 // |---x5---|
112 // | x6 |
113 // |---.. --|
114 // |---x31--|
115 // |---fp---|
116 // |---ra---|
117 int v0_offset_in_bytes(void) { return 0; }
118 int f0_offset_in_bytes(void) {
119 int f0_offset = 0;
120 #ifdef COMPILER2
121 if (_save_vectors) {
122 f0_offset += Matcher::scalable_vector_reg_size(T_INT) * VectorRegister::number_of_registers *
123 BytesPerInt;
124 }
125 #endif
126 return f0_offset;
127 }
128 int reserved_slot_offset_in_bytes(void) {
129 return f0_offset_in_bytes() +
130 FloatRegister::max_slots_per_register *
131 FloatRegister::number_of_registers *
132 BytesPerInt;
133 }
134
135 int reg_offset_in_bytes(Register r) {
136 assert (r->encoding() > 4, "ra, sp, gp and tp not saved");
137 return reserved_slot_offset_in_bytes() + (r->encoding() - 4 /* x1, x2, x3, x4 */) * wordSize;
138 }
139
140 int freg_offset_in_bytes(FloatRegister f) {
141 return f0_offset_in_bytes() + f->encoding() * wordSize;
142 }
143
144 int ra_offset_in_bytes(void) {
145 return reserved_slot_offset_in_bytes() +
146 (Register::number_of_registers - 3) *
147 Register::max_slots_per_register *
148 BytesPerInt;
149 }
150 };
151
152 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words, int* total_frame_words) {
153 int vector_size_in_bytes = 0;
154 int vector_size_in_slots = 0;
155 #ifdef COMPILER2
156 if (_save_vectors) {
157 vector_size_in_bytes += Matcher::scalable_vector_reg_size(T_BYTE);
158 vector_size_in_slots += Matcher::scalable_vector_reg_size(T_INT);
159 }
160 #endif
161
162 int frame_size_in_bytes = align_up(additional_frame_words * wordSize + ra_offset_in_bytes() + wordSize, 16);
163 // OopMap frame size is in compiler stack slots (jint's) not bytes or words
164 int frame_size_in_slots = frame_size_in_bytes / BytesPerInt;
165 // The caller will allocate additional_frame_words
166 int additional_frame_slots = additional_frame_words * wordSize / BytesPerInt;
167 // CodeBlob frame size is in words.
168 int frame_size_in_words = frame_size_in_bytes / wordSize;
169 *total_frame_words = frame_size_in_words;
170
171 // Save Integer, Float and Vector registers.
172 __ enter();
173 __ push_CPU_state(_save_vectors, vector_size_in_bytes);
174
175 // Set an oopmap for the call site. This oopmap will map all
176 // oop-registers and debug-info registers as callee-saved. This
177 // will allow deoptimization at this safepoint to find all possible
178 // debug-info recordings, as well as let GC find all oops.
179
180 OopMapSet *oop_maps = new OopMapSet();
181 OopMap* oop_map = new OopMap(frame_size_in_slots, 0);
182 assert_cond(oop_maps != nullptr && oop_map != nullptr);
183
184 int sp_offset_in_slots = 0;
185 int step_in_slots = 0;
186 if (_save_vectors) {
187 step_in_slots = vector_size_in_slots;
188 for (int i = 0; i < VectorRegister::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
189 VectorRegister r = as_VectorRegister(i);
190 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots), r->as_VMReg());
191 }
192 }
193
194 step_in_slots = FloatRegister::max_slots_per_register;
195 for (int i = 0; i < FloatRegister::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
196 FloatRegister r = as_FloatRegister(i);
197 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots), r->as_VMReg());
198 }
199
200 step_in_slots = Register::max_slots_per_register;
201 // skip the slot reserved for alignment, see MacroAssembler::push_reg;
202 // also skip x5 ~ x6 on the stack because they are caller-saved registers.
203 sp_offset_in_slots += Register::max_slots_per_register * 3;
204 // besides, we ignore x0 ~ x4 because push_CPU_state won't push them on the stack.
205 for (int i = 7; i < Register::number_of_registers; i++, sp_offset_in_slots += step_in_slots) {
206 Register r = as_Register(i);
207 if (r != xthread) {
208 oop_map->set_callee_saved(VMRegImpl::stack2reg(sp_offset_in_slots + additional_frame_slots), r->as_VMReg());
209 }
210 }
211
212 return oop_map;
213 }
214
215 void RegisterSaver::restore_live_registers(MacroAssembler* masm) {
216 #ifdef COMPILER2
217 __ pop_CPU_state(_save_vectors, Matcher::scalable_vector_reg_size(T_BYTE));
218 #else
219 #if !INCLUDE_JVMCI
220 assert(!_save_vectors, "vectors are generated only by C2 and JVMCI");
221 #endif
222 __ pop_CPU_state(_save_vectors);
223 #endif
224 __ leave();
225 }
226
227 // Is vector's size (in bytes) bigger than a size saved by default?
228 // riscv does not ovlerlay the floating-point registers on vector registers like aarch64.
229 bool SharedRuntime::is_wide_vector(int size) {
230 return UseRVV;
231 }
232
233 // ---------------------------------------------------------------------------
234 // Read the array of BasicTypes from a signature, and compute where the
235 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
236 // quantities. Values less than VMRegImpl::stack0 are registers, those above
237 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
238 // as framesizes are fixed.
239 // VMRegImpl::stack0 refers to the first slot 0(sp).
240 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher.
241 // Register up to Register::number_of_registers) are the 64-bit
242 // integer registers.
243
244 // Note: the INPUTS in sig_bt are in units of Java argument words,
245 // which are 64-bit. The OUTPUTS are in 32-bit units.
246
247 // The Java calling convention is a "shifted" version of the C ABI.
248 // By skipping the first C ABI register we can call non-static jni
249 // methods with small numbers of arguments without having to shuffle
250 // the arguments at all. Since we control the java ABI we ought to at
251 // least get some advantage out of it.
252
253 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
254 VMRegPair *regs,
255 int total_args_passed) {
256 // Create the mapping between argument positions and
257 // registers.
258 static const Register INT_ArgReg[Argument::n_int_register_parameters_j] = {
259 j_rarg0, j_rarg1, j_rarg2, j_rarg3,
260 j_rarg4, j_rarg5, j_rarg6, j_rarg7
261 };
262 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_j] = {
263 j_farg0, j_farg1, j_farg2, j_farg3,
264 j_farg4, j_farg5, j_farg6, j_farg7
265 };
266
267 uint int_args = 0;
268 uint fp_args = 0;
269 uint stk_args = 0; // inc by 2 each time
270
271 for (int i = 0; i < total_args_passed; i++) {
272 switch (sig_bt[i]) {
273 case T_BOOLEAN: // fall through
274 case T_CHAR: // fall through
275 case T_BYTE: // fall through
276 case T_SHORT: // fall through
277 case T_INT:
278 if (int_args < Argument::n_int_register_parameters_j) {
279 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
280 } else {
281 regs[i].set1(VMRegImpl::stack2reg(stk_args));
282 stk_args += 2;
283 }
284 break;
285 case T_VOID:
286 // halves of T_LONG or T_DOUBLE
287 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
288 regs[i].set_bad();
289 break;
290 case T_LONG: // fall through
291 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
292 case T_OBJECT: // fall through
293 case T_ARRAY: // fall through
294 case T_ADDRESS:
295 if (int_args < Argument::n_int_register_parameters_j) {
296 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
297 } else {
298 regs[i].set2(VMRegImpl::stack2reg(stk_args));
299 stk_args += 2;
300 }
301 break;
302 case T_FLOAT:
303 if (fp_args < Argument::n_float_register_parameters_j) {
304 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
305 } else {
306 regs[i].set1(VMRegImpl::stack2reg(stk_args));
307 stk_args += 2;
308 }
309 break;
310 case T_DOUBLE:
311 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
312 if (fp_args < Argument::n_float_register_parameters_j) {
313 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
314 } else {
315 regs[i].set2(VMRegImpl::stack2reg(stk_args));
316 stk_args += 2;
317 }
318 break;
319 default:
320 ShouldNotReachHere();
321 }
322 }
323
324 return align_up(stk_args, 2);
325 }
326
327 // Patch the callers callsite with entry to compiled code if it exists.
328 static void patch_callers_callsite(MacroAssembler *masm) {
329 Label L;
330 __ ld(t0, Address(xmethod, in_bytes(Method::code_offset())));
331 __ beqz(t0, L);
332
333 __ enter();
334 __ push_CPU_state();
335
336 // VM needs caller's callsite
337 // VM needs target method
338 // This needs to be a long call since we will relocate this adapter to
339 // the codeBuffer and it may not reach
340
341 #ifndef PRODUCT
342 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
343 #endif
344
345 __ mv(c_rarg0, xmethod);
346 __ mv(c_rarg1, ra);
347 RuntimeAddress target(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite));
348 __ relocate(target.rspec(), [&] {
349 int32_t offset;
350 __ la_patchable(t0, target, offset);
351 __ jalr(x1, t0, offset);
352 });
353
354 __ pop_CPU_state();
355 // restore sp
356 __ leave();
357 __ bind(L);
358 }
359
360 static void gen_c2i_adapter(MacroAssembler *masm,
361 int total_args_passed,
362 int comp_args_on_stack,
363 const BasicType *sig_bt,
364 const VMRegPair *regs,
365 Label& skip_fixup) {
366 // Before we get into the guts of the C2I adapter, see if we should be here
367 // at all. We've come from compiled code and are attempting to jump to the
368 // interpreter, which means the caller made a static call to get here
369 // (vcalls always get a compiled target if there is one). Check for a
370 // compiled target. If there is one, we need to patch the caller's call.
371 patch_callers_callsite(masm);
372
373 __ bind(skip_fixup);
374
375 int words_pushed = 0;
376
377 // Since all args are passed on the stack, total_args_passed *
378 // Interpreter::stackElementSize is the space we need.
379
380 int extraspace = total_args_passed * Interpreter::stackElementSize;
381
382 __ mv(x19_sender_sp, sp);
383
384 // stack is aligned, keep it that way
385 extraspace = align_up(extraspace, 2 * wordSize);
386
387 if (extraspace) {
388 __ sub(sp, sp, extraspace);
389 }
390
391 // Now write the args into the outgoing interpreter space
392 for (int i = 0; i < total_args_passed; i++) {
393 if (sig_bt[i] == T_VOID) {
394 assert(i > 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "missing half");
395 continue;
396 }
397
398 // offset to start parameters
399 int st_off = (total_args_passed - i - 1) * Interpreter::stackElementSize;
400 int next_off = st_off - Interpreter::stackElementSize;
401
402 // Say 4 args:
403 // i st_off
404 // 0 32 T_LONG
405 // 1 24 T_VOID
406 // 2 16 T_OBJECT
407 // 3 8 T_BOOL
408 // - 0 return address
409 //
410 // However to make thing extra confusing. Because we can fit a Java long/double in
411 // a single slot on a 64 bt vm and it would be silly to break them up, the interpreter
412 // leaves one slot empty and only stores to a single slot. In this case the
413 // slot that is occupied is the T_VOID slot. See I said it was confusing.
414
415 VMReg r_1 = regs[i].first();
416 VMReg r_2 = regs[i].second();
417 if (!r_1->is_valid()) {
418 assert(!r_2->is_valid(), "");
419 continue;
420 }
421 if (r_1->is_stack()) {
422 // memory to memory use t0
423 int ld_off = (r_1->reg2stack() * VMRegImpl::stack_slot_size
424 + extraspace
425 + words_pushed * wordSize);
426 if (!r_2->is_valid()) {
427 __ lwu(t0, Address(sp, ld_off));
428 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
429 } else {
430 __ ld(t0, Address(sp, ld_off), /*temp register*/esp);
431
432 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
433 // T_DOUBLE and T_LONG use two slots in the interpreter
434 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
435 // ld_off == LSW, ld_off+wordSize == MSW
436 // st_off == MSW, next_off == LSW
437 __ sd(t0, Address(sp, next_off), /*temp register*/esp);
438 #ifdef ASSERT
439 // Overwrite the unused slot with known junk
440 __ mv(t0, 0xdeadffffdeadaaaaul);
441 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
442 #endif /* ASSERT */
443 } else {
444 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
445 }
446 }
447 } else if (r_1->is_Register()) {
448 Register r = r_1->as_Register();
449 if (!r_2->is_valid()) {
450 // must be only an int (or less ) so move only 32bits to slot
451 __ sd(r, Address(sp, st_off));
452 } else {
453 // Two VMREgs|OptoRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
454 // T_DOUBLE and T_LONG use two slots in the interpreter
455 if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
456 // long/double in gpr
457 #ifdef ASSERT
458 // Overwrite the unused slot with known junk
459 __ mv(t0, 0xdeadffffdeadaaabul);
460 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
461 #endif /* ASSERT */
462 __ sd(r, Address(sp, next_off));
463 } else {
464 __ sd(r, Address(sp, st_off));
465 }
466 }
467 } else {
468 assert(r_1->is_FloatRegister(), "");
469 if (!r_2->is_valid()) {
470 // only a float use just part of the slot
471 __ fsw(r_1->as_FloatRegister(), Address(sp, st_off));
472 } else {
473 #ifdef ASSERT
474 // Overwrite the unused slot with known junk
475 __ mv(t0, 0xdeadffffdeadaaacul);
476 __ sd(t0, Address(sp, st_off), /*temp register*/esp);
477 #endif /* ASSERT */
478 __ fsd(r_1->as_FloatRegister(), Address(sp, next_off));
479 }
480 }
481 }
482
483 __ mv(esp, sp); // Interp expects args on caller's expression stack
484
485 __ ld(t0, Address(xmethod, in_bytes(Method::interpreter_entry_offset())));
486 __ jr(t0);
487 }
488
489 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
490 int total_args_passed,
491 int comp_args_on_stack,
492 const BasicType *sig_bt,
493 const VMRegPair *regs) {
494 // Note: x19_sender_sp contains the senderSP on entry. We must
495 // preserve it since we may do a i2c -> c2i transition if we lose a
496 // race where compiled code goes non-entrant while we get args
497 // ready.
498
499 // Cut-out for having no stack args.
500 int comp_words_on_stack = align_up(comp_args_on_stack * VMRegImpl::stack_slot_size, wordSize) >> LogBytesPerWord;
501 if (comp_args_on_stack != 0) {
502 __ sub(t0, sp, comp_words_on_stack * wordSize);
503 __ andi(sp, t0, -16);
504 }
505
506 // Will jump to the compiled code just as if compiled code was doing it.
507 // Pre-load the register-jump target early, to schedule it better.
508 __ ld(t1, Address(xmethod, in_bytes(Method::from_compiled_offset())));
509
510 #if INCLUDE_JVMCI
511 if (EnableJVMCI) {
512 // check if this call should be routed towards a specific entry point
513 __ ld(t0, Address(xthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
514 Label no_alternative_target;
515 __ beqz(t0, no_alternative_target);
516 __ mv(t1, t0);
517 __ sd(zr, Address(xthread, in_bytes(JavaThread::jvmci_alternate_call_target_offset())));
518 __ bind(no_alternative_target);
519 }
520 #endif // INCLUDE_JVMCI
521
522 // Now generate the shuffle code.
523 for (int i = 0; i < total_args_passed; i++) {
524 if (sig_bt[i] == T_VOID) {
525 assert(i > 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "missing half");
526 continue;
527 }
528
529 // Pick up 0, 1 or 2 words from SP+offset.
530
531 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
532 "scrambled load targets?");
533 // Load in argument order going down.
534 int ld_off = (total_args_passed - i - 1) * Interpreter::stackElementSize;
535 // Point to interpreter value (vs. tag)
536 int next_off = ld_off - Interpreter::stackElementSize;
537
538 VMReg r_1 = regs[i].first();
539 VMReg r_2 = regs[i].second();
540 if (!r_1->is_valid()) {
541 assert(!r_2->is_valid(), "");
542 continue;
543 }
544 if (r_1->is_stack()) {
545 // Convert stack slot to an SP offset (+ wordSize to account for return address )
546 int st_off = regs[i].first()->reg2stack() * VMRegImpl::stack_slot_size;
547 if (!r_2->is_valid()) {
548 __ lw(t0, Address(esp, ld_off));
549 __ sd(t0, Address(sp, st_off), /*temp register*/t2);
550 } else {
551 //
552 // We are using two optoregs. This can be either T_OBJECT,
553 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
554 // two slots but only uses one for thr T_LONG or T_DOUBLE case
555 // So we must adjust where to pick up the data to match the
556 // interpreter.
557 //
558 // Interpreter local[n] == MSW, local[n+1] == LSW however locals
559 // are accessed as negative so LSW is at LOW address
560
561 // ld_off is MSW so get LSW
562 const int offset = (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) ?
563 next_off : ld_off;
564 __ ld(t0, Address(esp, offset));
565 // st_off is LSW (i.e. reg.first())
566 __ sd(t0, Address(sp, st_off), /*temp register*/t2);
567 }
568 } else if (r_1->is_Register()) { // Register argument
569 Register r = r_1->as_Register();
570 if (r_2->is_valid()) {
571 //
572 // We are using two VMRegs. This can be either T_OBJECT,
573 // T_ADDRESS, T_LONG, or T_DOUBLE the interpreter allocates
574 // two slots but only uses one for thr T_LONG or T_DOUBLE case
575 // So we must adjust where to pick up the data to match the
576 // interpreter.
577
578 const int offset = (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) ?
579 next_off : ld_off;
580
581 // this can be a misaligned move
582 __ ld(r, Address(esp, offset));
583 } else {
584 // sign extend and use a full word?
585 __ lw(r, Address(esp, ld_off));
586 }
587 } else {
588 if (!r_2->is_valid()) {
589 __ flw(r_1->as_FloatRegister(), Address(esp, ld_off));
590 } else {
591 __ fld(r_1->as_FloatRegister(), Address(esp, next_off));
592 }
593 }
594 }
595
596 __ push_cont_fastpath(xthread); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about
597
598 // 6243940 We might end up in handle_wrong_method if
599 // the callee is deoptimized as we race thru here. If that
600 // happens we don't want to take a safepoint because the
601 // caller frame will look interpreted and arguments are now
602 // "compiled" so it is much better to make this transition
603 // invisible to the stack walking code. Unfortunately if
604 // we try and find the callee by normal means a safepoint
605 // is possible. So we stash the desired callee in the thread
606 // and the vm will find there should this case occur.
607
608 __ sd(xmethod, Address(xthread, JavaThread::callee_target_offset()));
609
610 __ jr(t1);
611 }
612
613 // ---------------------------------------------------------------
614 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
615 int total_args_passed,
616 int comp_args_on_stack,
617 const BasicType *sig_bt,
618 const VMRegPair *regs,
619 AdapterFingerPrint* fingerprint) {
620 address i2c_entry = __ pc();
621 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
622
623 address c2i_unverified_entry = __ pc();
624 Label skip_fixup;
625
626 Label ok;
627
628 const Register holder = t1;
629 const Register receiver = j_rarg0;
630 const Register tmp = t2; // A call-clobbered register not used for arg passing
631
632 // -------------------------------------------------------------------------
633 // Generate a C2I adapter. On entry we know xmethod holds the Method* during calls
634 // to the interpreter. The args start out packed in the compiled layout. They
635 // need to be unpacked into the interpreter layout. This will almost always
636 // require some stack space. We grow the current (compiled) stack, then repack
637 // the args. We finally end in a jump to the generic interpreter entry point.
638 // On exit from the interpreter, the interpreter will restore our SP (lest the
639 // compiled code, which relies solely on SP and not FP, get sick).
640
641 {
642 __ block_comment("c2i_unverified_entry {");
643 __ load_klass(t0, receiver, tmp);
644 __ ld(tmp, Address(holder, CompiledICHolder::holder_klass_offset()));
645 __ ld(xmethod, Address(holder, CompiledICHolder::holder_metadata_offset()));
646 __ beq(t0, tmp, ok);
647 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
648
649 __ bind(ok);
650 // Method might have been compiled since the call site was patched to
651 // interpreted; if that is the case treat it as a miss so we can get
652 // the call site corrected.
653 __ ld(t0, Address(xmethod, in_bytes(Method::code_offset())));
654 __ beqz(t0, skip_fixup);
655 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
656 __ block_comment("} c2i_unverified_entry");
657 }
658
659 address c2i_entry = __ pc();
660
661 // Class initialization barrier for static methods
662 address c2i_no_clinit_check_entry = nullptr;
663 if (VM_Version::supports_fast_class_init_checks()) {
664 Label L_skip_barrier;
665
666 { // Bypass the barrier for non-static methods
667 __ lwu(t0, Address(xmethod, Method::access_flags_offset()));
668 __ test_bit(t1, t0, exact_log2(JVM_ACC_STATIC));
669 __ beqz(t1, L_skip_barrier); // non-static
670 }
671
672 __ load_method_holder(t1, xmethod);
673 __ clinit_barrier(t1, t0, &L_skip_barrier);
674 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
675
676 __ bind(L_skip_barrier);
677 c2i_no_clinit_check_entry = __ pc();
678 }
679
680 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
681 bs->c2i_entry_barrier(masm);
682
683 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup);
684
685 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry, c2i_no_clinit_check_entry);
686 }
687
688 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
689 uint num_bits,
690 uint total_args_passed) {
691 Unimplemented();
692 return 0;
693 }
694
695 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
696 VMRegPair *regs,
697 VMRegPair *regs2,
698 int total_args_passed) {
699 assert(regs2 == nullptr, "not needed on riscv");
700
701 // We return the amount of VMRegImpl stack slots we need to reserve for all
702 // the arguments NOT counting out_preserve_stack_slots.
703
704 static const Register INT_ArgReg[Argument::n_int_register_parameters_c] = {
705 c_rarg0, c_rarg1, c_rarg2, c_rarg3,
706 c_rarg4, c_rarg5, c_rarg6, c_rarg7
707 };
708 static const FloatRegister FP_ArgReg[Argument::n_float_register_parameters_c] = {
709 c_farg0, c_farg1, c_farg2, c_farg3,
710 c_farg4, c_farg5, c_farg6, c_farg7
711 };
712
713 uint int_args = 0;
714 uint fp_args = 0;
715 uint stk_args = 0; // inc by 2 each time
716
717 for (int i = 0; i < total_args_passed; i++) {
718 switch (sig_bt[i]) {
719 case T_BOOLEAN: // fall through
720 case T_CHAR: // fall through
721 case T_BYTE: // fall through
722 case T_SHORT: // fall through
723 case T_INT:
724 if (int_args < Argument::n_int_register_parameters_c) {
725 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
726 } else {
727 regs[i].set1(VMRegImpl::stack2reg(stk_args));
728 stk_args += 2;
729 }
730 break;
731 case T_LONG: // fall through
732 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
733 case T_OBJECT: // fall through
734 case T_ARRAY: // fall through
735 case T_ADDRESS: // fall through
736 case T_METADATA:
737 if (int_args < Argument::n_int_register_parameters_c) {
738 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
739 } else {
740 regs[i].set2(VMRegImpl::stack2reg(stk_args));
741 stk_args += 2;
742 }
743 break;
744 case T_FLOAT:
745 if (fp_args < Argument::n_float_register_parameters_c) {
746 regs[i].set1(FP_ArgReg[fp_args++]->as_VMReg());
747 } else if (int_args < Argument::n_int_register_parameters_c) {
748 regs[i].set1(INT_ArgReg[int_args++]->as_VMReg());
749 } else {
750 regs[i].set1(VMRegImpl::stack2reg(stk_args));
751 stk_args += 2;
752 }
753 break;
754 case T_DOUBLE:
755 assert((i + 1) < total_args_passed && sig_bt[i + 1] == T_VOID, "expecting half");
756 if (fp_args < Argument::n_float_register_parameters_c) {
757 regs[i].set2(FP_ArgReg[fp_args++]->as_VMReg());
758 } else if (int_args < Argument::n_int_register_parameters_c) {
759 regs[i].set2(INT_ArgReg[int_args++]->as_VMReg());
760 } else {
761 regs[i].set2(VMRegImpl::stack2reg(stk_args));
762 stk_args += 2;
763 }
764 break;
765 case T_VOID: // Halves of longs and doubles
766 assert(i != 0 && (sig_bt[i - 1] == T_LONG || sig_bt[i - 1] == T_DOUBLE), "expecting half");
767 regs[i].set_bad();
768 break;
769 default:
770 ShouldNotReachHere();
771 }
772 }
773
774 return stk_args;
775 }
776
777 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
778 // We always ignore the frame_slots arg and just use the space just below frame pointer
779 // which by this time is free to use
780 switch (ret_type) {
781 case T_FLOAT:
782 __ fsw(f10, Address(fp, -3 * wordSize));
783 break;
784 case T_DOUBLE:
785 __ fsd(f10, Address(fp, -3 * wordSize));
786 break;
787 case T_VOID: break;
788 default: {
789 __ sd(x10, Address(fp, -3 * wordSize));
790 }
791 }
792 }
793
794 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
795 // We always ignore the frame_slots arg and just use the space just below frame pointer
796 // which by this time is free to use
797 switch (ret_type) {
798 case T_FLOAT:
799 __ flw(f10, Address(fp, -3 * wordSize));
800 break;
801 case T_DOUBLE:
802 __ fld(f10, Address(fp, -3 * wordSize));
803 break;
804 case T_VOID: break;
805 default: {
806 __ ld(x10, Address(fp, -3 * wordSize));
807 }
808 }
809 }
810
811 static void save_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 if (args[i].first()->is_FloatRegister()) {
817 __ addi(sp, sp, -2 * wordSize);
818 __ fsd(args[i].first()->as_FloatRegister(), Address(sp, 0));
819 }
820 }
821 __ push_reg(x, sp);
822 }
823
824 static void restore_args(MacroAssembler *masm, int arg_count, int first_arg, VMRegPair *args) {
825 RegSet x;
826 for ( int i = first_arg ; i < arg_count ; i++ ) {
827 if (args[i].first()->is_Register()) {
828 x = x + args[i].first()->as_Register();
829 } else {
830 ;
831 }
832 }
833 __ pop_reg(x, sp);
834 for ( int i = arg_count - 1 ; i >= first_arg ; i-- ) {
835 if (args[i].first()->is_Register()) {
836 ;
837 } else if (args[i].first()->is_FloatRegister()) {
838 __ fld(args[i].first()->as_FloatRegister(), Address(sp, 0));
839 __ add(sp, sp, 2 * wordSize);
840 }
841 }
842 }
843
844 static void verify_oop_args(MacroAssembler* masm,
845 const methodHandle& method,
846 const BasicType* sig_bt,
847 const VMRegPair* regs) {
848 const Register temp_reg = x9; // not part of any compiled calling seq
849 if (VerifyOops) {
850 for (int i = 0; i < method->size_of_parameters(); i++) {
851 if (sig_bt[i] == T_OBJECT ||
852 sig_bt[i] == T_ARRAY) {
853 VMReg r = regs[i].first();
854 assert(r->is_valid(), "bad oop arg");
855 if (r->is_stack()) {
856 __ ld(temp_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
857 __ verify_oop(temp_reg);
858 } else {
859 __ verify_oop(r->as_Register());
860 }
861 }
862 }
863 }
864 }
865
866 // on exit, sp points to the ContinuationEntry
867 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& stack_slots) {
868 assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
869 assert(in_bytes(ContinuationEntry::cont_offset()) % VMRegImpl::stack_slot_size == 0, "");
870 assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
871
872 stack_slots += (int)ContinuationEntry::size() / wordSize;
873 __ sub(sp, sp, (int)ContinuationEntry::size()); // place Continuation metadata
874
875 OopMap* map = new OopMap(((int)ContinuationEntry::size() + wordSize) / VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
876
877 __ ld(t0, Address(xthread, JavaThread::cont_entry_offset()));
878 __ sd(t0, Address(sp, ContinuationEntry::parent_offset()));
879 __ sd(sp, Address(xthread, JavaThread::cont_entry_offset()));
880
881 return map;
882 }
883
884 // on entry c_rarg1 points to the continuation
885 // sp points to ContinuationEntry
886 // c_rarg3 -- isVirtualThread
887 static void fill_continuation_entry(MacroAssembler* masm) {
888 #ifdef ASSERT
889 __ mv(t0, ContinuationEntry::cookie_value());
890 __ sw(t0, Address(sp, ContinuationEntry::cookie_offset()));
891 #endif
892
893 __ sd(c_rarg1, Address(sp, ContinuationEntry::cont_offset()));
894 __ sw(c_rarg3, Address(sp, ContinuationEntry::flags_offset()));
895 __ sd(zr, Address(sp, ContinuationEntry::chunk_offset()));
896 __ sw(zr, Address(sp, ContinuationEntry::argsize_offset()));
897 __ sw(zr, Address(sp, ContinuationEntry::pin_count_offset()));
898
899 __ ld(t0, Address(xthread, JavaThread::cont_fastpath_offset()));
900 __ sd(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
901 __ ld(t0, Address(xthread, JavaThread::held_monitor_count_offset()));
902 __ sd(t0, Address(sp, ContinuationEntry::parent_held_monitor_count_offset()));
903
904 __ sd(zr, Address(xthread, JavaThread::cont_fastpath_offset()));
905 __ sd(zr, Address(xthread, JavaThread::held_monitor_count_offset()));
906 }
907
908 // on entry, sp points to the ContinuationEntry
909 // on exit, fp points to the spilled fp + 2 * wordSize in the entry frame
910 static void continuation_enter_cleanup(MacroAssembler* masm) {
911 #ifndef PRODUCT
912 Label OK;
913 __ ld(t0, Address(xthread, JavaThread::cont_entry_offset()));
914 __ beq(sp, t0, OK);
915 __ stop("incorrect sp");
916 __ bind(OK);
917 #endif
918
919 __ ld(t0, Address(sp, ContinuationEntry::parent_cont_fastpath_offset()));
920 __ sd(t0, Address(xthread, JavaThread::cont_fastpath_offset()));
921 __ ld(t0, Address(sp, ContinuationEntry::parent_held_monitor_count_offset()));
922 __ sd(t0, Address(xthread, JavaThread::held_monitor_count_offset()));
923
924 __ ld(t0, Address(sp, ContinuationEntry::parent_offset()));
925 __ sd(t0, Address(xthread, JavaThread::cont_entry_offset()));
926 __ add(fp, sp, (int)ContinuationEntry::size() + 2 * wordSize /* 2 extra words to match up with leave() */);
927 }
928
929 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
930 // On entry: c_rarg1 -- the continuation object
931 // c_rarg2 -- isContinue
932 // c_rarg3 -- isVirtualThread
933 static void gen_continuation_enter(MacroAssembler* masm,
934 const methodHandle& method,
935 const BasicType* sig_bt,
936 const VMRegPair* regs,
937 int& exception_offset,
938 OopMapSet*oop_maps,
939 int& frame_complete,
940 int& stack_slots,
941 int& interpreted_entry_offset,
942 int& compiled_entry_offset) {
943 // verify_oop_args(masm, method, sig_bt, regs);
944 Address resolve(SharedRuntime::get_resolve_static_call_stub(), relocInfo::static_call_type);
945
946 address start = __ pc();
947
948 Label call_thaw, exit;
949
950 // i2i entry used at interp_only_mode only
951 interpreted_entry_offset = __ pc() - start;
952 {
953 #ifdef ASSERT
954 Label is_interp_only;
955 __ lw(t0, Address(xthread, JavaThread::interp_only_mode_offset()));
956 __ bnez(t0, is_interp_only);
957 __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
958 __ bind(is_interp_only);
959 #endif
960
961 // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
962 __ ld(c_rarg1, Address(esp, Interpreter::stackElementSize * 2));
963 __ ld(c_rarg2, Address(esp, Interpreter::stackElementSize * 1));
964 __ ld(c_rarg3, Address(esp, Interpreter::stackElementSize * 0));
965 __ push_cont_fastpath(xthread);
966
967 __ enter();
968 stack_slots = 2; // will be adjusted in setup
969 OopMap* map = continuation_enter_setup(masm, stack_slots);
970 // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
971 // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
972
973 fill_continuation_entry(masm);
974
975 __ bnez(c_rarg2, call_thaw);
976
977 // Make sure the call is patchable
978 __ align(NativeInstruction::instruction_size);
979
980 const address tr_call = __ trampoline_call(resolve);
981 if (tr_call == nullptr) {
982 fatal("CodeCache is full at gen_continuation_enter");
983 }
984
985 oop_maps->add_gc_map(__ pc() - start, map);
986 __ post_call_nop();
987
988 __ j(exit);
989
990 CodeBuffer* cbuf = masm->code_section()->outer();
991 address stub = CompiledStaticCall::emit_to_interp_stub(*cbuf, tr_call);
992 if (stub == nullptr) {
993 fatal("CodeCache is full at gen_continuation_enter");
994 }
995 }
996
997 // compiled entry
998 __ align(CodeEntryAlignment);
999 compiled_entry_offset = __ pc() - start;
1000
1001 __ enter();
1002 stack_slots = 2; // will be adjusted in setup
1003 OopMap* map = continuation_enter_setup(masm, stack_slots);
1004 frame_complete = __ pc() - start;
1005
1006 fill_continuation_entry(masm);
1007
1008 __ bnez(c_rarg2, call_thaw);
1009
1010 // Make sure the call is patchable
1011 __ align(NativeInstruction::instruction_size);
1012
1013 const address tr_call = __ trampoline_call(resolve);
1014 if (tr_call == nullptr) {
1015 fatal("CodeCache is full at gen_continuation_enter");
1016 }
1017
1018 oop_maps->add_gc_map(__ pc() - start, map);
1019 __ post_call_nop();
1020
1021 __ j(exit);
1022
1023 __ bind(call_thaw);
1024
1025 __ rt_call(CAST_FROM_FN_PTR(address, StubRoutines::cont_thaw()));
1026 oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1027 ContinuationEntry::_return_pc_offset = __ pc() - start;
1028 __ post_call_nop();
1029
1030 __ bind(exit);
1031 continuation_enter_cleanup(masm);
1032 __ leave();
1033 __ ret();
1034
1035 // exception handling
1036 exception_offset = __ pc() - start;
1037 {
1038 __ mv(x9, x10); // save return value contaning the exception oop in callee-saved x9
1039
1040 continuation_enter_cleanup(masm);
1041
1042 __ ld(c_rarg1, Address(fp, -1 * wordSize)); // return address
1043 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), xthread, c_rarg1);
1044
1045 // see OptoRuntime::generate_exception_blob: x10 -- exception oop, x13 -- exception pc
1046
1047 __ mv(x11, x10); // the exception handler
1048 __ mv(x10, x9); // restore return value contaning the exception oop
1049 __ verify_oop(x10);
1050
1051 __ leave();
1052 __ mv(x13, ra);
1053 __ jr(x11); // the exception handler
1054 }
1055
1056 CodeBuffer* cbuf = masm->code_section()->outer();
1057 address stub = CompiledStaticCall::emit_to_interp_stub(*cbuf, tr_call);
1058 if (stub == nullptr) {
1059 fatal("CodeCache is full at gen_continuation_enter");
1060 }
1061 }
1062
1063 static void gen_continuation_yield(MacroAssembler* masm,
1064 const methodHandle& method,
1065 const BasicType* sig_bt,
1066 const VMRegPair* regs,
1067 OopMapSet* oop_maps,
1068 int& frame_complete,
1069 int& stack_slots,
1070 int& compiled_entry_offset) {
1071 enum layout {
1072 fp_off,
1073 fp_off2,
1074 return_off,
1075 return_off2,
1076 framesize // inclusive of return address
1077 };
1078 // assert(is_even(framesize/2), "sp not 16-byte aligned");
1079
1080 stack_slots = framesize / VMRegImpl::slots_per_word;
1081 assert(stack_slots == 2, "recheck layout");
1082
1083 address start = __ pc();
1084
1085 compiled_entry_offset = __ pc() - start;
1086 __ enter();
1087
1088 __ mv(c_rarg1, sp);
1089
1090 frame_complete = __ pc() - start;
1091 address the_pc = __ pc();
1092
1093 __ 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
1094
1095 __ mv(c_rarg0, xthread);
1096 __ set_last_Java_frame(sp, fp, the_pc, t0);
1097 __ call_VM_leaf(Continuation::freeze_entry(), 2);
1098 __ reset_last_Java_frame(true);
1099
1100 Label pinned;
1101
1102 __ bnez(x10, pinned);
1103
1104 // We've succeeded, set sp to the ContinuationEntry
1105 __ ld(sp, Address(xthread, JavaThread::cont_entry_offset()));
1106 continuation_enter_cleanup(masm);
1107
1108 __ bind(pinned); // pinned -- return to caller
1109
1110 // handle pending exception thrown by freeze
1111 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1112 Label ok;
1113 __ beqz(t0, ok);
1114 __ leave();
1115 __ la(t0, RuntimeAddress(StubRoutines::forward_exception_entry()));
1116 __ jr(t0);
1117 __ bind(ok);
1118
1119 __ leave();
1120 __ ret();
1121
1122 OopMap* map = new OopMap(framesize, 1);
1123 oop_maps->add_gc_map(the_pc - start, map);
1124 }
1125
1126 static void gen_special_dispatch(MacroAssembler* masm,
1127 const methodHandle& method,
1128 const BasicType* sig_bt,
1129 const VMRegPair* regs) {
1130 verify_oop_args(masm, method, sig_bt, regs);
1131 vmIntrinsics::ID iid = method->intrinsic_id();
1132
1133 // Now write the args into the outgoing interpreter space
1134 bool has_receiver = false;
1135 Register receiver_reg = noreg;
1136 int member_arg_pos = -1;
1137 Register member_reg = noreg;
1138 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1139 if (ref_kind != 0) {
1140 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1141 member_reg = x9; // known to be free at this point
1142 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1143 } else if (iid == vmIntrinsics::_invokeBasic) {
1144 has_receiver = true;
1145 } else if (iid == vmIntrinsics::_linkToNative) {
1146 member_arg_pos = method->size_of_parameters() - 1; // trailing NativeEntryPoint argument
1147 member_reg = x9; // known to be free at this point
1148 } else {
1149 fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1150 }
1151
1152 if (member_reg != noreg) {
1153 // Load the member_arg into register, if necessary.
1154 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1155 VMReg r = regs[member_arg_pos].first();
1156 if (r->is_stack()) {
1157 __ ld(member_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1158 } else {
1159 // no data motion is needed
1160 member_reg = r->as_Register();
1161 }
1162 }
1163
1164 if (has_receiver) {
1165 // Make sure the receiver is loaded into a register.
1166 assert(method->size_of_parameters() > 0, "oob");
1167 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1168 VMReg r = regs[0].first();
1169 assert(r->is_valid(), "bad receiver arg");
1170 if (r->is_stack()) {
1171 // Porting note: This assumes that compiled calling conventions always
1172 // pass the receiver oop in a register. If this is not true on some
1173 // platform, pick a temp and load the receiver from stack.
1174 fatal("receiver always in a register");
1175 receiver_reg = x12; // known to be free at this point
1176 __ ld(receiver_reg, Address(sp, r->reg2stack() * VMRegImpl::stack_slot_size));
1177 } else {
1178 // no data motion is needed
1179 receiver_reg = r->as_Register();
1180 }
1181 }
1182
1183 // Figure out which address we are really jumping to:
1184 MethodHandles::generate_method_handle_dispatch(masm, iid,
1185 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1186 }
1187
1188 // ---------------------------------------------------------------------------
1189 // Generate a native wrapper for a given method. The method takes arguments
1190 // in the Java compiled code convention, marshals them to the native
1191 // convention (handlizes oops, etc), transitions to native, makes the call,
1192 // returns to java state (possibly blocking), unhandlizes any result and
1193 // returns.
1194 //
1195 // Critical native functions are a shorthand for the use of
1196 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1197 // functions. The wrapper is expected to unpack the arguments before
1198 // passing them to the callee and perform checks before and after the
1199 // native call to ensure that they GCLocker
1200 // lock_critical/unlock_critical semantics are followed. Some other
1201 // parts of JNI setup are skipped like the tear down of the JNI handle
1202 // block and the check for pending exceptions it's impossible for them
1203 // to be thrown.
1204 //
1205 // They are roughly structured like this:
1206 // if (GCLocker::needs_gc()) SharedRuntime::block_for_jni_critical()
1207 // tranistion to thread_in_native
1208 // unpack array arguments and call native entry point
1209 // check for safepoint in progress
1210 // check if any thread suspend flags are set
1211 // call into JVM and possible unlock the JNI critical
1212 // if a GC was suppressed while in the critical native.
1213 // transition back to thread_in_Java
1214 // return to caller
1215 //
1216 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1217 const methodHandle& method,
1218 int compile_id,
1219 BasicType* in_sig_bt,
1220 VMRegPair* in_regs,
1221 BasicType ret_type) {
1222 if (method->is_continuation_native_intrinsic()) {
1223 int exception_offset = -1;
1224 OopMapSet* oop_maps = new OopMapSet();
1225 int frame_complete = -1;
1226 int stack_slots = -1;
1227 int interpreted_entry_offset = -1;
1228 int vep_offset = -1;
1229 if (method->is_continuation_enter_intrinsic()) {
1230 gen_continuation_enter(masm,
1231 method,
1232 in_sig_bt,
1233 in_regs,
1234 exception_offset,
1235 oop_maps,
1236 frame_complete,
1237 stack_slots,
1238 interpreted_entry_offset,
1239 vep_offset);
1240 } else if (method->is_continuation_yield_intrinsic()) {
1241 gen_continuation_yield(masm,
1242 method,
1243 in_sig_bt,
1244 in_regs,
1245 oop_maps,
1246 frame_complete,
1247 stack_slots,
1248 vep_offset);
1249 } else {
1250 guarantee(false, "Unknown Continuation native intrinsic");
1251 }
1252
1253 #ifdef ASSERT
1254 if (method->is_continuation_enter_intrinsic()) {
1255 assert(interpreted_entry_offset != -1, "Must be set");
1256 assert(exception_offset != -1, "Must be set");
1257 } else {
1258 assert(interpreted_entry_offset == -1, "Must be unset");
1259 assert(exception_offset == -1, "Must be unset");
1260 }
1261 assert(frame_complete != -1, "Must be set");
1262 assert(stack_slots != -1, "Must be set");
1263 assert(vep_offset != -1, "Must be set");
1264 #endif
1265
1266 __ flush();
1267 nmethod* nm = nmethod::new_native_nmethod(method,
1268 compile_id,
1269 masm->code(),
1270 vep_offset,
1271 frame_complete,
1272 stack_slots,
1273 in_ByteSize(-1),
1274 in_ByteSize(-1),
1275 oop_maps,
1276 exception_offset);
1277 if (method->is_continuation_enter_intrinsic()) {
1278 ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
1279 } else if (method->is_continuation_yield_intrinsic()) {
1280 _cont_doYield_stub = nm;
1281 } else {
1282 guarantee(false, "Unknown Continuation native intrinsic");
1283 }
1284 return nm;
1285 }
1286
1287 if (method->is_method_handle_intrinsic()) {
1288 vmIntrinsics::ID iid = method->intrinsic_id();
1289 intptr_t start = (intptr_t)__ pc();
1290 int vep_offset = ((intptr_t)__ pc()) - start;
1291
1292 // First instruction must be a nop as it may need to be patched on deoptimisation
1293 {
1294 Assembler::IncompressibleRegion ir(masm); // keep the nop as 4 bytes for patching.
1295 MacroAssembler::assert_alignment(__ pc());
1296 __ nop(); // 4 bytes
1297 }
1298 gen_special_dispatch(masm,
1299 method,
1300 in_sig_bt,
1301 in_regs);
1302 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
1303 __ flush();
1304 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
1305 return nmethod::new_native_nmethod(method,
1306 compile_id,
1307 masm->code(),
1308 vep_offset,
1309 frame_complete,
1310 stack_slots / VMRegImpl::slots_per_word,
1311 in_ByteSize(-1),
1312 in_ByteSize(-1),
1313 (OopMapSet*)nullptr);
1314 }
1315 address native_func = method->native_function();
1316 assert(native_func != nullptr, "must have function");
1317
1318 // An OopMap for lock (and class if static)
1319 OopMapSet *oop_maps = new OopMapSet();
1320 assert_cond(oop_maps != nullptr);
1321 intptr_t start = (intptr_t)__ pc();
1322
1323 // We have received a description of where all the java arg are located
1324 // on entry to the wrapper. We need to convert these args to where
1325 // the jni function will expect them. To figure out where they go
1326 // we convert the java signature to a C signature by inserting
1327 // the hidden arguments as arg[0] and possibly arg[1] (static method)
1328
1329 const int total_in_args = method->size_of_parameters();
1330 int total_c_args = total_in_args + (method->is_static() ? 2 : 1);
1331
1332 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1333 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1334 BasicType* in_elem_bt = nullptr;
1335
1336 int argc = 0;
1337 out_sig_bt[argc++] = T_ADDRESS;
1338 if (method->is_static()) {
1339 out_sig_bt[argc++] = T_OBJECT;
1340 }
1341
1342 for (int i = 0; i < total_in_args ; i++) {
1343 out_sig_bt[argc++] = in_sig_bt[i];
1344 }
1345
1346 // Now figure out where the args must be stored and how much stack space
1347 // they require.
1348 int out_arg_slots = c_calling_convention(out_sig_bt, out_regs, nullptr, total_c_args);
1349
1350 // Compute framesize for the wrapper. We need to handlize all oops in
1351 // incoming registers
1352
1353 // Calculate the total number of stack slots we will need.
1354
1355 // First count the abi requirement plus all of the outgoing args
1356 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1357
1358 // Now the space for the inbound oop handle area
1359 int total_save_slots = 8 * VMRegImpl::slots_per_word; // 8 arguments passed in registers
1360
1361 int oop_handle_offset = stack_slots;
1362 stack_slots += total_save_slots;
1363
1364 // Now any space we need for handlizing a klass if static method
1365
1366 int klass_slot_offset = 0;
1367 int klass_offset = -1;
1368 int lock_slot_offset = 0;
1369 bool is_static = false;
1370
1371 if (method->is_static()) {
1372 klass_slot_offset = stack_slots;
1373 stack_slots += VMRegImpl::slots_per_word;
1374 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1375 is_static = true;
1376 }
1377
1378 // Plus a lock if needed
1379
1380 if (method->is_synchronized()) {
1381 lock_slot_offset = stack_slots;
1382 stack_slots += VMRegImpl::slots_per_word;
1383 }
1384
1385 // Now a place (+2) to save return values or temp during shuffling
1386 // + 4 for return address (which we own) and saved fp
1387 stack_slots += 6;
1388
1389 // Ok The space we have allocated will look like:
1390 //
1391 //
1392 // FP-> | |
1393 // | 2 slots (ra) |
1394 // | 2 slots (fp) |
1395 // |---------------------|
1396 // | 2 slots for moves |
1397 // |---------------------|
1398 // | lock box (if sync) |
1399 // |---------------------| <- lock_slot_offset
1400 // | klass (if static) |
1401 // |---------------------| <- klass_slot_offset
1402 // | oopHandle area |
1403 // |---------------------| <- oop_handle_offset (8 java arg registers)
1404 // | outbound memory |
1405 // | based arguments |
1406 // | |
1407 // |---------------------|
1408 // | |
1409 // SP-> | out_preserved_slots |
1410 //
1411 //
1412
1413
1414 // Now compute actual number of stack words we need rounding to make
1415 // stack properly aligned.
1416 stack_slots = align_up(stack_slots, StackAlignmentInSlots);
1417
1418 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1419
1420 // First thing make an ic check to see if we should even be here
1421
1422 // We are free to use all registers as temps without saving them and
1423 // restoring them except fp. fp is the only callee save register
1424 // as far as the interpreter and the compiler(s) are concerned.
1425
1426
1427 const Register ic_reg = t1;
1428 const Register receiver = j_rarg0;
1429
1430 Label hit;
1431 Label exception_pending;
1432
1433 __ verify_oop(receiver);
1434 assert_different_registers(ic_reg, receiver, t0, t2);
1435 __ cmp_klass(receiver, ic_reg, t0, t2 /* call-clobbered t2 as a tmp */, hit);
1436
1437 __ far_jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
1438
1439 // Verified entry point must be aligned
1440 __ align(8);
1441
1442 __ bind(hit);
1443
1444 int vep_offset = ((intptr_t)__ pc()) - start;
1445
1446 // If we have to make this method not-entrant we'll overwrite its
1447 // first instruction with a jump.
1448 {
1449 Assembler::IncompressibleRegion ir(masm); // keep the nop as 4 bytes for patching.
1450 MacroAssembler::assert_alignment(__ pc());
1451 __ nop(); // 4 bytes
1452 }
1453
1454 if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) {
1455 Label L_skip_barrier;
1456 __ mov_metadata(t1, method->method_holder()); // InstanceKlass*
1457 __ clinit_barrier(t1, t0, &L_skip_barrier);
1458 __ far_jump(RuntimeAddress(SharedRuntime::get_handle_wrong_method_stub()));
1459
1460 __ bind(L_skip_barrier);
1461 }
1462
1463 // Generate stack overflow check
1464 __ bang_stack_with_offset(checked_cast<int>(StackOverflow::stack_shadow_zone_size()));
1465
1466 // Generate a new frame for the wrapper.
1467 __ enter();
1468 // -2 because return address is already present and so is saved fp
1469 __ sub(sp, sp, stack_size - 2 * wordSize);
1470
1471 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1472 assert_cond(bs != nullptr);
1473 bs->nmethod_entry_barrier(masm, nullptr /* slow_path */, nullptr /* continuation */, nullptr /* guard */);
1474
1475 // Frame is now completed as far as size and linkage.
1476 int frame_complete = ((intptr_t)__ pc()) - start;
1477
1478 // We use x18 as the oop handle for the receiver/klass
1479 // It is callee save so it survives the call to native
1480
1481 const Register oop_handle_reg = x18;
1482
1483 //
1484 // We immediately shuffle the arguments so that any vm call we have to
1485 // make from here on out (sync slow path, jvmti, etc.) we will have
1486 // captured the oops from our caller and have a valid oopMap for
1487 // them.
1488
1489 // -----------------
1490 // The Grand Shuffle
1491
1492 // The Java calling convention is either equal (linux) or denser (win64) than the
1493 // c calling convention. However the because of the jni_env argument the c calling
1494 // convention always has at least one more (and two for static) arguments than Java.
1495 // Therefore if we move the args from java -> c backwards then we will never have
1496 // a register->register conflict and we don't have to build a dependency graph
1497 // and figure out how to break any cycles.
1498 //
1499
1500 // Record esp-based slot for receiver on stack for non-static methods
1501 int receiver_offset = -1;
1502
1503 // This is a trick. We double the stack slots so we can claim
1504 // the oops in the caller's frame. Since we are sure to have
1505 // more args than the caller doubling is enough to make
1506 // sure we can capture all the incoming oop args from the
1507 // caller.
1508 //
1509 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1510 assert_cond(map != nullptr);
1511
1512 int float_args = 0;
1513 int int_args = 0;
1514
1515 #ifdef ASSERT
1516 bool reg_destroyed[Register::number_of_registers];
1517 bool freg_destroyed[FloatRegister::number_of_registers];
1518 for ( int r = 0 ; r < Register::number_of_registers ; r++ ) {
1519 reg_destroyed[r] = false;
1520 }
1521 for ( int f = 0 ; f < FloatRegister::number_of_registers ; f++ ) {
1522 freg_destroyed[f] = false;
1523 }
1524
1525 #endif /* ASSERT */
1526
1527 // For JNI natives the incoming and outgoing registers are offset upwards.
1528 GrowableArray<int> arg_order(2 * total_in_args);
1529 VMRegPair tmp_vmreg;
1530 tmp_vmreg.set2(x9->as_VMReg());
1531
1532 for (int i = total_in_args - 1, c_arg = total_c_args - 1; i >= 0; i--, c_arg--) {
1533 arg_order.push(i);
1534 arg_order.push(c_arg);
1535 }
1536
1537 int temploc = -1;
1538 for (int ai = 0; ai < arg_order.length(); ai += 2) {
1539 int i = arg_order.at(ai);
1540 int c_arg = arg_order.at(ai + 1);
1541 __ block_comment(err_msg("mv %d -> %d", i, c_arg));
1542 assert(c_arg != -1 && i != -1, "wrong order");
1543 #ifdef ASSERT
1544 if (in_regs[i].first()->is_Register()) {
1545 assert(!reg_destroyed[in_regs[i].first()->as_Register()->encoding()], "destroyed reg!");
1546 } else if (in_regs[i].first()->is_FloatRegister()) {
1547 assert(!freg_destroyed[in_regs[i].first()->as_FloatRegister()->encoding()], "destroyed reg!");
1548 }
1549 if (out_regs[c_arg].first()->is_Register()) {
1550 reg_destroyed[out_regs[c_arg].first()->as_Register()->encoding()] = true;
1551 } else if (out_regs[c_arg].first()->is_FloatRegister()) {
1552 freg_destroyed[out_regs[c_arg].first()->as_FloatRegister()->encoding()] = true;
1553 }
1554 #endif /* ASSERT */
1555 switch (in_sig_bt[i]) {
1556 case T_ARRAY:
1557 case T_OBJECT:
1558 __ object_move(map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
1559 ((i == 0) && (!is_static)),
1560 &receiver_offset);
1561 int_args++;
1562 break;
1563 case T_VOID:
1564 break;
1565
1566 case T_FLOAT:
1567 __ float_move(in_regs[i], out_regs[c_arg]);
1568 float_args++;
1569 break;
1570
1571 case T_DOUBLE:
1572 assert( i + 1 < total_in_args &&
1573 in_sig_bt[i + 1] == T_VOID &&
1574 out_sig_bt[c_arg + 1] == T_VOID, "bad arg list");
1575 __ double_move(in_regs[i], out_regs[c_arg]);
1576 float_args++;
1577 break;
1578
1579 case T_LONG :
1580 __ long_move(in_regs[i], out_regs[c_arg]);
1581 int_args++;
1582 break;
1583
1584 case T_ADDRESS:
1585 assert(false, "found T_ADDRESS in java args");
1586 break;
1587
1588 default:
1589 __ move32_64(in_regs[i], out_regs[c_arg]);
1590 int_args++;
1591 }
1592 }
1593
1594 // point c_arg at the first arg that is already loaded in case we
1595 // need to spill before we call out
1596 int c_arg = total_c_args - total_in_args;
1597
1598 // Pre-load a static method's oop into c_rarg1.
1599 if (method->is_static()) {
1600
1601 // load oop into a register
1602 __ movoop(c_rarg1,
1603 JNIHandles::make_local(method->method_holder()->java_mirror()));
1604
1605 // Now handlize the static class mirror it's known not-null.
1606 __ sd(c_rarg1, Address(sp, klass_offset));
1607 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
1608
1609 // Now get the handle
1610 __ la(c_rarg1, Address(sp, klass_offset));
1611 // and protect the arg if we must spill
1612 c_arg--;
1613 }
1614
1615 // Change state to native (we save the return address in the thread, since it might not
1616 // be pushed on the stack when we do a stack traversal).
1617 // We use the same pc/oopMap repeatedly when we call out
1618
1619 Label native_return;
1620 __ set_last_Java_frame(sp, noreg, native_return, t0);
1621
1622 Label dtrace_method_entry, dtrace_method_entry_done;
1623 {
1624 ExternalAddress target((address)&DTraceMethodProbes);
1625 __ relocate(target.rspec(), [&] {
1626 int32_t offset;
1627 __ la_patchable(t0, target, offset);
1628 __ lbu(t0, Address(t0, offset));
1629 });
1630 __ bnez(t0, dtrace_method_entry);
1631 __ bind(dtrace_method_entry_done);
1632 }
1633
1634 // RedefineClasses() tracing support for obsolete method entry
1635 if (log_is_enabled(Trace, redefine, class, obsolete)) {
1636 // protect the args we've loaded
1637 save_args(masm, total_c_args, c_arg, out_regs);
1638 __ mov_metadata(c_rarg1, method());
1639 __ call_VM_leaf(
1640 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
1641 xthread, c_rarg1);
1642 restore_args(masm, total_c_args, c_arg, out_regs);
1643 }
1644
1645 // Lock a synchronized method
1646
1647 // Register definitions used by locking and unlocking
1648
1649 const Register swap_reg = x10;
1650 const Register obj_reg = x9; // Will contain the oop
1651 const Register lock_reg = x30; // Address of compiler lock object (BasicLock)
1652 const Register old_hdr = x30; // value of old header at unlock time
1653 const Register tmp = ra;
1654
1655 Label slow_path_lock;
1656 Label lock_done;
1657
1658 if (method->is_synchronized()) {
1659 Label count;
1660
1661 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes();
1662
1663 // Get the handle (the 2nd argument)
1664 __ mv(oop_handle_reg, c_rarg1);
1665
1666 // Get address of the box
1667
1668 __ la(lock_reg, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1669
1670 // Load the oop from the handle
1671 __ ld(obj_reg, Address(oop_handle_reg, 0));
1672
1673 if (LockingMode == LM_MONITOR) {
1674 __ j(slow_path_lock);
1675 } else if (LockingMode == LM_LEGACY) {
1676 // Load (object->mark() | 1) into swap_reg % x10
1677 __ ld(t0, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
1678 __ ori(swap_reg, t0, 1);
1679
1680 // Save (object->mark() | 1) into BasicLock's displaced header
1681 __ sd(swap_reg, Address(lock_reg, mark_word_offset));
1682
1683 // src -> dest if dest == x10 else x10 <- dest
1684 __ cmpxchg_obj_header(x10, lock_reg, obj_reg, t0, count, /*fallthrough*/nullptr);
1685
1686 // Test if the oopMark is an obvious stack pointer, i.e.,
1687 // 1) (mark & 3) == 0, and
1688 // 2) sp <= mark < mark + os::pagesize()
1689 // These 3 tests can be done by evaluating the following
1690 // expression: ((mark - sp) & (3 - os::vm_page_size())),
1691 // assuming both stack pointer and pagesize have their
1692 // least significant 2 bits clear.
1693 // NOTE: the oopMark is in swap_reg % 10 as the result of cmpxchg
1694
1695 __ sub(swap_reg, swap_reg, sp);
1696 __ andi(swap_reg, swap_reg, 3 - (int)os::vm_page_size());
1697
1698 // Save the test result, for recursive case, the result is zero
1699 __ sd(swap_reg, Address(lock_reg, mark_word_offset));
1700 __ bnez(swap_reg, slow_path_lock);
1701 } else {
1702 assert(LockingMode == LM_LIGHTWEIGHT, "");
1703 __ ld(swap_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
1704 __ fast_lock(obj_reg, swap_reg, tmp, t0, slow_path_lock);
1705 }
1706
1707 __ bind(count);
1708 __ increment(Address(xthread, JavaThread::held_monitor_count_offset()));
1709
1710 // Slow path will re-enter here
1711 __ bind(lock_done);
1712 }
1713
1714
1715 // Finally just about ready to make the JNI call
1716
1717 // get JNIEnv* which is first argument to native
1718 __ la(c_rarg0, Address(xthread, in_bytes(JavaThread::jni_environment_offset())));
1719
1720 // Now set thread in native
1721 __ la(t1, Address(xthread, JavaThread::thread_state_offset()));
1722 __ mv(t0, _thread_in_native);
1723 __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore);
1724 __ sw(t0, Address(t1));
1725
1726 __ rt_call(native_func);
1727
1728 __ bind(native_return);
1729
1730 intptr_t return_pc = (intptr_t) __ pc();
1731 oop_maps->add_gc_map(return_pc - start, map);
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 Label after_transition;
1740
1741 // Switch thread to "native transition" state before reading the synchronization state.
1742 // This additional state is necessary because reading and testing the synchronization
1743 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1744 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1745 // VM thread changes sync state to synchronizing and suspends threads for GC.
1746 // Thread A is resumed to finish this native method, but doesn't block here since it
1747 // didn't see any synchronization is progress, and escapes.
1748 __ mv(t0, _thread_in_native_trans);
1749
1750 __ sw(t0, Address(xthread, JavaThread::thread_state_offset()));
1751
1752 // Force this write out before the read below
1753 if (!UseSystemMemoryBarrier) {
1754 __ membar(MacroAssembler::AnyAny);
1755 }
1756
1757 // check for safepoint operation in progress and/or pending suspend requests
1758 {
1759 // We need an acquire here to ensure that any subsequent load of the
1760 // global SafepointSynchronize::_state flag is ordered after this load
1761 // of the thread-local polling word. We don't want this poll to
1762 // return false (i.e. not safepointing) and a later poll of the global
1763 // SafepointSynchronize::_state spuriously to return true.
1764 // This is to avoid a race when we're in a native->Java transition
1765 // racing the code which wakes up from a safepoint.
1766
1767 __ safepoint_poll(safepoint_in_progress, true /* at_return */, true /* acquire */, false /* in_nmethod */);
1768 __ lwu(t0, Address(xthread, JavaThread::suspend_flags_offset()));
1769 __ bnez(t0, safepoint_in_progress);
1770 __ bind(safepoint_in_progress_done);
1771 }
1772
1773 // change thread state
1774 __ la(t1, Address(xthread, JavaThread::thread_state_offset()));
1775 __ mv(t0, _thread_in_Java);
1776 __ membar(MacroAssembler::LoadStore | MacroAssembler::StoreStore);
1777 __ sw(t0, Address(t1));
1778 __ bind(after_transition);
1779
1780 Label reguard;
1781 Label reguard_done;
1782 __ lbu(t0, Address(xthread, JavaThread::stack_guard_state_offset()));
1783 __ mv(t1, StackOverflow::stack_guard_yellow_reserved_disabled);
1784 __ beq(t0, t1, reguard);
1785 __ bind(reguard_done);
1786
1787 // native result if any is live
1788
1789 // Unlock
1790 Label unlock_done;
1791 Label slow_path_unlock;
1792 if (method->is_synchronized()) {
1793
1794 // Get locked oop from the handle we passed to jni
1795 __ ld(obj_reg, Address(oop_handle_reg, 0));
1796
1797 Label done, not_recursive;
1798
1799 if (LockingMode == LM_LEGACY) {
1800 // Simple recursive lock?
1801 __ ld(t0, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1802 __ bnez(t0, not_recursive);
1803 __ decrement(Address(xthread, JavaThread::held_monitor_count_offset()));
1804 __ j(done);
1805 }
1806
1807 __ bind(not_recursive);
1808
1809 // Must save x10 if if it is live now because cmpxchg must use it
1810 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1811 save_native_result(masm, ret_type, stack_slots);
1812 }
1813
1814 if (LockingMode == LM_MONITOR) {
1815 __ j(slow_path_unlock);
1816 } else if (LockingMode == LM_LEGACY) {
1817 // get address of the stack lock
1818 __ la(x10, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1819 // get old displaced header
1820 __ ld(old_hdr, Address(x10, 0));
1821
1822 // Atomic swap old header if oop still contains the stack lock
1823 Label count;
1824 __ cmpxchg_obj_header(x10, old_hdr, obj_reg, t0, count, &slow_path_unlock);
1825 __ bind(count);
1826 __ decrement(Address(xthread, JavaThread::held_monitor_count_offset()));
1827 } else {
1828 assert(LockingMode == LM_LIGHTWEIGHT, "");
1829 __ ld(old_hdr, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
1830 __ test_bit(t0, old_hdr, exact_log2(markWord::monitor_value));
1831 __ bnez(t0, slow_path_unlock);
1832 __ fast_unlock(obj_reg, old_hdr, swap_reg, t0, slow_path_unlock);
1833 __ decrement(Address(xthread, JavaThread::held_monitor_count_offset()));
1834 }
1835
1836 // slow path re-enters here
1837 __ bind(unlock_done);
1838 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1839 restore_native_result(masm, ret_type, stack_slots);
1840 }
1841
1842 __ bind(done);
1843 }
1844
1845 Label dtrace_method_exit, dtrace_method_exit_done;
1846 {
1847 ExternalAddress target((address)&DTraceMethodProbes);
1848 __ relocate(target.rspec(), [&] {
1849 int32_t offset;
1850 __ la_patchable(t0, target, offset);
1851 __ lbu(t0, Address(t0, offset));
1852 });
1853 __ bnez(t0, dtrace_method_exit);
1854 __ bind(dtrace_method_exit_done);
1855 }
1856
1857 __ reset_last_Java_frame(false);
1858
1859 // Unbox oop result, e.g. JNIHandles::resolve result.
1860 if (is_reference_type(ret_type)) {
1861 __ resolve_jobject(x10, x11, x12);
1862 }
1863
1864 if (CheckJNICalls) {
1865 // clear_pending_jni_exception_check
1866 __ sd(zr, Address(xthread, JavaThread::pending_jni_exception_check_fn_offset()));
1867 }
1868
1869 // reset handle block
1870 __ ld(x12, Address(xthread, JavaThread::active_handles_offset()));
1871 __ sd(zr, Address(x12, JNIHandleBlock::top_offset()));
1872
1873 __ leave();
1874
1875 // Any exception pending?
1876 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1877 __ bnez(t0, exception_pending);
1878
1879 // We're done
1880 __ ret();
1881
1882 // Unexpected paths are out of line and go here
1883
1884 // forward the exception
1885 __ bind(exception_pending);
1886
1887 // and forward the exception
1888 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
1889
1890 // Slow path locking & unlocking
1891 if (method->is_synchronized()) {
1892
1893 __ block_comment("Slow path lock {");
1894 __ bind(slow_path_lock);
1895
1896 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM
1897 // args are (oop obj, BasicLock* lock, JavaThread* thread)
1898
1899 // protect the args we've loaded
1900 save_args(masm, total_c_args, c_arg, out_regs);
1901
1902 __ mv(c_rarg0, obj_reg);
1903 __ mv(c_rarg1, lock_reg);
1904 __ mv(c_rarg2, xthread);
1905
1906 // Not a leaf but we have last_Java_frame setup as we want
1907 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), 3);
1908 restore_args(masm, total_c_args, c_arg, out_regs);
1909
1910 #ifdef ASSERT
1911 { Label L;
1912 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1913 __ beqz(t0, L);
1914 __ stop("no pending exception allowed on exit from monitorenter");
1915 __ bind(L);
1916 }
1917 #endif
1918 __ j(lock_done);
1919
1920 __ block_comment("} Slow path lock");
1921
1922 __ block_comment("Slow path unlock {");
1923 __ bind(slow_path_unlock);
1924
1925 if (ret_type == T_FLOAT || ret_type == T_DOUBLE) {
1926 save_native_result(masm, ret_type, stack_slots);
1927 }
1928
1929 __ mv(c_rarg2, xthread);
1930 __ la(c_rarg1, Address(sp, lock_slot_offset * VMRegImpl::stack_slot_size));
1931 __ mv(c_rarg0, obj_reg);
1932
1933 // Save pending exception around call to VM (which contains an EXCEPTION_MARK)
1934 // NOTE that obj_reg == x9 currently
1935 __ ld(x9, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1936 __ sd(zr, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1937
1938 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C));
1939
1940 #ifdef ASSERT
1941 {
1942 Label L;
1943 __ ld(t0, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1944 __ beqz(t0, L);
1945 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C");
1946 __ bind(L);
1947 }
1948 #endif /* ASSERT */
1949
1950 __ sd(x9, Address(xthread, in_bytes(Thread::pending_exception_offset())));
1951
1952 if (ret_type == T_FLOAT || ret_type == T_DOUBLE) {
1953 restore_native_result(masm, ret_type, stack_slots);
1954 }
1955 __ j(unlock_done);
1956
1957 __ block_comment("} Slow path unlock");
1958
1959 } // synchronized
1960
1961 // SLOW PATH Reguard the stack if needed
1962
1963 __ bind(reguard);
1964 save_native_result(masm, ret_type, stack_slots);
1965 __ rt_call(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
1966 restore_native_result(masm, ret_type, stack_slots);
1967 // and continue
1968 __ j(reguard_done);
1969
1970 // SLOW PATH safepoint
1971 {
1972 __ block_comment("safepoint {");
1973 __ bind(safepoint_in_progress);
1974
1975 // Don't use call_VM as it will see a possible pending exception and forward it
1976 // and never return here preventing us from clearing _last_native_pc down below.
1977 //
1978 save_native_result(masm, ret_type, stack_slots);
1979 __ mv(c_rarg0, xthread);
1980 #ifndef PRODUCT
1981 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
1982 #endif
1983 RuntimeAddress target(CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans));
1984 __ relocate(target.rspec(), [&] {
1985 int32_t offset;
1986 __ la_patchable(t0, target, offset);
1987 __ jalr(x1, t0, offset);
1988 });
1989
1990 // Restore any method result value
1991 restore_native_result(masm, ret_type, stack_slots);
1992
1993 __ j(safepoint_in_progress_done);
1994 __ block_comment("} safepoint");
1995 }
1996
1997 // SLOW PATH dtrace support
1998 {
1999 __ block_comment("dtrace entry {");
2000 __ bind(dtrace_method_entry);
2001
2002 // We have all of the arguments setup at this point. We must not touch any register
2003 // argument registers at this point (what if we save/restore them there are no oop?
2004
2005 save_args(masm, total_c_args, c_arg, out_regs);
2006 __ mov_metadata(c_rarg1, method());
2007 __ call_VM_leaf(
2008 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
2009 xthread, c_rarg1);
2010 restore_args(masm, total_c_args, c_arg, out_regs);
2011 __ j(dtrace_method_entry_done);
2012 __ block_comment("} dtrace entry");
2013 }
2014
2015 {
2016 __ block_comment("dtrace exit {");
2017 __ bind(dtrace_method_exit);
2018 save_native_result(masm, ret_type, stack_slots);
2019 __ mov_metadata(c_rarg1, method());
2020 __ call_VM_leaf(
2021 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
2022 xthread, c_rarg1);
2023 restore_native_result(masm, ret_type, stack_slots);
2024 __ j(dtrace_method_exit_done);
2025 __ block_comment("} dtrace exit");
2026 }
2027
2028 __ flush();
2029
2030 nmethod *nm = nmethod::new_native_nmethod(method,
2031 compile_id,
2032 masm->code(),
2033 vep_offset,
2034 frame_complete,
2035 stack_slots / VMRegImpl::slots_per_word,
2036 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2037 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size),
2038 oop_maps);
2039 assert(nm != nullptr, "create native nmethod fail!");
2040 return nm;
2041 }
2042
2043 // this function returns the adjust size (in number of words) to a c2i adapter
2044 // activation for use during deoptimization
2045 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2046 assert(callee_locals >= callee_parameters,
2047 "test and remove; got more parms than locals");
2048 if (callee_locals < callee_parameters) {
2049 return 0; // No adjustment for negative locals
2050 }
2051 int diff = (callee_locals - callee_parameters) * Interpreter::stackElementWords;
2052 // diff is counted in stack words
2053 return align_up(diff, 2);
2054 }
2055
2056 //------------------------------generate_deopt_blob----------------------------
2057 void SharedRuntime::generate_deopt_blob() {
2058 // Allocate space for the code
2059 ResourceMark rm;
2060 // Setup code generation tools
2061 int pad = 0;
2062 #if INCLUDE_JVMCI
2063 if (EnableJVMCI) {
2064 pad += 512; // Increase the buffer size when compiling for JVMCI
2065 }
2066 #endif
2067 CodeBuffer buffer("deopt_blob", 2048 + pad, 1024);
2068 MacroAssembler* masm = new MacroAssembler(&buffer);
2069 int frame_size_in_words = -1;
2070 OopMap* map = nullptr;
2071 OopMapSet *oop_maps = new OopMapSet();
2072 assert_cond(masm != nullptr && oop_maps != nullptr);
2073 RegisterSaver reg_saver(COMPILER2_OR_JVMCI != 0);
2074
2075 // -------------
2076 // This code enters when returning to a de-optimized nmethod. A return
2077 // address has been pushed on the stack, and return values are in
2078 // registers.
2079 // If we are doing a normal deopt then we were called from the patched
2080 // nmethod from the point we returned to the nmethod. So the return
2081 // address on the stack is wrong by NativeCall::instruction_size
2082 // We will adjust the value so it looks like we have the original return
2083 // address on the stack (like when we eagerly deoptimized).
2084 // In the case of an exception pending when deoptimizing, we enter
2085 // with a return address on the stack that points after the call we patched
2086 // into the exception handler. We have the following register state from,
2087 // e.g., the forward exception stub (see stubGenerator_riscv.cpp).
2088 // x10: exception oop
2089 // x9: exception handler
2090 // x13: throwing pc
2091 // So in this case we simply jam x13 into the useless return address and
2092 // the stack looks just like we want.
2093 //
2094 // At this point we need to de-opt. We save the argument return
2095 // registers. We call the first C routine, fetch_unroll_info(). This
2096 // routine captures the return values and returns a structure which
2097 // describes the current frame size and the sizes of all replacement frames.
2098 // The current frame is compiled code and may contain many inlined
2099 // functions, each with their own JVM state. We pop the current frame, then
2100 // push all the new frames. Then we call the C routine unpack_frames() to
2101 // populate these frames. Finally unpack_frames() returns us the new target
2102 // address. Notice that callee-save registers are BLOWN here; they have
2103 // already been captured in the vframeArray at the time the return PC was
2104 // patched.
2105 address start = __ pc();
2106 Label cont;
2107
2108 // Prolog for non exception case!
2109
2110 // Save everything in sight.
2111 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2112
2113 // Normal deoptimization. Save exec mode for unpack_frames.
2114 __ mv(xcpool, Deoptimization::Unpack_deopt); // callee-saved
2115 __ j(cont);
2116
2117 int reexecute_offset = __ pc() - start;
2118 #if INCLUDE_JVMCI && !defined(COMPILER1)
2119 if (EnableJVMCI && UseJVMCICompiler) {
2120 // JVMCI does not use this kind of deoptimization
2121 __ should_not_reach_here();
2122 }
2123 #endif
2124
2125 // Reexecute case
2126 // return address is the pc describes what bci to do re-execute at
2127
2128 // No need to update map as each call to save_live_registers will produce identical oopmap
2129 (void) reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2130
2131 __ mv(xcpool, Deoptimization::Unpack_reexecute); // callee-saved
2132 __ j(cont);
2133
2134 #if INCLUDE_JVMCI
2135 Label after_fetch_unroll_info_call;
2136 int implicit_exception_uncommon_trap_offset = 0;
2137 int uncommon_trap_offset = 0;
2138
2139 if (EnableJVMCI) {
2140 implicit_exception_uncommon_trap_offset = __ pc() - start;
2141
2142 __ ld(ra, Address(xthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2143 __ sd(zr, Address(xthread, in_bytes(JavaThread::jvmci_implicit_exception_pc_offset())));
2144
2145 uncommon_trap_offset = __ pc() - start;
2146
2147 // Save everything in sight.
2148 reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2149 // fetch_unroll_info needs to call last_java_frame()
2150 Label retaddr;
2151 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2152
2153 __ lw(c_rarg1, Address(xthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2154 __ mv(t0, -1);
2155 __ sw(t0, Address(xthread, in_bytes(JavaThread::pending_deoptimization_offset())));
2156
2157 __ mv(xcpool, Deoptimization::Unpack_reexecute);
2158 __ mv(c_rarg0, xthread);
2159 __ orrw(c_rarg2, zr, xcpool); // exec mode
2160 RuntimeAddress target(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap));
2161 __ relocate(target.rspec(), [&] {
2162 int32_t offset;
2163 __ la_patchable(t0, target, offset);
2164 __ jalr(x1, t0, offset);
2165 });
2166 __ bind(retaddr);
2167 oop_maps->add_gc_map( __ pc()-start, map->deep_copy());
2168
2169 __ reset_last_Java_frame(false);
2170
2171 __ j(after_fetch_unroll_info_call);
2172 } // EnableJVMCI
2173 #endif // INCLUDE_JVMCI
2174
2175 int exception_offset = __ pc() - start;
2176
2177 // Prolog for exception case
2178
2179 // all registers are dead at this entry point, except for x10, and
2180 // x13 which contain the exception oop and exception pc
2181 // respectively. Set them in TLS and fall thru to the
2182 // unpack_with_exception_in_tls entry point.
2183
2184 __ sd(x13, Address(xthread, JavaThread::exception_pc_offset()));
2185 __ sd(x10, Address(xthread, JavaThread::exception_oop_offset()));
2186
2187 int exception_in_tls_offset = __ pc() - start;
2188
2189 // new implementation because exception oop is now passed in JavaThread
2190
2191 // Prolog for exception case
2192 // All registers must be preserved because they might be used by LinearScan
2193 // Exceptiop oop and throwing PC are passed in JavaThread
2194 // tos: stack at point of call to method that threw the exception (i.e. only
2195 // args are on the stack, no return address)
2196
2197 // The return address pushed by save_live_registers will be patched
2198 // later with the throwing pc. The correct value is not available
2199 // now because loading it from memory would destroy registers.
2200
2201 // NB: The SP at this point must be the SP of the method that is
2202 // being deoptimized. Deoptimization assumes that the frame created
2203 // here by save_live_registers is immediately below the method's SP.
2204 // This is a somewhat fragile mechanism.
2205
2206 // Save everything in sight.
2207 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2208
2209 // Now it is safe to overwrite any register
2210
2211 // Deopt during an exception. Save exec mode for unpack_frames.
2212 __ mv(xcpool, Deoptimization::Unpack_exception); // callee-saved
2213
2214 // load throwing pc from JavaThread and patch it as the return address
2215 // of the current frame. Then clear the field in JavaThread
2216
2217 __ ld(x13, Address(xthread, JavaThread::exception_pc_offset()));
2218 __ sd(x13, Address(fp, frame::return_addr_offset * wordSize));
2219 __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2220
2221 #ifdef ASSERT
2222 // verify that there is really an exception oop in JavaThread
2223 __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2224 __ verify_oop(x10);
2225
2226 // verify that there is no pending exception
2227 Label no_pending_exception;
2228 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2229 __ beqz(t0, no_pending_exception);
2230 __ stop("must not have pending exception here");
2231 __ bind(no_pending_exception);
2232 #endif
2233
2234 __ bind(cont);
2235
2236 // Call C code. Need thread and this frame, but NOT official VM entry
2237 // crud. We cannot block on this call, no GC can happen.
2238 //
2239 // UnrollBlock* fetch_unroll_info(JavaThread* thread)
2240
2241 // fetch_unroll_info needs to call last_java_frame().
2242
2243 Label retaddr;
2244 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2245 #ifdef ASSERT
2246 {
2247 Label L;
2248 __ ld(t0, Address(xthread,
2249 JavaThread::last_Java_fp_offset()));
2250 __ beqz(t0, L);
2251 __ stop("SharedRuntime::generate_deopt_blob: last_Java_fp not cleared");
2252 __ bind(L);
2253 }
2254 #endif // ASSERT
2255 __ mv(c_rarg0, xthread);
2256 __ mv(c_rarg1, xcpool);
2257 RuntimeAddress target(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info));
2258 __ relocate(target.rspec(), [&] {
2259 int32_t offset;
2260 __ la_patchable(t0, target, offset);
2261 __ jalr(x1, t0, offset);
2262 });
2263 __ bind(retaddr);
2264
2265 // Need to have an oopmap that tells fetch_unroll_info where to
2266 // find any register it might need.
2267 oop_maps->add_gc_map(__ pc() - start, map);
2268
2269 __ reset_last_Java_frame(false);
2270
2271 #if INCLUDE_JVMCI
2272 if (EnableJVMCI) {
2273 __ bind(after_fetch_unroll_info_call);
2274 }
2275 #endif
2276
2277 // Load UnrollBlock* into x15
2278 __ mv(x15, x10);
2279
2280 __ lwu(xcpool, Address(x15, Deoptimization::UnrollBlock::unpack_kind_offset()));
2281 Label noException;
2282 __ mv(t0, Deoptimization::Unpack_exception);
2283 __ bne(xcpool, t0, noException); // Was exception pending?
2284 __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2285 __ ld(x13, Address(xthread, JavaThread::exception_pc_offset()));
2286 __ sd(zr, Address(xthread, JavaThread::exception_oop_offset()));
2287 __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2288
2289 __ verify_oop(x10);
2290
2291 // Overwrite the result registers with the exception results.
2292 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2293
2294 __ bind(noException);
2295
2296 // Only register save data is on the stack.
2297 // Now restore the result registers. Everything else is either dead
2298 // or captured in the vframeArray.
2299
2300 // Restore fp result register
2301 __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2302 // Restore integer result register
2303 __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2304
2305 // Pop all of the register save area off the stack
2306 __ add(sp, sp, frame_size_in_words * wordSize);
2307
2308 // All of the register save area has been popped of the stack. Only the
2309 // return address remains.
2310
2311 // Pop all the frames we must move/replace.
2312 //
2313 // Frame picture (youngest to oldest)
2314 // 1: self-frame (no frame link)
2315 // 2: deopting frame (no frame link)
2316 // 3: caller of deopting frame (could be compiled/interpreted).
2317 //
2318 // Note: by leaving the return address of self-frame on the stack
2319 // and using the size of frame 2 to adjust the stack
2320 // when we are done the return to frame 3 will still be on the stack.
2321
2322 // Pop deoptimized frame
2323 __ lwu(x12, Address(x15, Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset()));
2324 __ sub(x12, x12, 2 * wordSize);
2325 __ add(sp, sp, x12);
2326 __ ld(fp, Address(sp, 0));
2327 __ ld(ra, Address(sp, wordSize));
2328 __ addi(sp, sp, 2 * wordSize);
2329 // RA should now be the return address to the caller (3)
2330
2331 #ifdef ASSERT
2332 // Compilers generate code that bang the stack by as much as the
2333 // interpreter would need. So this stack banging should never
2334 // trigger a fault. Verify that it does not on non product builds.
2335 __ lwu(x9, Address(x15, Deoptimization::UnrollBlock::total_frame_sizes_offset()));
2336 __ bang_stack_size(x9, x12);
2337 #endif
2338 // Load address of array of frame pcs into x12
2339 __ ld(x12, Address(x15, Deoptimization::UnrollBlock::frame_pcs_offset()));
2340
2341 // Load address of array of frame sizes into x14
2342 __ ld(x14, Address(x15, Deoptimization::UnrollBlock::frame_sizes_offset()));
2343
2344 // Load counter into x13
2345 __ lwu(x13, Address(x15, Deoptimization::UnrollBlock::number_of_frames_offset()));
2346
2347 // Now adjust the caller's stack to make up for the extra locals
2348 // but record the original sp so that we can save it in the skeletal interpreter
2349 // frame and the stack walking of interpreter_sender will get the unextended sp
2350 // value and not the "real" sp value.
2351
2352 const Register sender_sp = x16;
2353
2354 __ mv(sender_sp, sp);
2355 __ lwu(x9, Address(x15,
2356 Deoptimization::UnrollBlock::
2357 caller_adjustment_offset()));
2358 __ sub(sp, sp, x9);
2359
2360 // Push interpreter frames in a loop
2361 __ mv(t0, 0xDEADDEAD); // Make a recognizable pattern
2362 __ mv(t1, t0);
2363 Label loop;
2364 __ bind(loop);
2365 __ ld(x9, Address(x14, 0)); // Load frame size
2366 __ addi(x14, x14, wordSize);
2367 __ sub(x9, x9, 2 * wordSize); // We'll push pc and fp by hand
2368 __ ld(ra, Address(x12, 0)); // Load pc
2369 __ addi(x12, x12, wordSize);
2370 __ enter(); // Save old & set new fp
2371 __ sub(sp, sp, x9); // Prolog
2372 // This value is corrected by layout_activation_impl
2373 __ sd(zr, Address(fp, frame::interpreter_frame_last_sp_offset * wordSize));
2374 __ sd(sender_sp, Address(fp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable
2375 __ mv(sender_sp, sp); // Pass sender_sp to next frame
2376 __ addi(x13, x13, -1); // Decrement counter
2377 __ bnez(x13, loop);
2378
2379 // Re-push self-frame
2380 __ ld(ra, Address(x12));
2381 __ enter();
2382
2383 // Allocate a full sized register save area. We subtract 2 because
2384 // enter() just pushed 2 words
2385 __ sub(sp, sp, (frame_size_in_words - 2) * wordSize);
2386
2387 // Restore frame locals after moving the frame
2388 __ fsd(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2389 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2390
2391 // Call C code. Need thread but NOT official VM entry
2392 // crud. We cannot block on this call, no GC can happen. Call should
2393 // restore return values to their stack-slots with the new SP.
2394 //
2395 // void Deoptimization::unpack_frames(JavaThread* thread, int exec_mode)
2396
2397 // Use fp because the frames look interpreted now
2398 // Don't need the precise return PC here, just precise enough to point into this code blob.
2399 address the_pc = __ pc();
2400 __ set_last_Java_frame(sp, fp, the_pc, t0);
2401
2402 __ mv(c_rarg0, xthread);
2403 __ mv(c_rarg1, xcpool); // second arg: exec_mode
2404 target = RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames));
2405 __ relocate(target.rspec(), [&] {
2406 int32_t offset;
2407 __ la_patchable(t0, target, offset);
2408 __ jalr(x1, t0, offset);
2409 });
2410
2411 // Set an oopmap for the call site
2412 // Use the same PC we used for the last java frame
2413 oop_maps->add_gc_map(the_pc - start,
2414 new OopMap(frame_size_in_words, 0));
2415
2416 // Clear fp AND pc
2417 __ reset_last_Java_frame(true);
2418
2419 // Collect return values
2420 __ fld(f10, Address(sp, reg_saver.freg_offset_in_bytes(f10)));
2421 __ ld(x10, Address(sp, reg_saver.reg_offset_in_bytes(x10)));
2422
2423 // Pop self-frame.
2424 __ leave(); // Epilog
2425
2426 // Jump to interpreter
2427 __ ret();
2428
2429 // Make sure all code is generated
2430 masm->flush();
2431
2432 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words);
2433 assert(_deopt_blob != nullptr, "create deoptimization blob fail!");
2434 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
2435 #if INCLUDE_JVMCI
2436 if (EnableJVMCI) {
2437 _deopt_blob->set_uncommon_trap_offset(uncommon_trap_offset);
2438 _deopt_blob->set_implicit_exception_uncommon_trap_offset(implicit_exception_uncommon_trap_offset);
2439 }
2440 #endif
2441 }
2442
2443 // Number of stack slots between incoming argument block and the start of
2444 // a new frame. The PROLOG must add this many slots to the stack. The
2445 // EPILOG must remove this many slots.
2446 // RISCV needs two words for RA (return address) and FP (frame pointer).
2447 uint SharedRuntime::in_preserve_stack_slots() {
2448 return 2 * VMRegImpl::slots_per_word;
2449 }
2450
2451 uint SharedRuntime::out_preserve_stack_slots() {
2452 return 0;
2453 }
2454
2455 #ifdef COMPILER2
2456 //------------------------------generate_uncommon_trap_blob--------------------
2457 void SharedRuntime::generate_uncommon_trap_blob() {
2458 // Allocate space for the code
2459 ResourceMark rm;
2460 // Setup code generation tools
2461 CodeBuffer buffer("uncommon_trap_blob", 2048, 1024);
2462 MacroAssembler* masm = new MacroAssembler(&buffer);
2463 assert_cond(masm != nullptr);
2464
2465 assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned");
2466
2467 address start = __ pc();
2468
2469 // Push self-frame. We get here with a return address in RA
2470 // and sp should be 16 byte aligned
2471 // push fp and retaddr by hand
2472 __ addi(sp, sp, -2 * wordSize);
2473 __ sd(ra, Address(sp, wordSize));
2474 __ sd(fp, Address(sp, 0));
2475 // we don't expect an arg reg save area
2476 #ifndef PRODUCT
2477 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
2478 #endif
2479 // compiler left unloaded_class_index in j_rarg0 move to where the
2480 // runtime expects it.
2481 __ sign_extend(c_rarg1, j_rarg0, 32);
2482
2483 // we need to set the past SP to the stack pointer of the stub frame
2484 // and the pc to the address where this runtime call will return
2485 // although actually any pc in this code blob will do).
2486 Label retaddr;
2487 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2488
2489 // Call C code. Need thread but NOT official VM entry
2490 // crud. We cannot block on this call, no GC can happen. Call should
2491 // capture callee-saved registers as well as return values.
2492 //
2493 // UnrollBlock* uncommon_trap(JavaThread* thread, jint unloaded_class_index, jint exec_mode)
2494 //
2495 // n.b. 3 gp args, 0 fp args, integral return type
2496
2497 __ mv(c_rarg0, xthread);
2498 __ mv(c_rarg2, Deoptimization::Unpack_uncommon_trap);
2499 RuntimeAddress target(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap));
2500 __ relocate(target.rspec(), [&] {
2501 int32_t offset;
2502 __ la_patchable(t0, target, offset);
2503 __ jalr(x1, t0, offset);
2504 });
2505 __ bind(retaddr);
2506
2507 // Set an oopmap for the call site
2508 OopMapSet* oop_maps = new OopMapSet();
2509 OopMap* map = new OopMap(SimpleRuntimeFrame::framesize, 0);
2510 assert_cond(oop_maps != nullptr && map != nullptr);
2511
2512 // location of fp is known implicitly by the frame sender code
2513
2514 oop_maps->add_gc_map(__ pc() - start, map);
2515
2516 __ reset_last_Java_frame(false);
2517
2518 // move UnrollBlock* into x14
2519 __ mv(x14, x10);
2520
2521 #ifdef ASSERT
2522 { Label L;
2523 __ lwu(t0, Address(x14, Deoptimization::UnrollBlock::unpack_kind_offset()));
2524 __ mv(t1, Deoptimization::Unpack_uncommon_trap);
2525 __ beq(t0, t1, L);
2526 __ stop("SharedRuntime::generate_uncommon_trap_blob: expected Unpack_uncommon_trap");
2527 __ bind(L);
2528 }
2529 #endif
2530
2531 // Pop all the frames we must move/replace.
2532 //
2533 // Frame picture (youngest to oldest)
2534 // 1: self-frame (no frame link)
2535 // 2: deopting frame (no frame link)
2536 // 3: caller of deopting frame (could be compiled/interpreted).
2537
2538 __ add(sp, sp, (SimpleRuntimeFrame::framesize) << LogBytesPerInt); // Epilog!
2539
2540 // Pop deoptimized frame (int)
2541 __ lwu(x12, Address(x14,
2542 Deoptimization::UnrollBlock::
2543 size_of_deoptimized_frame_offset()));
2544 __ sub(x12, x12, 2 * wordSize);
2545 __ add(sp, sp, x12);
2546 __ ld(fp, Address(sp, 0));
2547 __ ld(ra, Address(sp, wordSize));
2548 __ addi(sp, sp, 2 * wordSize);
2549 // RA should now be the return address to the caller (3) frame
2550
2551 #ifdef ASSERT
2552 // Compilers generate code that bang the stack by as much as the
2553 // interpreter would need. So this stack banging should never
2554 // trigger a fault. Verify that it does not on non product builds.
2555 __ lwu(x11, Address(x14,
2556 Deoptimization::UnrollBlock::
2557 total_frame_sizes_offset()));
2558 __ bang_stack_size(x11, x12);
2559 #endif
2560
2561 // Load address of array of frame pcs into x12 (address*)
2562 __ ld(x12, Address(x14,
2563 Deoptimization::UnrollBlock::frame_pcs_offset()));
2564
2565 // Load address of array of frame sizes into x15 (intptr_t*)
2566 __ ld(x15, Address(x14,
2567 Deoptimization::UnrollBlock::
2568 frame_sizes_offset()));
2569
2570 // Counter
2571 __ lwu(x13, Address(x14,
2572 Deoptimization::UnrollBlock::
2573 number_of_frames_offset())); // (int)
2574
2575 // Now adjust the caller's stack to make up for the extra locals but
2576 // record the original sp so that we can save it in the skeletal
2577 // interpreter frame and the stack walking of interpreter_sender
2578 // will get the unextended sp value and not the "real" sp value.
2579
2580 const Register sender_sp = t1; // temporary register
2581
2582 __ lwu(x11, Address(x14,
2583 Deoptimization::UnrollBlock::
2584 caller_adjustment_offset())); // (int)
2585 __ mv(sender_sp, sp);
2586 __ sub(sp, sp, x11);
2587
2588 // Push interpreter frames in a loop
2589 Label loop;
2590 __ bind(loop);
2591 __ ld(x11, Address(x15, 0)); // Load frame size
2592 __ sub(x11, x11, 2 * wordSize); // We'll push pc and fp by hand
2593 __ ld(ra, Address(x12, 0)); // Save return address
2594 __ enter(); // and old fp & set new fp
2595 __ sub(sp, sp, x11); // Prolog
2596 __ sd(sender_sp, Address(fp, frame::interpreter_frame_sender_sp_offset * wordSize)); // Make it walkable
2597 // This value is corrected by layout_activation_impl
2598 __ sd(zr, Address(fp, frame::interpreter_frame_last_sp_offset * wordSize));
2599 __ mv(sender_sp, sp); // Pass sender_sp to next frame
2600 __ add(x15, x15, wordSize); // Bump array pointer (sizes)
2601 __ add(x12, x12, wordSize); // Bump array pointer (pcs)
2602 __ subw(x13, x13, 1); // Decrement counter
2603 __ bgtz(x13, loop);
2604 __ ld(ra, Address(x12, 0)); // save final return address
2605 // Re-push self-frame
2606 __ enter(); // & old fp & set new fp
2607
2608 // Use fp because the frames look interpreted now
2609 // Save "the_pc" since it cannot easily be retrieved using the last_java_SP after we aligned SP.
2610 // Don't need the precise return PC here, just precise enough to point into this code blob.
2611 address the_pc = __ pc();
2612 __ set_last_Java_frame(sp, fp, the_pc, t0);
2613
2614 // Call C code. Need thread but NOT official VM entry
2615 // crud. We cannot block on this call, no GC can happen. Call should
2616 // restore return values to their stack-slots with the new SP.
2617 //
2618 // BasicType unpack_frames(JavaThread* thread, int exec_mode)
2619 //
2620
2621 // n.b. 2 gp args, 0 fp args, integral return type
2622
2623 // sp should already be aligned
2624 __ mv(c_rarg0, xthread);
2625 __ mv(c_rarg1, Deoptimization::Unpack_uncommon_trap);
2626 target = RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames));
2627 __ relocate(target.rspec(), [&] {
2628 int32_t offset;
2629 __ la_patchable(t0, target, offset);
2630 __ jalr(x1, t0, offset);
2631 });
2632
2633 // Set an oopmap for the call site
2634 // Use the same PC we used for the last java frame
2635 oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0));
2636
2637 // Clear fp AND pc
2638 __ reset_last_Java_frame(true);
2639
2640 // Pop self-frame.
2641 __ leave(); // Epilog
2642
2643 // Jump to interpreter
2644 __ ret();
2645
2646 // Make sure all code is generated
2647 masm->flush();
2648
2649 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps,
2650 SimpleRuntimeFrame::framesize >> 1);
2651 }
2652 #endif // COMPILER2
2653
2654 //------------------------------generate_handler_blob------
2655 //
2656 // Generate a special Compile2Runtime blob that saves all registers,
2657 // and setup oopmap.
2658 //
2659 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) {
2660 ResourceMark rm;
2661 OopMapSet *oop_maps = new OopMapSet();
2662 assert_cond(oop_maps != nullptr);
2663 OopMap* map = nullptr;
2664
2665 // Allocate space for the code. Setup code generation tools.
2666 CodeBuffer buffer("handler_blob", 2048, 1024);
2667 MacroAssembler* masm = new MacroAssembler(&buffer);
2668 assert_cond(masm != nullptr);
2669
2670 address start = __ pc();
2671 address call_pc = nullptr;
2672 int frame_size_in_words = -1;
2673 bool cause_return = (poll_type == POLL_AT_RETURN);
2674 RegisterSaver reg_saver(poll_type == POLL_AT_VECTOR_LOOP /* save_vectors */);
2675
2676 // Save Integer and Float registers.
2677 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2678
2679 // The following is basically a call_VM. However, we need the precise
2680 // address of the call in order to generate an oopmap. Hence, we do all the
2681 // work ourselves.
2682
2683 Label retaddr;
2684 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2685
2686 // The return address must always be correct so that frame constructor never
2687 // sees an invalid pc.
2688
2689 if (!cause_return) {
2690 // overwrite the return address pushed by save_live_registers
2691 // Additionally, x18 is a callee-saved register so we can look at
2692 // it later to determine if someone changed the return address for
2693 // us!
2694 __ ld(x18, Address(xthread, JavaThread::saved_exception_pc_offset()));
2695 __ sd(x18, Address(fp, frame::return_addr_offset * wordSize));
2696 }
2697
2698 // Do the call
2699 __ mv(c_rarg0, xthread);
2700 RuntimeAddress target(call_ptr);
2701 __ relocate(target.rspec(), [&] {
2702 int32_t offset;
2703 __ la_patchable(t0, target, offset);
2704 __ jalr(x1, t0, offset);
2705 });
2706 __ bind(retaddr);
2707
2708 // Set an oopmap for the call site. This oopmap will map all
2709 // oop-registers and debug-info registers as callee-saved. This
2710 // will allow deoptimization at this safepoint to find all possible
2711 // debug-info recordings, as well as let GC find all oops.
2712
2713 oop_maps->add_gc_map( __ pc() - start, map);
2714
2715 Label noException;
2716
2717 __ reset_last_Java_frame(false);
2718
2719 __ membar(MacroAssembler::LoadLoad | MacroAssembler::LoadStore);
2720
2721 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2722 __ beqz(t0, noException);
2723
2724 // Exception pending
2725
2726 reg_saver.restore_live_registers(masm);
2727
2728 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2729
2730 // No exception case
2731 __ bind(noException);
2732
2733 Label no_adjust, bail;
2734 if (!cause_return) {
2735 // If our stashed return pc was modified by the runtime we avoid touching it
2736 __ ld(t0, Address(fp, frame::return_addr_offset * wordSize));
2737 __ bne(x18, t0, no_adjust);
2738
2739 #ifdef ASSERT
2740 // Verify the correct encoding of the poll we're about to skip.
2741 // See NativeInstruction::is_lwu_to_zr()
2742 __ lwu(t0, Address(x18));
2743 __ andi(t1, t0, 0b0000011);
2744 __ mv(t2, 0b0000011);
2745 __ bne(t1, t2, bail); // 0-6:0b0000011
2746 __ srli(t1, t0, 7);
2747 __ andi(t1, t1, 0b00000);
2748 __ bnez(t1, bail); // 7-11:0b00000
2749 __ srli(t1, t0, 12);
2750 __ andi(t1, t1, 0b110);
2751 __ mv(t2, 0b110);
2752 __ bne(t1, t2, bail); // 12-14:0b110
2753 #endif
2754 // Adjust return pc forward to step over the safepoint poll instruction
2755 __ add(x18, x18, NativeInstruction::instruction_size);
2756 __ sd(x18, Address(fp, frame::return_addr_offset * wordSize));
2757 }
2758
2759 __ bind(no_adjust);
2760 // Normal exit, restore registers and exit.
2761
2762 reg_saver.restore_live_registers(masm);
2763 __ ret();
2764
2765 #ifdef ASSERT
2766 __ bind(bail);
2767 __ stop("Attempting to adjust pc to skip safepoint poll but the return point is not what we expected");
2768 #endif
2769
2770 // Make sure all code is generated
2771 masm->flush();
2772
2773 // Fill-out other meta info
2774 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words);
2775 }
2776
2777 //
2778 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
2779 //
2780 // Generate a stub that calls into vm to find out the proper destination
2781 // of a java call. All the argument registers are live at this point
2782 // but since this is generic code we don't know what they are and the caller
2783 // must do any gc of the args.
2784 //
2785 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) {
2786 assert(StubRoutines::forward_exception_entry() != nullptr, "must be generated before");
2787
2788 // allocate space for the code
2789 ResourceMark rm;
2790
2791 CodeBuffer buffer(name, 1000, 512);
2792 MacroAssembler* masm = new MacroAssembler(&buffer);
2793 assert_cond(masm != nullptr);
2794
2795 int frame_size_in_words = -1;
2796 RegisterSaver reg_saver(false /* save_vectors */);
2797
2798 OopMapSet *oop_maps = new OopMapSet();
2799 assert_cond(oop_maps != nullptr);
2800 OopMap* map = nullptr;
2801
2802 int start = __ offset();
2803
2804 map = reg_saver.save_live_registers(masm, 0, &frame_size_in_words);
2805
2806 int frame_complete = __ offset();
2807
2808 {
2809 Label retaddr;
2810 __ set_last_Java_frame(sp, noreg, retaddr, t0);
2811
2812 __ mv(c_rarg0, xthread);
2813 RuntimeAddress target(destination);
2814 __ relocate(target.rspec(), [&] {
2815 int32_t offset;
2816 __ la_patchable(t0, target, offset);
2817 __ jalr(x1, t0, offset);
2818 });
2819 __ bind(retaddr);
2820 }
2821
2822 // Set an oopmap for the call site.
2823 // We need this not only for callee-saved registers, but also for volatile
2824 // registers that the compiler might be keeping live across a safepoint.
2825
2826 oop_maps->add_gc_map( __ offset() - start, map);
2827
2828 // x10 contains the address we are going to jump to assuming no exception got installed
2829
2830 // clear last_Java_sp
2831 __ reset_last_Java_frame(false);
2832 // check for pending exceptions
2833 Label pending;
2834 __ ld(t0, Address(xthread, Thread::pending_exception_offset()));
2835 __ bnez(t0, pending);
2836
2837 // get the returned Method*
2838 __ get_vm_result_2(xmethod, xthread);
2839 __ sd(xmethod, Address(sp, reg_saver.reg_offset_in_bytes(xmethod)));
2840
2841 // x10 is where we want to jump, overwrite t0 which is saved and temporary
2842 __ sd(x10, Address(sp, reg_saver.reg_offset_in_bytes(t0)));
2843 reg_saver.restore_live_registers(masm);
2844
2845 // We are back to the original state on entry and ready to go.
2846
2847 __ jr(t0);
2848
2849 // Pending exception after the safepoint
2850
2851 __ bind(pending);
2852
2853 reg_saver.restore_live_registers(masm);
2854
2855 // exception pending => remove activation and forward to exception handler
2856
2857 __ sd(zr, Address(xthread, JavaThread::vm_result_offset()));
2858
2859 __ ld(x10, Address(xthread, Thread::pending_exception_offset()));
2860 __ far_jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2861
2862 // -------------
2863 // make sure all code is generated
2864 masm->flush();
2865
2866 // return the blob
2867 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_words, oop_maps, true);
2868 }
2869
2870 #ifdef COMPILER2
2871 //------------------------------generate_exception_blob---------------------------
2872 // creates exception blob at the end
2873 // Using exception blob, this code is jumped from a compiled method.
2874 // (see emit_exception_handler in riscv.ad file)
2875 //
2876 // Given an exception pc at a call we call into the runtime for the
2877 // handler in this method. This handler might merely restore state
2878 // (i.e. callee save registers) unwind the frame and jump to the
2879 // exception handler for the nmethod if there is no Java level handler
2880 // for the nmethod.
2881 //
2882 // This code is entered with a jmp.
2883 //
2884 // Arguments:
2885 // x10: exception oop
2886 // x13: exception pc
2887 //
2888 // Results:
2889 // x10: exception oop
2890 // x13: exception pc in caller
2891 // destination: exception handler of caller
2892 //
2893 // Note: the exception pc MUST be at a call (precise debug information)
2894 // Registers x10, x13, x12, x14, x15, t0 are not callee saved.
2895 //
2896
2897 void OptoRuntime::generate_exception_blob() {
2898 assert(!OptoRuntime::is_callee_saved_register(R13_num), "");
2899 assert(!OptoRuntime::is_callee_saved_register(R10_num), "");
2900 assert(!OptoRuntime::is_callee_saved_register(R12_num), "");
2901
2902 assert(SimpleRuntimeFrame::framesize % 4 == 0, "sp not 16-byte aligned");
2903
2904 // Allocate space for the code
2905 ResourceMark rm;
2906 // Setup code generation tools
2907 CodeBuffer buffer("exception_blob", 2048, 1024);
2908 MacroAssembler* masm = new MacroAssembler(&buffer);
2909 assert_cond(masm != nullptr);
2910
2911 // TODO check various assumptions made here
2912 //
2913 // make sure we do so before running this
2914
2915 address start = __ pc();
2916
2917 // push fp and retaddr by hand
2918 // Exception pc is 'return address' for stack walker
2919 __ addi(sp, sp, -2 * wordSize);
2920 __ sd(ra, Address(sp, wordSize));
2921 __ sd(fp, Address(sp));
2922 // there are no callee save registers and we don't expect an
2923 // arg reg save area
2924 #ifndef PRODUCT
2925 assert(frame::arg_reg_save_area_bytes == 0, "not expecting frame reg save area");
2926 #endif
2927 // Store exception in Thread object. We cannot pass any arguments to the
2928 // handle_exception call, since we do not want to make any assumption
2929 // about the size of the frame where the exception happened in.
2930 __ sd(x10, Address(xthread, JavaThread::exception_oop_offset()));
2931 __ sd(x13, Address(xthread, JavaThread::exception_pc_offset()));
2932
2933 // This call does all the hard work. It checks if an exception handler
2934 // exists in the method.
2935 // If so, it returns the handler address.
2936 // If not, it prepares for stack-unwinding, restoring the callee-save
2937 // registers of the frame being removed.
2938 //
2939 // address OptoRuntime::handle_exception_C(JavaThread* thread)
2940 //
2941 // n.b. 1 gp arg, 0 fp args, integral return type
2942
2943 // the stack should always be aligned
2944 address the_pc = __ pc();
2945 __ set_last_Java_frame(sp, noreg, the_pc, t0);
2946 __ mv(c_rarg0, xthread);
2947 RuntimeAddress target(CAST_FROM_FN_PTR(address, OptoRuntime::handle_exception_C));
2948 __ relocate(target.rspec(), [&] {
2949 int32_t offset;
2950 __ la_patchable(t0, target, offset);
2951 __ jalr(x1, t0, offset);
2952 });
2953
2954
2955 // handle_exception_C is a special VM call which does not require an explicit
2956 // instruction sync afterwards.
2957
2958 // Set an oopmap for the call site. This oopmap will only be used if we
2959 // are unwinding the stack. Hence, all locations will be dead.
2960 // Callee-saved registers will be the same as the frame above (i.e.,
2961 // handle_exception_stub), since they were restored when we got the
2962 // exception.
2963
2964 OopMapSet* oop_maps = new OopMapSet();
2965 assert_cond(oop_maps != nullptr);
2966
2967 oop_maps->add_gc_map(the_pc - start, new OopMap(SimpleRuntimeFrame::framesize, 0));
2968
2969 __ reset_last_Java_frame(false);
2970
2971 // Restore callee-saved registers
2972
2973 // fp is an implicitly saved callee saved register (i.e. the calling
2974 // convention will save restore it in prolog/epilog) Other than that
2975 // there are no callee save registers now that adapter frames are gone.
2976 // and we dont' expect an arg reg save area
2977 __ ld(fp, Address(sp));
2978 __ ld(x13, Address(sp, wordSize));
2979 __ addi(sp, sp , 2 * wordSize);
2980
2981 // x10: exception handler
2982
2983 // We have a handler in x10 (could be deopt blob).
2984 __ mv(t0, x10);
2985
2986 // Get the exception oop
2987 __ ld(x10, Address(xthread, JavaThread::exception_oop_offset()));
2988 // Get the exception pc in case we are deoptimized
2989 __ ld(x14, Address(xthread, JavaThread::exception_pc_offset()));
2990 #ifdef ASSERT
2991 __ sd(zr, Address(xthread, JavaThread::exception_handler_pc_offset()));
2992 __ sd(zr, Address(xthread, JavaThread::exception_pc_offset()));
2993 #endif
2994 // Clear the exception oop so GC no longer processes it as a root.
2995 __ sd(zr, Address(xthread, JavaThread::exception_oop_offset()));
2996
2997 // x10: exception oop
2998 // t0: exception handler
2999 // x14: exception pc
3000 // Jump to handler
3001
3002 __ jr(t0);
3003
3004 // Make sure all code is generated
3005 masm->flush();
3006
3007 // Set exception blob
3008 _exception_blob = ExceptionBlob::create(&buffer, oop_maps, SimpleRuntimeFrame::framesize >> 1);
3009 }
3010 #endif // COMPILER2