1 /*
2 * Copyright (c) 1997, 2025, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2012, 2025 SAP SE. All rights reserved.
4 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
5 *
6 * This code is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 only, as
8 * published by the Free Software Foundation.
9 *
10 * This code is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
13 * version 2 for more details (a copy is included in the LICENSE file that
14 * accompanied this code).
15 *
16 * You should have received a copy of the GNU General Public License version
17 * 2 along with this work; if not, write to the Free Software Foundation,
18 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
19 *
20 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
21 * or visit www.oracle.com if you need additional information or have any
22 * questions.
23 *
24 */
25
26 #include "asm/macroAssembler.inline.hpp"
27 #include "code/debugInfoRec.hpp"
28 #include "code/compiledIC.hpp"
29 #include "code/vtableStubs.hpp"
30 #include "frame_ppc.hpp"
31 #include "compiler/oopMap.hpp"
32 #include "gc/shared/gcLocker.hpp"
33 #include "interpreter/interpreter.hpp"
34 #include "interpreter/interp_masm.hpp"
35 #include "memory/resourceArea.hpp"
36 #include "oops/klass.inline.hpp"
37 #include "prims/methodHandles.hpp"
38 #include "runtime/continuation.hpp"
39 #include "runtime/continuationEntry.inline.hpp"
40 #include "runtime/jniHandles.hpp"
41 #include "runtime/os.inline.hpp"
42 #include "runtime/safepointMechanism.hpp"
43 #include "runtime/sharedRuntime.hpp"
44 #include "runtime/signature.hpp"
45 #include "runtime/stubRoutines.hpp"
46 #include "runtime/timerTrace.hpp"
47 #include "runtime/vframeArray.hpp"
48 #include "utilities/align.hpp"
49 #include "utilities/macros.hpp"
50 #include "vmreg_ppc.inline.hpp"
51 #ifdef COMPILER1
52 #include "c1/c1_Runtime1.hpp"
53 #endif
54 #ifdef COMPILER2
55 #include "opto/ad.hpp"
56 #include "opto/runtime.hpp"
57 #endif
58
59 #include <alloca.h>
60
61 #define __ masm->
62
63 #ifdef PRODUCT
64 #define BLOCK_COMMENT(str) // nothing
65 #else
66 #define BLOCK_COMMENT(str) __ block_comment(str)
67 #endif
68
69 #define BIND(label) bind(label); BLOCK_COMMENT(#label ":")
70
71
72 class RegisterSaver {
73 // Used for saving volatile registers.
74 public:
75
76 // Support different return pc locations.
77 enum ReturnPCLocation {
78 return_pc_is_lr,
79 return_pc_is_pre_saved,
80 return_pc_is_thread_saved_exception_pc
81 };
82
83 static OopMap* push_frame_reg_args_and_save_live_registers(MacroAssembler* masm,
84 int* out_frame_size_in_bytes,
85 bool generate_oop_map,
86 int return_pc_adjustment,
87 ReturnPCLocation return_pc_location,
88 bool save_vectors = false);
89 static void restore_live_registers_and_pop_frame(MacroAssembler* masm,
90 int frame_size_in_bytes,
91 bool restore_ctr,
92 bool save_vectors = false);
93
94 static void push_frame_and_save_argument_registers(MacroAssembler* masm,
95 Register r_temp,
96 int frame_size,
97 int total_args,
98 const VMRegPair *regs, const VMRegPair *regs2 = nullptr);
99 static void restore_argument_registers_and_pop_frame(MacroAssembler*masm,
100 int frame_size,
101 int total_args,
102 const VMRegPair *regs, const VMRegPair *regs2 = nullptr);
103
104 // During deoptimization only the result registers need to be restored
105 // all the other values have already been extracted.
106 static void restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes);
107
108 // Constants and data structures:
109
110 typedef enum {
111 int_reg,
112 float_reg,
113 special_reg,
114 vec_reg
115 } RegisterType;
116
117 typedef enum {
118 reg_size = 8,
119 half_reg_size = reg_size / 2,
120 vec_reg_size = 16
121 } RegisterConstants;
122
123 typedef struct {
124 RegisterType reg_type;
125 int reg_num;
126 VMReg vmreg;
127 } LiveRegType;
128 };
129
130
131 #define RegisterSaver_LiveIntReg(regname) \
132 { RegisterSaver::int_reg, regname->encoding(), regname->as_VMReg() }
133
134 #define RegisterSaver_LiveFloatReg(regname) \
135 { RegisterSaver::float_reg, regname->encoding(), regname->as_VMReg() }
136
137 #define RegisterSaver_LiveSpecialReg(regname) \
138 { RegisterSaver::special_reg, regname->encoding(), regname->as_VMReg() }
139
140 #define RegisterSaver_LiveVecReg(regname) \
141 { RegisterSaver::vec_reg, regname->encoding(), regname->as_VMReg() }
142
143 static const RegisterSaver::LiveRegType RegisterSaver_LiveRegs[] = {
144 // Live registers which get spilled to the stack. Register
145 // positions in this array correspond directly to the stack layout.
146
147 //
148 // live special registers:
149 //
150 RegisterSaver_LiveSpecialReg(SR_CTR),
151 //
152 // live float registers:
153 //
154 RegisterSaver_LiveFloatReg( F0 ),
155 RegisterSaver_LiveFloatReg( F1 ),
156 RegisterSaver_LiveFloatReg( F2 ),
157 RegisterSaver_LiveFloatReg( F3 ),
158 RegisterSaver_LiveFloatReg( F4 ),
159 RegisterSaver_LiveFloatReg( F5 ),
160 RegisterSaver_LiveFloatReg( F6 ),
161 RegisterSaver_LiveFloatReg( F7 ),
162 RegisterSaver_LiveFloatReg( F8 ),
163 RegisterSaver_LiveFloatReg( F9 ),
164 RegisterSaver_LiveFloatReg( F10 ),
165 RegisterSaver_LiveFloatReg( F11 ),
166 RegisterSaver_LiveFloatReg( F12 ),
167 RegisterSaver_LiveFloatReg( F13 ),
168 RegisterSaver_LiveFloatReg( F14 ),
169 RegisterSaver_LiveFloatReg( F15 ),
170 RegisterSaver_LiveFloatReg( F16 ),
171 RegisterSaver_LiveFloatReg( F17 ),
172 RegisterSaver_LiveFloatReg( F18 ),
173 RegisterSaver_LiveFloatReg( F19 ),
174 RegisterSaver_LiveFloatReg( F20 ),
175 RegisterSaver_LiveFloatReg( F21 ),
176 RegisterSaver_LiveFloatReg( F22 ),
177 RegisterSaver_LiveFloatReg( F23 ),
178 RegisterSaver_LiveFloatReg( F24 ),
179 RegisterSaver_LiveFloatReg( F25 ),
180 RegisterSaver_LiveFloatReg( F26 ),
181 RegisterSaver_LiveFloatReg( F27 ),
182 RegisterSaver_LiveFloatReg( F28 ),
183 RegisterSaver_LiveFloatReg( F29 ),
184 RegisterSaver_LiveFloatReg( F30 ),
185 RegisterSaver_LiveFloatReg( F31 ),
186 //
187 // live integer registers:
188 //
189 RegisterSaver_LiveIntReg( R0 ),
190 //RegisterSaver_LiveIntReg( R1 ), // stack pointer
191 RegisterSaver_LiveIntReg( R2 ),
192 RegisterSaver_LiveIntReg( R3 ),
193 RegisterSaver_LiveIntReg( R4 ),
194 RegisterSaver_LiveIntReg( R5 ),
195 RegisterSaver_LiveIntReg( R6 ),
196 RegisterSaver_LiveIntReg( R7 ),
197 RegisterSaver_LiveIntReg( R8 ),
198 RegisterSaver_LiveIntReg( R9 ),
199 RegisterSaver_LiveIntReg( R10 ),
200 RegisterSaver_LiveIntReg( R11 ),
201 RegisterSaver_LiveIntReg( R12 ),
202 //RegisterSaver_LiveIntReg( R13 ), // system thread id
203 RegisterSaver_LiveIntReg( R14 ),
204 RegisterSaver_LiveIntReg( R15 ),
205 RegisterSaver_LiveIntReg( R16 ),
206 RegisterSaver_LiveIntReg( R17 ),
207 RegisterSaver_LiveIntReg( R18 ),
208 RegisterSaver_LiveIntReg( R19 ),
209 RegisterSaver_LiveIntReg( R20 ),
210 RegisterSaver_LiveIntReg( R21 ),
211 RegisterSaver_LiveIntReg( R22 ),
212 RegisterSaver_LiveIntReg( R23 ),
213 RegisterSaver_LiveIntReg( R24 ),
214 RegisterSaver_LiveIntReg( R25 ),
215 RegisterSaver_LiveIntReg( R26 ),
216 RegisterSaver_LiveIntReg( R27 ),
217 RegisterSaver_LiveIntReg( R28 ),
218 RegisterSaver_LiveIntReg( R29 ),
219 RegisterSaver_LiveIntReg( R30 ),
220 RegisterSaver_LiveIntReg( R31 ) // must be the last register (see save/restore functions below)
221 };
222
223 static const RegisterSaver::LiveRegType RegisterSaver_LiveVecRegs[] = {
224 //
225 // live vector registers (optional, only these ones are used by C2):
226 //
227 RegisterSaver_LiveVecReg( VR0 ),
228 RegisterSaver_LiveVecReg( VR1 ),
229 RegisterSaver_LiveVecReg( VR2 ),
230 RegisterSaver_LiveVecReg( VR3 ),
231 RegisterSaver_LiveVecReg( VR4 ),
232 RegisterSaver_LiveVecReg( VR5 ),
233 RegisterSaver_LiveVecReg( VR6 ),
234 RegisterSaver_LiveVecReg( VR7 ),
235 RegisterSaver_LiveVecReg( VR8 ),
236 RegisterSaver_LiveVecReg( VR9 ),
237 RegisterSaver_LiveVecReg( VR10 ),
238 RegisterSaver_LiveVecReg( VR11 ),
239 RegisterSaver_LiveVecReg( VR12 ),
240 RegisterSaver_LiveVecReg( VR13 ),
241 RegisterSaver_LiveVecReg( VR14 ),
242 RegisterSaver_LiveVecReg( VR15 ),
243 RegisterSaver_LiveVecReg( VR16 ),
244 RegisterSaver_LiveVecReg( VR17 ),
245 RegisterSaver_LiveVecReg( VR18 ),
246 RegisterSaver_LiveVecReg( VR19 ),
247 RegisterSaver_LiveVecReg( VR20 ),
248 RegisterSaver_LiveVecReg( VR21 ),
249 RegisterSaver_LiveVecReg( VR22 ),
250 RegisterSaver_LiveVecReg( VR23 ),
251 RegisterSaver_LiveVecReg( VR24 ),
252 RegisterSaver_LiveVecReg( VR25 ),
253 RegisterSaver_LiveVecReg( VR26 ),
254 RegisterSaver_LiveVecReg( VR27 ),
255 RegisterSaver_LiveVecReg( VR28 ),
256 RegisterSaver_LiveVecReg( VR29 ),
257 RegisterSaver_LiveVecReg( VR30 ),
258 RegisterSaver_LiveVecReg( VR31 )
259 };
260
261
262 OopMap* RegisterSaver::push_frame_reg_args_and_save_live_registers(MacroAssembler* masm,
263 int* out_frame_size_in_bytes,
264 bool generate_oop_map,
265 int return_pc_adjustment,
266 ReturnPCLocation return_pc_location,
267 bool save_vectors) {
268 // Push an abi_reg_args-frame and store all registers which may be live.
269 // If requested, create an OopMap: Record volatile registers as
270 // callee-save values in an OopMap so their save locations will be
271 // propagated to the RegisterMap of the caller frame during
272 // StackFrameStream construction (needed for deoptimization; see
273 // compiledVFrame::create_stack_value).
274 // If return_pc_adjustment != 0 adjust the return pc by return_pc_adjustment.
275 // Updated return pc is returned in R31 (if not return_pc_is_pre_saved).
276
277 // calculate frame size
278 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
279 sizeof(RegisterSaver::LiveRegType);
280 const int vecregstosave_num = save_vectors ? (sizeof(RegisterSaver_LiveVecRegs) /
281 sizeof(RegisterSaver::LiveRegType))
282 : 0;
283 const int register_save_size = regstosave_num * reg_size + vecregstosave_num * vec_reg_size;
284 const int frame_size_in_bytes = align_up(register_save_size, frame::alignment_in_bytes)
285 + frame::native_abi_reg_args_size;
286
287 *out_frame_size_in_bytes = frame_size_in_bytes;
288 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint);
289 const int register_save_offset = frame_size_in_bytes - register_save_size;
290
291 // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words.
292 OopMap* map = generate_oop_map ? new OopMap(frame_size_in_slots, 0) : nullptr;
293
294 BLOCK_COMMENT("push_frame_reg_args_and_save_live_registers {");
295
296 // push a new frame
297 __ push_frame(frame_size_in_bytes, noreg);
298
299 // Save some registers in the last (non-vector) slots of the new frame so we
300 // can use them as scratch regs or to determine the return pc.
301 __ std(R31, frame_size_in_bytes - reg_size - vecregstosave_num * vec_reg_size, R1_SP);
302 __ std(R30, frame_size_in_bytes - 2*reg_size - vecregstosave_num * vec_reg_size, R1_SP);
303
304 // save the flags
305 // Do the save_LR by hand and adjust the return pc if requested.
306 switch (return_pc_location) {
307 case return_pc_is_lr: __ mflr(R31); break;
308 case return_pc_is_pre_saved: assert(return_pc_adjustment == 0, "unsupported"); break;
309 case return_pc_is_thread_saved_exception_pc: __ ld(R31, thread_(saved_exception_pc)); break;
310 default: ShouldNotReachHere();
311 }
312 if (return_pc_location != return_pc_is_pre_saved) {
313 if (return_pc_adjustment != 0) {
314 __ addi(R31, R31, return_pc_adjustment);
315 }
316 __ std(R31, frame_size_in_bytes + _abi0(lr), R1_SP);
317 }
318
319 // save all registers (ints and floats)
320 int offset = register_save_offset;
321
322 for (int i = 0; i < regstosave_num; i++) {
323 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
324 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
325
326 switch (reg_type) {
327 case RegisterSaver::int_reg: {
328 if (reg_num < 30) { // We spilled R30-31 right at the beginning.
329 __ std(as_Register(reg_num), offset, R1_SP);
330 }
331 break;
332 }
333 case RegisterSaver::float_reg: {
334 __ stfd(as_FloatRegister(reg_num), offset, R1_SP);
335 break;
336 }
337 case RegisterSaver::special_reg: {
338 if (reg_num == SR_CTR.encoding()) {
339 __ mfctr(R30);
340 __ std(R30, offset, R1_SP);
341 } else {
342 Unimplemented();
343 }
344 break;
345 }
346 default:
347 ShouldNotReachHere();
348 }
349
350 if (generate_oop_map) {
351 map->set_callee_saved(VMRegImpl::stack2reg(offset >> 2),
352 RegisterSaver_LiveRegs[i].vmreg);
353 }
354 offset += reg_size;
355 }
356
357 // Note that generate_oop_map in the following loop is only used for the
358 // polling_page_vectors_safepoint_handler_blob.
359 // The order in which the vector contents are stored depends on Endianess and
360 // the utilized instructions (PowerArchitecturePPC64).
361 assert(is_aligned(offset, StackAlignmentInBytes), "should be");
362 if (PowerArchitecturePPC64 >= 10) {
363 assert(is_even(vecregstosave_num), "expectation");
364 for (int i = 0; i < vecregstosave_num; i += 2) {
365 int reg_num = RegisterSaver_LiveVecRegs[i].reg_num;
366 assert(RegisterSaver_LiveVecRegs[i + 1].reg_num == reg_num + 1, "or use other instructions!");
367
368 __ stxvp(as_VectorRegister(reg_num).to_vsr(), offset, R1_SP);
369 // Note: The contents were read in the same order (see loadV16_Power9 node in ppc.ad).
370 if (generate_oop_map) {
371 map->set_callee_saved(VMRegImpl::stack2reg(offset >> 2),
372 RegisterSaver_LiveVecRegs[i LITTLE_ENDIAN_ONLY(+1) ].vmreg);
373 map->set_callee_saved(VMRegImpl::stack2reg((offset + vec_reg_size) >> 2),
374 RegisterSaver_LiveVecRegs[i BIG_ENDIAN_ONLY(+1) ].vmreg);
375 }
376 offset += (2 * vec_reg_size);
377 }
378 } else {
379 for (int i = 0; i < vecregstosave_num; i++) {
380 int reg_num = RegisterSaver_LiveVecRegs[i].reg_num;
381
382 if (PowerArchitecturePPC64 >= 9) {
383 __ stxv(as_VectorRegister(reg_num)->to_vsr(), offset, R1_SP);
384 } else {
385 __ li(R31, offset);
386 __ stxvd2x(as_VectorRegister(reg_num)->to_vsr(), R31, R1_SP);
387 }
388 // Note: The contents were read in the same order (see loadV16_Power8 / loadV16_Power9 node in ppc.ad).
389 if (generate_oop_map) {
390 VMReg vsr = RegisterSaver_LiveVecRegs[i].vmreg;
391 map->set_callee_saved(VMRegImpl::stack2reg(offset >> 2), vsr);
392 }
393 offset += vec_reg_size;
394 }
395 }
396
397 assert(offset == frame_size_in_bytes, "consistency check");
398
399 BLOCK_COMMENT("} push_frame_reg_args_and_save_live_registers");
400
401 // And we're done.
402 return map;
403 }
404
405
406 // Pop the current frame and restore all the registers that we
407 // saved.
408 void RegisterSaver::restore_live_registers_and_pop_frame(MacroAssembler* masm,
409 int frame_size_in_bytes,
410 bool restore_ctr,
411 bool save_vectors) {
412 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
413 sizeof(RegisterSaver::LiveRegType);
414 const int vecregstosave_num = save_vectors ? (sizeof(RegisterSaver_LiveVecRegs) /
415 sizeof(RegisterSaver::LiveRegType))
416 : 0;
417 const int register_save_size = regstosave_num * reg_size + vecregstosave_num * vec_reg_size;
418
419 const int register_save_offset = frame_size_in_bytes - register_save_size;
420
421 BLOCK_COMMENT("restore_live_registers_and_pop_frame {");
422
423 // restore all registers (ints and floats)
424 int offset = register_save_offset;
425
426 for (int i = 0; i < regstosave_num; i++) {
427 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
428 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
429
430 switch (reg_type) {
431 case RegisterSaver::int_reg: {
432 if (reg_num != 31) // R31 restored at the end, it's the tmp reg!
433 __ ld(as_Register(reg_num), offset, R1_SP);
434 break;
435 }
436 case RegisterSaver::float_reg: {
437 __ lfd(as_FloatRegister(reg_num), offset, R1_SP);
438 break;
439 }
440 case RegisterSaver::special_reg: {
441 if (reg_num == SR_CTR.encoding()) {
442 if (restore_ctr) { // Nothing to do here if ctr already contains the next address.
443 __ ld(R31, offset, R1_SP);
444 __ mtctr(R31);
445 }
446 } else {
447 Unimplemented();
448 }
449 break;
450 }
451 default:
452 ShouldNotReachHere();
453 }
454 offset += reg_size;
455 }
456
457 assert(is_aligned(offset, StackAlignmentInBytes), "should be");
458 if (PowerArchitecturePPC64 >= 10) {
459 for (int i = 0; i < vecregstosave_num; i += 2) {
460 int reg_num = RegisterSaver_LiveVecRegs[i].reg_num;
461 assert(RegisterSaver_LiveVecRegs[i + 1].reg_num == reg_num + 1, "or use other instructions!");
462
463 __ lxvp(as_VectorRegister(reg_num).to_vsr(), offset, R1_SP);
464
465 offset += (2 * vec_reg_size);
466 }
467 } else {
468 for (int i = 0; i < vecregstosave_num; i++) {
469 int reg_num = RegisterSaver_LiveVecRegs[i].reg_num;
470
471 if (PowerArchitecturePPC64 >= 9) {
472 __ lxv(as_VectorRegister(reg_num).to_vsr(), offset, R1_SP);
473 } else {
474 __ li(R31, offset);
475 __ lxvd2x(as_VectorRegister(reg_num).to_vsr(), R31, R1_SP);
476 }
477
478 offset += vec_reg_size;
479 }
480 }
481
482 assert(offset == frame_size_in_bytes, "consistency check");
483
484 // restore link and the flags
485 __ ld(R31, frame_size_in_bytes + _abi0(lr), R1_SP);
486 __ mtlr(R31);
487
488 // restore scratch register's value
489 __ ld(R31, frame_size_in_bytes - reg_size - vecregstosave_num * vec_reg_size, R1_SP);
490
491 // pop the frame
492 __ addi(R1_SP, R1_SP, frame_size_in_bytes);
493
494 BLOCK_COMMENT("} restore_live_registers_and_pop_frame");
495 }
496
497 void RegisterSaver::push_frame_and_save_argument_registers(MacroAssembler* masm, Register r_temp,
498 int frame_size,int total_args, const VMRegPair *regs,
499 const VMRegPair *regs2) {
500 __ push_frame(frame_size, r_temp);
501 int st_off = frame_size - wordSize;
502 for (int i = 0; i < total_args; i++) {
503 VMReg r_1 = regs[i].first();
504 VMReg r_2 = regs[i].second();
505 if (!r_1->is_valid()) {
506 assert(!r_2->is_valid(), "");
507 continue;
508 }
509 if (r_1->is_Register()) {
510 Register r = r_1->as_Register();
511 __ std(r, st_off, R1_SP);
512 st_off -= wordSize;
513 } else if (r_1->is_FloatRegister()) {
514 FloatRegister f = r_1->as_FloatRegister();
515 __ stfd(f, st_off, R1_SP);
516 st_off -= wordSize;
517 }
518 }
519 if (regs2 != nullptr) {
520 for (int i = 0; i < total_args; i++) {
521 VMReg r_1 = regs2[i].first();
522 VMReg r_2 = regs2[i].second();
523 if (!r_1->is_valid()) {
524 assert(!r_2->is_valid(), "");
525 continue;
526 }
527 if (r_1->is_Register()) {
528 Register r = r_1->as_Register();
529 __ std(r, st_off, R1_SP);
530 st_off -= wordSize;
531 } else if (r_1->is_FloatRegister()) {
532 FloatRegister f = r_1->as_FloatRegister();
533 __ stfd(f, st_off, R1_SP);
534 st_off -= wordSize;
535 }
536 }
537 }
538 }
539
540 void RegisterSaver::restore_argument_registers_and_pop_frame(MacroAssembler*masm, int frame_size,
541 int total_args, const VMRegPair *regs,
542 const VMRegPair *regs2) {
543 int st_off = frame_size - wordSize;
544 for (int i = 0; i < total_args; i++) {
545 VMReg r_1 = regs[i].first();
546 VMReg r_2 = regs[i].second();
547 if (r_1->is_Register()) {
548 Register r = r_1->as_Register();
549 __ ld(r, st_off, R1_SP);
550 st_off -= wordSize;
551 } else if (r_1->is_FloatRegister()) {
552 FloatRegister f = r_1->as_FloatRegister();
553 __ lfd(f, st_off, R1_SP);
554 st_off -= wordSize;
555 }
556 }
557 if (regs2 != nullptr)
558 for (int i = 0; i < total_args; i++) {
559 VMReg r_1 = regs2[i].first();
560 VMReg r_2 = regs2[i].second();
561 if (r_1->is_Register()) {
562 Register r = r_1->as_Register();
563 __ ld(r, st_off, R1_SP);
564 st_off -= wordSize;
565 } else if (r_1->is_FloatRegister()) {
566 FloatRegister f = r_1->as_FloatRegister();
567 __ lfd(f, st_off, R1_SP);
568 st_off -= wordSize;
569 }
570 }
571 __ pop_frame();
572 }
573
574 // Restore the registers that might be holding a result.
575 void RegisterSaver::restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes) {
576 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
577 sizeof(RegisterSaver::LiveRegType);
578 const int register_save_size = regstosave_num * reg_size; // VS registers not relevant here.
579 const int register_save_offset = frame_size_in_bytes - register_save_size;
580
581 // restore all result registers (ints and floats)
582 int offset = register_save_offset;
583 for (int i = 0; i < regstosave_num; i++) {
584 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
585 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
586 switch (reg_type) {
587 case RegisterSaver::int_reg: {
588 if (as_Register(reg_num)==R3_RET) // int result_reg
589 __ ld(as_Register(reg_num), offset, R1_SP);
590 break;
591 }
592 case RegisterSaver::float_reg: {
593 if (as_FloatRegister(reg_num)==F1_RET) // float result_reg
594 __ lfd(as_FloatRegister(reg_num), offset, R1_SP);
595 break;
596 }
597 case RegisterSaver::special_reg: {
598 // Special registers don't hold a result.
599 break;
600 }
601 default:
602 ShouldNotReachHere();
603 }
604 offset += reg_size;
605 }
606
607 assert(offset == frame_size_in_bytes, "consistency check");
608 }
609
610 // Is vector's size (in bytes) bigger than a size saved by default?
611 bool SharedRuntime::is_wide_vector(int size) {
612 // Note, MaxVectorSize == 8/16 on PPC64.
613 assert(size <= (SuperwordUseVSX ? 16 : 8), "%d bytes vectors are not supported", size);
614 return size > 8;
615 }
616
617 static int reg2slot(VMReg r) {
618 return r->reg2stack() + SharedRuntime::out_preserve_stack_slots();
619 }
620
621 static int reg2offset(VMReg r) {
622 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
623 }
624
625 // ---------------------------------------------------------------------------
626 // Read the array of BasicTypes from a signature, and compute where the
627 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
628 // quantities. Values less than VMRegImpl::stack0 are registers, those above
629 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
630 // as framesizes are fixed.
631 // VMRegImpl::stack0 refers to the first slot 0(sp).
632 // and VMRegImpl::stack0+1 refers to the memory word 4-bytes higher. Register
633 // up to Register::number_of_registers) are the 64-bit
634 // integer registers.
635
636 // Note: the INPUTS in sig_bt are in units of Java argument words, which are
637 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit
638 // units regardless of build. Of course for i486 there is no 64 bit build
639
640 // The Java calling convention is a "shifted" version of the C ABI.
641 // By skipping the first C ABI register we can call non-static jni methods
642 // with small numbers of arguments without having to shuffle the arguments
643 // at all. Since we control the java ABI we ought to at least get some
644 // advantage out of it.
645
646 const VMReg java_iarg_reg[8] = {
647 R3->as_VMReg(),
648 R4->as_VMReg(),
649 R5->as_VMReg(),
650 R6->as_VMReg(),
651 R7->as_VMReg(),
652 R8->as_VMReg(),
653 R9->as_VMReg(),
654 R10->as_VMReg()
655 };
656
657 const VMReg java_farg_reg[13] = {
658 F1->as_VMReg(),
659 F2->as_VMReg(),
660 F3->as_VMReg(),
661 F4->as_VMReg(),
662 F5->as_VMReg(),
663 F6->as_VMReg(),
664 F7->as_VMReg(),
665 F8->as_VMReg(),
666 F9->as_VMReg(),
667 F10->as_VMReg(),
668 F11->as_VMReg(),
669 F12->as_VMReg(),
670 F13->as_VMReg()
671 };
672
673 const int num_java_iarg_registers = sizeof(java_iarg_reg) / sizeof(java_iarg_reg[0]);
674 const int num_java_farg_registers = sizeof(java_farg_reg) / sizeof(java_farg_reg[0]);
675
676 STATIC_ASSERT(num_java_iarg_registers == Argument::n_int_register_parameters_j);
677 STATIC_ASSERT(num_java_farg_registers == Argument::n_float_register_parameters_j);
678
679 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
680 VMRegPair *regs,
681 int total_args_passed) {
682 // C2c calling conventions for compiled-compiled calls.
683 // Put 8 ints/longs into registers _AND_ 13 float/doubles into
684 // registers _AND_ put the rest on the stack.
685
686 const int inc_stk_for_intfloat = 1; // 1 slots for ints and floats
687 const int inc_stk_for_longdouble = 2; // 2 slots for longs and doubles
688
689 int i;
690 VMReg reg;
691 int stk = 0;
692 int ireg = 0;
693 int freg = 0;
694
695 // We put the first 8 arguments into registers and the rest on the
696 // stack, float arguments are already in their argument registers
697 // due to c2c calling conventions (see calling_convention).
698 for (int i = 0; i < total_args_passed; ++i) {
699 switch(sig_bt[i]) {
700 case T_BOOLEAN:
701 case T_CHAR:
702 case T_BYTE:
703 case T_SHORT:
704 case T_INT:
705 if (ireg < num_java_iarg_registers) {
706 // Put int/ptr in register
707 reg = java_iarg_reg[ireg];
708 ++ireg;
709 } else {
710 // Put int/ptr on stack.
711 reg = VMRegImpl::stack2reg(stk);
712 stk += inc_stk_for_intfloat;
713 }
714 regs[i].set1(reg);
715 break;
716 case T_LONG:
717 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
718 if (ireg < num_java_iarg_registers) {
719 // Put long in register.
720 reg = java_iarg_reg[ireg];
721 ++ireg;
722 } else {
723 // Put long on stack. They must be aligned to 2 slots.
724 if (stk & 0x1) ++stk;
725 reg = VMRegImpl::stack2reg(stk);
726 stk += inc_stk_for_longdouble;
727 }
728 regs[i].set2(reg);
729 break;
730 case T_OBJECT:
731 case T_ARRAY:
732 case T_ADDRESS:
733 if (ireg < num_java_iarg_registers) {
734 // Put ptr in register.
735 reg = java_iarg_reg[ireg];
736 ++ireg;
737 } else {
738 // Put ptr on stack. Objects must be aligned to 2 slots too,
739 // because "64-bit pointers record oop-ishness on 2 aligned
740 // adjacent registers." (see OopFlow::build_oop_map).
741 if (stk & 0x1) ++stk;
742 reg = VMRegImpl::stack2reg(stk);
743 stk += inc_stk_for_longdouble;
744 }
745 regs[i].set2(reg);
746 break;
747 case T_FLOAT:
748 if (freg < num_java_farg_registers) {
749 // Put float in register.
750 reg = java_farg_reg[freg];
751 ++freg;
752 } else {
753 // Put float on stack.
754 reg = VMRegImpl::stack2reg(stk);
755 stk += inc_stk_for_intfloat;
756 }
757 regs[i].set1(reg);
758 break;
759 case T_DOUBLE:
760 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
761 if (freg < num_java_farg_registers) {
762 // Put double in register.
763 reg = java_farg_reg[freg];
764 ++freg;
765 } else {
766 // Put double on stack. They must be aligned to 2 slots.
767 if (stk & 0x1) ++stk;
768 reg = VMRegImpl::stack2reg(stk);
769 stk += inc_stk_for_longdouble;
770 }
771 regs[i].set2(reg);
772 break;
773 case T_VOID:
774 // Do not count halves.
775 regs[i].set_bad();
776 break;
777 default:
778 ShouldNotReachHere();
779 }
780 }
781 return stk;
782 }
783
784 #if defined(COMPILER1) || defined(COMPILER2)
785 // Calling convention for calling C code.
786 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
787 VMRegPair *regs,
788 int total_args_passed) {
789 // Calling conventions for C runtime calls and calls to JNI native methods.
790 //
791 // PPC64 convention: Hoist the first 8 int/ptr/long's in the first 8
792 // int regs, leaving int regs undefined if the arg is flt/dbl. Hoist
793 // the first 13 flt/dbl's in the first 13 fp regs but additionally
794 // copy flt/dbl to the stack if they are beyond the 8th argument.
795
796 const VMReg iarg_reg[8] = {
797 R3->as_VMReg(),
798 R4->as_VMReg(),
799 R5->as_VMReg(),
800 R6->as_VMReg(),
801 R7->as_VMReg(),
802 R8->as_VMReg(),
803 R9->as_VMReg(),
804 R10->as_VMReg()
805 };
806
807 const VMReg farg_reg[13] = {
808 F1->as_VMReg(),
809 F2->as_VMReg(),
810 F3->as_VMReg(),
811 F4->as_VMReg(),
812 F5->as_VMReg(),
813 F6->as_VMReg(),
814 F7->as_VMReg(),
815 F8->as_VMReg(),
816 F9->as_VMReg(),
817 F10->as_VMReg(),
818 F11->as_VMReg(),
819 F12->as_VMReg(),
820 F13->as_VMReg()
821 };
822
823 // Check calling conventions consistency.
824 assert(sizeof(iarg_reg) / sizeof(iarg_reg[0]) == Argument::n_int_register_parameters_c &&
825 sizeof(farg_reg) / sizeof(farg_reg[0]) == Argument::n_float_register_parameters_c,
826 "consistency");
827
828 const int additional_frame_header_slots = ((frame::native_abi_minframe_size - frame::jit_out_preserve_size)
829 / VMRegImpl::stack_slot_size);
830 const int float_offset_in_slots = Argument::float_on_stack_offset_in_bytes_c / VMRegImpl::stack_slot_size;
831
832 VMReg reg;
833 int arg = 0;
834 int freg = 0;
835 bool stack_used = false;
836
837 for (int i = 0; i < total_args_passed; ++i, ++arg) {
838 // Each argument corresponds to a slot in the Parameter Save Area (if not omitted)
839 int stk = (arg * 2) + additional_frame_header_slots;
840
841 switch(sig_bt[i]) {
842 //
843 // If arguments 0-7 are integers, they are passed in integer registers.
844 // Argument i is placed in iarg_reg[i].
845 //
846 case T_BOOLEAN:
847 case T_CHAR:
848 case T_BYTE:
849 case T_SHORT:
850 case T_INT:
851 // We must cast ints to longs and use full 64 bit stack slots
852 // here. Thus fall through, handle as long.
853 case T_LONG:
854 case T_OBJECT:
855 case T_ARRAY:
856 case T_ADDRESS:
857 case T_METADATA:
858 // Oops are already boxed if required (JNI).
859 if (arg < Argument::n_int_register_parameters_c) {
860 reg = iarg_reg[arg];
861 } else {
862 reg = VMRegImpl::stack2reg(stk);
863 stack_used = true;
864 }
865 regs[i].set2(reg);
866 break;
867
868 //
869 // Floats are treated differently from int regs: The first 13 float arguments
870 // are passed in registers (not the float args among the first 13 args).
871 // Thus argument i is NOT passed in farg_reg[i] if it is float. It is passed
872 // in farg_reg[j] if argument i is the j-th float argument of this call.
873 //
874 case T_FLOAT:
875 if (freg < Argument::n_float_register_parameters_c) {
876 // Put float in register ...
877 reg = farg_reg[freg];
878 ++freg;
879 } else {
880 // Put float on stack.
881 reg = VMRegImpl::stack2reg(stk + float_offset_in_slots);
882 stack_used = true;
883 }
884 regs[i].set1(reg);
885 break;
886 case T_DOUBLE:
887 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
888 if (freg < Argument::n_float_register_parameters_c) {
889 // Put double in register ...
890 reg = farg_reg[freg];
891 ++freg;
892 } else {
893 // Put double on stack.
894 reg = VMRegImpl::stack2reg(stk);
895 stack_used = true;
896 }
897 regs[i].set2(reg);
898 break;
899
900 case T_VOID:
901 // Do not count halves.
902 regs[i].set_bad();
903 --arg;
904 break;
905 default:
906 ShouldNotReachHere();
907 }
908 }
909
910 // Return size of the stack frame excluding the jit_out_preserve part in single-word slots.
911 #if defined(ABI_ELFv2)
912 assert(additional_frame_header_slots == 0, "ABIv2 shouldn't use extra slots");
913 // ABIv2 allows omitting the Parameter Save Area if the callee's prototype
914 // indicates that all parameters can be passed in registers.
915 return stack_used ? (arg * 2) : 0;
916 #else
917 // The Parameter Save Area needs to be at least 8 double-word slots for ABIv1.
918 // We have to add extra slots because ABIv1 uses a larger header.
919 return MAX2(arg, 8) * 2 + additional_frame_header_slots;
920 #endif
921 }
922 #endif // COMPILER2
923
924 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
925 uint num_bits,
926 uint total_args_passed) {
927 Unimplemented();
928 return 0;
929 }
930
931 static address gen_c2i_adapter(MacroAssembler *masm,
932 int total_args_passed,
933 int comp_args_on_stack,
934 const BasicType *sig_bt,
935 const VMRegPair *regs,
936 Label& call_interpreter,
937 const Register& ientry) {
938
939 address c2i_entrypoint;
940
941 const Register sender_SP = R21_sender_SP; // == R21_tmp1
942 const Register code = R22_tmp2;
943 //const Register ientry = R23_tmp3;
944 const Register value_regs[] = { R24_tmp4, R25_tmp5, R26_tmp6 };
945 const int num_value_regs = sizeof(value_regs) / sizeof(Register);
946 int value_regs_index = 0;
947
948 const Register return_pc = R27_tmp7;
949 const Register tmp = R28_tmp8;
950
951 assert_different_registers(sender_SP, code, ientry, return_pc, tmp);
952
953 // Adapter needs TOP_IJAVA_FRAME_ABI.
954 const int adapter_size = frame::top_ijava_frame_abi_size +
955 align_up(total_args_passed * wordSize, frame::alignment_in_bytes);
956
957 // regular (verified) c2i entry point
958 c2i_entrypoint = __ pc();
959
960 // Does compiled code exists? If yes, patch the caller's callsite.
961 __ ld(code, method_(code));
962 __ cmpdi(CR0, code, 0);
963 __ ld(ientry, method_(interpreter_entry)); // preloaded
964 __ beq(CR0, call_interpreter);
965
966
967 // Patch caller's callsite, method_(code) was not null which means that
968 // compiled code exists.
969 __ mflr(return_pc);
970 __ std(return_pc, _abi0(lr), R1_SP);
971 RegisterSaver::push_frame_and_save_argument_registers(masm, tmp, adapter_size, total_args_passed, regs);
972
973 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), R19_method, return_pc);
974
975 RegisterSaver::restore_argument_registers_and_pop_frame(masm, adapter_size, total_args_passed, regs);
976 __ ld(return_pc, _abi0(lr), R1_SP);
977 __ ld(ientry, method_(interpreter_entry)); // preloaded
978 __ mtlr(return_pc);
979
980
981 // Call the interpreter.
982 __ BIND(call_interpreter);
983 __ mtctr(ientry);
984
985 // Get a copy of the current SP for loading caller's arguments.
986 __ mr(sender_SP, R1_SP);
987
988 // Add space for the adapter.
989 __ resize_frame(-adapter_size, R12_scratch2);
990
991 int st_off = adapter_size - wordSize;
992
993 // Write the args into the outgoing interpreter space.
994 for (int i = 0; i < total_args_passed; i++) {
995 VMReg r_1 = regs[i].first();
996 VMReg r_2 = regs[i].second();
997 if (!r_1->is_valid()) {
998 assert(!r_2->is_valid(), "");
999 continue;
1000 }
1001 if (r_1->is_stack()) {
1002 Register tmp_reg = value_regs[value_regs_index];
1003 value_regs_index = (value_regs_index + 1) % num_value_regs;
1004 // The calling convention produces OptoRegs that ignore the out
1005 // preserve area (JIT's ABI). We must account for it here.
1006 int ld_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
1007 if (!r_2->is_valid()) {
1008 __ lwz(tmp_reg, ld_off, sender_SP);
1009 } else {
1010 __ ld(tmp_reg, ld_off, sender_SP);
1011 }
1012 // Pretend stack targets were loaded into tmp_reg.
1013 r_1 = tmp_reg->as_VMReg();
1014 }
1015
1016 if (r_1->is_Register()) {
1017 Register r = r_1->as_Register();
1018 if (!r_2->is_valid()) {
1019 __ stw(r, st_off, R1_SP);
1020 st_off-=wordSize;
1021 } else {
1022 // Longs are given 2 64-bit slots in the interpreter, but the
1023 // data is passed in only 1 slot.
1024 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
1025 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); )
1026 st_off-=wordSize;
1027 }
1028 __ std(r, st_off, R1_SP);
1029 st_off-=wordSize;
1030 }
1031 } else {
1032 assert(r_1->is_FloatRegister(), "");
1033 FloatRegister f = r_1->as_FloatRegister();
1034 if (!r_2->is_valid()) {
1035 __ stfs(f, st_off, R1_SP);
1036 st_off-=wordSize;
1037 } else {
1038 // In 64bit, doubles are given 2 64-bit slots in the interpreter, but the
1039 // data is passed in only 1 slot.
1040 // One of these should get known junk...
1041 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); )
1042 st_off-=wordSize;
1043 __ stfd(f, st_off, R1_SP);
1044 st_off-=wordSize;
1045 }
1046 }
1047 }
1048
1049 // Jump to the interpreter just as if interpreter was doing it.
1050
1051 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
1052
1053 // load TOS
1054 __ addi(R15_esp, R1_SP, st_off);
1055
1056 // Frame_manager expects initial_caller_sp (= SP without resize by c2i) in R21_tmp1.
1057 assert(sender_SP == R21_sender_SP, "passing initial caller's SP in wrong register");
1058 __ bctr();
1059
1060 return c2i_entrypoint;
1061 }
1062
1063 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
1064 int total_args_passed,
1065 int comp_args_on_stack,
1066 const BasicType *sig_bt,
1067 const VMRegPair *regs) {
1068
1069 // Load method's entry-point from method.
1070 __ ld(R12_scratch2, in_bytes(Method::from_compiled_offset()), R19_method);
1071 __ mtctr(R12_scratch2);
1072
1073 // We will only enter here from an interpreted frame and never from after
1074 // passing thru a c2i. Azul allowed this but we do not. If we lose the
1075 // race and use a c2i we will remain interpreted for the race loser(s).
1076 // This removes all sorts of headaches on the x86 side and also eliminates
1077 // the possibility of having c2i -> i2c -> c2i -> ... endless transitions.
1078
1079 // Note: r13 contains the senderSP on entry. We must preserve it since
1080 // we may do a i2c -> c2i transition if we lose a race where compiled
1081 // code goes non-entrant while we get args ready.
1082 // In addition we use r13 to locate all the interpreter args as
1083 // we must align the stack to 16 bytes on an i2c entry else we
1084 // lose alignment we expect in all compiled code and register
1085 // save code can segv when fxsave instructions find improperly
1086 // aligned stack pointer.
1087
1088 const Register ld_ptr = R15_esp;
1089 const Register value_regs[] = { R22_tmp2, R23_tmp3, R24_tmp4, R25_tmp5, R26_tmp6 };
1090 const int num_value_regs = sizeof(value_regs) / sizeof(Register);
1091 int value_regs_index = 0;
1092
1093 int ld_offset = total_args_passed*wordSize;
1094
1095 // Cut-out for having no stack args. Since up to 2 int/oop args are passed
1096 // in registers, we will occasionally have no stack args.
1097 int comp_words_on_stack = 0;
1098 if (comp_args_on_stack) {
1099 // Sig words on the stack are greater-than VMRegImpl::stack0. Those in
1100 // registers are below. By subtracting stack0, we either get a negative
1101 // number (all values in registers) or the maximum stack slot accessed.
1102
1103 // Convert 4-byte c2 stack slots to words.
1104 comp_words_on_stack = align_up(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
1105 // Round up to miminum stack alignment, in wordSize.
1106 comp_words_on_stack = align_up(comp_words_on_stack, 2);
1107 __ resize_frame(-comp_words_on_stack * wordSize, R11_scratch1);
1108 }
1109
1110 // Now generate the shuffle code. Pick up all register args and move the
1111 // rest through register value=Z_R12.
1112 BLOCK_COMMENT("Shuffle arguments");
1113 for (int i = 0; i < total_args_passed; i++) {
1114 if (sig_bt[i] == T_VOID) {
1115 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
1116 continue;
1117 }
1118
1119 // Pick up 0, 1 or 2 words from ld_ptr.
1120 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
1121 "scrambled load targets?");
1122 VMReg r_1 = regs[i].first();
1123 VMReg r_2 = regs[i].second();
1124 if (!r_1->is_valid()) {
1125 assert(!r_2->is_valid(), "");
1126 continue;
1127 }
1128 if (r_1->is_FloatRegister()) {
1129 if (!r_2->is_valid()) {
1130 __ lfs(r_1->as_FloatRegister(), ld_offset, ld_ptr);
1131 ld_offset-=wordSize;
1132 } else {
1133 // Skip the unused interpreter slot.
1134 __ lfd(r_1->as_FloatRegister(), ld_offset-wordSize, ld_ptr);
1135 ld_offset-=2*wordSize;
1136 }
1137 } else {
1138 Register r;
1139 if (r_1->is_stack()) {
1140 // Must do a memory to memory move thru "value".
1141 r = value_regs[value_regs_index];
1142 value_regs_index = (value_regs_index + 1) % num_value_regs;
1143 } else {
1144 r = r_1->as_Register();
1145 }
1146 if (!r_2->is_valid()) {
1147 // Not sure we need to do this but it shouldn't hurt.
1148 if (is_reference_type(sig_bt[i]) || sig_bt[i] == T_ADDRESS) {
1149 __ ld(r, ld_offset, ld_ptr);
1150 ld_offset-=wordSize;
1151 } else {
1152 __ lwz(r, ld_offset, ld_ptr);
1153 ld_offset-=wordSize;
1154 }
1155 } else {
1156 // In 64bit, longs are given 2 64-bit slots in the interpreter, but the
1157 // data is passed in only 1 slot.
1158 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
1159 ld_offset-=wordSize;
1160 }
1161 __ ld(r, ld_offset, ld_ptr);
1162 ld_offset-=wordSize;
1163 }
1164
1165 if (r_1->is_stack()) {
1166 // Now store value where the compiler expects it
1167 int st_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots())*VMRegImpl::stack_slot_size;
1168
1169 if (sig_bt[i] == T_INT || sig_bt[i] == T_FLOAT ||sig_bt[i] == T_BOOLEAN ||
1170 sig_bt[i] == T_SHORT || sig_bt[i] == T_CHAR || sig_bt[i] == T_BYTE) {
1171 __ stw(r, st_off, R1_SP);
1172 } else {
1173 __ std(r, st_off, R1_SP);
1174 }
1175 }
1176 }
1177 }
1178
1179 __ push_cont_fastpath(); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about
1180
1181 BLOCK_COMMENT("Store method");
1182 // Store method into thread->callee_target.
1183 // We might end up in handle_wrong_method if the callee is
1184 // deoptimized as we race thru here. If that happens we don't want
1185 // to take a safepoint because the caller frame will look
1186 // interpreted and arguments are now "compiled" so it is much better
1187 // to make this transition invisible to the stack walking
1188 // code. Unfortunately if we try and find the callee by normal means
1189 // a safepoint is possible. So we stash the desired callee in the
1190 // thread and the vm will find there should this case occur.
1191 __ std(R19_method, thread_(callee_target));
1192
1193 // Jump to the compiled code just as if compiled code was doing it.
1194 __ bctr();
1195 }
1196
1197 void SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
1198 int total_args_passed,
1199 int comp_args_on_stack,
1200 const BasicType *sig_bt,
1201 const VMRegPair *regs,
1202 AdapterHandlerEntry* handler) {
1203 address i2c_entry;
1204 address c2i_unverified_entry;
1205 address c2i_entry;
1206
1207
1208 // entry: i2c
1209
1210 __ align(CodeEntryAlignment);
1211 i2c_entry = __ pc();
1212 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
1213
1214
1215 // entry: c2i unverified
1216
1217 __ align(CodeEntryAlignment);
1218 BLOCK_COMMENT("c2i unverified entry");
1219 c2i_unverified_entry = __ pc();
1220
1221 // inline_cache contains a CompiledICData
1222 const Register ic = R19_inline_cache_reg;
1223 const Register ic_klass = R11_scratch1;
1224 const Register receiver_klass = R12_scratch2;
1225 const Register code = R21_tmp1;
1226 const Register ientry = R23_tmp3;
1227
1228 assert_different_registers(ic, ic_klass, receiver_klass, R3_ARG1, code, ientry);
1229 assert(R11_scratch1 == R11, "need prologue scratch register");
1230
1231 Label call_interpreter;
1232
1233 __ ic_check(4 /* end_alignment */);
1234 __ ld(R19_method, CompiledICData::speculated_method_offset(), ic);
1235 // Argument is valid and klass is as expected, continue.
1236
1237 __ ld(code, method_(code));
1238 __ cmpdi(CR0, code, 0);
1239 __ ld(ientry, method_(interpreter_entry)); // preloaded
1240 __ beq_predict_taken(CR0, call_interpreter);
1241
1242 // Branch to ic_miss_stub.
1243 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), relocInfo::runtime_call_type);
1244
1245 // entry: c2i
1246
1247 c2i_entry = __ pc();
1248
1249 // Class initialization barrier for static methods
1250 address c2i_no_clinit_check_entry = nullptr;
1251 if (VM_Version::supports_fast_class_init_checks()) {
1252 Label L_skip_barrier;
1253
1254 { // Bypass the barrier for non-static methods
1255 __ lhz(R0, in_bytes(Method::access_flags_offset()), R19_method);
1256 __ andi_(R0, R0, JVM_ACC_STATIC);
1257 __ beq(CR0, L_skip_barrier); // non-static
1258 }
1259
1260 Register klass = R11_scratch1;
1261 __ load_method_holder(klass, R19_method);
1262 __ clinit_barrier(klass, R16_thread, &L_skip_barrier /*L_fast_path*/);
1263
1264 __ load_const_optimized(klass, SharedRuntime::get_handle_wrong_method_stub(), R0);
1265 __ mtctr(klass);
1266 __ bctr();
1267
1268 __ bind(L_skip_barrier);
1269 c2i_no_clinit_check_entry = __ pc();
1270 }
1271
1272 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1273 bs->c2i_entry_barrier(masm, /* tmp register*/ ic_klass, /* tmp register*/ receiver_klass, /* tmp register*/ code);
1274
1275 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, call_interpreter, ientry);
1276
1277 handler->set_entry_points(i2c_entry, c2i_entry, c2i_unverified_entry, c2i_no_clinit_check_entry);
1278 return;
1279 }
1280
1281 // An oop arg. Must pass a handle not the oop itself.
1282 static void object_move(MacroAssembler* masm,
1283 int frame_size_in_slots,
1284 OopMap* oop_map, int oop_handle_offset,
1285 bool is_receiver, int* receiver_offset,
1286 VMRegPair src, VMRegPair dst,
1287 Register r_caller_sp, Register r_temp_1, Register r_temp_2) {
1288 assert(!is_receiver || (is_receiver && (*receiver_offset == -1)),
1289 "receiver has already been moved");
1290
1291 // We must pass a handle. First figure out the location we use as a handle.
1292
1293 if (src.first()->is_stack()) {
1294 // stack to stack or reg
1295
1296 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register();
1297 Label skip;
1298 const int oop_slot_in_callers_frame = reg2slot(src.first());
1299
1300 guarantee(!is_receiver, "expecting receiver in register");
1301 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot_in_callers_frame + frame_size_in_slots));
1302
1303 __ addi(r_handle, r_caller_sp, reg2offset(src.first()));
1304 __ ld( r_temp_2, reg2offset(src.first()), r_caller_sp);
1305 __ cmpdi(CR0, r_temp_2, 0);
1306 __ bne(CR0, skip);
1307 // Use a null handle if oop is null.
1308 __ li(r_handle, 0);
1309 __ bind(skip);
1310
1311 if (dst.first()->is_stack()) {
1312 // stack to stack
1313 __ std(r_handle, reg2offset(dst.first()), R1_SP);
1314 } else {
1315 // stack to reg
1316 // Nothing to do, r_handle is already the dst register.
1317 }
1318 } else {
1319 // reg to stack or reg
1320 const Register r_oop = src.first()->as_Register();
1321 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register();
1322 const int oop_slot = (r_oop->encoding()-R3_ARG1->encoding()) * VMRegImpl::slots_per_word
1323 + oop_handle_offset; // in slots
1324 const int oop_offset = oop_slot * VMRegImpl::stack_slot_size;
1325 Label skip;
1326
1327 if (is_receiver) {
1328 *receiver_offset = oop_offset;
1329 }
1330 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot));
1331
1332 __ std( r_oop, oop_offset, R1_SP);
1333 __ addi(r_handle, R1_SP, oop_offset);
1334
1335 __ cmpdi(CR0, r_oop, 0);
1336 __ bne(CR0, skip);
1337 // Use a null handle if oop is null.
1338 __ li(r_handle, 0);
1339 __ bind(skip);
1340
1341 if (dst.first()->is_stack()) {
1342 // reg to stack
1343 __ std(r_handle, reg2offset(dst.first()), R1_SP);
1344 } else {
1345 // reg to reg
1346 // Nothing to do, r_handle is already the dst register.
1347 }
1348 }
1349 }
1350
1351 static void int_move(MacroAssembler*masm,
1352 VMRegPair src, VMRegPair dst,
1353 Register r_caller_sp, Register r_temp) {
1354 assert(src.first()->is_valid(), "incoming must be int");
1355 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long");
1356
1357 if (src.first()->is_stack()) {
1358 if (dst.first()->is_stack()) {
1359 // stack to stack
1360 __ lwa(r_temp, reg2offset(src.first()), r_caller_sp);
1361 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1362 } else {
1363 // stack to reg
1364 __ lwa(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp);
1365 }
1366 } else if (dst.first()->is_stack()) {
1367 // reg to stack
1368 __ extsw(r_temp, src.first()->as_Register());
1369 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1370 } else {
1371 // reg to reg
1372 __ extsw(dst.first()->as_Register(), src.first()->as_Register());
1373 }
1374 }
1375
1376 static void long_move(MacroAssembler*masm,
1377 VMRegPair src, VMRegPair dst,
1378 Register r_caller_sp, Register r_temp) {
1379 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be long");
1380 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long");
1381
1382 if (src.first()->is_stack()) {
1383 if (dst.first()->is_stack()) {
1384 // stack to stack
1385 __ ld( r_temp, reg2offset(src.first()), r_caller_sp);
1386 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1387 } else {
1388 // stack to reg
1389 __ ld(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp);
1390 }
1391 } else if (dst.first()->is_stack()) {
1392 // reg to stack
1393 __ std(src.first()->as_Register(), reg2offset(dst.first()), R1_SP);
1394 } else {
1395 // reg to reg
1396 if (dst.first()->as_Register() != src.first()->as_Register())
1397 __ mr(dst.first()->as_Register(), src.first()->as_Register());
1398 }
1399 }
1400
1401 static void float_move(MacroAssembler*masm,
1402 VMRegPair src, VMRegPair dst,
1403 Register r_caller_sp, Register r_temp) {
1404 assert(src.first()->is_valid() && !src.second()->is_valid(), "incoming must be float");
1405 assert(dst.first()->is_valid() && !dst.second()->is_valid(), "outgoing must be float");
1406
1407 if (src.first()->is_stack()) {
1408 if (dst.first()->is_stack()) {
1409 // stack to stack
1410 __ lwz(r_temp, reg2offset(src.first()), r_caller_sp);
1411 __ stw(r_temp, reg2offset(dst.first()), R1_SP);
1412 } else {
1413 // stack to reg
1414 __ lfs(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp);
1415 }
1416 } else if (dst.first()->is_stack()) {
1417 // reg to stack
1418 __ stfs(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP);
1419 } else {
1420 // reg to reg
1421 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister())
1422 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister());
1423 }
1424 }
1425
1426 static void double_move(MacroAssembler*masm,
1427 VMRegPair src, VMRegPair dst,
1428 Register r_caller_sp, Register r_temp) {
1429 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be double");
1430 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be double");
1431
1432 if (src.first()->is_stack()) {
1433 if (dst.first()->is_stack()) {
1434 // stack to stack
1435 __ ld( r_temp, reg2offset(src.first()), r_caller_sp);
1436 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1437 } else {
1438 // stack to reg
1439 __ lfd(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp);
1440 }
1441 } else if (dst.first()->is_stack()) {
1442 // reg to stack
1443 __ stfd(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP);
1444 } else {
1445 // reg to reg
1446 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister())
1447 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister());
1448 }
1449 }
1450
1451 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1452 switch (ret_type) {
1453 case T_BOOLEAN:
1454 case T_CHAR:
1455 case T_BYTE:
1456 case T_SHORT:
1457 case T_INT:
1458 __ stw (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1459 break;
1460 case T_ARRAY:
1461 case T_OBJECT:
1462 case T_LONG:
1463 __ std (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1464 break;
1465 case T_FLOAT:
1466 __ stfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1467 break;
1468 case T_DOUBLE:
1469 __ stfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1470 break;
1471 case T_VOID:
1472 break;
1473 default:
1474 ShouldNotReachHere();
1475 break;
1476 }
1477 }
1478
1479 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1480 switch (ret_type) {
1481 case T_BOOLEAN:
1482 case T_CHAR:
1483 case T_BYTE:
1484 case T_SHORT:
1485 case T_INT:
1486 __ lwz(R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1487 break;
1488 case T_ARRAY:
1489 case T_OBJECT:
1490 case T_LONG:
1491 __ ld (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1492 break;
1493 case T_FLOAT:
1494 __ lfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1495 break;
1496 case T_DOUBLE:
1497 __ lfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1498 break;
1499 case T_VOID:
1500 break;
1501 default:
1502 ShouldNotReachHere();
1503 break;
1504 }
1505 }
1506
1507 static void verify_oop_args(MacroAssembler* masm,
1508 const methodHandle& method,
1509 const BasicType* sig_bt,
1510 const VMRegPair* regs) {
1511 Register temp_reg = R19_method; // not part of any compiled calling seq
1512 if (VerifyOops) {
1513 for (int i = 0; i < method->size_of_parameters(); i++) {
1514 if (is_reference_type(sig_bt[i])) {
1515 VMReg r = regs[i].first();
1516 assert(r->is_valid(), "bad oop arg");
1517 if (r->is_stack()) {
1518 __ ld(temp_reg, reg2offset(r), R1_SP);
1519 __ verify_oop(temp_reg, FILE_AND_LINE);
1520 } else {
1521 __ verify_oop(r->as_Register(), FILE_AND_LINE);
1522 }
1523 }
1524 }
1525 }
1526 }
1527
1528 static void gen_special_dispatch(MacroAssembler* masm,
1529 const methodHandle& method,
1530 const BasicType* sig_bt,
1531 const VMRegPair* regs) {
1532 verify_oop_args(masm, method, sig_bt, regs);
1533 vmIntrinsics::ID iid = method->intrinsic_id();
1534
1535 // Now write the args into the outgoing interpreter space
1536 bool has_receiver = false;
1537 Register receiver_reg = noreg;
1538 int member_arg_pos = -1;
1539 Register member_reg = noreg;
1540 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1541 if (ref_kind != 0) {
1542 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1543 member_reg = R19_method; // known to be free at this point
1544 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1545 } else if (iid == vmIntrinsics::_invokeBasic) {
1546 has_receiver = true;
1547 } else if (iid == vmIntrinsics::_linkToNative) {
1548 member_arg_pos = method->size_of_parameters() - 1; // trailing NativeEntryPoint argument
1549 member_reg = R19_method; // known to be free at this point
1550 } else {
1551 fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1552 }
1553
1554 if (member_reg != noreg) {
1555 // Load the member_arg into register, if necessary.
1556 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1557 VMReg r = regs[member_arg_pos].first();
1558 if (r->is_stack()) {
1559 __ ld(member_reg, reg2offset(r), R1_SP);
1560 } else {
1561 // no data motion is needed
1562 member_reg = r->as_Register();
1563 }
1564 }
1565
1566 if (has_receiver) {
1567 // Make sure the receiver is loaded into a register.
1568 assert(method->size_of_parameters() > 0, "oob");
1569 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1570 VMReg r = regs[0].first();
1571 assert(r->is_valid(), "bad receiver arg");
1572 if (r->is_stack()) {
1573 // Porting note: This assumes that compiled calling conventions always
1574 // pass the receiver oop in a register. If this is not true on some
1575 // platform, pick a temp and load the receiver from stack.
1576 fatal("receiver always in a register");
1577 receiver_reg = R11_scratch1; // TODO (hs24): is R11_scratch1 really free at this point?
1578 __ ld(receiver_reg, reg2offset(r), R1_SP);
1579 } else {
1580 // no data motion is needed
1581 receiver_reg = r->as_Register();
1582 }
1583 }
1584
1585 // Figure out which address we are really jumping to:
1586 MethodHandles::generate_method_handle_dispatch(masm, iid,
1587 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1588 }
1589
1590 //---------------------------- continuation_enter_setup ---------------------------
1591 //
1592 // Frame setup.
1593 //
1594 // Arguments:
1595 // None.
1596 //
1597 // Results:
1598 // R1_SP: pointer to blank ContinuationEntry in the pushed frame.
1599 //
1600 // Kills:
1601 // R0, R20
1602 //
1603 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& framesize_words) {
1604 assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
1605 assert(in_bytes(ContinuationEntry::cont_offset()) % VMRegImpl::stack_slot_size == 0, "");
1606 assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
1607
1608 const int frame_size_in_bytes = (int)ContinuationEntry::size();
1609 assert(is_aligned(frame_size_in_bytes, frame::alignment_in_bytes), "alignment error");
1610
1611 framesize_words = frame_size_in_bytes / wordSize;
1612
1613 DEBUG_ONLY(__ block_comment("setup {"));
1614 // Save return pc and push entry frame
1615 const Register return_pc = R20;
1616 __ mflr(return_pc);
1617 __ std(return_pc, _abi0(lr), R1_SP); // SP->lr = return_pc
1618 __ push_frame(frame_size_in_bytes , R0); // SP -= frame_size_in_bytes
1619
1620 OopMap* map = new OopMap((int)frame_size_in_bytes / VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
1621
1622 __ ld_ptr(R0, JavaThread::cont_entry_offset(), R16_thread);
1623 __ st_ptr(R1_SP, JavaThread::cont_entry_offset(), R16_thread);
1624 __ st_ptr(R0, ContinuationEntry::parent_offset(), R1_SP);
1625 DEBUG_ONLY(__ block_comment("} setup"));
1626
1627 return map;
1628 }
1629
1630 //---------------------------- fill_continuation_entry ---------------------------
1631 //
1632 // Initialize the new ContinuationEntry.
1633 //
1634 // Arguments:
1635 // R1_SP: pointer to blank Continuation entry
1636 // reg_cont_obj: pointer to the continuation
1637 // reg_flags: flags
1638 //
1639 // Results:
1640 // R1_SP: pointer to filled out ContinuationEntry
1641 //
1642 // Kills:
1643 // R8_ARG6, R9_ARG7, R10_ARG8
1644 //
1645 static void fill_continuation_entry(MacroAssembler* masm, Register reg_cont_obj, Register reg_flags) {
1646 assert_different_registers(reg_cont_obj, reg_flags);
1647 Register zero = R8_ARG6;
1648 Register tmp2 = R9_ARG7;
1649 Register tmp3 = R10_ARG8;
1650
1651 DEBUG_ONLY(__ block_comment("fill {"));
1652 #ifdef ASSERT
1653 __ load_const_optimized(tmp2, ContinuationEntry::cookie_value());
1654 __ stw(tmp2, in_bytes(ContinuationEntry::cookie_offset()), R1_SP);
1655 #endif //ASSERT
1656
1657 __ li(zero, 0);
1658 __ st_ptr(reg_cont_obj, ContinuationEntry::cont_offset(), R1_SP);
1659 __ stw(reg_flags, in_bytes(ContinuationEntry::flags_offset()), R1_SP);
1660 __ st_ptr(zero, ContinuationEntry::chunk_offset(), R1_SP);
1661 __ stw(zero, in_bytes(ContinuationEntry::argsize_offset()), R1_SP);
1662 __ stw(zero, in_bytes(ContinuationEntry::pin_count_offset()), R1_SP);
1663
1664 __ ld_ptr(tmp2, JavaThread::cont_fastpath_offset(), R16_thread);
1665 __ ld(tmp3, in_bytes(JavaThread::held_monitor_count_offset()), R16_thread);
1666 __ st_ptr(tmp2, ContinuationEntry::parent_cont_fastpath_offset(), R1_SP);
1667 __ std(tmp3, in_bytes(ContinuationEntry::parent_held_monitor_count_offset()), R1_SP);
1668
1669 __ st_ptr(zero, JavaThread::cont_fastpath_offset(), R16_thread);
1670 __ std(zero, in_bytes(JavaThread::held_monitor_count_offset()), R16_thread);
1671 DEBUG_ONLY(__ block_comment("} fill"));
1672 }
1673
1674 //---------------------------- continuation_enter_cleanup ---------------------------
1675 //
1676 // Copy corresponding attributes from the top ContinuationEntry to the JavaThread
1677 // before deleting it.
1678 //
1679 // Arguments:
1680 // R1_SP: pointer to the ContinuationEntry
1681 //
1682 // Results:
1683 // None.
1684 //
1685 // Kills:
1686 // R8_ARG6, R9_ARG7, R10_ARG8, R15_esp
1687 //
1688 static void continuation_enter_cleanup(MacroAssembler* masm) {
1689 Register tmp1 = R8_ARG6;
1690 Register tmp2 = R9_ARG7;
1691 Register tmp3 = R10_ARG8;
1692
1693 #ifdef ASSERT
1694 __ block_comment("clean {");
1695 __ ld_ptr(tmp1, JavaThread::cont_entry_offset(), R16_thread);
1696 __ cmpd(CR0, R1_SP, tmp1);
1697 __ asm_assert_eq(FILE_AND_LINE ": incorrect R1_SP");
1698 #endif
1699
1700 __ ld_ptr(tmp1, ContinuationEntry::parent_cont_fastpath_offset(), R1_SP);
1701 __ st_ptr(tmp1, JavaThread::cont_fastpath_offset(), R16_thread);
1702
1703 if (CheckJNICalls) {
1704 // Check if this is a virtual thread continuation
1705 Label L_skip_vthread_code;
1706 __ lwz(R0, in_bytes(ContinuationEntry::flags_offset()), R1_SP);
1707 __ cmpwi(CR0, R0, 0);
1708 __ beq(CR0, L_skip_vthread_code);
1709
1710 // If the held monitor count is > 0 and this vthread is terminating then
1711 // it failed to release a JNI monitor. So we issue the same log message
1712 // that JavaThread::exit does.
1713 __ ld(R0, in_bytes(JavaThread::jni_monitor_count_offset()), R16_thread);
1714 __ cmpdi(CR0, R0, 0);
1715 __ beq(CR0, L_skip_vthread_code);
1716
1717 // Save return value potentially containing the exception oop
1718 Register ex_oop = R15_esp; // nonvolatile register
1719 __ mr(ex_oop, R3_RET);
1720 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::log_jni_monitor_still_held));
1721 // Restore potental return value
1722 __ mr(R3_RET, ex_oop);
1723
1724 // For vthreads we have to explicitly zero the JNI monitor count of the carrier
1725 // on termination. The held count is implicitly zeroed below when we restore from
1726 // the parent held count (which has to be zero).
1727 __ li(tmp1, 0);
1728 __ std(tmp1, in_bytes(JavaThread::jni_monitor_count_offset()), R16_thread);
1729
1730 __ bind(L_skip_vthread_code);
1731 }
1732 #ifdef ASSERT
1733 else {
1734 // Check if this is a virtual thread continuation
1735 Label L_skip_vthread_code;
1736 __ lwz(R0, in_bytes(ContinuationEntry::flags_offset()), R1_SP);
1737 __ cmpwi(CR0, R0, 0);
1738 __ beq(CR0, L_skip_vthread_code);
1739
1740 // See comment just above. If not checking JNI calls the JNI count is only
1741 // needed for assertion checking.
1742 __ li(tmp1, 0);
1743 __ std(tmp1, in_bytes(JavaThread::jni_monitor_count_offset()), R16_thread);
1744
1745 __ bind(L_skip_vthread_code);
1746 }
1747 #endif
1748
1749 __ ld(tmp2, in_bytes(ContinuationEntry::parent_held_monitor_count_offset()), R1_SP);
1750 __ ld_ptr(tmp3, ContinuationEntry::parent_offset(), R1_SP);
1751 __ std(tmp2, in_bytes(JavaThread::held_monitor_count_offset()), R16_thread);
1752 __ st_ptr(tmp3, JavaThread::cont_entry_offset(), R16_thread);
1753 DEBUG_ONLY(__ block_comment("} clean"));
1754 }
1755
1756 static void check_continuation_enter_argument(VMReg actual_vmreg,
1757 Register expected_reg,
1758 const char* name) {
1759 assert(!actual_vmreg->is_stack(), "%s cannot be on stack", name);
1760 assert(actual_vmreg->as_Register() == expected_reg,
1761 "%s is in unexpected register: %s instead of %s",
1762 name, actual_vmreg->as_Register()->name(), expected_reg->name());
1763 }
1764
1765 static void gen_continuation_enter(MacroAssembler* masm,
1766 const VMRegPair* regs,
1767 int& exception_offset,
1768 OopMapSet* oop_maps,
1769 int& frame_complete,
1770 int& framesize_words,
1771 int& interpreted_entry_offset,
1772 int& compiled_entry_offset) {
1773
1774 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
1775 int pos_cont_obj = 0;
1776 int pos_is_cont = 1;
1777 int pos_is_virtual = 2;
1778
1779 // The platform-specific calling convention may present the arguments in various registers.
1780 // To simplify the rest of the code, we expect the arguments to reside at these known
1781 // registers, and we additionally check the placement here in case calling convention ever
1782 // changes.
1783 Register reg_cont_obj = R3_ARG1;
1784 Register reg_is_cont = R4_ARG2;
1785 Register reg_is_virtual = R5_ARG3;
1786
1787 check_continuation_enter_argument(regs[pos_cont_obj].first(), reg_cont_obj, "Continuation object");
1788 check_continuation_enter_argument(regs[pos_is_cont].first(), reg_is_cont, "isContinue");
1789 check_continuation_enter_argument(regs[pos_is_virtual].first(), reg_is_virtual, "isVirtualThread");
1790
1791 address resolve_static_call = SharedRuntime::get_resolve_static_call_stub();
1792
1793 address start = __ pc();
1794
1795 Label L_thaw, L_exit;
1796
1797 // i2i entry used at interp_only_mode only
1798 interpreted_entry_offset = __ pc() - start;
1799 {
1800 #ifdef ASSERT
1801 Label is_interp_only;
1802 __ lwz(R0, in_bytes(JavaThread::interp_only_mode_offset()), R16_thread);
1803 __ cmpwi(CR0, R0, 0);
1804 __ bne(CR0, is_interp_only);
1805 __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
1806 __ bind(is_interp_only);
1807 #endif
1808
1809 // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
1810 __ ld(reg_cont_obj, Interpreter::stackElementSize*3, R15_esp);
1811 __ lwz(reg_is_cont, Interpreter::stackElementSize*2, R15_esp);
1812 __ lwz(reg_is_virtual, Interpreter::stackElementSize*1, R15_esp);
1813
1814 __ push_cont_fastpath();
1815
1816 OopMap* map = continuation_enter_setup(masm, framesize_words);
1817
1818 // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
1819 // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
1820
1821 fill_continuation_entry(masm, reg_cont_obj, reg_is_virtual);
1822
1823 // If isContinue, call to thaw. Otherwise, call Continuation.enter(Continuation c, boolean isContinue)
1824 __ cmpwi(CR0, reg_is_cont, 0);
1825 __ bne(CR0, L_thaw);
1826
1827 // --- call Continuation.enter(Continuation c, boolean isContinue)
1828
1829 // Emit compiled static call. The call will be always resolved to the c2i
1830 // entry of Continuation.enter(Continuation c, boolean isContinue).
1831 // There are special cases in SharedRuntime::resolve_static_call_C() and
1832 // SharedRuntime::resolve_sub_helper_internal() to achieve this
1833 // See also corresponding call below.
1834 address c2i_call_pc = __ pc();
1835 int start_offset = __ offset();
1836 // Put the entry point as a constant into the constant pool.
1837 const address entry_point_toc_addr = __ address_constant(resolve_static_call, RelocationHolder::none);
1838 const int entry_point_toc_offset = __ offset_to_method_toc(entry_point_toc_addr);
1839 guarantee(entry_point_toc_addr != nullptr, "const section overflow");
1840
1841 // Emit the trampoline stub which will be related to the branch-and-link below.
1842 address stub = __ emit_trampoline_stub(entry_point_toc_offset, start_offset);
1843 guarantee(stub != nullptr, "no space for trampoline stub");
1844
1845 __ relocate(relocInfo::static_call_type);
1846 // Note: At this point we do not have the address of the trampoline
1847 // stub, and the entry point might be too far away for bl, so __ pc()
1848 // serves as dummy and the bl will be patched later.
1849 __ bl(__ pc());
1850 oop_maps->add_gc_map(__ pc() - start, map);
1851 __ post_call_nop();
1852
1853 __ b(L_exit);
1854
1855 // static stub for the call above
1856 stub = CompiledDirectCall::emit_to_interp_stub(masm, c2i_call_pc);
1857 guarantee(stub != nullptr, "no space for static stub");
1858 }
1859
1860 // compiled entry
1861 __ align(CodeEntryAlignment);
1862 compiled_entry_offset = __ pc() - start;
1863
1864 OopMap* map = continuation_enter_setup(masm, framesize_words);
1865
1866 // Frame is now completed as far as size and linkage.
1867 frame_complete =__ pc() - start;
1868
1869 fill_continuation_entry(masm, reg_cont_obj, reg_is_virtual);
1870
1871 // If isContinue, call to thaw. Otherwise, call Continuation.enter(Continuation c, boolean isContinue)
1872 __ cmpwi(CR0, reg_is_cont, 0);
1873 __ bne(CR0, L_thaw);
1874
1875 // --- call Continuation.enter(Continuation c, boolean isContinue)
1876
1877 // Emit compiled static call
1878 // The call needs to be resolved. There's a special case for this in
1879 // SharedRuntime::find_callee_info_helper() which calls
1880 // LinkResolver::resolve_continuation_enter() which resolves the call to
1881 // Continuation.enter(Continuation c, boolean isContinue).
1882 address call_pc = __ pc();
1883 int start_offset = __ offset();
1884 // Put the entry point as a constant into the constant pool.
1885 const address entry_point_toc_addr = __ address_constant(resolve_static_call, RelocationHolder::none);
1886 const int entry_point_toc_offset = __ offset_to_method_toc(entry_point_toc_addr);
1887 guarantee(entry_point_toc_addr != nullptr, "const section overflow");
1888
1889 // Emit the trampoline stub which will be related to the branch-and-link below.
1890 address stub = __ emit_trampoline_stub(entry_point_toc_offset, start_offset);
1891 guarantee(stub != nullptr, "no space for trampoline stub");
1892
1893 __ relocate(relocInfo::static_call_type);
1894 // Note: At this point we do not have the address of the trampoline
1895 // stub, and the entry point might be too far away for bl, so __ pc()
1896 // serves as dummy and the bl will be patched later.
1897 __ bl(__ pc());
1898 oop_maps->add_gc_map(__ pc() - start, map);
1899 __ post_call_nop();
1900
1901 __ b(L_exit);
1902
1903 // --- Thawing path
1904
1905 __ bind(L_thaw);
1906 ContinuationEntry::_thaw_call_pc_offset = __ pc() - start;
1907 __ add_const_optimized(R0, R29_TOC, MacroAssembler::offset_to_global_toc(StubRoutines::cont_thaw()));
1908 __ mtctr(R0);
1909 __ bctrl();
1910 oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1911 ContinuationEntry::_return_pc_offset = __ pc() - start;
1912 __ post_call_nop();
1913
1914 // --- Normal exit (resolve/thawing)
1915
1916 __ bind(L_exit);
1917 ContinuationEntry::_cleanup_offset = __ pc() - start;
1918 continuation_enter_cleanup(masm);
1919
1920 // Pop frame and return
1921 DEBUG_ONLY(__ ld_ptr(R0, 0, R1_SP));
1922 __ addi(R1_SP, R1_SP, framesize_words*wordSize);
1923 DEBUG_ONLY(__ cmpd(CR0, R0, R1_SP));
1924 __ asm_assert_eq(FILE_AND_LINE ": inconsistent frame size");
1925 __ ld(R0, _abi0(lr), R1_SP); // Return pc
1926 __ mtlr(R0);
1927 __ blr();
1928
1929 // --- Exception handling path
1930
1931 exception_offset = __ pc() - start;
1932
1933 continuation_enter_cleanup(masm);
1934 Register ex_pc = R17_tos; // nonvolatile register
1935 Register ex_oop = R15_esp; // nonvolatile register
1936 __ ld(ex_pc, _abi0(callers_sp), R1_SP); // Load caller's return pc
1937 __ ld(ex_pc, _abi0(lr), ex_pc);
1938 __ mr(ex_oop, R3_RET); // save return value containing the exception oop
1939 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), R16_thread, ex_pc);
1940 __ mtlr(R3_RET); // the exception handler
1941 __ ld(R1_SP, _abi0(callers_sp), R1_SP); // remove enterSpecial frame
1942
1943 // Continue at exception handler
1944 // See OptoRuntime::generate_exception_blob for register arguments
1945 __ mr(R3_ARG1, ex_oop); // pass exception oop
1946 __ mr(R4_ARG2, ex_pc); // pass exception pc
1947 __ blr();
1948
1949 // static stub for the call above
1950 stub = CompiledDirectCall::emit_to_interp_stub(masm, call_pc);
1951 guarantee(stub != nullptr, "no space for static stub");
1952 }
1953
1954 static void gen_continuation_yield(MacroAssembler* masm,
1955 const VMRegPair* regs,
1956 OopMapSet* oop_maps,
1957 int& frame_complete,
1958 int& framesize_words,
1959 int& compiled_entry_offset) {
1960 Register tmp = R10_ARG8;
1961
1962 const int framesize_bytes = (int)align_up((int)frame::native_abi_reg_args_size, frame::alignment_in_bytes);
1963 framesize_words = framesize_bytes / wordSize;
1964
1965 address start = __ pc();
1966 compiled_entry_offset = __ pc() - start;
1967
1968 // Save return pc and push entry frame
1969 __ mflr(tmp);
1970 __ std(tmp, _abi0(lr), R1_SP); // SP->lr = return_pc
1971 __ push_frame(framesize_bytes , R0); // SP -= frame_size_in_bytes
1972
1973 DEBUG_ONLY(__ block_comment("Frame Complete"));
1974 frame_complete = __ pc() - start;
1975 address last_java_pc = __ pc();
1976
1977 // This nop must be exactly at the PC we push into the frame info.
1978 // We use this nop for fast CodeBlob lookup, associate the OopMap
1979 // with it right away.
1980 __ post_call_nop();
1981 OopMap* map = new OopMap(framesize_bytes / VMRegImpl::stack_slot_size, 1);
1982 oop_maps->add_gc_map(last_java_pc - start, map);
1983
1984 __ calculate_address_from_global_toc(tmp, last_java_pc); // will be relocated
1985 __ set_last_Java_frame(R1_SP, tmp);
1986 __ call_VM_leaf(Continuation::freeze_entry(), R16_thread, R1_SP);
1987 __ reset_last_Java_frame();
1988
1989 Label L_pinned;
1990
1991 __ cmpwi(CR0, R3_RET, 0);
1992 __ bne(CR0, L_pinned);
1993
1994 // yield succeeded
1995
1996 // Pop frames of continuation including this stub's frame
1997 __ ld_ptr(R1_SP, JavaThread::cont_entry_offset(), R16_thread);
1998 // The frame pushed by gen_continuation_enter is on top now again
1999 continuation_enter_cleanup(masm);
2000
2001 // Pop frame and return
2002 Label L_return;
2003 __ bind(L_return);
2004 __ pop_frame();
2005 __ ld(R0, _abi0(lr), R1_SP); // Return pc
2006 __ mtlr(R0);
2007 __ blr();
2008
2009 // yield failed - continuation is pinned
2010
2011 __ bind(L_pinned);
2012
2013 // handle pending exception thrown by freeze
2014 __ ld(tmp, in_bytes(JavaThread::pending_exception_offset()), R16_thread);
2015 __ cmpdi(CR0, tmp, 0);
2016 __ beq(CR0, L_return); // return if no exception is pending
2017 __ pop_frame();
2018 __ ld(R0, _abi0(lr), R1_SP); // Return pc
2019 __ mtlr(R0);
2020 __ load_const_optimized(tmp, StubRoutines::forward_exception_entry(), R0);
2021 __ mtctr(tmp);
2022 __ bctr();
2023 }
2024
2025 void SharedRuntime::continuation_enter_cleanup(MacroAssembler* masm) {
2026 ::continuation_enter_cleanup(masm);
2027 }
2028
2029 // ---------------------------------------------------------------------------
2030 // Generate a native wrapper for a given method. The method takes arguments
2031 // in the Java compiled code convention, marshals them to the native
2032 // convention (handlizes oops, etc), transitions to native, makes the call,
2033 // returns to java state (possibly blocking), unhandlizes any result and
2034 // returns.
2035 //
2036 // Critical native functions are a shorthand for the use of
2037 // GetPrimtiveArrayCritical and disallow the use of any other JNI
2038 // functions. The wrapper is expected to unpack the arguments before
2039 // passing them to the callee. Critical native functions leave the state _in_Java,
2040 // since they cannot stop for GC.
2041 // Some other parts of JNI setup are skipped like the tear down of the JNI handle
2042 // block and the check for pending exceptions it's impossible for them
2043 // to be thrown.
2044 //
2045 nmethod *SharedRuntime::generate_native_wrapper(MacroAssembler *masm,
2046 const methodHandle& method,
2047 int compile_id,
2048 BasicType *in_sig_bt,
2049 VMRegPair *in_regs,
2050 BasicType ret_type) {
2051 if (method->is_continuation_native_intrinsic()) {
2052 int exception_offset = -1;
2053 OopMapSet* oop_maps = new OopMapSet();
2054 int frame_complete = -1;
2055 int stack_slots = -1;
2056 int interpreted_entry_offset = -1;
2057 int vep_offset = -1;
2058 if (method->is_continuation_enter_intrinsic()) {
2059 gen_continuation_enter(masm,
2060 in_regs,
2061 exception_offset,
2062 oop_maps,
2063 frame_complete,
2064 stack_slots,
2065 interpreted_entry_offset,
2066 vep_offset);
2067 } else if (method->is_continuation_yield_intrinsic()) {
2068 gen_continuation_yield(masm,
2069 in_regs,
2070 oop_maps,
2071 frame_complete,
2072 stack_slots,
2073 vep_offset);
2074 } else {
2075 guarantee(false, "Unknown Continuation native intrinsic");
2076 }
2077
2078 #ifdef ASSERT
2079 if (method->is_continuation_enter_intrinsic()) {
2080 assert(interpreted_entry_offset != -1, "Must be set");
2081 assert(exception_offset != -1, "Must be set");
2082 } else {
2083 assert(interpreted_entry_offset == -1, "Must be unset");
2084 assert(exception_offset == -1, "Must be unset");
2085 }
2086 assert(frame_complete != -1, "Must be set");
2087 assert(stack_slots != -1, "Must be set");
2088 assert(vep_offset != -1, "Must be set");
2089 #endif
2090
2091 __ flush();
2092 nmethod* nm = nmethod::new_native_nmethod(method,
2093 compile_id,
2094 masm->code(),
2095 vep_offset,
2096 frame_complete,
2097 stack_slots,
2098 in_ByteSize(-1),
2099 in_ByteSize(-1),
2100 oop_maps,
2101 exception_offset);
2102 if (nm == nullptr) return nm;
2103 if (method->is_continuation_enter_intrinsic()) {
2104 ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
2105 } else if (method->is_continuation_yield_intrinsic()) {
2106 _cont_doYield_stub = nm;
2107 }
2108 return nm;
2109 }
2110
2111 if (method->is_method_handle_intrinsic()) {
2112 vmIntrinsics::ID iid = method->intrinsic_id();
2113 intptr_t start = (intptr_t)__ pc();
2114 int vep_offset = ((intptr_t)__ pc()) - start;
2115 gen_special_dispatch(masm,
2116 method,
2117 in_sig_bt,
2118 in_regs);
2119 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
2120 __ flush();
2121 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
2122 return nmethod::new_native_nmethod(method,
2123 compile_id,
2124 masm->code(),
2125 vep_offset,
2126 frame_complete,
2127 stack_slots / VMRegImpl::slots_per_word,
2128 in_ByteSize(-1),
2129 in_ByteSize(-1),
2130 (OopMapSet*)nullptr);
2131 }
2132
2133 address native_func = method->native_function();
2134 assert(native_func != nullptr, "must have function");
2135
2136 // First, create signature for outgoing C call
2137 // --------------------------------------------------------------------------
2138
2139 int total_in_args = method->size_of_parameters();
2140 // We have received a description of where all the java args are located
2141 // on entry to the wrapper. We need to convert these args to where
2142 // the jni function will expect them. To figure out where they go
2143 // we convert the java signature to a C signature by inserting
2144 // the hidden arguments as arg[0] and possibly arg[1] (static method)
2145
2146 // Calculate the total number of C arguments and create arrays for the
2147 // signature and the outgoing registers.
2148 // On ppc64, we have two arrays for the outgoing registers, because
2149 // some floating-point arguments must be passed in registers _and_
2150 // in stack locations.
2151 bool method_is_static = method->is_static();
2152 int total_c_args = total_in_args + (method_is_static ? 2 : 1);
2153
2154 BasicType *out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
2155 VMRegPair *out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
2156
2157 // Create the signature for the C call:
2158 // 1) add the JNIEnv*
2159 // 2) add the class if the method is static
2160 // 3) copy the rest of the incoming signature (shifted by the number of
2161 // hidden arguments).
2162
2163 int argc = 0;
2164 out_sig_bt[argc++] = T_ADDRESS;
2165 if (method->is_static()) {
2166 out_sig_bt[argc++] = T_OBJECT;
2167 }
2168
2169 for (int i = 0; i < total_in_args ; i++ ) {
2170 out_sig_bt[argc++] = in_sig_bt[i];
2171 }
2172
2173
2174 // Compute the wrapper's frame size.
2175 // --------------------------------------------------------------------------
2176
2177 // Now figure out where the args must be stored and how much stack space
2178 // they require.
2179 //
2180 // Compute framesize for the wrapper. We need to handlize all oops in
2181 // incoming registers.
2182 //
2183 // Calculate the total number of stack slots we will need:
2184 // 1) abi requirements
2185 // 2) outgoing arguments
2186 // 3) space for inbound oop handle area
2187 // 4) space for handlizing a klass if static method
2188 // 5) space for a lock if synchronized method
2189 // 6) workspace for saving return values, int <-> float reg moves, etc.
2190 // 7) alignment
2191 //
2192 // Layout of the native wrapper frame:
2193 // (stack grows upwards, memory grows downwards)
2194 //
2195 // NW [ABI_REG_ARGS] <-- 1) R1_SP
2196 // [outgoing arguments] <-- 2) R1_SP + out_arg_slot_offset
2197 // [oopHandle area] <-- 3) R1_SP + oop_handle_offset
2198 // klass <-- 4) R1_SP + klass_offset
2199 // lock <-- 5) R1_SP + lock_offset
2200 // [workspace] <-- 6) R1_SP + workspace_offset
2201 // [alignment] (optional) <-- 7)
2202 // caller [JIT_TOP_ABI_48] <-- r_callers_sp
2203 //
2204 // - *_slot_offset Indicates offset from SP in number of stack slots.
2205 // - *_offset Indicates offset from SP in bytes.
2206
2207 int stack_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args) + // 1+2)
2208 SharedRuntime::out_preserve_stack_slots(); // See c_calling_convention.
2209
2210 // Now the space for the inbound oop handle area.
2211 int total_save_slots = num_java_iarg_registers * VMRegImpl::slots_per_word;
2212
2213 int oop_handle_slot_offset = stack_slots;
2214 stack_slots += total_save_slots; // 3)
2215
2216 int klass_slot_offset = 0;
2217 int klass_offset = -1;
2218 if (method_is_static) { // 4)
2219 klass_slot_offset = stack_slots;
2220 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
2221 stack_slots += VMRegImpl::slots_per_word;
2222 }
2223
2224 int lock_slot_offset = 0;
2225 int lock_offset = -1;
2226 if (method->is_synchronized()) { // 5)
2227 lock_slot_offset = stack_slots;
2228 lock_offset = lock_slot_offset * VMRegImpl::stack_slot_size;
2229 stack_slots += VMRegImpl::slots_per_word;
2230 }
2231
2232 int workspace_slot_offset = stack_slots; // 6)
2233 stack_slots += 2;
2234
2235 // Now compute actual number of stack words we need.
2236 // Rounding to make stack properly aligned.
2237 stack_slots = align_up(stack_slots, // 7)
2238 frame::alignment_in_bytes / VMRegImpl::stack_slot_size);
2239 int frame_size_in_bytes = stack_slots * VMRegImpl::stack_slot_size;
2240
2241
2242 // Now we can start generating code.
2243 // --------------------------------------------------------------------------
2244
2245 intptr_t start_pc = (intptr_t)__ pc();
2246 intptr_t vep_start_pc;
2247 intptr_t frame_done_pc;
2248
2249 Label handle_pending_exception;
2250 Label last_java_pc;
2251
2252 Register r_callers_sp = R21;
2253 Register r_temp_1 = R22;
2254 Register r_temp_2 = R23;
2255 Register r_temp_3 = R24;
2256 Register r_temp_4 = R25;
2257 Register r_temp_5 = R26;
2258 Register r_temp_6 = R27;
2259 Register r_last_java_pc = R28;
2260
2261 Register r_carg1_jnienv = noreg;
2262 Register r_carg2_classorobject = noreg;
2263 r_carg1_jnienv = out_regs[0].first()->as_Register();
2264 r_carg2_classorobject = out_regs[1].first()->as_Register();
2265
2266
2267 // Generate the Unverified Entry Point (UEP).
2268 // --------------------------------------------------------------------------
2269 assert(start_pc == (intptr_t)__ pc(), "uep must be at start");
2270
2271 // Check ic: object class == cached class?
2272 if (!method_is_static) {
2273 __ ic_check(4 /* end_alignment */);
2274 }
2275
2276 // Generate the Verified Entry Point (VEP).
2277 // --------------------------------------------------------------------------
2278 vep_start_pc = (intptr_t)__ pc();
2279
2280 if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) {
2281 Label L_skip_barrier;
2282 Register klass = r_temp_1;
2283 // Notify OOP recorder (don't need the relocation)
2284 AddressLiteral md = __ constant_metadata_address(method->method_holder());
2285 __ load_const_optimized(klass, md.value(), R0);
2286 __ clinit_barrier(klass, R16_thread, &L_skip_barrier /*L_fast_path*/);
2287
2288 __ load_const_optimized(klass, SharedRuntime::get_handle_wrong_method_stub(), R0);
2289 __ mtctr(klass);
2290 __ bctr();
2291
2292 __ bind(L_skip_barrier);
2293 }
2294
2295 __ save_LR(r_temp_1);
2296 __ generate_stack_overflow_check(frame_size_in_bytes); // Check before creating frame.
2297 __ mr(r_callers_sp, R1_SP); // Remember frame pointer.
2298 __ push_frame(frame_size_in_bytes, r_temp_1); // Push the c2n adapter's frame.
2299
2300 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
2301 bs->nmethod_entry_barrier(masm, r_temp_1);
2302
2303 frame_done_pc = (intptr_t)__ pc();
2304
2305 // Native nmethod wrappers never take possession of the oop arguments.
2306 // So the caller will gc the arguments.
2307 // The only thing we need an oopMap for is if the call is static.
2308 //
2309 // An OopMap for lock (and class if static), and one for the VM call itself.
2310 OopMapSet *oop_maps = new OopMapSet();
2311 OopMap *oop_map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
2312
2313 // Move arguments from register/stack to register/stack.
2314 // --------------------------------------------------------------------------
2315 //
2316 // We immediately shuffle the arguments so that for any vm call we have
2317 // to make from here on out (sync slow path, jvmti, etc.) we will have
2318 // captured the oops from our caller and have a valid oopMap for them.
2319 //
2320 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv*
2321 // (derived from JavaThread* which is in R16_thread) and, if static,
2322 // the class mirror instead of a receiver. This pretty much guarantees that
2323 // register layout will not match. We ignore these extra arguments during
2324 // the shuffle. The shuffle is described by the two calling convention
2325 // vectors we have in our possession. We simply walk the java vector to
2326 // get the source locations and the c vector to get the destinations.
2327
2328 // Record sp-based slot for receiver on stack for non-static methods.
2329 int receiver_offset = -1;
2330
2331 // We move the arguments backward because the floating point registers
2332 // destination will always be to a register with a greater or equal
2333 // register number or the stack.
2334 // in is the index of the incoming Java arguments
2335 // out is the index of the outgoing C arguments
2336
2337 #ifdef ASSERT
2338 bool reg_destroyed[Register::number_of_registers];
2339 bool freg_destroyed[FloatRegister::number_of_registers];
2340 for (int r = 0 ; r < Register::number_of_registers ; r++) {
2341 reg_destroyed[r] = false;
2342 }
2343 for (int f = 0 ; f < FloatRegister::number_of_registers ; f++) {
2344 freg_destroyed[f] = false;
2345 }
2346 #endif // ASSERT
2347
2348 for (int in = total_in_args - 1, out = total_c_args - 1; in >= 0 ; in--, out--) {
2349
2350 #ifdef ASSERT
2351 if (in_regs[in].first()->is_Register()) {
2352 assert(!reg_destroyed[in_regs[in].first()->as_Register()->encoding()], "ack!");
2353 } else if (in_regs[in].first()->is_FloatRegister()) {
2354 assert(!freg_destroyed[in_regs[in].first()->as_FloatRegister()->encoding()], "ack!");
2355 }
2356 if (out_regs[out].first()->is_Register()) {
2357 reg_destroyed[out_regs[out].first()->as_Register()->encoding()] = true;
2358 } else if (out_regs[out].first()->is_FloatRegister()) {
2359 freg_destroyed[out_regs[out].first()->as_FloatRegister()->encoding()] = true;
2360 }
2361 #endif // ASSERT
2362
2363 switch (in_sig_bt[in]) {
2364 case T_BOOLEAN:
2365 case T_CHAR:
2366 case T_BYTE:
2367 case T_SHORT:
2368 case T_INT:
2369 // Move int and do sign extension.
2370 int_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2371 break;
2372 case T_LONG:
2373 long_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2374 break;
2375 case T_ARRAY:
2376 case T_OBJECT:
2377 object_move(masm, stack_slots,
2378 oop_map, oop_handle_slot_offset,
2379 ((in == 0) && (!method_is_static)), &receiver_offset,
2380 in_regs[in], out_regs[out],
2381 r_callers_sp, r_temp_1, r_temp_2);
2382 break;
2383 case T_VOID:
2384 break;
2385 case T_FLOAT:
2386 float_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2387 break;
2388 case T_DOUBLE:
2389 double_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2390 break;
2391 case T_ADDRESS:
2392 fatal("found type (T_ADDRESS) in java args");
2393 break;
2394 default:
2395 ShouldNotReachHere();
2396 break;
2397 }
2398 }
2399
2400 // Pre-load a static method's oop into ARG2.
2401 // Used both by locking code and the normal JNI call code.
2402 if (method_is_static) {
2403 __ set_oop_constant(JNIHandles::make_local(method->method_holder()->java_mirror()),
2404 r_carg2_classorobject);
2405
2406 // Now handlize the static class mirror in carg2. It's known not-null.
2407 __ std(r_carg2_classorobject, klass_offset, R1_SP);
2408 oop_map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
2409 __ addi(r_carg2_classorobject, R1_SP, klass_offset);
2410 }
2411
2412 // Get JNIEnv* which is first argument to native.
2413 __ addi(r_carg1_jnienv, R16_thread, in_bytes(JavaThread::jni_environment_offset()));
2414
2415 // NOTE:
2416 //
2417 // We have all of the arguments setup at this point.
2418 // We MUST NOT touch any outgoing regs from this point on.
2419 // So if we must call out we must push a new frame.
2420
2421 // The last java pc will also be used as resume pc if this is the wrapper for wait0.
2422 // For this purpose the precise location matters but not for oopmap lookup.
2423 __ calculate_address_from_global_toc(r_last_java_pc, last_java_pc, true, true, true, true);
2424
2425 // Make sure that thread is non-volatile; it crosses a bunch of VM calls below.
2426 assert(R16_thread->is_nonvolatile(), "thread must be in non-volatile register");
2427
2428 # if 0
2429 // DTrace method entry
2430 # endif
2431
2432 // Lock a synchronized method.
2433 // --------------------------------------------------------------------------
2434
2435 if (method->is_synchronized()) {
2436 Register r_oop = r_temp_4;
2437 const Register r_box = r_temp_5;
2438 Label done, locked;
2439
2440 // Load the oop for the object or class. r_carg2_classorobject contains
2441 // either the handlized oop from the incoming arguments or the handlized
2442 // class mirror (if the method is static).
2443 __ ld(r_oop, 0, r_carg2_classorobject);
2444
2445 // Get the lock box slot's address.
2446 __ addi(r_box, R1_SP, lock_offset);
2447
2448 // Try fastpath for locking.
2449 if (LockingMode == LM_LIGHTWEIGHT) {
2450 // fast_lock kills r_temp_1, r_temp_2, r_temp_3.
2451 Register r_temp_3_or_noreg = UseObjectMonitorTable ? r_temp_3 : noreg;
2452 __ compiler_fast_lock_lightweight_object(CR0, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3_or_noreg);
2453 } else {
2454 // fast_lock kills r_temp_1, r_temp_2, r_temp_3.
2455 __ compiler_fast_lock_object(CR0, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3);
2456 }
2457 __ beq(CR0, locked);
2458
2459 // None of the above fast optimizations worked so we have to get into the
2460 // slow case of monitor enter. Inline a special case of call_VM that
2461 // disallows any pending_exception.
2462
2463 // Save argument registers and leave room for C-compatible ABI_REG_ARGS.
2464 int frame_size = frame::native_abi_reg_args_size + align_up(total_c_args * wordSize, frame::alignment_in_bytes);
2465 __ mr(R11_scratch1, R1_SP);
2466 RegisterSaver::push_frame_and_save_argument_registers(masm, R12_scratch2, frame_size, total_c_args, out_regs);
2467
2468 // Do the call.
2469 __ set_last_Java_frame(R11_scratch1, r_last_java_pc);
2470 assert(r_last_java_pc->is_nonvolatile(), "r_last_java_pc needs to be preserved accross complete_monitor_locking_C call");
2471 // The following call will not be preempted.
2472 // push_cont_fastpath forces freeze slow path in case we try to preempt where we will pin the
2473 // vthread to the carrier (see FreezeBase::recurse_freeze_native_frame()).
2474 __ push_cont_fastpath();
2475 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), r_oop, r_box, R16_thread);
2476 __ pop_cont_fastpath();
2477 __ reset_last_Java_frame();
2478
2479 RegisterSaver::restore_argument_registers_and_pop_frame(masm, frame_size, total_c_args, out_regs);
2480
2481 __ asm_assert_mem8_is_zero(thread_(pending_exception),
2482 "no pending exception allowed on exit from SharedRuntime::complete_monitor_locking_C");
2483
2484 __ bind(locked);
2485 }
2486
2487 __ set_last_Java_frame(R1_SP, r_last_java_pc);
2488
2489 // Publish thread state
2490 // --------------------------------------------------------------------------
2491
2492 // Transition from _thread_in_Java to _thread_in_native.
2493 __ li(R0, _thread_in_native);
2494 __ release();
2495 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2496 __ stw(R0, thread_(thread_state));
2497
2498
2499 // The JNI call
2500 // --------------------------------------------------------------------------
2501 __ call_c(native_func, relocInfo::runtime_call_type);
2502
2503
2504 // Now, we are back from the native code.
2505
2506
2507 // Unpack the native result.
2508 // --------------------------------------------------------------------------
2509
2510 // For int-types, we do any needed sign-extension required.
2511 // Care must be taken that the return values (R3_RET and F1_RET)
2512 // will survive any VM calls for blocking or unlocking.
2513 // An OOP result (handle) is done specially in the slow-path code.
2514
2515 switch (ret_type) {
2516 case T_VOID: break; // Nothing to do!
2517 case T_FLOAT: break; // Got it where we want it (unless slow-path).
2518 case T_DOUBLE: break; // Got it where we want it (unless slow-path).
2519 case T_LONG: break; // Got it where we want it (unless slow-path).
2520 case T_OBJECT: break; // Really a handle.
2521 // Cannot de-handlize until after reclaiming jvm_lock.
2522 case T_ARRAY: break;
2523
2524 case T_BOOLEAN: { // 0 -> false(0); !0 -> true(1)
2525 __ normalize_bool(R3_RET);
2526 break;
2527 }
2528 case T_BYTE: { // sign extension
2529 __ extsb(R3_RET, R3_RET);
2530 break;
2531 }
2532 case T_CHAR: { // unsigned result
2533 __ andi(R3_RET, R3_RET, 0xffff);
2534 break;
2535 }
2536 case T_SHORT: { // sign extension
2537 __ extsh(R3_RET, R3_RET);
2538 break;
2539 }
2540 case T_INT: // nothing to do
2541 break;
2542 default:
2543 ShouldNotReachHere();
2544 break;
2545 }
2546
2547 // Publish thread state
2548 // --------------------------------------------------------------------------
2549
2550 // Switch thread to "native transition" state before reading the
2551 // synchronization state. This additional state is necessary because reading
2552 // and testing the synchronization state is not atomic w.r.t. GC, as this
2553 // scenario demonstrates:
2554 // - Java thread A, in _thread_in_native state, loads _not_synchronized
2555 // and is preempted.
2556 // - VM thread changes sync state to synchronizing and suspends threads
2557 // for GC.
2558 // - Thread A is resumed to finish this native method, but doesn't block
2559 // here since it didn't see any synchronization in progress, and escapes.
2560
2561 // Transition from _thread_in_native to _thread_in_native_trans.
2562 __ li(R0, _thread_in_native_trans);
2563 __ release();
2564 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2565 __ stw(R0, thread_(thread_state));
2566
2567
2568 // Must we block?
2569 // --------------------------------------------------------------------------
2570
2571 // Block, if necessary, before resuming in _thread_in_Java state.
2572 // In order for GC to work, don't clear the last_Java_sp until after blocking.
2573 {
2574 Label no_block, sync;
2575
2576 // Force this write out before the read below.
2577 if (!UseSystemMemoryBarrier) {
2578 __ fence();
2579 }
2580
2581 Register sync_state_addr = r_temp_4;
2582 Register sync_state = r_temp_5;
2583 Register suspend_flags = r_temp_6;
2584
2585 // No synchronization in progress nor yet synchronized
2586 // (cmp-br-isync on one path, release (same as acquire on PPC64) on the other path).
2587 __ safepoint_poll(sync, sync_state, true /* at_return */, false /* in_nmethod */);
2588
2589 // Not suspended.
2590 // TODO: PPC port assert(4 == Thread::sz_suspend_flags(), "unexpected field size");
2591 __ lwz(suspend_flags, thread_(suspend_flags));
2592 __ cmpwi(CR1, suspend_flags, 0);
2593 __ beq(CR1, no_block);
2594
2595 // Block. Save any potential method result value before the operation and
2596 // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this
2597 // lets us share the oopMap we used when we went native rather than create
2598 // a distinct one for this pc.
2599 __ bind(sync);
2600 __ isync();
2601
2602 address entry_point =
2603 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans);
2604 save_native_result(masm, ret_type, workspace_slot_offset);
2605 __ call_VM_leaf(entry_point, R16_thread);
2606 restore_native_result(masm, ret_type, workspace_slot_offset);
2607
2608 __ bind(no_block);
2609
2610 // Publish thread state.
2611 // --------------------------------------------------------------------------
2612
2613 // Thread state is thread_in_native_trans. Any safepoint blocking has
2614 // already happened so we can now change state to _thread_in_Java.
2615
2616 // Transition from _thread_in_native_trans to _thread_in_Java.
2617 __ li(R0, _thread_in_Java);
2618 __ lwsync(); // Acquire safepoint and suspend state, release thread state.
2619 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2620 __ stw(R0, thread_(thread_state));
2621
2622 // Check preemption for Object.wait()
2623 if (LockingMode != LM_LEGACY && method->is_object_wait0()) {
2624 Label not_preempted;
2625 __ ld(R0, in_bytes(JavaThread::preempt_alternate_return_offset()), R16_thread);
2626 __ cmpdi(CR0, R0, 0);
2627 __ beq(CR0, not_preempted);
2628 __ mtlr(R0);
2629 __ li(R0, 0);
2630 __ std(R0, in_bytes(JavaThread::preempt_alternate_return_offset()), R16_thread);
2631 __ blr();
2632 __ bind(not_preempted);
2633 }
2634 __ bind(last_java_pc);
2635 // We use the same pc/oopMap repeatedly when we call out above.
2636 intptr_t oopmap_pc = (intptr_t) __ pc();
2637 oop_maps->add_gc_map(oopmap_pc - start_pc, oop_map);
2638 }
2639
2640 // Reguard any pages if necessary.
2641 // --------------------------------------------------------------------------
2642
2643 Label no_reguard;
2644 __ lwz(r_temp_1, thread_(stack_guard_state));
2645 __ cmpwi(CR0, r_temp_1, StackOverflow::stack_guard_yellow_reserved_disabled);
2646 __ bne(CR0, no_reguard);
2647
2648 save_native_result(masm, ret_type, workspace_slot_offset);
2649 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2650 restore_native_result(masm, ret_type, workspace_slot_offset);
2651
2652 __ bind(no_reguard);
2653
2654
2655 // Unlock
2656 // --------------------------------------------------------------------------
2657
2658 if (method->is_synchronized()) {
2659 const Register r_oop = r_temp_4;
2660 const Register r_box = r_temp_5;
2661 const Register r_exception = r_temp_6;
2662 Label done;
2663
2664 // Get oop and address of lock object box.
2665 if (method_is_static) {
2666 assert(klass_offset != -1, "");
2667 __ ld(r_oop, klass_offset, R1_SP);
2668 } else {
2669 assert(receiver_offset != -1, "");
2670 __ ld(r_oop, receiver_offset, R1_SP);
2671 }
2672 __ addi(r_box, R1_SP, lock_offset);
2673
2674 // Try fastpath for unlocking.
2675 if (LockingMode == LM_LIGHTWEIGHT) {
2676 __ compiler_fast_unlock_lightweight_object(CR0, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3);
2677 } else {
2678 __ compiler_fast_unlock_object(CR0, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3);
2679 }
2680 __ beq(CR0, done);
2681
2682 // Save and restore any potential method result value around the unlocking operation.
2683 save_native_result(masm, ret_type, workspace_slot_offset);
2684
2685 // Must save pending exception around the slow-path VM call. Since it's a
2686 // leaf call, the pending exception (if any) can be kept in a register.
2687 __ ld(r_exception, thread_(pending_exception));
2688 assert(r_exception->is_nonvolatile(), "exception register must be non-volatile");
2689 __ li(R0, 0);
2690 __ std(R0, thread_(pending_exception));
2691
2692 // Slow case of monitor enter.
2693 // Inline a special case of call_VM that disallows any pending_exception.
2694 // Arguments are (oop obj, BasicLock* lock, JavaThread* thread).
2695 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), r_oop, r_box, R16_thread);
2696
2697 __ asm_assert_mem8_is_zero(thread_(pending_exception),
2698 "no pending exception allowed on exit from SharedRuntime::complete_monitor_unlocking_C");
2699
2700 restore_native_result(masm, ret_type, workspace_slot_offset);
2701
2702 // Check_forward_pending_exception jump to forward_exception if any pending
2703 // exception is set. The forward_exception routine expects to see the
2704 // exception in pending_exception and not in a register. Kind of clumsy,
2705 // since all folks who branch to forward_exception must have tested
2706 // pending_exception first and hence have it in a register already.
2707 __ std(r_exception, thread_(pending_exception));
2708
2709 __ bind(done);
2710 }
2711
2712 # if 0
2713 // DTrace method exit
2714 # endif
2715
2716 // Clear "last Java frame" SP and PC.
2717 // --------------------------------------------------------------------------
2718
2719 // Last java frame won't be set if we're resuming after preemption
2720 bool maybe_preempted = LockingMode != LM_LEGACY && method->is_object_wait0();
2721 __ reset_last_Java_frame(!maybe_preempted /* check_last_java_sp */);
2722
2723 // Unbox oop result, e.g. JNIHandles::resolve value.
2724 // --------------------------------------------------------------------------
2725
2726 if (is_reference_type(ret_type)) {
2727 __ resolve_jobject(R3_RET, r_temp_1, r_temp_2, MacroAssembler::PRESERVATION_NONE);
2728 }
2729
2730 if (CheckJNICalls) {
2731 // clear_pending_jni_exception_check
2732 __ load_const_optimized(R0, 0L);
2733 __ st_ptr(R0, JavaThread::pending_jni_exception_check_fn_offset(), R16_thread);
2734 }
2735
2736 // Reset handle block.
2737 // --------------------------------------------------------------------------
2738 __ ld(r_temp_1, thread_(active_handles));
2739 // TODO: PPC port assert(4 == JNIHandleBlock::top_size_in_bytes(), "unexpected field size");
2740 __ li(r_temp_2, 0);
2741 __ stw(r_temp_2, in_bytes(JNIHandleBlock::top_offset()), r_temp_1);
2742
2743 // Prepare for return
2744 // --------------------------------------------------------------------------
2745 __ pop_frame();
2746 __ restore_LR(R11);
2747
2748 #if INCLUDE_JFR
2749 // We need to do a poll test after unwind in case the sampler
2750 // managed to sample the native frame after returning to Java.
2751 Label L_stub;
2752 int safepoint_offset = __ offset();
2753 if (!UseSIGTRAP) {
2754 __ relocate(relocInfo::poll_return_type);
2755 }
2756 __ safepoint_poll(L_stub, r_temp_2, true /* at_return */, true /* in_nmethod: frame already popped */);
2757 #endif // INCLUDE_JFR
2758
2759 // Check for pending exceptions.
2760 // --------------------------------------------------------------------------
2761 __ ld(r_temp_2, thread_(pending_exception));
2762 __ cmpdi(CR0, r_temp_2, 0);
2763 __ bne(CR0, handle_pending_exception);
2764
2765 // Return.
2766 __ blr();
2767
2768 // Handler for return safepoint (out-of-line).
2769 #if INCLUDE_JFR
2770 if (!UseSIGTRAP) {
2771 __ bind(L_stub);
2772 __ jump_to_polling_page_return_handler_blob(safepoint_offset);
2773 }
2774 #endif // INCLUDE_JFR
2775
2776 // Handler for pending exceptions (out-of-line).
2777 // --------------------------------------------------------------------------
2778 // Since this is a native call, we know the proper exception handler
2779 // is the empty function. We just pop this frame and then jump to
2780 // forward_exception_entry.
2781 __ bind(handle_pending_exception);
2782 __ b64_patchable((address)StubRoutines::forward_exception_entry(),
2783 relocInfo::runtime_call_type);
2784
2785 // Done.
2786 // --------------------------------------------------------------------------
2787
2788 __ flush();
2789
2790 nmethod *nm = nmethod::new_native_nmethod(method,
2791 compile_id,
2792 masm->code(),
2793 vep_start_pc-start_pc,
2794 frame_done_pc-start_pc,
2795 stack_slots / VMRegImpl::slots_per_word,
2796 (method_is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2797 in_ByteSize(lock_offset),
2798 oop_maps);
2799
2800 return nm;
2801 }
2802
2803 // This function returns the adjust size (in number of words) to a c2i adapter
2804 // activation for use during deoptimization.
2805 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2806 return align_up((callee_locals - callee_parameters) * Interpreter::stackElementWords, frame::frame_alignment_in_words);
2807 }
2808
2809 uint SharedRuntime::in_preserve_stack_slots() {
2810 return frame::jit_in_preserve_size / VMRegImpl::stack_slot_size;
2811 }
2812
2813 uint SharedRuntime::out_preserve_stack_slots() {
2814 #if defined(COMPILER1) || defined(COMPILER2)
2815 return frame::jit_out_preserve_size / VMRegImpl::stack_slot_size;
2816 #else
2817 return 0;
2818 #endif
2819 }
2820
2821 VMReg SharedRuntime::thread_register() {
2822 // On PPC virtual threads don't save the JavaThread* in their context (e.g. C1 stub frames).
2823 ShouldNotCallThis();
2824 return nullptr;
2825 }
2826
2827 #if defined(COMPILER1) || defined(COMPILER2)
2828 // Frame generation for deopt and uncommon trap blobs.
2829 static void push_skeleton_frame(MacroAssembler* masm, bool deopt,
2830 /* Read */
2831 Register unroll_block_reg,
2832 /* Update */
2833 Register frame_sizes_reg,
2834 Register number_of_frames_reg,
2835 Register pcs_reg,
2836 /* Invalidate */
2837 Register frame_size_reg,
2838 Register pc_reg) {
2839
2840 __ ld(pc_reg, 0, pcs_reg);
2841 __ ld(frame_size_reg, 0, frame_sizes_reg);
2842 __ std(pc_reg, _abi0(lr), R1_SP);
2843 __ push_frame(frame_size_reg, R0/*tmp*/);
2844 __ std(R1_SP, _ijava_state_neg(sender_sp), R1_SP);
2845 __ addi(number_of_frames_reg, number_of_frames_reg, -1);
2846 __ addi(frame_sizes_reg, frame_sizes_reg, wordSize);
2847 __ addi(pcs_reg, pcs_reg, wordSize);
2848 }
2849
2850 // Loop through the UnrollBlock info and create new frames.
2851 static void push_skeleton_frames(MacroAssembler* masm, bool deopt,
2852 /* read */
2853 Register unroll_block_reg,
2854 /* invalidate */
2855 Register frame_sizes_reg,
2856 Register number_of_frames_reg,
2857 Register pcs_reg,
2858 Register frame_size_reg,
2859 Register pc_reg) {
2860 Label loop;
2861
2862 // _number_of_frames is of type int (deoptimization.hpp)
2863 __ lwa(number_of_frames_reg,
2864 in_bytes(Deoptimization::UnrollBlock::number_of_frames_offset()),
2865 unroll_block_reg);
2866 __ ld(pcs_reg,
2867 in_bytes(Deoptimization::UnrollBlock::frame_pcs_offset()),
2868 unroll_block_reg);
2869 __ ld(frame_sizes_reg,
2870 in_bytes(Deoptimization::UnrollBlock::frame_sizes_offset()),
2871 unroll_block_reg);
2872
2873 // stack: (caller_of_deoptee, ...).
2874
2875 // At this point we either have an interpreter frame or a compiled
2876 // frame on top of stack. If it is a compiled frame we push a new c2i
2877 // adapter here
2878
2879 // Memorize top-frame stack-pointer.
2880 __ mr(frame_size_reg/*old_sp*/, R1_SP);
2881
2882 // Resize interpreter top frame OR C2I adapter.
2883
2884 // At this moment, the top frame (which is the caller of the deoptee) is
2885 // an interpreter frame or a newly pushed C2I adapter or an entry frame.
2886 // The top frame has a TOP_IJAVA_FRAME_ABI and the frame contains the
2887 // outgoing arguments.
2888 //
2889 // In order to push the interpreter frame for the deoptee, we need to
2890 // resize the top frame such that we are able to place the deoptee's
2891 // locals in the frame.
2892 // Additionally, we have to turn the top frame's TOP_IJAVA_FRAME_ABI
2893 // into a valid PARENT_IJAVA_FRAME_ABI.
2894
2895 __ lwa(R11_scratch1,
2896 in_bytes(Deoptimization::UnrollBlock::caller_adjustment_offset()),
2897 unroll_block_reg);
2898 __ neg(R11_scratch1, R11_scratch1);
2899
2900 // R11_scratch1 contains size of locals for frame resizing.
2901 // R12_scratch2 contains top frame's lr.
2902
2903 // Resize frame by complete frame size prevents TOC from being
2904 // overwritten by locals. A more stack space saving way would be
2905 // to copy the TOC to its location in the new abi.
2906 __ addi(R11_scratch1, R11_scratch1, - frame::parent_ijava_frame_abi_size);
2907
2908 // now, resize the frame
2909 __ resize_frame(R11_scratch1, pc_reg/*tmp*/);
2910
2911 // In the case where we have resized a c2i frame above, the optional
2912 // alignment below the locals has size 32 (why?).
2913 __ std(R12_scratch2, _abi0(lr), R1_SP);
2914
2915 // Initialize initial_caller_sp.
2916 __ std(frame_size_reg, _ijava_state_neg(sender_sp), R1_SP);
2917
2918 #ifdef ASSERT
2919 // Make sure that there is at least one entry in the array.
2920 __ cmpdi(CR0, number_of_frames_reg, 0);
2921 __ asm_assert_ne("array_size must be > 0");
2922 #endif
2923
2924 // Now push the new interpreter frames.
2925 //
2926 __ bind(loop);
2927 // Allocate a new frame, fill in the pc.
2928 push_skeleton_frame(masm, deopt,
2929 unroll_block_reg,
2930 frame_sizes_reg,
2931 number_of_frames_reg,
2932 pcs_reg,
2933 frame_size_reg,
2934 pc_reg);
2935 __ cmpdi(CR0, number_of_frames_reg, 0);
2936 __ bne(CR0, loop);
2937
2938 // Get the return address pointing into the template interpreter.
2939 __ ld(R0, 0, pcs_reg);
2940 // Store it in the top interpreter frame.
2941 __ std(R0, _abi0(lr), R1_SP);
2942 // Initialize frame_manager_lr of interpreter top frame.
2943 }
2944 #endif
2945
2946 void SharedRuntime::generate_deopt_blob() {
2947 // Allocate space for the code
2948 ResourceMark rm;
2949 // Setup code generation tools
2950 const char* name = SharedRuntime::stub_name(SharedStubId::deopt_id);
2951 CodeBuffer buffer(name, 2048, 1024);
2952 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer);
2953 Label exec_mode_initialized;
2954 int frame_size_in_words;
2955 OopMap* map = nullptr;
2956 OopMapSet *oop_maps = new OopMapSet();
2957
2958 // size of ABI112 plus spill slots for R3_RET and F1_RET.
2959 const int frame_size_in_bytes = frame::native_abi_reg_args_spill_size;
2960 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint);
2961 int first_frame_size_in_bytes = 0; // frame size of "unpack frame" for call to fetch_unroll_info.
2962
2963 const Register exec_mode_reg = R21_tmp1;
2964
2965 const address start = __ pc();
2966
2967 #if defined(COMPILER1) || defined(COMPILER2)
2968 // --------------------------------------------------------------------------
2969 // Prolog for non exception case!
2970
2971 // We have been called from the deopt handler of the deoptee.
2972 //
2973 // deoptee:
2974 // ...
2975 // call X
2976 // ...
2977 // deopt_handler: call_deopt_stub
2978 // cur. return pc --> ...
2979 //
2980 // So currently SR_LR points behind the call in the deopt handler.
2981 // We adjust it such that it points to the start of the deopt handler.
2982 // The return_pc has been stored in the frame of the deoptee and
2983 // will replace the address of the deopt_handler in the call
2984 // to Deoptimization::fetch_unroll_info below.
2985 // We can't grab a free register here, because all registers may
2986 // contain live values, so let the RegisterSaver do the adjustment
2987 // of the return pc.
2988 const int return_pc_adjustment_no_exception = -MacroAssembler::bl64_patchable_size;
2989
2990 // Push the "unpack frame"
2991 // Save everything in sight.
2992 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
2993 &first_frame_size_in_bytes,
2994 /*generate_oop_map=*/ true,
2995 return_pc_adjustment_no_exception,
2996 RegisterSaver::return_pc_is_lr);
2997 assert(map != nullptr, "OopMap must have been created");
2998
2999 __ li(exec_mode_reg, Deoptimization::Unpack_deopt);
3000 // Save exec mode for unpack_frames.
3001 __ b(exec_mode_initialized);
3002
3003 // --------------------------------------------------------------------------
3004 // Prolog for exception case
3005
3006 // An exception is pending.
3007 // We have been called with a return (interpreter) or a jump (exception blob).
3008 //
3009 // - R3_ARG1: exception oop
3010 // - R4_ARG2: exception pc
3011
3012 int exception_offset = __ pc() - start;
3013
3014 BLOCK_COMMENT("Prolog for exception case");
3015
3016 // Store exception oop and pc in thread (location known to GC).
3017 // This is needed since the call to "fetch_unroll_info()" may safepoint.
3018 __ std(R3_ARG1, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
3019 __ std(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
3020 __ std(R4_ARG2, _abi0(lr), R1_SP);
3021
3022 // Vanilla deoptimization with an exception pending in exception_oop.
3023 int exception_in_tls_offset = __ pc() - start;
3024
3025 // Push the "unpack frame".
3026 // Save everything in sight.
3027 RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
3028 &first_frame_size_in_bytes,
3029 /*generate_oop_map=*/ false,
3030 /*return_pc_adjustment_exception=*/ 0,
3031 RegisterSaver::return_pc_is_pre_saved);
3032
3033 // Deopt during an exception. Save exec mode for unpack_frames.
3034 __ li(exec_mode_reg, Deoptimization::Unpack_exception);
3035
3036 // fall through
3037
3038 int reexecute_offset = 0;
3039 #ifdef COMPILER1
3040 __ b(exec_mode_initialized);
3041
3042 // Reexecute entry, similar to c2 uncommon trap
3043 reexecute_offset = __ pc() - start;
3044
3045 RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
3046 &first_frame_size_in_bytes,
3047 /*generate_oop_map=*/ false,
3048 /*return_pc_adjustment_reexecute=*/ 0,
3049 RegisterSaver::return_pc_is_pre_saved);
3050 __ li(exec_mode_reg, Deoptimization::Unpack_reexecute);
3051 #endif
3052
3053 // --------------------------------------------------------------------------
3054 __ BIND(exec_mode_initialized);
3055
3056 const Register unroll_block_reg = R22_tmp2;
3057
3058 // We need to set `last_Java_frame' because `fetch_unroll_info' will
3059 // call `last_Java_frame()'. The value of the pc in the frame is not
3060 // particularly important. It just needs to identify this blob.
3061 __ set_last_Java_frame(R1_SP, noreg);
3062
3063 // With EscapeAnalysis turned on, this call may safepoint!
3064 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), R16_thread, exec_mode_reg);
3065 address calls_return_pc = __ last_calls_return_pc();
3066 // Set an oopmap for the call site that describes all our saved registers.
3067 oop_maps->add_gc_map(calls_return_pc - start, map);
3068
3069 __ reset_last_Java_frame();
3070 // Save the return value.
3071 __ mr(unroll_block_reg, R3_RET);
3072
3073 // Restore only the result registers that have been saved
3074 // by save_volatile_registers(...).
3075 RegisterSaver::restore_result_registers(masm, first_frame_size_in_bytes);
3076
3077 // reload the exec mode from the UnrollBlock (it might have changed)
3078 __ lwz(exec_mode_reg, in_bytes(Deoptimization::UnrollBlock::unpack_kind_offset()), unroll_block_reg);
3079 // In excp_deopt_mode, restore and clear exception oop which we
3080 // stored in the thread during exception entry above. The exception
3081 // oop will be the return value of this stub.
3082 Label skip_restore_excp;
3083 __ cmpdi(CR0, exec_mode_reg, Deoptimization::Unpack_exception);
3084 __ bne(CR0, skip_restore_excp);
3085 __ ld(R3_RET, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
3086 __ ld(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
3087 __ li(R0, 0);
3088 __ std(R0, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
3089 __ std(R0, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
3090 __ BIND(skip_restore_excp);
3091
3092 __ pop_frame();
3093
3094 // stack: (deoptee, optional i2c, caller of deoptee, ...).
3095
3096 // pop the deoptee's frame
3097 __ pop_frame();
3098
3099 // stack: (caller_of_deoptee, ...).
3100
3101 // Freezing continuation frames requires that the caller is trimmed to unextended sp if compiled.
3102 // If not compiled the loaded value is equal to the current SP (see frame::initial_deoptimization_info())
3103 // and the frame is effectively not resized.
3104 Register caller_sp = R23_tmp3;
3105 __ ld_ptr(caller_sp, Deoptimization::UnrollBlock::initial_info_offset(), unroll_block_reg);
3106 __ resize_frame_absolute(caller_sp, R24_tmp4, R25_tmp5);
3107
3108 // Loop through the `UnrollBlock' info and create interpreter frames.
3109 push_skeleton_frames(masm, true/*deopt*/,
3110 unroll_block_reg,
3111 R23_tmp3,
3112 R24_tmp4,
3113 R25_tmp5,
3114 R26_tmp6,
3115 R27_tmp7);
3116
3117 // stack: (skeletal interpreter frame, ..., optional skeletal
3118 // interpreter frame, optional c2i, caller of deoptee, ...).
3119
3120 // push an `unpack_frame' taking care of float / int return values.
3121 __ push_frame(frame_size_in_bytes, R0/*tmp*/);
3122
3123 // stack: (unpack frame, skeletal interpreter frame, ..., optional
3124 // skeletal interpreter frame, optional c2i, caller of deoptee,
3125 // ...).
3126
3127 // Spill live volatile registers since we'll do a call.
3128 __ std( R3_RET, _native_abi_reg_args_spill(spill_ret), R1_SP);
3129 __ stfd(F1_RET, _native_abi_reg_args_spill(spill_fret), R1_SP);
3130
3131 // Let the unpacker layout information in the skeletal frames just
3132 // allocated.
3133 __ calculate_address_from_global_toc(R3_RET, calls_return_pc, true, true, true, true);
3134 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R3_RET);
3135 // This is a call to a LEAF method, so no oop map is required.
3136 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames),
3137 R16_thread/*thread*/, exec_mode_reg/*exec_mode*/);
3138 __ reset_last_Java_frame();
3139
3140 // Restore the volatiles saved above.
3141 __ ld( R3_RET, _native_abi_reg_args_spill(spill_ret), R1_SP);
3142 __ lfd(F1_RET, _native_abi_reg_args_spill(spill_fret), R1_SP);
3143
3144 // Pop the unpack frame.
3145 __ pop_frame();
3146 __ restore_LR(R0);
3147
3148 // stack: (top interpreter frame, ..., optional interpreter frame,
3149 // optional c2i, caller of deoptee, ...).
3150
3151 // Initialize R14_state.
3152 __ restore_interpreter_state(R11_scratch1);
3153 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
3154
3155 // Return to the interpreter entry point.
3156 __ blr();
3157 __ flush();
3158 #else // COMPILER2
3159 __ unimplemented("deopt blob needed only with compiler");
3160 int exception_offset = __ pc() - start;
3161 #endif // COMPILER2
3162
3163 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset,
3164 reexecute_offset, first_frame_size_in_bytes / wordSize);
3165 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
3166 }
3167
3168 #ifdef COMPILER2
3169 UncommonTrapBlob* OptoRuntime::generate_uncommon_trap_blob() {
3170 // Allocate space for the code.
3171 ResourceMark rm;
3172 // Setup code generation tools.
3173 const char* name = OptoRuntime::stub_name(OptoStubId::uncommon_trap_id);
3174 CodeBuffer buffer(name, 2048, 1024);
3175 if (buffer.blob() == nullptr) {
3176 return nullptr;
3177 }
3178 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer);
3179 address start = __ pc();
3180
3181 Register unroll_block_reg = R21_tmp1;
3182 Register klass_index_reg = R22_tmp2;
3183 Register unc_trap_reg = R23_tmp3;
3184 Register r_return_pc = R27_tmp7;
3185
3186 OopMapSet* oop_maps = new OopMapSet();
3187 int frame_size_in_bytes = frame::native_abi_reg_args_size;
3188 OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0);
3189
3190 // stack: (deoptee, optional i2c, caller_of_deoptee, ...).
3191
3192 // Push a dummy `unpack_frame' and call
3193 // `Deoptimization::uncommon_trap' to pack the compiled frame into a
3194 // vframe array and return the `UnrollBlock' information.
3195
3196 // Save LR to compiled frame.
3197 __ save_LR(R11_scratch1);
3198
3199 // Push an "uncommon_trap" frame.
3200 __ push_frame_reg_args(0, R11_scratch1);
3201
3202 // stack: (unpack frame, deoptee, optional i2c, caller_of_deoptee, ...).
3203
3204 // Set the `unpack_frame' as last_Java_frame.
3205 // `Deoptimization::uncommon_trap' expects it and considers its
3206 // sender frame as the deoptee frame.
3207 // Remember the offset of the instruction whose address will be
3208 // moved to R11_scratch1.
3209 address gc_map_pc = __ pc();
3210 __ calculate_address_from_global_toc(r_return_pc, gc_map_pc, true, true, true, true);
3211 __ set_last_Java_frame(/*sp*/R1_SP, r_return_pc);
3212
3213 __ mr(klass_index_reg, R3);
3214 __ li(R5_ARG3, Deoptimization::Unpack_uncommon_trap);
3215 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap),
3216 R16_thread, klass_index_reg, R5_ARG3);
3217
3218 // Set an oopmap for the call site.
3219 oop_maps->add_gc_map(gc_map_pc - start, map);
3220
3221 __ reset_last_Java_frame();
3222
3223 // Pop the `unpack frame'.
3224 __ pop_frame();
3225
3226 // stack: (deoptee, optional i2c, caller_of_deoptee, ...).
3227
3228 // Save the return value.
3229 __ mr(unroll_block_reg, R3_RET);
3230
3231 // Pop the uncommon_trap frame.
3232 __ pop_frame();
3233
3234 // stack: (caller_of_deoptee, ...).
3235
3236 #ifdef ASSERT
3237 __ lwz(R22_tmp2, in_bytes(Deoptimization::UnrollBlock::unpack_kind_offset()), unroll_block_reg);
3238 __ cmpdi(CR0, R22_tmp2, (unsigned)Deoptimization::Unpack_uncommon_trap);
3239 __ asm_assert_eq("OptoRuntime::generate_uncommon_trap_blob: expected Unpack_uncommon_trap");
3240 #endif
3241
3242 // Freezing continuation frames requires that the caller is trimmed to unextended sp if compiled.
3243 // If not compiled the loaded value is equal to the current SP (see frame::initial_deoptimization_info())
3244 // and the frame is effectively not resized.
3245 Register caller_sp = R23_tmp3;
3246 __ ld_ptr(caller_sp, Deoptimization::UnrollBlock::initial_info_offset(), unroll_block_reg);
3247 __ resize_frame_absolute(caller_sp, R24_tmp4, R25_tmp5);
3248
3249 // Allocate new interpreter frame(s) and possibly a c2i adapter
3250 // frame.
3251 push_skeleton_frames(masm, false/*deopt*/,
3252 unroll_block_reg,
3253 R22_tmp2,
3254 R23_tmp3,
3255 R24_tmp4,
3256 R25_tmp5,
3257 R26_tmp6);
3258
3259 // stack: (skeletal interpreter frame, ..., optional skeletal
3260 // interpreter frame, optional c2i, caller of deoptee, ...).
3261
3262 // Push a dummy `unpack_frame' taking care of float return values.
3263 // Call `Deoptimization::unpack_frames' to layout information in the
3264 // interpreter frames just created.
3265
3266 // Push a simple "unpack frame" here.
3267 __ push_frame_reg_args(0, R11_scratch1);
3268
3269 // stack: (unpack frame, skeletal interpreter frame, ..., optional
3270 // skeletal interpreter frame, optional c2i, caller of deoptee,
3271 // ...).
3272
3273 // Set the "unpack_frame" as last_Java_frame.
3274 __ set_last_Java_frame(/*sp*/R1_SP, r_return_pc);
3275
3276 // Indicate it is the uncommon trap case.
3277 __ li(unc_trap_reg, Deoptimization::Unpack_uncommon_trap);
3278 // Let the unpacker layout information in the skeletal frames just
3279 // allocated.
3280 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames),
3281 R16_thread, unc_trap_reg);
3282
3283 __ reset_last_Java_frame();
3284 // Pop the `unpack frame'.
3285 __ pop_frame();
3286 // Restore LR from top interpreter frame.
3287 __ restore_LR(R11_scratch1);
3288
3289 // stack: (top interpreter frame, ..., optional interpreter frame,
3290 // optional c2i, caller of deoptee, ...).
3291
3292 __ restore_interpreter_state(R11_scratch1);
3293 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
3294
3295 // Return to the interpreter entry point.
3296 __ blr();
3297
3298 masm->flush();
3299
3300 return UncommonTrapBlob::create(&buffer, oop_maps, frame_size_in_bytes/wordSize);
3301 }
3302 #endif // COMPILER2
3303
3304 // Generate a special Compile2Runtime blob that saves all registers, and setup oopmap.
3305 SafepointBlob* SharedRuntime::generate_handler_blob(SharedStubId id, address call_ptr) {
3306 assert(StubRoutines::forward_exception_entry() != nullptr,
3307 "must be generated before");
3308 assert(is_polling_page_id(id), "expected a polling page stub id");
3309
3310 ResourceMark rm;
3311 OopMapSet *oop_maps = new OopMapSet();
3312 OopMap* map;
3313
3314 // Allocate space for the code. Setup code generation tools.
3315 const char* name = SharedRuntime::stub_name(id);
3316 CodeBuffer buffer(name, 2048, 1024);
3317 MacroAssembler* masm = new MacroAssembler(&buffer);
3318
3319 address start = __ pc();
3320 int frame_size_in_bytes = 0;
3321
3322 RegisterSaver::ReturnPCLocation return_pc_location;
3323 bool cause_return = (id == SharedStubId::polling_page_return_handler_id);
3324 if (cause_return) {
3325 // Nothing to do here. The frame has already been popped in MachEpilogNode.
3326 // Register LR already contains the return pc.
3327 return_pc_location = RegisterSaver::return_pc_is_pre_saved;
3328 } else {
3329 // Use thread()->saved_exception_pc() as return pc.
3330 return_pc_location = RegisterSaver::return_pc_is_thread_saved_exception_pc;
3331 }
3332
3333 bool save_vectors = (id == SharedStubId::polling_page_vectors_safepoint_handler_id);
3334
3335 // Save registers, fpu state, and flags. Set R31 = return pc.
3336 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
3337 &frame_size_in_bytes,
3338 /*generate_oop_map=*/ true,
3339 /*return_pc_adjustment=*/0,
3340 return_pc_location, save_vectors);
3341
3342 // The following is basically a call_VM. However, we need the precise
3343 // address of the call in order to generate an oopmap. Hence, we do all the
3344 // work ourselves.
3345 __ set_last_Java_frame(/*sp=*/R1_SP, /*pc=*/noreg);
3346
3347 // The return address must always be correct so that the frame constructor
3348 // never sees an invalid pc.
3349
3350 // Do the call
3351 __ call_VM_leaf(call_ptr, R16_thread);
3352 address calls_return_pc = __ last_calls_return_pc();
3353
3354 // Set an oopmap for the call site. This oopmap will map all
3355 // oop-registers and debug-info registers as callee-saved. This
3356 // will allow deoptimization at this safepoint to find all possible
3357 // debug-info recordings, as well as let GC find all oops.
3358 oop_maps->add_gc_map(calls_return_pc - start, map);
3359
3360 Label noException;
3361
3362 // Clear the last Java frame.
3363 __ reset_last_Java_frame();
3364
3365 BLOCK_COMMENT(" Check pending exception.");
3366 const Register pending_exception = R0;
3367 __ ld(pending_exception, thread_(pending_exception));
3368 __ cmpdi(CR0, pending_exception, 0);
3369 __ beq(CR0, noException);
3370
3371 // Exception pending
3372 RegisterSaver::restore_live_registers_and_pop_frame(masm,
3373 frame_size_in_bytes,
3374 /*restore_ctr=*/true, save_vectors);
3375
3376 BLOCK_COMMENT(" Jump to forward_exception_entry.");
3377 // Jump to forward_exception_entry, with the issuing PC in LR
3378 // so it looks like the original nmethod called forward_exception_entry.
3379 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
3380
3381 // No exception case.
3382 __ BIND(noException);
3383
3384 if (!cause_return) {
3385 Label no_adjust;
3386 // If our stashed return pc was modified by the runtime we avoid touching it
3387 __ ld(R0, frame_size_in_bytes + _abi0(lr), R1_SP);
3388 __ cmpd(CR0, R0, R31);
3389 __ bne(CR0, no_adjust);
3390
3391 // Adjust return pc forward to step over the safepoint poll instruction
3392 __ addi(R31, R31, 4);
3393 __ std(R31, frame_size_in_bytes + _abi0(lr), R1_SP);
3394
3395 __ bind(no_adjust);
3396 }
3397
3398 // Normal exit, restore registers and exit.
3399 RegisterSaver::restore_live_registers_and_pop_frame(masm,
3400 frame_size_in_bytes,
3401 /*restore_ctr=*/true, save_vectors);
3402
3403 __ blr();
3404
3405 // Make sure all code is generated
3406 masm->flush();
3407
3408 // Fill-out other meta info
3409 // CodeBlob frame size is in words.
3410 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_bytes / wordSize);
3411 }
3412
3413 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss)
3414 //
3415 // Generate a stub that calls into the vm to find out the proper destination
3416 // of a java call. All the argument registers are live at this point
3417 // but since this is generic code we don't know what they are and the caller
3418 // must do any gc of the args.
3419 //
3420 RuntimeStub* SharedRuntime::generate_resolve_blob(SharedStubId id, address destination) {
3421 assert(is_resolve_id(id), "expected a resolve stub id");
3422
3423 // allocate space for the code
3424 ResourceMark rm;
3425
3426 const char* name = SharedRuntime::stub_name(id);
3427 CodeBuffer buffer(name, 1000, 512);
3428 MacroAssembler* masm = new MacroAssembler(&buffer);
3429
3430 int frame_size_in_bytes;
3431
3432 OopMapSet *oop_maps = new OopMapSet();
3433 OopMap* map = nullptr;
3434
3435 address start = __ pc();
3436
3437 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
3438 &frame_size_in_bytes,
3439 /*generate_oop_map*/ true,
3440 /*return_pc_adjustment*/ 0,
3441 RegisterSaver::return_pc_is_lr);
3442
3443 // Use noreg as last_Java_pc, the return pc will be reconstructed
3444 // from the physical frame.
3445 __ set_last_Java_frame(/*sp*/R1_SP, noreg);
3446
3447 int frame_complete = __ offset();
3448
3449 // Pass R19_method as 2nd (optional) argument, used by
3450 // counter_overflow_stub.
3451 __ call_VM_leaf(destination, R16_thread, R19_method);
3452 address calls_return_pc = __ last_calls_return_pc();
3453 // Set an oopmap for the call site.
3454 // We need this not only for callee-saved registers, but also for volatile
3455 // registers that the compiler might be keeping live across a safepoint.
3456 // Create the oopmap for the call's return pc.
3457 oop_maps->add_gc_map(calls_return_pc - start, map);
3458
3459 // R3_RET contains the address we are going to jump to assuming no exception got installed.
3460
3461 // clear last_Java_sp
3462 __ reset_last_Java_frame();
3463
3464 // Check for pending exceptions.
3465 BLOCK_COMMENT("Check for pending exceptions.");
3466 Label pending;
3467 __ ld(R11_scratch1, thread_(pending_exception));
3468 __ cmpdi(CR0, R11_scratch1, 0);
3469 __ bne(CR0, pending);
3470
3471 __ mtctr(R3_RET); // Ctr will not be touched by restore_live_registers_and_pop_frame.
3472
3473 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ false);
3474
3475 // Get the returned method.
3476 __ get_vm_result_metadata(R19_method);
3477
3478 __ bctr();
3479
3480
3481 // Pending exception after the safepoint.
3482 __ BIND(pending);
3483
3484 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ true);
3485
3486 // exception pending => remove activation and forward to exception handler
3487
3488 __ li(R11_scratch1, 0);
3489 __ ld(R3_ARG1, thread_(pending_exception));
3490 __ std(R11_scratch1, in_bytes(JavaThread::vm_result_oop_offset()), R16_thread);
3491 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
3492
3493 // -------------
3494 // Make sure all code is generated.
3495 masm->flush();
3496
3497 // return the blob
3498 // frame_size_words or bytes??
3499 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_bytes/wordSize,
3500 oop_maps, true);
3501 }
3502
3503 // Continuation point for throwing of implicit exceptions that are
3504 // not handled in the current activation. Fabricates an exception
3505 // oop and initiates normal exception dispatching in this
3506 // frame. Only callee-saved registers are preserved (through the
3507 // normal register window / RegisterMap handling). If the compiler
3508 // needs all registers to be preserved between the fault point and
3509 // the exception handler then it must assume responsibility for that
3510 // in AbstractCompiler::continuation_for_implicit_null_exception or
3511 // continuation_for_implicit_division_by_zero_exception. All other
3512 // implicit exceptions (e.g., NullPointerException or
3513 // AbstractMethodError on entry) are either at call sites or
3514 // otherwise assume that stack unwinding will be initiated, so
3515 // caller saved registers were assumed volatile in the compiler.
3516 //
3517 // Note that we generate only this stub into a RuntimeStub, because
3518 // it needs to be properly traversed and ignored during GC, so we
3519 // change the meaning of the "__" macro within this method.
3520 //
3521 // Note: the routine set_pc_not_at_call_for_caller in
3522 // SharedRuntime.cpp requires that this code be generated into a
3523 // RuntimeStub.
3524 RuntimeStub* SharedRuntime::generate_throw_exception(SharedStubId id, address runtime_entry) {
3525 assert(is_throw_id(id), "expected a throw stub id");
3526
3527 const char* name = SharedRuntime::stub_name(id);
3528
3529 ResourceMark rm;
3530 const char* timer_msg = "SharedRuntime generate_throw_exception";
3531 TraceTime timer(timer_msg, TRACETIME_LOG(Info, startuptime));
3532
3533 CodeBuffer code(name, 1024 DEBUG_ONLY(+ 512), 0);
3534 MacroAssembler* masm = new MacroAssembler(&code);
3535
3536 OopMapSet* oop_maps = new OopMapSet();
3537 int frame_size_in_bytes = frame::native_abi_reg_args_size;
3538 OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0);
3539
3540 address start = __ pc();
3541
3542 __ save_LR(R11_scratch1);
3543
3544 // Push a frame.
3545 __ push_frame_reg_args(0, R11_scratch1);
3546
3547 address frame_complete_pc = __ pc();
3548
3549 // Note that we always have a runtime stub frame on the top of
3550 // stack by this point. Remember the offset of the instruction
3551 // whose address will be moved to R11_scratch1.
3552 address gc_map_pc = __ get_PC_trash_LR(R11_scratch1);
3553
3554 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R11_scratch1);
3555
3556 __ mr(R3_ARG1, R16_thread);
3557 __ call_c(runtime_entry);
3558
3559 // Set an oopmap for the call site.
3560 oop_maps->add_gc_map((int)(gc_map_pc - start), map);
3561
3562 __ reset_last_Java_frame();
3563
3564 #ifdef ASSERT
3565 // Make sure that this code is only executed if there is a pending
3566 // exception.
3567 {
3568 Label L;
3569 __ ld(R0,
3570 in_bytes(Thread::pending_exception_offset()),
3571 R16_thread);
3572 __ cmpdi(CR0, R0, 0);
3573 __ bne(CR0, L);
3574 __ stop("SharedRuntime::throw_exception: no pending exception");
3575 __ bind(L);
3576 }
3577 #endif
3578
3579 // Pop frame.
3580 __ pop_frame();
3581
3582 __ restore_LR(R11_scratch1);
3583
3584 __ load_const(R11_scratch1, StubRoutines::forward_exception_entry());
3585 __ mtctr(R11_scratch1);
3586 __ bctr();
3587
3588 // Create runtime stub with OopMap.
3589 RuntimeStub* stub =
3590 RuntimeStub::new_runtime_stub(name, &code,
3591 /*frame_complete=*/ (int)(frame_complete_pc - start),
3592 frame_size_in_bytes/wordSize,
3593 oop_maps,
3594 false);
3595 return stub;
3596 }
3597
3598 //------------------------------Montgomery multiplication------------------------
3599 //
3600
3601 // Subtract 0:b from carry:a. Return carry.
3602 static unsigned long
3603 sub(unsigned long a[], unsigned long b[], unsigned long carry, long len) {
3604 long i = 0;
3605 unsigned long tmp, tmp2;
3606 __asm__ __volatile__ (
3607 "subfc %[tmp], %[tmp], %[tmp] \n" // pre-set CA
3608 "mtctr %[len] \n"
3609 "0: \n"
3610 "ldx %[tmp], %[i], %[a] \n"
3611 "ldx %[tmp2], %[i], %[b] \n"
3612 "subfe %[tmp], %[tmp2], %[tmp] \n" // subtract extended
3613 "stdx %[tmp], %[i], %[a] \n"
3614 "addi %[i], %[i], 8 \n"
3615 "bdnz 0b \n"
3616 "addme %[tmp], %[carry] \n" // carry + CA - 1
3617 : [i]"+b"(i), [tmp]"=&r"(tmp), [tmp2]"=&r"(tmp2)
3618 : [a]"r"(a), [b]"r"(b), [carry]"r"(carry), [len]"r"(len)
3619 : "ctr", "xer", "memory"
3620 );
3621 return tmp;
3622 }
3623
3624 // Multiply (unsigned) Long A by Long B, accumulating the double-
3625 // length result into the accumulator formed of T0, T1, and T2.
3626 inline void MACC(unsigned long A, unsigned long B, unsigned long &T0, unsigned long &T1, unsigned long &T2) {
3627 unsigned long hi, lo;
3628 __asm__ __volatile__ (
3629 "mulld %[lo], %[A], %[B] \n"
3630 "mulhdu %[hi], %[A], %[B] \n"
3631 "addc %[T0], %[T0], %[lo] \n"
3632 "adde %[T1], %[T1], %[hi] \n"
3633 "addze %[T2], %[T2] \n"
3634 : [hi]"=&r"(hi), [lo]"=&r"(lo), [T0]"+r"(T0), [T1]"+r"(T1), [T2]"+r"(T2)
3635 : [A]"r"(A), [B]"r"(B)
3636 : "xer"
3637 );
3638 }
3639
3640 // As above, but add twice the double-length result into the
3641 // accumulator.
3642 inline void MACC2(unsigned long A, unsigned long B, unsigned long &T0, unsigned long &T1, unsigned long &T2) {
3643 unsigned long hi, lo;
3644 __asm__ __volatile__ (
3645 "mulld %[lo], %[A], %[B] \n"
3646 "mulhdu %[hi], %[A], %[B] \n"
3647 "addc %[T0], %[T0], %[lo] \n"
3648 "adde %[T1], %[T1], %[hi] \n"
3649 "addze %[T2], %[T2] \n"
3650 "addc %[T0], %[T0], %[lo] \n"
3651 "adde %[T1], %[T1], %[hi] \n"
3652 "addze %[T2], %[T2] \n"
3653 : [hi]"=&r"(hi), [lo]"=&r"(lo), [T0]"+r"(T0), [T1]"+r"(T1), [T2]"+r"(T2)
3654 : [A]"r"(A), [B]"r"(B)
3655 : "xer"
3656 );
3657 }
3658
3659 // Fast Montgomery multiplication. The derivation of the algorithm is
3660 // in "A Cryptographic Library for the Motorola DSP56000,
3661 // Dusse and Kaliski, Proc. EUROCRYPT 90, pp. 230-237".
3662 static void
3663 montgomery_multiply(unsigned long a[], unsigned long b[], unsigned long n[],
3664 unsigned long m[], unsigned long inv, int len) {
3665 unsigned long t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator
3666 int i;
3667
3668 assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply");
3669
3670 for (i = 0; i < len; i++) {
3671 int j;
3672 for (j = 0; j < i; j++) {
3673 MACC(a[j], b[i-j], t0, t1, t2);
3674 MACC(m[j], n[i-j], t0, t1, t2);
3675 }
3676 MACC(a[i], b[0], t0, t1, t2);
3677 m[i] = t0 * inv;
3678 MACC(m[i], n[0], t0, t1, t2);
3679
3680 assert(t0 == 0, "broken Montgomery multiply");
3681
3682 t0 = t1; t1 = t2; t2 = 0;
3683 }
3684
3685 for (i = len; i < 2*len; i++) {
3686 int j;
3687 for (j = i-len+1; j < len; j++) {
3688 MACC(a[j], b[i-j], t0, t1, t2);
3689 MACC(m[j], n[i-j], t0, t1, t2);
3690 }
3691 m[i-len] = t0;
3692 t0 = t1; t1 = t2; t2 = 0;
3693 }
3694
3695 while (t0) {
3696 t0 = sub(m, n, t0, len);
3697 }
3698 }
3699
3700 // Fast Montgomery squaring. This uses asymptotically 25% fewer
3701 // multiplies so it should be up to 25% faster than Montgomery
3702 // multiplication. However, its loop control is more complex and it
3703 // may actually run slower on some machines.
3704 static void
3705 montgomery_square(unsigned long a[], unsigned long n[],
3706 unsigned long m[], unsigned long inv, int len) {
3707 unsigned long t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator
3708 int i;
3709
3710 assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply");
3711
3712 for (i = 0; i < len; i++) {
3713 int j;
3714 int end = (i+1)/2;
3715 for (j = 0; j < end; j++) {
3716 MACC2(a[j], a[i-j], t0, t1, t2);
3717 MACC(m[j], n[i-j], t0, t1, t2);
3718 }
3719 if ((i & 1) == 0) {
3720 MACC(a[j], a[j], t0, t1, t2);
3721 }
3722 for (; j < i; j++) {
3723 MACC(m[j], n[i-j], t0, t1, t2);
3724 }
3725 m[i] = t0 * inv;
3726 MACC(m[i], n[0], t0, t1, t2);
3727
3728 assert(t0 == 0, "broken Montgomery square");
3729
3730 t0 = t1; t1 = t2; t2 = 0;
3731 }
3732
3733 for (i = len; i < 2*len; i++) {
3734 int start = i-len+1;
3735 int end = start + (len - start)/2;
3736 int j;
3737 for (j = start; j < end; j++) {
3738 MACC2(a[j], a[i-j], t0, t1, t2);
3739 MACC(m[j], n[i-j], t0, t1, t2);
3740 }
3741 if ((i & 1) == 0) {
3742 MACC(a[j], a[j], t0, t1, t2);
3743 }
3744 for (; j < len; j++) {
3745 MACC(m[j], n[i-j], t0, t1, t2);
3746 }
3747 m[i-len] = t0;
3748 t0 = t1; t1 = t2; t2 = 0;
3749 }
3750
3751 while (t0) {
3752 t0 = sub(m, n, t0, len);
3753 }
3754 }
3755
3756 // The threshold at which squaring is advantageous was determined
3757 // experimentally on an i7-3930K (Ivy Bridge) CPU @ 3.5GHz.
3758 // Doesn't seem to be relevant for Power8 so we use the same value.
3759 #define MONTGOMERY_SQUARING_THRESHOLD 64
3760
3761 // Copy len longwords from s to d, word-swapping as we go. The
3762 // destination array is reversed.
3763 static void reverse_words(unsigned long *s, unsigned long *d, int len) {
3764 d += len;
3765 while(len-- > 0) {
3766 d--;
3767 unsigned long s_val = *s;
3768 // Swap words in a longword on little endian machines.
3769 #ifdef VM_LITTLE_ENDIAN
3770 s_val = (s_val << 32) | (s_val >> 32);
3771 #endif
3772 *d = s_val;
3773 s++;
3774 }
3775 }
3776
3777 void SharedRuntime::montgomery_multiply(jint *a_ints, jint *b_ints, jint *n_ints,
3778 jint len, jlong inv,
3779 jint *m_ints) {
3780 len = len & 0x7fffFFFF; // C2 does not respect int to long conversion for stub calls.
3781 assert(len % 2 == 0, "array length in montgomery_multiply must be even");
3782 int longwords = len/2;
3783
3784 // Make very sure we don't use so much space that the stack might
3785 // overflow. 512 jints corresponds to an 16384-bit integer and
3786 // will use here a total of 8k bytes of stack space.
3787 int divisor = sizeof(unsigned long) * 4;
3788 guarantee(longwords <= 8192 / divisor, "must be");
3789 int total_allocation = longwords * sizeof (unsigned long) * 4;
3790 unsigned long *scratch = (unsigned long *)alloca(total_allocation);
3791
3792 // Local scratch arrays
3793 unsigned long
3794 *a = scratch + 0 * longwords,
3795 *b = scratch + 1 * longwords,
3796 *n = scratch + 2 * longwords,
3797 *m = scratch + 3 * longwords;
3798
3799 reverse_words((unsigned long *)a_ints, a, longwords);
3800 reverse_words((unsigned long *)b_ints, b, longwords);
3801 reverse_words((unsigned long *)n_ints, n, longwords);
3802
3803 ::montgomery_multiply(a, b, n, m, (unsigned long)inv, longwords);
3804
3805 reverse_words(m, (unsigned long *)m_ints, longwords);
3806 }
3807
3808 void SharedRuntime::montgomery_square(jint *a_ints, jint *n_ints,
3809 jint len, jlong inv,
3810 jint *m_ints) {
3811 len = len & 0x7fffFFFF; // C2 does not respect int to long conversion for stub calls.
3812 assert(len % 2 == 0, "array length in montgomery_square must be even");
3813 int longwords = len/2;
3814
3815 // Make very sure we don't use so much space that the stack might
3816 // overflow. 512 jints corresponds to an 16384-bit integer and
3817 // will use here a total of 6k bytes of stack space.
3818 int divisor = sizeof(unsigned long) * 3;
3819 guarantee(longwords <= (8192 / divisor), "must be");
3820 int total_allocation = longwords * sizeof (unsigned long) * 3;
3821 unsigned long *scratch = (unsigned long *)alloca(total_allocation);
3822
3823 // Local scratch arrays
3824 unsigned long
3825 *a = scratch + 0 * longwords,
3826 *n = scratch + 1 * longwords,
3827 *m = scratch + 2 * longwords;
3828
3829 reverse_words((unsigned long *)a_ints, a, longwords);
3830 reverse_words((unsigned long *)n_ints, n, longwords);
3831
3832 if (len >= MONTGOMERY_SQUARING_THRESHOLD) {
3833 ::montgomery_square(a, n, m, (unsigned long)inv, longwords);
3834 } else {
3835 ::montgomery_multiply(a, a, n, m, (unsigned long)inv, longwords);
3836 }
3837
3838 reverse_words(m, (unsigned long *)m_ints, longwords);
3839 }
3840
3841 #if INCLUDE_JFR
3842
3843 // For c2: c_rarg0 is junk, call to runtime to write a checkpoint.
3844 // It returns a jobject handle to the event writer.
3845 // The handle is dereferenced and the return value is the event writer oop.
3846 RuntimeStub* SharedRuntime::generate_jfr_write_checkpoint() {
3847 const char* name = SharedRuntime::stub_name(SharedStubId::jfr_write_checkpoint_id);
3848 CodeBuffer code(name, 512, 64);
3849 MacroAssembler* masm = new MacroAssembler(&code);
3850
3851 Register tmp1 = R10_ARG8;
3852 Register tmp2 = R9_ARG7;
3853
3854 int framesize = frame::native_abi_reg_args_size / VMRegImpl::stack_slot_size;
3855 address start = __ pc();
3856 __ mflr(tmp1);
3857 __ std(tmp1, _abi0(lr), R1_SP); // save return pc
3858 __ push_frame_reg_args(0, tmp1);
3859 int frame_complete = __ pc() - start;
3860 __ set_last_Java_frame(R1_SP, noreg);
3861 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::write_checkpoint), R16_thread);
3862 address calls_return_pc = __ last_calls_return_pc();
3863 __ reset_last_Java_frame();
3864 // The handle is dereferenced through a load barrier.
3865 __ resolve_global_jobject(R3_RET, tmp1, tmp2, MacroAssembler::PRESERVATION_NONE);
3866 __ pop_frame();
3867 __ ld(tmp1, _abi0(lr), R1_SP);
3868 __ mtlr(tmp1);
3869 __ blr();
3870
3871 OopMapSet* oop_maps = new OopMapSet();
3872 OopMap* map = new OopMap(framesize, 0);
3873 oop_maps->add_gc_map(calls_return_pc - start, map);
3874
3875 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
3876 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
3877 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3878 oop_maps, false);
3879 return stub;
3880 }
3881
3882 // For c2: call to return a leased buffer.
3883 RuntimeStub* SharedRuntime::generate_jfr_return_lease() {
3884 const char* name = SharedRuntime::stub_name(SharedStubId::jfr_return_lease_id);
3885 CodeBuffer code(name, 512, 64);
3886 MacroAssembler* masm = new MacroAssembler(&code);
3887
3888 Register tmp1 = R10_ARG8;
3889 Register tmp2 = R9_ARG7;
3890
3891 int framesize = frame::native_abi_reg_args_size / VMRegImpl::stack_slot_size;
3892 address start = __ pc();
3893 __ mflr(tmp1);
3894 __ std(tmp1, _abi0(lr), R1_SP); // save return pc
3895 __ push_frame_reg_args(0, tmp1);
3896 int frame_complete = __ pc() - start;
3897 __ set_last_Java_frame(R1_SP, noreg);
3898 __ call_VM_leaf(CAST_FROM_FN_PTR(address, JfrIntrinsicSupport::return_lease), R16_thread);
3899 address calls_return_pc = __ last_calls_return_pc();
3900 __ reset_last_Java_frame();
3901 __ pop_frame();
3902 __ ld(tmp1, _abi0(lr), R1_SP);
3903 __ mtlr(tmp1);
3904 __ blr();
3905
3906 OopMapSet* oop_maps = new OopMapSet();
3907 OopMap* map = new OopMap(framesize, 0);
3908 oop_maps->add_gc_map(calls_return_pc - start, map);
3909
3910 RuntimeStub* stub = // codeBlob framesize is in words (not VMRegImpl::slot_size)
3911 RuntimeStub::new_runtime_stub(name, &code, frame_complete,
3912 (framesize >> (LogBytesPerWord - LogBytesPerInt)),
3913 oop_maps, false);
3914 return stub;
3915 }
3916
3917 #endif // INCLUDE_JFR