1 /*
2 * Copyright (c) 1997, 2024, Oracle and/or its affiliates. All rights reserved.
3 * Copyright (c) 2012, 2024 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 "precompiled.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "code/debugInfoRec.hpp"
29 #include "code/compiledIC.hpp"
30 #include "code/vtableStubs.hpp"
31 #include "frame_ppc.hpp"
32 #include "compiler/oopMap.hpp"
33 #include "gc/shared/gcLocker.hpp"
34 #include "interpreter/interpreter.hpp"
35 #include "interpreter/interp_masm.hpp"
36 #include "memory/resourceArea.hpp"
37 #include "oops/klass.inline.hpp"
38 #include "prims/methodHandles.hpp"
39 #include "runtime/continuation.hpp"
40 #include "runtime/continuationEntry.inline.hpp"
41 #include "runtime/jniHandles.hpp"
42 #include "runtime/os.inline.hpp"
43 #include "runtime/safepointMechanism.hpp"
44 #include "runtime/sharedRuntime.hpp"
45 #include "runtime/signature.hpp"
46 #include "runtime/stubRoutines.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 vs_reg
115 } RegisterType;
116
117 typedef enum {
118 reg_size = 8,
119 half_reg_size = reg_size / 2,
120 vs_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_LiveVSReg(regname) \
141 { RegisterSaver::vs_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_LiveVSRegs[] = {
224 //
225 // live vector scalar registers (optional, only these ones are used by C2):
226 //
227 RegisterSaver_LiveVSReg( VSR32 ),
228 RegisterSaver_LiveVSReg( VSR33 ),
229 RegisterSaver_LiveVSReg( VSR34 ),
230 RegisterSaver_LiveVSReg( VSR35 ),
231 RegisterSaver_LiveVSReg( VSR36 ),
232 RegisterSaver_LiveVSReg( VSR37 ),
233 RegisterSaver_LiveVSReg( VSR38 ),
234 RegisterSaver_LiveVSReg( VSR39 ),
235 RegisterSaver_LiveVSReg( VSR40 ),
236 RegisterSaver_LiveVSReg( VSR41 ),
237 RegisterSaver_LiveVSReg( VSR42 ),
238 RegisterSaver_LiveVSReg( VSR43 ),
239 RegisterSaver_LiveVSReg( VSR44 ),
240 RegisterSaver_LiveVSReg( VSR45 ),
241 RegisterSaver_LiveVSReg( VSR46 ),
242 RegisterSaver_LiveVSReg( VSR47 ),
243 RegisterSaver_LiveVSReg( VSR48 ),
244 RegisterSaver_LiveVSReg( VSR49 ),
245 RegisterSaver_LiveVSReg( VSR50 ),
246 RegisterSaver_LiveVSReg( VSR51 )
247 };
248
249
250 OopMap* RegisterSaver::push_frame_reg_args_and_save_live_registers(MacroAssembler* masm,
251 int* out_frame_size_in_bytes,
252 bool generate_oop_map,
253 int return_pc_adjustment,
254 ReturnPCLocation return_pc_location,
255 bool save_vectors) {
256 // Push an abi_reg_args-frame and store all registers which may be live.
257 // If requested, create an OopMap: Record volatile registers as
258 // callee-save values in an OopMap so their save locations will be
259 // propagated to the RegisterMap of the caller frame during
260 // StackFrameStream construction (needed for deoptimization; see
261 // compiledVFrame::create_stack_value).
262 // If return_pc_adjustment != 0 adjust the return pc by return_pc_adjustment.
263 // Updated return pc is returned in R31 (if not return_pc_is_pre_saved).
264
265 // calculate frame size
266 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
267 sizeof(RegisterSaver::LiveRegType);
268 const int vsregstosave_num = save_vectors ? (sizeof(RegisterSaver_LiveVSRegs) /
269 sizeof(RegisterSaver::LiveRegType))
270 : 0;
271 const int register_save_size = regstosave_num * reg_size + vsregstosave_num * vs_reg_size;
272 const int frame_size_in_bytes = align_up(register_save_size, frame::alignment_in_bytes)
273 + frame::native_abi_reg_args_size;
274
275 *out_frame_size_in_bytes = frame_size_in_bytes;
276 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint);
277 const int register_save_offset = frame_size_in_bytes - register_save_size;
278
279 // OopMap frame size is in c2 stack slots (sizeof(jint)) not bytes or words.
280 OopMap* map = generate_oop_map ? new OopMap(frame_size_in_slots, 0) : nullptr;
281
282 BLOCK_COMMENT("push_frame_reg_args_and_save_live_registers {");
283
284 // push a new frame
285 __ push_frame(frame_size_in_bytes, noreg);
286
287 // Save some registers in the last (non-vector) slots of the new frame so we
288 // can use them as scratch regs or to determine the return pc.
289 __ std(R31, frame_size_in_bytes - reg_size - vsregstosave_num * vs_reg_size, R1_SP);
290 __ std(R30, frame_size_in_bytes - 2*reg_size - vsregstosave_num * vs_reg_size, R1_SP);
291
292 // save the flags
293 // Do the save_LR by hand and adjust the return pc if requested.
294 switch (return_pc_location) {
295 case return_pc_is_lr: __ mflr(R31); break;
296 case return_pc_is_pre_saved: assert(return_pc_adjustment == 0, "unsupported"); break;
297 case return_pc_is_thread_saved_exception_pc: __ ld(R31, thread_(saved_exception_pc)); break;
298 default: ShouldNotReachHere();
299 }
300 if (return_pc_location != return_pc_is_pre_saved) {
301 if (return_pc_adjustment != 0) {
302 __ addi(R31, R31, return_pc_adjustment);
303 }
304 __ std(R31, frame_size_in_bytes + _abi0(lr), R1_SP);
305 }
306
307 // save all registers (ints and floats)
308 int offset = register_save_offset;
309
310 for (int i = 0; i < regstosave_num; i++) {
311 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
312 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
313
314 switch (reg_type) {
315 case RegisterSaver::int_reg: {
316 if (reg_num < 30) { // We spilled R30-31 right at the beginning.
317 __ std(as_Register(reg_num), offset, R1_SP);
318 }
319 break;
320 }
321 case RegisterSaver::float_reg: {
322 __ stfd(as_FloatRegister(reg_num), offset, R1_SP);
323 break;
324 }
325 case RegisterSaver::special_reg: {
326 if (reg_num == SR_CTR.encoding()) {
327 __ mfctr(R30);
328 __ std(R30, offset, R1_SP);
329 } else {
330 Unimplemented();
331 }
332 break;
333 }
334 default:
335 ShouldNotReachHere();
336 }
337
338 if (generate_oop_map) {
339 map->set_callee_saved(VMRegImpl::stack2reg(offset>>2),
340 RegisterSaver_LiveRegs[i].vmreg);
341 map->set_callee_saved(VMRegImpl::stack2reg((offset + half_reg_size)>>2),
342 RegisterSaver_LiveRegs[i].vmreg->next());
343 }
344 offset += reg_size;
345 }
346
347 for (int i = 0; i < vsregstosave_num; i++) {
348 int reg_num = RegisterSaver_LiveVSRegs[i].reg_num;
349 int reg_type = RegisterSaver_LiveVSRegs[i].reg_type;
350
351 __ li(R30, offset);
352 __ stxvd2x(as_VectorSRegister(reg_num), R30, R1_SP);
353
354 if (generate_oop_map) {
355 map->set_callee_saved(VMRegImpl::stack2reg(offset>>2),
356 RegisterSaver_LiveVSRegs[i].vmreg);
357 }
358 offset += vs_reg_size;
359 }
360
361 assert(offset == frame_size_in_bytes, "consistency check");
362
363 BLOCK_COMMENT("} push_frame_reg_args_and_save_live_registers");
364
365 // And we're done.
366 return map;
367 }
368
369
370 // Pop the current frame and restore all the registers that we
371 // saved.
372 void RegisterSaver::restore_live_registers_and_pop_frame(MacroAssembler* masm,
373 int frame_size_in_bytes,
374 bool restore_ctr,
375 bool save_vectors) {
376 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
377 sizeof(RegisterSaver::LiveRegType);
378 const int vsregstosave_num = save_vectors ? (sizeof(RegisterSaver_LiveVSRegs) /
379 sizeof(RegisterSaver::LiveRegType))
380 : 0;
381 const int register_save_size = regstosave_num * reg_size + vsregstosave_num * vs_reg_size;
382
383 const int register_save_offset = frame_size_in_bytes - register_save_size;
384
385 BLOCK_COMMENT("restore_live_registers_and_pop_frame {");
386
387 // restore all registers (ints and floats)
388 int offset = register_save_offset;
389
390 for (int i = 0; i < regstosave_num; i++) {
391 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
392 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
393
394 switch (reg_type) {
395 case RegisterSaver::int_reg: {
396 if (reg_num != 31) // R31 restored at the end, it's the tmp reg!
397 __ ld(as_Register(reg_num), offset, R1_SP);
398 break;
399 }
400 case RegisterSaver::float_reg: {
401 __ lfd(as_FloatRegister(reg_num), offset, R1_SP);
402 break;
403 }
404 case RegisterSaver::special_reg: {
405 if (reg_num == SR_CTR.encoding()) {
406 if (restore_ctr) { // Nothing to do here if ctr already contains the next address.
407 __ ld(R31, offset, R1_SP);
408 __ mtctr(R31);
409 }
410 } else {
411 Unimplemented();
412 }
413 break;
414 }
415 default:
416 ShouldNotReachHere();
417 }
418 offset += reg_size;
419 }
420
421 for (int i = 0; i < vsregstosave_num; i++) {
422 int reg_num = RegisterSaver_LiveVSRegs[i].reg_num;
423 int reg_type = RegisterSaver_LiveVSRegs[i].reg_type;
424
425 __ li(R31, offset);
426 __ lxvd2x(as_VectorSRegister(reg_num), R31, R1_SP);
427
428 offset += vs_reg_size;
429 }
430
431 assert(offset == frame_size_in_bytes, "consistency check");
432
433 // restore link and the flags
434 __ ld(R31, frame_size_in_bytes + _abi0(lr), R1_SP);
435 __ mtlr(R31);
436
437 // restore scratch register's value
438 __ ld(R31, frame_size_in_bytes - reg_size - vsregstosave_num * vs_reg_size, R1_SP);
439
440 // pop the frame
441 __ addi(R1_SP, R1_SP, frame_size_in_bytes);
442
443 BLOCK_COMMENT("} restore_live_registers_and_pop_frame");
444 }
445
446 void RegisterSaver::push_frame_and_save_argument_registers(MacroAssembler* masm, Register r_temp,
447 int frame_size,int total_args, const VMRegPair *regs,
448 const VMRegPair *regs2) {
449 __ push_frame(frame_size, r_temp);
450 int st_off = frame_size - wordSize;
451 for (int i = 0; i < total_args; i++) {
452 VMReg r_1 = regs[i].first();
453 VMReg r_2 = regs[i].second();
454 if (!r_1->is_valid()) {
455 assert(!r_2->is_valid(), "");
456 continue;
457 }
458 if (r_1->is_Register()) {
459 Register r = r_1->as_Register();
460 __ std(r, st_off, R1_SP);
461 st_off -= wordSize;
462 } else if (r_1->is_FloatRegister()) {
463 FloatRegister f = r_1->as_FloatRegister();
464 __ stfd(f, st_off, R1_SP);
465 st_off -= wordSize;
466 }
467 }
468 if (regs2 != nullptr) {
469 for (int i = 0; i < total_args; i++) {
470 VMReg r_1 = regs2[i].first();
471 VMReg r_2 = regs2[i].second();
472 if (!r_1->is_valid()) {
473 assert(!r_2->is_valid(), "");
474 continue;
475 }
476 if (r_1->is_Register()) {
477 Register r = r_1->as_Register();
478 __ std(r, st_off, R1_SP);
479 st_off -= wordSize;
480 } else if (r_1->is_FloatRegister()) {
481 FloatRegister f = r_1->as_FloatRegister();
482 __ stfd(f, st_off, R1_SP);
483 st_off -= wordSize;
484 }
485 }
486 }
487 }
488
489 void RegisterSaver::restore_argument_registers_and_pop_frame(MacroAssembler*masm, int frame_size,
490 int total_args, const VMRegPair *regs,
491 const VMRegPair *regs2) {
492 int st_off = frame_size - wordSize;
493 for (int i = 0; i < total_args; i++) {
494 VMReg r_1 = regs[i].first();
495 VMReg r_2 = regs[i].second();
496 if (r_1->is_Register()) {
497 Register r = r_1->as_Register();
498 __ ld(r, st_off, R1_SP);
499 st_off -= wordSize;
500 } else if (r_1->is_FloatRegister()) {
501 FloatRegister f = r_1->as_FloatRegister();
502 __ lfd(f, st_off, R1_SP);
503 st_off -= wordSize;
504 }
505 }
506 if (regs2 != nullptr)
507 for (int i = 0; i < total_args; i++) {
508 VMReg r_1 = regs2[i].first();
509 VMReg r_2 = regs2[i].second();
510 if (r_1->is_Register()) {
511 Register r = r_1->as_Register();
512 __ ld(r, st_off, R1_SP);
513 st_off -= wordSize;
514 } else if (r_1->is_FloatRegister()) {
515 FloatRegister f = r_1->as_FloatRegister();
516 __ lfd(f, st_off, R1_SP);
517 st_off -= wordSize;
518 }
519 }
520 __ pop_frame();
521 }
522
523 // Restore the registers that might be holding a result.
524 void RegisterSaver::restore_result_registers(MacroAssembler* masm, int frame_size_in_bytes) {
525 const int regstosave_num = sizeof(RegisterSaver_LiveRegs) /
526 sizeof(RegisterSaver::LiveRegType);
527 const int register_save_size = regstosave_num * reg_size; // VS registers not relevant here.
528 const int register_save_offset = frame_size_in_bytes - register_save_size;
529
530 // restore all result registers (ints and floats)
531 int offset = register_save_offset;
532 for (int i = 0; i < regstosave_num; i++) {
533 int reg_num = RegisterSaver_LiveRegs[i].reg_num;
534 int reg_type = RegisterSaver_LiveRegs[i].reg_type;
535 switch (reg_type) {
536 case RegisterSaver::int_reg: {
537 if (as_Register(reg_num)==R3_RET) // int result_reg
538 __ ld(as_Register(reg_num), offset, R1_SP);
539 break;
540 }
541 case RegisterSaver::float_reg: {
542 if (as_FloatRegister(reg_num)==F1_RET) // float result_reg
543 __ lfd(as_FloatRegister(reg_num), offset, R1_SP);
544 break;
545 }
546 case RegisterSaver::special_reg: {
547 // Special registers don't hold a result.
548 break;
549 }
550 default:
551 ShouldNotReachHere();
552 }
553 offset += reg_size;
554 }
555
556 assert(offset == frame_size_in_bytes, "consistency check");
557 }
558
559 // Is vector's size (in bytes) bigger than a size saved by default?
560 bool SharedRuntime::is_wide_vector(int size) {
561 // Note, MaxVectorSize == 8/16 on PPC64.
562 assert(size <= (SuperwordUseVSX ? 16 : 8), "%d bytes vectors are not supported", size);
563 return size > 8;
564 }
565
566 static int reg2slot(VMReg r) {
567 return r->reg2stack() + SharedRuntime::out_preserve_stack_slots();
568 }
569
570 static int reg2offset(VMReg r) {
571 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
572 }
573
574 // ---------------------------------------------------------------------------
575 // Read the array of BasicTypes from a signature, and compute where the
576 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
577 // quantities. Values less than VMRegImpl::stack0 are registers, those above
578 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
579 // as framesizes are fixed.
580 // VMRegImpl::stack0 refers to the first slot 0(sp).
581 // and VMRegImpl::stack0+1 refers to the memory word 4-bytes higher. Register
582 // up to Register::number_of_registers) are the 64-bit
583 // integer registers.
584
585 // Note: the INPUTS in sig_bt are in units of Java argument words, which are
586 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit
587 // units regardless of build. Of course for i486 there is no 64 bit build
588
589 // The Java calling convention is a "shifted" version of the C ABI.
590 // By skipping the first C ABI register we can call non-static jni methods
591 // with small numbers of arguments without having to shuffle the arguments
592 // at all. Since we control the java ABI we ought to at least get some
593 // advantage out of it.
594
595 const VMReg java_iarg_reg[8] = {
596 R3->as_VMReg(),
597 R4->as_VMReg(),
598 R5->as_VMReg(),
599 R6->as_VMReg(),
600 R7->as_VMReg(),
601 R8->as_VMReg(),
602 R9->as_VMReg(),
603 R10->as_VMReg()
604 };
605
606 const VMReg java_farg_reg[13] = {
607 F1->as_VMReg(),
608 F2->as_VMReg(),
609 F3->as_VMReg(),
610 F4->as_VMReg(),
611 F5->as_VMReg(),
612 F6->as_VMReg(),
613 F7->as_VMReg(),
614 F8->as_VMReg(),
615 F9->as_VMReg(),
616 F10->as_VMReg(),
617 F11->as_VMReg(),
618 F12->as_VMReg(),
619 F13->as_VMReg()
620 };
621
622 const int num_java_iarg_registers = sizeof(java_iarg_reg) / sizeof(java_iarg_reg[0]);
623 const int num_java_farg_registers = sizeof(java_farg_reg) / sizeof(java_farg_reg[0]);
624
625 STATIC_ASSERT(num_java_iarg_registers == Argument::n_int_register_parameters_j);
626 STATIC_ASSERT(num_java_farg_registers == Argument::n_float_register_parameters_j);
627
628 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
629 VMRegPair *regs,
630 int total_args_passed) {
631 // C2c calling conventions for compiled-compiled calls.
632 // Put 8 ints/longs into registers _AND_ 13 float/doubles into
633 // registers _AND_ put the rest on the stack.
634
635 const int inc_stk_for_intfloat = 1; // 1 slots for ints and floats
636 const int inc_stk_for_longdouble = 2; // 2 slots for longs and doubles
637
638 int i;
639 VMReg reg;
640 int stk = 0;
641 int ireg = 0;
642 int freg = 0;
643
644 // We put the first 8 arguments into registers and the rest on the
645 // stack, float arguments are already in their argument registers
646 // due to c2c calling conventions (see calling_convention).
647 for (int i = 0; i < total_args_passed; ++i) {
648 switch(sig_bt[i]) {
649 case T_BOOLEAN:
650 case T_CHAR:
651 case T_BYTE:
652 case T_SHORT:
653 case T_INT:
654 if (ireg < num_java_iarg_registers) {
655 // Put int/ptr in register
656 reg = java_iarg_reg[ireg];
657 ++ireg;
658 } else {
659 // Put int/ptr on stack.
660 reg = VMRegImpl::stack2reg(stk);
661 stk += inc_stk_for_intfloat;
662 }
663 regs[i].set1(reg);
664 break;
665 case T_LONG:
666 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
667 if (ireg < num_java_iarg_registers) {
668 // Put long in register.
669 reg = java_iarg_reg[ireg];
670 ++ireg;
671 } else {
672 // Put long on stack. They must be aligned to 2 slots.
673 if (stk & 0x1) ++stk;
674 reg = VMRegImpl::stack2reg(stk);
675 stk += inc_stk_for_longdouble;
676 }
677 regs[i].set2(reg);
678 break;
679 case T_OBJECT:
680 case T_ARRAY:
681 case T_ADDRESS:
682 if (ireg < num_java_iarg_registers) {
683 // Put ptr in register.
684 reg = java_iarg_reg[ireg];
685 ++ireg;
686 } else {
687 // Put ptr on stack. Objects must be aligned to 2 slots too,
688 // because "64-bit pointers record oop-ishness on 2 aligned
689 // adjacent registers." (see OopFlow::build_oop_map).
690 if (stk & 0x1) ++stk;
691 reg = VMRegImpl::stack2reg(stk);
692 stk += inc_stk_for_longdouble;
693 }
694 regs[i].set2(reg);
695 break;
696 case T_FLOAT:
697 if (freg < num_java_farg_registers) {
698 // Put float in register.
699 reg = java_farg_reg[freg];
700 ++freg;
701 } else {
702 // Put float on stack.
703 reg = VMRegImpl::stack2reg(stk);
704 stk += inc_stk_for_intfloat;
705 }
706 regs[i].set1(reg);
707 break;
708 case T_DOUBLE:
709 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
710 if (freg < num_java_farg_registers) {
711 // Put double in register.
712 reg = java_farg_reg[freg];
713 ++freg;
714 } else {
715 // Put double on stack. They must be aligned to 2 slots.
716 if (stk & 0x1) ++stk;
717 reg = VMRegImpl::stack2reg(stk);
718 stk += inc_stk_for_longdouble;
719 }
720 regs[i].set2(reg);
721 break;
722 case T_VOID:
723 // Do not count halves.
724 regs[i].set_bad();
725 break;
726 default:
727 ShouldNotReachHere();
728 }
729 }
730 return stk;
731 }
732
733 #if defined(COMPILER1) || defined(COMPILER2)
734 // Calling convention for calling C code.
735 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
736 VMRegPair *regs,
737 int total_args_passed) {
738 // Calling conventions for C runtime calls and calls to JNI native methods.
739 //
740 // PPC64 convention: Hoist the first 8 int/ptr/long's in the first 8
741 // int regs, leaving int regs undefined if the arg is flt/dbl. Hoist
742 // the first 13 flt/dbl's in the first 13 fp regs but additionally
743 // copy flt/dbl to the stack if they are beyond the 8th argument.
744
745 const VMReg iarg_reg[8] = {
746 R3->as_VMReg(),
747 R4->as_VMReg(),
748 R5->as_VMReg(),
749 R6->as_VMReg(),
750 R7->as_VMReg(),
751 R8->as_VMReg(),
752 R9->as_VMReg(),
753 R10->as_VMReg()
754 };
755
756 const VMReg farg_reg[13] = {
757 F1->as_VMReg(),
758 F2->as_VMReg(),
759 F3->as_VMReg(),
760 F4->as_VMReg(),
761 F5->as_VMReg(),
762 F6->as_VMReg(),
763 F7->as_VMReg(),
764 F8->as_VMReg(),
765 F9->as_VMReg(),
766 F10->as_VMReg(),
767 F11->as_VMReg(),
768 F12->as_VMReg(),
769 F13->as_VMReg()
770 };
771
772 // Check calling conventions consistency.
773 assert(sizeof(iarg_reg) / sizeof(iarg_reg[0]) == Argument::n_int_register_parameters_c &&
774 sizeof(farg_reg) / sizeof(farg_reg[0]) == Argument::n_float_register_parameters_c,
775 "consistency");
776
777 const int additional_frame_header_slots = ((frame::native_abi_minframe_size - frame::jit_out_preserve_size)
778 / VMRegImpl::stack_slot_size);
779 const int float_offset_in_slots = Argument::float_on_stack_offset_in_bytes_c / VMRegImpl::stack_slot_size;
780
781 VMReg reg;
782 int arg = 0;
783 int freg = 0;
784 bool stack_used = false;
785
786 for (int i = 0; i < total_args_passed; ++i, ++arg) {
787 // Each argument corresponds to a slot in the Parameter Save Area (if not omitted)
788 int stk = (arg * 2) + additional_frame_header_slots;
789
790 switch(sig_bt[i]) {
791 //
792 // If arguments 0-7 are integers, they are passed in integer registers.
793 // Argument i is placed in iarg_reg[i].
794 //
795 case T_BOOLEAN:
796 case T_CHAR:
797 case T_BYTE:
798 case T_SHORT:
799 case T_INT:
800 // We must cast ints to longs and use full 64 bit stack slots
801 // here. Thus fall through, handle as long.
802 case T_LONG:
803 case T_OBJECT:
804 case T_ARRAY:
805 case T_ADDRESS:
806 case T_METADATA:
807 // Oops are already boxed if required (JNI).
808 if (arg < Argument::n_int_register_parameters_c) {
809 reg = iarg_reg[arg];
810 } else {
811 reg = VMRegImpl::stack2reg(stk);
812 stack_used = true;
813 }
814 regs[i].set2(reg);
815 break;
816
817 //
818 // Floats are treated differently from int regs: The first 13 float arguments
819 // are passed in registers (not the float args among the first 13 args).
820 // Thus argument i is NOT passed in farg_reg[i] if it is float. It is passed
821 // in farg_reg[j] if argument i is the j-th float argument of this call.
822 //
823 case T_FLOAT:
824 if (freg < Argument::n_float_register_parameters_c) {
825 // Put float in register ...
826 reg = farg_reg[freg];
827 ++freg;
828 } else {
829 // Put float on stack.
830 reg = VMRegImpl::stack2reg(stk + float_offset_in_slots);
831 stack_used = true;
832 }
833 regs[i].set1(reg);
834 break;
835 case T_DOUBLE:
836 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "expecting half");
837 if (freg < Argument::n_float_register_parameters_c) {
838 // Put double in register ...
839 reg = farg_reg[freg];
840 ++freg;
841 } else {
842 // Put double on stack.
843 reg = VMRegImpl::stack2reg(stk);
844 stack_used = true;
845 }
846 regs[i].set2(reg);
847 break;
848
849 case T_VOID:
850 // Do not count halves.
851 regs[i].set_bad();
852 --arg;
853 break;
854 default:
855 ShouldNotReachHere();
856 }
857 }
858
859 // Return size of the stack frame excluding the jit_out_preserve part in single-word slots.
860 #if defined(ABI_ELFv2)
861 assert(additional_frame_header_slots == 0, "ABIv2 shouldn't use extra slots");
862 // ABIv2 allows omitting the Parameter Save Area if the callee's prototype
863 // indicates that all parameters can be passed in registers.
864 return stack_used ? (arg * 2) : 0;
865 #else
866 // The Parameter Save Area needs to be at least 8 double-word slots for ABIv1.
867 // We have to add extra slots because ABIv1 uses a larger header.
868 return MAX2(arg, 8) * 2 + additional_frame_header_slots;
869 #endif
870 }
871 #endif // COMPILER2
872
873 int SharedRuntime::vector_calling_convention(VMRegPair *regs,
874 uint num_bits,
875 uint total_args_passed) {
876 Unimplemented();
877 return 0;
878 }
879
880 static address gen_c2i_adapter(MacroAssembler *masm,
881 int total_args_passed,
882 int comp_args_on_stack,
883 const BasicType *sig_bt,
884 const VMRegPair *regs,
885 Label& call_interpreter,
886 const Register& ientry) {
887
888 address c2i_entrypoint;
889
890 const Register sender_SP = R21_sender_SP; // == R21_tmp1
891 const Register code = R22_tmp2;
892 //const Register ientry = R23_tmp3;
893 const Register value_regs[] = { R24_tmp4, R25_tmp5, R26_tmp6 };
894 const int num_value_regs = sizeof(value_regs) / sizeof(Register);
895 int value_regs_index = 0;
896
897 const Register return_pc = R27_tmp7;
898 const Register tmp = R28_tmp8;
899
900 assert_different_registers(sender_SP, code, ientry, return_pc, tmp);
901
902 // Adapter needs TOP_IJAVA_FRAME_ABI.
903 const int adapter_size = frame::top_ijava_frame_abi_size +
904 align_up(total_args_passed * wordSize, frame::alignment_in_bytes);
905
906 // regular (verified) c2i entry point
907 c2i_entrypoint = __ pc();
908
909 // Does compiled code exists? If yes, patch the caller's callsite.
910 __ ld(code, method_(code));
911 __ cmpdi(CCR0, code, 0);
912 __ ld(ientry, method_(interpreter_entry)); // preloaded
913 __ beq(CCR0, call_interpreter);
914
915
916 // Patch caller's callsite, method_(code) was not null which means that
917 // compiled code exists.
918 __ mflr(return_pc);
919 __ std(return_pc, _abi0(lr), R1_SP);
920 RegisterSaver::push_frame_and_save_argument_registers(masm, tmp, adapter_size, total_args_passed, regs);
921
922 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite), R19_method, return_pc);
923
924 RegisterSaver::restore_argument_registers_and_pop_frame(masm, adapter_size, total_args_passed, regs);
925 __ ld(return_pc, _abi0(lr), R1_SP);
926 __ ld(ientry, method_(interpreter_entry)); // preloaded
927 __ mtlr(return_pc);
928
929
930 // Call the interpreter.
931 __ BIND(call_interpreter);
932 __ mtctr(ientry);
933
934 // Get a copy of the current SP for loading caller's arguments.
935 __ mr(sender_SP, R1_SP);
936
937 // Add space for the adapter.
938 __ resize_frame(-adapter_size, R12_scratch2);
939
940 int st_off = adapter_size - wordSize;
941
942 // Write the args into the outgoing interpreter space.
943 for (int i = 0; i < total_args_passed; i++) {
944 VMReg r_1 = regs[i].first();
945 VMReg r_2 = regs[i].second();
946 if (!r_1->is_valid()) {
947 assert(!r_2->is_valid(), "");
948 continue;
949 }
950 if (r_1->is_stack()) {
951 Register tmp_reg = value_regs[value_regs_index];
952 value_regs_index = (value_regs_index + 1) % num_value_regs;
953 // The calling convention produces OptoRegs that ignore the out
954 // preserve area (JIT's ABI). We must account for it here.
955 int ld_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
956 if (!r_2->is_valid()) {
957 __ lwz(tmp_reg, ld_off, sender_SP);
958 } else {
959 __ ld(tmp_reg, ld_off, sender_SP);
960 }
961 // Pretend stack targets were loaded into tmp_reg.
962 r_1 = tmp_reg->as_VMReg();
963 }
964
965 if (r_1->is_Register()) {
966 Register r = r_1->as_Register();
967 if (!r_2->is_valid()) {
968 __ stw(r, st_off, R1_SP);
969 st_off-=wordSize;
970 } else {
971 // Longs are given 2 64-bit slots in the interpreter, but the
972 // data is passed in only 1 slot.
973 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
974 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); )
975 st_off-=wordSize;
976 }
977 __ std(r, st_off, R1_SP);
978 st_off-=wordSize;
979 }
980 } else {
981 assert(r_1->is_FloatRegister(), "");
982 FloatRegister f = r_1->as_FloatRegister();
983 if (!r_2->is_valid()) {
984 __ stfs(f, st_off, R1_SP);
985 st_off-=wordSize;
986 } else {
987 // In 64bit, doubles are given 2 64-bit slots in the interpreter, but the
988 // data is passed in only 1 slot.
989 // One of these should get known junk...
990 DEBUG_ONLY( __ li(tmp, 0); __ std(tmp, st_off, R1_SP); )
991 st_off-=wordSize;
992 __ stfd(f, st_off, R1_SP);
993 st_off-=wordSize;
994 }
995 }
996 }
997
998 // Jump to the interpreter just as if interpreter was doing it.
999
1000 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
1001
1002 // load TOS
1003 __ addi(R15_esp, R1_SP, st_off);
1004
1005 // Frame_manager expects initial_caller_sp (= SP without resize by c2i) in R21_tmp1.
1006 assert(sender_SP == R21_sender_SP, "passing initial caller's SP in wrong register");
1007 __ bctr();
1008
1009 return c2i_entrypoint;
1010 }
1011
1012 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
1013 int total_args_passed,
1014 int comp_args_on_stack,
1015 const BasicType *sig_bt,
1016 const VMRegPair *regs) {
1017
1018 // Load method's entry-point from method.
1019 __ ld(R12_scratch2, in_bytes(Method::from_compiled_offset()), R19_method);
1020 __ mtctr(R12_scratch2);
1021
1022 // We will only enter here from an interpreted frame and never from after
1023 // passing thru a c2i. Azul allowed this but we do not. If we lose the
1024 // race and use a c2i we will remain interpreted for the race loser(s).
1025 // This removes all sorts of headaches on the x86 side and also eliminates
1026 // the possibility of having c2i -> i2c -> c2i -> ... endless transitions.
1027
1028 // Note: r13 contains the senderSP on entry. We must preserve it since
1029 // we may do a i2c -> c2i transition if we lose a race where compiled
1030 // code goes non-entrant while we get args ready.
1031 // In addition we use r13 to locate all the interpreter args as
1032 // we must align the stack to 16 bytes on an i2c entry else we
1033 // lose alignment we expect in all compiled code and register
1034 // save code can segv when fxsave instructions find improperly
1035 // aligned stack pointer.
1036
1037 const Register ld_ptr = R15_esp;
1038 const Register value_regs[] = { R22_tmp2, R23_tmp3, R24_tmp4, R25_tmp5, R26_tmp6 };
1039 const int num_value_regs = sizeof(value_regs) / sizeof(Register);
1040 int value_regs_index = 0;
1041
1042 int ld_offset = total_args_passed*wordSize;
1043
1044 // Cut-out for having no stack args. Since up to 2 int/oop args are passed
1045 // in registers, we will occasionally have no stack args.
1046 int comp_words_on_stack = 0;
1047 if (comp_args_on_stack) {
1048 // Sig words on the stack are greater-than VMRegImpl::stack0. Those in
1049 // registers are below. By subtracting stack0, we either get a negative
1050 // number (all values in registers) or the maximum stack slot accessed.
1051
1052 // Convert 4-byte c2 stack slots to words.
1053 comp_words_on_stack = align_up(comp_args_on_stack*VMRegImpl::stack_slot_size, wordSize)>>LogBytesPerWord;
1054 // Round up to miminum stack alignment, in wordSize.
1055 comp_words_on_stack = align_up(comp_words_on_stack, 2);
1056 __ resize_frame(-comp_words_on_stack * wordSize, R11_scratch1);
1057 }
1058
1059 // Now generate the shuffle code. Pick up all register args and move the
1060 // rest through register value=Z_R12.
1061 BLOCK_COMMENT("Shuffle arguments");
1062 for (int i = 0; i < total_args_passed; i++) {
1063 if (sig_bt[i] == T_VOID) {
1064 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
1065 continue;
1066 }
1067
1068 // Pick up 0, 1 or 2 words from ld_ptr.
1069 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
1070 "scrambled load targets?");
1071 VMReg r_1 = regs[i].first();
1072 VMReg r_2 = regs[i].second();
1073 if (!r_1->is_valid()) {
1074 assert(!r_2->is_valid(), "");
1075 continue;
1076 }
1077 if (r_1->is_FloatRegister()) {
1078 if (!r_2->is_valid()) {
1079 __ lfs(r_1->as_FloatRegister(), ld_offset, ld_ptr);
1080 ld_offset-=wordSize;
1081 } else {
1082 // Skip the unused interpreter slot.
1083 __ lfd(r_1->as_FloatRegister(), ld_offset-wordSize, ld_ptr);
1084 ld_offset-=2*wordSize;
1085 }
1086 } else {
1087 Register r;
1088 if (r_1->is_stack()) {
1089 // Must do a memory to memory move thru "value".
1090 r = value_regs[value_regs_index];
1091 value_regs_index = (value_regs_index + 1) % num_value_regs;
1092 } else {
1093 r = r_1->as_Register();
1094 }
1095 if (!r_2->is_valid()) {
1096 // Not sure we need to do this but it shouldn't hurt.
1097 if (is_reference_type(sig_bt[i]) || sig_bt[i] == T_ADDRESS) {
1098 __ ld(r, ld_offset, ld_ptr);
1099 ld_offset-=wordSize;
1100 } else {
1101 __ lwz(r, ld_offset, ld_ptr);
1102 ld_offset-=wordSize;
1103 }
1104 } else {
1105 // In 64bit, longs are given 2 64-bit slots in the interpreter, but the
1106 // data is passed in only 1 slot.
1107 if (sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
1108 ld_offset-=wordSize;
1109 }
1110 __ ld(r, ld_offset, ld_ptr);
1111 ld_offset-=wordSize;
1112 }
1113
1114 if (r_1->is_stack()) {
1115 // Now store value where the compiler expects it
1116 int st_off = (r_1->reg2stack() + SharedRuntime::out_preserve_stack_slots())*VMRegImpl::stack_slot_size;
1117
1118 if (sig_bt[i] == T_INT || sig_bt[i] == T_FLOAT ||sig_bt[i] == T_BOOLEAN ||
1119 sig_bt[i] == T_SHORT || sig_bt[i] == T_CHAR || sig_bt[i] == T_BYTE) {
1120 __ stw(r, st_off, R1_SP);
1121 } else {
1122 __ std(r, st_off, R1_SP);
1123 }
1124 }
1125 }
1126 }
1127
1128 __ push_cont_fastpath(); // Set JavaThread::_cont_fastpath to the sp of the oldest interpreted frame we know about
1129
1130 BLOCK_COMMENT("Store method");
1131 // Store method into thread->callee_target.
1132 // We might end up in handle_wrong_method if the callee is
1133 // deoptimized as we race thru here. If that happens we don't want
1134 // to take a safepoint because the caller frame will look
1135 // interpreted and arguments are now "compiled" so it is much better
1136 // to make this transition invisible to the stack walking
1137 // code. Unfortunately if we try and find the callee by normal means
1138 // a safepoint is possible. So we stash the desired callee in the
1139 // thread and the vm will find there should this case occur.
1140 __ std(R19_method, thread_(callee_target));
1141
1142 // Jump to the compiled code just as if compiled code was doing it.
1143 __ bctr();
1144 }
1145
1146 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
1147 int total_args_passed,
1148 int comp_args_on_stack,
1149 const BasicType *sig_bt,
1150 const VMRegPair *regs,
1151 AdapterFingerPrint* fingerprint) {
1152 address i2c_entry;
1153 address c2i_unverified_entry;
1154 address c2i_entry;
1155
1156
1157 // entry: i2c
1158
1159 __ align(CodeEntryAlignment);
1160 i2c_entry = __ pc();
1161 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
1162
1163
1164 // entry: c2i unverified
1165
1166 __ align(CodeEntryAlignment);
1167 BLOCK_COMMENT("c2i unverified entry");
1168 c2i_unverified_entry = __ pc();
1169
1170 // inline_cache contains a CompiledICData
1171 const Register ic = R19_inline_cache_reg;
1172 const Register ic_klass = R11_scratch1;
1173 const Register receiver_klass = R12_scratch2;
1174 const Register code = R21_tmp1;
1175 const Register ientry = R23_tmp3;
1176
1177 assert_different_registers(ic, ic_klass, receiver_klass, R3_ARG1, code, ientry);
1178 assert(R11_scratch1 == R11, "need prologue scratch register");
1179
1180 Label call_interpreter;
1181
1182 __ ic_check(4 /* end_alignment */);
1183 __ ld(R19_method, CompiledICData::speculated_method_offset(), ic);
1184 // Argument is valid and klass is as expected, continue.
1185
1186 __ ld(code, method_(code));
1187 __ cmpdi(CCR0, code, 0);
1188 __ ld(ientry, method_(interpreter_entry)); // preloaded
1189 __ beq_predict_taken(CCR0, call_interpreter);
1190
1191 // Branch to ic_miss_stub.
1192 __ b64_patchable((address)SharedRuntime::get_ic_miss_stub(), relocInfo::runtime_call_type);
1193
1194 // entry: c2i
1195
1196 c2i_entry = __ pc();
1197
1198 // Class initialization barrier for static methods
1199 address c2i_no_clinit_check_entry = nullptr;
1200 if (VM_Version::supports_fast_class_init_checks()) {
1201 Label L_skip_barrier;
1202
1203 { // Bypass the barrier for non-static methods
1204 __ lwz(R0, in_bytes(Method::access_flags_offset()), R19_method);
1205 __ andi_(R0, R0, JVM_ACC_STATIC);
1206 __ beq(CCR0, L_skip_barrier); // non-static
1207 }
1208
1209 Register klass = R11_scratch1;
1210 __ load_method_holder(klass, R19_method);
1211 __ clinit_barrier(klass, R16_thread, &L_skip_barrier /*L_fast_path*/);
1212
1213 __ load_const_optimized(klass, SharedRuntime::get_handle_wrong_method_stub(), R0);
1214 __ mtctr(klass);
1215 __ bctr();
1216
1217 __ bind(L_skip_barrier);
1218 c2i_no_clinit_check_entry = __ pc();
1219 }
1220
1221 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
1222 bs->c2i_entry_barrier(masm, /* tmp register*/ ic_klass, /* tmp register*/ receiver_klass, /* tmp register*/ code);
1223
1224 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, call_interpreter, ientry);
1225
1226 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry,
1227 c2i_no_clinit_check_entry);
1228 }
1229
1230 // An oop arg. Must pass a handle not the oop itself.
1231 static void object_move(MacroAssembler* masm,
1232 int frame_size_in_slots,
1233 OopMap* oop_map, int oop_handle_offset,
1234 bool is_receiver, int* receiver_offset,
1235 VMRegPair src, VMRegPair dst,
1236 Register r_caller_sp, Register r_temp_1, Register r_temp_2) {
1237 assert(!is_receiver || (is_receiver && (*receiver_offset == -1)),
1238 "receiver has already been moved");
1239
1240 // We must pass a handle. First figure out the location we use as a handle.
1241
1242 if (src.first()->is_stack()) {
1243 // stack to stack or reg
1244
1245 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register();
1246 Label skip;
1247 const int oop_slot_in_callers_frame = reg2slot(src.first());
1248
1249 guarantee(!is_receiver, "expecting receiver in register");
1250 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot_in_callers_frame + frame_size_in_slots));
1251
1252 __ addi(r_handle, r_caller_sp, reg2offset(src.first()));
1253 __ ld( r_temp_2, reg2offset(src.first()), r_caller_sp);
1254 __ cmpdi(CCR0, r_temp_2, 0);
1255 __ bne(CCR0, skip);
1256 // Use a null handle if oop is null.
1257 __ li(r_handle, 0);
1258 __ bind(skip);
1259
1260 if (dst.first()->is_stack()) {
1261 // stack to stack
1262 __ std(r_handle, reg2offset(dst.first()), R1_SP);
1263 } else {
1264 // stack to reg
1265 // Nothing to do, r_handle is already the dst register.
1266 }
1267 } else {
1268 // reg to stack or reg
1269 const Register r_oop = src.first()->as_Register();
1270 const Register r_handle = dst.first()->is_stack() ? r_temp_1 : dst.first()->as_Register();
1271 const int oop_slot = (r_oop->encoding()-R3_ARG1->encoding()) * VMRegImpl::slots_per_word
1272 + oop_handle_offset; // in slots
1273 const int oop_offset = oop_slot * VMRegImpl::stack_slot_size;
1274 Label skip;
1275
1276 if (is_receiver) {
1277 *receiver_offset = oop_offset;
1278 }
1279 oop_map->set_oop(VMRegImpl::stack2reg(oop_slot));
1280
1281 __ std( r_oop, oop_offset, R1_SP);
1282 __ addi(r_handle, R1_SP, oop_offset);
1283
1284 __ cmpdi(CCR0, r_oop, 0);
1285 __ bne(CCR0, skip);
1286 // Use a null handle if oop is null.
1287 __ li(r_handle, 0);
1288 __ bind(skip);
1289
1290 if (dst.first()->is_stack()) {
1291 // reg to stack
1292 __ std(r_handle, reg2offset(dst.first()), R1_SP);
1293 } else {
1294 // reg to reg
1295 // Nothing to do, r_handle is already the dst register.
1296 }
1297 }
1298 }
1299
1300 static void int_move(MacroAssembler*masm,
1301 VMRegPair src, VMRegPair dst,
1302 Register r_caller_sp, Register r_temp) {
1303 assert(src.first()->is_valid(), "incoming must be int");
1304 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long");
1305
1306 if (src.first()->is_stack()) {
1307 if (dst.first()->is_stack()) {
1308 // stack to stack
1309 __ lwa(r_temp, reg2offset(src.first()), r_caller_sp);
1310 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1311 } else {
1312 // stack to reg
1313 __ lwa(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp);
1314 }
1315 } else if (dst.first()->is_stack()) {
1316 // reg to stack
1317 __ extsw(r_temp, src.first()->as_Register());
1318 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1319 } else {
1320 // reg to reg
1321 __ extsw(dst.first()->as_Register(), src.first()->as_Register());
1322 }
1323 }
1324
1325 static void long_move(MacroAssembler*masm,
1326 VMRegPair src, VMRegPair dst,
1327 Register r_caller_sp, Register r_temp) {
1328 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be long");
1329 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be long");
1330
1331 if (src.first()->is_stack()) {
1332 if (dst.first()->is_stack()) {
1333 // stack to stack
1334 __ ld( r_temp, reg2offset(src.first()), r_caller_sp);
1335 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1336 } else {
1337 // stack to reg
1338 __ ld(dst.first()->as_Register(), reg2offset(src.first()), r_caller_sp);
1339 }
1340 } else if (dst.first()->is_stack()) {
1341 // reg to stack
1342 __ std(src.first()->as_Register(), reg2offset(dst.first()), R1_SP);
1343 } else {
1344 // reg to reg
1345 if (dst.first()->as_Register() != src.first()->as_Register())
1346 __ mr(dst.first()->as_Register(), src.first()->as_Register());
1347 }
1348 }
1349
1350 static void float_move(MacroAssembler*masm,
1351 VMRegPair src, VMRegPair dst,
1352 Register r_caller_sp, Register r_temp) {
1353 assert(src.first()->is_valid() && !src.second()->is_valid(), "incoming must be float");
1354 assert(dst.first()->is_valid() && !dst.second()->is_valid(), "outgoing must be float");
1355
1356 if (src.first()->is_stack()) {
1357 if (dst.first()->is_stack()) {
1358 // stack to stack
1359 __ lwz(r_temp, reg2offset(src.first()), r_caller_sp);
1360 __ stw(r_temp, reg2offset(dst.first()), R1_SP);
1361 } else {
1362 // stack to reg
1363 __ lfs(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp);
1364 }
1365 } else if (dst.first()->is_stack()) {
1366 // reg to stack
1367 __ stfs(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP);
1368 } else {
1369 // reg to reg
1370 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister())
1371 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister());
1372 }
1373 }
1374
1375 static void double_move(MacroAssembler*masm,
1376 VMRegPair src, VMRegPair dst,
1377 Register r_caller_sp, Register r_temp) {
1378 assert(src.first()->is_valid() && src.second() == src.first()->next(), "incoming must be double");
1379 assert(dst.first()->is_valid() && dst.second() == dst.first()->next(), "outgoing must be double");
1380
1381 if (src.first()->is_stack()) {
1382 if (dst.first()->is_stack()) {
1383 // stack to stack
1384 __ ld( r_temp, reg2offset(src.first()), r_caller_sp);
1385 __ std(r_temp, reg2offset(dst.first()), R1_SP);
1386 } else {
1387 // stack to reg
1388 __ lfd(dst.first()->as_FloatRegister(), reg2offset(src.first()), r_caller_sp);
1389 }
1390 } else if (dst.first()->is_stack()) {
1391 // reg to stack
1392 __ stfd(src.first()->as_FloatRegister(), reg2offset(dst.first()), R1_SP);
1393 } else {
1394 // reg to reg
1395 if (dst.first()->as_FloatRegister() != src.first()->as_FloatRegister())
1396 __ fmr(dst.first()->as_FloatRegister(), src.first()->as_FloatRegister());
1397 }
1398 }
1399
1400 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1401 switch (ret_type) {
1402 case T_BOOLEAN:
1403 case T_CHAR:
1404 case T_BYTE:
1405 case T_SHORT:
1406 case T_INT:
1407 __ stw (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1408 break;
1409 case T_ARRAY:
1410 case T_OBJECT:
1411 case T_LONG:
1412 __ std (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1413 break;
1414 case T_FLOAT:
1415 __ stfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1416 break;
1417 case T_DOUBLE:
1418 __ stfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1419 break;
1420 case T_VOID:
1421 break;
1422 default:
1423 ShouldNotReachHere();
1424 break;
1425 }
1426 }
1427
1428 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1429 switch (ret_type) {
1430 case T_BOOLEAN:
1431 case T_CHAR:
1432 case T_BYTE:
1433 case T_SHORT:
1434 case T_INT:
1435 __ lwz(R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1436 break;
1437 case T_ARRAY:
1438 case T_OBJECT:
1439 case T_LONG:
1440 __ ld (R3_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1441 break;
1442 case T_FLOAT:
1443 __ lfs(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1444 break;
1445 case T_DOUBLE:
1446 __ lfd(F1_RET, frame_slots*VMRegImpl::stack_slot_size, R1_SP);
1447 break;
1448 case T_VOID:
1449 break;
1450 default:
1451 ShouldNotReachHere();
1452 break;
1453 }
1454 }
1455
1456 static void verify_oop_args(MacroAssembler* masm,
1457 const methodHandle& method,
1458 const BasicType* sig_bt,
1459 const VMRegPair* regs) {
1460 Register temp_reg = R19_method; // not part of any compiled calling seq
1461 if (VerifyOops) {
1462 for (int i = 0; i < method->size_of_parameters(); i++) {
1463 if (is_reference_type(sig_bt[i])) {
1464 VMReg r = regs[i].first();
1465 assert(r->is_valid(), "bad oop arg");
1466 if (r->is_stack()) {
1467 __ ld(temp_reg, reg2offset(r), R1_SP);
1468 __ verify_oop(temp_reg, FILE_AND_LINE);
1469 } else {
1470 __ verify_oop(r->as_Register(), FILE_AND_LINE);
1471 }
1472 }
1473 }
1474 }
1475 }
1476
1477 static void gen_special_dispatch(MacroAssembler* masm,
1478 const methodHandle& method,
1479 const BasicType* sig_bt,
1480 const VMRegPair* regs) {
1481 verify_oop_args(masm, method, sig_bt, regs);
1482 vmIntrinsics::ID iid = method->intrinsic_id();
1483
1484 // Now write the args into the outgoing interpreter space
1485 bool has_receiver = false;
1486 Register receiver_reg = noreg;
1487 int member_arg_pos = -1;
1488 Register member_reg = noreg;
1489 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1490 if (ref_kind != 0) {
1491 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1492 member_reg = R19_method; // known to be free at this point
1493 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1494 } else if (iid == vmIntrinsics::_invokeBasic) {
1495 has_receiver = true;
1496 } else if (iid == vmIntrinsics::_linkToNative) {
1497 member_arg_pos = method->size_of_parameters() - 1; // trailing NativeEntryPoint argument
1498 member_reg = R19_method; // known to be free at this point
1499 } else {
1500 fatal("unexpected intrinsic id %d", vmIntrinsics::as_int(iid));
1501 }
1502
1503 if (member_reg != noreg) {
1504 // Load the member_arg into register, if necessary.
1505 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1506 VMReg r = regs[member_arg_pos].first();
1507 if (r->is_stack()) {
1508 __ ld(member_reg, reg2offset(r), R1_SP);
1509 } else {
1510 // no data motion is needed
1511 member_reg = r->as_Register();
1512 }
1513 }
1514
1515 if (has_receiver) {
1516 // Make sure the receiver is loaded into a register.
1517 assert(method->size_of_parameters() > 0, "oob");
1518 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1519 VMReg r = regs[0].first();
1520 assert(r->is_valid(), "bad receiver arg");
1521 if (r->is_stack()) {
1522 // Porting note: This assumes that compiled calling conventions always
1523 // pass the receiver oop in a register. If this is not true on some
1524 // platform, pick a temp and load the receiver from stack.
1525 fatal("receiver always in a register");
1526 receiver_reg = R11_scratch1; // TODO (hs24): is R11_scratch1 really free at this point?
1527 __ ld(receiver_reg, reg2offset(r), R1_SP);
1528 } else {
1529 // no data motion is needed
1530 receiver_reg = r->as_Register();
1531 }
1532 }
1533
1534 // Figure out which address we are really jumping to:
1535 MethodHandles::generate_method_handle_dispatch(masm, iid,
1536 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1537 }
1538
1539 //---------------------------- continuation_enter_setup ---------------------------
1540 //
1541 // Frame setup.
1542 //
1543 // Arguments:
1544 // None.
1545 //
1546 // Results:
1547 // R1_SP: pointer to blank ContinuationEntry in the pushed frame.
1548 //
1549 // Kills:
1550 // R0, R20
1551 //
1552 static OopMap* continuation_enter_setup(MacroAssembler* masm, int& framesize_words) {
1553 assert(ContinuationEntry::size() % VMRegImpl::stack_slot_size == 0, "");
1554 assert(in_bytes(ContinuationEntry::cont_offset()) % VMRegImpl::stack_slot_size == 0, "");
1555 assert(in_bytes(ContinuationEntry::chunk_offset()) % VMRegImpl::stack_slot_size == 0, "");
1556
1557 const int frame_size_in_bytes = (int)ContinuationEntry::size();
1558 assert(is_aligned(frame_size_in_bytes, frame::alignment_in_bytes), "alignment error");
1559
1560 framesize_words = frame_size_in_bytes / wordSize;
1561
1562 DEBUG_ONLY(__ block_comment("setup {"));
1563 // Save return pc and push entry frame
1564 const Register return_pc = R20;
1565 __ mflr(return_pc);
1566 __ std(return_pc, _abi0(lr), R1_SP); // SP->lr = return_pc
1567 __ push_frame(frame_size_in_bytes , R0); // SP -= frame_size_in_bytes
1568
1569 OopMap* map = new OopMap((int)frame_size_in_bytes / VMRegImpl::stack_slot_size, 0 /* arg_slots*/);
1570
1571 __ ld_ptr(R0, JavaThread::cont_entry_offset(), R16_thread);
1572 __ st_ptr(R1_SP, JavaThread::cont_entry_offset(), R16_thread);
1573 __ st_ptr(R0, ContinuationEntry::parent_offset(), R1_SP);
1574 DEBUG_ONLY(__ block_comment("} setup"));
1575
1576 return map;
1577 }
1578
1579 //---------------------------- fill_continuation_entry ---------------------------
1580 //
1581 // Initialize the new ContinuationEntry.
1582 //
1583 // Arguments:
1584 // R1_SP: pointer to blank Continuation entry
1585 // reg_cont_obj: pointer to the continuation
1586 // reg_flags: flags
1587 //
1588 // Results:
1589 // R1_SP: pointer to filled out ContinuationEntry
1590 //
1591 // Kills:
1592 // R8_ARG6, R9_ARG7, R10_ARG8
1593 //
1594 static void fill_continuation_entry(MacroAssembler* masm, Register reg_cont_obj, Register reg_flags) {
1595 assert_different_registers(reg_cont_obj, reg_flags);
1596 Register zero = R8_ARG6;
1597 Register tmp2 = R9_ARG7;
1598 Register tmp3 = R10_ARG8;
1599
1600 DEBUG_ONLY(__ block_comment("fill {"));
1601 #ifdef ASSERT
1602 __ load_const_optimized(tmp2, ContinuationEntry::cookie_value());
1603 __ stw(tmp2, in_bytes(ContinuationEntry::cookie_offset()), R1_SP);
1604 #endif //ASSERT
1605
1606 __ li(zero, 0);
1607 __ st_ptr(reg_cont_obj, ContinuationEntry::cont_offset(), R1_SP);
1608 __ stw(reg_flags, in_bytes(ContinuationEntry::flags_offset()), R1_SP);
1609 __ st_ptr(zero, ContinuationEntry::chunk_offset(), R1_SP);
1610 __ stw(zero, in_bytes(ContinuationEntry::argsize_offset()), R1_SP);
1611 __ stw(zero, in_bytes(ContinuationEntry::pin_count_offset()), R1_SP);
1612
1613 __ ld_ptr(tmp2, JavaThread::cont_fastpath_offset(), R16_thread);
1614 __ ld(tmp3, in_bytes(JavaThread::held_monitor_count_offset()), R16_thread);
1615 __ st_ptr(tmp2, ContinuationEntry::parent_cont_fastpath_offset(), R1_SP);
1616 __ std(tmp3, in_bytes(ContinuationEntry::parent_held_monitor_count_offset()), R1_SP);
1617
1618 __ st_ptr(zero, JavaThread::cont_fastpath_offset(), R16_thread);
1619 __ std(zero, in_bytes(JavaThread::held_monitor_count_offset()), R16_thread);
1620 DEBUG_ONLY(__ block_comment("} fill"));
1621 }
1622
1623 //---------------------------- continuation_enter_cleanup ---------------------------
1624 //
1625 // Copy corresponding attributes from the top ContinuationEntry to the JavaThread
1626 // before deleting it.
1627 //
1628 // Arguments:
1629 // R1_SP: pointer to the ContinuationEntry
1630 //
1631 // Results:
1632 // None.
1633 //
1634 // Kills:
1635 // R8_ARG6, R9_ARG7, R10_ARG8, R15_esp
1636 //
1637 static void continuation_enter_cleanup(MacroAssembler* masm) {
1638 Register tmp1 = R8_ARG6;
1639 Register tmp2 = R9_ARG7;
1640 Register tmp3 = R10_ARG8;
1641
1642 #ifdef ASSERT
1643 __ block_comment("clean {");
1644 __ ld_ptr(tmp1, JavaThread::cont_entry_offset(), R16_thread);
1645 __ cmpd(CCR0, R1_SP, tmp1);
1646 __ asm_assert_eq(FILE_AND_LINE ": incorrect R1_SP");
1647 #endif
1648
1649 __ ld_ptr(tmp1, ContinuationEntry::parent_cont_fastpath_offset(), R1_SP);
1650 __ st_ptr(tmp1, JavaThread::cont_fastpath_offset(), R16_thread);
1651
1652 if (CheckJNICalls) {
1653 // Check if this is a virtual thread continuation
1654 Label L_skip_vthread_code;
1655 __ lwz(R0, in_bytes(ContinuationEntry::flags_offset()), R1_SP);
1656 __ cmpwi(CCR0, R0, 0);
1657 __ beq(CCR0, L_skip_vthread_code);
1658
1659 // If the held monitor count is > 0 and this vthread is terminating then
1660 // it failed to release a JNI monitor. So we issue the same log message
1661 // that JavaThread::exit does.
1662 __ ld(R0, in_bytes(JavaThread::jni_monitor_count_offset()), R16_thread);
1663 __ cmpdi(CCR0, R0, 0);
1664 __ beq(CCR0, L_skip_vthread_code);
1665
1666 // Save return value potentially containing the exception oop
1667 Register ex_oop = R15_esp; // nonvolatile register
1668 __ mr(ex_oop, R3_RET);
1669 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::log_jni_monitor_still_held));
1670 // Restore potental return value
1671 __ mr(R3_RET, ex_oop);
1672
1673 // For vthreads we have to explicitly zero the JNI monitor count of the carrier
1674 // on termination. The held count is implicitly zeroed below when we restore from
1675 // the parent held count (which has to be zero).
1676 __ li(tmp1, 0);
1677 __ std(tmp1, in_bytes(JavaThread::jni_monitor_count_offset()), R16_thread);
1678
1679 __ bind(L_skip_vthread_code);
1680 }
1681 #ifdef ASSERT
1682 else {
1683 // Check if this is a virtual thread continuation
1684 Label L_skip_vthread_code;
1685 __ lwz(R0, in_bytes(ContinuationEntry::flags_offset()), R1_SP);
1686 __ cmpwi(CCR0, R0, 0);
1687 __ beq(CCR0, L_skip_vthread_code);
1688
1689 // See comment just above. If not checking JNI calls the JNI count is only
1690 // needed for assertion checking.
1691 __ li(tmp1, 0);
1692 __ std(tmp1, in_bytes(JavaThread::jni_monitor_count_offset()), R16_thread);
1693
1694 __ bind(L_skip_vthread_code);
1695 }
1696 #endif
1697
1698 __ ld(tmp2, in_bytes(ContinuationEntry::parent_held_monitor_count_offset()), R1_SP);
1699 __ ld_ptr(tmp3, ContinuationEntry::parent_offset(), R1_SP);
1700 __ std(tmp2, in_bytes(JavaThread::held_monitor_count_offset()), R16_thread);
1701 __ st_ptr(tmp3, JavaThread::cont_entry_offset(), R16_thread);
1702 DEBUG_ONLY(__ block_comment("} clean"));
1703 }
1704
1705 static void check_continuation_enter_argument(VMReg actual_vmreg,
1706 Register expected_reg,
1707 const char* name) {
1708 assert(!actual_vmreg->is_stack(), "%s cannot be on stack", name);
1709 assert(actual_vmreg->as_Register() == expected_reg,
1710 "%s is in unexpected register: %s instead of %s",
1711 name, actual_vmreg->as_Register()->name(), expected_reg->name());
1712 }
1713
1714 static void gen_continuation_enter(MacroAssembler* masm,
1715 const VMRegPair* regs,
1716 int& exception_offset,
1717 OopMapSet* oop_maps,
1718 int& frame_complete,
1719 int& framesize_words,
1720 int& interpreted_entry_offset,
1721 int& compiled_entry_offset) {
1722
1723 // enterSpecial(Continuation c, boolean isContinue, boolean isVirtualThread)
1724 int pos_cont_obj = 0;
1725 int pos_is_cont = 1;
1726 int pos_is_virtual = 2;
1727
1728 // The platform-specific calling convention may present the arguments in various registers.
1729 // To simplify the rest of the code, we expect the arguments to reside at these known
1730 // registers, and we additionally check the placement here in case calling convention ever
1731 // changes.
1732 Register reg_cont_obj = R3_ARG1;
1733 Register reg_is_cont = R4_ARG2;
1734 Register reg_is_virtual = R5_ARG3;
1735
1736 check_continuation_enter_argument(regs[pos_cont_obj].first(), reg_cont_obj, "Continuation object");
1737 check_continuation_enter_argument(regs[pos_is_cont].first(), reg_is_cont, "isContinue");
1738 check_continuation_enter_argument(regs[pos_is_virtual].first(), reg_is_virtual, "isVirtualThread");
1739
1740 address resolve_static_call = SharedRuntime::get_resolve_static_call_stub();
1741
1742 address start = __ pc();
1743
1744 Label L_thaw, L_exit;
1745
1746 // i2i entry used at interp_only_mode only
1747 interpreted_entry_offset = __ pc() - start;
1748 {
1749 #ifdef ASSERT
1750 Label is_interp_only;
1751 __ lwz(R0, in_bytes(JavaThread::interp_only_mode_offset()), R16_thread);
1752 __ cmpwi(CCR0, R0, 0);
1753 __ bne(CCR0, is_interp_only);
1754 __ stop("enterSpecial interpreter entry called when not in interp_only_mode");
1755 __ bind(is_interp_only);
1756 #endif
1757
1758 // Read interpreter arguments into registers (this is an ad-hoc i2c adapter)
1759 __ ld(reg_cont_obj, Interpreter::stackElementSize*3, R15_esp);
1760 __ lwz(reg_is_cont, Interpreter::stackElementSize*2, R15_esp);
1761 __ lwz(reg_is_virtual, Interpreter::stackElementSize*1, R15_esp);
1762
1763 __ push_cont_fastpath();
1764
1765 OopMap* map = continuation_enter_setup(masm, framesize_words);
1766
1767 // The frame is complete here, but we only record it for the compiled entry, so the frame would appear unsafe,
1768 // but that's okay because at the very worst we'll miss an async sample, but we're in interp_only_mode anyway.
1769
1770 fill_continuation_entry(masm, reg_cont_obj, reg_is_virtual);
1771
1772 // If isContinue, call to thaw. Otherwise, call Continuation.enter(Continuation c, boolean isContinue)
1773 __ cmpwi(CCR0, reg_is_cont, 0);
1774 __ bne(CCR0, L_thaw);
1775
1776 // --- call Continuation.enter(Continuation c, boolean isContinue)
1777
1778 // Emit compiled static call. The call will be always resolved to the c2i
1779 // entry of Continuation.enter(Continuation c, boolean isContinue).
1780 // There are special cases in SharedRuntime::resolve_static_call_C() and
1781 // SharedRuntime::resolve_sub_helper_internal() to achieve this
1782 // See also corresponding call below.
1783 address c2i_call_pc = __ pc();
1784 int start_offset = __ offset();
1785 // Put the entry point as a constant into the constant pool.
1786 const address entry_point_toc_addr = __ address_constant(resolve_static_call, RelocationHolder::none);
1787 const int entry_point_toc_offset = __ offset_to_method_toc(entry_point_toc_addr);
1788 guarantee(entry_point_toc_addr != nullptr, "const section overflow");
1789
1790 // Emit the trampoline stub which will be related to the branch-and-link below.
1791 address stub = __ emit_trampoline_stub(entry_point_toc_offset, start_offset);
1792 guarantee(stub != nullptr, "no space for trampoline stub");
1793
1794 __ relocate(relocInfo::static_call_type);
1795 // Note: At this point we do not have the address of the trampoline
1796 // stub, and the entry point might be too far away for bl, so __ pc()
1797 // serves as dummy and the bl will be patched later.
1798 __ bl(__ pc());
1799 oop_maps->add_gc_map(__ pc() - start, map);
1800 __ post_call_nop();
1801
1802 __ b(L_exit);
1803
1804 // static stub for the call above
1805 stub = CompiledDirectCall::emit_to_interp_stub(masm, c2i_call_pc);
1806 guarantee(stub != nullptr, "no space for static stub");
1807 }
1808
1809 // compiled entry
1810 __ align(CodeEntryAlignment);
1811 compiled_entry_offset = __ pc() - start;
1812
1813 OopMap* map = continuation_enter_setup(masm, framesize_words);
1814
1815 // Frame is now completed as far as size and linkage.
1816 frame_complete =__ pc() - start;
1817
1818 fill_continuation_entry(masm, reg_cont_obj, reg_is_virtual);
1819
1820 // If isContinue, call to thaw. Otherwise, call Continuation.enter(Continuation c, boolean isContinue)
1821 __ cmpwi(CCR0, reg_is_cont, 0);
1822 __ bne(CCR0, L_thaw);
1823
1824 // --- call Continuation.enter(Continuation c, boolean isContinue)
1825
1826 // Emit compiled static call
1827 // The call needs to be resolved. There's a special case for this in
1828 // SharedRuntime::find_callee_info_helper() which calls
1829 // LinkResolver::resolve_continuation_enter() which resolves the call to
1830 // Continuation.enter(Continuation c, boolean isContinue).
1831 address call_pc = __ pc();
1832 int start_offset = __ offset();
1833 // Put the entry point as a constant into the constant pool.
1834 const address entry_point_toc_addr = __ address_constant(resolve_static_call, RelocationHolder::none);
1835 const int entry_point_toc_offset = __ offset_to_method_toc(entry_point_toc_addr);
1836 guarantee(entry_point_toc_addr != nullptr, "const section overflow");
1837
1838 // Emit the trampoline stub which will be related to the branch-and-link below.
1839 address stub = __ emit_trampoline_stub(entry_point_toc_offset, start_offset);
1840 guarantee(stub != nullptr, "no space for trampoline stub");
1841
1842 __ relocate(relocInfo::static_call_type);
1843 // Note: At this point we do not have the address of the trampoline
1844 // stub, and the entry point might be too far away for bl, so __ pc()
1845 // serves as dummy and the bl will be patched later.
1846 __ bl(__ pc());
1847 oop_maps->add_gc_map(__ pc() - start, map);
1848 __ post_call_nop();
1849
1850 __ b(L_exit);
1851
1852 // --- Thawing path
1853
1854 __ bind(L_thaw);
1855 __ add_const_optimized(R0, R29_TOC, MacroAssembler::offset_to_global_toc(StubRoutines::cont_thaw()));
1856 __ mtctr(R0);
1857 __ bctrl();
1858 oop_maps->add_gc_map(__ pc() - start, map->deep_copy());
1859 ContinuationEntry::_return_pc_offset = __ pc() - start;
1860 __ post_call_nop();
1861
1862 // --- Normal exit (resolve/thawing)
1863
1864 __ bind(L_exit);
1865 continuation_enter_cleanup(masm);
1866
1867 // Pop frame and return
1868 DEBUG_ONLY(__ ld_ptr(R0, 0, R1_SP));
1869 __ addi(R1_SP, R1_SP, framesize_words*wordSize);
1870 DEBUG_ONLY(__ cmpd(CCR0, R0, R1_SP));
1871 __ asm_assert_eq(FILE_AND_LINE ": inconsistent frame size");
1872 __ ld(R0, _abi0(lr), R1_SP); // Return pc
1873 __ mtlr(R0);
1874 __ blr();
1875
1876 // --- Exception handling path
1877
1878 exception_offset = __ pc() - start;
1879
1880 continuation_enter_cleanup(masm);
1881 Register ex_pc = R17_tos; // nonvolatile register
1882 Register ex_oop = R15_esp; // nonvolatile register
1883 __ ld(ex_pc, _abi0(callers_sp), R1_SP); // Load caller's return pc
1884 __ ld(ex_pc, _abi0(lr), ex_pc);
1885 __ mr(ex_oop, R3_RET); // save return value containing the exception oop
1886 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address), R16_thread, ex_pc);
1887 __ mtlr(R3_RET); // the exception handler
1888 __ ld(R1_SP, _abi0(callers_sp), R1_SP); // remove enterSpecial frame
1889
1890 // Continue at exception handler
1891 // See OptoRuntime::generate_exception_blob for register arguments
1892 __ mr(R3_ARG1, ex_oop); // pass exception oop
1893 __ mr(R4_ARG2, ex_pc); // pass exception pc
1894 __ blr();
1895
1896 // static stub for the call above
1897 stub = CompiledDirectCall::emit_to_interp_stub(masm, call_pc);
1898 guarantee(stub != nullptr, "no space for static stub");
1899 }
1900
1901 static void gen_continuation_yield(MacroAssembler* masm,
1902 const VMRegPair* regs,
1903 OopMapSet* oop_maps,
1904 int& frame_complete,
1905 int& framesize_words,
1906 int& compiled_entry_offset) {
1907 Register tmp = R10_ARG8;
1908
1909 const int framesize_bytes = (int)align_up((int)frame::native_abi_reg_args_size, frame::alignment_in_bytes);
1910 framesize_words = framesize_bytes / wordSize;
1911
1912 address start = __ pc();
1913 compiled_entry_offset = __ pc() - start;
1914
1915 // Save return pc and push entry frame
1916 __ mflr(tmp);
1917 __ std(tmp, _abi0(lr), R1_SP); // SP->lr = return_pc
1918 __ push_frame(framesize_bytes , R0); // SP -= frame_size_in_bytes
1919
1920 DEBUG_ONLY(__ block_comment("Frame Complete"));
1921 frame_complete = __ pc() - start;
1922 address last_java_pc = __ pc();
1923
1924 // This nop must be exactly at the PC we push into the frame info.
1925 // We use this nop for fast CodeBlob lookup, associate the OopMap
1926 // with it right away.
1927 __ post_call_nop();
1928 OopMap* map = new OopMap(framesize_bytes / VMRegImpl::stack_slot_size, 1);
1929 oop_maps->add_gc_map(last_java_pc - start, map);
1930
1931 __ calculate_address_from_global_toc(tmp, last_java_pc); // will be relocated
1932 __ set_last_Java_frame(R1_SP, tmp);
1933 __ call_VM_leaf(Continuation::freeze_entry(), R16_thread, R1_SP);
1934 __ reset_last_Java_frame();
1935
1936 Label L_pinned;
1937
1938 __ cmpwi(CCR0, R3_RET, 0);
1939 __ bne(CCR0, L_pinned);
1940
1941 // yield succeeded
1942
1943 // Pop frames of continuation including this stub's frame
1944 __ ld_ptr(R1_SP, JavaThread::cont_entry_offset(), R16_thread);
1945 // The frame pushed by gen_continuation_enter is on top now again
1946 continuation_enter_cleanup(masm);
1947
1948 // Pop frame and return
1949 Label L_return;
1950 __ bind(L_return);
1951 __ pop_frame();
1952 __ ld(R0, _abi0(lr), R1_SP); // Return pc
1953 __ mtlr(R0);
1954 __ blr();
1955
1956 // yield failed - continuation is pinned
1957
1958 __ bind(L_pinned);
1959
1960 // handle pending exception thrown by freeze
1961 __ ld(tmp, in_bytes(JavaThread::pending_exception_offset()), R16_thread);
1962 __ cmpdi(CCR0, tmp, 0);
1963 __ beq(CCR0, L_return); // return if no exception is pending
1964 __ pop_frame();
1965 __ ld(R0, _abi0(lr), R1_SP); // Return pc
1966 __ mtlr(R0);
1967 __ load_const_optimized(tmp, StubRoutines::forward_exception_entry(), R0);
1968 __ mtctr(tmp);
1969 __ bctr();
1970 }
1971
1972 // ---------------------------------------------------------------------------
1973 // Generate a native wrapper for a given method. The method takes arguments
1974 // in the Java compiled code convention, marshals them to the native
1975 // convention (handlizes oops, etc), transitions to native, makes the call,
1976 // returns to java state (possibly blocking), unhandlizes any result and
1977 // returns.
1978 //
1979 // Critical native functions are a shorthand for the use of
1980 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1981 // functions. The wrapper is expected to unpack the arguments before
1982 // passing them to the callee. Critical native functions leave the state _in_Java,
1983 // since they cannot stop for GC.
1984 // Some other parts of JNI setup are skipped like the tear down of the JNI handle
1985 // block and the check for pending exceptions it's impossible for them
1986 // to be thrown.
1987 //
1988 nmethod *SharedRuntime::generate_native_wrapper(MacroAssembler *masm,
1989 const methodHandle& method,
1990 int compile_id,
1991 BasicType *in_sig_bt,
1992 VMRegPair *in_regs,
1993 BasicType ret_type) {
1994 if (method->is_continuation_native_intrinsic()) {
1995 int exception_offset = -1;
1996 OopMapSet* oop_maps = new OopMapSet();
1997 int frame_complete = -1;
1998 int stack_slots = -1;
1999 int interpreted_entry_offset = -1;
2000 int vep_offset = -1;
2001 if (method->is_continuation_enter_intrinsic()) {
2002 gen_continuation_enter(masm,
2003 in_regs,
2004 exception_offset,
2005 oop_maps,
2006 frame_complete,
2007 stack_slots,
2008 interpreted_entry_offset,
2009 vep_offset);
2010 } else if (method->is_continuation_yield_intrinsic()) {
2011 gen_continuation_yield(masm,
2012 in_regs,
2013 oop_maps,
2014 frame_complete,
2015 stack_slots,
2016 vep_offset);
2017 } else {
2018 guarantee(false, "Unknown Continuation native intrinsic");
2019 }
2020
2021 #ifdef ASSERT
2022 if (method->is_continuation_enter_intrinsic()) {
2023 assert(interpreted_entry_offset != -1, "Must be set");
2024 assert(exception_offset != -1, "Must be set");
2025 } else {
2026 assert(interpreted_entry_offset == -1, "Must be unset");
2027 assert(exception_offset == -1, "Must be unset");
2028 }
2029 assert(frame_complete != -1, "Must be set");
2030 assert(stack_slots != -1, "Must be set");
2031 assert(vep_offset != -1, "Must be set");
2032 #endif
2033
2034 __ flush();
2035 nmethod* nm = nmethod::new_native_nmethod(method,
2036 compile_id,
2037 masm->code(),
2038 vep_offset,
2039 frame_complete,
2040 stack_slots,
2041 in_ByteSize(-1),
2042 in_ByteSize(-1),
2043 oop_maps,
2044 exception_offset);
2045 if (nm == nullptr) return nm;
2046 if (method->is_continuation_enter_intrinsic()) {
2047 ContinuationEntry::set_enter_code(nm, interpreted_entry_offset);
2048 } else if (method->is_continuation_yield_intrinsic()) {
2049 _cont_doYield_stub = nm;
2050 }
2051 return nm;
2052 }
2053
2054 if (method->is_method_handle_intrinsic()) {
2055 vmIntrinsics::ID iid = method->intrinsic_id();
2056 intptr_t start = (intptr_t)__ pc();
2057 int vep_offset = ((intptr_t)__ pc()) - start;
2058 gen_special_dispatch(masm,
2059 method,
2060 in_sig_bt,
2061 in_regs);
2062 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
2063 __ flush();
2064 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
2065 return nmethod::new_native_nmethod(method,
2066 compile_id,
2067 masm->code(),
2068 vep_offset,
2069 frame_complete,
2070 stack_slots / VMRegImpl::slots_per_word,
2071 in_ByteSize(-1),
2072 in_ByteSize(-1),
2073 (OopMapSet*)nullptr);
2074 }
2075
2076 address native_func = method->native_function();
2077 assert(native_func != nullptr, "must have function");
2078
2079 // First, create signature for outgoing C call
2080 // --------------------------------------------------------------------------
2081
2082 int total_in_args = method->size_of_parameters();
2083 // We have received a description of where all the java args are located
2084 // on entry to the wrapper. We need to convert these args to where
2085 // the jni function will expect them. To figure out where they go
2086 // we convert the java signature to a C signature by inserting
2087 // the hidden arguments as arg[0] and possibly arg[1] (static method)
2088
2089 // Calculate the total number of C arguments and create arrays for the
2090 // signature and the outgoing registers.
2091 // On ppc64, we have two arrays for the outgoing registers, because
2092 // some floating-point arguments must be passed in registers _and_
2093 // in stack locations.
2094 bool method_is_static = method->is_static();
2095 int total_c_args = total_in_args + (method_is_static ? 2 : 1);
2096
2097 BasicType *out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
2098 VMRegPair *out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
2099 BasicType* in_elem_bt = nullptr;
2100
2101 // Create the signature for the C call:
2102 // 1) add the JNIEnv*
2103 // 2) add the class if the method is static
2104 // 3) copy the rest of the incoming signature (shifted by the number of
2105 // hidden arguments).
2106
2107 int argc = 0;
2108 out_sig_bt[argc++] = T_ADDRESS;
2109 if (method->is_static()) {
2110 out_sig_bt[argc++] = T_OBJECT;
2111 }
2112
2113 for (int i = 0; i < total_in_args ; i++ ) {
2114 out_sig_bt[argc++] = in_sig_bt[i];
2115 }
2116
2117
2118 // Compute the wrapper's frame size.
2119 // --------------------------------------------------------------------------
2120
2121 // Now figure out where the args must be stored and how much stack space
2122 // they require.
2123 //
2124 // Compute framesize for the wrapper. We need to handlize all oops in
2125 // incoming registers.
2126 //
2127 // Calculate the total number of stack slots we will need:
2128 // 1) abi requirements
2129 // 2) outgoing arguments
2130 // 3) space for inbound oop handle area
2131 // 4) space for handlizing a klass if static method
2132 // 5) space for a lock if synchronized method
2133 // 6) workspace for saving return values, int <-> float reg moves, etc.
2134 // 7) alignment
2135 //
2136 // Layout of the native wrapper frame:
2137 // (stack grows upwards, memory grows downwards)
2138 //
2139 // NW [ABI_REG_ARGS] <-- 1) R1_SP
2140 // [outgoing arguments] <-- 2) R1_SP + out_arg_slot_offset
2141 // [oopHandle area] <-- 3) R1_SP + oop_handle_offset
2142 // klass <-- 4) R1_SP + klass_offset
2143 // lock <-- 5) R1_SP + lock_offset
2144 // [workspace] <-- 6) R1_SP + workspace_offset
2145 // [alignment] (optional) <-- 7)
2146 // caller [JIT_TOP_ABI_48] <-- r_callers_sp
2147 //
2148 // - *_slot_offset Indicates offset from SP in number of stack slots.
2149 // - *_offset Indicates offset from SP in bytes.
2150
2151 int stack_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args) + // 1+2)
2152 SharedRuntime::out_preserve_stack_slots(); // See c_calling_convention.
2153
2154 // Now the space for the inbound oop handle area.
2155 int total_save_slots = num_java_iarg_registers * VMRegImpl::slots_per_word;
2156
2157 int oop_handle_slot_offset = stack_slots;
2158 stack_slots += total_save_slots; // 3)
2159
2160 int klass_slot_offset = 0;
2161 int klass_offset = -1;
2162 if (method_is_static) { // 4)
2163 klass_slot_offset = stack_slots;
2164 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
2165 stack_slots += VMRegImpl::slots_per_word;
2166 }
2167
2168 int lock_slot_offset = 0;
2169 int lock_offset = -1;
2170 if (method->is_synchronized()) { // 5)
2171 lock_slot_offset = stack_slots;
2172 lock_offset = lock_slot_offset * VMRegImpl::stack_slot_size;
2173 stack_slots += VMRegImpl::slots_per_word;
2174 }
2175
2176 int workspace_slot_offset = stack_slots; // 6)
2177 stack_slots += 2;
2178
2179 // Now compute actual number of stack words we need.
2180 // Rounding to make stack properly aligned.
2181 stack_slots = align_up(stack_slots, // 7)
2182 frame::alignment_in_bytes / VMRegImpl::stack_slot_size);
2183 int frame_size_in_bytes = stack_slots * VMRegImpl::stack_slot_size;
2184
2185
2186 // Now we can start generating code.
2187 // --------------------------------------------------------------------------
2188
2189 intptr_t start_pc = (intptr_t)__ pc();
2190 intptr_t vep_start_pc;
2191 intptr_t frame_done_pc;
2192 intptr_t oopmap_pc;
2193
2194 Label handle_pending_exception;
2195
2196 Register r_callers_sp = R21;
2197 Register r_temp_1 = R22;
2198 Register r_temp_2 = R23;
2199 Register r_temp_3 = R24;
2200 Register r_temp_4 = R25;
2201 Register r_temp_5 = R26;
2202 Register r_temp_6 = R27;
2203 Register r_return_pc = R28;
2204
2205 Register r_carg1_jnienv = noreg;
2206 Register r_carg2_classorobject = noreg;
2207 r_carg1_jnienv = out_regs[0].first()->as_Register();
2208 r_carg2_classorobject = out_regs[1].first()->as_Register();
2209
2210
2211 // Generate the Unverified Entry Point (UEP).
2212 // --------------------------------------------------------------------------
2213 assert(start_pc == (intptr_t)__ pc(), "uep must be at start");
2214
2215 // Check ic: object class == cached class?
2216 if (!method_is_static) {
2217 __ ic_check(4 /* end_alignment */);
2218 }
2219
2220 // Generate the Verified Entry Point (VEP).
2221 // --------------------------------------------------------------------------
2222 vep_start_pc = (intptr_t)__ pc();
2223
2224 if (VM_Version::supports_fast_class_init_checks() && method->needs_clinit_barrier()) {
2225 Label L_skip_barrier;
2226 Register klass = r_temp_1;
2227 // Notify OOP recorder (don't need the relocation)
2228 AddressLiteral md = __ constant_metadata_address(method->method_holder());
2229 __ load_const_optimized(klass, md.value(), R0);
2230 __ clinit_barrier(klass, R16_thread, &L_skip_barrier /*L_fast_path*/);
2231
2232 __ load_const_optimized(klass, SharedRuntime::get_handle_wrong_method_stub(), R0);
2233 __ mtctr(klass);
2234 __ bctr();
2235
2236 __ bind(L_skip_barrier);
2237 }
2238
2239 __ save_LR(r_temp_1);
2240 __ generate_stack_overflow_check(frame_size_in_bytes); // Check before creating frame.
2241 __ mr(r_callers_sp, R1_SP); // Remember frame pointer.
2242 __ push_frame(frame_size_in_bytes, r_temp_1); // Push the c2n adapter's frame.
2243
2244 BarrierSetAssembler* bs = BarrierSet::barrier_set()->barrier_set_assembler();
2245 bs->nmethod_entry_barrier(masm, r_temp_1);
2246
2247 frame_done_pc = (intptr_t)__ pc();
2248
2249 // Native nmethod wrappers never take possession of the oop arguments.
2250 // So the caller will gc the arguments.
2251 // The only thing we need an oopMap for is if the call is static.
2252 //
2253 // An OopMap for lock (and class if static), and one for the VM call itself.
2254 OopMapSet *oop_maps = new OopMapSet();
2255 OopMap *oop_map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
2256
2257 // Move arguments from register/stack to register/stack.
2258 // --------------------------------------------------------------------------
2259 //
2260 // We immediately shuffle the arguments so that for any vm call we have
2261 // to make from here on out (sync slow path, jvmti, etc.) we will have
2262 // captured the oops from our caller and have a valid oopMap for them.
2263 //
2264 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv*
2265 // (derived from JavaThread* which is in R16_thread) and, if static,
2266 // the class mirror instead of a receiver. This pretty much guarantees that
2267 // register layout will not match. We ignore these extra arguments during
2268 // the shuffle. The shuffle is described by the two calling convention
2269 // vectors we have in our possession. We simply walk the java vector to
2270 // get the source locations and the c vector to get the destinations.
2271
2272 // Record sp-based slot for receiver on stack for non-static methods.
2273 int receiver_offset = -1;
2274
2275 // We move the arguments backward because the floating point registers
2276 // destination will always be to a register with a greater or equal
2277 // register number or the stack.
2278 // in is the index of the incoming Java arguments
2279 // out is the index of the outgoing C arguments
2280
2281 #ifdef ASSERT
2282 bool reg_destroyed[Register::number_of_registers];
2283 bool freg_destroyed[FloatRegister::number_of_registers];
2284 for (int r = 0 ; r < Register::number_of_registers ; r++) {
2285 reg_destroyed[r] = false;
2286 }
2287 for (int f = 0 ; f < FloatRegister::number_of_registers ; f++) {
2288 freg_destroyed[f] = false;
2289 }
2290 #endif // ASSERT
2291
2292 for (int in = total_in_args - 1, out = total_c_args - 1; in >= 0 ; in--, out--) {
2293
2294 #ifdef ASSERT
2295 if (in_regs[in].first()->is_Register()) {
2296 assert(!reg_destroyed[in_regs[in].first()->as_Register()->encoding()], "ack!");
2297 } else if (in_regs[in].first()->is_FloatRegister()) {
2298 assert(!freg_destroyed[in_regs[in].first()->as_FloatRegister()->encoding()], "ack!");
2299 }
2300 if (out_regs[out].first()->is_Register()) {
2301 reg_destroyed[out_regs[out].first()->as_Register()->encoding()] = true;
2302 } else if (out_regs[out].first()->is_FloatRegister()) {
2303 freg_destroyed[out_regs[out].first()->as_FloatRegister()->encoding()] = true;
2304 }
2305 #endif // ASSERT
2306
2307 switch (in_sig_bt[in]) {
2308 case T_BOOLEAN:
2309 case T_CHAR:
2310 case T_BYTE:
2311 case T_SHORT:
2312 case T_INT:
2313 // Move int and do sign extension.
2314 int_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2315 break;
2316 case T_LONG:
2317 long_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2318 break;
2319 case T_ARRAY:
2320 case T_OBJECT:
2321 object_move(masm, stack_slots,
2322 oop_map, oop_handle_slot_offset,
2323 ((in == 0) && (!method_is_static)), &receiver_offset,
2324 in_regs[in], out_regs[out],
2325 r_callers_sp, r_temp_1, r_temp_2);
2326 break;
2327 case T_VOID:
2328 break;
2329 case T_FLOAT:
2330 float_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2331 break;
2332 case T_DOUBLE:
2333 double_move(masm, in_regs[in], out_regs[out], r_callers_sp, r_temp_1);
2334 break;
2335 case T_ADDRESS:
2336 fatal("found type (T_ADDRESS) in java args");
2337 break;
2338 default:
2339 ShouldNotReachHere();
2340 break;
2341 }
2342 }
2343
2344 // Pre-load a static method's oop into ARG2.
2345 // Used both by locking code and the normal JNI call code.
2346 if (method_is_static) {
2347 __ set_oop_constant(JNIHandles::make_local(method->method_holder()->java_mirror()),
2348 r_carg2_classorobject);
2349
2350 // Now handlize the static class mirror in carg2. It's known not-null.
2351 __ std(r_carg2_classorobject, klass_offset, R1_SP);
2352 oop_map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
2353 __ addi(r_carg2_classorobject, R1_SP, klass_offset);
2354 }
2355
2356 // Get JNIEnv* which is first argument to native.
2357 __ addi(r_carg1_jnienv, R16_thread, in_bytes(JavaThread::jni_environment_offset()));
2358
2359 // NOTE:
2360 //
2361 // We have all of the arguments setup at this point.
2362 // We MUST NOT touch any outgoing regs from this point on.
2363 // So if we must call out we must push a new frame.
2364
2365 // Get current pc for oopmap, and load it patchable relative to global toc.
2366 oopmap_pc = (intptr_t) __ pc();
2367 __ calculate_address_from_global_toc(r_return_pc, (address)oopmap_pc, true, true, true, true);
2368
2369 // We use the same pc/oopMap repeatedly when we call out.
2370 oop_maps->add_gc_map(oopmap_pc - start_pc, oop_map);
2371
2372 // r_return_pc now has the pc loaded that we will use when we finally call
2373 // to native.
2374
2375 // Make sure that thread is non-volatile; it crosses a bunch of VM calls below.
2376 assert(R16_thread->is_nonvolatile(), "thread must be in non-volatile register");
2377
2378 # if 0
2379 // DTrace method entry
2380 # endif
2381
2382 // Lock a synchronized method.
2383 // --------------------------------------------------------------------------
2384
2385 if (method->is_synchronized()) {
2386 Register r_oop = r_temp_4;
2387 const Register r_box = r_temp_5;
2388 Label done, locked;
2389
2390 // Load the oop for the object or class. r_carg2_classorobject contains
2391 // either the handlized oop from the incoming arguments or the handlized
2392 // class mirror (if the method is static).
2393 __ ld(r_oop, 0, r_carg2_classorobject);
2394
2395 // Get the lock box slot's address.
2396 __ addi(r_box, R1_SP, lock_offset);
2397
2398 // Try fastpath for locking.
2399 if (LockingMode == LM_LIGHTWEIGHT) {
2400 // fast_lock kills r_temp_1, r_temp_2, r_temp_3.
2401 __ compiler_fast_lock_lightweight_object(CCR0, r_oop, r_temp_1, r_temp_2, r_temp_3);
2402 } else {
2403 // fast_lock kills r_temp_1, r_temp_2, r_temp_3.
2404 __ compiler_fast_lock_object(CCR0, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3);
2405 }
2406 __ beq(CCR0, locked);
2407
2408 // None of the above fast optimizations worked so we have to get into the
2409 // slow case of monitor enter. Inline a special case of call_VM that
2410 // disallows any pending_exception.
2411
2412 // Save argument registers and leave room for C-compatible ABI_REG_ARGS.
2413 int frame_size = frame::native_abi_reg_args_size + align_up(total_c_args * wordSize, frame::alignment_in_bytes);
2414 __ mr(R11_scratch1, R1_SP);
2415 RegisterSaver::push_frame_and_save_argument_registers(masm, R12_scratch2, frame_size, total_c_args, out_regs);
2416
2417 // Do the call.
2418 __ set_last_Java_frame(R11_scratch1, r_return_pc);
2419 assert(r_return_pc->is_nonvolatile(), "expecting return pc to be in non-volatile register");
2420 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C), r_oop, r_box, R16_thread);
2421 __ reset_last_Java_frame();
2422
2423 RegisterSaver::restore_argument_registers_and_pop_frame(masm, frame_size, total_c_args, out_regs);
2424
2425 __ asm_assert_mem8_is_zero(thread_(pending_exception),
2426 "no pending exception allowed on exit from SharedRuntime::complete_monitor_locking_C");
2427
2428 __ bind(locked);
2429 }
2430
2431 // Use that pc we placed in r_return_pc a while back as the current frame anchor.
2432 __ set_last_Java_frame(R1_SP, r_return_pc);
2433
2434 // Publish thread state
2435 // --------------------------------------------------------------------------
2436
2437 // Transition from _thread_in_Java to _thread_in_native.
2438 __ li(R0, _thread_in_native);
2439 __ release();
2440 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2441 __ stw(R0, thread_(thread_state));
2442
2443
2444 // The JNI call
2445 // --------------------------------------------------------------------------
2446 #if defined(ABI_ELFv2)
2447 __ call_c(native_func, relocInfo::runtime_call_type);
2448 #else
2449 FunctionDescriptor* fd_native_method = (FunctionDescriptor*) native_func;
2450 __ call_c(fd_native_method, relocInfo::runtime_call_type);
2451 #endif
2452
2453
2454 // Now, we are back from the native code.
2455
2456
2457 // Unpack the native result.
2458 // --------------------------------------------------------------------------
2459
2460 // For int-types, we do any needed sign-extension required.
2461 // Care must be taken that the return values (R3_RET and F1_RET)
2462 // will survive any VM calls for blocking or unlocking.
2463 // An OOP result (handle) is done specially in the slow-path code.
2464
2465 switch (ret_type) {
2466 case T_VOID: break; // Nothing to do!
2467 case T_FLOAT: break; // Got it where we want it (unless slow-path).
2468 case T_DOUBLE: break; // Got it where we want it (unless slow-path).
2469 case T_LONG: break; // Got it where we want it (unless slow-path).
2470 case T_OBJECT: break; // Really a handle.
2471 // Cannot de-handlize until after reclaiming jvm_lock.
2472 case T_ARRAY: break;
2473
2474 case T_BOOLEAN: { // 0 -> false(0); !0 -> true(1)
2475 __ normalize_bool(R3_RET);
2476 break;
2477 }
2478 case T_BYTE: { // sign extension
2479 __ extsb(R3_RET, R3_RET);
2480 break;
2481 }
2482 case T_CHAR: { // unsigned result
2483 __ andi(R3_RET, R3_RET, 0xffff);
2484 break;
2485 }
2486 case T_SHORT: { // sign extension
2487 __ extsh(R3_RET, R3_RET);
2488 break;
2489 }
2490 case T_INT: // nothing to do
2491 break;
2492 default:
2493 ShouldNotReachHere();
2494 break;
2495 }
2496
2497 Label after_transition;
2498
2499 // Publish thread state
2500 // --------------------------------------------------------------------------
2501
2502 // Switch thread to "native transition" state before reading the
2503 // synchronization state. This additional state is necessary because reading
2504 // and testing the synchronization state is not atomic w.r.t. GC, as this
2505 // scenario demonstrates:
2506 // - Java thread A, in _thread_in_native state, loads _not_synchronized
2507 // and is preempted.
2508 // - VM thread changes sync state to synchronizing and suspends threads
2509 // for GC.
2510 // - Thread A is resumed to finish this native method, but doesn't block
2511 // here since it didn't see any synchronization in progress, and escapes.
2512
2513 // Transition from _thread_in_native to _thread_in_native_trans.
2514 __ li(R0, _thread_in_native_trans);
2515 __ release();
2516 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2517 __ stw(R0, thread_(thread_state));
2518
2519
2520 // Must we block?
2521 // --------------------------------------------------------------------------
2522
2523 // Block, if necessary, before resuming in _thread_in_Java state.
2524 // In order for GC to work, don't clear the last_Java_sp until after blocking.
2525 {
2526 Label no_block, sync;
2527
2528 // Force this write out before the read below.
2529 if (!UseSystemMemoryBarrier) {
2530 __ fence();
2531 }
2532
2533 Register sync_state_addr = r_temp_4;
2534 Register sync_state = r_temp_5;
2535 Register suspend_flags = r_temp_6;
2536
2537 // No synchronization in progress nor yet synchronized
2538 // (cmp-br-isync on one path, release (same as acquire on PPC64) on the other path).
2539 __ safepoint_poll(sync, sync_state, true /* at_return */, false /* in_nmethod */);
2540
2541 // Not suspended.
2542 // TODO: PPC port assert(4 == Thread::sz_suspend_flags(), "unexpected field size");
2543 __ lwz(suspend_flags, thread_(suspend_flags));
2544 __ cmpwi(CCR1, suspend_flags, 0);
2545 __ beq(CCR1, no_block);
2546
2547 // Block. Save any potential method result value before the operation and
2548 // use a leaf call to leave the last_Java_frame setup undisturbed. Doing this
2549 // lets us share the oopMap we used when we went native rather than create
2550 // a distinct one for this pc.
2551 __ bind(sync);
2552 __ isync();
2553
2554 address entry_point =
2555 CAST_FROM_FN_PTR(address, JavaThread::check_special_condition_for_native_trans);
2556 save_native_result(masm, ret_type, workspace_slot_offset);
2557 __ call_VM_leaf(entry_point, R16_thread);
2558 restore_native_result(masm, ret_type, workspace_slot_offset);
2559
2560 __ bind(no_block);
2561
2562 // Publish thread state.
2563 // --------------------------------------------------------------------------
2564
2565 // Thread state is thread_in_native_trans. Any safepoint blocking has
2566 // already happened so we can now change state to _thread_in_Java.
2567
2568 // Transition from _thread_in_native_trans to _thread_in_Java.
2569 __ li(R0, _thread_in_Java);
2570 __ lwsync(); // Acquire safepoint and suspend state, release thread state.
2571 // TODO: PPC port assert(4 == JavaThread::sz_thread_state(), "unexpected field size");
2572 __ stw(R0, thread_(thread_state));
2573 __ bind(after_transition);
2574 }
2575
2576 // Reguard any pages if necessary.
2577 // --------------------------------------------------------------------------
2578
2579 Label no_reguard;
2580 __ lwz(r_temp_1, thread_(stack_guard_state));
2581 __ cmpwi(CCR0, r_temp_1, StackOverflow::stack_guard_yellow_reserved_disabled);
2582 __ bne(CCR0, no_reguard);
2583
2584 save_native_result(masm, ret_type, workspace_slot_offset);
2585 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages));
2586 restore_native_result(masm, ret_type, workspace_slot_offset);
2587
2588 __ bind(no_reguard);
2589
2590
2591 // Unlock
2592 // --------------------------------------------------------------------------
2593
2594 if (method->is_synchronized()) {
2595 const Register r_oop = r_temp_4;
2596 const Register r_box = r_temp_5;
2597 const Register r_exception = r_temp_6;
2598 Label done;
2599
2600 // Get oop and address of lock object box.
2601 if (method_is_static) {
2602 assert(klass_offset != -1, "");
2603 __ ld(r_oop, klass_offset, R1_SP);
2604 } else {
2605 assert(receiver_offset != -1, "");
2606 __ ld(r_oop, receiver_offset, R1_SP);
2607 }
2608 __ addi(r_box, R1_SP, lock_offset);
2609
2610 // Try fastpath for unlocking.
2611 if (LockingMode == LM_LIGHTWEIGHT) {
2612 __ compiler_fast_unlock_lightweight_object(CCR0, r_oop, r_temp_1, r_temp_2, r_temp_3);
2613 } else {
2614 __ compiler_fast_unlock_object(CCR0, r_oop, r_box, r_temp_1, r_temp_2, r_temp_3);
2615 }
2616 __ beq(CCR0, done);
2617
2618 // Save and restore any potential method result value around the unlocking operation.
2619 save_native_result(masm, ret_type, workspace_slot_offset);
2620
2621 // Must save pending exception around the slow-path VM call. Since it's a
2622 // leaf call, the pending exception (if any) can be kept in a register.
2623 __ ld(r_exception, thread_(pending_exception));
2624 assert(r_exception->is_nonvolatile(), "exception register must be non-volatile");
2625 __ li(R0, 0);
2626 __ std(R0, thread_(pending_exception));
2627
2628 // Slow case of monitor enter.
2629 // Inline a special case of call_VM that disallows any pending_exception.
2630 // Arguments are (oop obj, BasicLock* lock, JavaThread* thread).
2631 __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C), r_oop, r_box, R16_thread);
2632
2633 __ asm_assert_mem8_is_zero(thread_(pending_exception),
2634 "no pending exception allowed on exit from SharedRuntime::complete_monitor_unlocking_C");
2635
2636 restore_native_result(masm, ret_type, workspace_slot_offset);
2637
2638 // Check_forward_pending_exception jump to forward_exception if any pending
2639 // exception is set. The forward_exception routine expects to see the
2640 // exception in pending_exception and not in a register. Kind of clumsy,
2641 // since all folks who branch to forward_exception must have tested
2642 // pending_exception first and hence have it in a register already.
2643 __ std(r_exception, thread_(pending_exception));
2644
2645 __ bind(done);
2646 }
2647
2648 # if 0
2649 // DTrace method exit
2650 # endif
2651
2652 // Clear "last Java frame" SP and PC.
2653 // --------------------------------------------------------------------------
2654
2655 __ reset_last_Java_frame();
2656
2657 // Unbox oop result, e.g. JNIHandles::resolve value.
2658 // --------------------------------------------------------------------------
2659
2660 if (is_reference_type(ret_type)) {
2661 __ resolve_jobject(R3_RET, r_temp_1, r_temp_2, MacroAssembler::PRESERVATION_NONE);
2662 }
2663
2664 if (CheckJNICalls) {
2665 // clear_pending_jni_exception_check
2666 __ load_const_optimized(R0, 0L);
2667 __ st_ptr(R0, JavaThread::pending_jni_exception_check_fn_offset(), R16_thread);
2668 }
2669
2670 // Reset handle block.
2671 // --------------------------------------------------------------------------
2672 __ ld(r_temp_1, thread_(active_handles));
2673 // TODO: PPC port assert(4 == JNIHandleBlock::top_size_in_bytes(), "unexpected field size");
2674 __ li(r_temp_2, 0);
2675 __ stw(r_temp_2, in_bytes(JNIHandleBlock::top_offset()), r_temp_1);
2676
2677
2678 // Check for pending exceptions.
2679 // --------------------------------------------------------------------------
2680 __ ld(r_temp_2, thread_(pending_exception));
2681 __ cmpdi(CCR0, r_temp_2, 0);
2682 __ bne(CCR0, handle_pending_exception);
2683
2684 // Return
2685 // --------------------------------------------------------------------------
2686
2687 __ pop_frame();
2688 __ restore_LR(R11);
2689 __ blr();
2690
2691
2692 // Handler for pending exceptions (out-of-line).
2693 // --------------------------------------------------------------------------
2694 // Since this is a native call, we know the proper exception handler
2695 // is the empty function. We just pop this frame and then jump to
2696 // forward_exception_entry.
2697 __ bind(handle_pending_exception);
2698
2699 __ pop_frame();
2700 __ restore_LR(R11);
2701 __ b64_patchable((address)StubRoutines::forward_exception_entry(),
2702 relocInfo::runtime_call_type);
2703
2704 // Done.
2705 // --------------------------------------------------------------------------
2706
2707 __ flush();
2708
2709 nmethod *nm = nmethod::new_native_nmethod(method,
2710 compile_id,
2711 masm->code(),
2712 vep_start_pc-start_pc,
2713 frame_done_pc-start_pc,
2714 stack_slots / VMRegImpl::slots_per_word,
2715 (method_is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2716 in_ByteSize(lock_offset),
2717 oop_maps);
2718
2719 return nm;
2720 }
2721
2722 // This function returns the adjust size (in number of words) to a c2i adapter
2723 // activation for use during deoptimization.
2724 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals) {
2725 return align_up((callee_locals - callee_parameters) * Interpreter::stackElementWords, frame::frame_alignment_in_words);
2726 }
2727
2728 uint SharedRuntime::in_preserve_stack_slots() {
2729 return frame::jit_in_preserve_size / VMRegImpl::stack_slot_size;
2730 }
2731
2732 uint SharedRuntime::out_preserve_stack_slots() {
2733 #if defined(COMPILER1) || defined(COMPILER2)
2734 return frame::jit_out_preserve_size / VMRegImpl::stack_slot_size;
2735 #else
2736 return 0;
2737 #endif
2738 }
2739
2740 #if defined(COMPILER1) || defined(COMPILER2)
2741 // Frame generation for deopt and uncommon trap blobs.
2742 static void push_skeleton_frame(MacroAssembler* masm, bool deopt,
2743 /* Read */
2744 Register unroll_block_reg,
2745 /* Update */
2746 Register frame_sizes_reg,
2747 Register number_of_frames_reg,
2748 Register pcs_reg,
2749 /* Invalidate */
2750 Register frame_size_reg,
2751 Register pc_reg) {
2752
2753 __ ld(pc_reg, 0, pcs_reg);
2754 __ ld(frame_size_reg, 0, frame_sizes_reg);
2755 __ std(pc_reg, _abi0(lr), R1_SP);
2756 __ push_frame(frame_size_reg, R0/*tmp*/);
2757 __ std(R1_SP, _ijava_state_neg(sender_sp), R1_SP);
2758 __ addi(number_of_frames_reg, number_of_frames_reg, -1);
2759 __ addi(frame_sizes_reg, frame_sizes_reg, wordSize);
2760 __ addi(pcs_reg, pcs_reg, wordSize);
2761 }
2762
2763 // Loop through the UnrollBlock info and create new frames.
2764 static void push_skeleton_frames(MacroAssembler* masm, bool deopt,
2765 /* read */
2766 Register unroll_block_reg,
2767 /* invalidate */
2768 Register frame_sizes_reg,
2769 Register number_of_frames_reg,
2770 Register pcs_reg,
2771 Register frame_size_reg,
2772 Register pc_reg) {
2773 Label loop;
2774
2775 // _number_of_frames is of type int (deoptimization.hpp)
2776 __ lwa(number_of_frames_reg,
2777 in_bytes(Deoptimization::UnrollBlock::number_of_frames_offset()),
2778 unroll_block_reg);
2779 __ ld(pcs_reg,
2780 in_bytes(Deoptimization::UnrollBlock::frame_pcs_offset()),
2781 unroll_block_reg);
2782 __ ld(frame_sizes_reg,
2783 in_bytes(Deoptimization::UnrollBlock::frame_sizes_offset()),
2784 unroll_block_reg);
2785
2786 // stack: (caller_of_deoptee, ...).
2787
2788 // At this point we either have an interpreter frame or a compiled
2789 // frame on top of stack. If it is a compiled frame we push a new c2i
2790 // adapter here
2791
2792 // Memorize top-frame stack-pointer.
2793 __ mr(frame_size_reg/*old_sp*/, R1_SP);
2794
2795 // Resize interpreter top frame OR C2I adapter.
2796
2797 // At this moment, the top frame (which is the caller of the deoptee) is
2798 // an interpreter frame or a newly pushed C2I adapter or an entry frame.
2799 // The top frame has a TOP_IJAVA_FRAME_ABI and the frame contains the
2800 // outgoing arguments.
2801 //
2802 // In order to push the interpreter frame for the deoptee, we need to
2803 // resize the top frame such that we are able to place the deoptee's
2804 // locals in the frame.
2805 // Additionally, we have to turn the top frame's TOP_IJAVA_FRAME_ABI
2806 // into a valid PARENT_IJAVA_FRAME_ABI.
2807
2808 __ lwa(R11_scratch1,
2809 in_bytes(Deoptimization::UnrollBlock::caller_adjustment_offset()),
2810 unroll_block_reg);
2811 __ neg(R11_scratch1, R11_scratch1);
2812
2813 // R11_scratch1 contains size of locals for frame resizing.
2814 // R12_scratch2 contains top frame's lr.
2815
2816 // Resize frame by complete frame size prevents TOC from being
2817 // overwritten by locals. A more stack space saving way would be
2818 // to copy the TOC to its location in the new abi.
2819 __ addi(R11_scratch1, R11_scratch1, - frame::parent_ijava_frame_abi_size);
2820
2821 // now, resize the frame
2822 __ resize_frame(R11_scratch1, pc_reg/*tmp*/);
2823
2824 // In the case where we have resized a c2i frame above, the optional
2825 // alignment below the locals has size 32 (why?).
2826 __ std(R12_scratch2, _abi0(lr), R1_SP);
2827
2828 // Initialize initial_caller_sp.
2829 __ std(frame_size_reg, _ijava_state_neg(sender_sp), R1_SP);
2830
2831 #ifdef ASSERT
2832 // Make sure that there is at least one entry in the array.
2833 __ cmpdi(CCR0, number_of_frames_reg, 0);
2834 __ asm_assert_ne("array_size must be > 0");
2835 #endif
2836
2837 // Now push the new interpreter frames.
2838 //
2839 __ bind(loop);
2840 // Allocate a new frame, fill in the pc.
2841 push_skeleton_frame(masm, deopt,
2842 unroll_block_reg,
2843 frame_sizes_reg,
2844 number_of_frames_reg,
2845 pcs_reg,
2846 frame_size_reg,
2847 pc_reg);
2848 __ cmpdi(CCR0, number_of_frames_reg, 0);
2849 __ bne(CCR0, loop);
2850
2851 // Get the return address pointing into the frame manager.
2852 __ ld(R0, 0, pcs_reg);
2853 // Store it in the top interpreter frame.
2854 __ std(R0, _abi0(lr), R1_SP);
2855 // Initialize frame_manager_lr of interpreter top frame.
2856 }
2857 #endif
2858
2859 void SharedRuntime::generate_deopt_blob() {
2860 // Allocate space for the code
2861 ResourceMark rm;
2862 // Setup code generation tools
2863 CodeBuffer buffer("deopt_blob", 2048, 1024);
2864 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer);
2865 Label exec_mode_initialized;
2866 int frame_size_in_words;
2867 OopMap* map = nullptr;
2868 OopMapSet *oop_maps = new OopMapSet();
2869
2870 // size of ABI112 plus spill slots for R3_RET and F1_RET.
2871 const int frame_size_in_bytes = frame::native_abi_reg_args_spill_size;
2872 const int frame_size_in_slots = frame_size_in_bytes / sizeof(jint);
2873 int first_frame_size_in_bytes = 0; // frame size of "unpack frame" for call to fetch_unroll_info.
2874
2875 const Register exec_mode_reg = R21_tmp1;
2876
2877 const address start = __ pc();
2878
2879 #if defined(COMPILER1) || defined(COMPILER2)
2880 // --------------------------------------------------------------------------
2881 // Prolog for non exception case!
2882
2883 // We have been called from the deopt handler of the deoptee.
2884 //
2885 // deoptee:
2886 // ...
2887 // call X
2888 // ...
2889 // deopt_handler: call_deopt_stub
2890 // cur. return pc --> ...
2891 //
2892 // So currently SR_LR points behind the call in the deopt handler.
2893 // We adjust it such that it points to the start of the deopt handler.
2894 // The return_pc has been stored in the frame of the deoptee and
2895 // will replace the address of the deopt_handler in the call
2896 // to Deoptimization::fetch_unroll_info below.
2897 // We can't grab a free register here, because all registers may
2898 // contain live values, so let the RegisterSaver do the adjustment
2899 // of the return pc.
2900 const int return_pc_adjustment_no_exception = -MacroAssembler::bl64_patchable_size;
2901
2902 // Push the "unpack frame"
2903 // Save everything in sight.
2904 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
2905 &first_frame_size_in_bytes,
2906 /*generate_oop_map=*/ true,
2907 return_pc_adjustment_no_exception,
2908 RegisterSaver::return_pc_is_lr);
2909 assert(map != nullptr, "OopMap must have been created");
2910
2911 __ li(exec_mode_reg, Deoptimization::Unpack_deopt);
2912 // Save exec mode for unpack_frames.
2913 __ b(exec_mode_initialized);
2914
2915 // --------------------------------------------------------------------------
2916 // Prolog for exception case
2917
2918 // An exception is pending.
2919 // We have been called with a return (interpreter) or a jump (exception blob).
2920 //
2921 // - R3_ARG1: exception oop
2922 // - R4_ARG2: exception pc
2923
2924 int exception_offset = __ pc() - start;
2925
2926 BLOCK_COMMENT("Prolog for exception case");
2927
2928 // Store exception oop and pc in thread (location known to GC).
2929 // This is needed since the call to "fetch_unroll_info()" may safepoint.
2930 __ std(R3_ARG1, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
2931 __ std(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
2932 __ std(R4_ARG2, _abi0(lr), R1_SP);
2933
2934 // Vanilla deoptimization with an exception pending in exception_oop.
2935 int exception_in_tls_offset = __ pc() - start;
2936
2937 // Push the "unpack frame".
2938 // Save everything in sight.
2939 RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
2940 &first_frame_size_in_bytes,
2941 /*generate_oop_map=*/ false,
2942 /*return_pc_adjustment_exception=*/ 0,
2943 RegisterSaver::return_pc_is_pre_saved);
2944
2945 // Deopt during an exception. Save exec mode for unpack_frames.
2946 __ li(exec_mode_reg, Deoptimization::Unpack_exception);
2947
2948 // fall through
2949
2950 int reexecute_offset = 0;
2951 #ifdef COMPILER1
2952 __ b(exec_mode_initialized);
2953
2954 // Reexecute entry, similar to c2 uncommon trap
2955 reexecute_offset = __ pc() - start;
2956
2957 RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
2958 &first_frame_size_in_bytes,
2959 /*generate_oop_map=*/ false,
2960 /*return_pc_adjustment_reexecute=*/ 0,
2961 RegisterSaver::return_pc_is_pre_saved);
2962 __ li(exec_mode_reg, Deoptimization::Unpack_reexecute);
2963 #endif
2964
2965 // --------------------------------------------------------------------------
2966 __ BIND(exec_mode_initialized);
2967
2968 const Register unroll_block_reg = R22_tmp2;
2969
2970 // We need to set `last_Java_frame' because `fetch_unroll_info' will
2971 // call `last_Java_frame()'. The value of the pc in the frame is not
2972 // particularly important. It just needs to identify this blob.
2973 __ set_last_Java_frame(R1_SP, noreg);
2974
2975 // With EscapeAnalysis turned on, this call may safepoint!
2976 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info), R16_thread, exec_mode_reg);
2977 address calls_return_pc = __ last_calls_return_pc();
2978 // Set an oopmap for the call site that describes all our saved registers.
2979 oop_maps->add_gc_map(calls_return_pc - start, map);
2980
2981 __ reset_last_Java_frame();
2982 // Save the return value.
2983 __ mr(unroll_block_reg, R3_RET);
2984
2985 // Restore only the result registers that have been saved
2986 // by save_volatile_registers(...).
2987 RegisterSaver::restore_result_registers(masm, first_frame_size_in_bytes);
2988
2989 // reload the exec mode from the UnrollBlock (it might have changed)
2990 __ lwz(exec_mode_reg, in_bytes(Deoptimization::UnrollBlock::unpack_kind_offset()), unroll_block_reg);
2991 // In excp_deopt_mode, restore and clear exception oop which we
2992 // stored in the thread during exception entry above. The exception
2993 // oop will be the return value of this stub.
2994 Label skip_restore_excp;
2995 __ cmpdi(CCR0, exec_mode_reg, Deoptimization::Unpack_exception);
2996 __ bne(CCR0, skip_restore_excp);
2997 __ ld(R3_RET, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
2998 __ ld(R4_ARG2, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
2999 __ li(R0, 0);
3000 __ std(R0, in_bytes(JavaThread::exception_pc_offset()), R16_thread);
3001 __ std(R0, in_bytes(JavaThread::exception_oop_offset()), R16_thread);
3002 __ BIND(skip_restore_excp);
3003
3004 __ pop_frame();
3005
3006 // stack: (deoptee, optional i2c, caller of deoptee, ...).
3007
3008 // pop the deoptee's frame
3009 __ pop_frame();
3010
3011 // stack: (caller_of_deoptee, ...).
3012
3013 // Freezing continuation frames requires that the caller is trimmed to unextended sp if compiled.
3014 // If not compiled the loaded value is equal to the current SP (see frame::initial_deoptimization_info())
3015 // and the frame is effectively not resized.
3016 Register caller_sp = R23_tmp3;
3017 __ ld_ptr(caller_sp, Deoptimization::UnrollBlock::initial_info_offset(), unroll_block_reg);
3018 __ resize_frame_absolute(caller_sp, R24_tmp4, R25_tmp5);
3019
3020 // Loop through the `UnrollBlock' info and create interpreter frames.
3021 push_skeleton_frames(masm, true/*deopt*/,
3022 unroll_block_reg,
3023 R23_tmp3,
3024 R24_tmp4,
3025 R25_tmp5,
3026 R26_tmp6,
3027 R27_tmp7);
3028
3029 // stack: (skeletal interpreter frame, ..., optional skeletal
3030 // interpreter frame, optional c2i, caller of deoptee, ...).
3031
3032 // push an `unpack_frame' taking care of float / int return values.
3033 __ push_frame(frame_size_in_bytes, R0/*tmp*/);
3034
3035 // stack: (unpack frame, skeletal interpreter frame, ..., optional
3036 // skeletal interpreter frame, optional c2i, caller of deoptee,
3037 // ...).
3038
3039 // Spill live volatile registers since we'll do a call.
3040 __ std( R3_RET, _native_abi_reg_args_spill(spill_ret), R1_SP);
3041 __ stfd(F1_RET, _native_abi_reg_args_spill(spill_fret), R1_SP);
3042
3043 // Let the unpacker layout information in the skeletal frames just
3044 // allocated.
3045 __ calculate_address_from_global_toc(R3_RET, calls_return_pc, true, true, true, true);
3046 __ set_last_Java_frame(/*sp*/R1_SP, /*pc*/R3_RET);
3047 // This is a call to a LEAF method, so no oop map is required.
3048 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames),
3049 R16_thread/*thread*/, exec_mode_reg/*exec_mode*/);
3050 __ reset_last_Java_frame();
3051
3052 // Restore the volatiles saved above.
3053 __ ld( R3_RET, _native_abi_reg_args_spill(spill_ret), R1_SP);
3054 __ lfd(F1_RET, _native_abi_reg_args_spill(spill_fret), R1_SP);
3055
3056 // Pop the unpack frame.
3057 __ pop_frame();
3058 __ restore_LR(R0);
3059
3060 // stack: (top interpreter frame, ..., optional interpreter frame,
3061 // optional c2i, caller of deoptee, ...).
3062
3063 // Initialize R14_state.
3064 __ restore_interpreter_state(R11_scratch1);
3065 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
3066
3067 // Return to the interpreter entry point.
3068 __ blr();
3069 __ flush();
3070 #else // COMPILER2
3071 __ unimplemented("deopt blob needed only with compiler");
3072 int exception_offset = __ pc() - start;
3073 #endif // COMPILER2
3074
3075 _deopt_blob = DeoptimizationBlob::create(&buffer, oop_maps, 0, exception_offset,
3076 reexecute_offset, first_frame_size_in_bytes / wordSize);
3077 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
3078 }
3079
3080 #ifdef COMPILER2
3081 void SharedRuntime::generate_uncommon_trap_blob() {
3082 // Allocate space for the code.
3083 ResourceMark rm;
3084 // Setup code generation tools.
3085 CodeBuffer buffer("uncommon_trap_blob", 2048, 1024);
3086 InterpreterMacroAssembler* masm = new InterpreterMacroAssembler(&buffer);
3087 address start = __ pc();
3088
3089 Register unroll_block_reg = R21_tmp1;
3090 Register klass_index_reg = R22_tmp2;
3091 Register unc_trap_reg = R23_tmp3;
3092 Register r_return_pc = R27_tmp7;
3093
3094 OopMapSet* oop_maps = new OopMapSet();
3095 int frame_size_in_bytes = frame::native_abi_reg_args_size;
3096 OopMap* map = new OopMap(frame_size_in_bytes / sizeof(jint), 0);
3097
3098 // stack: (deoptee, optional i2c, caller_of_deoptee, ...).
3099
3100 // Push a dummy `unpack_frame' and call
3101 // `Deoptimization::uncommon_trap' to pack the compiled frame into a
3102 // vframe array and return the `UnrollBlock' information.
3103
3104 // Save LR to compiled frame.
3105 __ save_LR(R11_scratch1);
3106
3107 // Push an "uncommon_trap" frame.
3108 __ push_frame_reg_args(0, R11_scratch1);
3109
3110 // stack: (unpack frame, deoptee, optional i2c, caller_of_deoptee, ...).
3111
3112 // Set the `unpack_frame' as last_Java_frame.
3113 // `Deoptimization::uncommon_trap' expects it and considers its
3114 // sender frame as the deoptee frame.
3115 // Remember the offset of the instruction whose address will be
3116 // moved to R11_scratch1.
3117 address gc_map_pc = __ pc();
3118 __ calculate_address_from_global_toc(r_return_pc, gc_map_pc, true, true, true, true);
3119 __ set_last_Java_frame(/*sp*/R1_SP, r_return_pc);
3120
3121 __ mr(klass_index_reg, R3);
3122 __ li(R5_ARG3, Deoptimization::Unpack_uncommon_trap);
3123 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap),
3124 R16_thread, klass_index_reg, R5_ARG3);
3125
3126 // Set an oopmap for the call site.
3127 oop_maps->add_gc_map(gc_map_pc - start, map);
3128
3129 __ reset_last_Java_frame();
3130
3131 // Pop the `unpack frame'.
3132 __ pop_frame();
3133
3134 // stack: (deoptee, optional i2c, caller_of_deoptee, ...).
3135
3136 // Save the return value.
3137 __ mr(unroll_block_reg, R3_RET);
3138
3139 // Pop the uncommon_trap frame.
3140 __ pop_frame();
3141
3142 // stack: (caller_of_deoptee, ...).
3143
3144 #ifdef ASSERT
3145 __ lwz(R22_tmp2, in_bytes(Deoptimization::UnrollBlock::unpack_kind_offset()), unroll_block_reg);
3146 __ cmpdi(CCR0, R22_tmp2, (unsigned)Deoptimization::Unpack_uncommon_trap);
3147 __ asm_assert_eq("SharedRuntime::generate_deopt_blob: expected Unpack_uncommon_trap");
3148 #endif
3149
3150 // Freezing continuation frames requires that the caller is trimmed to unextended sp if compiled.
3151 // If not compiled the loaded value is equal to the current SP (see frame::initial_deoptimization_info())
3152 // and the frame is effectively not resized.
3153 Register caller_sp = R23_tmp3;
3154 __ ld_ptr(caller_sp, Deoptimization::UnrollBlock::initial_info_offset(), unroll_block_reg);
3155 __ resize_frame_absolute(caller_sp, R24_tmp4, R25_tmp5);
3156
3157 // Allocate new interpreter frame(s) and possibly a c2i adapter
3158 // frame.
3159 push_skeleton_frames(masm, false/*deopt*/,
3160 unroll_block_reg,
3161 R22_tmp2,
3162 R23_tmp3,
3163 R24_tmp4,
3164 R25_tmp5,
3165 R26_tmp6);
3166
3167 // stack: (skeletal interpreter frame, ..., optional skeletal
3168 // interpreter frame, optional c2i, caller of deoptee, ...).
3169
3170 // Push a dummy `unpack_frame' taking care of float return values.
3171 // Call `Deoptimization::unpack_frames' to layout information in the
3172 // interpreter frames just created.
3173
3174 // Push a simple "unpack frame" here.
3175 __ push_frame_reg_args(0, R11_scratch1);
3176
3177 // stack: (unpack frame, skeletal interpreter frame, ..., optional
3178 // skeletal interpreter frame, optional c2i, caller of deoptee,
3179 // ...).
3180
3181 // Set the "unpack_frame" as last_Java_frame.
3182 __ set_last_Java_frame(/*sp*/R1_SP, r_return_pc);
3183
3184 // Indicate it is the uncommon trap case.
3185 __ li(unc_trap_reg, Deoptimization::Unpack_uncommon_trap);
3186 // Let the unpacker layout information in the skeletal frames just
3187 // allocated.
3188 __ call_VM_leaf(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames),
3189 R16_thread, unc_trap_reg);
3190
3191 __ reset_last_Java_frame();
3192 // Pop the `unpack frame'.
3193 __ pop_frame();
3194 // Restore LR from top interpreter frame.
3195 __ restore_LR(R11_scratch1);
3196
3197 // stack: (top interpreter frame, ..., optional interpreter frame,
3198 // optional c2i, caller of deoptee, ...).
3199
3200 __ restore_interpreter_state(R11_scratch1);
3201 __ load_const_optimized(R25_templateTableBase, (address)Interpreter::dispatch_table((TosState)0), R11_scratch1);
3202
3203 // Return to the interpreter entry point.
3204 __ blr();
3205
3206 masm->flush();
3207
3208 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, frame_size_in_bytes/wordSize);
3209 }
3210 #endif // COMPILER2
3211
3212 // Generate a special Compile2Runtime blob that saves all registers, and setup oopmap.
3213 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) {
3214 assert(StubRoutines::forward_exception_entry() != nullptr,
3215 "must be generated before");
3216
3217 ResourceMark rm;
3218 OopMapSet *oop_maps = new OopMapSet();
3219 OopMap* map;
3220
3221 // Allocate space for the code. Setup code generation tools.
3222 CodeBuffer buffer("handler_blob", 2048, 1024);
3223 MacroAssembler* masm = new MacroAssembler(&buffer);
3224
3225 address start = __ pc();
3226 int frame_size_in_bytes = 0;
3227
3228 RegisterSaver::ReturnPCLocation return_pc_location;
3229 bool cause_return = (poll_type == POLL_AT_RETURN);
3230 if (cause_return) {
3231 // Nothing to do here. The frame has already been popped in MachEpilogNode.
3232 // Register LR already contains the return pc.
3233 return_pc_location = RegisterSaver::return_pc_is_pre_saved;
3234 } else {
3235 // Use thread()->saved_exception_pc() as return pc.
3236 return_pc_location = RegisterSaver::return_pc_is_thread_saved_exception_pc;
3237 }
3238
3239 bool save_vectors = (poll_type == POLL_AT_VECTOR_LOOP);
3240
3241 // Save registers, fpu state, and flags. Set R31 = return pc.
3242 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
3243 &frame_size_in_bytes,
3244 /*generate_oop_map=*/ true,
3245 /*return_pc_adjustment=*/0,
3246 return_pc_location, save_vectors);
3247
3248 // The following is basically a call_VM. However, we need the precise
3249 // address of the call in order to generate an oopmap. Hence, we do all the
3250 // work ourselves.
3251 __ set_last_Java_frame(/*sp=*/R1_SP, /*pc=*/noreg);
3252
3253 // The return address must always be correct so that the frame constructor
3254 // never sees an invalid pc.
3255
3256 // Do the call
3257 __ call_VM_leaf(call_ptr, R16_thread);
3258 address calls_return_pc = __ last_calls_return_pc();
3259
3260 // Set an oopmap for the call site. This oopmap will map all
3261 // oop-registers and debug-info registers as callee-saved. This
3262 // will allow deoptimization at this safepoint to find all possible
3263 // debug-info recordings, as well as let GC find all oops.
3264 oop_maps->add_gc_map(calls_return_pc - start, map);
3265
3266 Label noException;
3267
3268 // Clear the last Java frame.
3269 __ reset_last_Java_frame();
3270
3271 BLOCK_COMMENT(" Check pending exception.");
3272 const Register pending_exception = R0;
3273 __ ld(pending_exception, thread_(pending_exception));
3274 __ cmpdi(CCR0, pending_exception, 0);
3275 __ beq(CCR0, noException);
3276
3277 // Exception pending
3278 RegisterSaver::restore_live_registers_and_pop_frame(masm,
3279 frame_size_in_bytes,
3280 /*restore_ctr=*/true, save_vectors);
3281
3282 BLOCK_COMMENT(" Jump to forward_exception_entry.");
3283 // Jump to forward_exception_entry, with the issuing PC in LR
3284 // so it looks like the original nmethod called forward_exception_entry.
3285 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
3286
3287 // No exception case.
3288 __ BIND(noException);
3289
3290 if (!cause_return) {
3291 Label no_adjust;
3292 // If our stashed return pc was modified by the runtime we avoid touching it
3293 __ ld(R0, frame_size_in_bytes + _abi0(lr), R1_SP);
3294 __ cmpd(CCR0, R0, R31);
3295 __ bne(CCR0, no_adjust);
3296
3297 // Adjust return pc forward to step over the safepoint poll instruction
3298 __ addi(R31, R31, 4);
3299 __ std(R31, frame_size_in_bytes + _abi0(lr), R1_SP);
3300
3301 __ bind(no_adjust);
3302 }
3303
3304 // Normal exit, restore registers and exit.
3305 RegisterSaver::restore_live_registers_and_pop_frame(masm,
3306 frame_size_in_bytes,
3307 /*restore_ctr=*/true, save_vectors);
3308
3309 __ blr();
3310
3311 // Make sure all code is generated
3312 masm->flush();
3313
3314 // Fill-out other meta info
3315 // CodeBlob frame size is in words.
3316 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_bytes / wordSize);
3317 }
3318
3319 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss)
3320 //
3321 // Generate a stub that calls into the vm to find out the proper destination
3322 // of a java call. All the argument registers are live at this point
3323 // but since this is generic code we don't know what they are and the caller
3324 // must do any gc of the args.
3325 //
3326 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) {
3327
3328 // allocate space for the code
3329 ResourceMark rm;
3330
3331 CodeBuffer buffer(name, 1000, 512);
3332 MacroAssembler* masm = new MacroAssembler(&buffer);
3333
3334 int frame_size_in_bytes;
3335
3336 OopMapSet *oop_maps = new OopMapSet();
3337 OopMap* map = nullptr;
3338
3339 address start = __ pc();
3340
3341 map = RegisterSaver::push_frame_reg_args_and_save_live_registers(masm,
3342 &frame_size_in_bytes,
3343 /*generate_oop_map*/ true,
3344 /*return_pc_adjustment*/ 0,
3345 RegisterSaver::return_pc_is_lr);
3346
3347 // Use noreg as last_Java_pc, the return pc will be reconstructed
3348 // from the physical frame.
3349 __ set_last_Java_frame(/*sp*/R1_SP, noreg);
3350
3351 int frame_complete = __ offset();
3352
3353 // Pass R19_method as 2nd (optional) argument, used by
3354 // counter_overflow_stub.
3355 __ call_VM_leaf(destination, R16_thread, R19_method);
3356 address calls_return_pc = __ last_calls_return_pc();
3357 // Set an oopmap for the call site.
3358 // We need this not only for callee-saved registers, but also for volatile
3359 // registers that the compiler might be keeping live across a safepoint.
3360 // Create the oopmap for the call's return pc.
3361 oop_maps->add_gc_map(calls_return_pc - start, map);
3362
3363 // R3_RET contains the address we are going to jump to assuming no exception got installed.
3364
3365 // clear last_Java_sp
3366 __ reset_last_Java_frame();
3367
3368 // Check for pending exceptions.
3369 BLOCK_COMMENT("Check for pending exceptions.");
3370 Label pending;
3371 __ ld(R11_scratch1, thread_(pending_exception));
3372 __ cmpdi(CCR0, R11_scratch1, 0);
3373 __ bne(CCR0, pending);
3374
3375 __ mtctr(R3_RET); // Ctr will not be touched by restore_live_registers_and_pop_frame.
3376
3377 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ false);
3378
3379 // Get the returned method.
3380 __ get_vm_result_2(R19_method);
3381
3382 __ bctr();
3383
3384
3385 // Pending exception after the safepoint.
3386 __ BIND(pending);
3387
3388 RegisterSaver::restore_live_registers_and_pop_frame(masm, frame_size_in_bytes, /*restore_ctr*/ true);
3389
3390 // exception pending => remove activation and forward to exception handler
3391
3392 __ li(R11_scratch1, 0);
3393 __ ld(R3_ARG1, thread_(pending_exception));
3394 __ std(R11_scratch1, in_bytes(JavaThread::vm_result_offset()), R16_thread);
3395 __ b64_patchable(StubRoutines::forward_exception_entry(), relocInfo::runtime_call_type);
3396
3397 // -------------
3398 // Make sure all code is generated.
3399 masm->flush();
3400
3401 // return the blob
3402 // frame_size_words or bytes??
3403 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_in_bytes/wordSize,
3404 oop_maps, true);
3405 }
3406
3407
3408 //------------------------------Montgomery multiplication------------------------
3409 //
3410
3411 // Subtract 0:b from carry:a. Return carry.
3412 static unsigned long
3413 sub(unsigned long a[], unsigned long b[], unsigned long carry, long len) {
3414 long i = 0;
3415 unsigned long tmp, tmp2;
3416 __asm__ __volatile__ (
3417 "subfc %[tmp], %[tmp], %[tmp] \n" // pre-set CA
3418 "mtctr %[len] \n"
3419 "0: \n"
3420 "ldx %[tmp], %[i], %[a] \n"
3421 "ldx %[tmp2], %[i], %[b] \n"
3422 "subfe %[tmp], %[tmp2], %[tmp] \n" // subtract extended
3423 "stdx %[tmp], %[i], %[a] \n"
3424 "addi %[i], %[i], 8 \n"
3425 "bdnz 0b \n"
3426 "addme %[tmp], %[carry] \n" // carry + CA - 1
3427 : [i]"+b"(i), [tmp]"=&r"(tmp), [tmp2]"=&r"(tmp2)
3428 : [a]"r"(a), [b]"r"(b), [carry]"r"(carry), [len]"r"(len)
3429 : "ctr", "xer", "memory"
3430 );
3431 return tmp;
3432 }
3433
3434 // Multiply (unsigned) Long A by Long B, accumulating the double-
3435 // length result into the accumulator formed of T0, T1, and T2.
3436 inline void MACC(unsigned long A, unsigned long B, unsigned long &T0, unsigned long &T1, unsigned long &T2) {
3437 unsigned long hi, lo;
3438 __asm__ __volatile__ (
3439 "mulld %[lo], %[A], %[B] \n"
3440 "mulhdu %[hi], %[A], %[B] \n"
3441 "addc %[T0], %[T0], %[lo] \n"
3442 "adde %[T1], %[T1], %[hi] \n"
3443 "addze %[T2], %[T2] \n"
3444 : [hi]"=&r"(hi), [lo]"=&r"(lo), [T0]"+r"(T0), [T1]"+r"(T1), [T2]"+r"(T2)
3445 : [A]"r"(A), [B]"r"(B)
3446 : "xer"
3447 );
3448 }
3449
3450 // As above, but add twice the double-length result into the
3451 // accumulator.
3452 inline void MACC2(unsigned long A, unsigned long B, unsigned long &T0, unsigned long &T1, unsigned long &T2) {
3453 unsigned long hi, lo;
3454 __asm__ __volatile__ (
3455 "mulld %[lo], %[A], %[B] \n"
3456 "mulhdu %[hi], %[A], %[B] \n"
3457 "addc %[T0], %[T0], %[lo] \n"
3458 "adde %[T1], %[T1], %[hi] \n"
3459 "addze %[T2], %[T2] \n"
3460 "addc %[T0], %[T0], %[lo] \n"
3461 "adde %[T1], %[T1], %[hi] \n"
3462 "addze %[T2], %[T2] \n"
3463 : [hi]"=&r"(hi), [lo]"=&r"(lo), [T0]"+r"(T0), [T1]"+r"(T1), [T2]"+r"(T2)
3464 : [A]"r"(A), [B]"r"(B)
3465 : "xer"
3466 );
3467 }
3468
3469 // Fast Montgomery multiplication. The derivation of the algorithm is
3470 // in "A Cryptographic Library for the Motorola DSP56000,
3471 // Dusse and Kaliski, Proc. EUROCRYPT 90, pp. 230-237".
3472 static void
3473 montgomery_multiply(unsigned long a[], unsigned long b[], unsigned long n[],
3474 unsigned long m[], unsigned long inv, int len) {
3475 unsigned long t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator
3476 int i;
3477
3478 assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply");
3479
3480 for (i = 0; i < len; i++) {
3481 int j;
3482 for (j = 0; j < i; j++) {
3483 MACC(a[j], b[i-j], t0, t1, t2);
3484 MACC(m[j], n[i-j], t0, t1, t2);
3485 }
3486 MACC(a[i], b[0], t0, t1, t2);
3487 m[i] = t0 * inv;
3488 MACC(m[i], n[0], t0, t1, t2);
3489
3490 assert(t0 == 0, "broken Montgomery multiply");
3491
3492 t0 = t1; t1 = t2; t2 = 0;
3493 }
3494
3495 for (i = len; i < 2*len; i++) {
3496 int j;
3497 for (j = i-len+1; j < len; j++) {
3498 MACC(a[j], b[i-j], t0, t1, t2);
3499 MACC(m[j], n[i-j], t0, t1, t2);
3500 }
3501 m[i-len] = t0;
3502 t0 = t1; t1 = t2; t2 = 0;
3503 }
3504
3505 while (t0) {
3506 t0 = sub(m, n, t0, len);
3507 }
3508 }
3509
3510 // Fast Montgomery squaring. This uses asymptotically 25% fewer
3511 // multiplies so it should be up to 25% faster than Montgomery
3512 // multiplication. However, its loop control is more complex and it
3513 // may actually run slower on some machines.
3514 static void
3515 montgomery_square(unsigned long a[], unsigned long n[],
3516 unsigned long m[], unsigned long inv, int len) {
3517 unsigned long t0 = 0, t1 = 0, t2 = 0; // Triple-precision accumulator
3518 int i;
3519
3520 assert(inv * n[0] == -1UL, "broken inverse in Montgomery multiply");
3521
3522 for (i = 0; i < len; i++) {
3523 int j;
3524 int end = (i+1)/2;
3525 for (j = 0; j < end; j++) {
3526 MACC2(a[j], a[i-j], t0, t1, t2);
3527 MACC(m[j], n[i-j], t0, t1, t2);
3528 }
3529 if ((i & 1) == 0) {
3530 MACC(a[j], a[j], t0, t1, t2);
3531 }
3532 for (; j < i; j++) {
3533 MACC(m[j], n[i-j], t0, t1, t2);
3534 }
3535 m[i] = t0 * inv;
3536 MACC(m[i], n[0], t0, t1, t2);
3537
3538 assert(t0 == 0, "broken Montgomery square");
3539
3540 t0 = t1; t1 = t2; t2 = 0;
3541 }
3542
3543 for (i = len; i < 2*len; i++) {
3544 int start = i-len+1;
3545 int end = start + (len - start)/2;
3546 int j;
3547 for (j = start; j < end; j++) {
3548 MACC2(a[j], a[i-j], t0, t1, t2);
3549 MACC(m[j], n[i-j], t0, t1, t2);
3550 }
3551 if ((i & 1) == 0) {
3552 MACC(a[j], a[j], t0, t1, t2);
3553 }
3554 for (; j < len; j++) {
3555 MACC(m[j], n[i-j], t0, t1, t2);
3556 }
3557 m[i-len] = t0;
3558 t0 = t1; t1 = t2; t2 = 0;
3559 }
3560
3561 while (t0) {
3562 t0 = sub(m, n, t0, len);
3563 }
3564 }
3565
3566 // The threshold at which squaring is advantageous was determined
3567 // experimentally on an i7-3930K (Ivy Bridge) CPU @ 3.5GHz.
3568 // Doesn't seem to be relevant for Power8 so we use the same value.
3569 #define MONTGOMERY_SQUARING_THRESHOLD 64
3570
3571 // Copy len longwords from s to d, word-swapping as we go. The
3572 // destination array is reversed.
3573 static void reverse_words(unsigned long *s, unsigned long *d, int len) {
3574 d += len;
3575 while(len-- > 0) {
3576 d--;
3577 unsigned long s_val = *s;
3578 // Swap words in a longword on little endian machines.
3579 #ifdef VM_LITTLE_ENDIAN
3580 s_val = (s_val << 32) | (s_val >> 32);
3581 #endif
3582 *d = s_val;
3583 s++;
3584 }
3585 }
3586
3587 void SharedRuntime::montgomery_multiply(jint *a_ints, jint *b_ints, jint *n_ints,
3588 jint len, jlong inv,
3589 jint *m_ints) {
3590 len = len & 0x7fffFFFF; // C2 does not respect int to long conversion for stub calls.
3591 assert(len % 2 == 0, "array length in montgomery_multiply must be even");
3592 int longwords = len/2;
3593
3594 // Make very sure we don't use so much space that the stack might
3595 // overflow. 512 jints corresponds to an 16384-bit integer and
3596 // will use here a total of 8k bytes of stack space.
3597 int divisor = sizeof(unsigned long) * 4;
3598 guarantee(longwords <= 8192 / divisor, "must be");
3599 int total_allocation = longwords * sizeof (unsigned long) * 4;
3600 unsigned long *scratch = (unsigned long *)alloca(total_allocation);
3601
3602 // Local scratch arrays
3603 unsigned long
3604 *a = scratch + 0 * longwords,
3605 *b = scratch + 1 * longwords,
3606 *n = scratch + 2 * longwords,
3607 *m = scratch + 3 * longwords;
3608
3609 reverse_words((unsigned long *)a_ints, a, longwords);
3610 reverse_words((unsigned long *)b_ints, b, longwords);
3611 reverse_words((unsigned long *)n_ints, n, longwords);
3612
3613 ::montgomery_multiply(a, b, n, m, (unsigned long)inv, longwords);
3614
3615 reverse_words(m, (unsigned long *)m_ints, longwords);
3616 }
3617
3618 void SharedRuntime::montgomery_square(jint *a_ints, jint *n_ints,
3619 jint len, jlong inv,
3620 jint *m_ints) {
3621 len = len & 0x7fffFFFF; // C2 does not respect int to long conversion for stub calls.
3622 assert(len % 2 == 0, "array length in montgomery_square must be even");
3623 int longwords = len/2;
3624
3625 // Make very sure we don't use so much space that the stack might
3626 // overflow. 512 jints corresponds to an 16384-bit integer and
3627 // will use here a total of 6k bytes of stack space.
3628 int divisor = sizeof(unsigned long) * 3;
3629 guarantee(longwords <= (8192 / divisor), "must be");
3630 int total_allocation = longwords * sizeof (unsigned long) * 3;
3631 unsigned long *scratch = (unsigned long *)alloca(total_allocation);
3632
3633 // Local scratch arrays
3634 unsigned long
3635 *a = scratch + 0 * longwords,
3636 *n = scratch + 1 * longwords,
3637 *m = scratch + 2 * longwords;
3638
3639 reverse_words((unsigned long *)a_ints, a, longwords);
3640 reverse_words((unsigned long *)n_ints, n, longwords);
3641
3642 if (len >= MONTGOMERY_SQUARING_THRESHOLD) {
3643 ::montgomery_square(a, n, m, (unsigned long)inv, longwords);
3644 } else {
3645 ::montgomery_multiply(a, a, n, m, (unsigned long)inv, longwords);
3646 }
3647
3648 reverse_words(m, (unsigned long *)m_ints, longwords);
3649 }