1 /*
2 * Copyright (c) 2003, 2017, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
24
25 #include "precompiled.hpp"
26 #include "asm/macroAssembler.hpp"
27 #include "asm/macroAssembler.inline.hpp"
28 #include "code/debugInfoRec.hpp"
29 #include "code/icBuffer.hpp"
30 #include "code/vtableStubs.hpp"
31 #include "interpreter/interpreter.hpp"
32 #include "logging/log.hpp"
33 #include "memory/resourceArea.hpp"
34 #include "oops/compiledICHolder.hpp"
35 #include "runtime/sharedRuntime.hpp"
36 #include "runtime/vframeArray.hpp"
37 #include "vmreg_x86.inline.hpp"
38 #ifdef COMPILER1
39 #include "c1/c1_Runtime1.hpp"
40 #endif
41 #ifdef COMPILER2
42 #include "opto/runtime.hpp"
43 #endif
44 #include "vm_version_x86.hpp"
45
46 #define __ masm->
47
48 const int StackAlignmentInSlots = StackAlignmentInBytes / VMRegImpl::stack_slot_size;
49
50 class RegisterSaver {
51 // Capture info about frame layout
52 #define DEF_XMM_OFFS(regnum) xmm ## regnum ## _off = xmm_off + (regnum)*16/BytesPerInt, xmm ## regnum ## H_off
53 enum layout {
54 fpu_state_off = 0,
55 fpu_state_end = fpu_state_off+FPUStateSizeInWords,
56 st0_off, st0H_off,
57 st1_off, st1H_off,
58 st2_off, st2H_off,
59 st3_off, st3H_off,
60 st4_off, st4H_off,
61 st5_off, st5H_off,
62 st6_off, st6H_off,
63 st7_off, st7H_off,
64 xmm_off,
65 DEF_XMM_OFFS(0),
66 DEF_XMM_OFFS(1),
67 DEF_XMM_OFFS(2),
68 DEF_XMM_OFFS(3),
69 DEF_XMM_OFFS(4),
70 DEF_XMM_OFFS(5),
71 DEF_XMM_OFFS(6),
72 DEF_XMM_OFFS(7),
73 flags_off = xmm7_off + 16/BytesPerInt + 1, // 16-byte stack alignment fill word
74 rdi_off,
75 rsi_off,
76 ignore_off, // extra copy of rbp,
77 rsp_off,
78 rbx_off,
79 rdx_off,
80 rcx_off,
81 rax_off,
82 // The frame sender code expects that rbp will be in the "natural" place and
83 // will override any oopMap setting for it. We must therefore force the layout
84 // so that it agrees with the frame sender code.
85 rbp_off,
86 return_off, // slot for return address
87 reg_save_size };
88 enum { FPU_regs_live = flags_off - fpu_state_end };
89
90 public:
91
92 static OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words,
93 int* total_frame_words, bool verify_fpu = true, bool save_vectors = false);
94 static void restore_live_registers(MacroAssembler* masm, bool restore_vectors = false);
95
96 static int rax_offset() { return rax_off; }
97 static int rbx_offset() { return rbx_off; }
98
99 // Offsets into the register save area
100 // Used by deoptimization when it is managing result register
101 // values on its own
102
103 static int raxOffset(void) { return rax_off; }
104 static int rdxOffset(void) { return rdx_off; }
105 static int rbxOffset(void) { return rbx_off; }
106 static int xmm0Offset(void) { return xmm0_off; }
107 // This really returns a slot in the fp save area, which one is not important
108 static int fpResultOffset(void) { return st0_off; }
109
110 // During deoptimization only the result register need to be restored
111 // all the other values have already been extracted.
112
113 static void restore_result_registers(MacroAssembler* masm);
114
115 };
116
117 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words,
118 int* total_frame_words, bool verify_fpu, bool save_vectors) {
119 int num_xmm_regs = XMMRegisterImpl::number_of_registers;
120 int ymm_bytes = num_xmm_regs * 16;
121 int zmm_bytes = num_xmm_regs * 32;
122 #ifdef COMPILER2
123 if (save_vectors) {
124 assert(UseAVX > 0, "Vectors larger than 16 byte long are supported only with AVX");
125 assert(MaxVectorSize <= 64, "Only up to 64 byte long vectors are supported");
126 // Save upper half of YMM registers
127 int vect_bytes = ymm_bytes;
128 if (UseAVX > 2) {
129 // Save upper half of ZMM registers as well
130 vect_bytes += zmm_bytes;
131 }
132 additional_frame_words += vect_bytes / wordSize;
133 }
134 #else
135 assert(!save_vectors, "vectors are generated only by C2");
136 #endif
137 int frame_size_in_bytes = (reg_save_size + additional_frame_words) * wordSize;
138 int frame_words = frame_size_in_bytes / wordSize;
139 *total_frame_words = frame_words;
140
141 assert(FPUStateSizeInWords == 27, "update stack layout");
142
143 // save registers, fpu state, and flags
144 // We assume caller has already has return address slot on the stack
145 // We push epb twice in this sequence because we want the real rbp,
146 // to be under the return like a normal enter and we want to use pusha
147 // We push by hand instead of pusing push
148 __ enter();
149 __ pusha();
150 __ pushf();
151 __ subptr(rsp,FPU_regs_live*wordSize); // Push FPU registers space
152 __ push_FPU_state(); // Save FPU state & init
153
154 if (verify_fpu) {
155 // Some stubs may have non standard FPU control word settings so
156 // only check and reset the value when it required to be the
157 // standard value. The safepoint blob in particular can be used
158 // in methods which are using the 24 bit control word for
159 // optimized float math.
160
161 #ifdef ASSERT
162 // Make sure the control word has the expected value
163 Label ok;
164 __ cmpw(Address(rsp, 0), StubRoutines::fpu_cntrl_wrd_std());
165 __ jccb(Assembler::equal, ok);
166 __ stop("corrupted control word detected");
167 __ bind(ok);
168 #endif
169
170 // Reset the control word to guard against exceptions being unmasked
171 // since fstp_d can cause FPU stack underflow exceptions. Write it
172 // into the on stack copy and then reload that to make sure that the
173 // current and future values are correct.
174 __ movw(Address(rsp, 0), StubRoutines::fpu_cntrl_wrd_std());
175 }
176
177 __ frstor(Address(rsp, 0));
178 if (!verify_fpu) {
179 // Set the control word so that exceptions are masked for the
180 // following code.
181 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std()));
182 }
183
184 int off = st0_off;
185 int delta = st1_off - off;
186
187 // Save the FPU registers in de-opt-able form
188 for (int n = 0; n < FloatRegisterImpl::number_of_registers; n++) {
189 __ fstp_d(Address(rsp, off*wordSize));
190 off += delta;
191 }
192
193 off = xmm0_off;
194 delta = xmm1_off - off;
195 if(UseSSE == 1) {
196 // Save the XMM state
197 for (int n = 0; n < num_xmm_regs; n++) {
198 __ movflt(Address(rsp, off*wordSize), as_XMMRegister(n));
199 off += delta;
200 }
201 } else if(UseSSE >= 2) {
202 // Save whole 128bit (16 bytes) XMM registers
203 for (int n = 0; n < num_xmm_regs; n++) {
204 __ movdqu(Address(rsp, off*wordSize), as_XMMRegister(n));
205 off += delta;
206 }
207 }
208
209 if (save_vectors) {
210 __ subptr(rsp, ymm_bytes);
211 // Save upper half of YMM registers
212 for (int n = 0; n < num_xmm_regs; n++) {
213 __ vextractf128_high(Address(rsp, n*16), as_XMMRegister(n));
214 }
215 if (UseAVX > 2) {
216 __ subptr(rsp, zmm_bytes);
217 // Save upper half of ZMM registers
218 for (int n = 0; n < num_xmm_regs; n++) {
219 __ vextractf64x4_high(Address(rsp, n*32), as_XMMRegister(n));
220 }
221 }
222 }
223 __ vzeroupper();
224
225 // Set an oopmap for the call site. This oopmap will map all
226 // oop-registers and debug-info registers as callee-saved. This
227 // will allow deoptimization at this safepoint to find all possible
228 // debug-info recordings, as well as let GC find all oops.
229
230 OopMapSet *oop_maps = new OopMapSet();
231 OopMap* map = new OopMap( frame_words, 0 );
232
233 #define STACK_OFFSET(x) VMRegImpl::stack2reg((x) + additional_frame_words)
234 #define NEXTREG(x) (x)->as_VMReg()->next()
235
236 map->set_callee_saved(STACK_OFFSET(rax_off), rax->as_VMReg());
237 map->set_callee_saved(STACK_OFFSET(rcx_off), rcx->as_VMReg());
238 map->set_callee_saved(STACK_OFFSET(rdx_off), rdx->as_VMReg());
239 map->set_callee_saved(STACK_OFFSET(rbx_off), rbx->as_VMReg());
240 // rbp, location is known implicitly, no oopMap
241 map->set_callee_saved(STACK_OFFSET(rsi_off), rsi->as_VMReg());
242 map->set_callee_saved(STACK_OFFSET(rdi_off), rdi->as_VMReg());
243 // %%% This is really a waste but we'll keep things as they were for now for the upper component
244 off = st0_off;
245 delta = st1_off - off;
246 for (int n = 0; n < FloatRegisterImpl::number_of_registers; n++) {
247 FloatRegister freg_name = as_FloatRegister(n);
248 map->set_callee_saved(STACK_OFFSET(off), freg_name->as_VMReg());
249 map->set_callee_saved(STACK_OFFSET(off+1), NEXTREG(freg_name));
250 off += delta;
251 }
252 off = xmm0_off;
253 delta = xmm1_off - off;
254 for (int n = 0; n < num_xmm_regs; n++) {
255 XMMRegister xmm_name = as_XMMRegister(n);
256 map->set_callee_saved(STACK_OFFSET(off), xmm_name->as_VMReg());
257 map->set_callee_saved(STACK_OFFSET(off+1), NEXTREG(xmm_name));
258 off += delta;
259 }
260 #undef NEXTREG
261 #undef STACK_OFFSET
262
263 return map;
264 }
265
266 void RegisterSaver::restore_live_registers(MacroAssembler* masm, bool restore_vectors) {
267 int num_xmm_regs = XMMRegisterImpl::number_of_registers;
268 int ymm_bytes = num_xmm_regs * 16;
269 int zmm_bytes = num_xmm_regs * 32;
270 // Recover XMM & FPU state
271 int additional_frame_bytes = 0;
272 #ifdef COMPILER2
273 if (restore_vectors) {
274 assert(UseAVX > 0, "Vectors larger than 16 byte long are supported only with AVX");
275 assert(MaxVectorSize <= 64, "Only up to 64 byte long vectors are supported");
276 // Save upper half of YMM registers
277 additional_frame_bytes = ymm_bytes;
278 if (UseAVX > 2) {
279 // Save upper half of ZMM registers as well
280 additional_frame_bytes += zmm_bytes;
281 }
282 }
283 #else
284 assert(!restore_vectors, "vectors are generated only by C2");
285 #endif
286
287 int off = xmm0_off;
288 int delta = xmm1_off - off;
289
290 __ vzeroupper();
291
292 if (UseSSE == 1) {
293 // Restore XMM registers
294 assert(additional_frame_bytes == 0, "");
295 for (int n = 0; n < num_xmm_regs; n++) {
296 __ movflt(as_XMMRegister(n), Address(rsp, off*wordSize));
297 off += delta;
298 }
299 } else if (UseSSE >= 2) {
300 // Restore whole 128bit (16 bytes) XMM registers. Do this before restoring YMM and
301 // ZMM because the movdqu instruction zeros the upper part of the XMM register.
302 for (int n = 0; n < num_xmm_regs; n++) {
303 __ movdqu(as_XMMRegister(n), Address(rsp, off*wordSize+additional_frame_bytes));
304 off += delta;
305 }
306 }
307
308 if (restore_vectors) {
309 if (UseAVX > 2) {
310 // Restore upper half of ZMM registers.
311 for (int n = 0; n < num_xmm_regs; n++) {
312 __ vinsertf64x4_high(as_XMMRegister(n), Address(rsp, n*32));
313 }
314 __ addptr(rsp, zmm_bytes);
315 }
316 // Restore upper half of YMM registers.
317 for (int n = 0; n < num_xmm_regs; n++) {
318 __ vinsertf128_high(as_XMMRegister(n), Address(rsp, n*16));
319 }
320 __ addptr(rsp, ymm_bytes);
321 }
322
323 __ pop_FPU_state();
324 __ addptr(rsp, FPU_regs_live*wordSize); // Pop FPU registers
325
326 __ popf();
327 __ popa();
328 // Get the rbp, described implicitly by the frame sender code (no oopMap)
329 __ pop(rbp);
330 }
331
332 void RegisterSaver::restore_result_registers(MacroAssembler* masm) {
333
334 // Just restore result register. Only used by deoptimization. By
335 // now any callee save register that needs to be restore to a c2
336 // caller of the deoptee has been extracted into the vframeArray
337 // and will be stuffed into the c2i adapter we create for later
338 // restoration so only result registers need to be restored here.
339 //
340
341 __ frstor(Address(rsp, 0)); // Restore fpu state
342
343 // Recover XMM & FPU state
344 if( UseSSE == 1 ) {
345 __ movflt(xmm0, Address(rsp, xmm0_off*wordSize));
346 } else if( UseSSE >= 2 ) {
347 __ movdbl(xmm0, Address(rsp, xmm0_off*wordSize));
348 }
349 __ movptr(rax, Address(rsp, rax_off*wordSize));
350 __ movptr(rdx, Address(rsp, rdx_off*wordSize));
351 // Pop all of the register save are off the stack except the return address
352 __ addptr(rsp, return_off * wordSize);
353 }
354
355 // Is vector's size (in bytes) bigger than a size saved by default?
356 // 16 bytes XMM registers are saved by default using SSE2 movdqu instructions.
357 // Note, MaxVectorSize == 0 with UseSSE < 2 and vectors are not generated.
358 bool SharedRuntime::is_wide_vector(int size) {
359 return size > 16;
360 }
361
362 size_t SharedRuntime::trampoline_size() {
363 return 16;
364 }
365
366 void SharedRuntime::generate_trampoline(MacroAssembler *masm, address destination) {
367 __ jump(RuntimeAddress(destination));
368 }
369
370 // The java_calling_convention describes stack locations as ideal slots on
371 // a frame with no abi restrictions. Since we must observe abi restrictions
372 // (like the placement of the register window) the slots must be biased by
373 // the following value.
374 static int reg2offset_in(VMReg r) {
375 // Account for saved rbp, and return address
376 // This should really be in_preserve_stack_slots
377 return (r->reg2stack() + 2) * VMRegImpl::stack_slot_size;
378 }
379
380 static int reg2offset_out(VMReg r) {
381 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
382 }
383
384 // ---------------------------------------------------------------------------
385 // Read the array of BasicTypes from a signature, and compute where the
386 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
387 // quantities. Values less than SharedInfo::stack0 are registers, those above
388 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
389 // as framesizes are fixed.
390 // VMRegImpl::stack0 refers to the first slot 0(sp).
391 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. Register
392 // up to RegisterImpl::number_of_registers) are the 32-bit
393 // integer registers.
394
395 // Pass first two oop/int args in registers ECX and EDX.
396 // Pass first two float/double args in registers XMM0 and XMM1.
397 // Doubles have precedence, so if you pass a mix of floats and doubles
398 // the doubles will grab the registers before the floats will.
399
400 // Note: the INPUTS in sig_bt are in units of Java argument words, which are
401 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit
402 // units regardless of build. Of course for i486 there is no 64 bit build
403
404
405 // ---------------------------------------------------------------------------
406 // The compiled Java calling convention.
407 // Pass first two oop/int args in registers ECX and EDX.
408 // Pass first two float/double args in registers XMM0 and XMM1.
409 // Doubles have precedence, so if you pass a mix of floats and doubles
410 // the doubles will grab the registers before the floats will.
411 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
412 VMRegPair *regs,
413 int total_args_passed,
414 int is_outgoing) {
415 uint stack = 0; // Starting stack position for args on stack
416
417
418 // Pass first two oop/int args in registers ECX and EDX.
419 uint reg_arg0 = 9999;
420 uint reg_arg1 = 9999;
421
422 // Pass first two float/double args in registers XMM0 and XMM1.
423 // Doubles have precedence, so if you pass a mix of floats and doubles
424 // the doubles will grab the registers before the floats will.
425 // CNC - TURNED OFF FOR non-SSE.
426 // On Intel we have to round all doubles (and most floats) at
427 // call sites by storing to the stack in any case.
428 // UseSSE=0 ==> Don't Use ==> 9999+0
429 // UseSSE=1 ==> Floats only ==> 9999+1
430 // UseSSE>=2 ==> Floats or doubles ==> 9999+2
431 enum { fltarg_dontuse = 9999+0, fltarg_float_only = 9999+1, fltarg_flt_dbl = 9999+2 };
432 uint fargs = (UseSSE>=2) ? 2 : UseSSE;
433 uint freg_arg0 = 9999+fargs;
434 uint freg_arg1 = 9999+fargs;
435
436 // Pass doubles & longs aligned on the stack. First count stack slots for doubles
437 int i;
438 for( i = 0; i < total_args_passed; i++) {
439 if( sig_bt[i] == T_DOUBLE ) {
440 // first 2 doubles go in registers
441 if( freg_arg0 == fltarg_flt_dbl ) freg_arg0 = i;
442 else if( freg_arg1 == fltarg_flt_dbl ) freg_arg1 = i;
443 else // Else double is passed low on the stack to be aligned.
444 stack += 2;
445 } else if( sig_bt[i] == T_LONG ) {
446 stack += 2;
447 }
448 }
449 int dstack = 0; // Separate counter for placing doubles
450
451 // Now pick where all else goes.
452 for( i = 0; i < total_args_passed; i++) {
453 // From the type and the argument number (count) compute the location
454 switch( sig_bt[i] ) {
455 case T_SHORT:
456 case T_CHAR:
457 case T_BYTE:
458 case T_BOOLEAN:
459 case T_INT:
460 case T_ARRAY:
461 case T_OBJECT:
462 case T_ADDRESS:
463 if( reg_arg0 == 9999 ) {
464 reg_arg0 = i;
465 regs[i].set1(rcx->as_VMReg());
466 } else if( reg_arg1 == 9999 ) {
467 reg_arg1 = i;
468 regs[i].set1(rdx->as_VMReg());
469 } else {
470 regs[i].set1(VMRegImpl::stack2reg(stack++));
471 }
472 break;
473 case T_FLOAT:
474 if( freg_arg0 == fltarg_flt_dbl || freg_arg0 == fltarg_float_only ) {
475 freg_arg0 = i;
476 regs[i].set1(xmm0->as_VMReg());
477 } else if( freg_arg1 == fltarg_flt_dbl || freg_arg1 == fltarg_float_only ) {
478 freg_arg1 = i;
479 regs[i].set1(xmm1->as_VMReg());
480 } else {
481 regs[i].set1(VMRegImpl::stack2reg(stack++));
482 }
483 break;
484 case T_LONG:
485 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "missing Half" );
486 regs[i].set2(VMRegImpl::stack2reg(dstack));
487 dstack += 2;
488 break;
489 case T_DOUBLE:
490 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "missing Half" );
491 if( freg_arg0 == (uint)i ) {
492 regs[i].set2(xmm0->as_VMReg());
493 } else if( freg_arg1 == (uint)i ) {
494 regs[i].set2(xmm1->as_VMReg());
495 } else {
496 regs[i].set2(VMRegImpl::stack2reg(dstack));
497 dstack += 2;
498 }
499 break;
500 case T_VOID: regs[i].set_bad(); break;
501 break;
502 default:
503 ShouldNotReachHere();
504 break;
505 }
506 }
507
508 // return value can be odd number of VMRegImpl stack slots make multiple of 2
509 return align_up(stack, 2);
510 }
511
512 // Patch the callers callsite with entry to compiled code if it exists.
513 static void patch_callers_callsite(MacroAssembler *masm) {
514 Label L;
515 __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), (int32_t)NULL_WORD);
516 __ jcc(Assembler::equal, L);
517 // Schedule the branch target address early.
518 // Call into the VM to patch the caller, then jump to compiled callee
519 // rax, isn't live so capture return address while we easily can
520 __ movptr(rax, Address(rsp, 0));
521 __ pusha();
522 __ pushf();
523
524 if (UseSSE == 1) {
525 __ subptr(rsp, 2*wordSize);
526 __ movflt(Address(rsp, 0), xmm0);
527 __ movflt(Address(rsp, wordSize), xmm1);
528 }
529 if (UseSSE >= 2) {
530 __ subptr(rsp, 4*wordSize);
531 __ movdbl(Address(rsp, 0), xmm0);
532 __ movdbl(Address(rsp, 2*wordSize), xmm1);
533 }
534 #ifdef COMPILER2
535 // C2 may leave the stack dirty if not in SSE2+ mode
536 if (UseSSE >= 2) {
537 __ verify_FPU(0, "c2i transition should have clean FPU stack");
538 } else {
539 __ empty_FPU_stack();
540 }
541 #endif /* COMPILER2 */
542
543 // VM needs caller's callsite
544 __ push(rax);
545 // VM needs target method
546 __ push(rbx);
547 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite)));
548 __ addptr(rsp, 2*wordSize);
549
550 if (UseSSE == 1) {
551 __ movflt(xmm0, Address(rsp, 0));
552 __ movflt(xmm1, Address(rsp, wordSize));
553 __ addptr(rsp, 2*wordSize);
554 }
555 if (UseSSE >= 2) {
556 __ movdbl(xmm0, Address(rsp, 0));
557 __ movdbl(xmm1, Address(rsp, 2*wordSize));
558 __ addptr(rsp, 4*wordSize);
559 }
560
561 __ popf();
562 __ popa();
563 __ bind(L);
564 }
565
566
567 static void move_c2i_double(MacroAssembler *masm, XMMRegister r, int st_off) {
568 int next_off = st_off - Interpreter::stackElementSize;
569 __ movdbl(Address(rsp, next_off), r);
570 }
571
572 static void gen_c2i_adapter(MacroAssembler *masm,
573 int total_args_passed,
574 int comp_args_on_stack,
575 const BasicType *sig_bt,
576 const VMRegPair *regs,
577 Label& skip_fixup) {
578 // Before we get into the guts of the C2I adapter, see if we should be here
579 // at all. We've come from compiled code and are attempting to jump to the
580 // interpreter, which means the caller made a static call to get here
581 // (vcalls always get a compiled target if there is one). Check for a
582 // compiled target. If there is one, we need to patch the caller's call.
583 patch_callers_callsite(masm);
584
585 __ bind(skip_fixup);
586
587 #ifdef COMPILER2
588 // C2 may leave the stack dirty if not in SSE2+ mode
589 if (UseSSE >= 2) {
590 __ verify_FPU(0, "c2i transition should have clean FPU stack");
591 } else {
592 __ empty_FPU_stack();
593 }
594 #endif /* COMPILER2 */
595
596 // Since all args are passed on the stack, total_args_passed * interpreter_
597 // stack_element_size is the
598 // space we need.
599 int extraspace = total_args_passed * Interpreter::stackElementSize;
600
601 // Get return address
602 __ pop(rax);
603
604 // set senderSP value
605 __ movptr(rsi, rsp);
606
607 __ subptr(rsp, extraspace);
608
609 // Now write the args into the outgoing interpreter space
610 for (int i = 0; i < total_args_passed; i++) {
611 if (sig_bt[i] == T_VOID) {
612 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
613 continue;
614 }
615
616 // st_off points to lowest address on stack.
617 int st_off = ((total_args_passed - 1) - i) * Interpreter::stackElementSize;
618 int next_off = st_off - Interpreter::stackElementSize;
619
620 // Say 4 args:
621 // i st_off
622 // 0 12 T_LONG
623 // 1 8 T_VOID
624 // 2 4 T_OBJECT
625 // 3 0 T_BOOL
626 VMReg r_1 = regs[i].first();
627 VMReg r_2 = regs[i].second();
628 if (!r_1->is_valid()) {
629 assert(!r_2->is_valid(), "");
630 continue;
631 }
632
633 if (r_1->is_stack()) {
634 // memory to memory use fpu stack top
635 int ld_off = r_1->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
636
637 if (!r_2->is_valid()) {
638 __ movl(rdi, Address(rsp, ld_off));
639 __ movptr(Address(rsp, st_off), rdi);
640 } else {
641
642 // ld_off == LSW, ld_off+VMRegImpl::stack_slot_size == MSW
643 // st_off == MSW, st_off-wordSize == LSW
644
645 __ movptr(rdi, Address(rsp, ld_off));
646 __ movptr(Address(rsp, next_off), rdi);
647 #ifndef _LP64
648 __ movptr(rdi, Address(rsp, ld_off + wordSize));
649 __ movptr(Address(rsp, st_off), rdi);
650 #else
651 #ifdef ASSERT
652 // Overwrite the unused slot with known junk
653 __ mov64(rax, CONST64(0xdeadffffdeadaaaa));
654 __ movptr(Address(rsp, st_off), rax);
655 #endif /* ASSERT */
656 #endif // _LP64
657 }
658 } else if (r_1->is_Register()) {
659 Register r = r_1->as_Register();
660 if (!r_2->is_valid()) {
661 __ movl(Address(rsp, st_off), r);
662 } else {
663 // long/double in gpr
664 NOT_LP64(ShouldNotReachHere());
665 // Two VMRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
666 // T_DOUBLE and T_LONG use two slots in the interpreter
667 if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
668 // long/double in gpr
669 #ifdef ASSERT
670 // Overwrite the unused slot with known junk
671 LP64_ONLY(__ mov64(rax, CONST64(0xdeadffffdeadaaab)));
672 __ movptr(Address(rsp, st_off), rax);
673 #endif /* ASSERT */
674 __ movptr(Address(rsp, next_off), r);
675 } else {
676 __ movptr(Address(rsp, st_off), r);
677 }
678 }
679 } else {
680 assert(r_1->is_XMMRegister(), "");
681 if (!r_2->is_valid()) {
682 __ movflt(Address(rsp, st_off), r_1->as_XMMRegister());
683 } else {
684 assert(sig_bt[i] == T_DOUBLE || sig_bt[i] == T_LONG, "wrong type");
685 move_c2i_double(masm, r_1->as_XMMRegister(), st_off);
686 }
687 }
688 }
689
690 // Schedule the branch target address early.
691 __ movptr(rcx, Address(rbx, in_bytes(Method::interpreter_entry_offset())));
692 // And repush original return address
693 __ push(rax);
694 __ jmp(rcx);
695 }
696
697
698 static void move_i2c_double(MacroAssembler *masm, XMMRegister r, Register saved_sp, int ld_off) {
699 int next_val_off = ld_off - Interpreter::stackElementSize;
700 __ movdbl(r, Address(saved_sp, next_val_off));
701 }
702
703 static void range_check(MacroAssembler* masm, Register pc_reg, Register temp_reg,
704 address code_start, address code_end,
705 Label& L_ok) {
706 Label L_fail;
707 __ lea(temp_reg, ExternalAddress(code_start));
708 __ cmpptr(pc_reg, temp_reg);
709 __ jcc(Assembler::belowEqual, L_fail);
710 __ lea(temp_reg, ExternalAddress(code_end));
711 __ cmpptr(pc_reg, temp_reg);
712 __ jcc(Assembler::below, L_ok);
713 __ bind(L_fail);
714 }
715
716 void SharedRuntime::gen_i2c_adapter(MacroAssembler *masm,
717 int total_args_passed,
718 int comp_args_on_stack,
719 const BasicType *sig_bt,
720 const VMRegPair *regs) {
721 // Note: rsi contains the senderSP on entry. We must preserve it since
722 // we may do a i2c -> c2i transition if we lose a race where compiled
723 // code goes non-entrant while we get args ready.
724
725 // Adapters can be frameless because they do not require the caller
726 // to perform additional cleanup work, such as correcting the stack pointer.
727 // An i2c adapter is frameless because the *caller* frame, which is interpreted,
728 // routinely repairs its own stack pointer (from interpreter_frame_last_sp),
729 // even if a callee has modified the stack pointer.
730 // A c2i adapter is frameless because the *callee* frame, which is interpreted,
731 // routinely repairs its caller's stack pointer (from sender_sp, which is set
732 // up via the senderSP register).
733 // In other words, if *either* the caller or callee is interpreted, we can
734 // get the stack pointer repaired after a call.
735 // This is why c2i and i2c adapters cannot be indefinitely composed.
736 // In particular, if a c2i adapter were to somehow call an i2c adapter,
737 // both caller and callee would be compiled methods, and neither would
738 // clean up the stack pointer changes performed by the two adapters.
739 // If this happens, control eventually transfers back to the compiled
740 // caller, but with an uncorrected stack, causing delayed havoc.
741
742 // Pick up the return address
743 __ movptr(rax, Address(rsp, 0));
744
745 if (VerifyAdapterCalls &&
746 (Interpreter::code() != NULL || StubRoutines::code1() != NULL)) {
747 // So, let's test for cascading c2i/i2c adapters right now.
748 // assert(Interpreter::contains($return_addr) ||
749 // StubRoutines::contains($return_addr),
750 // "i2c adapter must return to an interpreter frame");
751 __ block_comment("verify_i2c { ");
752 Label L_ok;
753 if (Interpreter::code() != NULL)
754 range_check(masm, rax, rdi,
755 Interpreter::code()->code_start(), Interpreter::code()->code_end(),
756 L_ok);
757 if (StubRoutines::code1() != NULL)
758 range_check(masm, rax, rdi,
759 StubRoutines::code1()->code_begin(), StubRoutines::code1()->code_end(),
760 L_ok);
761 if (StubRoutines::code2() != NULL)
762 range_check(masm, rax, rdi,
763 StubRoutines::code2()->code_begin(), StubRoutines::code2()->code_end(),
764 L_ok);
765 const char* msg = "i2c adapter must return to an interpreter frame";
766 __ block_comment(msg);
767 __ stop(msg);
768 __ bind(L_ok);
769 __ block_comment("} verify_i2ce ");
770 }
771
772 // Must preserve original SP for loading incoming arguments because
773 // we need to align the outgoing SP for compiled code.
774 __ movptr(rdi, rsp);
775
776 // Cut-out for having no stack args. Since up to 2 int/oop args are passed
777 // in registers, we will occasionally have no stack args.
778 int comp_words_on_stack = 0;
779 if (comp_args_on_stack) {
780 // Sig words on the stack are greater-than VMRegImpl::stack0. Those in
781 // registers are below. By subtracting stack0, we either get a negative
782 // number (all values in registers) or the maximum stack slot accessed.
783 // int comp_args_on_stack = VMRegImpl::reg2stack(max_arg);
784 // Convert 4-byte stack slots to words.
785 comp_words_on_stack = align_up(comp_args_on_stack*4, wordSize)>>LogBytesPerWord;
786 // Round up to miminum stack alignment, in wordSize
787 comp_words_on_stack = align_up(comp_words_on_stack, 2);
788 __ subptr(rsp, comp_words_on_stack * wordSize);
789 }
790
791 // Align the outgoing SP
792 __ andptr(rsp, -(StackAlignmentInBytes));
793
794 // push the return address on the stack (note that pushing, rather
795 // than storing it, yields the correct frame alignment for the callee)
796 __ push(rax);
797
798 // Put saved SP in another register
799 const Register saved_sp = rax;
800 __ movptr(saved_sp, rdi);
801
802
803 // Will jump to the compiled code just as if compiled code was doing it.
804 // Pre-load the register-jump target early, to schedule it better.
805 __ movptr(rdi, Address(rbx, in_bytes(Method::from_compiled_offset())));
806
807 // Now generate the shuffle code. Pick up all register args and move the
808 // rest through the floating point stack top.
809 for (int i = 0; i < total_args_passed; i++) {
810 if (sig_bt[i] == T_VOID) {
811 // Longs and doubles are passed in native word order, but misaligned
812 // in the 32-bit build.
813 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
814 continue;
815 }
816
817 // Pick up 0, 1 or 2 words from SP+offset.
818
819 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
820 "scrambled load targets?");
821 // Load in argument order going down.
822 int ld_off = (total_args_passed - i) * Interpreter::stackElementSize;
823 // Point to interpreter value (vs. tag)
824 int next_off = ld_off - Interpreter::stackElementSize;
825 //
826 //
827 //
828 VMReg r_1 = regs[i].first();
829 VMReg r_2 = regs[i].second();
830 if (!r_1->is_valid()) {
831 assert(!r_2->is_valid(), "");
832 continue;
833 }
834 if (r_1->is_stack()) {
835 // Convert stack slot to an SP offset (+ wordSize to account for return address )
836 int st_off = regs[i].first()->reg2stack()*VMRegImpl::stack_slot_size + wordSize;
837
838 // We can use rsi as a temp here because compiled code doesn't need rsi as an input
839 // and if we end up going thru a c2i because of a miss a reasonable value of rsi
840 // we be generated.
841 if (!r_2->is_valid()) {
842 // __ fld_s(Address(saved_sp, ld_off));
843 // __ fstp_s(Address(rsp, st_off));
844 __ movl(rsi, Address(saved_sp, ld_off));
845 __ movptr(Address(rsp, st_off), rsi);
846 } else {
847 // Interpreter local[n] == MSW, local[n+1] == LSW however locals
848 // are accessed as negative so LSW is at LOW address
849
850 // ld_off is MSW so get LSW
851 // st_off is LSW (i.e. reg.first())
852 // __ fld_d(Address(saved_sp, next_off));
853 // __ fstp_d(Address(rsp, st_off));
854 //
855 // We are using two VMRegs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE
856 // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case
857 // So we must adjust where to pick up the data to match the interpreter.
858 //
859 // Interpreter local[n] == MSW, local[n+1] == LSW however locals
860 // are accessed as negative so LSW is at LOW address
861
862 // ld_off is MSW so get LSW
863 const int offset = (NOT_LP64(true ||) sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)?
864 next_off : ld_off;
865 __ movptr(rsi, Address(saved_sp, offset));
866 __ movptr(Address(rsp, st_off), rsi);
867 #ifndef _LP64
868 __ movptr(rsi, Address(saved_sp, ld_off));
869 __ movptr(Address(rsp, st_off + wordSize), rsi);
870 #endif // _LP64
871 }
872 } else if (r_1->is_Register()) { // Register argument
873 Register r = r_1->as_Register();
874 assert(r != rax, "must be different");
875 if (r_2->is_valid()) {
876 //
877 // We are using two VMRegs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE
878 // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case
879 // So we must adjust where to pick up the data to match the interpreter.
880
881 const int offset = (NOT_LP64(true ||) sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)?
882 next_off : ld_off;
883
884 // this can be a misaligned move
885 __ movptr(r, Address(saved_sp, offset));
886 #ifndef _LP64
887 assert(r_2->as_Register() != rax, "need another temporary register");
888 // Remember r_1 is low address (and LSB on x86)
889 // So r_2 gets loaded from high address regardless of the platform
890 __ movptr(r_2->as_Register(), Address(saved_sp, ld_off));
891 #endif // _LP64
892 } else {
893 __ movl(r, Address(saved_sp, ld_off));
894 }
895 } else {
896 assert(r_1->is_XMMRegister(), "");
897 if (!r_2->is_valid()) {
898 __ movflt(r_1->as_XMMRegister(), Address(saved_sp, ld_off));
899 } else {
900 move_i2c_double(masm, r_1->as_XMMRegister(), saved_sp, ld_off);
901 }
902 }
903 }
904
905 // 6243940 We might end up in handle_wrong_method if
906 // the callee is deoptimized as we race thru here. If that
907 // happens we don't want to take a safepoint because the
908 // caller frame will look interpreted and arguments are now
909 // "compiled" so it is much better to make this transition
910 // invisible to the stack walking code. Unfortunately if
911 // we try and find the callee by normal means a safepoint
912 // is possible. So we stash the desired callee in the thread
913 // and the vm will find there should this case occur.
914
915 __ get_thread(rax);
916 __ movptr(Address(rax, JavaThread::callee_target_offset()), rbx);
917
918 // move Method* to rax, in case we end up in an c2i adapter.
919 // the c2i adapters expect Method* in rax, (c2) because c2's
920 // resolve stubs return the result (the method) in rax,.
921 // I'd love to fix this.
922 __ mov(rax, rbx);
923
924 __ jmp(rdi);
925 }
926
927 // ---------------------------------------------------------------
928 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
929 int total_args_passed,
930 int comp_args_on_stack,
931 const BasicType *sig_bt,
932 const VMRegPair *regs,
933 AdapterFingerPrint* fingerprint) {
934 address i2c_entry = __ pc();
935
936 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
937
938 // -------------------------------------------------------------------------
939 // Generate a C2I adapter. On entry we know rbx, holds the Method* during calls
940 // to the interpreter. The args start out packed in the compiled layout. They
941 // need to be unpacked into the interpreter layout. This will almost always
942 // require some stack space. We grow the current (compiled) stack, then repack
943 // the args. We finally end in a jump to the generic interpreter entry point.
944 // On exit from the interpreter, the interpreter will restore our SP (lest the
945 // compiled code, which relys solely on SP and not EBP, get sick).
946
947 address c2i_unverified_entry = __ pc();
948 Label skip_fixup;
949
950 Register holder = rax;
951 Register receiver = rcx;
952 Register temp = rbx;
953
954 {
955
956 Label missed;
957 __ movptr(temp, Address(receiver, oopDesc::klass_offset_in_bytes()));
958 __ cmpptr(temp, Address(holder, CompiledICHolder::holder_klass_offset()));
959 __ movptr(rbx, Address(holder, CompiledICHolder::holder_method_offset()));
960 __ jcc(Assembler::notEqual, missed);
961 // Method might have been compiled since the call site was patched to
962 // interpreted if that is the case treat it as a miss so we can get
963 // the call site corrected.
964 __ cmpptr(Address(rbx, in_bytes(Method::code_offset())), (int32_t)NULL_WORD);
965 __ jcc(Assembler::equal, skip_fixup);
966
967 __ bind(missed);
968 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
969 }
970
971 address c2i_entry = __ pc();
972
973 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup);
974
975 __ flush();
976 return AdapterHandlerLibrary::new_entry(fingerprint, i2c_entry, c2i_entry, c2i_unverified_entry);
977 }
978
979 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
980 VMRegPair *regs,
981 VMRegPair *regs2,
982 int total_args_passed) {
983 assert(regs2 == NULL, "not needed on x86");
984 // We return the amount of VMRegImpl stack slots we need to reserve for all
985 // the arguments NOT counting out_preserve_stack_slots.
986
987 uint stack = 0; // All arguments on stack
988
989 for( int i = 0; i < total_args_passed; i++) {
990 // From the type and the argument number (count) compute the location
991 switch( sig_bt[i] ) {
992 case T_BOOLEAN:
993 case T_CHAR:
994 case T_FLOAT:
995 case T_BYTE:
996 case T_SHORT:
997 case T_INT:
998 case T_OBJECT:
999 case T_ARRAY:
1000 case T_ADDRESS:
1001 case T_METADATA:
1002 regs[i].set1(VMRegImpl::stack2reg(stack++));
1003 break;
1004 case T_LONG:
1005 case T_DOUBLE: // The stack numbering is reversed from Java
1006 // Since C arguments do not get reversed, the ordering for
1007 // doubles on the stack must be opposite the Java convention
1008 assert((i + 1) < total_args_passed && sig_bt[i+1] == T_VOID, "missing Half" );
1009 regs[i].set2(VMRegImpl::stack2reg(stack));
1010 stack += 2;
1011 break;
1012 case T_VOID: regs[i].set_bad(); break;
1013 default:
1014 ShouldNotReachHere();
1015 break;
1016 }
1017 }
1018 return stack;
1019 }
1020
1021 // A simple move of integer like type
1022 static void simple_move32(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1023 if (src.first()->is_stack()) {
1024 if (dst.first()->is_stack()) {
1025 // stack to stack
1026 // __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1027 // __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1028 __ movl2ptr(rax, Address(rbp, reg2offset_in(src.first())));
1029 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax);
1030 } else {
1031 // stack to reg
1032 __ movl2ptr(dst.first()->as_Register(), Address(rbp, reg2offset_in(src.first())));
1033 }
1034 } else if (dst.first()->is_stack()) {
1035 // reg to stack
1036 // no need to sign extend on 64bit
1037 __ movptr(Address(rsp, reg2offset_out(dst.first())), src.first()->as_Register());
1038 } else {
1039 if (dst.first() != src.first()) {
1040 __ mov(dst.first()->as_Register(), src.first()->as_Register());
1041 }
1042 }
1043 }
1044
1045 // An oop arg. Must pass a handle not the oop itself
1046 static void object_move(MacroAssembler* masm,
1047 OopMap* map,
1048 int oop_handle_offset,
1049 int framesize_in_slots,
1050 VMRegPair src,
1051 VMRegPair dst,
1052 bool is_receiver,
1053 int* receiver_offset) {
1054
1055 // Because of the calling conventions we know that src can be a
1056 // register or a stack location. dst can only be a stack location.
1057
1058 assert(dst.first()->is_stack(), "must be stack");
1059 // must pass a handle. First figure out the location we use as a handle
1060
1061 if (src.first()->is_stack()) {
1062 // Oop is already on the stack as an argument
1063 Register rHandle = rax;
1064 Label nil;
1065 __ xorptr(rHandle, rHandle);
1066 __ cmpptr(Address(rbp, reg2offset_in(src.first())), (int32_t)NULL_WORD);
1067 __ jcc(Assembler::equal, nil);
1068 __ lea(rHandle, Address(rbp, reg2offset_in(src.first())));
1069 __ bind(nil);
1070 __ movptr(Address(rsp, reg2offset_out(dst.first())), rHandle);
1071
1072 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots();
1073 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots));
1074 if (is_receiver) {
1075 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size;
1076 }
1077 } else {
1078 // Oop is in an a register we must store it to the space we reserve
1079 // on the stack for oop_handles
1080 const Register rOop = src.first()->as_Register();
1081 const Register rHandle = rax;
1082 int oop_slot = (rOop == rcx ? 0 : 1) * VMRegImpl::slots_per_word + oop_handle_offset;
1083 int offset = oop_slot*VMRegImpl::stack_slot_size;
1084 Label skip;
1085 __ movptr(Address(rsp, offset), rOop);
1086 map->set_oop(VMRegImpl::stack2reg(oop_slot));
1087 __ xorptr(rHandle, rHandle);
1088 __ cmpptr(rOop, (int32_t)NULL_WORD);
1089 __ jcc(Assembler::equal, skip);
1090 __ lea(rHandle, Address(rsp, offset));
1091 __ bind(skip);
1092 // Store the handle parameter
1093 __ movptr(Address(rsp, reg2offset_out(dst.first())), rHandle);
1094 if (is_receiver) {
1095 *receiver_offset = offset;
1096 }
1097 }
1098 }
1099
1100 // A float arg may have to do float reg int reg conversion
1101 static void float_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1102 assert(!src.second()->is_valid() && !dst.second()->is_valid(), "bad float_move");
1103
1104 // Because of the calling convention we know that src is either a stack location
1105 // or an xmm register. dst can only be a stack location.
1106
1107 assert(dst.first()->is_stack() && ( src.first()->is_stack() || src.first()->is_XMMRegister()), "bad parameters");
1108
1109 if (src.first()->is_stack()) {
1110 __ movl(rax, Address(rbp, reg2offset_in(src.first())));
1111 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax);
1112 } else {
1113 // reg to stack
1114 __ movflt(Address(rsp, reg2offset_out(dst.first())), src.first()->as_XMMRegister());
1115 }
1116 }
1117
1118 // A long move
1119 static void long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1120
1121 // The only legal possibility for a long_move VMRegPair is:
1122 // 1: two stack slots (possibly unaligned)
1123 // as neither the java or C calling convention will use registers
1124 // for longs.
1125
1126 if (src.first()->is_stack() && dst.first()->is_stack()) {
1127 assert(src.second()->is_stack() && dst.second()->is_stack(), "must be all stack");
1128 __ movptr(rax, Address(rbp, reg2offset_in(src.first())));
1129 NOT_LP64(__ movptr(rbx, Address(rbp, reg2offset_in(src.second()))));
1130 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax);
1131 NOT_LP64(__ movptr(Address(rsp, reg2offset_out(dst.second())), rbx));
1132 } else {
1133 ShouldNotReachHere();
1134 }
1135 }
1136
1137 // A double move
1138 static void double_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1139
1140 // The only legal possibilities for a double_move VMRegPair are:
1141 // The painful thing here is that like long_move a VMRegPair might be
1142
1143 // Because of the calling convention we know that src is either
1144 // 1: a single physical register (xmm registers only)
1145 // 2: two stack slots (possibly unaligned)
1146 // dst can only be a pair of stack slots.
1147
1148 assert(dst.first()->is_stack() && (src.first()->is_XMMRegister() || src.first()->is_stack()), "bad args");
1149
1150 if (src.first()->is_stack()) {
1151 // source is all stack
1152 __ movptr(rax, Address(rbp, reg2offset_in(src.first())));
1153 NOT_LP64(__ movptr(rbx, Address(rbp, reg2offset_in(src.second()))));
1154 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax);
1155 NOT_LP64(__ movptr(Address(rsp, reg2offset_out(dst.second())), rbx));
1156 } else {
1157 // reg to stack
1158 // No worries about stack alignment
1159 __ movdbl(Address(rsp, reg2offset_out(dst.first())), src.first()->as_XMMRegister());
1160 }
1161 }
1162
1163
1164 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1165 // We always ignore the frame_slots arg and just use the space just below frame pointer
1166 // which by this time is free to use
1167 switch (ret_type) {
1168 case T_FLOAT:
1169 __ fstp_s(Address(rbp, -wordSize));
1170 break;
1171 case T_DOUBLE:
1172 __ fstp_d(Address(rbp, -2*wordSize));
1173 break;
1174 case T_VOID: break;
1175 case T_LONG:
1176 __ movptr(Address(rbp, -wordSize), rax);
1177 NOT_LP64(__ movptr(Address(rbp, -2*wordSize), rdx));
1178 break;
1179 default: {
1180 __ movptr(Address(rbp, -wordSize), rax);
1181 }
1182 }
1183 }
1184
1185 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1186 // We always ignore the frame_slots arg and just use the space just below frame pointer
1187 // which by this time is free to use
1188 switch (ret_type) {
1189 case T_FLOAT:
1190 __ fld_s(Address(rbp, -wordSize));
1191 break;
1192 case T_DOUBLE:
1193 __ fld_d(Address(rbp, -2*wordSize));
1194 break;
1195 case T_LONG:
1196 __ movptr(rax, Address(rbp, -wordSize));
1197 NOT_LP64(__ movptr(rdx, Address(rbp, -2*wordSize)));
1198 break;
1199 case T_VOID: break;
1200 default: {
1201 __ movptr(rax, Address(rbp, -wordSize));
1202 }
1203 }
1204 }
1205
1206
1207 static void save_or_restore_arguments(MacroAssembler* masm,
1208 const int stack_slots,
1209 const int total_in_args,
1210 const int arg_save_area,
1211 OopMap* map,
1212 VMRegPair* in_regs,
1213 BasicType* in_sig_bt) {
1214 // if map is non-NULL then the code should store the values,
1215 // otherwise it should load them.
1216 int handle_index = 0;
1217 // Save down double word first
1218 for ( int i = 0; i < total_in_args; i++) {
1219 if (in_regs[i].first()->is_XMMRegister() && in_sig_bt[i] == T_DOUBLE) {
1220 int slot = handle_index * VMRegImpl::slots_per_word + arg_save_area;
1221 int offset = slot * VMRegImpl::stack_slot_size;
1222 handle_index += 2;
1223 assert(handle_index <= stack_slots, "overflow");
1224 if (map != NULL) {
1225 __ movdbl(Address(rsp, offset), in_regs[i].first()->as_XMMRegister());
1226 } else {
1227 __ movdbl(in_regs[i].first()->as_XMMRegister(), Address(rsp, offset));
1228 }
1229 }
1230 if (in_regs[i].first()->is_Register() && in_sig_bt[i] == T_LONG) {
1231 int slot = handle_index * VMRegImpl::slots_per_word + arg_save_area;
1232 int offset = slot * VMRegImpl::stack_slot_size;
1233 handle_index += 2;
1234 assert(handle_index <= stack_slots, "overflow");
1235 if (map != NULL) {
1236 __ movl(Address(rsp, offset), in_regs[i].first()->as_Register());
1237 if (in_regs[i].second()->is_Register()) {
1238 __ movl(Address(rsp, offset + 4), in_regs[i].second()->as_Register());
1239 }
1240 } else {
1241 __ movl(in_regs[i].first()->as_Register(), Address(rsp, offset));
1242 if (in_regs[i].second()->is_Register()) {
1243 __ movl(in_regs[i].second()->as_Register(), Address(rsp, offset + 4));
1244 }
1245 }
1246 }
1247 }
1248 // Save or restore single word registers
1249 for ( int i = 0; i < total_in_args; i++) {
1250 if (in_regs[i].first()->is_Register()) {
1251 int slot = handle_index++ * VMRegImpl::slots_per_word + arg_save_area;
1252 int offset = slot * VMRegImpl::stack_slot_size;
1253 assert(handle_index <= stack_slots, "overflow");
1254 if (in_sig_bt[i] == T_ARRAY && map != NULL) {
1255 map->set_oop(VMRegImpl::stack2reg(slot));;
1256 }
1257
1258 // Value is in an input register pass we must flush it to the stack
1259 const Register reg = in_regs[i].first()->as_Register();
1260 switch (in_sig_bt[i]) {
1261 case T_ARRAY:
1262 if (map != NULL) {
1263 __ movptr(Address(rsp, offset), reg);
1264 } else {
1265 __ movptr(reg, Address(rsp, offset));
1266 }
1267 break;
1268 case T_BOOLEAN:
1269 case T_CHAR:
1270 case T_BYTE:
1271 case T_SHORT:
1272 case T_INT:
1273 if (map != NULL) {
1274 __ movl(Address(rsp, offset), reg);
1275 } else {
1276 __ movl(reg, Address(rsp, offset));
1277 }
1278 break;
1279 case T_OBJECT:
1280 default: ShouldNotReachHere();
1281 }
1282 } else if (in_regs[i].first()->is_XMMRegister()) {
1283 if (in_sig_bt[i] == T_FLOAT) {
1284 int slot = handle_index++ * VMRegImpl::slots_per_word + arg_save_area;
1285 int offset = slot * VMRegImpl::stack_slot_size;
1286 assert(handle_index <= stack_slots, "overflow");
1287 if (map != NULL) {
1288 __ movflt(Address(rsp, offset), in_regs[i].first()->as_XMMRegister());
1289 } else {
1290 __ movflt(in_regs[i].first()->as_XMMRegister(), Address(rsp, offset));
1291 }
1292 }
1293 } else if (in_regs[i].first()->is_stack()) {
1294 if (in_sig_bt[i] == T_ARRAY && map != NULL) {
1295 int offset_in_older_frame = in_regs[i].first()->reg2stack() + SharedRuntime::out_preserve_stack_slots();
1296 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + stack_slots));
1297 }
1298 }
1299 }
1300 }
1301
1302 // Check GCLocker::needs_gc and enter the runtime if it's true. This
1303 // keeps a new JNI critical region from starting until a GC has been
1304 // forced. Save down any oops in registers and describe them in an
1305 // OopMap.
1306 static void check_needs_gc_for_critical_native(MacroAssembler* masm,
1307 Register thread,
1308 int stack_slots,
1309 int total_c_args,
1310 int total_in_args,
1311 int arg_save_area,
1312 OopMapSet* oop_maps,
1313 VMRegPair* in_regs,
1314 BasicType* in_sig_bt) {
1315 __ block_comment("check GCLocker::needs_gc");
1316 Label cont;
1317 __ cmp8(ExternalAddress((address)GCLocker::needs_gc_address()), false);
1318 __ jcc(Assembler::equal, cont);
1319
1320 // Save down any incoming oops and call into the runtime to halt for a GC
1321
1322 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1323
1324 save_or_restore_arguments(masm, stack_slots, total_in_args,
1325 arg_save_area, map, in_regs, in_sig_bt);
1326
1327 address the_pc = __ pc();
1328 oop_maps->add_gc_map( __ offset(), map);
1329 __ set_last_Java_frame(thread, rsp, noreg, the_pc);
1330
1331 __ block_comment("block_for_jni_critical");
1332 __ push(thread);
1333 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::block_for_jni_critical)));
1334 __ increment(rsp, wordSize);
1335
1336 __ get_thread(thread);
1337 __ reset_last_Java_frame(thread, false);
1338
1339 save_or_restore_arguments(masm, stack_slots, total_in_args,
1340 arg_save_area, NULL, in_regs, in_sig_bt);
1341
1342 __ bind(cont);
1343 #ifdef ASSERT
1344 if (StressCriticalJNINatives) {
1345 // Stress register saving
1346 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1347 save_or_restore_arguments(masm, stack_slots, total_in_args,
1348 arg_save_area, map, in_regs, in_sig_bt);
1349 // Destroy argument registers
1350 for (int i = 0; i < total_in_args - 1; i++) {
1351 if (in_regs[i].first()->is_Register()) {
1352 const Register reg = in_regs[i].first()->as_Register();
1353 __ xorptr(reg, reg);
1354 } else if (in_regs[i].first()->is_XMMRegister()) {
1355 __ xorpd(in_regs[i].first()->as_XMMRegister(), in_regs[i].first()->as_XMMRegister());
1356 } else if (in_regs[i].first()->is_FloatRegister()) {
1357 ShouldNotReachHere();
1358 } else if (in_regs[i].first()->is_stack()) {
1359 // Nothing to do
1360 } else {
1361 ShouldNotReachHere();
1362 }
1363 if (in_sig_bt[i] == T_LONG || in_sig_bt[i] == T_DOUBLE) {
1364 i++;
1365 }
1366 }
1367
1368 save_or_restore_arguments(masm, stack_slots, total_in_args,
1369 arg_save_area, NULL, in_regs, in_sig_bt);
1370 }
1371 #endif
1372 }
1373
1374 // Unpack an array argument into a pointer to the body and the length
1375 // if the array is non-null, otherwise pass 0 for both.
1376 static void unpack_array_argument(MacroAssembler* masm, VMRegPair reg, BasicType in_elem_type, VMRegPair body_arg, VMRegPair length_arg) {
1377 Register tmp_reg = rax;
1378 assert(!body_arg.first()->is_Register() || body_arg.first()->as_Register() != tmp_reg,
1379 "possible collision");
1380 assert(!length_arg.first()->is_Register() || length_arg.first()->as_Register() != tmp_reg,
1381 "possible collision");
1382
1383 // Pass the length, ptr pair
1384 Label is_null, done;
1385 VMRegPair tmp(tmp_reg->as_VMReg());
1386 if (reg.first()->is_stack()) {
1387 // Load the arg up from the stack
1388 simple_move32(masm, reg, tmp);
1389 reg = tmp;
1390 }
1391 __ testptr(reg.first()->as_Register(), reg.first()->as_Register());
1392 __ jccb(Assembler::equal, is_null);
1393 __ lea(tmp_reg, Address(reg.first()->as_Register(), arrayOopDesc::base_offset_in_bytes(in_elem_type)));
1394 simple_move32(masm, tmp, body_arg);
1395 // load the length relative to the body.
1396 __ movl(tmp_reg, Address(tmp_reg, arrayOopDesc::length_offset_in_bytes() -
1397 arrayOopDesc::base_offset_in_bytes(in_elem_type)));
1398 simple_move32(masm, tmp, length_arg);
1399 __ jmpb(done);
1400 __ bind(is_null);
1401 // Pass zeros
1402 __ xorptr(tmp_reg, tmp_reg);
1403 simple_move32(masm, tmp, body_arg);
1404 simple_move32(masm, tmp, length_arg);
1405 __ bind(done);
1406 }
1407
1408 static void verify_oop_args(MacroAssembler* masm,
1409 methodHandle method,
1410 const BasicType* sig_bt,
1411 const VMRegPair* regs) {
1412 Register temp_reg = rbx; // not part of any compiled calling seq
1413 if (VerifyOops) {
1414 for (int i = 0; i < method->size_of_parameters(); i++) {
1415 if (sig_bt[i] == T_OBJECT ||
1416 sig_bt[i] == T_ARRAY) {
1417 VMReg r = regs[i].first();
1418 assert(r->is_valid(), "bad oop arg");
1419 if (r->is_stack()) {
1420 __ movptr(temp_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize));
1421 __ verify_oop(temp_reg);
1422 } else {
1423 __ verify_oop(r->as_Register());
1424 }
1425 }
1426 }
1427 }
1428 }
1429
1430 static void gen_special_dispatch(MacroAssembler* masm,
1431 methodHandle method,
1432 const BasicType* sig_bt,
1433 const VMRegPair* regs) {
1434 verify_oop_args(masm, method, sig_bt, regs);
1435 vmIntrinsics::ID iid = method->intrinsic_id();
1436
1437 // Now write the args into the outgoing interpreter space
1438 bool has_receiver = false;
1439 Register receiver_reg = noreg;
1440 int member_arg_pos = -1;
1441 Register member_reg = noreg;
1442 int ref_kind = MethodHandles::signature_polymorphic_intrinsic_ref_kind(iid);
1443 if (ref_kind != 0) {
1444 member_arg_pos = method->size_of_parameters() - 1; // trailing MemberName argument
1445 member_reg = rbx; // known to be free at this point
1446 has_receiver = MethodHandles::ref_kind_has_receiver(ref_kind);
1447 } else if (iid == vmIntrinsics::_invokeBasic) {
1448 has_receiver = true;
1449 } else {
1450 fatal("unexpected intrinsic id %d", iid);
1451 }
1452
1453 if (member_reg != noreg) {
1454 // Load the member_arg into register, if necessary.
1455 SharedRuntime::check_member_name_argument_is_last_argument(method, sig_bt, regs);
1456 VMReg r = regs[member_arg_pos].first();
1457 if (r->is_stack()) {
1458 __ movptr(member_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize));
1459 } else {
1460 // no data motion is needed
1461 member_reg = r->as_Register();
1462 }
1463 }
1464
1465 if (has_receiver) {
1466 // Make sure the receiver is loaded into a register.
1467 assert(method->size_of_parameters() > 0, "oob");
1468 assert(sig_bt[0] == T_OBJECT, "receiver argument must be an object");
1469 VMReg r = regs[0].first();
1470 assert(r->is_valid(), "bad receiver arg");
1471 if (r->is_stack()) {
1472 // Porting note: This assumes that compiled calling conventions always
1473 // pass the receiver oop in a register. If this is not true on some
1474 // platform, pick a temp and load the receiver from stack.
1475 fatal("receiver always in a register");
1476 receiver_reg = rcx; // known to be free at this point
1477 __ movptr(receiver_reg, Address(rsp, r->reg2stack() * VMRegImpl::stack_slot_size + wordSize));
1478 } else {
1479 // no data motion is needed
1480 receiver_reg = r->as_Register();
1481 }
1482 }
1483
1484 // Figure out which address we are really jumping to:
1485 MethodHandles::generate_method_handle_dispatch(masm, iid,
1486 receiver_reg, member_reg, /*for_compiler_entry:*/ true);
1487 }
1488
1489 // ---------------------------------------------------------------------------
1490 // Generate a native wrapper for a given method. The method takes arguments
1491 // in the Java compiled code convention, marshals them to the native
1492 // convention (handlizes oops, etc), transitions to native, makes the call,
1493 // returns to java state (possibly blocking), unhandlizes any result and
1494 // returns.
1495 //
1496 // Critical native functions are a shorthand for the use of
1497 // GetPrimtiveArrayCritical and disallow the use of any other JNI
1498 // functions. The wrapper is expected to unpack the arguments before
1499 // passing them to the callee and perform checks before and after the
1500 // native call to ensure that they GCLocker
1501 // lock_critical/unlock_critical semantics are followed. Some other
1502 // parts of JNI setup are skipped like the tear down of the JNI handle
1503 // block and the check for pending exceptions it's impossible for them
1504 // to be thrown.
1505 //
1506 // They are roughly structured like this:
1507 // if (GCLocker::needs_gc())
1508 // SharedRuntime::block_for_jni_critical();
1509 // tranistion to thread_in_native
1510 // unpack arrray arguments and call native entry point
1511 // check for safepoint in progress
1512 // check if any thread suspend flags are set
1513 // call into JVM and possible unlock the JNI critical
1514 // if a GC was suppressed while in the critical native.
1515 // transition back to thread_in_Java
1516 // return to caller
1517 //
1518 nmethod* SharedRuntime::generate_native_wrapper(MacroAssembler* masm,
1519 const methodHandle& method,
1520 int compile_id,
1521 BasicType* in_sig_bt,
1522 VMRegPair* in_regs,
1523 BasicType ret_type) {
1524 if (method->is_method_handle_intrinsic()) {
1525 vmIntrinsics::ID iid = method->intrinsic_id();
1526 intptr_t start = (intptr_t)__ pc();
1527 int vep_offset = ((intptr_t)__ pc()) - start;
1528 gen_special_dispatch(masm,
1529 method,
1530 in_sig_bt,
1531 in_regs);
1532 int frame_complete = ((intptr_t)__ pc()) - start; // not complete, period
1533 __ flush();
1534 int stack_slots = SharedRuntime::out_preserve_stack_slots(); // no out slots at all, actually
1535 return nmethod::new_native_nmethod(method,
1536 compile_id,
1537 masm->code(),
1538 vep_offset,
1539 frame_complete,
1540 stack_slots / VMRegImpl::slots_per_word,
1541 in_ByteSize(-1),
1542 in_ByteSize(-1),
1543 (OopMapSet*)NULL);
1544 }
1545 bool is_critical_native = true;
1546 address native_func = method->critical_native_function();
1547 if (native_func == NULL) {
1548 native_func = method->native_function();
1549 is_critical_native = false;
1550 }
1551 assert(native_func != NULL, "must have function");
1552
1553 // An OopMap for lock (and class if static)
1554 OopMapSet *oop_maps = new OopMapSet();
1555
1556 // We have received a description of where all the java arg are located
1557 // on entry to the wrapper. We need to convert these args to where
1558 // the jni function will expect them. To figure out where they go
1559 // we convert the java signature to a C signature by inserting
1560 // the hidden arguments as arg[0] and possibly arg[1] (static method)
1561
1562 const int total_in_args = method->size_of_parameters();
1563 int total_c_args = total_in_args;
1564 if (!is_critical_native) {
1565 total_c_args += 1;
1566 if (method->is_static()) {
1567 total_c_args++;
1568 }
1569 } else {
1570 for (int i = 0; i < total_in_args; i++) {
1571 if (in_sig_bt[i] == T_ARRAY) {
1572 total_c_args++;
1573 }
1574 }
1575 }
1576
1577 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1578 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1579 BasicType* in_elem_bt = NULL;
1580
1581 int argc = 0;
1582 if (!is_critical_native) {
1583 out_sig_bt[argc++] = T_ADDRESS;
1584 if (method->is_static()) {
1585 out_sig_bt[argc++] = T_OBJECT;
1586 }
1587
1588 for (int i = 0; i < total_in_args ; i++ ) {
1589 out_sig_bt[argc++] = in_sig_bt[i];
1590 }
1591 } else {
1592 Thread* THREAD = Thread::current();
1593 in_elem_bt = NEW_RESOURCE_ARRAY(BasicType, total_in_args);
1594 SignatureStream ss(method->signature());
1595 for (int i = 0; i < total_in_args ; i++ ) {
1596 if (in_sig_bt[i] == T_ARRAY) {
1597 // Arrays are passed as int, elem* pair
1598 out_sig_bt[argc++] = T_INT;
1599 out_sig_bt[argc++] = T_ADDRESS;
1600 Symbol* atype = ss.as_symbol(CHECK_NULL);
1601 const char* at = atype->as_C_string();
1602 if (strlen(at) == 2) {
1603 assert(at[0] == '[', "must be");
1604 switch (at[1]) {
1605 case 'B': in_elem_bt[i] = T_BYTE; break;
1606 case 'C': in_elem_bt[i] = T_CHAR; break;
1607 case 'D': in_elem_bt[i] = T_DOUBLE; break;
1608 case 'F': in_elem_bt[i] = T_FLOAT; break;
1609 case 'I': in_elem_bt[i] = T_INT; break;
1610 case 'J': in_elem_bt[i] = T_LONG; break;
1611 case 'S': in_elem_bt[i] = T_SHORT; break;
1612 case 'Z': in_elem_bt[i] = T_BOOLEAN; break;
1613 default: ShouldNotReachHere();
1614 }
1615 }
1616 } else {
1617 out_sig_bt[argc++] = in_sig_bt[i];
1618 in_elem_bt[i] = T_VOID;
1619 }
1620 if (in_sig_bt[i] != T_VOID) {
1621 assert(in_sig_bt[i] == ss.type(), "must match");
1622 ss.next();
1623 }
1624 }
1625 }
1626
1627 // Now figure out where the args must be stored and how much stack space
1628 // they require.
1629 int out_arg_slots;
1630 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, NULL, total_c_args);
1631
1632 // Compute framesize for the wrapper. We need to handlize all oops in
1633 // registers a max of 2 on x86.
1634
1635 // Calculate the total number of stack slots we will need.
1636
1637 // First count the abi requirement plus all of the outgoing args
1638 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1639
1640 // Now the space for the inbound oop handle area
1641 int total_save_slots = 2 * VMRegImpl::slots_per_word; // 2 arguments passed in registers
1642 if (is_critical_native) {
1643 // Critical natives may have to call out so they need a save area
1644 // for register arguments.
1645 int double_slots = 0;
1646 int single_slots = 0;
1647 for ( int i = 0; i < total_in_args; i++) {
1648 if (in_regs[i].first()->is_Register()) {
1649 const Register reg = in_regs[i].first()->as_Register();
1650 switch (in_sig_bt[i]) {
1651 case T_ARRAY: // critical array (uses 2 slots on LP64)
1652 case T_BOOLEAN:
1653 case T_BYTE:
1654 case T_SHORT:
1655 case T_CHAR:
1656 case T_INT: single_slots++; break;
1657 case T_LONG: double_slots++; break;
1658 default: ShouldNotReachHere();
1659 }
1660 } else if (in_regs[i].first()->is_XMMRegister()) {
1661 switch (in_sig_bt[i]) {
1662 case T_FLOAT: single_slots++; break;
1663 case T_DOUBLE: double_slots++; break;
1664 default: ShouldNotReachHere();
1665 }
1666 } else if (in_regs[i].first()->is_FloatRegister()) {
1667 ShouldNotReachHere();
1668 }
1669 }
1670 total_save_slots = double_slots * 2 + single_slots;
1671 // align the save area
1672 if (double_slots != 0) {
1673 stack_slots = align_up(stack_slots, 2);
1674 }
1675 }
1676
1677 int oop_handle_offset = stack_slots;
1678 stack_slots += total_save_slots;
1679
1680 // Now any space we need for handlizing a klass if static method
1681
1682 int klass_slot_offset = 0;
1683 int klass_offset = -1;
1684 int lock_slot_offset = 0;
1685 bool is_static = false;
1686
1687 if (method->is_static()) {
1688 klass_slot_offset = stack_slots;
1689 stack_slots += VMRegImpl::slots_per_word;
1690 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1691 is_static = true;
1692 }
1693
1694 // Plus a lock if needed
1695
1696 if (method->is_synchronized()) {
1697 lock_slot_offset = stack_slots;
1698 stack_slots += VMRegImpl::slots_per_word;
1699 }
1700
1701 // Now a place (+2) to save return values or temp during shuffling
1702 // + 2 for return address (which we own) and saved rbp,
1703 stack_slots += 4;
1704
1705 // Ok The space we have allocated will look like:
1706 //
1707 //
1708 // FP-> | |
1709 // |---------------------|
1710 // | 2 slots for moves |
1711 // |---------------------|
1712 // | lock box (if sync) |
1713 // |---------------------| <- lock_slot_offset (-lock_slot_rbp_offset)
1714 // | klass (if static) |
1715 // |---------------------| <- klass_slot_offset
1716 // | oopHandle area |
1717 // |---------------------| <- oop_handle_offset (a max of 2 registers)
1718 // | outbound memory |
1719 // | based arguments |
1720 // | |
1721 // |---------------------|
1722 // | |
1723 // SP-> | out_preserved_slots |
1724 //
1725 //
1726 // ****************************************************************************
1727 // WARNING - on Windows Java Natives use pascal calling convention and pop the
1728 // arguments off of the stack after the jni call. Before the call we can use
1729 // instructions that are SP relative. After the jni call we switch to FP
1730 // relative instructions instead of re-adjusting the stack on windows.
1731 // ****************************************************************************
1732
1733
1734 // Now compute actual number of stack words we need rounding to make
1735 // stack properly aligned.
1736 stack_slots = align_up(stack_slots, StackAlignmentInSlots);
1737
1738 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1739
1740 intptr_t start = (intptr_t)__ pc();
1741
1742 // First thing make an ic check to see if we should even be here
1743
1744 // We are free to use all registers as temps without saving them and
1745 // restoring them except rbp. rbp is the only callee save register
1746 // as far as the interpreter and the compiler(s) are concerned.
1747
1748
1749 const Register ic_reg = rax;
1750 const Register receiver = rcx;
1751 Label hit;
1752 Label exception_pending;
1753
1754 __ verify_oop(receiver);
1755 __ cmpptr(ic_reg, Address(receiver, oopDesc::klass_offset_in_bytes()));
1756 __ jcc(Assembler::equal, hit);
1757
1758 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
1759
1760 // verified entry must be aligned for code patching.
1761 // and the first 5 bytes must be in the same cache line
1762 // if we align at 8 then we will be sure 5 bytes are in the same line
1763 __ align(8);
1764
1765 __ bind(hit);
1766
1767 int vep_offset = ((intptr_t)__ pc()) - start;
1768
1769 #ifdef COMPILER1
1770 // For Object.hashCode, System.identityHashCode try to pull hashCode from object header if available.
1771 if ((InlineObjectHash && method->intrinsic_id() == vmIntrinsics::_hashCode) || (method->intrinsic_id() == vmIntrinsics::_identityHashCode)) {
1772 inline_check_hashcode_from_object_header(masm, method, rcx /*obj_reg*/, rax /*result*/);
1773 }
1774 #endif // COMPILER1
1775
1776 // The instruction at the verified entry point must be 5 bytes or longer
1777 // because it can be patched on the fly by make_non_entrant. The stack bang
1778 // instruction fits that requirement.
1779
1780 // Generate stack overflow check
1781
1782 if (UseStackBanging) {
1783 __ bang_stack_with_offset((int)JavaThread::stack_shadow_zone_size());
1784 } else {
1785 // need a 5 byte instruction to allow MT safe patching to non-entrant
1786 __ fat_nop();
1787 }
1788
1789 // Generate a new frame for the wrapper.
1790 __ enter();
1791 // -2 because return address is already present and so is saved rbp
1792 __ subptr(rsp, stack_size - 2*wordSize);
1793
1794 // Frame is now completed as far as size and linkage.
1795 int frame_complete = ((intptr_t)__ pc()) - start;
1796
1797 if (UseRTMLocking) {
1798 // Abort RTM transaction before calling JNI
1799 // because critical section will be large and will be
1800 // aborted anyway. Also nmethod could be deoptimized.
1801 __ xabort(0);
1802 }
1803
1804 // Calculate the difference between rsp and rbp,. We need to know it
1805 // after the native call because on windows Java Natives will pop
1806 // the arguments and it is painful to do rsp relative addressing
1807 // in a platform independent way. So after the call we switch to
1808 // rbp, relative addressing.
1809
1810 int fp_adjustment = stack_size - 2*wordSize;
1811
1812 #ifdef COMPILER2
1813 // C2 may leave the stack dirty if not in SSE2+ mode
1814 if (UseSSE >= 2) {
1815 __ verify_FPU(0, "c2i transition should have clean FPU stack");
1816 } else {
1817 __ empty_FPU_stack();
1818 }
1819 #endif /* COMPILER2 */
1820
1821 // Compute the rbp, offset for any slots used after the jni call
1822
1823 int lock_slot_rbp_offset = (lock_slot_offset*VMRegImpl::stack_slot_size) - fp_adjustment;
1824
1825 // We use rdi as a thread pointer because it is callee save and
1826 // if we load it once it is usable thru the entire wrapper
1827 const Register thread = rdi;
1828
1829 // We use rsi as the oop handle for the receiver/klass
1830 // It is callee save so it survives the call to native
1831
1832 const Register oop_handle_reg = rsi;
1833
1834 __ get_thread(thread);
1835
1836 if (is_critical_native) {
1837 check_needs_gc_for_critical_native(masm, thread, stack_slots, total_c_args, total_in_args,
1838 oop_handle_offset, oop_maps, in_regs, in_sig_bt);
1839 }
1840
1841 //
1842 // We immediately shuffle the arguments so that any vm call we have to
1843 // make from here on out (sync slow path, jvmti, etc.) we will have
1844 // captured the oops from our caller and have a valid oopMap for
1845 // them.
1846
1847 // -----------------
1848 // The Grand Shuffle
1849 //
1850 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv*
1851 // and, if static, the class mirror instead of a receiver. This pretty much
1852 // guarantees that register layout will not match (and x86 doesn't use reg
1853 // parms though amd does). Since the native abi doesn't use register args
1854 // and the java conventions does we don't have to worry about collisions.
1855 // All of our moved are reg->stack or stack->stack.
1856 // We ignore the extra arguments during the shuffle and handle them at the
1857 // last moment. The shuffle is described by the two calling convention
1858 // vectors we have in our possession. We simply walk the java vector to
1859 // get the source locations and the c vector to get the destinations.
1860
1861 int c_arg = is_critical_native ? 0 : (method->is_static() ? 2 : 1 );
1862
1863 // Record rsp-based slot for receiver on stack for non-static methods
1864 int receiver_offset = -1;
1865
1866 // This is a trick. We double the stack slots so we can claim
1867 // the oops in the caller's frame. Since we are sure to have
1868 // more args than the caller doubling is enough to make
1869 // sure we can capture all the incoming oop args from the
1870 // caller.
1871 //
1872 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1873
1874 // Mark location of rbp,
1875 // map->set_callee_saved(VMRegImpl::stack2reg( stack_slots - 2), stack_slots * 2, 0, rbp->as_VMReg());
1876
1877 // We know that we only have args in at most two integer registers (rcx, rdx). So rax, rbx
1878 // Are free to temporaries if we have to do stack to steck moves.
1879 // All inbound args are referenced based on rbp, and all outbound args via rsp.
1880
1881 for (int i = 0; i < total_in_args ; i++, c_arg++ ) {
1882 switch (in_sig_bt[i]) {
1883 case T_ARRAY:
1884 if (is_critical_native) {
1885 unpack_array_argument(masm, in_regs[i], in_elem_bt[i], out_regs[c_arg + 1], out_regs[c_arg]);
1886 c_arg++;
1887 break;
1888 }
1889 case T_OBJECT:
1890 assert(!is_critical_native, "no oop arguments");
1891 object_move(masm, map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
1892 ((i == 0) && (!is_static)),
1893 &receiver_offset);
1894 break;
1895 case T_VOID:
1896 break;
1897
1898 case T_FLOAT:
1899 float_move(masm, in_regs[i], out_regs[c_arg]);
1900 break;
1901
1902 case T_DOUBLE:
1903 assert( i + 1 < total_in_args &&
1904 in_sig_bt[i + 1] == T_VOID &&
1905 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
1906 double_move(masm, in_regs[i], out_regs[c_arg]);
1907 break;
1908
1909 case T_LONG :
1910 long_move(masm, in_regs[i], out_regs[c_arg]);
1911 break;
1912
1913 case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
1914
1915 default:
1916 simple_move32(masm, in_regs[i], out_regs[c_arg]);
1917 }
1918 }
1919
1920 // Pre-load a static method's oop into rsi. Used both by locking code and
1921 // the normal JNI call code.
1922 if (method->is_static() && !is_critical_native) {
1923
1924 // load opp into a register
1925 __ movoop(oop_handle_reg, JNIHandles::make_local(method->method_holder()->java_mirror()));
1926
1927 // Now handlize the static class mirror it's known not-null.
1928 __ movptr(Address(rsp, klass_offset), oop_handle_reg);
1929 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
1930
1931 // Now get the handle
1932 __ lea(oop_handle_reg, Address(rsp, klass_offset));
1933 // store the klass handle as second argument
1934 __ movptr(Address(rsp, wordSize), oop_handle_reg);
1935 }
1936
1937 // Change state to native (we save the return address in the thread, since it might not
1938 // be pushed on the stack when we do a a stack traversal). It is enough that the pc()
1939 // points into the right code segment. It does not have to be the correct return pc.
1940 // We use the same pc/oopMap repeatedly when we call out
1941
1942 intptr_t the_pc = (intptr_t) __ pc();
1943 oop_maps->add_gc_map(the_pc - start, map);
1944
1945 __ set_last_Java_frame(thread, rsp, noreg, (address)the_pc);
1946
1947
1948 // We have all of the arguments setup at this point. We must not touch any register
1949 // argument registers at this point (what if we save/restore them there are no oop?
1950
1951 {
1952 SkipIfEqual skip_if(masm, &DTraceMethodProbes, 0);
1953 __ mov_metadata(rax, method());
1954 __ call_VM_leaf(
1955 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
1956 thread, rax);
1957 }
1958
1959 // RedefineClasses() tracing support for obsolete method entry
1960 if (log_is_enabled(Trace, redefine, class, obsolete)) {
1961 __ mov_metadata(rax, method());
1962 __ call_VM_leaf(
1963 CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
1964 thread, rax);
1965 }
1966
1967 // These are register definitions we need for locking/unlocking
1968 const Register swap_reg = rax; // Must use rax, for cmpxchg instruction
1969 const Register obj_reg = rcx; // Will contain the oop
1970 const Register lock_reg = rdx; // Address of compiler lock object (BasicLock)
1971
1972 Label slow_path_lock;
1973 Label lock_done;
1974
1975 // Lock a synchronized method
1976 if (method->is_synchronized()) {
1977 assert(!is_critical_native, "unhandled");
1978
1979
1980 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes();
1981
1982 // Get the handle (the 2nd argument)
1983 __ movptr(oop_handle_reg, Address(rsp, wordSize));
1984
1985 // Get address of the box
1986
1987 __ lea(lock_reg, Address(rbp, lock_slot_rbp_offset));
1988
1989 // Load the oop from the handle
1990 __ movptr(obj_reg, Address(oop_handle_reg, 0));
1991
1992 if (UseBiasedLocking) {
1993 // Note that oop_handle_reg is trashed during this call
1994 __ biased_locking_enter(lock_reg, obj_reg, swap_reg, oop_handle_reg, false, lock_done, &slow_path_lock);
1995 }
1996
1997 // Load immediate 1 into swap_reg %rax,
1998 __ movptr(swap_reg, 1);
1999
2000 // Load (object->mark() | 1) into swap_reg %rax,
2001 __ orptr(swap_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
2002
2003 // Save (object->mark() | 1) into BasicLock's displaced header
2004 __ movptr(Address(lock_reg, mark_word_offset), swap_reg);
2005
2006 if (os::is_MP()) {
2007 __ lock();
2008 }
2009
2010 // src -> dest iff dest == rax, else rax, <- dest
2011 // *obj_reg = lock_reg iff *obj_reg == rax, else rax, = *(obj_reg)
2012 __ cmpxchgptr(lock_reg, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
2013 __ jcc(Assembler::equal, lock_done);
2014
2015 // Test if the oopMark is an obvious stack pointer, i.e.,
2016 // 1) (mark & 3) == 0, and
2017 // 2) rsp <= mark < mark + os::pagesize()
2018 // These 3 tests can be done by evaluating the following
2019 // expression: ((mark - rsp) & (3 - os::vm_page_size())),
2020 // assuming both stack pointer and pagesize have their
2021 // least significant 2 bits clear.
2022 // NOTE: the oopMark is in swap_reg %rax, as the result of cmpxchg
2023
2024 __ subptr(swap_reg, rsp);
2025 __ andptr(swap_reg, 3 - os::vm_page_size());
2026
2027 // Save the test result, for recursive case, the result is zero
2028 __ movptr(Address(lock_reg, mark_word_offset), swap_reg);
2029 __ jcc(Assembler::notEqual, slow_path_lock);
2030 // Slow path will re-enter here
2031 __ bind(lock_done);
2032
2033 if (UseBiasedLocking) {
2034 // Re-fetch oop_handle_reg as we trashed it above
2035 __ movptr(oop_handle_reg, Address(rsp, wordSize));
2036 }
2037 }
2038
2039
2040 // Finally just about ready to make the JNI call
2041
2042
2043 // get JNIEnv* which is first argument to native
2044 if (!is_critical_native) {
2045 __ lea(rdx, Address(thread, in_bytes(JavaThread::jni_environment_offset())));
2046 __ movptr(Address(rsp, 0), rdx);
2047 }
2048
2049 // Now set thread in native
2050 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_native);
2051
2052 __ call(RuntimeAddress(native_func));
2053
2054 // Verify or restore cpu control state after JNI call
2055 __ restore_cpu_control_state_after_jni();
2056
2057 // WARNING - on Windows Java Natives use pascal calling convention and pop the
2058 // arguments off of the stack. We could just re-adjust the stack pointer here
2059 // and continue to do SP relative addressing but we instead switch to FP
2060 // relative addressing.
2061
2062 // Unpack native results.
2063 switch (ret_type) {
2064 case T_BOOLEAN: __ c2bool(rax); break;
2065 case T_CHAR : __ andptr(rax, 0xFFFF); break;
2066 case T_BYTE : __ sign_extend_byte (rax); break;
2067 case T_SHORT : __ sign_extend_short(rax); break;
2068 case T_INT : /* nothing to do */ break;
2069 case T_DOUBLE :
2070 case T_FLOAT :
2071 // Result is in st0 we'll save as needed
2072 break;
2073 case T_ARRAY: // Really a handle
2074 case T_OBJECT: // Really a handle
2075 break; // can't de-handlize until after safepoint check
2076 case T_VOID: break;
2077 case T_LONG: break;
2078 default : ShouldNotReachHere();
2079 }
2080
2081 // Switch thread to "native transition" state before reading the synchronization state.
2082 // This additional state is necessary because reading and testing the synchronization
2083 // state is not atomic w.r.t. GC, as this scenario demonstrates:
2084 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
2085 // VM thread changes sync state to synchronizing and suspends threads for GC.
2086 // Thread A is resumed to finish this native method, but doesn't block here since it
2087 // didn't see any synchronization is progress, and escapes.
2088 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_native_trans);
2089
2090 if(os::is_MP()) {
2091 if (UseMembar) {
2092 // Force this write out before the read below
2093 __ membar(Assembler::Membar_mask_bits(
2094 Assembler::LoadLoad | Assembler::LoadStore |
2095 Assembler::StoreLoad | Assembler::StoreStore));
2096 } else {
2097 // Write serialization page so VM thread can do a pseudo remote membar.
2098 // We use the current thread pointer to calculate a thread specific
2099 // offset to write to within the page. This minimizes bus traffic
2100 // due to cache line collision.
2101 __ serialize_memory(thread, rcx);
2102 }
2103 }
2104
2105 if (AlwaysRestoreFPU) {
2106 // Make sure the control word is correct.
2107 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std()));
2108 }
2109
2110 Label after_transition;
2111
2112 // check for safepoint operation in progress and/or pending suspend requests
2113 { Label Continue;
2114
2115 __ cmp32(ExternalAddress((address)SafepointSynchronize::address_of_state()),
2116 SafepointSynchronize::_not_synchronized);
2117
2118 Label L;
2119 __ jcc(Assembler::notEqual, L);
2120 __ cmpl(Address(thread, JavaThread::suspend_flags_offset()), 0);
2121 __ jcc(Assembler::equal, Continue);
2122 __ bind(L);
2123
2124 // Don't use call_VM as it will see a possible pending exception and forward it
2125 // and never return here preventing us from clearing _last_native_pc down below.
2126 // Also can't use call_VM_leaf either as it will check to see if rsi & rdi are
2127 // preserved and correspond to the bcp/locals pointers. So we do a runtime call
2128 // by hand.
2129 //
2130 __ vzeroupper();
2131
2132 save_native_result(masm, ret_type, stack_slots);
2133 __ push(thread);
2134 if (!is_critical_native) {
2135 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address,
2136 JavaThread::check_special_condition_for_native_trans)));
2137 } else {
2138 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address,
2139 JavaThread::check_special_condition_for_native_trans_and_transition)));
2140 }
2141 __ increment(rsp, wordSize);
2142 // Restore any method result value
2143 restore_native_result(masm, ret_type, stack_slots);
2144
2145 if (is_critical_native) {
2146 // The call above performed the transition to thread_in_Java so
2147 // skip the transition logic below.
2148 __ jmpb(after_transition);
2149 }
2150
2151 __ bind(Continue);
2152 }
2153
2154 // change thread state
2155 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_Java);
2156 __ bind(after_transition);
2157
2158 Label reguard;
2159 Label reguard_done;
2160 __ cmpl(Address(thread, JavaThread::stack_guard_state_offset()), JavaThread::stack_guard_yellow_reserved_disabled);
2161 __ jcc(Assembler::equal, reguard);
2162
2163 // slow path reguard re-enters here
2164 __ bind(reguard_done);
2165
2166 // Handle possible exception (will unlock if necessary)
2167
2168 // native result if any is live
2169
2170 // Unlock
2171 Label slow_path_unlock;
2172 Label unlock_done;
2173 if (method->is_synchronized()) {
2174
2175 Label done;
2176
2177 // Get locked oop from the handle we passed to jni
2178 __ movptr(obj_reg, Address(oop_handle_reg, 0));
2179
2180 if (UseBiasedLocking) {
2181 __ biased_locking_exit(obj_reg, rbx, done);
2182 }
2183
2184 // Simple recursive lock?
2185
2186 __ cmpptr(Address(rbp, lock_slot_rbp_offset), (int32_t)NULL_WORD);
2187 __ jcc(Assembler::equal, done);
2188
2189 // Must save rax, if if it is live now because cmpxchg must use it
2190 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
2191 save_native_result(masm, ret_type, stack_slots);
2192 }
2193
2194 // get old displaced header
2195 __ movptr(rbx, Address(rbp, lock_slot_rbp_offset));
2196
2197 // get address of the stack lock
2198 __ lea(rax, Address(rbp, lock_slot_rbp_offset));
2199
2200 // Atomic swap old header if oop still contains the stack lock
2201 if (os::is_MP()) {
2202 __ lock();
2203 }
2204
2205 // src -> dest iff dest == rax, else rax, <- dest
2206 // *obj_reg = rbx, iff *obj_reg == rax, else rax, = *(obj_reg)
2207 __ cmpxchgptr(rbx, Address(obj_reg, oopDesc::mark_offset_in_bytes()));
2208 __ jcc(Assembler::notEqual, slow_path_unlock);
2209
2210 // slow path re-enters here
2211 __ bind(unlock_done);
2212 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
2213 restore_native_result(masm, ret_type, stack_slots);
2214 }
2215
2216 __ bind(done);
2217
2218 }
2219
2220 {
2221 SkipIfEqual skip_if(masm, &DTraceMethodProbes, 0);
2222 // Tell dtrace about this method exit
2223 save_native_result(masm, ret_type, stack_slots);
2224 __ mov_metadata(rax, method());
2225 __ call_VM_leaf(
2226 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
2227 thread, rax);
2228 restore_native_result(masm, ret_type, stack_slots);
2229 }
2230
2231 // We can finally stop using that last_Java_frame we setup ages ago
2232
2233 __ reset_last_Java_frame(thread, false);
2234
2235 // Unbox oop result, e.g. JNIHandles::resolve value.
2236 if (ret_type == T_OBJECT || ret_type == T_ARRAY) {
2237 __ resolve_jobject(rax /* value */,
2238 thread /* thread */,
2239 rcx /* tmp */);
2240 }
2241
2242 if (CheckJNICalls) {
2243 // clear_pending_jni_exception_check
2244 __ movptr(Address(thread, JavaThread::pending_jni_exception_check_fn_offset()), NULL_WORD);
2245 }
2246
2247 if (!is_critical_native) {
2248 // reset handle block
2249 __ movptr(rcx, Address(thread, JavaThread::active_handles_offset()));
2250 __ movl(Address(rcx, JNIHandleBlock::top_offset_in_bytes()), NULL_WORD);
2251
2252 // Any exception pending?
2253 __ cmpptr(Address(thread, in_bytes(Thread::pending_exception_offset())), (int32_t)NULL_WORD);
2254 __ jcc(Assembler::notEqual, exception_pending);
2255 }
2256
2257 // no exception, we're almost done
2258
2259 // check that only result value is on FPU stack
2260 __ verify_FPU(ret_type == T_FLOAT || ret_type == T_DOUBLE ? 1 : 0, "native_wrapper normal exit");
2261
2262 // Fixup floating pointer results so that result looks like a return from a compiled method
2263 if (ret_type == T_FLOAT) {
2264 if (UseSSE >= 1) {
2265 // Pop st0 and store as float and reload into xmm register
2266 __ fstp_s(Address(rbp, -4));
2267 __ movflt(xmm0, Address(rbp, -4));
2268 }
2269 } else if (ret_type == T_DOUBLE) {
2270 if (UseSSE >= 2) {
2271 // Pop st0 and store as double and reload into xmm register
2272 __ fstp_d(Address(rbp, -8));
2273 __ movdbl(xmm0, Address(rbp, -8));
2274 }
2275 }
2276
2277 // Return
2278
2279 __ leave();
2280 __ ret(0);
2281
2282 // Unexpected paths are out of line and go here
2283
2284 // Slow path locking & unlocking
2285 if (method->is_synchronized()) {
2286
2287 // BEGIN Slow path lock
2288
2289 __ bind(slow_path_lock);
2290
2291 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM
2292 // args are (oop obj, BasicLock* lock, JavaThread* thread)
2293 __ push(thread);
2294 __ push(lock_reg);
2295 __ push(obj_reg);
2296 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C)));
2297 __ addptr(rsp, 3*wordSize);
2298
2299 #ifdef ASSERT
2300 { Label L;
2301 __ cmpptr(Address(thread, in_bytes(Thread::pending_exception_offset())), (int)NULL_WORD);
2302 __ jcc(Assembler::equal, L);
2303 __ stop("no pending exception allowed on exit from monitorenter");
2304 __ bind(L);
2305 }
2306 #endif
2307 __ jmp(lock_done);
2308
2309 // END Slow path lock
2310
2311 // BEGIN Slow path unlock
2312 __ bind(slow_path_unlock);
2313 __ vzeroupper();
2314 // Slow path unlock
2315
2316 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) {
2317 save_native_result(masm, ret_type, stack_slots);
2318 }
2319 // Save pending exception around call to VM (which contains an EXCEPTION_MARK)
2320
2321 __ pushptr(Address(thread, in_bytes(Thread::pending_exception_offset())));
2322 __ movptr(Address(thread, in_bytes(Thread::pending_exception_offset())), NULL_WORD);
2323
2324
2325 // should be a peal
2326 // +wordSize because of the push above
2327 // args are (oop obj, BasicLock* lock, JavaThread* thread)
2328 __ push(thread);
2329 __ lea(rax, Address(rbp, lock_slot_rbp_offset));
2330 __ push(rax);
2331
2332 __ push(obj_reg);
2333 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C)));
2334 __ addptr(rsp, 3*wordSize);
2335 #ifdef ASSERT
2336 {
2337 Label L;
2338 __ cmpptr(Address(thread, in_bytes(Thread::pending_exception_offset())), (int32_t)NULL_WORD);
2339 __ jcc(Assembler::equal, L);
2340 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C");
2341 __ bind(L);
2342 }
2343 #endif /* ASSERT */
2344
2345 __ popptr(Address(thread, in_bytes(Thread::pending_exception_offset())));
2346
2347 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) {
2348 restore_native_result(masm, ret_type, stack_slots);
2349 }
2350 __ jmp(unlock_done);
2351 // END Slow path unlock
2352
2353 }
2354
2355 // SLOW PATH Reguard the stack if needed
2356
2357 __ bind(reguard);
2358 __ vzeroupper();
2359 save_native_result(masm, ret_type, stack_slots);
2360 {
2361 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages)));
2362 }
2363 restore_native_result(masm, ret_type, stack_slots);
2364 __ jmp(reguard_done);
2365
2366
2367 // BEGIN EXCEPTION PROCESSING
2368
2369 if (!is_critical_native) {
2370 // Forward the exception
2371 __ bind(exception_pending);
2372
2373 // remove possible return value from FPU register stack
2374 __ empty_FPU_stack();
2375
2376 // pop our frame
2377 __ leave();
2378 // and forward the exception
2379 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2380 }
2381
2382 __ flush();
2383
2384 nmethod *nm = nmethod::new_native_nmethod(method,
2385 compile_id,
2386 masm->code(),
2387 vep_offset,
2388 frame_complete,
2389 stack_slots / VMRegImpl::slots_per_word,
2390 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
2391 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size),
2392 oop_maps);
2393
2394 if (is_critical_native) {
2395 nm->set_lazy_critical_native(true);
2396 }
2397
2398 return nm;
2399
2400 }
2401
2402 // this function returns the adjust size (in number of words) to a c2i adapter
2403 // activation for use during deoptimization
2404 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals ) {
2405 return (callee_locals - callee_parameters) * Interpreter::stackElementWords;
2406 }
2407
2408
2409 uint SharedRuntime::out_preserve_stack_slots() {
2410 return 0;
2411 }
2412
2413 //------------------------------generate_deopt_blob----------------------------
2414 void SharedRuntime::generate_deopt_blob() {
2415 // allocate space for the code
2416 ResourceMark rm;
2417 // setup code generation tools
2418 // note: the buffer code size must account for StackShadowPages=50
2419 CodeBuffer buffer("deopt_blob", 1536, 1024);
2420 MacroAssembler* masm = new MacroAssembler(&buffer);
2421 int frame_size_in_words;
2422 OopMap* map = NULL;
2423 // Account for the extra args we place on the stack
2424 // by the time we call fetch_unroll_info
2425 const int additional_words = 2; // deopt kind, thread
2426
2427 OopMapSet *oop_maps = new OopMapSet();
2428
2429 // -------------
2430 // This code enters when returning to a de-optimized nmethod. A return
2431 // address has been pushed on the the stack, and return values are in
2432 // registers.
2433 // If we are doing a normal deopt then we were called from the patched
2434 // nmethod from the point we returned to the nmethod. So the return
2435 // address on the stack is wrong by NativeCall::instruction_size
2436 // We will adjust the value to it looks like we have the original return
2437 // address on the stack (like when we eagerly deoptimized).
2438 // In the case of an exception pending with deoptimized then we enter
2439 // with a return address on the stack that points after the call we patched
2440 // into the exception handler. We have the following register state:
2441 // rax,: exception
2442 // rbx,: exception handler
2443 // rdx: throwing pc
2444 // So in this case we simply jam rdx into the useless return address and
2445 // the stack looks just like we want.
2446 //
2447 // At this point we need to de-opt. We save the argument return
2448 // registers. We call the first C routine, fetch_unroll_info(). This
2449 // routine captures the return values and returns a structure which
2450 // describes the current frame size and the sizes of all replacement frames.
2451 // The current frame is compiled code and may contain many inlined
2452 // functions, each with their own JVM state. We pop the current frame, then
2453 // push all the new frames. Then we call the C routine unpack_frames() to
2454 // populate these frames. Finally unpack_frames() returns us the new target
2455 // address. Notice that callee-save registers are BLOWN here; they have
2456 // already been captured in the vframeArray at the time the return PC was
2457 // patched.
2458 address start = __ pc();
2459 Label cont;
2460
2461 // Prolog for non exception case!
2462
2463 // Save everything in sight.
2464
2465 map = RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words, false);
2466 // Normal deoptimization
2467 __ push(Deoptimization::Unpack_deopt);
2468 __ jmp(cont);
2469
2470 int reexecute_offset = __ pc() - start;
2471
2472 // Reexecute case
2473 // return address is the pc describes what bci to do re-execute at
2474
2475 // No need to update map as each call to save_live_registers will produce identical oopmap
2476 (void) RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words, false);
2477
2478 __ push(Deoptimization::Unpack_reexecute);
2479 __ jmp(cont);
2480
2481 int exception_offset = __ pc() - start;
2482
2483 // Prolog for exception case
2484
2485 // all registers are dead at this entry point, except for rax, and
2486 // rdx which contain the exception oop and exception pc
2487 // respectively. Set them in TLS and fall thru to the
2488 // unpack_with_exception_in_tls entry point.
2489
2490 __ get_thread(rdi);
2491 __ movptr(Address(rdi, JavaThread::exception_pc_offset()), rdx);
2492 __ movptr(Address(rdi, JavaThread::exception_oop_offset()), rax);
2493
2494 int exception_in_tls_offset = __ pc() - start;
2495
2496 // new implementation because exception oop is now passed in JavaThread
2497
2498 // Prolog for exception case
2499 // All registers must be preserved because they might be used by LinearScan
2500 // Exceptiop oop and throwing PC are passed in JavaThread
2501 // tos: stack at point of call to method that threw the exception (i.e. only
2502 // args are on the stack, no return address)
2503
2504 // make room on stack for the return address
2505 // It will be patched later with the throwing pc. The correct value is not
2506 // available now because loading it from memory would destroy registers.
2507 __ push(0);
2508
2509 // Save everything in sight.
2510
2511 // No need to update map as each call to save_live_registers will produce identical oopmap
2512 (void) RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words, false);
2513
2514 // Now it is safe to overwrite any register
2515
2516 // store the correct deoptimization type
2517 __ push(Deoptimization::Unpack_exception);
2518
2519 // load throwing pc from JavaThread and patch it as the return address
2520 // of the current frame. Then clear the field in JavaThread
2521 __ get_thread(rdi);
2522 __ movptr(rdx, Address(rdi, JavaThread::exception_pc_offset()));
2523 __ movptr(Address(rbp, wordSize), rdx);
2524 __ movptr(Address(rdi, JavaThread::exception_pc_offset()), NULL_WORD);
2525
2526 #ifdef ASSERT
2527 // verify that there is really an exception oop in JavaThread
2528 __ movptr(rax, Address(rdi, JavaThread::exception_oop_offset()));
2529 __ verify_oop(rax);
2530
2531 // verify that there is no pending exception
2532 Label no_pending_exception;
2533 __ movptr(rax, Address(rdi, Thread::pending_exception_offset()));
2534 __ testptr(rax, rax);
2535 __ jcc(Assembler::zero, no_pending_exception);
2536 __ stop("must not have pending exception here");
2537 __ bind(no_pending_exception);
2538 #endif
2539
2540 __ bind(cont);
2541
2542 // Compiled code leaves the floating point stack dirty, empty it.
2543 __ empty_FPU_stack();
2544
2545
2546 // Call C code. Need thread and this frame, but NOT official VM entry
2547 // crud. We cannot block on this call, no GC can happen.
2548 __ get_thread(rcx);
2549 __ push(rcx);
2550 // fetch_unroll_info needs to call last_java_frame()
2551 __ set_last_Java_frame(rcx, noreg, noreg, NULL);
2552
2553 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info)));
2554
2555 // Need to have an oopmap that tells fetch_unroll_info where to
2556 // find any register it might need.
2557
2558 oop_maps->add_gc_map( __ pc()-start, map);
2559
2560 // Discard args to fetch_unroll_info
2561 __ pop(rcx);
2562 __ pop(rcx);
2563
2564 __ get_thread(rcx);
2565 __ reset_last_Java_frame(rcx, false);
2566
2567 // Load UnrollBlock into EDI
2568 __ mov(rdi, rax);
2569
2570 // Move the unpack kind to a safe place in the UnrollBlock because
2571 // we are very short of registers
2572
2573 Address unpack_kind(rdi, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes());
2574 // retrieve the deopt kind from the UnrollBlock.
2575 __ movl(rax, unpack_kind);
2576
2577 Label noException;
2578 __ cmpl(rax, Deoptimization::Unpack_exception); // Was exception pending?
2579 __ jcc(Assembler::notEqual, noException);
2580 __ movptr(rax, Address(rcx, JavaThread::exception_oop_offset()));
2581 __ movptr(rdx, Address(rcx, JavaThread::exception_pc_offset()));
2582 __ movptr(Address(rcx, JavaThread::exception_oop_offset()), NULL_WORD);
2583 __ movptr(Address(rcx, JavaThread::exception_pc_offset()), NULL_WORD);
2584
2585 __ verify_oop(rax);
2586
2587 // Overwrite the result registers with the exception results.
2588 __ movptr(Address(rsp, RegisterSaver::raxOffset()*wordSize), rax);
2589 __ movptr(Address(rsp, RegisterSaver::rdxOffset()*wordSize), rdx);
2590
2591 __ bind(noException);
2592
2593 // Stack is back to only having register save data on the stack.
2594 // Now restore the result registers. Everything else is either dead or captured
2595 // in the vframeArray.
2596
2597 RegisterSaver::restore_result_registers(masm);
2598
2599 // Non standard control word may be leaked out through a safepoint blob, and we can
2600 // deopt at a poll point with the non standard control word. However, we should make
2601 // sure the control word is correct after restore_result_registers.
2602 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std()));
2603
2604 // All of the register save area has been popped of the stack. Only the
2605 // return address remains.
2606
2607 // Pop all the frames we must move/replace.
2608 //
2609 // Frame picture (youngest to oldest)
2610 // 1: self-frame (no frame link)
2611 // 2: deopting frame (no frame link)
2612 // 3: caller of deopting frame (could be compiled/interpreted).
2613 //
2614 // Note: by leaving the return address of self-frame on the stack
2615 // and using the size of frame 2 to adjust the stack
2616 // when we are done the return to frame 3 will still be on the stack.
2617
2618 // Pop deoptimized frame
2619 __ addptr(rsp, Address(rdi,Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset_in_bytes()));
2620
2621 // sp should be pointing at the return address to the caller (3)
2622
2623 // Pick up the initial fp we should save
2624 // restore rbp before stack bang because if stack overflow is thrown it needs to be pushed (and preserved)
2625 __ movptr(rbp, Address(rdi, Deoptimization::UnrollBlock::initial_info_offset_in_bytes()));
2626
2627 #ifdef ASSERT
2628 // Compilers generate code that bang the stack by as much as the
2629 // interpreter would need. So this stack banging should never
2630 // trigger a fault. Verify that it does not on non product builds.
2631 if (UseStackBanging) {
2632 __ movl(rbx, Address(rdi ,Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes()));
2633 __ bang_stack_size(rbx, rcx);
2634 }
2635 #endif
2636
2637 // Load array of frame pcs into ECX
2638 __ movptr(rcx,Address(rdi,Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes()));
2639
2640 __ pop(rsi); // trash the old pc
2641
2642 // Load array of frame sizes into ESI
2643 __ movptr(rsi,Address(rdi,Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes()));
2644
2645 Address counter(rdi, Deoptimization::UnrollBlock::counter_temp_offset_in_bytes());
2646
2647 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes()));
2648 __ movl(counter, rbx);
2649
2650 // Now adjust the caller's stack to make up for the extra locals
2651 // but record the original sp so that we can save it in the skeletal interpreter
2652 // frame and the stack walking of interpreter_sender will get the unextended sp
2653 // value and not the "real" sp value.
2654
2655 Address sp_temp(rdi, Deoptimization::UnrollBlock::sender_sp_temp_offset_in_bytes());
2656 __ movptr(sp_temp, rsp);
2657 __ movl2ptr(rbx, Address(rdi, Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes()));
2658 __ subptr(rsp, rbx);
2659
2660 // Push interpreter frames in a loop
2661 Label loop;
2662 __ bind(loop);
2663 __ movptr(rbx, Address(rsi, 0)); // Load frame size
2664 __ subptr(rbx, 2*wordSize); // we'll push pc and rbp, by hand
2665 __ pushptr(Address(rcx, 0)); // save return address
2666 __ enter(); // save old & set new rbp,
2667 __ subptr(rsp, rbx); // Prolog!
2668 __ movptr(rbx, sp_temp); // sender's sp
2669 // This value is corrected by layout_activation_impl
2670 __ movptr(Address(rbp, frame::interpreter_frame_last_sp_offset * wordSize), NULL_WORD);
2671 __ movptr(Address(rbp, frame::interpreter_frame_sender_sp_offset * wordSize), rbx); // Make it walkable
2672 __ movptr(sp_temp, rsp); // pass to next frame
2673 __ addptr(rsi, wordSize); // Bump array pointer (sizes)
2674 __ addptr(rcx, wordSize); // Bump array pointer (pcs)
2675 __ decrementl(counter); // decrement counter
2676 __ jcc(Assembler::notZero, loop);
2677 __ pushptr(Address(rcx, 0)); // save final return address
2678
2679 // Re-push self-frame
2680 __ enter(); // save old & set new rbp,
2681
2682 // Return address and rbp, are in place
2683 // We'll push additional args later. Just allocate a full sized
2684 // register save area
2685 __ subptr(rsp, (frame_size_in_words-additional_words - 2) * wordSize);
2686
2687 // Restore frame locals after moving the frame
2688 __ movptr(Address(rsp, RegisterSaver::raxOffset()*wordSize), rax);
2689 __ movptr(Address(rsp, RegisterSaver::rdxOffset()*wordSize), rdx);
2690 __ fstp_d(Address(rsp, RegisterSaver::fpResultOffset()*wordSize)); // Pop float stack and store in local
2691 if( UseSSE>=2 ) __ movdbl(Address(rsp, RegisterSaver::xmm0Offset()*wordSize), xmm0);
2692 if( UseSSE==1 ) __ movflt(Address(rsp, RegisterSaver::xmm0Offset()*wordSize), xmm0);
2693
2694 // Set up the args to unpack_frame
2695
2696 __ pushl(unpack_kind); // get the unpack_kind value
2697 __ get_thread(rcx);
2698 __ push(rcx);
2699
2700 // set last_Java_sp, last_Java_fp
2701 __ set_last_Java_frame(rcx, noreg, rbp, NULL);
2702
2703 // Call C code. Need thread but NOT official VM entry
2704 // crud. We cannot block on this call, no GC can happen. Call should
2705 // restore return values to their stack-slots with the new SP.
2706 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames)));
2707 // Set an oopmap for the call site
2708 oop_maps->add_gc_map( __ pc()-start, new OopMap( frame_size_in_words, 0 ));
2709
2710 // rax, contains the return result type
2711 __ push(rax);
2712
2713 __ get_thread(rcx);
2714 __ reset_last_Java_frame(rcx, false);
2715
2716 // Collect return values
2717 __ movptr(rax,Address(rsp, (RegisterSaver::raxOffset() + additional_words + 1)*wordSize));
2718 __ movptr(rdx,Address(rsp, (RegisterSaver::rdxOffset() + additional_words + 1)*wordSize));
2719
2720 // Clear floating point stack before returning to interpreter
2721 __ empty_FPU_stack();
2722
2723 // Check if we should push the float or double return value.
2724 Label results_done, yes_double_value;
2725 __ cmpl(Address(rsp, 0), T_DOUBLE);
2726 __ jcc (Assembler::zero, yes_double_value);
2727 __ cmpl(Address(rsp, 0), T_FLOAT);
2728 __ jcc (Assembler::notZero, results_done);
2729
2730 // return float value as expected by interpreter
2731 if( UseSSE>=1 ) __ movflt(xmm0, Address(rsp, (RegisterSaver::xmm0Offset() + additional_words + 1)*wordSize));
2732 else __ fld_d(Address(rsp, (RegisterSaver::fpResultOffset() + additional_words + 1)*wordSize));
2733 __ jmp(results_done);
2734
2735 // return double value as expected by interpreter
2736 __ bind(yes_double_value);
2737 if( UseSSE>=2 ) __ movdbl(xmm0, Address(rsp, (RegisterSaver::xmm0Offset() + additional_words + 1)*wordSize));
2738 else __ fld_d(Address(rsp, (RegisterSaver::fpResultOffset() + additional_words + 1)*wordSize));
2739
2740 __ bind(results_done);
2741
2742 // Pop self-frame.
2743 __ leave(); // Epilog!
2744
2745 // Jump to interpreter
2746 __ ret(0);
2747
2748 // -------------
2749 // make sure all code is generated
2750 masm->flush();
2751
2752 _deopt_blob = DeoptimizationBlob::create( &buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words);
2753 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
2754 }
2755
2756
2757 #ifdef COMPILER2
2758 //------------------------------generate_uncommon_trap_blob--------------------
2759 void SharedRuntime::generate_uncommon_trap_blob() {
2760 // allocate space for the code
2761 ResourceMark rm;
2762 // setup code generation tools
2763 CodeBuffer buffer("uncommon_trap_blob", 512, 512);
2764 MacroAssembler* masm = new MacroAssembler(&buffer);
2765
2766 enum frame_layout {
2767 arg0_off, // thread sp + 0 // Arg location for
2768 arg1_off, // unloaded_class_index sp + 1 // calling C
2769 arg2_off, // exec_mode sp + 2
2770 // The frame sender code expects that rbp will be in the "natural" place and
2771 // will override any oopMap setting for it. We must therefore force the layout
2772 // so that it agrees with the frame sender code.
2773 rbp_off, // callee saved register sp + 3
2774 return_off, // slot for return address sp + 4
2775 framesize
2776 };
2777
2778 address start = __ pc();
2779
2780 if (UseRTMLocking) {
2781 // Abort RTM transaction before possible nmethod deoptimization.
2782 __ xabort(0);
2783 }
2784
2785 // Push self-frame.
2786 __ subptr(rsp, return_off*wordSize); // Epilog!
2787
2788 // rbp, is an implicitly saved callee saved register (i.e. the calling
2789 // convention will save restore it in prolog/epilog) Other than that
2790 // there are no callee save registers no that adapter frames are gone.
2791 __ movptr(Address(rsp, rbp_off*wordSize), rbp);
2792
2793 // Clear the floating point exception stack
2794 __ empty_FPU_stack();
2795
2796 // set last_Java_sp
2797 __ get_thread(rdx);
2798 __ set_last_Java_frame(rdx, noreg, noreg, NULL);
2799
2800 // Call C code. Need thread but NOT official VM entry
2801 // crud. We cannot block on this call, no GC can happen. Call should
2802 // capture callee-saved registers as well as return values.
2803 __ movptr(Address(rsp, arg0_off*wordSize), rdx);
2804 // argument already in ECX
2805 __ movl(Address(rsp, arg1_off*wordSize),rcx);
2806 __ movl(Address(rsp, arg2_off*wordSize), Deoptimization::Unpack_uncommon_trap);
2807 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap)));
2808
2809 // Set an oopmap for the call site
2810 OopMapSet *oop_maps = new OopMapSet();
2811 OopMap* map = new OopMap( framesize, 0 );
2812 // No oopMap for rbp, it is known implicitly
2813
2814 oop_maps->add_gc_map( __ pc()-start, map);
2815
2816 __ get_thread(rcx);
2817
2818 __ reset_last_Java_frame(rcx, false);
2819
2820 // Load UnrollBlock into EDI
2821 __ movptr(rdi, rax);
2822
2823 #ifdef ASSERT
2824 { Label L;
2825 __ cmpptr(Address(rdi, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes()),
2826 (int32_t)Deoptimization::Unpack_uncommon_trap);
2827 __ jcc(Assembler::equal, L);
2828 __ stop("SharedRuntime::generate_deopt_blob: expected Unpack_uncommon_trap");
2829 __ bind(L);
2830 }
2831 #endif
2832
2833 // Pop all the frames we must move/replace.
2834 //
2835 // Frame picture (youngest to oldest)
2836 // 1: self-frame (no frame link)
2837 // 2: deopting frame (no frame link)
2838 // 3: caller of deopting frame (could be compiled/interpreted).
2839
2840 // Pop self-frame. We have no frame, and must rely only on EAX and ESP.
2841 __ addptr(rsp,(framesize-1)*wordSize); // Epilog!
2842
2843 // Pop deoptimized frame
2844 __ movl2ptr(rcx, Address(rdi,Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset_in_bytes()));
2845 __ addptr(rsp, rcx);
2846
2847 // sp should be pointing at the return address to the caller (3)
2848
2849 // Pick up the initial fp we should save
2850 // restore rbp before stack bang because if stack overflow is thrown it needs to be pushed (and preserved)
2851 __ movptr(rbp, Address(rdi, Deoptimization::UnrollBlock::initial_info_offset_in_bytes()));
2852
2853 #ifdef ASSERT
2854 // Compilers generate code that bang the stack by as much as the
2855 // interpreter would need. So this stack banging should never
2856 // trigger a fault. Verify that it does not on non product builds.
2857 if (UseStackBanging) {
2858 __ movl(rbx, Address(rdi ,Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes()));
2859 __ bang_stack_size(rbx, rcx);
2860 }
2861 #endif
2862
2863 // Load array of frame pcs into ECX
2864 __ movl(rcx,Address(rdi,Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes()));
2865
2866 __ pop(rsi); // trash the pc
2867
2868 // Load array of frame sizes into ESI
2869 __ movptr(rsi,Address(rdi,Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes()));
2870
2871 Address counter(rdi, Deoptimization::UnrollBlock::counter_temp_offset_in_bytes());
2872
2873 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes()));
2874 __ movl(counter, rbx);
2875
2876 // Now adjust the caller's stack to make up for the extra locals
2877 // but record the original sp so that we can save it in the skeletal interpreter
2878 // frame and the stack walking of interpreter_sender will get the unextended sp
2879 // value and not the "real" sp value.
2880
2881 Address sp_temp(rdi, Deoptimization::UnrollBlock::sender_sp_temp_offset_in_bytes());
2882 __ movptr(sp_temp, rsp);
2883 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes()));
2884 __ subptr(rsp, rbx);
2885
2886 // Push interpreter frames in a loop
2887 Label loop;
2888 __ bind(loop);
2889 __ movptr(rbx, Address(rsi, 0)); // Load frame size
2890 __ subptr(rbx, 2*wordSize); // we'll push pc and rbp, by hand
2891 __ pushptr(Address(rcx, 0)); // save return address
2892 __ enter(); // save old & set new rbp,
2893 __ subptr(rsp, rbx); // Prolog!
2894 __ movptr(rbx, sp_temp); // sender's sp
2895 // This value is corrected by layout_activation_impl
2896 __ movptr(Address(rbp, frame::interpreter_frame_last_sp_offset * wordSize), NULL_WORD );
2897 __ movptr(Address(rbp, frame::interpreter_frame_sender_sp_offset * wordSize), rbx); // Make it walkable
2898 __ movptr(sp_temp, rsp); // pass to next frame
2899 __ addptr(rsi, wordSize); // Bump array pointer (sizes)
2900 __ addptr(rcx, wordSize); // Bump array pointer (pcs)
2901 __ decrementl(counter); // decrement counter
2902 __ jcc(Assembler::notZero, loop);
2903 __ pushptr(Address(rcx, 0)); // save final return address
2904
2905 // Re-push self-frame
2906 __ enter(); // save old & set new rbp,
2907 __ subptr(rsp, (framesize-2) * wordSize); // Prolog!
2908
2909
2910 // set last_Java_sp, last_Java_fp
2911 __ get_thread(rdi);
2912 __ set_last_Java_frame(rdi, noreg, rbp, NULL);
2913
2914 // Call C code. Need thread but NOT official VM entry
2915 // crud. We cannot block on this call, no GC can happen. Call should
2916 // restore return values to their stack-slots with the new SP.
2917 __ movptr(Address(rsp,arg0_off*wordSize),rdi);
2918 __ movl(Address(rsp,arg1_off*wordSize), Deoptimization::Unpack_uncommon_trap);
2919 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames)));
2920 // Set an oopmap for the call site
2921 oop_maps->add_gc_map( __ pc()-start, new OopMap( framesize, 0 ) );
2922
2923 __ get_thread(rdi);
2924 __ reset_last_Java_frame(rdi, true);
2925
2926 // Pop self-frame.
2927 __ leave(); // Epilog!
2928
2929 // Jump to interpreter
2930 __ ret(0);
2931
2932 // -------------
2933 // make sure all code is generated
2934 masm->flush();
2935
2936 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, framesize);
2937 }
2938 #endif // COMPILER2
2939
2940 //------------------------------generate_handler_blob------
2941 //
2942 // Generate a special Compile2Runtime blob that saves all registers,
2943 // setup oopmap, and calls safepoint code to stop the compiled code for
2944 // a safepoint.
2945 //
2946 SafepointBlob* SharedRuntime::generate_handler_blob(address call_ptr, int poll_type) {
2947
2948 // Account for thread arg in our frame
2949 const int additional_words = 1;
2950 int frame_size_in_words;
2951
2952 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before");
2953
2954 ResourceMark rm;
2955 OopMapSet *oop_maps = new OopMapSet();
2956 OopMap* map;
2957
2958 // allocate space for the code
2959 // setup code generation tools
2960 CodeBuffer buffer("handler_blob", 1024, 512);
2961 MacroAssembler* masm = new MacroAssembler(&buffer);
2962
2963 const Register java_thread = rdi; // callee-saved for VC++
2964 address start = __ pc();
2965 address call_pc = NULL;
2966 bool cause_return = (poll_type == POLL_AT_RETURN);
2967 bool save_vectors = (poll_type == POLL_AT_VECTOR_LOOP);
2968
2969 if (UseRTMLocking) {
2970 // Abort RTM transaction before calling runtime
2971 // because critical section will be large and will be
2972 // aborted anyway. Also nmethod could be deoptimized.
2973 __ xabort(0);
2974 }
2975
2976 // If cause_return is true we are at a poll_return and there is
2977 // the return address on the stack to the caller on the nmethod
2978 // that is safepoint. We can leave this return on the stack and
2979 // effectively complete the return and safepoint in the caller.
2980 // Otherwise we push space for a return address that the safepoint
2981 // handler will install later to make the stack walking sensible.
2982 if (!cause_return)
2983 __ push(rbx); // Make room for return address (or push it again)
2984
2985 map = RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words, false, save_vectors);
2986
2987 // The following is basically a call_VM. However, we need the precise
2988 // address of the call in order to generate an oopmap. Hence, we do all the
2989 // work ourselves.
2990
2991 // Push thread argument and setup last_Java_sp
2992 __ get_thread(java_thread);
2993 __ push(java_thread);
2994 __ set_last_Java_frame(java_thread, noreg, noreg, NULL);
2995
2996 // if this was not a poll_return then we need to correct the return address now.
2997 if (!cause_return) {
2998 __ movptr(rax, Address(java_thread, JavaThread::saved_exception_pc_offset()));
2999 __ movptr(Address(rbp, wordSize), rax);
3000 }
3001
3002 // do the call
3003 __ call(RuntimeAddress(call_ptr));
3004
3005 // Set an oopmap for the call site. This oopmap will map all
3006 // oop-registers and debug-info registers as callee-saved. This
3007 // will allow deoptimization at this safepoint to find all possible
3008 // debug-info recordings, as well as let GC find all oops.
3009
3010 oop_maps->add_gc_map( __ pc() - start, map);
3011
3012 // Discard arg
3013 __ pop(rcx);
3014
3015 Label noException;
3016
3017 // Clear last_Java_sp again
3018 __ get_thread(java_thread);
3019 __ reset_last_Java_frame(java_thread, false);
3020
3021 __ cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
3022 __ jcc(Assembler::equal, noException);
3023
3024 // Exception pending
3025 RegisterSaver::restore_live_registers(masm, save_vectors);
3026
3027 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3028
3029 __ bind(noException);
3030
3031 // Normal exit, register restoring and exit
3032 RegisterSaver::restore_live_registers(masm, save_vectors);
3033
3034 __ ret(0);
3035
3036 // make sure all code is generated
3037 masm->flush();
3038
3039 // Fill-out other meta info
3040 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words);
3041 }
3042
3043 //
3044 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
3045 //
3046 // Generate a stub that calls into vm to find out the proper destination
3047 // of a java call. All the argument registers are live at this point
3048 // but since this is generic code we don't know what they are and the caller
3049 // must do any gc of the args.
3050 //
3051 RuntimeStub* SharedRuntime::generate_resolve_blob(address destination, const char* name) {
3052 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before");
3053
3054 // allocate space for the code
3055 ResourceMark rm;
3056
3057 CodeBuffer buffer(name, 1000, 512);
3058 MacroAssembler* masm = new MacroAssembler(&buffer);
3059
3060 int frame_size_words;
3061 enum frame_layout {
3062 thread_off,
3063 extra_words };
3064
3065 OopMapSet *oop_maps = new OopMapSet();
3066 OopMap* map = NULL;
3067
3068 int start = __ offset();
3069
3070 map = RegisterSaver::save_live_registers(masm, extra_words, &frame_size_words);
3071
3072 int frame_complete = __ offset();
3073
3074 const Register thread = rdi;
3075 __ get_thread(rdi);
3076
3077 __ push(thread);
3078 __ set_last_Java_frame(thread, noreg, rbp, NULL);
3079
3080 __ call(RuntimeAddress(destination));
3081
3082
3083 // Set an oopmap for the call site.
3084 // We need this not only for callee-saved registers, but also for volatile
3085 // registers that the compiler might be keeping live across a safepoint.
3086
3087 oop_maps->add_gc_map( __ offset() - start, map);
3088
3089 // rax, contains the address we are going to jump to assuming no exception got installed
3090
3091 __ addptr(rsp, wordSize);
3092
3093 // clear last_Java_sp
3094 __ reset_last_Java_frame(thread, true);
3095 // check for pending exceptions
3096 Label pending;
3097 __ cmpptr(Address(thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
3098 __ jcc(Assembler::notEqual, pending);
3099
3100 // get the returned Method*
3101 __ get_vm_result_2(rbx, thread);
3102 __ movptr(Address(rsp, RegisterSaver::rbx_offset() * wordSize), rbx);
3103
3104 __ movptr(Address(rsp, RegisterSaver::rax_offset() * wordSize), rax);
3105
3106 RegisterSaver::restore_live_registers(masm);
3107
3108 // We are back the the original state on entry and ready to go.
3109
3110 __ jmp(rax);
3111
3112 // Pending exception after the safepoint
3113
3114 __ bind(pending);
3115
3116 RegisterSaver::restore_live_registers(masm);
3117
3118 // exception pending => remove activation and forward to exception handler
3119
3120 __ get_thread(thread);
3121 __ movptr(Address(thread, JavaThread::vm_result_offset()), NULL_WORD);
3122 __ movptr(rax, Address(thread, Thread::pending_exception_offset()));
3123 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3124
3125 // -------------
3126 // make sure all code is generated
3127 masm->flush();
3128
3129 // return the blob
3130 // frame_size_words or bytes??
3131 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_words, oop_maps, true);
3132 }