Tue, 14 Oct 2008 15:10:26 -0700
6532536: Optimize arraycopy stubs for Intel cpus
Summary: Use SSE2 movdqu in arraycopy stubs on newest Intel's cpus
Reviewed-by: rasbold
1 /*
2 * Copyright 2003-2008 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_sharedRuntime_x86_32.cpp.incl"
28 #define __ masm->
29 #ifdef COMPILER2
30 UncommonTrapBlob *SharedRuntime::_uncommon_trap_blob;
31 #endif // COMPILER2
33 DeoptimizationBlob *SharedRuntime::_deopt_blob;
34 SafepointBlob *SharedRuntime::_polling_page_safepoint_handler_blob;
35 SafepointBlob *SharedRuntime::_polling_page_return_handler_blob;
36 RuntimeStub* SharedRuntime::_wrong_method_blob;
37 RuntimeStub* SharedRuntime::_ic_miss_blob;
38 RuntimeStub* SharedRuntime::_resolve_opt_virtual_call_blob;
39 RuntimeStub* SharedRuntime::_resolve_virtual_call_blob;
40 RuntimeStub* SharedRuntime::_resolve_static_call_blob;
42 class RegisterSaver {
43 enum { FPU_regs_live = 8 /*for the FPU stack*/+8/*eight more for XMM registers*/ };
44 // Capture info about frame layout
45 enum layout {
46 fpu_state_off = 0,
47 fpu_state_end = fpu_state_off+FPUStateSizeInWords-1,
48 st0_off, st0H_off,
49 st1_off, st1H_off,
50 st2_off, st2H_off,
51 st3_off, st3H_off,
52 st4_off, st4H_off,
53 st5_off, st5H_off,
54 st6_off, st6H_off,
55 st7_off, st7H_off,
57 xmm0_off, xmm0H_off,
58 xmm1_off, xmm1H_off,
59 xmm2_off, xmm2H_off,
60 xmm3_off, xmm3H_off,
61 xmm4_off, xmm4H_off,
62 xmm5_off, xmm5H_off,
63 xmm6_off, xmm6H_off,
64 xmm7_off, xmm7H_off,
65 flags_off,
66 rdi_off,
67 rsi_off,
68 ignore_off, // extra copy of rbp,
69 rsp_off,
70 rbx_off,
71 rdx_off,
72 rcx_off,
73 rax_off,
74 // The frame sender code expects that rbp will be in the "natural" place and
75 // will override any oopMap setting for it. We must therefore force the layout
76 // so that it agrees with the frame sender code.
77 rbp_off,
78 return_off, // slot for return address
79 reg_save_size };
82 public:
84 static OopMap* save_live_registers(MacroAssembler* masm, int additional_frame_words,
85 int* total_frame_words, bool verify_fpu = true);
86 static void restore_live_registers(MacroAssembler* masm);
88 static int rax_offset() { return rax_off; }
89 static int rbx_offset() { return rbx_off; }
91 // Offsets into the register save area
92 // Used by deoptimization when it is managing result register
93 // values on its own
95 static int raxOffset(void) { return rax_off; }
96 static int rdxOffset(void) { return rdx_off; }
97 static int rbxOffset(void) { return rbx_off; }
98 static int xmm0Offset(void) { return xmm0_off; }
99 // This really returns a slot in the fp save area, which one is not important
100 static int fpResultOffset(void) { return st0_off; }
102 // During deoptimization only the result register need to be restored
103 // all the other values have already been extracted.
105 static void restore_result_registers(MacroAssembler* masm);
107 };
109 OopMap* RegisterSaver::save_live_registers(MacroAssembler* masm, int additional_frame_words,
110 int* total_frame_words, bool verify_fpu) {
112 int frame_size_in_bytes = (reg_save_size + additional_frame_words) * wordSize;
113 int frame_words = frame_size_in_bytes / wordSize;
114 *total_frame_words = frame_words;
116 assert(FPUStateSizeInWords == 27, "update stack layout");
118 // save registers, fpu state, and flags
119 // We assume caller has already has return address slot on the stack
120 // We push epb twice in this sequence because we want the real rbp,
121 // to be under the return like a normal enter and we want to use pusha
122 // We push by hand instead of pusing push
123 __ enter();
124 __ pusha();
125 __ pushf();
126 __ subptr(rsp,FPU_regs_live*sizeof(jdouble)); // Push FPU registers space
127 __ push_FPU_state(); // Save FPU state & init
129 if (verify_fpu) {
130 // Some stubs may have non standard FPU control word settings so
131 // only check and reset the value when it required to be the
132 // standard value. The safepoint blob in particular can be used
133 // in methods which are using the 24 bit control word for
134 // optimized float math.
136 #ifdef ASSERT
137 // Make sure the control word has the expected value
138 Label ok;
139 __ cmpw(Address(rsp, 0), StubRoutines::fpu_cntrl_wrd_std());
140 __ jccb(Assembler::equal, ok);
141 __ stop("corrupted control word detected");
142 __ bind(ok);
143 #endif
145 // Reset the control word to guard against exceptions being unmasked
146 // since fstp_d can cause FPU stack underflow exceptions. Write it
147 // into the on stack copy and then reload that to make sure that the
148 // current and future values are correct.
149 __ movw(Address(rsp, 0), StubRoutines::fpu_cntrl_wrd_std());
150 }
152 __ frstor(Address(rsp, 0));
153 if (!verify_fpu) {
154 // Set the control word so that exceptions are masked for the
155 // following code.
156 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std()));
157 }
159 // Save the FPU registers in de-opt-able form
161 __ fstp_d(Address(rsp, st0_off*wordSize)); // st(0)
162 __ fstp_d(Address(rsp, st1_off*wordSize)); // st(1)
163 __ fstp_d(Address(rsp, st2_off*wordSize)); // st(2)
164 __ fstp_d(Address(rsp, st3_off*wordSize)); // st(3)
165 __ fstp_d(Address(rsp, st4_off*wordSize)); // st(4)
166 __ fstp_d(Address(rsp, st5_off*wordSize)); // st(5)
167 __ fstp_d(Address(rsp, st6_off*wordSize)); // st(6)
168 __ fstp_d(Address(rsp, st7_off*wordSize)); // st(7)
170 if( UseSSE == 1 ) { // Save the XMM state
171 __ movflt(Address(rsp,xmm0_off*wordSize),xmm0);
172 __ movflt(Address(rsp,xmm1_off*wordSize),xmm1);
173 __ movflt(Address(rsp,xmm2_off*wordSize),xmm2);
174 __ movflt(Address(rsp,xmm3_off*wordSize),xmm3);
175 __ movflt(Address(rsp,xmm4_off*wordSize),xmm4);
176 __ movflt(Address(rsp,xmm5_off*wordSize),xmm5);
177 __ movflt(Address(rsp,xmm6_off*wordSize),xmm6);
178 __ movflt(Address(rsp,xmm7_off*wordSize),xmm7);
179 } else if( UseSSE >= 2 ) {
180 __ movdbl(Address(rsp,xmm0_off*wordSize),xmm0);
181 __ movdbl(Address(rsp,xmm1_off*wordSize),xmm1);
182 __ movdbl(Address(rsp,xmm2_off*wordSize),xmm2);
183 __ movdbl(Address(rsp,xmm3_off*wordSize),xmm3);
184 __ movdbl(Address(rsp,xmm4_off*wordSize),xmm4);
185 __ movdbl(Address(rsp,xmm5_off*wordSize),xmm5);
186 __ movdbl(Address(rsp,xmm6_off*wordSize),xmm6);
187 __ movdbl(Address(rsp,xmm7_off*wordSize),xmm7);
188 }
190 // Set an oopmap for the call site. This oopmap will map all
191 // oop-registers and debug-info registers as callee-saved. This
192 // will allow deoptimization at this safepoint to find all possible
193 // debug-info recordings, as well as let GC find all oops.
195 OopMapSet *oop_maps = new OopMapSet();
196 OopMap* map = new OopMap( frame_words, 0 );
198 #define STACK_OFFSET(x) VMRegImpl::stack2reg((x) + additional_frame_words)
200 map->set_callee_saved(STACK_OFFSET( rax_off), rax->as_VMReg());
201 map->set_callee_saved(STACK_OFFSET( rcx_off), rcx->as_VMReg());
202 map->set_callee_saved(STACK_OFFSET( rdx_off), rdx->as_VMReg());
203 map->set_callee_saved(STACK_OFFSET( rbx_off), rbx->as_VMReg());
204 // rbp, location is known implicitly, no oopMap
205 map->set_callee_saved(STACK_OFFSET( rsi_off), rsi->as_VMReg());
206 map->set_callee_saved(STACK_OFFSET( rdi_off), rdi->as_VMReg());
207 map->set_callee_saved(STACK_OFFSET(st0_off), as_FloatRegister(0)->as_VMReg());
208 map->set_callee_saved(STACK_OFFSET(st1_off), as_FloatRegister(1)->as_VMReg());
209 map->set_callee_saved(STACK_OFFSET(st2_off), as_FloatRegister(2)->as_VMReg());
210 map->set_callee_saved(STACK_OFFSET(st3_off), as_FloatRegister(3)->as_VMReg());
211 map->set_callee_saved(STACK_OFFSET(st4_off), as_FloatRegister(4)->as_VMReg());
212 map->set_callee_saved(STACK_OFFSET(st5_off), as_FloatRegister(5)->as_VMReg());
213 map->set_callee_saved(STACK_OFFSET(st6_off), as_FloatRegister(6)->as_VMReg());
214 map->set_callee_saved(STACK_OFFSET(st7_off), as_FloatRegister(7)->as_VMReg());
215 map->set_callee_saved(STACK_OFFSET(xmm0_off), xmm0->as_VMReg());
216 map->set_callee_saved(STACK_OFFSET(xmm1_off), xmm1->as_VMReg());
217 map->set_callee_saved(STACK_OFFSET(xmm2_off), xmm2->as_VMReg());
218 map->set_callee_saved(STACK_OFFSET(xmm3_off), xmm3->as_VMReg());
219 map->set_callee_saved(STACK_OFFSET(xmm4_off), xmm4->as_VMReg());
220 map->set_callee_saved(STACK_OFFSET(xmm5_off), xmm5->as_VMReg());
221 map->set_callee_saved(STACK_OFFSET(xmm6_off), xmm6->as_VMReg());
222 map->set_callee_saved(STACK_OFFSET(xmm7_off), xmm7->as_VMReg());
223 // %%% This is really a waste but we'll keep things as they were for now
224 if (true) {
225 #define NEXTREG(x) (x)->as_VMReg()->next()
226 map->set_callee_saved(STACK_OFFSET(st0H_off), NEXTREG(as_FloatRegister(0)));
227 map->set_callee_saved(STACK_OFFSET(st1H_off), NEXTREG(as_FloatRegister(1)));
228 map->set_callee_saved(STACK_OFFSET(st2H_off), NEXTREG(as_FloatRegister(2)));
229 map->set_callee_saved(STACK_OFFSET(st3H_off), NEXTREG(as_FloatRegister(3)));
230 map->set_callee_saved(STACK_OFFSET(st4H_off), NEXTREG(as_FloatRegister(4)));
231 map->set_callee_saved(STACK_OFFSET(st5H_off), NEXTREG(as_FloatRegister(5)));
232 map->set_callee_saved(STACK_OFFSET(st6H_off), NEXTREG(as_FloatRegister(6)));
233 map->set_callee_saved(STACK_OFFSET(st7H_off), NEXTREG(as_FloatRegister(7)));
234 map->set_callee_saved(STACK_OFFSET(xmm0H_off), NEXTREG(xmm0));
235 map->set_callee_saved(STACK_OFFSET(xmm1H_off), NEXTREG(xmm1));
236 map->set_callee_saved(STACK_OFFSET(xmm2H_off), NEXTREG(xmm2));
237 map->set_callee_saved(STACK_OFFSET(xmm3H_off), NEXTREG(xmm3));
238 map->set_callee_saved(STACK_OFFSET(xmm4H_off), NEXTREG(xmm4));
239 map->set_callee_saved(STACK_OFFSET(xmm5H_off), NEXTREG(xmm5));
240 map->set_callee_saved(STACK_OFFSET(xmm6H_off), NEXTREG(xmm6));
241 map->set_callee_saved(STACK_OFFSET(xmm7H_off), NEXTREG(xmm7));
242 #undef NEXTREG
243 #undef STACK_OFFSET
244 }
246 return map;
248 }
250 void RegisterSaver::restore_live_registers(MacroAssembler* masm) {
252 // Recover XMM & FPU state
253 if( UseSSE == 1 ) {
254 __ movflt(xmm0,Address(rsp,xmm0_off*wordSize));
255 __ movflt(xmm1,Address(rsp,xmm1_off*wordSize));
256 __ movflt(xmm2,Address(rsp,xmm2_off*wordSize));
257 __ movflt(xmm3,Address(rsp,xmm3_off*wordSize));
258 __ movflt(xmm4,Address(rsp,xmm4_off*wordSize));
259 __ movflt(xmm5,Address(rsp,xmm5_off*wordSize));
260 __ movflt(xmm6,Address(rsp,xmm6_off*wordSize));
261 __ movflt(xmm7,Address(rsp,xmm7_off*wordSize));
262 } else if( UseSSE >= 2 ) {
263 __ movdbl(xmm0,Address(rsp,xmm0_off*wordSize));
264 __ movdbl(xmm1,Address(rsp,xmm1_off*wordSize));
265 __ movdbl(xmm2,Address(rsp,xmm2_off*wordSize));
266 __ movdbl(xmm3,Address(rsp,xmm3_off*wordSize));
267 __ movdbl(xmm4,Address(rsp,xmm4_off*wordSize));
268 __ movdbl(xmm5,Address(rsp,xmm5_off*wordSize));
269 __ movdbl(xmm6,Address(rsp,xmm6_off*wordSize));
270 __ movdbl(xmm7,Address(rsp,xmm7_off*wordSize));
271 }
272 __ pop_FPU_state();
273 __ addptr(rsp, FPU_regs_live*sizeof(jdouble)); // Pop FPU registers
275 __ popf();
276 __ popa();
277 // Get the rbp, described implicitly by the frame sender code (no oopMap)
278 __ pop(rbp);
280 }
282 void RegisterSaver::restore_result_registers(MacroAssembler* masm) {
284 // Just restore result register. Only used by deoptimization. By
285 // now any callee save register that needs to be restore to a c2
286 // caller of the deoptee has been extracted into the vframeArray
287 // and will be stuffed into the c2i adapter we create for later
288 // restoration so only result registers need to be restored here.
289 //
291 __ frstor(Address(rsp, 0)); // Restore fpu state
293 // Recover XMM & FPU state
294 if( UseSSE == 1 ) {
295 __ movflt(xmm0, Address(rsp, xmm0_off*wordSize));
296 } else if( UseSSE >= 2 ) {
297 __ movdbl(xmm0, Address(rsp, xmm0_off*wordSize));
298 }
299 __ movptr(rax, Address(rsp, rax_off*wordSize));
300 __ movptr(rdx, Address(rsp, rdx_off*wordSize));
301 // Pop all of the register save are off the stack except the return address
302 __ addptr(rsp, return_off * wordSize);
303 }
305 // The java_calling_convention describes stack locations as ideal slots on
306 // a frame with no abi restrictions. Since we must observe abi restrictions
307 // (like the placement of the register window) the slots must be biased by
308 // the following value.
309 static int reg2offset_in(VMReg r) {
310 // Account for saved rbp, and return address
311 // This should really be in_preserve_stack_slots
312 return (r->reg2stack() + 2) * VMRegImpl::stack_slot_size;
313 }
315 static int reg2offset_out(VMReg r) {
316 return (r->reg2stack() + SharedRuntime::out_preserve_stack_slots()) * VMRegImpl::stack_slot_size;
317 }
319 // ---------------------------------------------------------------------------
320 // Read the array of BasicTypes from a signature, and compute where the
321 // arguments should go. Values in the VMRegPair regs array refer to 4-byte
322 // quantities. Values less than SharedInfo::stack0 are registers, those above
323 // refer to 4-byte stack slots. All stack slots are based off of the stack pointer
324 // as framesizes are fixed.
325 // VMRegImpl::stack0 refers to the first slot 0(sp).
326 // and VMRegImpl::stack0+1 refers to the memory word 4-byes higher. Register
327 // up to RegisterImpl::number_of_registers) are the 32-bit
328 // integer registers.
330 // Pass first two oop/int args in registers ECX and EDX.
331 // Pass first two float/double args in registers XMM0 and XMM1.
332 // Doubles have precedence, so if you pass a mix of floats and doubles
333 // the doubles will grab the registers before the floats will.
335 // Note: the INPUTS in sig_bt are in units of Java argument words, which are
336 // either 32-bit or 64-bit depending on the build. The OUTPUTS are in 32-bit
337 // units regardless of build. Of course for i486 there is no 64 bit build
340 // ---------------------------------------------------------------------------
341 // The compiled Java calling convention.
342 // Pass first two oop/int args in registers ECX and EDX.
343 // Pass first two float/double args in registers XMM0 and XMM1.
344 // Doubles have precedence, so if you pass a mix of floats and doubles
345 // the doubles will grab the registers before the floats will.
346 int SharedRuntime::java_calling_convention(const BasicType *sig_bt,
347 VMRegPair *regs,
348 int total_args_passed,
349 int is_outgoing) {
350 uint stack = 0; // Starting stack position for args on stack
353 // Pass first two oop/int args in registers ECX and EDX.
354 uint reg_arg0 = 9999;
355 uint reg_arg1 = 9999;
357 // Pass first two float/double args in registers XMM0 and XMM1.
358 // Doubles have precedence, so if you pass a mix of floats and doubles
359 // the doubles will grab the registers before the floats will.
360 // CNC - TURNED OFF FOR non-SSE.
361 // On Intel we have to round all doubles (and most floats) at
362 // call sites by storing to the stack in any case.
363 // UseSSE=0 ==> Don't Use ==> 9999+0
364 // UseSSE=1 ==> Floats only ==> 9999+1
365 // UseSSE>=2 ==> Floats or doubles ==> 9999+2
366 enum { fltarg_dontuse = 9999+0, fltarg_float_only = 9999+1, fltarg_flt_dbl = 9999+2 };
367 uint fargs = (UseSSE>=2) ? 2 : UseSSE;
368 uint freg_arg0 = 9999+fargs;
369 uint freg_arg1 = 9999+fargs;
371 // Pass doubles & longs aligned on the stack. First count stack slots for doubles
372 int i;
373 for( i = 0; i < total_args_passed; i++) {
374 if( sig_bt[i] == T_DOUBLE ) {
375 // first 2 doubles go in registers
376 if( freg_arg0 == fltarg_flt_dbl ) freg_arg0 = i;
377 else if( freg_arg1 == fltarg_flt_dbl ) freg_arg1 = i;
378 else // Else double is passed low on the stack to be aligned.
379 stack += 2;
380 } else if( sig_bt[i] == T_LONG ) {
381 stack += 2;
382 }
383 }
384 int dstack = 0; // Separate counter for placing doubles
386 // Now pick where all else goes.
387 for( i = 0; i < total_args_passed; i++) {
388 // From the type and the argument number (count) compute the location
389 switch( sig_bt[i] ) {
390 case T_SHORT:
391 case T_CHAR:
392 case T_BYTE:
393 case T_BOOLEAN:
394 case T_INT:
395 case T_ARRAY:
396 case T_OBJECT:
397 case T_ADDRESS:
398 if( reg_arg0 == 9999 ) {
399 reg_arg0 = i;
400 regs[i].set1(rcx->as_VMReg());
401 } else if( reg_arg1 == 9999 ) {
402 reg_arg1 = i;
403 regs[i].set1(rdx->as_VMReg());
404 } else {
405 regs[i].set1(VMRegImpl::stack2reg(stack++));
406 }
407 break;
408 case T_FLOAT:
409 if( freg_arg0 == fltarg_flt_dbl || freg_arg0 == fltarg_float_only ) {
410 freg_arg0 = i;
411 regs[i].set1(xmm0->as_VMReg());
412 } else if( freg_arg1 == fltarg_flt_dbl || freg_arg1 == fltarg_float_only ) {
413 freg_arg1 = i;
414 regs[i].set1(xmm1->as_VMReg());
415 } else {
416 regs[i].set1(VMRegImpl::stack2reg(stack++));
417 }
418 break;
419 case T_LONG:
420 assert(sig_bt[i+1] == T_VOID, "missing Half" );
421 regs[i].set2(VMRegImpl::stack2reg(dstack));
422 dstack += 2;
423 break;
424 case T_DOUBLE:
425 assert(sig_bt[i+1] == T_VOID, "missing Half" );
426 if( freg_arg0 == (uint)i ) {
427 regs[i].set2(xmm0->as_VMReg());
428 } else if( freg_arg1 == (uint)i ) {
429 regs[i].set2(xmm1->as_VMReg());
430 } else {
431 regs[i].set2(VMRegImpl::stack2reg(dstack));
432 dstack += 2;
433 }
434 break;
435 case T_VOID: regs[i].set_bad(); break;
436 break;
437 default:
438 ShouldNotReachHere();
439 break;
440 }
441 }
443 // return value can be odd number of VMRegImpl stack slots make multiple of 2
444 return round_to(stack, 2);
445 }
447 // Patch the callers callsite with entry to compiled code if it exists.
448 static void patch_callers_callsite(MacroAssembler *masm) {
449 Label L;
450 __ verify_oop(rbx);
451 __ cmpptr(Address(rbx, in_bytes(methodOopDesc::code_offset())), (int32_t)NULL_WORD);
452 __ jcc(Assembler::equal, L);
453 // Schedule the branch target address early.
454 // Call into the VM to patch the caller, then jump to compiled callee
455 // rax, isn't live so capture return address while we easily can
456 __ movptr(rax, Address(rsp, 0));
457 __ pusha();
458 __ pushf();
460 if (UseSSE == 1) {
461 __ subptr(rsp, 2*wordSize);
462 __ movflt(Address(rsp, 0), xmm0);
463 __ movflt(Address(rsp, wordSize), xmm1);
464 }
465 if (UseSSE >= 2) {
466 __ subptr(rsp, 4*wordSize);
467 __ movdbl(Address(rsp, 0), xmm0);
468 __ movdbl(Address(rsp, 2*wordSize), xmm1);
469 }
470 #ifdef COMPILER2
471 // C2 may leave the stack dirty if not in SSE2+ mode
472 if (UseSSE >= 2) {
473 __ verify_FPU(0, "c2i transition should have clean FPU stack");
474 } else {
475 __ empty_FPU_stack();
476 }
477 #endif /* COMPILER2 */
479 // VM needs caller's callsite
480 __ push(rax);
481 // VM needs target method
482 __ push(rbx);
483 __ verify_oop(rbx);
484 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::fixup_callers_callsite)));
485 __ addptr(rsp, 2*wordSize);
487 if (UseSSE == 1) {
488 __ movflt(xmm0, Address(rsp, 0));
489 __ movflt(xmm1, Address(rsp, wordSize));
490 __ addptr(rsp, 2*wordSize);
491 }
492 if (UseSSE >= 2) {
493 __ movdbl(xmm0, Address(rsp, 0));
494 __ movdbl(xmm1, Address(rsp, 2*wordSize));
495 __ addptr(rsp, 4*wordSize);
496 }
498 __ popf();
499 __ popa();
500 __ bind(L);
501 }
504 // Helper function to put tags in interpreter stack.
505 static void tag_stack(MacroAssembler *masm, const BasicType sig, int st_off) {
506 if (TaggedStackInterpreter) {
507 int tag_offset = st_off + Interpreter::expr_tag_offset_in_bytes(0);
508 if (sig == T_OBJECT || sig == T_ARRAY) {
509 __ movptr(Address(rsp, tag_offset), frame::TagReference);
510 } else if (sig == T_LONG || sig == T_DOUBLE) {
511 int next_tag_offset = st_off + Interpreter::expr_tag_offset_in_bytes(1);
512 __ movptr(Address(rsp, next_tag_offset), frame::TagValue);
513 __ movptr(Address(rsp, tag_offset), frame::TagValue);
514 } else {
515 __ movptr(Address(rsp, tag_offset), frame::TagValue);
516 }
517 }
518 }
520 // Double and long values with Tagged stacks are not contiguous.
521 static void move_c2i_double(MacroAssembler *masm, XMMRegister r, int st_off) {
522 int next_off = st_off - Interpreter::stackElementSize();
523 if (TaggedStackInterpreter) {
524 __ movdbl(Address(rsp, next_off), r);
525 // Move top half up and put tag in the middle.
526 __ movl(rdi, Address(rsp, next_off+wordSize));
527 __ movl(Address(rsp, st_off), rdi);
528 tag_stack(masm, T_DOUBLE, next_off);
529 } else {
530 __ movdbl(Address(rsp, next_off), r);
531 }
532 }
534 static void gen_c2i_adapter(MacroAssembler *masm,
535 int total_args_passed,
536 int comp_args_on_stack,
537 const BasicType *sig_bt,
538 const VMRegPair *regs,
539 Label& skip_fixup) {
540 // Before we get into the guts of the C2I adapter, see if we should be here
541 // at all. We've come from compiled code and are attempting to jump to the
542 // interpreter, which means the caller made a static call to get here
543 // (vcalls always get a compiled target if there is one). Check for a
544 // compiled target. If there is one, we need to patch the caller's call.
545 patch_callers_callsite(masm);
547 __ bind(skip_fixup);
549 #ifdef COMPILER2
550 // C2 may leave the stack dirty if not in SSE2+ mode
551 if (UseSSE >= 2) {
552 __ verify_FPU(0, "c2i transition should have clean FPU stack");
553 } else {
554 __ empty_FPU_stack();
555 }
556 #endif /* COMPILER2 */
558 // Since all args are passed on the stack, total_args_passed * interpreter_
559 // stack_element_size is the
560 // space we need.
561 int extraspace = total_args_passed * Interpreter::stackElementSize();
563 // Get return address
564 __ pop(rax);
566 // set senderSP value
567 __ movptr(rsi, rsp);
569 __ subptr(rsp, extraspace);
571 // Now write the args into the outgoing interpreter space
572 for (int i = 0; i < total_args_passed; i++) {
573 if (sig_bt[i] == T_VOID) {
574 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
575 continue;
576 }
578 // st_off points to lowest address on stack.
579 int st_off = ((total_args_passed - 1) - i) * Interpreter::stackElementSize();
580 int next_off = st_off - Interpreter::stackElementSize();
582 // Say 4 args:
583 // i st_off
584 // 0 12 T_LONG
585 // 1 8 T_VOID
586 // 2 4 T_OBJECT
587 // 3 0 T_BOOL
588 VMReg r_1 = regs[i].first();
589 VMReg r_2 = regs[i].second();
590 if (!r_1->is_valid()) {
591 assert(!r_2->is_valid(), "");
592 continue;
593 }
595 if (r_1->is_stack()) {
596 // memory to memory use fpu stack top
597 int ld_off = r_1->reg2stack() * VMRegImpl::stack_slot_size + extraspace;
599 if (!r_2->is_valid()) {
600 __ movl(rdi, Address(rsp, ld_off));
601 __ movptr(Address(rsp, st_off), rdi);
602 tag_stack(masm, sig_bt[i], st_off);
603 } else {
605 // ld_off == LSW, ld_off+VMRegImpl::stack_slot_size == MSW
606 // st_off == MSW, st_off-wordSize == LSW
608 __ movptr(rdi, Address(rsp, ld_off));
609 __ movptr(Address(rsp, next_off), rdi);
610 #ifndef _LP64
611 __ movptr(rdi, Address(rsp, ld_off + wordSize));
612 __ movptr(Address(rsp, st_off), rdi);
613 #else
614 #ifdef ASSERT
615 // Overwrite the unused slot with known junk
616 __ mov64(rax, CONST64(0xdeadffffdeadaaaa));
617 __ movptr(Address(rsp, st_off), rax);
618 #endif /* ASSERT */
619 #endif // _LP64
620 tag_stack(masm, sig_bt[i], next_off);
621 }
622 } else if (r_1->is_Register()) {
623 Register r = r_1->as_Register();
624 if (!r_2->is_valid()) {
625 __ movl(Address(rsp, st_off), r);
626 tag_stack(masm, sig_bt[i], st_off);
627 } else {
628 // long/double in gpr
629 NOT_LP64(ShouldNotReachHere());
630 // Two VMRegs can be T_OBJECT, T_ADDRESS, T_DOUBLE, T_LONG
631 // T_DOUBLE and T_LONG use two slots in the interpreter
632 if ( sig_bt[i] == T_LONG || sig_bt[i] == T_DOUBLE) {
633 // long/double in gpr
634 #ifdef ASSERT
635 // Overwrite the unused slot with known junk
636 LP64_ONLY(__ mov64(rax, CONST64(0xdeadffffdeadaaab)));
637 __ movptr(Address(rsp, st_off), rax);
638 #endif /* ASSERT */
639 __ movptr(Address(rsp, next_off), r);
640 tag_stack(masm, sig_bt[i], next_off);
641 } else {
642 __ movptr(Address(rsp, st_off), r);
643 tag_stack(masm, sig_bt[i], st_off);
644 }
645 }
646 } else {
647 assert(r_1->is_XMMRegister(), "");
648 if (!r_2->is_valid()) {
649 __ movflt(Address(rsp, st_off), r_1->as_XMMRegister());
650 tag_stack(masm, sig_bt[i], st_off);
651 } else {
652 assert(sig_bt[i] == T_DOUBLE || sig_bt[i] == T_LONG, "wrong type");
653 move_c2i_double(masm, r_1->as_XMMRegister(), st_off);
654 }
655 }
656 }
658 // Schedule the branch target address early.
659 __ movptr(rcx, Address(rbx, in_bytes(methodOopDesc::interpreter_entry_offset())));
660 // And repush original return address
661 __ push(rax);
662 __ jmp(rcx);
663 }
666 // For tagged stacks, double or long value aren't contiguous on the stack
667 // so get them contiguous for the xmm load
668 static void move_i2c_double(MacroAssembler *masm, XMMRegister r, Register saved_sp, int ld_off) {
669 int next_val_off = ld_off - Interpreter::stackElementSize();
670 if (TaggedStackInterpreter) {
671 // use tag slot temporarily for MSW
672 __ movptr(rsi, Address(saved_sp, ld_off));
673 __ movptr(Address(saved_sp, next_val_off+wordSize), rsi);
674 __ movdbl(r, Address(saved_sp, next_val_off));
675 // restore tag
676 __ movptr(Address(saved_sp, next_val_off+wordSize), frame::TagValue);
677 } else {
678 __ movdbl(r, Address(saved_sp, next_val_off));
679 }
680 }
682 static void gen_i2c_adapter(MacroAssembler *masm,
683 int total_args_passed,
684 int comp_args_on_stack,
685 const BasicType *sig_bt,
686 const VMRegPair *regs) {
687 // we're being called from the interpreter but need to find the
688 // compiled return entry point. The return address on the stack
689 // should point at it and we just need to pull the old value out.
690 // load up the pointer to the compiled return entry point and
691 // rewrite our return pc. The code is arranged like so:
692 //
693 // .word Interpreter::return_sentinel
694 // .word address_of_compiled_return_point
695 // return_entry_point: blah_blah_blah
696 //
697 // So we can find the appropriate return point by loading up the word
698 // just prior to the current return address we have on the stack.
699 //
700 // We will only enter here from an interpreted frame and never from after
701 // passing thru a c2i. Azul allowed this but we do not. If we lose the
702 // race and use a c2i we will remain interpreted for the race loser(s).
703 // This removes all sorts of headaches on the x86 side and also eliminates
704 // the possibility of having c2i -> i2c -> c2i -> ... endless transitions.
707 // Note: rsi contains the senderSP on entry. We must preserve it since
708 // we may do a i2c -> c2i transition if we lose a race where compiled
709 // code goes non-entrant while we get args ready.
711 // Pick up the return address
712 __ movptr(rax, Address(rsp, 0));
714 // If UseSSE >= 2 then no cleanup is needed on the return to the
715 // interpreter so skip fixing up the return entry point unless
716 // VerifyFPU is enabled.
717 if (UseSSE < 2 || VerifyFPU) {
718 Label skip, chk_int;
719 // If we were called from the call stub we need to do a little bit different
720 // cleanup than if the interpreter returned to the call stub.
722 ExternalAddress stub_return_address(StubRoutines::_call_stub_return_address);
723 __ cmpptr(rax, stub_return_address.addr());
724 __ jcc(Assembler::notEqual, chk_int);
725 assert(StubRoutines::x86::get_call_stub_compiled_return() != NULL, "must be set");
726 __ lea(rax, ExternalAddress(StubRoutines::x86::get_call_stub_compiled_return()));
727 __ jmp(skip);
729 // It must be the interpreter since we never get here via a c2i (unlike Azul)
731 __ bind(chk_int);
732 #ifdef ASSERT
733 {
734 Label ok;
735 __ cmpl(Address(rax, -2*wordSize), Interpreter::return_sentinel);
736 __ jcc(Assembler::equal, ok);
737 __ int3();
738 __ bind(ok);
739 }
740 #endif // ASSERT
741 __ movptr(rax, Address(rax, -wordSize));
742 __ bind(skip);
743 }
745 // rax, now contains the compiled return entry point which will do an
746 // cleanup needed for the return from compiled to interpreted.
748 // Must preserve original SP for loading incoming arguments because
749 // we need to align the outgoing SP for compiled code.
750 __ movptr(rdi, rsp);
752 // Cut-out for having no stack args. Since up to 2 int/oop args are passed
753 // in registers, we will occasionally have no stack args.
754 int comp_words_on_stack = 0;
755 if (comp_args_on_stack) {
756 // Sig words on the stack are greater-than VMRegImpl::stack0. Those in
757 // registers are below. By subtracting stack0, we either get a negative
758 // number (all values in registers) or the maximum stack slot accessed.
759 // int comp_args_on_stack = VMRegImpl::reg2stack(max_arg);
760 // Convert 4-byte stack slots to words.
761 comp_words_on_stack = round_to(comp_args_on_stack*4, wordSize)>>LogBytesPerWord;
762 // Round up to miminum stack alignment, in wordSize
763 comp_words_on_stack = round_to(comp_words_on_stack, 2);
764 __ subptr(rsp, comp_words_on_stack * wordSize);
765 }
767 // Align the outgoing SP
768 __ andptr(rsp, -(StackAlignmentInBytes));
770 // push the return address on the stack (note that pushing, rather
771 // than storing it, yields the correct frame alignment for the callee)
772 __ push(rax);
774 // Put saved SP in another register
775 const Register saved_sp = rax;
776 __ movptr(saved_sp, rdi);
779 // Will jump to the compiled code just as if compiled code was doing it.
780 // Pre-load the register-jump target early, to schedule it better.
781 __ movptr(rdi, Address(rbx, in_bytes(methodOopDesc::from_compiled_offset())));
783 // Now generate the shuffle code. Pick up all register args and move the
784 // rest through the floating point stack top.
785 for (int i = 0; i < total_args_passed; i++) {
786 if (sig_bt[i] == T_VOID) {
787 // Longs and doubles are passed in native word order, but misaligned
788 // in the 32-bit build.
789 assert(i > 0 && (sig_bt[i-1] == T_LONG || sig_bt[i-1] == T_DOUBLE), "missing half");
790 continue;
791 }
793 // Pick up 0, 1 or 2 words from SP+offset.
795 assert(!regs[i].second()->is_valid() || regs[i].first()->next() == regs[i].second(),
796 "scrambled load targets?");
797 // Load in argument order going down.
798 int ld_off = (total_args_passed - i)*Interpreter::stackElementSize() + Interpreter::value_offset_in_bytes();
799 // Point to interpreter value (vs. tag)
800 int next_off = ld_off - Interpreter::stackElementSize();
801 //
802 //
803 //
804 VMReg r_1 = regs[i].first();
805 VMReg r_2 = regs[i].second();
806 if (!r_1->is_valid()) {
807 assert(!r_2->is_valid(), "");
808 continue;
809 }
810 if (r_1->is_stack()) {
811 // Convert stack slot to an SP offset (+ wordSize to account for return address )
812 int st_off = regs[i].first()->reg2stack()*VMRegImpl::stack_slot_size + wordSize;
814 // We can use rsi as a temp here because compiled code doesn't need rsi as an input
815 // and if we end up going thru a c2i because of a miss a reasonable value of rsi
816 // we be generated.
817 if (!r_2->is_valid()) {
818 // __ fld_s(Address(saved_sp, ld_off));
819 // __ fstp_s(Address(rsp, st_off));
820 __ movl(rsi, Address(saved_sp, ld_off));
821 __ movptr(Address(rsp, st_off), rsi);
822 } else {
823 // Interpreter local[n] == MSW, local[n+1] == LSW however locals
824 // are accessed as negative so LSW is at LOW address
826 // ld_off is MSW so get LSW
827 // st_off is LSW (i.e. reg.first())
828 // __ fld_d(Address(saved_sp, next_off));
829 // __ fstp_d(Address(rsp, st_off));
830 //
831 // We are using two VMRegs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE
832 // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case
833 // So we must adjust where to pick up the data to match the interpreter.
834 //
835 // Interpreter local[n] == MSW, local[n+1] == LSW however locals
836 // are accessed as negative so LSW is at LOW address
838 // ld_off is MSW so get LSW
839 const int offset = (NOT_LP64(true ||) sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)?
840 next_off : ld_off;
841 __ movptr(rsi, Address(saved_sp, offset));
842 __ movptr(Address(rsp, st_off), rsi);
843 #ifndef _LP64
844 __ movptr(rsi, Address(saved_sp, ld_off));
845 __ movptr(Address(rsp, st_off + wordSize), rsi);
846 #endif // _LP64
847 }
848 } else if (r_1->is_Register()) { // Register argument
849 Register r = r_1->as_Register();
850 assert(r != rax, "must be different");
851 if (r_2->is_valid()) {
852 //
853 // We are using two VMRegs. This can be either T_OBJECT, T_ADDRESS, T_LONG, or T_DOUBLE
854 // the interpreter allocates two slots but only uses one for thr T_LONG or T_DOUBLE case
855 // So we must adjust where to pick up the data to match the interpreter.
857 const int offset = (NOT_LP64(true ||) sig_bt[i]==T_LONG||sig_bt[i]==T_DOUBLE)?
858 next_off : ld_off;
860 // this can be a misaligned move
861 __ movptr(r, Address(saved_sp, offset));
862 #ifndef _LP64
863 assert(r_2->as_Register() != rax, "need another temporary register");
864 // Remember r_1 is low address (and LSB on x86)
865 // So r_2 gets loaded from high address regardless of the platform
866 __ movptr(r_2->as_Register(), Address(saved_sp, ld_off));
867 #endif // _LP64
868 } else {
869 __ movl(r, Address(saved_sp, ld_off));
870 }
871 } else {
872 assert(r_1->is_XMMRegister(), "");
873 if (!r_2->is_valid()) {
874 __ movflt(r_1->as_XMMRegister(), Address(saved_sp, ld_off));
875 } else {
876 move_i2c_double(masm, r_1->as_XMMRegister(), saved_sp, ld_off);
877 }
878 }
879 }
881 // 6243940 We might end up in handle_wrong_method if
882 // the callee is deoptimized as we race thru here. If that
883 // happens we don't want to take a safepoint because the
884 // caller frame will look interpreted and arguments are now
885 // "compiled" so it is much better to make this transition
886 // invisible to the stack walking code. Unfortunately if
887 // we try and find the callee by normal means a safepoint
888 // is possible. So we stash the desired callee in the thread
889 // and the vm will find there should this case occur.
891 __ get_thread(rax);
892 __ movptr(Address(rax, JavaThread::callee_target_offset()), rbx);
894 // move methodOop to rax, in case we end up in an c2i adapter.
895 // the c2i adapters expect methodOop in rax, (c2) because c2's
896 // resolve stubs return the result (the method) in rax,.
897 // I'd love to fix this.
898 __ mov(rax, rbx);
900 __ jmp(rdi);
901 }
903 // ---------------------------------------------------------------
904 AdapterHandlerEntry* SharedRuntime::generate_i2c2i_adapters(MacroAssembler *masm,
905 int total_args_passed,
906 int comp_args_on_stack,
907 const BasicType *sig_bt,
908 const VMRegPair *regs) {
909 address i2c_entry = __ pc();
911 gen_i2c_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs);
913 // -------------------------------------------------------------------------
914 // Generate a C2I adapter. On entry we know rbx, holds the methodOop during calls
915 // to the interpreter. The args start out packed in the compiled layout. They
916 // need to be unpacked into the interpreter layout. This will almost always
917 // require some stack space. We grow the current (compiled) stack, then repack
918 // the args. We finally end in a jump to the generic interpreter entry point.
919 // On exit from the interpreter, the interpreter will restore our SP (lest the
920 // compiled code, which relys solely on SP and not EBP, get sick).
922 address c2i_unverified_entry = __ pc();
923 Label skip_fixup;
925 Register holder = rax;
926 Register receiver = rcx;
927 Register temp = rbx;
929 {
931 Label missed;
933 __ verify_oop(holder);
934 __ movptr(temp, Address(receiver, oopDesc::klass_offset_in_bytes()));
935 __ verify_oop(temp);
937 __ cmpptr(temp, Address(holder, compiledICHolderOopDesc::holder_klass_offset()));
938 __ movptr(rbx, Address(holder, compiledICHolderOopDesc::holder_method_offset()));
939 __ jcc(Assembler::notEqual, missed);
940 // Method might have been compiled since the call site was patched to
941 // interpreted if that is the case treat it as a miss so we can get
942 // the call site corrected.
943 __ cmpptr(Address(rbx, in_bytes(methodOopDesc::code_offset())), (int32_t)NULL_WORD);
944 __ jcc(Assembler::equal, skip_fixup);
946 __ bind(missed);
947 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
948 }
950 address c2i_entry = __ pc();
952 gen_c2i_adapter(masm, total_args_passed, comp_args_on_stack, sig_bt, regs, skip_fixup);
954 __ flush();
955 return new AdapterHandlerEntry(i2c_entry, c2i_entry, c2i_unverified_entry);
956 }
958 int SharedRuntime::c_calling_convention(const BasicType *sig_bt,
959 VMRegPair *regs,
960 int total_args_passed) {
961 // We return the amount of VMRegImpl stack slots we need to reserve for all
962 // the arguments NOT counting out_preserve_stack_slots.
964 uint stack = 0; // All arguments on stack
966 for( int i = 0; i < total_args_passed; i++) {
967 // From the type and the argument number (count) compute the location
968 switch( sig_bt[i] ) {
969 case T_BOOLEAN:
970 case T_CHAR:
971 case T_FLOAT:
972 case T_BYTE:
973 case T_SHORT:
974 case T_INT:
975 case T_OBJECT:
976 case T_ARRAY:
977 case T_ADDRESS:
978 regs[i].set1(VMRegImpl::stack2reg(stack++));
979 break;
980 case T_LONG:
981 case T_DOUBLE: // The stack numbering is reversed from Java
982 // Since C arguments do not get reversed, the ordering for
983 // doubles on the stack must be opposite the Java convention
984 assert(sig_bt[i+1] == T_VOID, "missing Half" );
985 regs[i].set2(VMRegImpl::stack2reg(stack));
986 stack += 2;
987 break;
988 case T_VOID: regs[i].set_bad(); break;
989 default:
990 ShouldNotReachHere();
991 break;
992 }
993 }
994 return stack;
995 }
997 // A simple move of integer like type
998 static void simple_move32(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
999 if (src.first()->is_stack()) {
1000 if (dst.first()->is_stack()) {
1001 // stack to stack
1002 // __ ld(FP, reg2offset(src.first()) + STACK_BIAS, L5);
1003 // __ st(L5, SP, reg2offset(dst.first()) + STACK_BIAS);
1004 __ movl2ptr(rax, Address(rbp, reg2offset_in(src.first())));
1005 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax);
1006 } else {
1007 // stack to reg
1008 __ movl2ptr(dst.first()->as_Register(), Address(rbp, reg2offset_in(src.first())));
1009 }
1010 } else if (dst.first()->is_stack()) {
1011 // reg to stack
1012 // no need to sign extend on 64bit
1013 __ movptr(Address(rsp, reg2offset_out(dst.first())), src.first()->as_Register());
1014 } else {
1015 if (dst.first() != src.first()) {
1016 __ mov(dst.first()->as_Register(), src.first()->as_Register());
1017 }
1018 }
1019 }
1021 // An oop arg. Must pass a handle not the oop itself
1022 static void object_move(MacroAssembler* masm,
1023 OopMap* map,
1024 int oop_handle_offset,
1025 int framesize_in_slots,
1026 VMRegPair src,
1027 VMRegPair dst,
1028 bool is_receiver,
1029 int* receiver_offset) {
1031 // Because of the calling conventions we know that src can be a
1032 // register or a stack location. dst can only be a stack location.
1034 assert(dst.first()->is_stack(), "must be stack");
1035 // must pass a handle. First figure out the location we use as a handle
1037 if (src.first()->is_stack()) {
1038 // Oop is already on the stack as an argument
1039 Register rHandle = rax;
1040 Label nil;
1041 __ xorptr(rHandle, rHandle);
1042 __ cmpptr(Address(rbp, reg2offset_in(src.first())), (int32_t)NULL_WORD);
1043 __ jcc(Assembler::equal, nil);
1044 __ lea(rHandle, Address(rbp, reg2offset_in(src.first())));
1045 __ bind(nil);
1046 __ movptr(Address(rsp, reg2offset_out(dst.first())), rHandle);
1048 int offset_in_older_frame = src.first()->reg2stack() + SharedRuntime::out_preserve_stack_slots();
1049 map->set_oop(VMRegImpl::stack2reg(offset_in_older_frame + framesize_in_slots));
1050 if (is_receiver) {
1051 *receiver_offset = (offset_in_older_frame + framesize_in_slots) * VMRegImpl::stack_slot_size;
1052 }
1053 } else {
1054 // Oop is in an a register we must store it to the space we reserve
1055 // on the stack for oop_handles
1056 const Register rOop = src.first()->as_Register();
1057 const Register rHandle = rax;
1058 int oop_slot = (rOop == rcx ? 0 : 1) * VMRegImpl::slots_per_word + oop_handle_offset;
1059 int offset = oop_slot*VMRegImpl::stack_slot_size;
1060 Label skip;
1061 __ movptr(Address(rsp, offset), rOop);
1062 map->set_oop(VMRegImpl::stack2reg(oop_slot));
1063 __ xorptr(rHandle, rHandle);
1064 __ cmpptr(rOop, (int32_t)NULL_WORD);
1065 __ jcc(Assembler::equal, skip);
1066 __ lea(rHandle, Address(rsp, offset));
1067 __ bind(skip);
1068 // Store the handle parameter
1069 __ movptr(Address(rsp, reg2offset_out(dst.first())), rHandle);
1070 if (is_receiver) {
1071 *receiver_offset = offset;
1072 }
1073 }
1074 }
1076 // A float arg may have to do float reg int reg conversion
1077 static void float_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1078 assert(!src.second()->is_valid() && !dst.second()->is_valid(), "bad float_move");
1080 // Because of the calling convention we know that src is either a stack location
1081 // or an xmm register. dst can only be a stack location.
1083 assert(dst.first()->is_stack() && ( src.first()->is_stack() || src.first()->is_XMMRegister()), "bad parameters");
1085 if (src.first()->is_stack()) {
1086 __ movl(rax, Address(rbp, reg2offset_in(src.first())));
1087 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax);
1088 } else {
1089 // reg to stack
1090 __ movflt(Address(rsp, reg2offset_out(dst.first())), src.first()->as_XMMRegister());
1091 }
1092 }
1094 // A long move
1095 static void long_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1097 // The only legal possibility for a long_move VMRegPair is:
1098 // 1: two stack slots (possibly unaligned)
1099 // as neither the java or C calling convention will use registers
1100 // for longs.
1102 if (src.first()->is_stack() && dst.first()->is_stack()) {
1103 assert(src.second()->is_stack() && dst.second()->is_stack(), "must be all stack");
1104 __ movptr(rax, Address(rbp, reg2offset_in(src.first())));
1105 NOT_LP64(__ movptr(rbx, Address(rbp, reg2offset_in(src.second()))));
1106 __ movptr(Address(rsp, reg2offset_out(dst.first())), rax);
1107 NOT_LP64(__ movptr(Address(rsp, reg2offset_out(dst.second())), rbx));
1108 } else {
1109 ShouldNotReachHere();
1110 }
1111 }
1113 // A double move
1114 static void double_move(MacroAssembler* masm, VMRegPair src, VMRegPair dst) {
1116 // The only legal possibilities for a double_move VMRegPair are:
1117 // The painful thing here is that like long_move a VMRegPair might be
1119 // Because of the calling convention we know that src is either
1120 // 1: a single physical register (xmm registers only)
1121 // 2: two stack slots (possibly unaligned)
1122 // dst can only be a pair of stack slots.
1124 assert(dst.first()->is_stack() && (src.first()->is_XMMRegister() || src.first()->is_stack()), "bad args");
1126 if (src.first()->is_stack()) {
1127 // source is 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 // reg to stack
1134 // No worries about stack alignment
1135 __ movdbl(Address(rsp, reg2offset_out(dst.first())), src.first()->as_XMMRegister());
1136 }
1137 }
1140 void SharedRuntime::save_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1141 // We always ignore the frame_slots arg and just use the space just below frame pointer
1142 // which by this time is free to use
1143 switch (ret_type) {
1144 case T_FLOAT:
1145 __ fstp_s(Address(rbp, -wordSize));
1146 break;
1147 case T_DOUBLE:
1148 __ fstp_d(Address(rbp, -2*wordSize));
1149 break;
1150 case T_VOID: break;
1151 case T_LONG:
1152 __ movptr(Address(rbp, -wordSize), rax);
1153 NOT_LP64(__ movptr(Address(rbp, -2*wordSize), rdx));
1154 break;
1155 default: {
1156 __ movptr(Address(rbp, -wordSize), rax);
1157 }
1158 }
1159 }
1161 void SharedRuntime::restore_native_result(MacroAssembler *masm, BasicType ret_type, int frame_slots) {
1162 // We always ignore the frame_slots arg and just use the space just below frame pointer
1163 // which by this time is free to use
1164 switch (ret_type) {
1165 case T_FLOAT:
1166 __ fld_s(Address(rbp, -wordSize));
1167 break;
1168 case T_DOUBLE:
1169 __ fld_d(Address(rbp, -2*wordSize));
1170 break;
1171 case T_LONG:
1172 __ movptr(rax, Address(rbp, -wordSize));
1173 NOT_LP64(__ movptr(rdx, Address(rbp, -2*wordSize)));
1174 break;
1175 case T_VOID: break;
1176 default: {
1177 __ movptr(rax, Address(rbp, -wordSize));
1178 }
1179 }
1180 }
1182 // ---------------------------------------------------------------------------
1183 // Generate a native wrapper for a given method. The method takes arguments
1184 // in the Java compiled code convention, marshals them to the native
1185 // convention (handlizes oops, etc), transitions to native, makes the call,
1186 // returns to java state (possibly blocking), unhandlizes any result and
1187 // returns.
1188 nmethod *SharedRuntime::generate_native_wrapper(MacroAssembler *masm,
1189 methodHandle method,
1190 int total_in_args,
1191 int comp_args_on_stack,
1192 BasicType *in_sig_bt,
1193 VMRegPair *in_regs,
1194 BasicType ret_type) {
1196 // An OopMap for lock (and class if static)
1197 OopMapSet *oop_maps = new OopMapSet();
1199 // We have received a description of where all the java arg are located
1200 // on entry to the wrapper. We need to convert these args to where
1201 // the jni function will expect them. To figure out where they go
1202 // we convert the java signature to a C signature by inserting
1203 // the hidden arguments as arg[0] and possibly arg[1] (static method)
1205 int total_c_args = total_in_args + 1;
1206 if (method->is_static()) {
1207 total_c_args++;
1208 }
1210 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_c_args);
1211 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_c_args);
1213 int argc = 0;
1214 out_sig_bt[argc++] = T_ADDRESS;
1215 if (method->is_static()) {
1216 out_sig_bt[argc++] = T_OBJECT;
1217 }
1219 int i;
1220 for (i = 0; i < total_in_args ; i++ ) {
1221 out_sig_bt[argc++] = in_sig_bt[i];
1222 }
1225 // Now figure out where the args must be stored and how much stack space
1226 // they require (neglecting out_preserve_stack_slots but space for storing
1227 // the 1st six register arguments). It's weird see int_stk_helper.
1228 //
1229 int out_arg_slots;
1230 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
1232 // Compute framesize for the wrapper. We need to handlize all oops in
1233 // registers a max of 2 on x86.
1235 // Calculate the total number of stack slots we will need.
1237 // First count the abi requirement plus all of the outgoing args
1238 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
1240 // Now the space for the inbound oop handle area
1242 int oop_handle_offset = stack_slots;
1243 stack_slots += 2*VMRegImpl::slots_per_word;
1245 // Now any space we need for handlizing a klass if static method
1247 int klass_slot_offset = 0;
1248 int klass_offset = -1;
1249 int lock_slot_offset = 0;
1250 bool is_static = false;
1251 int oop_temp_slot_offset = 0;
1253 if (method->is_static()) {
1254 klass_slot_offset = stack_slots;
1255 stack_slots += VMRegImpl::slots_per_word;
1256 klass_offset = klass_slot_offset * VMRegImpl::stack_slot_size;
1257 is_static = true;
1258 }
1260 // Plus a lock if needed
1262 if (method->is_synchronized()) {
1263 lock_slot_offset = stack_slots;
1264 stack_slots += VMRegImpl::slots_per_word;
1265 }
1267 // Now a place (+2) to save return values or temp during shuffling
1268 // + 2 for return address (which we own) and saved rbp,
1269 stack_slots += 4;
1271 // Ok The space we have allocated will look like:
1272 //
1273 //
1274 // FP-> | |
1275 // |---------------------|
1276 // | 2 slots for moves |
1277 // |---------------------|
1278 // | lock box (if sync) |
1279 // |---------------------| <- lock_slot_offset (-lock_slot_rbp_offset)
1280 // | klass (if static) |
1281 // |---------------------| <- klass_slot_offset
1282 // | oopHandle area |
1283 // |---------------------| <- oop_handle_offset (a max of 2 registers)
1284 // | outbound memory |
1285 // | based arguments |
1286 // | |
1287 // |---------------------|
1288 // | |
1289 // SP-> | out_preserved_slots |
1290 //
1291 //
1292 // ****************************************************************************
1293 // WARNING - on Windows Java Natives use pascal calling convention and pop the
1294 // arguments off of the stack after the jni call. Before the call we can use
1295 // instructions that are SP relative. After the jni call we switch to FP
1296 // relative instructions instead of re-adjusting the stack on windows.
1297 // ****************************************************************************
1300 // Now compute actual number of stack words we need rounding to make
1301 // stack properly aligned.
1302 stack_slots = round_to(stack_slots, 2 * VMRegImpl::slots_per_word);
1304 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
1306 intptr_t start = (intptr_t)__ pc();
1308 // First thing make an ic check to see if we should even be here
1310 // We are free to use all registers as temps without saving them and
1311 // restoring them except rbp,. rbp, is the only callee save register
1312 // as far as the interpreter and the compiler(s) are concerned.
1315 const Register ic_reg = rax;
1316 const Register receiver = rcx;
1317 Label hit;
1318 Label exception_pending;
1321 __ verify_oop(receiver);
1322 __ cmpptr(ic_reg, Address(receiver, oopDesc::klass_offset_in_bytes()));
1323 __ jcc(Assembler::equal, hit);
1325 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
1327 // verified entry must be aligned for code patching.
1328 // and the first 5 bytes must be in the same cache line
1329 // if we align at 8 then we will be sure 5 bytes are in the same line
1330 __ align(8);
1332 __ bind(hit);
1334 int vep_offset = ((intptr_t)__ pc()) - start;
1336 #ifdef COMPILER1
1337 if (InlineObjectHash && method->intrinsic_id() == vmIntrinsics::_hashCode) {
1338 // Object.hashCode can pull the hashCode from the header word
1339 // instead of doing a full VM transition once it's been computed.
1340 // Since hashCode is usually polymorphic at call sites we can't do
1341 // this optimization at the call site without a lot of work.
1342 Label slowCase;
1343 Register receiver = rcx;
1344 Register result = rax;
1345 __ movptr(result, Address(receiver, oopDesc::mark_offset_in_bytes()));
1347 // check if locked
1348 __ testptr(result, markOopDesc::unlocked_value);
1349 __ jcc (Assembler::zero, slowCase);
1351 if (UseBiasedLocking) {
1352 // Check if biased and fall through to runtime if so
1353 __ testptr(result, markOopDesc::biased_lock_bit_in_place);
1354 __ jcc (Assembler::notZero, slowCase);
1355 }
1357 // get hash
1358 __ andptr(result, markOopDesc::hash_mask_in_place);
1359 // test if hashCode exists
1360 __ jcc (Assembler::zero, slowCase);
1361 __ shrptr(result, markOopDesc::hash_shift);
1362 __ ret(0);
1363 __ bind (slowCase);
1364 }
1365 #endif // COMPILER1
1367 // The instruction at the verified entry point must be 5 bytes or longer
1368 // because it can be patched on the fly by make_non_entrant. The stack bang
1369 // instruction fits that requirement.
1371 // Generate stack overflow check
1373 if (UseStackBanging) {
1374 __ bang_stack_with_offset(StackShadowPages*os::vm_page_size());
1375 } else {
1376 // need a 5 byte instruction to allow MT safe patching to non-entrant
1377 __ fat_nop();
1378 }
1380 // Generate a new frame for the wrapper.
1381 __ enter();
1382 // -2 because return address is already present and so is saved rbp,
1383 __ subptr(rsp, stack_size - 2*wordSize);
1385 // Frame is now completed as far a size and linkage.
1387 int frame_complete = ((intptr_t)__ pc()) - start;
1389 // Calculate the difference between rsp and rbp,. We need to know it
1390 // after the native call because on windows Java Natives will pop
1391 // the arguments and it is painful to do rsp relative addressing
1392 // in a platform independent way. So after the call we switch to
1393 // rbp, relative addressing.
1395 int fp_adjustment = stack_size - 2*wordSize;
1397 #ifdef COMPILER2
1398 // C2 may leave the stack dirty if not in SSE2+ mode
1399 if (UseSSE >= 2) {
1400 __ verify_FPU(0, "c2i transition should have clean FPU stack");
1401 } else {
1402 __ empty_FPU_stack();
1403 }
1404 #endif /* COMPILER2 */
1406 // Compute the rbp, offset for any slots used after the jni call
1408 int lock_slot_rbp_offset = (lock_slot_offset*VMRegImpl::stack_slot_size) - fp_adjustment;
1409 int oop_temp_slot_rbp_offset = (oop_temp_slot_offset*VMRegImpl::stack_slot_size) - fp_adjustment;
1411 // We use rdi as a thread pointer because it is callee save and
1412 // if we load it once it is usable thru the entire wrapper
1413 const Register thread = rdi;
1415 // We use rsi as the oop handle for the receiver/klass
1416 // It is callee save so it survives the call to native
1418 const Register oop_handle_reg = rsi;
1420 __ get_thread(thread);
1423 //
1424 // We immediately shuffle the arguments so that any vm call we have to
1425 // make from here on out (sync slow path, jvmti, etc.) we will have
1426 // captured the oops from our caller and have a valid oopMap for
1427 // them.
1429 // -----------------
1430 // The Grand Shuffle
1431 //
1432 // Natives require 1 or 2 extra arguments over the normal ones: the JNIEnv*
1433 // and, if static, the class mirror instead of a receiver. This pretty much
1434 // guarantees that register layout will not match (and x86 doesn't use reg
1435 // parms though amd does). Since the native abi doesn't use register args
1436 // and the java conventions does we don't have to worry about collisions.
1437 // All of our moved are reg->stack or stack->stack.
1438 // We ignore the extra arguments during the shuffle and handle them at the
1439 // last moment. The shuffle is described by the two calling convention
1440 // vectors we have in our possession. We simply walk the java vector to
1441 // get the source locations and the c vector to get the destinations.
1443 int c_arg = method->is_static() ? 2 : 1 ;
1445 // Record rsp-based slot for receiver on stack for non-static methods
1446 int receiver_offset = -1;
1448 // This is a trick. We double the stack slots so we can claim
1449 // the oops in the caller's frame. Since we are sure to have
1450 // more args than the caller doubling is enough to make
1451 // sure we can capture all the incoming oop args from the
1452 // caller.
1453 //
1454 OopMap* map = new OopMap(stack_slots * 2, 0 /* arg_slots*/);
1456 // Mark location of rbp,
1457 // map->set_callee_saved(VMRegImpl::stack2reg( stack_slots - 2), stack_slots * 2, 0, rbp->as_VMReg());
1459 // We know that we only have args in at most two integer registers (rcx, rdx). So rax, rbx
1460 // Are free to temporaries if we have to do stack to steck moves.
1461 // All inbound args are referenced based on rbp, and all outbound args via rsp.
1463 for (i = 0; i < total_in_args ; i++, c_arg++ ) {
1464 switch (in_sig_bt[i]) {
1465 case T_ARRAY:
1466 case T_OBJECT:
1467 object_move(masm, map, oop_handle_offset, stack_slots, in_regs[i], out_regs[c_arg],
1468 ((i == 0) && (!is_static)),
1469 &receiver_offset);
1470 break;
1471 case T_VOID:
1472 break;
1474 case T_FLOAT:
1475 float_move(masm, in_regs[i], out_regs[c_arg]);
1476 break;
1478 case T_DOUBLE:
1479 assert( i + 1 < total_in_args &&
1480 in_sig_bt[i + 1] == T_VOID &&
1481 out_sig_bt[c_arg+1] == T_VOID, "bad arg list");
1482 double_move(masm, in_regs[i], out_regs[c_arg]);
1483 break;
1485 case T_LONG :
1486 long_move(masm, in_regs[i], out_regs[c_arg]);
1487 break;
1489 case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
1491 default:
1492 simple_move32(masm, in_regs[i], out_regs[c_arg]);
1493 }
1494 }
1496 // Pre-load a static method's oop into rsi. Used both by locking code and
1497 // the normal JNI call code.
1498 if (method->is_static()) {
1500 // load opp into a register
1501 __ movoop(oop_handle_reg, JNIHandles::make_local(Klass::cast(method->method_holder())->java_mirror()));
1503 // Now handlize the static class mirror it's known not-null.
1504 __ movptr(Address(rsp, klass_offset), oop_handle_reg);
1505 map->set_oop(VMRegImpl::stack2reg(klass_slot_offset));
1507 // Now get the handle
1508 __ lea(oop_handle_reg, Address(rsp, klass_offset));
1509 // store the klass handle as second argument
1510 __ movptr(Address(rsp, wordSize), oop_handle_reg);
1511 }
1513 // Change state to native (we save the return address in the thread, since it might not
1514 // be pushed on the stack when we do a a stack traversal). It is enough that the pc()
1515 // points into the right code segment. It does not have to be the correct return pc.
1516 // We use the same pc/oopMap repeatedly when we call out
1518 intptr_t the_pc = (intptr_t) __ pc();
1519 oop_maps->add_gc_map(the_pc - start, map);
1521 __ set_last_Java_frame(thread, rsp, noreg, (address)the_pc);
1524 // We have all of the arguments setup at this point. We must not touch any register
1525 // argument registers at this point (what if we save/restore them there are no oop?
1527 {
1528 SkipIfEqual skip_if(masm, &DTraceMethodProbes, 0);
1529 __ movoop(rax, JNIHandles::make_local(method()));
1530 __ call_VM_leaf(
1531 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry),
1532 thread, rax);
1533 }
1536 // These are register definitions we need for locking/unlocking
1537 const Register swap_reg = rax; // Must use rax, for cmpxchg instruction
1538 const Register obj_reg = rcx; // Will contain the oop
1539 const Register lock_reg = rdx; // Address of compiler lock object (BasicLock)
1541 Label slow_path_lock;
1542 Label lock_done;
1544 // Lock a synchronized method
1545 if (method->is_synchronized()) {
1548 const int mark_word_offset = BasicLock::displaced_header_offset_in_bytes();
1550 // Get the handle (the 2nd argument)
1551 __ movptr(oop_handle_reg, Address(rsp, wordSize));
1553 // Get address of the box
1555 __ lea(lock_reg, Address(rbp, lock_slot_rbp_offset));
1557 // Load the oop from the handle
1558 __ movptr(obj_reg, Address(oop_handle_reg, 0));
1560 if (UseBiasedLocking) {
1561 // Note that oop_handle_reg is trashed during this call
1562 __ biased_locking_enter(lock_reg, obj_reg, swap_reg, oop_handle_reg, false, lock_done, &slow_path_lock);
1563 }
1565 // Load immediate 1 into swap_reg %rax,
1566 __ movptr(swap_reg, 1);
1568 // Load (object->mark() | 1) into swap_reg %rax,
1569 __ orptr(swap_reg, Address(obj_reg, 0));
1571 // Save (object->mark() | 1) into BasicLock's displaced header
1572 __ movptr(Address(lock_reg, mark_word_offset), swap_reg);
1574 if (os::is_MP()) {
1575 __ lock();
1576 }
1578 // src -> dest iff dest == rax, else rax, <- dest
1579 // *obj_reg = lock_reg iff *obj_reg == rax, else rax, = *(obj_reg)
1580 __ cmpxchgptr(lock_reg, Address(obj_reg, 0));
1581 __ jcc(Assembler::equal, lock_done);
1583 // Test if the oopMark is an obvious stack pointer, i.e.,
1584 // 1) (mark & 3) == 0, and
1585 // 2) rsp <= mark < mark + os::pagesize()
1586 // These 3 tests can be done by evaluating the following
1587 // expression: ((mark - rsp) & (3 - os::vm_page_size())),
1588 // assuming both stack pointer and pagesize have their
1589 // least significant 2 bits clear.
1590 // NOTE: the oopMark is in swap_reg %rax, as the result of cmpxchg
1592 __ subptr(swap_reg, rsp);
1593 __ andptr(swap_reg, 3 - os::vm_page_size());
1595 // Save the test result, for recursive case, the result is zero
1596 __ movptr(Address(lock_reg, mark_word_offset), swap_reg);
1597 __ jcc(Assembler::notEqual, slow_path_lock);
1598 // Slow path will re-enter here
1599 __ bind(lock_done);
1601 if (UseBiasedLocking) {
1602 // Re-fetch oop_handle_reg as we trashed it above
1603 __ movptr(oop_handle_reg, Address(rsp, wordSize));
1604 }
1605 }
1608 // Finally just about ready to make the JNI call
1611 // get JNIEnv* which is first argument to native
1613 __ lea(rdx, Address(thread, in_bytes(JavaThread::jni_environment_offset())));
1614 __ movptr(Address(rsp, 0), rdx);
1616 // Now set thread in native
1617 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_native);
1619 __ call(RuntimeAddress(method->native_function()));
1621 // WARNING - on Windows Java Natives use pascal calling convention and pop the
1622 // arguments off of the stack. We could just re-adjust the stack pointer here
1623 // and continue to do SP relative addressing but we instead switch to FP
1624 // relative addressing.
1626 // Unpack native results.
1627 switch (ret_type) {
1628 case T_BOOLEAN: __ c2bool(rax); break;
1629 case T_CHAR : __ andptr(rax, 0xFFFF); break;
1630 case T_BYTE : __ sign_extend_byte (rax); break;
1631 case T_SHORT : __ sign_extend_short(rax); break;
1632 case T_INT : /* nothing to do */ break;
1633 case T_DOUBLE :
1634 case T_FLOAT :
1635 // Result is in st0 we'll save as needed
1636 break;
1637 case T_ARRAY: // Really a handle
1638 case T_OBJECT: // Really a handle
1639 break; // can't de-handlize until after safepoint check
1640 case T_VOID: break;
1641 case T_LONG: break;
1642 default : ShouldNotReachHere();
1643 }
1645 // Switch thread to "native transition" state before reading the synchronization state.
1646 // This additional state is necessary because reading and testing the synchronization
1647 // state is not atomic w.r.t. GC, as this scenario demonstrates:
1648 // Java thread A, in _thread_in_native state, loads _not_synchronized and is preempted.
1649 // VM thread changes sync state to synchronizing and suspends threads for GC.
1650 // Thread A is resumed to finish this native method, but doesn't block here since it
1651 // didn't see any synchronization is progress, and escapes.
1652 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_native_trans);
1654 if(os::is_MP()) {
1655 if (UseMembar) {
1656 // Force this write out before the read below
1657 __ membar(Assembler::Membar_mask_bits(
1658 Assembler::LoadLoad | Assembler::LoadStore |
1659 Assembler::StoreLoad | Assembler::StoreStore));
1660 } else {
1661 // Write serialization page so VM thread can do a pseudo remote membar.
1662 // We use the current thread pointer to calculate a thread specific
1663 // offset to write to within the page. This minimizes bus traffic
1664 // due to cache line collision.
1665 __ serialize_memory(thread, rcx);
1666 }
1667 }
1669 if (AlwaysRestoreFPU) {
1670 // Make sure the control word is correct.
1671 __ fldcw(ExternalAddress(StubRoutines::addr_fpu_cntrl_wrd_std()));
1672 }
1674 // check for safepoint operation in progress and/or pending suspend requests
1675 { Label Continue;
1677 __ cmp32(ExternalAddress((address)SafepointSynchronize::address_of_state()),
1678 SafepointSynchronize::_not_synchronized);
1680 Label L;
1681 __ jcc(Assembler::notEqual, L);
1682 __ cmpl(Address(thread, JavaThread::suspend_flags_offset()), 0);
1683 __ jcc(Assembler::equal, Continue);
1684 __ bind(L);
1686 // Don't use call_VM as it will see a possible pending exception and forward it
1687 // and never return here preventing us from clearing _last_native_pc down below.
1688 // Also can't use call_VM_leaf either as it will check to see if rsi & rdi are
1689 // preserved and correspond to the bcp/locals pointers. So we do a runtime call
1690 // by hand.
1691 //
1692 save_native_result(masm, ret_type, stack_slots);
1693 __ push(thread);
1694 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address,
1695 JavaThread::check_special_condition_for_native_trans)));
1696 __ increment(rsp, wordSize);
1697 // Restore any method result value
1698 restore_native_result(masm, ret_type, stack_slots);
1700 __ bind(Continue);
1701 }
1703 // change thread state
1704 __ movl(Address(thread, JavaThread::thread_state_offset()), _thread_in_Java);
1706 Label reguard;
1707 Label reguard_done;
1708 __ cmpl(Address(thread, JavaThread::stack_guard_state_offset()), JavaThread::stack_guard_yellow_disabled);
1709 __ jcc(Assembler::equal, reguard);
1711 // slow path reguard re-enters here
1712 __ bind(reguard_done);
1714 // Handle possible exception (will unlock if necessary)
1716 // native result if any is live
1718 // Unlock
1719 Label slow_path_unlock;
1720 Label unlock_done;
1721 if (method->is_synchronized()) {
1723 Label done;
1725 // Get locked oop from the handle we passed to jni
1726 __ movptr(obj_reg, Address(oop_handle_reg, 0));
1728 if (UseBiasedLocking) {
1729 __ biased_locking_exit(obj_reg, rbx, done);
1730 }
1732 // Simple recursive lock?
1734 __ cmpptr(Address(rbp, lock_slot_rbp_offset), (int32_t)NULL_WORD);
1735 __ jcc(Assembler::equal, done);
1737 // Must save rax, if if it is live now because cmpxchg must use it
1738 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1739 save_native_result(masm, ret_type, stack_slots);
1740 }
1742 // get old displaced header
1743 __ movptr(rbx, Address(rbp, lock_slot_rbp_offset));
1745 // get address of the stack lock
1746 __ lea(rax, Address(rbp, lock_slot_rbp_offset));
1748 // Atomic swap old header if oop still contains the stack lock
1749 if (os::is_MP()) {
1750 __ lock();
1751 }
1753 // src -> dest iff dest == rax, else rax, <- dest
1754 // *obj_reg = rbx, iff *obj_reg == rax, else rax, = *(obj_reg)
1755 __ cmpxchgptr(rbx, Address(obj_reg, 0));
1756 __ jcc(Assembler::notEqual, slow_path_unlock);
1758 // slow path re-enters here
1759 __ bind(unlock_done);
1760 if (ret_type != T_FLOAT && ret_type != T_DOUBLE && ret_type != T_VOID) {
1761 restore_native_result(masm, ret_type, stack_slots);
1762 }
1764 __ bind(done);
1766 }
1768 {
1769 SkipIfEqual skip_if(masm, &DTraceMethodProbes, 0);
1770 // Tell dtrace about this method exit
1771 save_native_result(masm, ret_type, stack_slots);
1772 __ movoop(rax, JNIHandles::make_local(method()));
1773 __ call_VM_leaf(
1774 CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit),
1775 thread, rax);
1776 restore_native_result(masm, ret_type, stack_slots);
1777 }
1779 // We can finally stop using that last_Java_frame we setup ages ago
1781 __ reset_last_Java_frame(thread, false, true);
1783 // Unpack oop result
1784 if (ret_type == T_OBJECT || ret_type == T_ARRAY) {
1785 Label L;
1786 __ cmpptr(rax, (int32_t)NULL_WORD);
1787 __ jcc(Assembler::equal, L);
1788 __ movptr(rax, Address(rax, 0));
1789 __ bind(L);
1790 __ verify_oop(rax);
1791 }
1793 // reset handle block
1794 __ movptr(rcx, Address(thread, JavaThread::active_handles_offset()));
1796 __ movptr(Address(rcx, JNIHandleBlock::top_offset_in_bytes()), (int32_t)NULL_WORD);
1798 // Any exception pending?
1799 __ cmpptr(Address(thread, in_bytes(Thread::pending_exception_offset())), (int32_t)NULL_WORD);
1800 __ jcc(Assembler::notEqual, exception_pending);
1803 // no exception, we're almost done
1805 // check that only result value is on FPU stack
1806 __ verify_FPU(ret_type == T_FLOAT || ret_type == T_DOUBLE ? 1 : 0, "native_wrapper normal exit");
1808 // Fixup floating pointer results so that result looks like a return from a compiled method
1809 if (ret_type == T_FLOAT) {
1810 if (UseSSE >= 1) {
1811 // Pop st0 and store as float and reload into xmm register
1812 __ fstp_s(Address(rbp, -4));
1813 __ movflt(xmm0, Address(rbp, -4));
1814 }
1815 } else if (ret_type == T_DOUBLE) {
1816 if (UseSSE >= 2) {
1817 // Pop st0 and store as double and reload into xmm register
1818 __ fstp_d(Address(rbp, -8));
1819 __ movdbl(xmm0, Address(rbp, -8));
1820 }
1821 }
1823 // Return
1825 __ leave();
1826 __ ret(0);
1828 // Unexpected paths are out of line and go here
1830 // Slow path locking & unlocking
1831 if (method->is_synchronized()) {
1833 // BEGIN Slow path lock
1835 __ bind(slow_path_lock);
1837 // has last_Java_frame setup. No exceptions so do vanilla call not call_VM
1838 // args are (oop obj, BasicLock* lock, JavaThread* thread)
1839 __ push(thread);
1840 __ push(lock_reg);
1841 __ push(obj_reg);
1842 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_locking_C)));
1843 __ addptr(rsp, 3*wordSize);
1845 #ifdef ASSERT
1846 { Label L;
1847 __ cmpptr(Address(thread, in_bytes(Thread::pending_exception_offset())), (int)NULL_WORD);
1848 __ jcc(Assembler::equal, L);
1849 __ stop("no pending exception allowed on exit from monitorenter");
1850 __ bind(L);
1851 }
1852 #endif
1853 __ jmp(lock_done);
1855 // END Slow path lock
1857 // BEGIN Slow path unlock
1858 __ bind(slow_path_unlock);
1860 // Slow path unlock
1862 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) {
1863 save_native_result(masm, ret_type, stack_slots);
1864 }
1865 // Save pending exception around call to VM (which contains an EXCEPTION_MARK)
1867 __ pushptr(Address(thread, in_bytes(Thread::pending_exception_offset())));
1868 __ movptr(Address(thread, in_bytes(Thread::pending_exception_offset())), (int32_t)NULL_WORD);
1871 // should be a peal
1872 // +wordSize because of the push above
1873 __ lea(rax, Address(rbp, lock_slot_rbp_offset));
1874 __ push(rax);
1876 __ push(obj_reg);
1877 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::complete_monitor_unlocking_C)));
1878 __ addptr(rsp, 2*wordSize);
1879 #ifdef ASSERT
1880 {
1881 Label L;
1882 __ cmpptr(Address(thread, in_bytes(Thread::pending_exception_offset())), (int32_t)NULL_WORD);
1883 __ jcc(Assembler::equal, L);
1884 __ stop("no pending exception allowed on exit complete_monitor_unlocking_C");
1885 __ bind(L);
1886 }
1887 #endif /* ASSERT */
1889 __ popptr(Address(thread, in_bytes(Thread::pending_exception_offset())));
1891 if (ret_type == T_FLOAT || ret_type == T_DOUBLE ) {
1892 restore_native_result(masm, ret_type, stack_slots);
1893 }
1894 __ jmp(unlock_done);
1895 // END Slow path unlock
1897 }
1899 // SLOW PATH Reguard the stack if needed
1901 __ bind(reguard);
1902 save_native_result(masm, ret_type, stack_slots);
1903 {
1904 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, SharedRuntime::reguard_yellow_pages)));
1905 }
1906 restore_native_result(masm, ret_type, stack_slots);
1907 __ jmp(reguard_done);
1910 // BEGIN EXCEPTION PROCESSING
1912 // Forward the exception
1913 __ bind(exception_pending);
1915 // remove possible return value from FPU register stack
1916 __ empty_FPU_stack();
1918 // pop our frame
1919 __ leave();
1920 // and forward the exception
1921 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
1923 __ flush();
1925 nmethod *nm = nmethod::new_native_nmethod(method,
1926 masm->code(),
1927 vep_offset,
1928 frame_complete,
1929 stack_slots / VMRegImpl::slots_per_word,
1930 (is_static ? in_ByteSize(klass_offset) : in_ByteSize(receiver_offset)),
1931 in_ByteSize(lock_slot_offset*VMRegImpl::stack_slot_size),
1932 oop_maps);
1933 return nm;
1935 }
1937 #ifdef HAVE_DTRACE_H
1938 // ---------------------------------------------------------------------------
1939 // Generate a dtrace nmethod for a given signature. The method takes arguments
1940 // in the Java compiled code convention, marshals them to the native
1941 // abi and then leaves nops at the position you would expect to call a native
1942 // function. When the probe is enabled the nops are replaced with a trap
1943 // instruction that dtrace inserts and the trace will cause a notification
1944 // to dtrace.
1945 //
1946 // The probes are only able to take primitive types and java/lang/String as
1947 // arguments. No other java types are allowed. Strings are converted to utf8
1948 // strings so that from dtrace point of view java strings are converted to C
1949 // strings. There is an arbitrary fixed limit on the total space that a method
1950 // can use for converting the strings. (256 chars per string in the signature).
1951 // So any java string larger then this is truncated.
1953 nmethod *SharedRuntime::generate_dtrace_nmethod(
1954 MacroAssembler *masm, methodHandle method) {
1956 // generate_dtrace_nmethod is guarded by a mutex so we are sure to
1957 // be single threaded in this method.
1958 assert(AdapterHandlerLibrary_lock->owned_by_self(), "must be");
1960 // Fill in the signature array, for the calling-convention call.
1961 int total_args_passed = method->size_of_parameters();
1963 BasicType* in_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_args_passed);
1964 VMRegPair *in_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_args_passed);
1966 // The signature we are going to use for the trap that dtrace will see
1967 // java/lang/String is converted. We drop "this" and any other object
1968 // is converted to NULL. (A one-slot java/lang/Long object reference
1969 // is converted to a two-slot long, which is why we double the allocation).
1970 BasicType* out_sig_bt = NEW_RESOURCE_ARRAY(BasicType, total_args_passed * 2);
1971 VMRegPair* out_regs = NEW_RESOURCE_ARRAY(VMRegPair, total_args_passed * 2);
1973 int i=0;
1974 int total_strings = 0;
1975 int first_arg_to_pass = 0;
1976 int total_c_args = 0;
1978 if( !method->is_static() ) { // Pass in receiver first
1979 in_sig_bt[i++] = T_OBJECT;
1980 first_arg_to_pass = 1;
1981 }
1983 // We need to convert the java args to where a native (non-jni) function
1984 // would expect them. To figure out where they go we convert the java
1985 // signature to a C signature.
1987 SignatureStream ss(method->signature());
1988 for ( ; !ss.at_return_type(); ss.next()) {
1989 BasicType bt = ss.type();
1990 in_sig_bt[i++] = bt; // Collect remaining bits of signature
1991 out_sig_bt[total_c_args++] = bt;
1992 if( bt == T_OBJECT) {
1993 symbolOop s = ss.as_symbol_or_null();
1994 if (s == vmSymbols::java_lang_String()) {
1995 total_strings++;
1996 out_sig_bt[total_c_args-1] = T_ADDRESS;
1997 } else if (s == vmSymbols::java_lang_Boolean() ||
1998 s == vmSymbols::java_lang_Character() ||
1999 s == vmSymbols::java_lang_Byte() ||
2000 s == vmSymbols::java_lang_Short() ||
2001 s == vmSymbols::java_lang_Integer() ||
2002 s == vmSymbols::java_lang_Float()) {
2003 out_sig_bt[total_c_args-1] = T_INT;
2004 } else if (s == vmSymbols::java_lang_Long() ||
2005 s == vmSymbols::java_lang_Double()) {
2006 out_sig_bt[total_c_args-1] = T_LONG;
2007 out_sig_bt[total_c_args++] = T_VOID;
2008 }
2009 } else if ( bt == T_LONG || bt == T_DOUBLE ) {
2010 in_sig_bt[i++] = T_VOID; // Longs & doubles take 2 Java slots
2011 out_sig_bt[total_c_args++] = T_VOID;
2012 }
2013 }
2015 assert(i==total_args_passed, "validly parsed signature");
2017 // Now get the compiled-Java layout as input arguments
2018 int comp_args_on_stack;
2019 comp_args_on_stack = SharedRuntime::java_calling_convention(
2020 in_sig_bt, in_regs, total_args_passed, false);
2022 // Now figure out where the args must be stored and how much stack space
2023 // they require (neglecting out_preserve_stack_slots).
2025 int out_arg_slots;
2026 out_arg_slots = c_calling_convention(out_sig_bt, out_regs, total_c_args);
2028 // Calculate the total number of stack slots we will need.
2030 // First count the abi requirement plus all of the outgoing args
2031 int stack_slots = SharedRuntime::out_preserve_stack_slots() + out_arg_slots;
2033 // Now space for the string(s) we must convert
2035 int* string_locs = NEW_RESOURCE_ARRAY(int, total_strings + 1);
2036 for (i = 0; i < total_strings ; i++) {
2037 string_locs[i] = stack_slots;
2038 stack_slots += max_dtrace_string_size / VMRegImpl::stack_slot_size;
2039 }
2041 // + 2 for return address (which we own) and saved rbp,
2043 stack_slots += 2;
2045 // Ok The space we have allocated will look like:
2046 //
2047 //
2048 // FP-> | |
2049 // |---------------------|
2050 // | string[n] |
2051 // |---------------------| <- string_locs[n]
2052 // | string[n-1] |
2053 // |---------------------| <- string_locs[n-1]
2054 // | ... |
2055 // | ... |
2056 // |---------------------| <- string_locs[1]
2057 // | string[0] |
2058 // |---------------------| <- string_locs[0]
2059 // | outbound memory |
2060 // | based arguments |
2061 // | |
2062 // |---------------------|
2063 // | |
2064 // SP-> | out_preserved_slots |
2065 //
2066 //
2068 // Now compute actual number of stack words we need rounding to make
2069 // stack properly aligned.
2070 stack_slots = round_to(stack_slots, 2 * VMRegImpl::slots_per_word);
2072 int stack_size = stack_slots * VMRegImpl::stack_slot_size;
2074 intptr_t start = (intptr_t)__ pc();
2076 // First thing make an ic check to see if we should even be here
2078 // We are free to use all registers as temps without saving them and
2079 // restoring them except rbp. rbp, is the only callee save register
2080 // as far as the interpreter and the compiler(s) are concerned.
2082 const Register ic_reg = rax;
2083 const Register receiver = rcx;
2084 Label hit;
2085 Label exception_pending;
2088 __ verify_oop(receiver);
2089 __ cmpl(ic_reg, Address(receiver, oopDesc::klass_offset_in_bytes()));
2090 __ jcc(Assembler::equal, hit);
2092 __ jump(RuntimeAddress(SharedRuntime::get_ic_miss_stub()));
2094 // verified entry must be aligned for code patching.
2095 // and the first 5 bytes must be in the same cache line
2096 // if we align at 8 then we will be sure 5 bytes are in the same line
2097 __ align(8);
2099 __ bind(hit);
2101 int vep_offset = ((intptr_t)__ pc()) - start;
2104 // The instruction at the verified entry point must be 5 bytes or longer
2105 // because it can be patched on the fly by make_non_entrant. The stack bang
2106 // instruction fits that requirement.
2108 // Generate stack overflow check
2111 if (UseStackBanging) {
2112 if (stack_size <= StackShadowPages*os::vm_page_size()) {
2113 __ bang_stack_with_offset(StackShadowPages*os::vm_page_size());
2114 } else {
2115 __ movl(rax, stack_size);
2116 __ bang_stack_size(rax, rbx);
2117 }
2118 } else {
2119 // need a 5 byte instruction to allow MT safe patching to non-entrant
2120 __ fat_nop();
2121 }
2123 assert(((int)__ pc() - start - vep_offset) >= 5,
2124 "valid size for make_non_entrant");
2126 // Generate a new frame for the wrapper.
2127 __ enter();
2129 // -2 because return address is already present and so is saved rbp,
2130 if (stack_size - 2*wordSize != 0) {
2131 __ subl(rsp, stack_size - 2*wordSize);
2132 }
2134 // Frame is now completed as far a size and linkage.
2136 int frame_complete = ((intptr_t)__ pc()) - start;
2138 // First thing we do store all the args as if we are doing the call.
2139 // Since the C calling convention is stack based that ensures that
2140 // all the Java register args are stored before we need to convert any
2141 // string we might have.
2143 int sid = 0;
2144 int c_arg, j_arg;
2145 int string_reg = 0;
2147 for (j_arg = first_arg_to_pass, c_arg = 0 ;
2148 j_arg < total_args_passed ; j_arg++, c_arg++ ) {
2150 VMRegPair src = in_regs[j_arg];
2151 VMRegPair dst = out_regs[c_arg];
2152 assert(dst.first()->is_stack() || in_sig_bt[j_arg] == T_VOID,
2153 "stack based abi assumed");
2155 switch (in_sig_bt[j_arg]) {
2157 case T_ARRAY:
2158 case T_OBJECT:
2159 if (out_sig_bt[c_arg] == T_ADDRESS) {
2160 // Any register based arg for a java string after the first
2161 // will be destroyed by the call to get_utf so we store
2162 // the original value in the location the utf string address
2163 // will eventually be stored.
2164 if (src.first()->is_reg()) {
2165 if (string_reg++ != 0) {
2166 simple_move32(masm, src, dst);
2167 }
2168 }
2169 } else if (out_sig_bt[c_arg] == T_INT || out_sig_bt[c_arg] == T_LONG) {
2170 // need to unbox a one-word value
2171 Register in_reg = rax;
2172 if ( src.first()->is_reg() ) {
2173 in_reg = src.first()->as_Register();
2174 } else {
2175 simple_move32(masm, src, in_reg->as_VMReg());
2176 }
2177 Label skipUnbox;
2178 __ movl(Address(rsp, reg2offset_out(dst.first())), NULL_WORD);
2179 if ( out_sig_bt[c_arg] == T_LONG ) {
2180 __ movl(Address(rsp, reg2offset_out(dst.second())), NULL_WORD);
2181 }
2182 __ testl(in_reg, in_reg);
2183 __ jcc(Assembler::zero, skipUnbox);
2184 assert(dst.first()->is_stack() &&
2185 (!dst.second()->is_valid() || dst.second()->is_stack()),
2186 "value(s) must go into stack slots");
2188 BasicType bt = out_sig_bt[c_arg];
2189 int box_offset = java_lang_boxing_object::value_offset_in_bytes(bt);
2190 if ( bt == T_LONG ) {
2191 __ movl(rbx, Address(in_reg,
2192 box_offset + VMRegImpl::stack_slot_size));
2193 __ movl(Address(rsp, reg2offset_out(dst.second())), rbx);
2194 }
2195 __ movl(in_reg, Address(in_reg, box_offset));
2196 __ movl(Address(rsp, reg2offset_out(dst.first())), in_reg);
2197 __ bind(skipUnbox);
2198 } else {
2199 // Convert the arg to NULL
2200 __ movl(Address(rsp, reg2offset_out(dst.first())), NULL_WORD);
2201 }
2202 if (out_sig_bt[c_arg] == T_LONG) {
2203 assert(out_sig_bt[c_arg+1] == T_VOID, "must be");
2204 ++c_arg; // Move over the T_VOID To keep the loop indices in sync
2205 }
2206 break;
2208 case T_VOID:
2209 break;
2211 case T_FLOAT:
2212 float_move(masm, src, dst);
2213 break;
2215 case T_DOUBLE:
2216 assert( j_arg + 1 < total_args_passed &&
2217 in_sig_bt[j_arg + 1] == T_VOID, "bad arg list");
2218 double_move(masm, src, dst);
2219 break;
2221 case T_LONG :
2222 long_move(masm, src, dst);
2223 break;
2225 case T_ADDRESS: assert(false, "found T_ADDRESS in java args");
2227 default:
2228 simple_move32(masm, src, dst);
2229 }
2230 }
2232 // Now we must convert any string we have to utf8
2233 //
2235 for (sid = 0, j_arg = first_arg_to_pass, c_arg = 0 ;
2236 sid < total_strings ; j_arg++, c_arg++ ) {
2238 if (out_sig_bt[c_arg] == T_ADDRESS) {
2240 Address utf8_addr = Address(
2241 rsp, string_locs[sid++] * VMRegImpl::stack_slot_size);
2242 __ leal(rax, utf8_addr);
2244 // The first string we find might still be in the original java arg
2245 // register
2246 VMReg orig_loc = in_regs[j_arg].first();
2247 Register string_oop;
2249 // This is where the argument will eventually reside
2250 Address dest = Address(rsp, reg2offset_out(out_regs[c_arg].first()));
2252 if (sid == 1 && orig_loc->is_reg()) {
2253 string_oop = orig_loc->as_Register();
2254 assert(string_oop != rax, "smashed arg");
2255 } else {
2257 if (orig_loc->is_reg()) {
2258 // Get the copy of the jls object
2259 __ movl(rcx, dest);
2260 } else {
2261 // arg is still in the original location
2262 __ movl(rcx, Address(rbp, reg2offset_in(orig_loc)));
2263 }
2264 string_oop = rcx;
2266 }
2267 Label nullString;
2268 __ movl(dest, NULL_WORD);
2269 __ testl(string_oop, string_oop);
2270 __ jcc(Assembler::zero, nullString);
2272 // Now we can store the address of the utf string as the argument
2273 __ movl(dest, rax);
2275 // And do the conversion
2276 __ call_VM_leaf(CAST_FROM_FN_PTR(
2277 address, SharedRuntime::get_utf), string_oop, rax);
2278 __ bind(nullString);
2279 }
2281 if (in_sig_bt[j_arg] == T_OBJECT && out_sig_bt[c_arg] == T_LONG) {
2282 assert(out_sig_bt[c_arg+1] == T_VOID, "must be");
2283 ++c_arg; // Move over the T_VOID To keep the loop indices in sync
2284 }
2285 }
2288 // Ok now we are done. Need to place the nop that dtrace wants in order to
2289 // patch in the trap
2291 int patch_offset = ((intptr_t)__ pc()) - start;
2293 __ nop();
2296 // Return
2298 __ leave();
2299 __ ret(0);
2301 __ flush();
2303 nmethod *nm = nmethod::new_dtrace_nmethod(
2304 method, masm->code(), vep_offset, patch_offset, frame_complete,
2305 stack_slots / VMRegImpl::slots_per_word);
2306 return nm;
2308 }
2310 #endif // HAVE_DTRACE_H
2312 // this function returns the adjust size (in number of words) to a c2i adapter
2313 // activation for use during deoptimization
2314 int Deoptimization::last_frame_adjust(int callee_parameters, int callee_locals ) {
2315 return (callee_locals - callee_parameters) * Interpreter::stackElementWords();
2316 }
2319 uint SharedRuntime::out_preserve_stack_slots() {
2320 return 0;
2321 }
2324 //------------------------------generate_deopt_blob----------------------------
2325 void SharedRuntime::generate_deopt_blob() {
2326 // allocate space for the code
2327 ResourceMark rm;
2328 // setup code generation tools
2329 CodeBuffer buffer("deopt_blob", 1024, 1024);
2330 MacroAssembler* masm = new MacroAssembler(&buffer);
2331 int frame_size_in_words;
2332 OopMap* map = NULL;
2333 // Account for the extra args we place on the stack
2334 // by the time we call fetch_unroll_info
2335 const int additional_words = 2; // deopt kind, thread
2337 OopMapSet *oop_maps = new OopMapSet();
2339 // -------------
2340 // This code enters when returning to a de-optimized nmethod. A return
2341 // address has been pushed on the the stack, and return values are in
2342 // registers.
2343 // If we are doing a normal deopt then we were called from the patched
2344 // nmethod from the point we returned to the nmethod. So the return
2345 // address on the stack is wrong by NativeCall::instruction_size
2346 // We will adjust the value to it looks like we have the original return
2347 // address on the stack (like when we eagerly deoptimized).
2348 // In the case of an exception pending with deoptimized then we enter
2349 // with a return address on the stack that points after the call we patched
2350 // into the exception handler. We have the following register state:
2351 // rax,: exception
2352 // rbx,: exception handler
2353 // rdx: throwing pc
2354 // So in this case we simply jam rdx into the useless return address and
2355 // the stack looks just like we want.
2356 //
2357 // At this point we need to de-opt. We save the argument return
2358 // registers. We call the first C routine, fetch_unroll_info(). This
2359 // routine captures the return values and returns a structure which
2360 // describes the current frame size and the sizes of all replacement frames.
2361 // The current frame is compiled code and may contain many inlined
2362 // functions, each with their own JVM state. We pop the current frame, then
2363 // push all the new frames. Then we call the C routine unpack_frames() to
2364 // populate these frames. Finally unpack_frames() returns us the new target
2365 // address. Notice that callee-save registers are BLOWN here; they have
2366 // already been captured in the vframeArray at the time the return PC was
2367 // patched.
2368 address start = __ pc();
2369 Label cont;
2371 // Prolog for non exception case!
2373 // Save everything in sight.
2375 map = RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words);
2376 // Normal deoptimization
2377 __ push(Deoptimization::Unpack_deopt);
2378 __ jmp(cont);
2380 int reexecute_offset = __ pc() - start;
2382 // Reexecute case
2383 // return address is the pc describes what bci to do re-execute at
2385 // No need to update map as each call to save_live_registers will produce identical oopmap
2386 (void) RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words);
2388 __ push(Deoptimization::Unpack_reexecute);
2389 __ jmp(cont);
2391 int exception_offset = __ pc() - start;
2393 // Prolog for exception case
2395 // all registers are dead at this entry point, except for rax, and
2396 // rdx which contain the exception oop and exception pc
2397 // respectively. Set them in TLS and fall thru to the
2398 // unpack_with_exception_in_tls entry point.
2400 __ get_thread(rdi);
2401 __ movptr(Address(rdi, JavaThread::exception_pc_offset()), rdx);
2402 __ movptr(Address(rdi, JavaThread::exception_oop_offset()), rax);
2404 int exception_in_tls_offset = __ pc() - start;
2406 // new implementation because exception oop is now passed in JavaThread
2408 // Prolog for exception case
2409 // All registers must be preserved because they might be used by LinearScan
2410 // Exceptiop oop and throwing PC are passed in JavaThread
2411 // tos: stack at point of call to method that threw the exception (i.e. only
2412 // args are on the stack, no return address)
2414 // make room on stack for the return address
2415 // It will be patched later with the throwing pc. The correct value is not
2416 // available now because loading it from memory would destroy registers.
2417 __ push(0);
2419 // Save everything in sight.
2421 // No need to update map as each call to save_live_registers will produce identical oopmap
2422 (void) RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words);
2424 // Now it is safe to overwrite any register
2426 // store the correct deoptimization type
2427 __ push(Deoptimization::Unpack_exception);
2429 // load throwing pc from JavaThread and patch it as the return address
2430 // of the current frame. Then clear the field in JavaThread
2431 __ get_thread(rdi);
2432 __ movptr(rdx, Address(rdi, JavaThread::exception_pc_offset()));
2433 __ movptr(Address(rbp, wordSize), rdx);
2434 __ movptr(Address(rdi, JavaThread::exception_pc_offset()), (int32_t)NULL_WORD);
2436 #ifdef ASSERT
2437 // verify that there is really an exception oop in JavaThread
2438 __ movptr(rax, Address(rdi, JavaThread::exception_oop_offset()));
2439 __ verify_oop(rax);
2441 // verify that there is no pending exception
2442 Label no_pending_exception;
2443 __ movptr(rax, Address(rdi, Thread::pending_exception_offset()));
2444 __ testptr(rax, rax);
2445 __ jcc(Assembler::zero, no_pending_exception);
2446 __ stop("must not have pending exception here");
2447 __ bind(no_pending_exception);
2448 #endif
2450 __ bind(cont);
2452 // Compiled code leaves the floating point stack dirty, empty it.
2453 __ empty_FPU_stack();
2456 // Call C code. Need thread and this frame, but NOT official VM entry
2457 // crud. We cannot block on this call, no GC can happen.
2458 __ get_thread(rcx);
2459 __ push(rcx);
2460 // fetch_unroll_info needs to call last_java_frame()
2461 __ set_last_Java_frame(rcx, noreg, noreg, NULL);
2463 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::fetch_unroll_info)));
2465 // Need to have an oopmap that tells fetch_unroll_info where to
2466 // find any register it might need.
2468 oop_maps->add_gc_map( __ pc()-start, map);
2470 // Discard arg to fetch_unroll_info
2471 __ pop(rcx);
2473 __ get_thread(rcx);
2474 __ reset_last_Java_frame(rcx, false, false);
2476 // Load UnrollBlock into EDI
2477 __ mov(rdi, rax);
2479 // Move the unpack kind to a safe place in the UnrollBlock because
2480 // we are very short of registers
2482 Address unpack_kind(rdi, Deoptimization::UnrollBlock::unpack_kind_offset_in_bytes());
2483 // retrieve the deopt kind from where we left it.
2484 __ pop(rax);
2485 __ movl(unpack_kind, rax); // save the unpack_kind value
2487 Label noException;
2488 __ cmpl(rax, Deoptimization::Unpack_exception); // Was exception pending?
2489 __ jcc(Assembler::notEqual, noException);
2490 __ movptr(rax, Address(rcx, JavaThread::exception_oop_offset()));
2491 __ movptr(rdx, Address(rcx, JavaThread::exception_pc_offset()));
2492 __ movptr(Address(rcx, JavaThread::exception_oop_offset()), (int32_t)NULL_WORD);
2493 __ movptr(Address(rcx, JavaThread::exception_pc_offset()), (int32_t)NULL_WORD);
2495 __ verify_oop(rax);
2497 // Overwrite the result registers with the exception results.
2498 __ movptr(Address(rsp, RegisterSaver::raxOffset()*wordSize), rax);
2499 __ movptr(Address(rsp, RegisterSaver::rdxOffset()*wordSize), rdx);
2501 __ bind(noException);
2503 // Stack is back to only having register save data on the stack.
2504 // Now restore the result registers. Everything else is either dead or captured
2505 // in the vframeArray.
2507 RegisterSaver::restore_result_registers(masm);
2509 // All of the register save area has been popped of the stack. Only the
2510 // return address remains.
2512 // Pop all the frames we must move/replace.
2513 //
2514 // Frame picture (youngest to oldest)
2515 // 1: self-frame (no frame link)
2516 // 2: deopting frame (no frame link)
2517 // 3: caller of deopting frame (could be compiled/interpreted).
2518 //
2519 // Note: by leaving the return address of self-frame on the stack
2520 // and using the size of frame 2 to adjust the stack
2521 // when we are done the return to frame 3 will still be on the stack.
2523 // Pop deoptimized frame
2524 __ addptr(rsp, Address(rdi,Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset_in_bytes()));
2526 // sp should be pointing at the return address to the caller (3)
2528 // Stack bang to make sure there's enough room for these interpreter frames.
2529 if (UseStackBanging) {
2530 __ movl(rbx, Address(rdi ,Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes()));
2531 __ bang_stack_size(rbx, rcx);
2532 }
2534 // Load array of frame pcs into ECX
2535 __ movptr(rcx,Address(rdi,Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes()));
2537 __ pop(rsi); // trash the old pc
2539 // Load array of frame sizes into ESI
2540 __ movptr(rsi,Address(rdi,Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes()));
2542 Address counter(rdi, Deoptimization::UnrollBlock::counter_temp_offset_in_bytes());
2544 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes()));
2545 __ movl(counter, rbx);
2547 // Pick up the initial fp we should save
2548 __ movptr(rbp, Address(rdi, Deoptimization::UnrollBlock::initial_fp_offset_in_bytes()));
2550 // Now adjust the caller's stack to make up for the extra locals
2551 // but record the original sp so that we can save it in the skeletal interpreter
2552 // frame and the stack walking of interpreter_sender will get the unextended sp
2553 // value and not the "real" sp value.
2555 Address sp_temp(rdi, Deoptimization::UnrollBlock::sender_sp_temp_offset_in_bytes());
2556 __ movptr(sp_temp, rsp);
2557 __ movl2ptr(rbx, Address(rdi, Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes()));
2558 __ subptr(rsp, rbx);
2560 // Push interpreter frames in a loop
2561 Label loop;
2562 __ bind(loop);
2563 __ movptr(rbx, Address(rsi, 0)); // Load frame size
2564 #ifdef CC_INTERP
2565 __ subptr(rbx, 4*wordSize); // we'll push pc and ebp by hand and
2566 #ifdef ASSERT
2567 __ push(0xDEADDEAD); // Make a recognizable pattern
2568 __ push(0xDEADDEAD);
2569 #else /* ASSERT */
2570 __ subptr(rsp, 2*wordSize); // skip the "static long no_param"
2571 #endif /* ASSERT */
2572 #else /* CC_INTERP */
2573 __ subptr(rbx, 2*wordSize); // we'll push pc and rbp, by hand
2574 #endif /* CC_INTERP */
2575 __ pushptr(Address(rcx, 0)); // save return address
2576 __ enter(); // save old & set new rbp,
2577 __ subptr(rsp, rbx); // Prolog!
2578 __ movptr(rbx, sp_temp); // sender's sp
2579 #ifdef CC_INTERP
2580 __ movptr(Address(rbp,
2581 -(sizeof(BytecodeInterpreter)) + in_bytes(byte_offset_of(BytecodeInterpreter, _sender_sp))),
2582 rbx); // Make it walkable
2583 #else /* CC_INTERP */
2584 // This value is corrected by layout_activation_impl
2585 __ movptr(Address(rbp, frame::interpreter_frame_last_sp_offset * wordSize), (int32_t)NULL_WORD );
2586 __ movptr(Address(rbp, frame::interpreter_frame_sender_sp_offset * wordSize), rbx); // Make it walkable
2587 #endif /* CC_INTERP */
2588 __ movptr(sp_temp, rsp); // pass to next frame
2589 __ addptr(rsi, wordSize); // Bump array pointer (sizes)
2590 __ addptr(rcx, wordSize); // Bump array pointer (pcs)
2591 __ decrementl(counter); // decrement counter
2592 __ jcc(Assembler::notZero, loop);
2593 __ pushptr(Address(rcx, 0)); // save final return address
2595 // Re-push self-frame
2596 __ enter(); // save old & set new rbp,
2598 // Return address and rbp, are in place
2599 // We'll push additional args later. Just allocate a full sized
2600 // register save area
2601 __ subptr(rsp, (frame_size_in_words-additional_words - 2) * wordSize);
2603 // Restore frame locals after moving the frame
2604 __ movptr(Address(rsp, RegisterSaver::raxOffset()*wordSize), rax);
2605 __ movptr(Address(rsp, RegisterSaver::rdxOffset()*wordSize), rdx);
2606 __ fstp_d(Address(rsp, RegisterSaver::fpResultOffset()*wordSize)); // Pop float stack and store in local
2607 if( UseSSE>=2 ) __ movdbl(Address(rsp, RegisterSaver::xmm0Offset()*wordSize), xmm0);
2608 if( UseSSE==1 ) __ movflt(Address(rsp, RegisterSaver::xmm0Offset()*wordSize), xmm0);
2610 // Set up the args to unpack_frame
2612 __ pushl(unpack_kind); // get the unpack_kind value
2613 __ get_thread(rcx);
2614 __ push(rcx);
2616 // set last_Java_sp, last_Java_fp
2617 __ set_last_Java_frame(rcx, noreg, rbp, NULL);
2619 // Call C code. Need thread but NOT official VM entry
2620 // crud. We cannot block on this call, no GC can happen. Call should
2621 // restore return values to their stack-slots with the new SP.
2622 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames)));
2623 // Set an oopmap for the call site
2624 oop_maps->add_gc_map( __ pc()-start, new OopMap( frame_size_in_words, 0 ));
2626 // rax, contains the return result type
2627 __ push(rax);
2629 __ get_thread(rcx);
2630 __ reset_last_Java_frame(rcx, false, false);
2632 // Collect return values
2633 __ movptr(rax,Address(rsp, (RegisterSaver::raxOffset() + additional_words + 1)*wordSize));
2634 __ movptr(rdx,Address(rsp, (RegisterSaver::rdxOffset() + additional_words + 1)*wordSize));
2636 // Clear floating point stack before returning to interpreter
2637 __ empty_FPU_stack();
2639 // Check if we should push the float or double return value.
2640 Label results_done, yes_double_value;
2641 __ cmpl(Address(rsp, 0), T_DOUBLE);
2642 __ jcc (Assembler::zero, yes_double_value);
2643 __ cmpl(Address(rsp, 0), T_FLOAT);
2644 __ jcc (Assembler::notZero, results_done);
2646 // return float value as expected by interpreter
2647 if( UseSSE>=1 ) __ movflt(xmm0, Address(rsp, (RegisterSaver::xmm0Offset() + additional_words + 1)*wordSize));
2648 else __ fld_d(Address(rsp, (RegisterSaver::fpResultOffset() + additional_words + 1)*wordSize));
2649 __ jmp(results_done);
2651 // return double value as expected by interpreter
2652 __ bind(yes_double_value);
2653 if( UseSSE>=2 ) __ movdbl(xmm0, Address(rsp, (RegisterSaver::xmm0Offset() + additional_words + 1)*wordSize));
2654 else __ fld_d(Address(rsp, (RegisterSaver::fpResultOffset() + additional_words + 1)*wordSize));
2656 __ bind(results_done);
2658 // Pop self-frame.
2659 __ leave(); // Epilog!
2661 // Jump to interpreter
2662 __ ret(0);
2664 // -------------
2665 // make sure all code is generated
2666 masm->flush();
2668 _deopt_blob = DeoptimizationBlob::create( &buffer, oop_maps, 0, exception_offset, reexecute_offset, frame_size_in_words);
2669 _deopt_blob->set_unpack_with_exception_in_tls_offset(exception_in_tls_offset);
2670 }
2673 #ifdef COMPILER2
2674 //------------------------------generate_uncommon_trap_blob--------------------
2675 void SharedRuntime::generate_uncommon_trap_blob() {
2676 // allocate space for the code
2677 ResourceMark rm;
2678 // setup code generation tools
2679 CodeBuffer buffer("uncommon_trap_blob", 512, 512);
2680 MacroAssembler* masm = new MacroAssembler(&buffer);
2682 enum frame_layout {
2683 arg0_off, // thread sp + 0 // Arg location for
2684 arg1_off, // unloaded_class_index sp + 1 // calling C
2685 // The frame sender code expects that rbp will be in the "natural" place and
2686 // will override any oopMap setting for it. We must therefore force the layout
2687 // so that it agrees with the frame sender code.
2688 rbp_off, // callee saved register sp + 2
2689 return_off, // slot for return address sp + 3
2690 framesize
2691 };
2693 address start = __ pc();
2694 // Push self-frame.
2695 __ subptr(rsp, return_off*wordSize); // Epilog!
2697 // rbp, is an implicitly saved callee saved register (i.e. the calling
2698 // convention will save restore it in prolog/epilog) Other than that
2699 // there are no callee save registers no that adapter frames are gone.
2700 __ movptr(Address(rsp, rbp_off*wordSize), rbp);
2702 // Clear the floating point exception stack
2703 __ empty_FPU_stack();
2705 // set last_Java_sp
2706 __ get_thread(rdx);
2707 __ set_last_Java_frame(rdx, noreg, noreg, NULL);
2709 // Call C code. Need thread but NOT official VM entry
2710 // crud. We cannot block on this call, no GC can happen. Call should
2711 // capture callee-saved registers as well as return values.
2712 __ movptr(Address(rsp, arg0_off*wordSize), rdx);
2713 // argument already in ECX
2714 __ movl(Address(rsp, arg1_off*wordSize),rcx);
2715 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::uncommon_trap)));
2717 // Set an oopmap for the call site
2718 OopMapSet *oop_maps = new OopMapSet();
2719 OopMap* map = new OopMap( framesize, 0 );
2720 // No oopMap for rbp, it is known implicitly
2722 oop_maps->add_gc_map( __ pc()-start, map);
2724 __ get_thread(rcx);
2726 __ reset_last_Java_frame(rcx, false, false);
2728 // Load UnrollBlock into EDI
2729 __ movptr(rdi, rax);
2731 // Pop all the frames we must move/replace.
2732 //
2733 // Frame picture (youngest to oldest)
2734 // 1: self-frame (no frame link)
2735 // 2: deopting frame (no frame link)
2736 // 3: caller of deopting frame (could be compiled/interpreted).
2738 // Pop self-frame. We have no frame, and must rely only on EAX and ESP.
2739 __ addptr(rsp,(framesize-1)*wordSize); // Epilog!
2741 // Pop deoptimized frame
2742 __ movl2ptr(rcx, Address(rdi,Deoptimization::UnrollBlock::size_of_deoptimized_frame_offset_in_bytes()));
2743 __ addptr(rsp, rcx);
2745 // sp should be pointing at the return address to the caller (3)
2747 // Stack bang to make sure there's enough room for these interpreter frames.
2748 if (UseStackBanging) {
2749 __ movl(rbx, Address(rdi ,Deoptimization::UnrollBlock::total_frame_sizes_offset_in_bytes()));
2750 __ bang_stack_size(rbx, rcx);
2751 }
2754 // Load array of frame pcs into ECX
2755 __ movl(rcx,Address(rdi,Deoptimization::UnrollBlock::frame_pcs_offset_in_bytes()));
2757 __ pop(rsi); // trash the pc
2759 // Load array of frame sizes into ESI
2760 __ movptr(rsi,Address(rdi,Deoptimization::UnrollBlock::frame_sizes_offset_in_bytes()));
2762 Address counter(rdi, Deoptimization::UnrollBlock::counter_temp_offset_in_bytes());
2764 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock::number_of_frames_offset_in_bytes()));
2765 __ movl(counter, rbx);
2767 // Pick up the initial fp we should save
2768 __ movptr(rbp, Address(rdi, Deoptimization::UnrollBlock::initial_fp_offset_in_bytes()));
2770 // Now adjust the caller's stack to make up for the extra locals
2771 // but record the original sp so that we can save it in the skeletal interpreter
2772 // frame and the stack walking of interpreter_sender will get the unextended sp
2773 // value and not the "real" sp value.
2775 Address sp_temp(rdi, Deoptimization::UnrollBlock::sender_sp_temp_offset_in_bytes());
2776 __ movptr(sp_temp, rsp);
2777 __ movl(rbx, Address(rdi, Deoptimization::UnrollBlock::caller_adjustment_offset_in_bytes()));
2778 __ subptr(rsp, rbx);
2780 // Push interpreter frames in a loop
2781 Label loop;
2782 __ bind(loop);
2783 __ movptr(rbx, Address(rsi, 0)); // Load frame size
2784 #ifdef CC_INTERP
2785 __ subptr(rbx, 4*wordSize); // we'll push pc and ebp by hand and
2786 #ifdef ASSERT
2787 __ push(0xDEADDEAD); // Make a recognizable pattern
2788 __ push(0xDEADDEAD); // (parm to RecursiveInterpreter...)
2789 #else /* ASSERT */
2790 __ subptr(rsp, 2*wordSize); // skip the "static long no_param"
2791 #endif /* ASSERT */
2792 #else /* CC_INTERP */
2793 __ subptr(rbx, 2*wordSize); // we'll push pc and rbp, by hand
2794 #endif /* CC_INTERP */
2795 __ pushptr(Address(rcx, 0)); // save return address
2796 __ enter(); // save old & set new rbp,
2797 __ subptr(rsp, rbx); // Prolog!
2798 __ movptr(rbx, sp_temp); // sender's sp
2799 #ifdef CC_INTERP
2800 __ movptr(Address(rbp,
2801 -(sizeof(BytecodeInterpreter)) + in_bytes(byte_offset_of(BytecodeInterpreter, _sender_sp))),
2802 rbx); // Make it walkable
2803 #else /* CC_INTERP */
2804 // This value is corrected by layout_activation_impl
2805 __ movptr(Address(rbp, frame::interpreter_frame_last_sp_offset * wordSize), (int32_t)NULL_WORD );
2806 __ movptr(Address(rbp, frame::interpreter_frame_sender_sp_offset * wordSize), rbx); // Make it walkable
2807 #endif /* CC_INTERP */
2808 __ movptr(sp_temp, rsp); // pass to next frame
2809 __ addptr(rsi, wordSize); // Bump array pointer (sizes)
2810 __ addptr(rcx, wordSize); // Bump array pointer (pcs)
2811 __ decrementl(counter); // decrement counter
2812 __ jcc(Assembler::notZero, loop);
2813 __ pushptr(Address(rcx, 0)); // save final return address
2815 // Re-push self-frame
2816 __ enter(); // save old & set new rbp,
2817 __ subptr(rsp, (framesize-2) * wordSize); // Prolog!
2820 // set last_Java_sp, last_Java_fp
2821 __ get_thread(rdi);
2822 __ set_last_Java_frame(rdi, noreg, rbp, NULL);
2824 // Call C code. Need thread but NOT official VM entry
2825 // crud. We cannot block on this call, no GC can happen. Call should
2826 // restore return values to their stack-slots with the new SP.
2827 __ movptr(Address(rsp,arg0_off*wordSize),rdi);
2828 __ movl(Address(rsp,arg1_off*wordSize), Deoptimization::Unpack_uncommon_trap);
2829 __ call(RuntimeAddress(CAST_FROM_FN_PTR(address, Deoptimization::unpack_frames)));
2830 // Set an oopmap for the call site
2831 oop_maps->add_gc_map( __ pc()-start, new OopMap( framesize, 0 ) );
2833 __ get_thread(rdi);
2834 __ reset_last_Java_frame(rdi, true, false);
2836 // Pop self-frame.
2837 __ leave(); // Epilog!
2839 // Jump to interpreter
2840 __ ret(0);
2842 // -------------
2843 // make sure all code is generated
2844 masm->flush();
2846 _uncommon_trap_blob = UncommonTrapBlob::create(&buffer, oop_maps, framesize);
2847 }
2848 #endif // COMPILER2
2850 //------------------------------generate_handler_blob------
2851 //
2852 // Generate a special Compile2Runtime blob that saves all registers,
2853 // setup oopmap, and calls safepoint code to stop the compiled code for
2854 // a safepoint.
2855 //
2856 static SafepointBlob* generate_handler_blob(address call_ptr, bool cause_return) {
2858 // Account for thread arg in our frame
2859 const int additional_words = 1;
2860 int frame_size_in_words;
2862 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before");
2864 ResourceMark rm;
2865 OopMapSet *oop_maps = new OopMapSet();
2866 OopMap* map;
2868 // allocate space for the code
2869 // setup code generation tools
2870 CodeBuffer buffer("handler_blob", 1024, 512);
2871 MacroAssembler* masm = new MacroAssembler(&buffer);
2873 const Register java_thread = rdi; // callee-saved for VC++
2874 address start = __ pc();
2875 address call_pc = NULL;
2877 // If cause_return is true we are at a poll_return and there is
2878 // the return address on the stack to the caller on the nmethod
2879 // that is safepoint. We can leave this return on the stack and
2880 // effectively complete the return and safepoint in the caller.
2881 // Otherwise we push space for a return address that the safepoint
2882 // handler will install later to make the stack walking sensible.
2883 if( !cause_return )
2884 __ push(rbx); // Make room for return address (or push it again)
2886 map = RegisterSaver::save_live_registers(masm, additional_words, &frame_size_in_words, false);
2888 // The following is basically a call_VM. However, we need the precise
2889 // address of the call in order to generate an oopmap. Hence, we do all the
2890 // work ourselves.
2892 // Push thread argument and setup last_Java_sp
2893 __ get_thread(java_thread);
2894 __ push(java_thread);
2895 __ set_last_Java_frame(java_thread, noreg, noreg, NULL);
2897 // if this was not a poll_return then we need to correct the return address now.
2898 if( !cause_return ) {
2899 __ movptr(rax, Address(java_thread, JavaThread::saved_exception_pc_offset()));
2900 __ movptr(Address(rbp, wordSize), rax);
2901 }
2903 // do the call
2904 __ call(RuntimeAddress(call_ptr));
2906 // Set an oopmap for the call site. This oopmap will map all
2907 // oop-registers and debug-info registers as callee-saved. This
2908 // will allow deoptimization at this safepoint to find all possible
2909 // debug-info recordings, as well as let GC find all oops.
2911 oop_maps->add_gc_map( __ pc() - start, map);
2913 // Discard arg
2914 __ pop(rcx);
2916 Label noException;
2918 // Clear last_Java_sp again
2919 __ get_thread(java_thread);
2920 __ reset_last_Java_frame(java_thread, false, false);
2922 __ cmpptr(Address(java_thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
2923 __ jcc(Assembler::equal, noException);
2925 // Exception pending
2927 RegisterSaver::restore_live_registers(masm);
2929 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
2931 __ bind(noException);
2933 // Normal exit, register restoring and exit
2934 RegisterSaver::restore_live_registers(masm);
2936 __ ret(0);
2938 // make sure all code is generated
2939 masm->flush();
2941 // Fill-out other meta info
2942 return SafepointBlob::create(&buffer, oop_maps, frame_size_in_words);
2943 }
2945 //
2946 // generate_resolve_blob - call resolution (static/virtual/opt-virtual/ic-miss
2947 //
2948 // Generate a stub that calls into vm to find out the proper destination
2949 // of a java call. All the argument registers are live at this point
2950 // but since this is generic code we don't know what they are and the caller
2951 // must do any gc of the args.
2952 //
2953 static RuntimeStub* generate_resolve_blob(address destination, const char* name) {
2954 assert (StubRoutines::forward_exception_entry() != NULL, "must be generated before");
2956 // allocate space for the code
2957 ResourceMark rm;
2959 CodeBuffer buffer(name, 1000, 512);
2960 MacroAssembler* masm = new MacroAssembler(&buffer);
2962 int frame_size_words;
2963 enum frame_layout {
2964 thread_off,
2965 extra_words };
2967 OopMapSet *oop_maps = new OopMapSet();
2968 OopMap* map = NULL;
2970 int start = __ offset();
2972 map = RegisterSaver::save_live_registers(masm, extra_words, &frame_size_words);
2974 int frame_complete = __ offset();
2976 const Register thread = rdi;
2977 __ get_thread(rdi);
2979 __ push(thread);
2980 __ set_last_Java_frame(thread, noreg, rbp, NULL);
2982 __ call(RuntimeAddress(destination));
2985 // Set an oopmap for the call site.
2986 // We need this not only for callee-saved registers, but also for volatile
2987 // registers that the compiler might be keeping live across a safepoint.
2989 oop_maps->add_gc_map( __ offset() - start, map);
2991 // rax, contains the address we are going to jump to assuming no exception got installed
2993 __ addptr(rsp, wordSize);
2995 // clear last_Java_sp
2996 __ reset_last_Java_frame(thread, true, false);
2997 // check for pending exceptions
2998 Label pending;
2999 __ cmpptr(Address(thread, Thread::pending_exception_offset()), (int32_t)NULL_WORD);
3000 __ jcc(Assembler::notEqual, pending);
3002 // get the returned methodOop
3003 __ movptr(rbx, Address(thread, JavaThread::vm_result_offset()));
3004 __ movptr(Address(rsp, RegisterSaver::rbx_offset() * wordSize), rbx);
3006 __ movptr(Address(rsp, RegisterSaver::rax_offset() * wordSize), rax);
3008 RegisterSaver::restore_live_registers(masm);
3010 // We are back the the original state on entry and ready to go.
3012 __ jmp(rax);
3014 // Pending exception after the safepoint
3016 __ bind(pending);
3018 RegisterSaver::restore_live_registers(masm);
3020 // exception pending => remove activation and forward to exception handler
3022 __ get_thread(thread);
3023 __ movptr(Address(thread, JavaThread::vm_result_offset()), (int32_t)NULL_WORD);
3024 __ movptr(rax, Address(thread, Thread::pending_exception_offset()));
3025 __ jump(RuntimeAddress(StubRoutines::forward_exception_entry()));
3027 // -------------
3028 // make sure all code is generated
3029 masm->flush();
3031 // return the blob
3032 // frame_size_words or bytes??
3033 return RuntimeStub::new_runtime_stub(name, &buffer, frame_complete, frame_size_words, oop_maps, true);
3034 }
3036 void SharedRuntime::generate_stubs() {
3038 _wrong_method_blob = generate_resolve_blob(CAST_FROM_FN_PTR(address, SharedRuntime::handle_wrong_method),
3039 "wrong_method_stub");
3041 _ic_miss_blob = generate_resolve_blob(CAST_FROM_FN_PTR(address, SharedRuntime::handle_wrong_method_ic_miss),
3042 "ic_miss_stub");
3044 _resolve_opt_virtual_call_blob = generate_resolve_blob(CAST_FROM_FN_PTR(address, SharedRuntime::resolve_opt_virtual_call_C),
3045 "resolve_opt_virtual_call");
3047 _resolve_virtual_call_blob = generate_resolve_blob(CAST_FROM_FN_PTR(address, SharedRuntime::resolve_virtual_call_C),
3048 "resolve_virtual_call");
3050 _resolve_static_call_blob = generate_resolve_blob(CAST_FROM_FN_PTR(address, SharedRuntime::resolve_static_call_C),
3051 "resolve_static_call");
3053 _polling_page_safepoint_handler_blob =
3054 generate_handler_blob(CAST_FROM_FN_PTR(address,
3055 SafepointSynchronize::handle_polling_page_exception), false);
3057 _polling_page_return_handler_blob =
3058 generate_handler_blob(CAST_FROM_FN_PTR(address,
3059 SafepointSynchronize::handle_polling_page_exception), true);
3061 generate_deopt_blob();
3062 #ifdef COMPILER2
3063 generate_uncommon_trap_blob();
3064 #endif // COMPILER2
3065 }