2 * tramp-amd64.c: JIT trampoline code for amd64
5 * Dietmar Maurer (dietmar@ximian.com)
6 * Zoltan Varga (vargaz@gmail.com)
8 * (C) 2001 Ximian, Inc.
9 * Copyright 2003-2011 Novell, Inc (http://www.novell.com)
10 * Copyright 2011 Xamarin, Inc (http://www.xamarin.com)
16 #include <mono/metadata/abi-details.h>
17 #include <mono/metadata/appdomain.h>
18 #include <mono/metadata/marshal.h>
19 #include <mono/metadata/tabledefs.h>
20 #include <mono/metadata/mono-debug-debugger.h>
21 #include <mono/metadata/monitor.h>
22 #include <mono/metadata/profiler-private.h>
23 #include <mono/metadata/gc-internal.h>
24 #include <mono/arch/amd64/amd64-codegen.h>
26 #include <mono/utils/memcheck.h>
29 #include "mini-amd64.h"
30 #include "debugger-agent.h"
32 #if defined(__native_client_codegen__) && defined(__native_client__)
34 #include <nacl/nacl_dyncode.h>
37 #define ALIGN_TO(val,align) ((((guint64)val) + ((align) - 1)) & ~((align) - 1))
39 #define IS_REX(inst) (((inst) >= 0x40) && ((inst) <= 0x4f))
42 * mono_arch_get_unbox_trampoline:
44 * @addr: pointer to native code for @m
46 * when value type methods are called through the vtable we need to unbox the
47 * this argument. This method returns a pointer to a trampoline which does
48 * unboxing before calling the method
51 mono_arch_get_unbox_trampoline (MonoMethod *m, gpointer addr)
54 int this_reg, size = NACL_SIZE (20, 32);
56 MonoDomain *domain = mono_domain_get ();
58 this_reg = mono_arch_get_this_arg_reg (NULL);
60 start = code = mono_domain_code_reserve (domain, size);
62 amd64_alu_reg_imm (code, X86_ADD, this_reg, sizeof (MonoObject));
63 /* FIXME: Optimize this */
64 amd64_mov_reg_imm (code, AMD64_RAX, addr);
65 amd64_jump_reg (code, AMD64_RAX);
66 g_assert ((code - start) < size);
68 nacl_domain_code_validate (domain, &start, size, &code);
70 mono_arch_flush_icache (start, code - start);
71 mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_UNBOX_TRAMPOLINE, m);
77 * mono_arch_get_static_rgctx_trampoline:
79 * Create a trampoline which sets RGCTX_REG to MRGCTX, then jumps to ADDR.
82 mono_arch_get_static_rgctx_trampoline (MonoMethod *m, MonoMethodRuntimeGenericContext *mrgctx, gpointer addr)
87 MonoDomain *domain = mono_domain_get ();
89 #ifdef MONO_ARCH_NOMAP32BIT
92 /* AOTed code could still have a non-32 bit address */
93 if ((((guint64)addr) >> 32) == 0)
94 buf_len = NACL_SIZE (16, 32);
96 buf_len = NACL_SIZE (30, 32);
99 start = code = mono_domain_code_reserve (domain, buf_len);
101 amd64_mov_reg_imm (code, MONO_ARCH_RGCTX_REG, mrgctx);
102 amd64_jump_code (code, addr);
103 g_assert ((code - start) < buf_len);
105 nacl_domain_code_validate (domain, &start, buf_len, &code);
106 mono_arch_flush_icache (start, code - start);
107 mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_GENERICS_TRAMPOLINE, NULL);
113 mono_arch_get_llvm_imt_trampoline (MonoDomain *domain, MonoMethod *m, int vt_offset)
115 guint8 *code, *start;
121 start = code = mono_domain_code_reserve (domain, buf_len);
123 this_reg = mono_arch_get_this_arg_reg (NULL);
126 amd64_mov_reg_imm (code, MONO_ARCH_IMT_REG, m);
127 /* Load vtable address */
128 amd64_mov_reg_membase (code, AMD64_RAX, this_reg, 0, 8);
129 amd64_jump_membase (code, AMD64_RAX, vt_offset);
132 g_assert ((code - start) < buf_len);
134 nacl_domain_code_validate (domain, &start, buf_len, &code);
136 mono_arch_flush_icache (start, code - start);
137 mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_IMT_TRAMPOLINE, NULL);
143 * mono_arch_patch_callsite:
145 * Patch the callsite whose address is given by ORIG_CODE so it calls ADDR. ORIG_CODE
146 * points to the pc right after the call.
149 mono_arch_patch_callsite (guint8 *method_start, guint8 *orig_code, guint8 *addr)
151 #if defined(__default_codegen__)
154 gboolean can_write = mono_breakpoint_clean_code (method_start, orig_code, 14, buf, sizeof (buf));
158 /* mov 64-bit imm into r11 (followed by call reg?) or direct call*/
159 if (((code [-13] == 0x49) && (code [-12] == 0xbb)) || (code [-5] == 0xe8)) {
160 if (code [-5] != 0xe8) {
162 InterlockedExchangePointer ((gpointer*)(orig_code - 11), addr);
163 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 11, sizeof (gpointer));
166 gboolean disp_32bit = ((((gint64)addr - (gint64)orig_code)) < (1 << 30)) && ((((gint64)addr - (gint64)orig_code)) > -(1 << 30));
168 if ((((guint64)(addr)) >> 32) != 0 && !disp_32bit) {
170 * This might happen with LLVM or when calling AOTed code. Create a thunk.
172 guint8 *thunk_start, *thunk_code;
174 thunk_start = thunk_code = mono_domain_code_reserve (mono_domain_get (), 32);
175 amd64_jump_membase (thunk_code, AMD64_RIP, 0);
176 *(guint64*)thunk_code = (guint64)addr;
178 g_assert ((((guint64)(addr)) >> 32) == 0);
179 mono_arch_flush_icache (thunk_start, thunk_code - thunk_start);
180 mono_profiler_code_buffer_new (thunk_start, thunk_code - thunk_start, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
183 InterlockedExchange ((gint32*)(orig_code - 4), ((gint64)addr - (gint64)orig_code));
184 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, 4);
188 else if ((code [-7] == 0x41) && (code [-6] == 0xff) && (code [-5] == 0x15)) {
189 /* call *<OFFSET>(%rip) */
190 gpointer *got_entry = (gpointer*)((guint8*)orig_code + (*(guint32*)(orig_code - 4)));
192 InterlockedExchangePointer (got_entry, addr);
193 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, sizeof (gpointer));
196 #elif defined(__native_client__)
197 /* These are essentially the same 2 cases as above, modified for NaCl*/
199 /* Target must be bundle-aligned */
200 g_assert (((guint32)addr & kNaClAlignmentMask) == 0);
201 /* Return target must be bundle-aligned */
202 g_assert (((guint32)orig_code & kNaClAlignmentMask) == 0);
204 if (orig_code[-5] == 0xe8) {
207 gint32 offset = (gint32)addr - (gint32)orig_code;
208 guint8 buf[sizeof(gint32)];
209 *((gint32*)(buf)) = offset;
210 ret = nacl_dyncode_modify (orig_code - sizeof(gint32), buf, sizeof(gint32));
214 else if (is_nacl_call_reg_sequence (orig_code - 10) && orig_code[-16] == 0x41 && orig_code[-15] == 0xbb) {
216 guint8 buf[sizeof(gint32)];
217 *((gint32 *)(buf)) = addr;
218 /* orig_code[-14] is the start of the immediate. */
219 ret = nacl_dyncode_modify (orig_code - 14, buf, sizeof(gint32));
223 g_assert_not_reached ();
231 mono_arch_create_llvm_native_thunk (MonoDomain *domain, guint8 *addr)
234 * The caller is LLVM code and the call displacement might exceed 32 bits. We can't determine the caller address, so
235 * we add a thunk every time.
236 * Since the caller is also allocated using the domain code manager, hopefully the displacement will fit into 32 bits.
237 * FIXME: Avoid this if possible if !MONO_ARCH_NOMAP32BIT and ADDR is 32 bits.
239 guint8 *thunk_start, *thunk_code;
241 thunk_start = thunk_code = mono_domain_code_reserve (mono_domain_get (), 32);
242 amd64_jump_membase (thunk_code, AMD64_RIP, 0);
243 *(guint64*)thunk_code = (guint64)addr;
245 mono_arch_flush_icache (thunk_start, thunk_code - thunk_start);
246 mono_profiler_code_buffer_new (thunk_start, thunk_code - thunk_start, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
251 mono_arch_patch_plt_entry (guint8 *code, gpointer *got, mgreg_t *regs, guint8 *addr)
254 gpointer *plt_jump_table_entry;
256 #if defined(__default_codegen__)
257 /* A PLT entry: jmp *<DISP>(%rip) */
258 g_assert (code [0] == 0xff);
259 g_assert (code [1] == 0x25);
261 disp = *(gint32*)(code + 2);
263 plt_jump_table_entry = (gpointer*)(code + 6 + disp);
264 #elif defined(__native_client_codegen__)
266 /* mov <DISP>(%rip), %r11d */
269 /* Verify the 'mov' */
270 g_assert (code [0] == 0x45);
271 g_assert (code [1] == 0x8b);
272 g_assert (code [2] == 0x1d);
274 disp = *(gint32*)(code + 3);
276 /* 7 = 3 (mov opcode) + 4 (disp) */
277 /* This needs to resolve to the target of the RIP-relative offset */
278 plt_jump_table_entry = (gpointer*)(code + 7 + disp);
280 #endif /* __native_client_codegen__ */
283 InterlockedExchangePointer (plt_jump_table_entry, addr);
287 get_vcall_slot (guint8 *code, mgreg_t *regs, int *displacement)
291 MonoJitInfo *ji = NULL;
294 /* code - 9 might be before the start of the method */
295 /* FIXME: Avoid this expensive call somehow */
296 ji = mono_jit_info_table_find (mono_domain_get (), (char*)code);
299 mono_breakpoint_clean_code (ji ? ji->code_start : NULL, code, 9, buf, sizeof (buf));
306 if ((code [0] == 0x41) && (code [1] == 0xff) && (code [2] == 0x15)) {
307 /* call OFFSET(%rip) */
308 g_assert_not_reached ();
309 *displacement = *(guint32*)(code + 3);
310 return (gpointer*)(code + disp + 7);
312 g_assert_not_reached ();
318 get_vcall_slot_addr (guint8* code, mgreg_t *regs)
322 vt = get_vcall_slot (code, regs, &displacement);
325 return (gpointer*)((char*)vt + displacement);
329 mono_arch_nullify_class_init_trampoline (guint8 *code, mgreg_t *regs)
332 MonoJitInfo *ji = NULL;
334 gpointer tramp = mini_get_nullified_class_init_trampoline ();
337 /* code - 7 might be before the start of the method */
338 /* FIXME: Avoid this expensive call somehow */
339 ji = mono_jit_info_table_find (mono_domain_get (), (char*)code);
342 can_write = mono_breakpoint_clean_code (ji ? ji->code_start : NULL, code, 7, buf, sizeof (buf));
348 * A given byte sequence can match more than case here, so we have to be
349 * really careful about the ordering of the cases. Longer sequences
352 if ((buf [0] == 0x41) && (buf [1] == 0xff) && (buf [2] == 0x15)) {
353 gpointer *vtable_slot;
355 /* call *<OFFSET>(%rip) */
356 vtable_slot = get_vcall_slot_addr (code, regs);
357 g_assert (vtable_slot);
359 *vtable_slot = tramp;
360 } else if (buf [2] == 0xe8) {
362 //guint8 *buf = code - 2;
365 * It would be better to replace the call with nops, but that doesn't seem
366 * to work on SMP machines even when the whole call is inside a cache line.
367 * Patching the call address seems to work.
377 mono_arch_patch_callsite (code - 5, code, tramp);
378 } else if ((buf [5] == 0xff) && x86_modrm_mod (buf [6]) == 3 && x86_modrm_reg (buf [6]) == 2) {
380 /* Generated by the LLVM JIT or on platforms without MAP_32BIT set */
381 mono_arch_patch_callsite (code - 13, code, tramp);
382 } else if (buf [4] == 0x90 || buf [5] == 0xeb || buf [6] == 0x66) {
383 /* Already changed by another thread */
386 printf ("Invalid trampoline sequence: %x %x %x %x %x %x %x\n", buf [0], buf [1], buf [2], buf [3],
387 buf [4], buf [5], buf [6]);
388 g_assert_not_reached ();
393 stack_unaligned (MonoTrampolineType tramp_type)
395 printf ("%d\n", tramp_type);
396 g_assert_not_reached ();
400 mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInfo **info, gboolean aot)
403 guint8 *buf, *code, *tramp, *br [2], *r11_save_code, *after_r11_save_code;
404 int i, lmf_offset, offset, res_offset, arg_offset, rax_offset, tramp_offset, ctx_offset, saved_regs_offset;
405 int saved_fpregs_offset, rbp_offset, framesize, orig_rsp_to_rbp_offset, cfa_offset;
407 GSList *unwind_ops = NULL;
408 MonoJumpInfo *ji = NULL;
409 const guint kMaxCodeSize = NACL_SIZE (600, 600*2);
411 #if defined(__native_client_codegen__)
412 const guint kNaClTrampOffset = 17;
415 if (tramp_type == MONO_TRAMPOLINE_JUMP || tramp_type == MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD)
420 code = buf = mono_global_codeman_reserve (kMaxCodeSize);
422 /* Compute stack frame size and offsets */
424 rbp_offset = -offset;
426 offset += sizeof(mgreg_t);
427 rax_offset = -offset;
429 offset += sizeof(mgreg_t);
430 tramp_offset = -offset;
432 offset += sizeof(gpointer);
433 arg_offset = -offset;
435 offset += sizeof(mgreg_t);
436 res_offset = -offset;
438 offset += sizeof (MonoContext);
439 ctx_offset = -offset;
440 saved_regs_offset = ctx_offset + MONO_STRUCT_OFFSET (MonoContext, gregs);
441 saved_fpregs_offset = ctx_offset + MONO_STRUCT_OFFSET (MonoContext, fregs);
443 offset += sizeof (MonoLMFTramp);
444 lmf_offset = -offset;
446 framesize = ALIGN_TO (offset, MONO_ARCH_FRAME_ALIGNMENT);
448 orig_rsp_to_rbp_offset = 0;
449 r11_save_code = code;
450 /* Reserve space for the mov_membase_reg to save R11 */
452 after_r11_save_code = code;
454 // CFA = sp + 16 (the trampoline address is on the stack)
456 mono_add_unwind_op_def_cfa (unwind_ops, code, buf, AMD64_RSP, 16);
457 // IP saved at CFA - 8
458 mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RIP, -8);
460 /* Pop the return address off the stack */
461 amd64_pop_reg (code, AMD64_R11);
462 orig_rsp_to_rbp_offset += sizeof(mgreg_t);
464 cfa_offset -= sizeof(mgreg_t);
465 mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
468 * Allocate a new stack frame
470 amd64_push_reg (code, AMD64_RBP);
471 cfa_offset += sizeof(mgreg_t);
472 mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
473 mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RBP, - cfa_offset);
475 orig_rsp_to_rbp_offset -= sizeof(mgreg_t);
476 amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, sizeof(mgreg_t));
477 mono_add_unwind_op_def_cfa_reg (unwind_ops, code, buf, AMD64_RBP);
478 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, framesize);
480 /* Compute the trampoline address from the return address */
482 #if defined(__default_codegen__)
483 /* 7 = length of call *<offset>(rip) */
484 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 7);
485 #elif defined(__native_client_codegen__)
486 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, kNaClTrampOffset);
489 /* 5 = length of amd64_call_membase () */
490 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 5);
492 amd64_mov_membase_reg (code, AMD64_RBP, tramp_offset, AMD64_R11, sizeof(gpointer));
494 /* Save all registers */
495 for (i = 0; i < AMD64_NREG; ++i) {
496 if (i == AMD64_RBP) {
497 /* RAX is already saved */
498 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, rbp_offset, sizeof(mgreg_t));
499 amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), AMD64_RAX, sizeof(mgreg_t));
500 } else if (i == AMD64_RIP) {
502 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, 8, sizeof(gpointer));
504 amd64_mov_reg_imm (code, AMD64_R11, 0);
505 amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), AMD64_R11, sizeof(mgreg_t));
506 } else if (i == AMD64_RSP) {
507 amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, sizeof(mgreg_t));
508 amd64_alu_reg_imm (code, X86_ADD, AMD64_R11, framesize + 16);
509 amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), AMD64_R11, sizeof(mgreg_t));
510 } else if (i != AMD64_R11) {
511 amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), i, sizeof(mgreg_t));
513 /* We have to save R11 right at the start of
514 the trampoline code because it's used as a
516 amd64_mov_membase_reg (r11_save_code, AMD64_RSP, saved_regs_offset + orig_rsp_to_rbp_offset + (i * sizeof(mgreg_t)), i, sizeof(mgreg_t));
517 g_assert (r11_save_code == after_r11_save_code);
520 for (i = 0; i < 8; ++i)
521 amd64_movsd_membase_reg (code, AMD64_RBP, saved_fpregs_offset + (i * sizeof(mgreg_t)), i);
523 /* Check that the stack is aligned */
524 #if defined(__default_codegen__)
525 amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, sizeof (mgreg_t));
526 amd64_alu_reg_imm (code, X86_AND, AMD64_R11, 15);
527 amd64_alu_reg_imm (code, X86_CMP, AMD64_R11, 0);
529 amd64_branch_disp (code, X86_CC_Z, 0, FALSE);
531 amd64_mov_reg_imm (code, AMD64_R11, 0);
532 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, 8);
534 amd64_mov_reg_imm (code, MONO_AMD64_ARG_REG1, tramp_type);
535 amd64_mov_reg_imm (code, AMD64_R11, stack_unaligned);
536 amd64_call_reg (code, AMD64_R11);
538 mono_amd64_patch (br [0], code);
539 //amd64_breakpoint (code);
542 if (tramp_type != MONO_TRAMPOLINE_GENERIC_CLASS_INIT &&
543 tramp_type != MONO_TRAMPOLINE_MONITOR_ENTER &&
544 tramp_type != MONO_TRAMPOLINE_MONITOR_ENTER_V4 &&
545 tramp_type != MONO_TRAMPOLINE_MONITOR_EXIT &&
546 tramp_type != MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD) {
547 /* Obtain the trampoline argument which is encoded in the instruction stream */
549 /* Load the GOT offset */
550 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, sizeof(gpointer));
551 #if defined(__default_codegen__)
553 * r11 points to a call *<offset>(%rip) instruction, load the
554 * pc-relative offset from the instruction itself.
556 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 3, 4);
557 /* 7 is the length of the call, 8 is the offset to the next got slot */
558 amd64_alu_reg_imm_size (code, X86_ADD, AMD64_RAX, 7 + sizeof (gpointer), sizeof(gpointer));
559 #elif defined(__native_client_codegen__)
560 /* The arg is hidden in a "push imm32" instruction, */
561 /* add one to skip the opcode. */
562 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, kNaClTrampOffset+1, 4);
564 /* Compute the address of the GOT slot */
565 amd64_alu_reg_reg_size (code, X86_ADD, AMD64_R11, AMD64_RAX, sizeof(gpointer));
567 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, sizeof(gpointer));
569 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, sizeof(gpointer));
570 #if defined(__default_codegen__)
571 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 5, 1);
572 amd64_widen_reg (code, AMD64_RAX, AMD64_RAX, TRUE, FALSE);
573 amd64_alu_reg_imm_size (code, X86_CMP, AMD64_RAX, 4, 1);
575 x86_branch8 (code, X86_CC_NE, 6, FALSE);
576 /* 32 bit immediate */
577 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 4);
579 x86_jump8 (code, 10);
580 /* 64 bit immediate */
581 mono_amd64_patch (br [0], code);
582 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 8);
583 mono_amd64_patch (br [1], code);
584 #elif defined(__native_client_codegen__)
585 /* All args are 32-bit pointers in NaCl */
586 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 4);
589 amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, sizeof(gpointer));
591 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, saved_regs_offset + (MONO_AMD64_ARG_REG1 * sizeof(mgreg_t)), sizeof(mgreg_t));
592 amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, sizeof(gpointer));
599 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, 8, sizeof(gpointer));
601 amd64_mov_reg_imm (code, AMD64_R11, 0);
602 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rip), AMD64_R11, sizeof(mgreg_t));
604 amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, sizeof(mgreg_t));
605 amd64_alu_reg_imm (code, X86_ADD, AMD64_R11, framesize + 16);
606 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, rsp), AMD64_R11, sizeof(mgreg_t));
607 /* Save pointer to context */
608 amd64_lea_membase (code, AMD64_R11, AMD64_RBP, ctx_offset);
609 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + MONO_STRUCT_OFFSET (MonoLMFTramp, ctx), AMD64_R11, sizeof(mgreg_t));
612 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_get_lmf_addr");
614 amd64_mov_reg_imm (code, AMD64_R11, mono_get_lmf_addr);
616 amd64_call_reg (code, AMD64_R11);
619 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + MONO_STRUCT_OFFSET (MonoLMFTramp, lmf_addr), AMD64_RAX, sizeof(gpointer));
620 /* Save previous_lmf */
621 /* Set the lowest bit to signal that this LMF has the ip field set */
622 /* Set the third lowest bit to signal that this is a MonoLMFTramp structure */
623 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RAX, 0, sizeof(gpointer));
624 amd64_alu_reg_imm_size (code, X86_ADD, AMD64_R11, 0x5, sizeof(gpointer));
625 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_R11, sizeof(gpointer));
627 amd64_lea_membase (code, AMD64_R11, AMD64_RBP, lmf_offset);
628 amd64_mov_membase_reg (code, AMD64_RAX, 0, AMD64_R11, sizeof(gpointer));
632 /* Arg1 is the pointer to the saved registers */
633 amd64_lea_membase (code, AMD64_ARG_REG1, AMD64_RBP, saved_regs_offset);
635 /* Arg2 is the address of the calling code */
637 amd64_mov_reg_membase (code, AMD64_ARG_REG2, AMD64_RBP, 8, sizeof(gpointer));
639 amd64_mov_reg_imm (code, AMD64_ARG_REG2, 0);
641 /* Arg3 is the method/vtable ptr */
642 amd64_mov_reg_membase (code, AMD64_ARG_REG3, AMD64_RBP, arg_offset, sizeof(gpointer));
644 /* Arg4 is the trampoline address */
645 amd64_mov_reg_membase (code, AMD64_ARG_REG4, AMD64_RBP, tramp_offset, sizeof(gpointer));
648 char *icall_name = g_strdup_printf ("trampoline_func_%d", tramp_type);
649 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, icall_name);
651 tramp = (guint8*)mono_get_trampoline_func (tramp_type);
652 amd64_mov_reg_imm (code, AMD64_R11, tramp);
654 amd64_call_reg (code, AMD64_R11);
656 /* Check for thread interruption */
657 /* This is not perf critical code so no need to check the interrupt flag */
659 * Have to call the _force_ variant, since there could be a protected wrapper on the top of the stack.
661 amd64_mov_membase_reg (code, AMD64_RBP, res_offset, AMD64_RAX, sizeof(mgreg_t));
663 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_thread_force_interruption_checkpoint");
665 amd64_mov_reg_imm (code, AMD64_R11, (guint8*)mono_thread_force_interruption_checkpoint);
667 amd64_call_reg (code, AMD64_R11);
669 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, res_offset, sizeof(mgreg_t));
672 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RBP, lmf_offset + MONO_STRUCT_OFFSET (MonoLMF, previous_lmf), sizeof(gpointer));
673 amd64_alu_reg_imm_size (code, X86_SUB, AMD64_RCX, 0x5, sizeof(gpointer));
674 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, lmf_offset + MONO_STRUCT_OFFSET (MonoLMFTramp, lmf_addr), sizeof(gpointer));
675 amd64_mov_membase_reg (code, AMD64_R11, 0, AMD64_RCX, sizeof(gpointer));
678 * Save rax to the stack, after the leave instruction, this will become part of
681 amd64_mov_membase_reg (code, AMD64_RBP, rax_offset, AMD64_RAX, sizeof(mgreg_t));
683 /* Restore argument registers, r10 (imt method/rgxtx)
684 and rax (needed for direct calls to C vararg functions). */
685 for (i = 0; i < AMD64_NREG; ++i)
686 if (AMD64_IS_ARGUMENT_REG (i) || i == AMD64_R10 || i == AMD64_RAX)
687 amd64_mov_reg_membase (code, i, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), sizeof(mgreg_t));
688 for (i = 0; i < 8; ++i)
689 amd64_movsd_reg_membase (code, i, AMD64_RBP, saved_fpregs_offset + (i * sizeof(mgreg_t)));
694 if (MONO_TRAMPOLINE_TYPE_MUST_RETURN (tramp_type)) {
696 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, rax_offset - sizeof(mgreg_t), sizeof(mgreg_t));
699 /* call the compiled method using the saved rax */
700 amd64_jump_membase (code, AMD64_RSP, rax_offset - sizeof(mgreg_t));
703 g_assert ((code - buf) <= kMaxCodeSize);
705 nacl_global_codeman_validate (&buf, kMaxCodeSize, &code);
707 mono_arch_flush_icache (buf, code - buf);
708 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
710 tramp_name = mono_get_generic_trampoline_name (tramp_type);
711 *info = mono_tramp_info_create (tramp_name, buf, code - buf, ji, unwind_ops);
718 mono_arch_get_nullified_class_init_trampoline (MonoTrampInfo **info)
721 int size = NACL_SIZE (16, 32);
723 code = buf = mono_global_codeman_reserve (size);
726 nacl_global_codeman_validate(&buf, size, &code);
728 mono_arch_flush_icache (buf, code - buf);
729 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
731 *info = mono_tramp_info_create ("nullified_class_init_trampoline", buf, code - buf, NULL, NULL);
737 mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len)
739 guint8 *code, *buf, *tramp;
741 gboolean far_addr = FALSE;
743 tramp = mono_get_trampoline_code (tramp_type);
745 #if defined(__default_codegen__)
746 if ((((guint64)arg1) >> 32) == 0)
751 code = buf = mono_domain_code_reserve_align (domain, size, 1);
753 if (((gint64)tramp - (gint64)code) >> 31 != 0 && ((gint64)tramp - (gint64)code) >> 31 != -1) {
754 #ifndef MONO_ARCH_NOMAP32BIT
755 g_assert_not_reached ();
759 code = buf = mono_domain_code_reserve_align (domain, size, 1);
761 #elif defined(__native_client_codegen__)
763 /* Aligning the call site below could */
764 /* add up to kNaClAlignment-1 bytes */
765 size += (kNaClAlignment-1);
766 size = NACL_BUNDLE_ALIGN_UP (size);
767 buf = mono_domain_code_reserve_align (domain, size, kNaClAlignment);
772 amd64_mov_reg_imm (code, AMD64_R11, tramp);
773 amd64_call_reg (code, AMD64_R11);
775 amd64_call_code (code, tramp);
777 /* The trampoline code will obtain the argument from the instruction stream */
778 #if defined(__default_codegen__)
779 if ((((guint64)arg1) >> 32) == 0) {
781 *(guint32*)(code + 1) = (gint64)arg1;
785 *(guint64*)(code + 1) = (gint64)arg1;
788 #elif defined(__native_client_codegen__)
789 /* For NaCl, all tramp args are 32-bit because they're pointers */
790 *code = 0x68; /* push imm32 */
791 *(guint32*)(code + 1) = (gint32)arg1;
795 g_assert ((code - buf) <= size);
800 nacl_domain_code_validate(domain, &buf, size, &code);
802 mono_arch_flush_icache (buf, size);
803 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_SPECIFIC_TRAMPOLINE, mono_get_generic_trampoline_simple_name (tramp_type));
809 mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot, MonoTrampInfo **info, gboolean aot)
813 guint8 **rgctx_null_jumps;
818 MonoJumpInfo *ji = NULL;
819 GSList *unwind_ops = NULL;
821 mrgctx = MONO_RGCTX_SLOT_IS_MRGCTX (slot);
822 index = MONO_RGCTX_SLOT_INDEX (slot);
824 index += MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT / sizeof (gpointer);
825 for (depth = 0; ; ++depth) {
826 int size = mono_class_rgctx_get_array_size (depth, mrgctx);
828 if (index < size - 1)
833 tramp_size = NACL_SIZE (64 + 8 * depth, 128 + 8 * depth);
835 code = buf = mono_global_codeman_reserve (tramp_size);
837 unwind_ops = mono_arch_get_cie_program ();
839 rgctx_null_jumps = g_malloc (sizeof (guint8*) * (depth + 2));
843 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
845 /* load rgctx ptr from vtable */
846 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_ARG_REG1, MONO_STRUCT_OFFSET (MonoVTable, runtime_generic_context), sizeof(gpointer));
847 /* is the rgctx ptr null? */
848 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
849 /* if yes, jump to actual trampoline */
850 rgctx_null_jumps [0] = code;
851 amd64_branch8 (code, X86_CC_Z, -1, 1);
854 for (i = 0; i < depth; ++i) {
855 /* load ptr to next array */
856 if (mrgctx && i == 0)
857 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT, sizeof(gpointer));
859 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, 0, sizeof(gpointer));
860 /* is the ptr null? */
861 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
862 /* if yes, jump to actual trampoline */
863 rgctx_null_jumps [i + 1] = code;
864 amd64_branch8 (code, X86_CC_Z, -1, 1);
868 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, sizeof (gpointer) * (index + 1), sizeof(gpointer));
869 /* is the slot null? */
870 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
871 /* if yes, jump to actual trampoline */
872 rgctx_null_jumps [depth + 1] = code;
873 amd64_branch8 (code, X86_CC_Z, -1, 1);
874 /* otherwise return */
877 for (i = mrgctx ? 1 : 0; i <= depth + 1; ++i)
878 mono_amd64_patch (rgctx_null_jumps [i], code);
880 g_free (rgctx_null_jumps);
882 /* move the rgctx pointer to the VTABLE register */
883 amd64_mov_reg_reg (code, MONO_ARCH_VTABLE_REG, AMD64_ARG_REG1, sizeof(gpointer));
886 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, g_strdup_printf ("specific_trampoline_lazy_fetch_%u", slot));
887 amd64_jump_reg (code, AMD64_R11);
889 tramp = mono_arch_create_specific_trampoline (GUINT_TO_POINTER (slot), MONO_TRAMPOLINE_RGCTX_LAZY_FETCH, mono_get_root_domain (), NULL);
891 /* jump to the actual trampoline */
892 amd64_jump_code (code, tramp);
895 nacl_global_codeman_validate (&buf, tramp_size, &code);
896 mono_arch_flush_icache (buf, code - buf);
897 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_GENERICS_TRAMPOLINE, NULL);
899 g_assert (code - buf <= tramp_size);
901 char *name = mono_get_rgctx_fetch_trampoline_name (slot);
902 *info = mono_tramp_info_create (name, buf, code - buf, ji, unwind_ops);
908 #ifdef MONO_ARCH_MONITOR_OBJECT_REG
911 mono_arch_create_monitor_enter_trampoline (MonoTrampInfo **info, gboolean is_v4, gboolean aot)
915 guint8 *jump_obj_null, *jump_sync_null, *jump_cmpxchg_failed, *jump_other_owner, *jump_tid, *jump_sync_thin_hash = NULL;
916 guint8 *jump_lock_taken_true = NULL;
918 int status_offset, nest_offset;
919 MonoJumpInfo *ji = NULL;
920 GSList *unwind_ops = NULL;
921 int obj_reg = MONO_AMD64_ARG_REG1;
922 int lock_taken_reg = MONO_AMD64_ARG_REG2;
923 int sync_reg = MONO_AMD64_ARG_REG3;
924 int tid_reg = MONO_AMD64_ARG_REG4;
925 int status_reg = AMD64_RAX;
927 g_assert (MONO_ARCH_MONITOR_OBJECT_REG == obj_reg);
928 #ifdef MONO_ARCH_MONITOR_LOCK_TAKEN_REG
929 g_assert (MONO_ARCH_MONITOR_LOCK_TAKEN_REG == lock_taken_reg);
934 mono_monitor_threads_sync_members_offset (&status_offset, &nest_offset);
935 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (status_offset) == sizeof (guint32));
936 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
937 status_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (status_offset);
938 nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
942 code = buf = mono_global_codeman_reserve (tramp_size);
944 unwind_ops = mono_arch_get_cie_program ();
946 if (!aot && mono_thread_get_tls_offset () != -1) {
947 /* MonoObject* obj is in obj_reg */
949 amd64_test_reg_reg (code, obj_reg, obj_reg);
950 /* if yes, jump to actual trampoline */
951 jump_obj_null = code;
952 amd64_branch8 (code, X86_CC_Z, -1, 1);
955 amd64_test_membase_imm (code, lock_taken_reg, 0, 1);
956 /* if *lock_taken is 1, jump to actual trampoline */
957 jump_lock_taken_true = code;
958 x86_branch8 (code, X86_CC_NZ, -1, 1);
961 /* load obj->synchronization to sync_reg */
962 amd64_mov_reg_membase (code, sync_reg, obj_reg, MONO_STRUCT_OFFSET (MonoObject, synchronisation), 8);
964 if (mono_gc_is_moving ()) {
965 /*if bit zero is set it's a thin hash*/
966 /*FIXME use testb encoding*/
967 amd64_test_reg_imm (code, sync_reg, 0x01);
968 jump_sync_thin_hash = code;
969 amd64_branch8 (code, X86_CC_NE, -1, 1);
971 /*clear bits used by the gc*/
972 amd64_alu_reg_imm (code, X86_AND, sync_reg, ~0x3);
975 /* is synchronization null? */
976 amd64_test_reg_reg (code, sync_reg, sync_reg);
977 /* if yes, jump to actual trampoline */
978 jump_sync_null = code;
979 amd64_branch8 (code, X86_CC_Z, -1, 1);
981 /* load MonoInternalThread* into tid_reg */
982 code = mono_amd64_emit_tls_get (code, tid_reg, mono_thread_get_tls_offset ());
983 /* load TID into tid_reg */
984 amd64_mov_reg_membase (code, tid_reg, tid_reg, MONO_STRUCT_OFFSET (MonoInternalThread, small_id), 4);
986 /* is synchronization->owner free */
987 amd64_mov_reg_membase (code, status_reg, sync_reg, status_offset, 4);
988 amd64_test_reg_imm_size (code, status_reg, OWNER_MASK, 4);
989 /* if not, jump to next case */
991 amd64_branch8 (code, X86_CC_NZ, -1, 1);
993 /* if yes, try a compare-exchange with the TID */
994 g_assert (tid_reg != X86_EAX);
995 /* Form new status in tid_reg */
996 amd64_alu_reg_reg_size (code, X86_OR, tid_reg, status_reg, 4);
997 /* compare and exchange */
998 amd64_prefix (code, X86_LOCK_PREFIX);
999 amd64_cmpxchg_membase_reg_size (code, sync_reg, status_offset, tid_reg, 4);
1000 /* if not successful, jump to actual trampoline */
1001 jump_cmpxchg_failed = code;
1002 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1003 /* if successful, return */
1005 amd64_mov_membase_imm (code, lock_taken_reg, 0, 1, 1);
1008 /* next case: synchronization->owner is not null */
1009 x86_patch (jump_tid, code);
1010 /* is synchronization->owner == TID? */
1011 amd64_alu_reg_imm_size (code, X86_AND, status_reg, OWNER_MASK, 4);
1012 amd64_alu_reg_reg_size (code, X86_CMP, status_reg, tid_reg, 4);
1013 /* if not, jump to actual trampoline */
1014 jump_other_owner = code;
1015 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1016 /* if yes, increment nest */
1017 amd64_inc_membase_size (code, sync_reg, nest_offset, 4);
1020 amd64_mov_membase_imm (code, lock_taken_reg, 0, 1, 1);
1023 x86_patch (jump_obj_null, code);
1024 if (jump_sync_thin_hash)
1025 x86_patch (jump_sync_thin_hash, code);
1026 x86_patch (jump_sync_null, code);
1027 x86_patch (jump_cmpxchg_failed, code);
1028 x86_patch (jump_other_owner, code);
1030 x86_patch (jump_lock_taken_true, code);
1033 /* jump to the actual trampoline */
1034 if (MONO_AMD64_ARG_REG1 != obj_reg)
1035 amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, obj_reg, sizeof (mgreg_t));
1039 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_enter_v4");
1041 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_enter");
1042 amd64_jump_reg (code, AMD64_R11);
1045 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_ENTER_V4, mono_get_root_domain (), NULL);
1047 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_ENTER, mono_get_root_domain (), NULL);
1049 /* jump to the actual trampoline */
1050 amd64_jump_code (code, tramp);
1053 nacl_global_codeman_validate (&buf, tramp_size, &code);
1055 mono_arch_flush_icache (code, code - buf);
1056 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_MONITOR, NULL);
1057 g_assert (code - buf <= tramp_size);
1060 *info = mono_tramp_info_create ("monitor_enter_v4_trampoline", buf, code - buf, ji, unwind_ops);
1062 *info = mono_tramp_info_create ("monitor_enter_trampoline", buf, code - buf, ji, unwind_ops);
1068 mono_arch_create_monitor_exit_trampoline (MonoTrampInfo **info, gboolean aot)
1072 guint8 *jump_obj_null, *jump_have_waiters, *jump_sync_null, *jump_not_owned, *jump_cmpxchg_failed;
1073 guint8 *jump_next, *jump_sync_thin_hash = NULL;
1075 int status_offset, nest_offset;
1076 MonoJumpInfo *ji = NULL;
1077 GSList *unwind_ops = NULL;
1078 int obj_reg = MONO_AMD64_ARG_REG1;
1079 int sync_reg = MONO_AMD64_ARG_REG2;
1080 int status_reg = MONO_AMD64_ARG_REG3;
1082 g_assert (obj_reg == MONO_ARCH_MONITOR_OBJECT_REG);
1084 mono_monitor_threads_sync_members_offset (&status_offset, &nest_offset);
1085 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (status_offset) == sizeof (guint32));
1086 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
1087 status_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (status_offset);
1088 nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
1092 code = buf = mono_global_codeman_reserve (tramp_size);
1094 unwind_ops = mono_arch_get_cie_program ();
1096 if (!aot && mono_thread_get_tls_offset () != -1) {
1097 /* MonoObject* obj is in obj_reg */
1099 amd64_test_reg_reg (code, obj_reg, obj_reg);
1100 /* if yes, jump to actual trampoline */
1101 jump_obj_null = code;
1102 amd64_branch8 (code, X86_CC_Z, -1, 1);
1104 /* load obj->synchronization to RCX */
1105 amd64_mov_reg_membase (code, sync_reg, obj_reg, MONO_STRUCT_OFFSET (MonoObject, synchronisation), 8);
1107 if (mono_gc_is_moving ()) {
1108 /*if bit zero is set it's a thin hash*/
1109 /*FIXME use testb encoding*/
1110 amd64_test_reg_imm (code, sync_reg, 0x01);
1111 jump_sync_thin_hash = code;
1112 amd64_branch8 (code, X86_CC_NE, -1, 1);
1114 /*clear bits used by the gc*/
1115 amd64_alu_reg_imm (code, X86_AND, sync_reg, ~0x3);
1118 /* is synchronization null? */
1119 amd64_test_reg_reg (code, sync_reg, sync_reg);
1120 /* if yes, jump to actual trampoline */
1121 jump_sync_null = code;
1122 amd64_branch8 (code, X86_CC_Z, -1, 1);
1124 /* next case: synchronization is not null */
1125 /* load MonoInternalThread* into RAX */
1126 code = mono_amd64_emit_tls_get (code, AMD64_RAX, mono_thread_get_tls_offset ());
1127 /* load TID into RAX */
1128 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, MONO_STRUCT_OFFSET (MonoInternalThread, small_id), 4);
1129 /* is synchronization->owner == TID */
1130 amd64_mov_reg_membase (code, status_reg, sync_reg, status_offset, 4);
1131 amd64_alu_reg_reg_size (code, X86_XOR, AMD64_RAX, status_reg, 4);
1132 amd64_test_reg_imm_size (code, AMD64_RAX, OWNER_MASK, 4);
1134 /* if no, jump to actual trampoline */
1135 jump_not_owned = code;
1136 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1138 /* next case: synchronization->owner == TID */
1139 /* is synchronization->nest == 1 */
1140 amd64_alu_membase_imm_size (code, X86_CMP, sync_reg, nest_offset, 1, 4);
1141 /* if not, jump to next case */
1143 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1144 /* if yes, is synchronization->entry_count greater than zero */
1145 amd64_test_reg_imm_size (code, status_reg, ENTRY_COUNT_WAITERS, 4);
1146 /* if not, jump to actual trampoline */
1147 jump_have_waiters = code;
1148 amd64_branch8 (code, X86_CC_NZ, -1 , 1);
1149 /* if yes, try to set synchronization->owner to null and return */
1150 g_assert (status_reg != AMD64_RAX);
1151 /* old status in RAX */
1152 amd64_mov_reg_reg (code, AMD64_RAX, status_reg, 4);
1153 /* form new status */
1154 amd64_alu_reg_imm_size (code, X86_AND, status_reg, ENTRY_COUNT_MASK, 4);
1155 /* compare and exchange */
1156 amd64_prefix (code, X86_LOCK_PREFIX);
1157 amd64_cmpxchg_membase_reg_size (code, sync_reg, status_offset, status_reg, 4);
1158 /* if not successful, jump to actual trampoline */
1159 jump_cmpxchg_failed = code;
1160 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1163 /* next case: synchronization->nest is not 1 */
1164 x86_patch (jump_next, code);
1165 /* decrease synchronization->nest and return */
1166 amd64_dec_membase_size (code, sync_reg, nest_offset, 4);
1169 if (jump_sync_thin_hash)
1170 x86_patch (jump_sync_thin_hash, code);
1171 x86_patch (jump_obj_null, code);
1172 x86_patch (jump_have_waiters, code);
1173 x86_patch (jump_not_owned, code);
1174 x86_patch (jump_cmpxchg_failed, code);
1175 x86_patch (jump_sync_null, code);
1178 /* jump to the actual trampoline */
1179 if (MONO_AMD64_ARG_REG1 != obj_reg)
1180 amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, obj_reg, sizeof (mgreg_t));
1183 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_exit");
1184 amd64_jump_reg (code, AMD64_R11);
1186 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_EXIT, mono_get_root_domain (), NULL);
1187 amd64_jump_code (code, tramp);
1190 nacl_global_codeman_validate (&buf, tramp_size, &code);
1192 mono_arch_flush_icache (code, code - buf);
1193 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_MONITOR, NULL);
1194 g_assert (code - buf <= tramp_size);
1196 *info = mono_tramp_info_create ("monitor_exit_trampoline", buf, code - buf, ji, unwind_ops);
1204 mono_arch_create_monitor_enter_trampoline (MonoTrampInfo **info, gboolean is_v4, gboolean aot)
1206 g_assert_not_reached ();
1211 mono_arch_create_monitor_exit_trampoline (MonoTrampInfo **info, gboolean aot)
1213 g_assert_not_reached ();
1220 mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
1222 /* FIXME: This is not thread safe */
1223 guint8 *code = ji->code_start;
1225 amd64_mov_reg_imm (code, AMD64_ARG_REG1, func_arg);
1226 amd64_mov_reg_imm (code, AMD64_R11, func);
1228 x86_push_imm (code, (guint64)func_arg);
1229 amd64_call_reg (code, AMD64_R11);
1234 handler_block_trampoline_helper (gpointer *ptr)
1236 MonoJitTlsData *jit_tls = mono_native_tls_get_value (mono_jit_tls_id);
1237 *ptr = jit_tls->handler_block_return_address;
1241 mono_arch_create_handler_block_trampoline (MonoTrampInfo **info, gboolean aot)
1243 guint8 *tramp = mono_get_trampoline_code (MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD);
1245 int tramp_size = 64;
1246 MonoJumpInfo *ji = NULL;
1247 GSList *unwind_ops = NULL;
1251 code = buf = mono_global_codeman_reserve (tramp_size);
1254 This trampoline restore the call chain of the handler block then jumps into the code that deals with it.
1256 if (mono_get_jit_tls_offset () != -1) {
1257 code = mono_amd64_emit_tls_get (code, MONO_AMD64_ARG_REG1, mono_get_jit_tls_offset ());
1258 amd64_mov_reg_membase (code, MONO_AMD64_ARG_REG1, MONO_AMD64_ARG_REG1, MONO_STRUCT_OFFSET (MonoJitTlsData, handler_block_return_address), 8);
1259 /* Simulate a call */
1260 amd64_push_reg (code, AMD64_RAX);
1261 amd64_jump_code (code, tramp);
1263 /*Slow path uses a c helper*/
1264 amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RSP, 8);
1265 amd64_mov_reg_imm (code, AMD64_RAX, tramp);
1266 amd64_push_reg (code, AMD64_RAX);
1267 amd64_push_reg (code, AMD64_RAX);
1268 amd64_jump_code (code, handler_block_trampoline_helper);
1271 mono_arch_flush_icache (buf, code - buf);
1272 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
1273 g_assert (code - buf <= tramp_size);
1275 *info = mono_tramp_info_create ("handler_block_trampoline", buf, code - buf, ji, unwind_ops);
1281 * mono_arch_get_call_target:
1283 * Return the address called by the code before CODE if exists.
1286 mono_arch_get_call_target (guint8 *code)
1288 if (code [-5] == 0xe8) {
1289 gint32 disp = *(gint32*)(code - 4);
1290 guint8 *target = code + disp;
1299 * mono_arch_get_plt_info_offset:
1301 * Return the PLT info offset belonging to the plt entry PLT_ENTRY.
1304 mono_arch_get_plt_info_offset (guint8 *plt_entry, mgreg_t *regs, guint8 *code)
1306 #if defined(__native_client__) || defined(__native_client_codegen__)
1307 /* 18 = 3 (mov opcode) + 4 (disp) + 10 (nacljmp) + 1 (push opcode) */
1308 /* See aot-compiler.c arch_emit_plt_entry for details. */
1309 return *(guint32*)(plt_entry + 18);
1311 return *(guint32*)(plt_entry + 6);
1316 * mono_arch_create_sdb_trampoline:
1318 * Return a trampoline which captures the current context, passes it to
1319 * debugger_agent_single_step_from_context ()/debugger_agent_breakpoint_from_context (),
1320 * then restores the (potentially changed) context.
1323 mono_arch_create_sdb_trampoline (gboolean single_step, MonoTrampInfo **info, gboolean aot)
1325 int tramp_size = 256;
1326 int i, framesize, ctx_offset, cfa_offset, gregs_offset;
1328 GSList *unwind_ops = NULL;
1329 MonoJumpInfo *ji = NULL;
1331 code = buf = mono_global_codeman_reserve (tramp_size);
1333 framesize = sizeof (MonoContext);
1334 framesize = ALIGN_TO (framesize, MONO_ARCH_FRAME_ALIGNMENT);
1338 mono_add_unwind_op_def_cfa (unwind_ops, code, buf, AMD64_RSP, 8);
1339 // IP saved at CFA - 8
1340 mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RIP, -cfa_offset);
1342 amd64_push_reg (code, AMD64_RBP);
1343 cfa_offset += sizeof(mgreg_t);
1344 mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
1345 mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RBP, - cfa_offset);
1347 amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, sizeof(mgreg_t));
1348 mono_add_unwind_op_def_cfa_reg (unwind_ops, code, buf, AMD64_RBP);
1349 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, framesize);
1352 gregs_offset = ctx_offset + MONO_STRUCT_OFFSET (MonoContext, gregs);
1354 /* Initialize a MonoContext structure on the stack */
1355 for (i = 0; i < AMD64_NREG; ++i) {
1356 if (i != AMD64_RIP && i != AMD64_RSP && i != AMD64_RBP)
1357 amd64_mov_membase_reg (code, AMD64_RSP, gregs_offset + (i * sizeof (mgreg_t)), i, sizeof (mgreg_t));
1359 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, 0, sizeof (mgreg_t));
1360 amd64_mov_membase_reg (code, AMD64_RSP, gregs_offset + (AMD64_RBP * sizeof (mgreg_t)), AMD64_R11, sizeof (mgreg_t));
1361 amd64_lea_membase (code, AMD64_R11, AMD64_RBP, 2 * sizeof (mgreg_t));
1362 amd64_mov_membase_reg (code, AMD64_RSP, gregs_offset + (AMD64_RSP * sizeof (mgreg_t)), AMD64_R11, sizeof (mgreg_t));
1363 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, sizeof (mgreg_t), sizeof (mgreg_t));
1364 amd64_mov_membase_reg (code, AMD64_RSP, gregs_offset + (AMD64_RIP * sizeof (mgreg_t)), AMD64_R11, sizeof (mgreg_t));
1366 /* Call the single step/breakpoint function in sdb */
1367 amd64_lea_membase (code, AMD64_ARG_REG1, AMD64_RSP, ctx_offset);
1371 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "debugger_agent_single_step_from_context");
1373 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "debugger_agent_breakpoint_from_context");
1374 amd64_call_reg (code, AMD64_R11);
1377 amd64_call_code (code, debugger_agent_single_step_from_context);
1379 amd64_call_code (code, debugger_agent_breakpoint_from_context);
1382 /* Restore registers from ctx */
1383 for (i = 0; i < AMD64_NREG; ++i) {
1384 if (i != AMD64_RIP && i != AMD64_RSP && i != AMD64_RBP)
1385 amd64_mov_reg_membase (code, AMD64_RSP, i, gregs_offset + (i * sizeof (mgreg_t)), sizeof (mgreg_t));
1387 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RSP, gregs_offset + (AMD64_RBP * sizeof (mgreg_t)), sizeof (mgreg_t));
1388 amd64_mov_membase_reg (code, AMD64_RBP, 0, AMD64_R11, sizeof (mgreg_t));
1389 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RSP, gregs_offset + (AMD64_RIP * sizeof (mgreg_t)), sizeof (mgreg_t));
1390 amd64_mov_membase_reg (code, AMD64_RBP, sizeof (mgreg_t), AMD64_R11, sizeof (mgreg_t));
1395 mono_arch_flush_icache (code, code - buf);
1396 g_assert (code - buf <= tramp_size);
1398 const char *tramp_name = single_step ? "sdb_single_step_trampoline" : "sdb_breakpoint_trampoline";
1399 *info = mono_tramp_info_create (tramp_name, buf, code - buf, ji, unwind_ops);