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/appdomain.h>
17 #include <mono/metadata/marshal.h>
18 #include <mono/metadata/tabledefs.h>
19 #include <mono/metadata/mono-debug-debugger.h>
20 #include <mono/metadata/monitor.h>
21 #include <mono/metadata/monitor.h>
22 #include <mono/metadata/gc-internal.h>
23 #include <mono/arch/amd64/amd64-codegen.h>
25 #include <mono/utils/memcheck.h>
28 #include "mini-amd64.h"
30 #if defined(__native_client_codegen__) && defined(__native_client__)
32 #include <nacl/nacl_dyncode.h>
35 #define IS_REX(inst) (((inst) >= 0x40) && ((inst) <= 0x4f))
38 * mono_arch_get_unbox_trampoline:
40 * @addr: pointer to native code for @m
42 * when value type methods are called through the vtable we need to unbox the
43 * this argument. This method returns a pointer to a trampoline which does
44 * unboxing before calling the method
47 mono_arch_get_unbox_trampoline (MonoMethod *m, gpointer addr)
50 int this_reg, size = NACL_SIZE (20, 32);
52 MonoDomain *domain = mono_domain_get ();
54 this_reg = mono_arch_get_this_arg_reg (NULL);
56 start = code = mono_domain_code_reserve (domain, size);
58 amd64_alu_reg_imm (code, X86_ADD, this_reg, sizeof (MonoObject));
59 /* FIXME: Optimize this */
60 amd64_mov_reg_imm (code, AMD64_RAX, addr);
61 amd64_jump_reg (code, AMD64_RAX);
62 g_assert ((code - start) < size);
64 nacl_domain_code_validate (domain, &start, size, &code);
66 mono_arch_flush_icache (start, code - start);
72 * mono_arch_get_static_rgctx_trampoline:
74 * Create a trampoline which sets RGCTX_REG to MRGCTX, then jumps to ADDR.
77 mono_arch_get_static_rgctx_trampoline (MonoMethod *m, MonoMethodRuntimeGenericContext *mrgctx, gpointer addr)
82 MonoDomain *domain = mono_domain_get ();
84 #ifdef MONO_ARCH_NOMAP32BIT
87 /* AOTed code could still have a non-32 bit address */
88 if ((((guint64)addr) >> 32) == 0)
89 buf_len = NACL_SIZE (16, 32);
91 buf_len = NACL_SIZE (30, 32);
94 start = code = mono_domain_code_reserve (domain, buf_len);
96 amd64_mov_reg_imm (code, MONO_ARCH_RGCTX_REG, mrgctx);
97 amd64_jump_code (code, addr);
98 g_assert ((code - start) < buf_len);
100 nacl_domain_code_validate (domain, &start, buf_len, &code);
101 mono_arch_flush_icache (start, code - start);
107 mono_arch_get_llvm_imt_trampoline (MonoDomain *domain, MonoMethod *m, int vt_offset)
109 guint8 *code, *start;
115 start = code = mono_domain_code_reserve (domain, buf_len);
117 this_reg = mono_arch_get_this_arg_reg (NULL);
120 amd64_mov_reg_imm (code, MONO_ARCH_IMT_REG, m);
121 /* Load vtable address */
122 amd64_mov_reg_membase (code, AMD64_RAX, this_reg, 0, 8);
123 amd64_jump_membase (code, AMD64_RAX, vt_offset);
126 g_assert ((code - start) < buf_len);
128 nacl_domain_code_validate (domain, &start, buf_len, &code);
130 mono_arch_flush_icache (start, code - start);
136 * mono_arch_patch_callsite:
138 * Patch the callsite whose address is given by ORIG_CODE so it calls ADDR. ORIG_CODE
139 * points to the pc right after the call.
142 mono_arch_patch_callsite (guint8 *method_start, guint8 *orig_code, guint8 *addr)
144 #if defined(__default_codegen__)
147 gboolean can_write = mono_breakpoint_clean_code (method_start, orig_code, 14, buf, sizeof (buf));
151 /* mov 64-bit imm into r11 (followed by call reg?) or direct call*/
152 if (((code [-13] == 0x49) && (code [-12] == 0xbb)) || (code [-5] == 0xe8)) {
153 if (code [-5] != 0xe8) {
155 InterlockedExchangePointer ((gpointer*)(orig_code - 11), addr);
156 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 11, sizeof (gpointer));
159 gboolean disp_32bit = ((((gint64)addr - (gint64)orig_code)) < (1 << 30)) && ((((gint64)addr - (gint64)orig_code)) > -(1 << 30));
161 if ((((guint64)(addr)) >> 32) != 0 && !disp_32bit) {
162 #ifdef MONO_ARCH_NOMAP32BIT
163 /* Print some diagnostics */
164 MonoJitInfo *ji = mono_jit_info_table_find (mono_domain_get (), (char*)orig_code);
166 fprintf (stderr, "At %s, offset 0x%zx\n", mono_method_full_name (jinfo_get_method (ji), TRUE), (guint8*)orig_code - (guint8*)ji->code_start);
167 fprintf (stderr, "Addr: %p\n", addr);
168 ji = mono_jit_info_table_find (mono_domain_get (), (char*)addr);
170 fprintf (stderr, "Callee: %s\n", mono_method_full_name (jinfo_get_method (ji), TRUE));
171 g_assert_not_reached ();
174 * This might happen when calling AOTed code. Create a thunk.
176 guint8 *thunk_start, *thunk_code;
178 thunk_start = thunk_code = mono_domain_code_reserve (mono_domain_get (), 32);
179 amd64_jump_membase (thunk_code, AMD64_RIP, 0);
180 *(guint64*)thunk_code = (guint64)addr;
182 g_assert ((((guint64)(addr)) >> 32) == 0);
183 mono_arch_flush_icache (thunk_start, thunk_code - thunk_start);
187 InterlockedExchange ((gint32*)(orig_code - 4), ((gint64)addr - (gint64)orig_code));
188 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, 4);
192 else if ((code [-7] == 0x41) && (code [-6] == 0xff) && (code [-5] == 0x15)) {
193 /* call *<OFFSET>(%rip) */
194 gpointer *got_entry = (gpointer*)((guint8*)orig_code + (*(guint32*)(orig_code - 4)));
196 InterlockedExchangePointer (got_entry, addr);
197 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, sizeof (gpointer));
200 #elif defined(__native_client__)
201 /* These are essentially the same 2 cases as above, modified for NaCl*/
203 /* Target must be bundle-aligned */
204 g_assert (((guint32)addr & kNaClAlignmentMask) == 0);
205 /* Return target must be bundle-aligned */
206 g_assert (((guint32)orig_code & kNaClAlignmentMask) == 0);
208 if (orig_code[-5] == 0xe8) {
211 gint32 offset = (gint32)addr - (gint32)orig_code;
212 guint8 buf[sizeof(gint32)];
213 *((gint32*)(buf)) = offset;
214 ret = nacl_dyncode_modify (orig_code - sizeof(gint32), buf, sizeof(gint32));
218 else if (is_nacl_call_reg_sequence (orig_code - 10) && orig_code[-16] == 0x41 && orig_code[-15] == 0xbb) {
220 guint8 buf[sizeof(gint32)];
221 *((gint32 *)(buf)) = addr;
222 /* orig_code[-14] is the start of the immediate. */
223 ret = nacl_dyncode_modify (orig_code - 14, buf, sizeof(gint32));
227 g_assert_not_reached ();
235 mono_arch_create_llvm_native_thunk (MonoDomain *domain, guint8 *addr)
238 * The caller is LLVM code and the call displacement might exceed 32 bits. We can't determine the caller address, so
239 * we add a thunk every time.
240 * Since the caller is also allocated using the domain code manager, hopefully the displacement will fit into 32 bits.
241 * FIXME: Avoid this if possible if !MONO_ARCH_NOMAP32BIT and ADDR is 32 bits.
243 guint8 *thunk_start, *thunk_code;
245 thunk_start = thunk_code = mono_domain_code_reserve (mono_domain_get (), 32);
246 amd64_jump_membase (thunk_code, AMD64_RIP, 0);
247 *(guint64*)thunk_code = (guint64)addr;
249 mono_arch_flush_icache (thunk_start, thunk_code - thunk_start);
254 mono_arch_patch_plt_entry (guint8 *code, gpointer *got, mgreg_t *regs, guint8 *addr)
257 gpointer *plt_jump_table_entry;
259 #if defined(__default_codegen__)
260 /* A PLT entry: jmp *<DISP>(%rip) */
261 g_assert (code [0] == 0xff);
262 g_assert (code [1] == 0x25);
264 disp = *(gint32*)(code + 2);
266 plt_jump_table_entry = (gpointer*)(code + 6 + disp);
267 #elif defined(__native_client_codegen__)
269 /* mov <DISP>(%rip), %r11d */
272 /* Verify the 'mov' */
273 g_assert (code [0] == 0x45);
274 g_assert (code [1] == 0x8b);
275 g_assert (code [2] == 0x1d);
277 disp = *(gint32*)(code + 3);
279 /* 7 = 3 (mov opcode) + 4 (disp) */
280 /* This needs to resolve to the target of the RIP-relative offset */
281 plt_jump_table_entry = (gpointer*)(code + 7 + disp);
283 #endif /* __native_client_codegen__ */
286 InterlockedExchangePointer (plt_jump_table_entry, addr);
290 get_vcall_slot (guint8 *code, mgreg_t *regs, int *displacement)
294 MonoJitInfo *ji = NULL;
297 /* code - 9 might be before the start of the method */
298 /* FIXME: Avoid this expensive call somehow */
299 ji = mono_jit_info_table_find (mono_domain_get (), (char*)code);
302 mono_breakpoint_clean_code (ji ? ji->code_start : NULL, code, 9, buf, sizeof (buf));
309 if ((code [0] == 0x41) && (code [1] == 0xff) && (code [2] == 0x15)) {
310 /* call OFFSET(%rip) */
311 g_assert_not_reached ();
312 *displacement = *(guint32*)(code + 3);
313 return (gpointer*)(code + disp + 7);
315 g_assert_not_reached ();
321 get_vcall_slot_addr (guint8* code, mgreg_t *regs)
325 vt = get_vcall_slot (code, regs, &displacement);
328 return (gpointer*)((char*)vt + displacement);
332 mono_arch_nullify_class_init_trampoline (guint8 *code, mgreg_t *regs)
335 MonoJitInfo *ji = NULL;
337 gpointer tramp = mini_get_nullified_class_init_trampoline ();
340 /* code - 7 might be before the start of the method */
341 /* FIXME: Avoid this expensive call somehow */
342 ji = mono_jit_info_table_find (mono_domain_get (), (char*)code);
345 can_write = mono_breakpoint_clean_code (ji ? ji->code_start : NULL, code, 7, buf, sizeof (buf));
351 * A given byte sequence can match more than case here, so we have to be
352 * really careful about the ordering of the cases. Longer sequences
355 if ((buf [0] == 0x41) && (buf [1] == 0xff) && (buf [2] == 0x15)) {
356 gpointer *vtable_slot;
358 /* call *<OFFSET>(%rip) */
359 vtable_slot = get_vcall_slot_addr (code, regs);
360 g_assert (vtable_slot);
362 *vtable_slot = tramp;
363 } else if (buf [2] == 0xe8) {
365 //guint8 *buf = code - 2;
368 * It would be better to replace the call with nops, but that doesn't seem
369 * to work on SMP machines even when the whole call is inside a cache line.
370 * Patching the call address seems to work.
380 mono_arch_patch_callsite (code - 5, code, tramp);
381 } else if ((buf [5] == 0xff) && x86_modrm_mod (buf [6]) == 3 && x86_modrm_reg (buf [6]) == 2) {
383 /* Generated by the LLVM JIT or on platforms without MAP_32BIT set */
384 mono_arch_patch_callsite (code - 13, code, tramp);
385 } else if (buf [4] == 0x90 || buf [5] == 0xeb || buf [6] == 0x66) {
386 /* Already changed by another thread */
389 printf ("Invalid trampoline sequence: %x %x %x %x %x %x %x\n", buf [0], buf [1], buf [2], buf [3],
390 buf [4], buf [5], buf [6]);
391 g_assert_not_reached ();
396 mono_arch_nullify_plt_entry (guint8 *code, mgreg_t *regs)
398 mono_arch_patch_plt_entry (code, NULL, regs, mini_get_nullified_class_init_trampoline ());
402 stack_unaligned (MonoTrampolineType tramp_type)
404 printf ("%d\n", tramp_type);
405 g_assert_not_reached ();
409 mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInfo **info, gboolean aot)
412 guint8 *buf, *code, *tramp, *br [2], *r11_save_code, *after_r11_save_code;
413 int i, lmf_offset, offset, res_offset, arg_offset, rax_offset, tramp_offset, saved_regs_offset;
414 int saved_fpregs_offset, rbp_offset, framesize, orig_rsp_to_rbp_offset, cfa_offset;
416 GSList *unwind_ops = NULL;
417 MonoJumpInfo *ji = NULL;
418 const guint kMaxCodeSize = NACL_SIZE (600, 600*2);
420 #if defined(__native_client_codegen__)
421 const guint kNaClTrampOffset = 17;
424 if (tramp_type == MONO_TRAMPOLINE_JUMP || tramp_type == MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD)
429 code = buf = mono_global_codeman_reserve (kMaxCodeSize);
431 framesize = kMaxCodeSize + sizeof (MonoLMFTramp);
432 framesize = (framesize + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
434 orig_rsp_to_rbp_offset = 0;
435 r11_save_code = code;
436 /* Reserve 5 bytes for the mov_membase_reg to save R11 */
438 after_r11_save_code = code;
440 // CFA = sp + 16 (the trampoline address is on the stack)
442 mono_add_unwind_op_def_cfa (unwind_ops, code, buf, AMD64_RSP, 16);
443 // IP saved at CFA - 8
444 mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RIP, -8);
446 /* Pop the return address off the stack */
447 amd64_pop_reg (code, AMD64_R11);
448 orig_rsp_to_rbp_offset += sizeof(mgreg_t);
450 cfa_offset -= sizeof(mgreg_t);
451 mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
454 * Allocate a new stack frame
456 amd64_push_reg (code, AMD64_RBP);
457 cfa_offset += sizeof(mgreg_t);
458 mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
459 mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RBP, - cfa_offset);
461 orig_rsp_to_rbp_offset -= sizeof(mgreg_t);
462 amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, sizeof(mgreg_t));
463 mono_add_unwind_op_def_cfa_reg (unwind_ops, code, buf, AMD64_RBP);
464 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, framesize);
467 rbp_offset = - offset;
469 offset += sizeof(mgreg_t);
470 rax_offset = - offset;
472 offset += sizeof(mgreg_t);
473 tramp_offset = - offset;
475 offset += sizeof(gpointer);
476 arg_offset = - offset;
478 /* Compute the trampoline address from the return address */
480 #if defined(__default_codegen__)
481 /* 7 = length of call *<offset>(rip) */
482 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 7);
483 #elif defined(__native_client_codegen__)
484 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, kNaClTrampOffset);
487 /* 5 = length of amd64_call_membase () */
488 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 5);
490 amd64_mov_membase_reg (code, AMD64_RBP, tramp_offset, AMD64_R11, sizeof(gpointer));
492 offset += sizeof(mgreg_t);
493 res_offset = - offset;
495 /* Save all registers */
497 offset += AMD64_NREG * sizeof(mgreg_t);
498 saved_regs_offset = - offset;
499 for (i = 0; i < AMD64_NREG; ++i) {
500 if (i == AMD64_RBP) {
501 /* RAX is already saved */
502 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, rbp_offset, sizeof(mgreg_t));
503 amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), AMD64_RAX, sizeof(mgreg_t));
504 } else if (i != AMD64_R11) {
505 amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), i, sizeof(mgreg_t));
507 /* We have to save R11 right at the start of
508 the trampoline code because it's used as a
510 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));
511 g_assert (r11_save_code == after_r11_save_code);
514 offset += 8 * sizeof(mgreg_t);
515 saved_fpregs_offset = - offset;
516 for (i = 0; i < 8; ++i)
517 amd64_movsd_membase_reg (code, AMD64_RBP, saved_fpregs_offset + (i * sizeof(mgreg_t)), i);
519 /* Check that the stack is aligned */
520 #if defined(__default_codegen__)
521 amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, sizeof (mgreg_t));
522 amd64_alu_reg_imm (code, X86_AND, AMD64_R11, 15);
523 amd64_alu_reg_imm (code, X86_CMP, AMD64_R11, 0);
525 amd64_branch_disp (code, X86_CC_Z, 0, FALSE);
527 amd64_mov_reg_imm (code, AMD64_R11, 0);
528 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, 8);
530 amd64_mov_reg_imm (code, AMD64_RDI, tramp_type);
531 amd64_mov_reg_imm (code, AMD64_R11, stack_unaligned);
532 amd64_call_reg (code, AMD64_R11);
534 mono_amd64_patch (br [0], code);
535 //amd64_breakpoint (code);
538 if (tramp_type != MONO_TRAMPOLINE_GENERIC_CLASS_INIT &&
539 tramp_type != MONO_TRAMPOLINE_MONITOR_ENTER &&
540 tramp_type != MONO_TRAMPOLINE_MONITOR_EXIT &&
541 tramp_type != MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD) {
542 /* Obtain the trampoline argument which is encoded in the instruction stream */
544 /* Load the GOT offset */
545 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, sizeof(gpointer));
546 #if defined(__default_codegen__)
547 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 7, 4);
548 #elif defined(__native_client_codegen__)
549 /* The arg is hidden in a "push imm32" instruction, */
550 /* add one to skip the opcode. */
551 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, kNaClTrampOffset+1, 4);
553 /* Compute the address of the GOT slot */
554 amd64_alu_reg_reg_size (code, X86_ADD, AMD64_R11, AMD64_RAX, sizeof(gpointer));
556 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, sizeof(gpointer));
558 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, sizeof(gpointer));
559 #if defined(__default_codegen__)
560 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 5, 1);
561 amd64_widen_reg (code, AMD64_RAX, AMD64_RAX, TRUE, FALSE);
562 amd64_alu_reg_imm_size (code, X86_CMP, AMD64_RAX, 4, 1);
564 x86_branch8 (code, X86_CC_NE, 6, FALSE);
565 /* 32 bit immediate */
566 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 4);
568 x86_jump8 (code, 10);
569 /* 64 bit immediate */
570 mono_amd64_patch (br [0], code);
571 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 8);
572 mono_amd64_patch (br [1], code);
573 #elif defined(__native_client_codegen__)
574 /* All args are 32-bit pointers in NaCl */
575 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 4);
578 amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, sizeof(gpointer));
580 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, saved_regs_offset + (MONO_AMD64_ARG_REG1 * sizeof(mgreg_t)), sizeof(mgreg_t));
581 amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, sizeof(gpointer));
586 offset += sizeof (MonoLMFTramp);
587 lmf_offset = - offset;
591 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, 8, sizeof(gpointer));
593 amd64_mov_reg_imm (code, AMD64_R11, 0);
594 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rip), AMD64_R11, sizeof(mgreg_t));
596 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RSP, framesize, sizeof(mgreg_t));
597 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbp), AMD64_R11, sizeof(mgreg_t));
599 amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, sizeof(mgreg_t));
600 amd64_alu_reg_imm (code, X86_ADD, AMD64_R11, framesize + 16);
601 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsp), AMD64_R11, sizeof(mgreg_t));
602 /* Save pointer to registers */
603 amd64_lea_membase (code, AMD64_R11, AMD64_RBP, saved_regs_offset);
604 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMFTramp, regs), AMD64_R11, sizeof(mgreg_t));
607 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_get_lmf_addr");
609 amd64_mov_reg_imm (code, AMD64_R11, mono_get_lmf_addr);
611 amd64_call_reg (code, AMD64_R11);
614 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMFTramp, lmf_addr), AMD64_RAX, sizeof(gpointer));
615 /* Save previous_lmf */
616 /* Set the lowest bit to signal that this LMF has the ip field set */
617 /* Set the third lowest bit to signal that this is a MonoLMFTramp structure */
618 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RAX, 0, sizeof(gpointer));
619 amd64_alu_reg_imm_size (code, X86_ADD, AMD64_R11, 0x5, sizeof(gpointer));
620 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_R11, sizeof(gpointer));
622 amd64_lea_membase (code, AMD64_R11, AMD64_RBP, lmf_offset);
623 amd64_mov_membase_reg (code, AMD64_RAX, 0, AMD64_R11, sizeof(gpointer));
627 /* Arg1 is the pointer to the saved registers */
628 amd64_lea_membase (code, AMD64_ARG_REG1, AMD64_RBP, saved_regs_offset);
630 /* Arg2 is the address of the calling code */
632 amd64_mov_reg_membase (code, AMD64_ARG_REG2, AMD64_RBP, 8, sizeof(gpointer));
634 amd64_mov_reg_imm (code, AMD64_ARG_REG2, 0);
636 /* Arg3 is the method/vtable ptr */
637 amd64_mov_reg_membase (code, AMD64_ARG_REG3, AMD64_RBP, arg_offset, sizeof(gpointer));
639 /* Arg4 is the trampoline address */
640 amd64_mov_reg_membase (code, AMD64_ARG_REG4, AMD64_RBP, tramp_offset, sizeof(gpointer));
643 char *icall_name = g_strdup_printf ("trampoline_func_%d", tramp_type);
644 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, icall_name);
646 tramp = (guint8*)mono_get_trampoline_func (tramp_type);
647 amd64_mov_reg_imm (code, AMD64_R11, tramp);
649 amd64_call_reg (code, AMD64_R11);
651 /* Check for thread interruption */
652 /* This is not perf critical code so no need to check the interrupt flag */
654 * Have to call the _force_ variant, since there could be a protected wrapper on the top of the stack.
656 amd64_mov_membase_reg (code, AMD64_RBP, res_offset, AMD64_RAX, sizeof(mgreg_t));
658 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_thread_force_interruption_checkpoint");
660 amd64_mov_reg_imm (code, AMD64_R11, (guint8*)mono_thread_force_interruption_checkpoint);
662 amd64_call_reg (code, AMD64_R11);
664 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, res_offset, sizeof(mgreg_t));
667 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), sizeof(gpointer));
668 amd64_alu_reg_imm_size (code, X86_SUB, AMD64_RCX, 0x5, sizeof(gpointer));
669 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMFTramp, lmf_addr), sizeof(gpointer));
670 amd64_mov_membase_reg (code, AMD64_R11, 0, AMD64_RCX, sizeof(gpointer));
673 * Save rax to the stack, after the leave instruction, this will become part of
676 amd64_mov_membase_reg (code, AMD64_RBP, rax_offset, AMD64_RAX, sizeof(mgreg_t));
678 /* Restore argument registers, r10 (imt method/rgxtx)
679 and rax (needed for direct calls to C vararg functions). */
680 for (i = 0; i < AMD64_NREG; ++i)
681 if (AMD64_IS_ARGUMENT_REG (i) || i == AMD64_R10 || i == AMD64_RAX)
682 amd64_mov_reg_membase (code, i, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), sizeof(mgreg_t));
684 for (i = 0; i < 8; ++i)
685 amd64_movsd_reg_membase (code, i, AMD64_RBP, saved_fpregs_offset + (i * sizeof(mgreg_t)));
690 if (MONO_TRAMPOLINE_TYPE_MUST_RETURN (tramp_type)) {
692 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, rax_offset - sizeof(mgreg_t), sizeof(mgreg_t));
695 /* call the compiled method using the saved rax */
696 amd64_jump_membase (code, AMD64_RSP, rax_offset - sizeof(mgreg_t));
699 g_assert ((code - buf) <= kMaxCodeSize);
701 nacl_global_codeman_validate (&buf, kMaxCodeSize, &code);
703 mono_arch_flush_icache (buf, code - buf);
706 tramp_name = mono_get_generic_trampoline_name (tramp_type);
707 *info = mono_tramp_info_create (tramp_name, buf, code - buf, ji, unwind_ops);
715 mono_arch_get_nullified_class_init_trampoline (MonoTrampInfo **info)
718 int size = NACL_SIZE (16, 32);
720 code = buf = mono_global_codeman_reserve (size);
723 nacl_global_codeman_validate(&buf, size, &code);
725 mono_arch_flush_icache (buf, code - buf);
728 *info = mono_tramp_info_create ("nullified_class_init_trampoline", buf, code - buf, NULL, NULL);
734 mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len)
736 guint8 *code, *buf, *tramp;
738 gboolean far_addr = FALSE;
740 tramp = mono_get_trampoline_code (tramp_type);
742 #if defined(__default_codegen__)
743 if ((((guint64)arg1) >> 32) == 0)
748 code = buf = mono_domain_code_reserve_align (domain, size, 1);
750 if (((gint64)tramp - (gint64)code) >> 31 != 0 && ((gint64)tramp - (gint64)code) >> 31 != -1) {
751 #ifndef MONO_ARCH_NOMAP32BIT
752 g_assert_not_reached ();
756 code = buf = mono_domain_code_reserve_align (domain, size, 1);
758 #elif defined(__native_client_codegen__)
760 /* Aligning the call site below could */
761 /* add up to kNaClAlignment-1 bytes */
762 size += (kNaClAlignment-1);
763 size = NACL_BUNDLE_ALIGN_UP (size);
764 buf = mono_domain_code_reserve_align (domain, size, kNaClAlignment);
769 amd64_mov_reg_imm (code, AMD64_R11, tramp);
770 amd64_call_reg (code, AMD64_R11);
772 amd64_call_code (code, tramp);
774 /* The trampoline code will obtain the argument from the instruction stream */
775 #if defined(__default_codegen__)
776 if ((((guint64)arg1) >> 32) == 0) {
778 *(guint32*)(code + 1) = (gint64)arg1;
782 *(guint64*)(code + 1) = (gint64)arg1;
785 #elif defined(__native_client_codegen__)
786 /* For NaCl, all tramp args are 32-bit because they're pointers */
787 *code = 0x68; /* push imm32 */
788 *(guint32*)(code + 1) = (gint32)arg1;
792 g_assert ((code - buf) <= size);
797 nacl_domain_code_validate(domain, &buf, size, &code);
799 mono_arch_flush_icache (buf, size);
805 mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot, MonoTrampInfo **info, gboolean aot)
809 guint8 **rgctx_null_jumps;
814 MonoJumpInfo *ji = NULL;
815 GSList *unwind_ops = NULL;
817 mrgctx = MONO_RGCTX_SLOT_IS_MRGCTX (slot);
818 index = MONO_RGCTX_SLOT_INDEX (slot);
820 index += MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT / sizeof (gpointer);
821 for (depth = 0; ; ++depth) {
822 int size = mono_class_rgctx_get_array_size (depth, mrgctx);
824 if (index < size - 1)
829 tramp_size = NACL_SIZE (64 + 8 * depth, 128 + 8 * depth);
831 code = buf = mono_global_codeman_reserve (tramp_size);
833 unwind_ops = mono_arch_get_cie_program ();
835 rgctx_null_jumps = g_malloc (sizeof (guint8*) * (depth + 2));
839 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
841 /* load rgctx ptr from vtable */
842 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_ARG_REG1, G_STRUCT_OFFSET (MonoVTable, runtime_generic_context), sizeof(gpointer));
843 /* is the rgctx ptr null? */
844 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
845 /* if yes, jump to actual trampoline */
846 rgctx_null_jumps [0] = code;
847 amd64_branch8 (code, X86_CC_Z, -1, 1);
850 for (i = 0; i < depth; ++i) {
851 /* load ptr to next array */
852 if (mrgctx && i == 0)
853 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT, sizeof(gpointer));
855 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, 0, sizeof(gpointer));
856 /* is the ptr null? */
857 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
858 /* if yes, jump to actual trampoline */
859 rgctx_null_jumps [i + 1] = code;
860 amd64_branch8 (code, X86_CC_Z, -1, 1);
864 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, sizeof (gpointer) * (index + 1), sizeof(gpointer));
865 /* is the slot null? */
866 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
867 /* if yes, jump to actual trampoline */
868 rgctx_null_jumps [depth + 1] = code;
869 amd64_branch8 (code, X86_CC_Z, -1, 1);
870 /* otherwise return */
873 for (i = mrgctx ? 1 : 0; i <= depth + 1; ++i)
874 mono_amd64_patch (rgctx_null_jumps [i], code);
876 g_free (rgctx_null_jumps);
878 /* move the rgctx pointer to the VTABLE register */
879 amd64_mov_reg_reg (code, MONO_ARCH_VTABLE_REG, AMD64_ARG_REG1, sizeof(gpointer));
882 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, g_strdup_printf ("specific_trampoline_lazy_fetch_%u", slot));
883 amd64_jump_reg (code, AMD64_R11);
885 tramp = mono_arch_create_specific_trampoline (GUINT_TO_POINTER (slot), MONO_TRAMPOLINE_RGCTX_LAZY_FETCH, mono_get_root_domain (), NULL);
887 /* jump to the actual trampoline */
888 amd64_jump_code (code, tramp);
891 nacl_global_codeman_validate (&buf, tramp_size, &code);
892 mono_arch_flush_icache (buf, code - buf);
894 g_assert (code - buf <= tramp_size);
897 char *name = mono_get_rgctx_fetch_trampoline_name (slot);
898 *info = mono_tramp_info_create (name, buf, code - buf, ji, unwind_ops);
906 mono_arch_create_generic_class_init_trampoline (MonoTrampInfo **info, gboolean aot)
910 static int byte_offset = -1;
911 static guint8 bitmask;
914 GSList *unwind_ops = NULL;
915 MonoJumpInfo *ji = NULL;
919 code = buf = mono_global_codeman_reserve (tramp_size);
922 mono_marshal_find_bitfield_offset (MonoVTable, initialized, &byte_offset, &bitmask);
924 amd64_test_membase_imm_size (code, MONO_AMD64_ARG_REG1, byte_offset, bitmask, 1);
926 amd64_branch8 (code, X86_CC_Z, -1, 1);
930 x86_patch (jump, code);
933 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_generic_class_init");
934 amd64_jump_reg (code, AMD64_R11);
936 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_GENERIC_CLASS_INIT, mono_get_root_domain (), NULL);
938 /* jump to the actual trampoline */
939 amd64_jump_code (code, tramp);
942 nacl_global_codeman_validate (&buf, tramp_size, &code);
944 mono_arch_flush_icache (buf, code - buf);
946 g_assert (code - buf <= tramp_size);
949 *info = mono_tramp_info_create ("generic_class_init_trampoline", buf, code - buf, ji, unwind_ops);
954 #ifdef MONO_ARCH_MONITOR_OBJECT_REG
957 mono_arch_create_monitor_enter_trampoline (MonoTrampInfo **info, gboolean aot)
961 guint8 *jump_obj_null, *jump_sync_null, *jump_cmpxchg_failed, *jump_other_owner, *jump_tid, *jump_sync_thin_hash = NULL;
963 int owner_offset, nest_offset, dummy;
964 MonoJumpInfo *ji = NULL;
965 GSList *unwind_ops = NULL;
967 g_assert (MONO_ARCH_MONITOR_OBJECT_REG == AMD64_RDI);
969 mono_monitor_threads_sync_members_offset (&owner_offset, &nest_offset, &dummy);
970 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (owner_offset) == sizeof (gpointer));
971 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
972 owner_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (owner_offset);
973 nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
977 code = buf = mono_global_codeman_reserve (tramp_size);
979 unwind_ops = mono_arch_get_cie_program ();
981 if (mono_thread_get_tls_offset () != -1) {
982 /* MonoObject* obj is in RDI */
984 amd64_test_reg_reg (code, AMD64_RDI, AMD64_RDI);
985 /* if yes, jump to actual trampoline */
986 jump_obj_null = code;
987 amd64_branch8 (code, X86_CC_Z, -1, 1);
989 /* load obj->synchronization to RCX */
990 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RDI, G_STRUCT_OFFSET (MonoObject, synchronisation), 8);
992 if (mono_gc_is_moving ()) {
993 /*if bit zero is set it's a thin hash*/
994 /*FIXME use testb encoding*/
995 amd64_test_reg_imm (code, AMD64_RCX, 0x01);
996 jump_sync_thin_hash = code;
997 amd64_branch8 (code, X86_CC_NE, -1, 1);
999 /*clear bits used by the gc*/
1000 amd64_alu_reg_imm (code, X86_AND, AMD64_RCX, ~0x3);
1003 /* is synchronization null? */
1004 amd64_test_reg_reg (code, AMD64_RCX, AMD64_RCX);
1005 /* if yes, jump to actual trampoline */
1006 jump_sync_null = code;
1007 amd64_branch8 (code, X86_CC_Z, -1, 1);
1009 /* load MonoInternalThread* into RDX */
1010 code = mono_amd64_emit_tls_get (code, AMD64_RDX, mono_thread_get_tls_offset ());
1011 /* load TID into RDX */
1012 amd64_mov_reg_membase (code, AMD64_RDX, AMD64_RDX, G_STRUCT_OFFSET (MonoInternalThread, tid), 8);
1014 /* is synchronization->owner null? */
1015 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, owner_offset, 0, 8);
1016 /* if not, jump to next case */
1018 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1020 /* if yes, try a compare-exchange with the TID */
1022 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
1023 /* compare and exchange */
1024 amd64_prefix (code, X86_LOCK_PREFIX);
1025 amd64_cmpxchg_membase_reg_size (code, AMD64_RCX, owner_offset, AMD64_RDX, 8);
1026 /* if not successful, jump to actual trampoline */
1027 jump_cmpxchg_failed = code;
1028 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1029 /* if successful, return */
1032 /* next case: synchronization->owner is not null */
1033 x86_patch (jump_tid, code);
1034 /* is synchronization->owner == TID? */
1035 amd64_alu_membase_reg_size (code, X86_CMP, AMD64_RCX, owner_offset, AMD64_RDX, 8);
1036 /* if not, jump to actual trampoline */
1037 jump_other_owner = code;
1038 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1039 /* if yes, increment nest */
1040 amd64_inc_membase_size (code, AMD64_RCX, nest_offset, 4);
1044 x86_patch (jump_obj_null, code);
1045 if (jump_sync_thin_hash)
1046 x86_patch (jump_sync_thin_hash, code);
1047 x86_patch (jump_sync_null, code);
1048 x86_patch (jump_cmpxchg_failed, code);
1049 x86_patch (jump_other_owner, code);
1052 /* jump to the actual trampoline */
1053 #if MONO_AMD64_ARG_REG1 != AMD64_RDI
1054 amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RDI);
1058 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_enter");
1059 amd64_jump_reg (code, AMD64_R11);
1061 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_ENTER, mono_get_root_domain (), NULL);
1063 /* jump to the actual trampoline */
1064 amd64_jump_code (code, tramp);
1067 nacl_global_codeman_validate (&buf, tramp_size, &code);
1069 mono_arch_flush_icache (code, code - buf);
1070 g_assert (code - buf <= tramp_size);
1073 *info = mono_tramp_info_create ("monitor_enter_trampoline", buf, code - buf, ji, unwind_ops);
1079 mono_arch_create_monitor_exit_trampoline (MonoTrampInfo **info, gboolean aot)
1083 guint8 *jump_obj_null, *jump_have_waiters, *jump_sync_null, *jump_not_owned, *jump_sync_thin_hash = NULL;
1086 int owner_offset, nest_offset, entry_count_offset;
1087 MonoJumpInfo *ji = NULL;
1088 GSList *unwind_ops = NULL;
1090 g_assert (MONO_ARCH_MONITOR_OBJECT_REG == AMD64_RDI);
1092 mono_monitor_threads_sync_members_offset (&owner_offset, &nest_offset, &entry_count_offset);
1093 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (owner_offset) == sizeof (gpointer));
1094 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
1095 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (entry_count_offset) == sizeof (gint32));
1096 owner_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (owner_offset);
1097 nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
1098 entry_count_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (entry_count_offset);
1102 code = buf = mono_global_codeman_reserve (tramp_size);
1104 unwind_ops = mono_arch_get_cie_program ();
1106 if (mono_thread_get_tls_offset () != -1) {
1107 /* MonoObject* obj is in RDI */
1109 amd64_test_reg_reg (code, AMD64_RDI, AMD64_RDI);
1110 /* if yes, jump to actual trampoline */
1111 jump_obj_null = code;
1112 amd64_branch8 (code, X86_CC_Z, -1, 1);
1114 /* load obj->synchronization to RCX */
1115 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RDI, G_STRUCT_OFFSET (MonoObject, synchronisation), 8);
1117 if (mono_gc_is_moving ()) {
1118 /*if bit zero is set it's a thin hash*/
1119 /*FIXME use testb encoding*/
1120 amd64_test_reg_imm (code, AMD64_RCX, 0x01);
1121 jump_sync_thin_hash = code;
1122 amd64_branch8 (code, X86_CC_NE, -1, 1);
1124 /*clear bits used by the gc*/
1125 amd64_alu_reg_imm (code, X86_AND, AMD64_RCX, ~0x3);
1128 /* is synchronization null? */
1129 amd64_test_reg_reg (code, AMD64_RCX, AMD64_RCX);
1130 /* if yes, jump to actual trampoline */
1131 jump_sync_null = code;
1132 amd64_branch8 (code, X86_CC_Z, -1, 1);
1134 /* next case: synchronization is not null */
1135 /* load MonoInternalThread* into RDX */
1136 code = mono_amd64_emit_tls_get (code, AMD64_RDX, mono_thread_get_tls_offset ());
1137 /* load TID into RDX */
1138 amd64_mov_reg_membase (code, AMD64_RDX, AMD64_RDX, G_STRUCT_OFFSET (MonoInternalThread, tid), 8);
1139 /* is synchronization->owner == TID */
1140 amd64_alu_membase_reg_size (code, X86_CMP, AMD64_RCX, owner_offset, AMD64_RDX, 8);
1141 /* if no, jump to actual trampoline */
1142 jump_not_owned = code;
1143 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1145 /* next case: synchronization->owner == TID */
1146 /* is synchronization->nest == 1 */
1147 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, nest_offset, 1, 4);
1148 /* if not, jump to next case */
1150 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1151 /* if yes, is synchronization->entry_count zero? */
1152 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, entry_count_offset, 0, 4);
1153 /* if not, jump to actual trampoline */
1154 jump_have_waiters = code;
1155 amd64_branch8 (code, X86_CC_NZ, -1 , 1);
1156 /* if yes, set synchronization->owner to null and return */
1157 amd64_mov_membase_imm (code, AMD64_RCX, owner_offset, 0, 8);
1160 /* next case: synchronization->nest is not 1 */
1161 x86_patch (jump_next, code);
1162 /* decrease synchronization->nest and return */
1163 amd64_dec_membase_size (code, AMD64_RCX, nest_offset, 4);
1166 x86_patch (jump_obj_null, code);
1167 x86_patch (jump_have_waiters, code);
1168 x86_patch (jump_not_owned, code);
1169 x86_patch (jump_sync_null, code);
1172 /* jump to the actual trampoline */
1173 #if MONO_AMD64_ARG_REG1 != AMD64_RDI
1174 amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RDI);
1178 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_exit");
1179 amd64_jump_reg (code, AMD64_R11);
1181 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_EXIT, mono_get_root_domain (), NULL);
1182 amd64_jump_code (code, tramp);
1185 nacl_global_codeman_validate (&buf, tramp_size, &code);
1187 mono_arch_flush_icache (code, code - buf);
1188 g_assert (code - buf <= tramp_size);
1191 *info = mono_tramp_info_create ("monitor_exit_trampoline", buf, code - buf, ji, unwind_ops);
1198 mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
1200 /* FIXME: This is not thread safe */
1201 guint8 *code = ji->code_start;
1203 amd64_mov_reg_imm (code, AMD64_ARG_REG1, func_arg);
1204 amd64_mov_reg_imm (code, AMD64_R11, func);
1206 x86_push_imm (code, (guint64)func_arg);
1207 amd64_call_reg (code, AMD64_R11);
1212 handler_block_trampoline_helper (gpointer *ptr)
1214 MonoJitTlsData *jit_tls = mono_native_tls_get_value (mono_jit_tls_id);
1215 *ptr = jit_tls->handler_block_return_address;
1219 mono_arch_create_handler_block_trampoline (MonoTrampInfo **info, gboolean aot)
1221 guint8 *tramp = mono_get_trampoline_code (MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD);
1223 int tramp_size = 64;
1224 MonoJumpInfo *ji = NULL;
1225 GSList *unwind_ops = NULL;
1229 code = buf = mono_global_codeman_reserve (tramp_size);
1232 This trampoline restore the call chain of the handler block then jumps into the code that deals with it.
1235 if (mono_get_jit_tls_offset () != -1) {
1236 code = mono_amd64_emit_tls_get (code, AMD64_RDI, mono_get_jit_tls_offset ());
1237 amd64_mov_reg_membase (code, AMD64_RDI, AMD64_RDI, G_STRUCT_OFFSET (MonoJitTlsData, handler_block_return_address), 8);
1238 /* Simulate a call */
1239 amd64_push_reg (code, AMD64_RAX);
1240 amd64_jump_code (code, tramp);
1242 /*Slow path uses a c helper*/
1243 amd64_mov_reg_reg (code, AMD64_RDI, AMD64_RSP, 8);
1244 amd64_mov_reg_imm (code, AMD64_RAX, tramp);
1245 amd64_push_reg (code, AMD64_RAX);
1246 amd64_jump_code (code, handler_block_trampoline_helper);
1249 mono_arch_flush_icache (buf, code - buf);
1250 g_assert (code - buf <= tramp_size);
1253 *info = mono_tramp_info_create ("handler_block_trampoline", buf, code - buf, ji, unwind_ops);
1259 * mono_arch_get_call_target:
1261 * Return the address called by the code before CODE if exists.
1264 mono_arch_get_call_target (guint8 *code)
1266 if (code [-5] == 0xe8) {
1267 guint32 disp = *(guint32*)(code - 4);
1268 guint8 *target = code + disp;
1277 * mono_arch_get_plt_info_offset:
1279 * Return the PLT info offset belonging to the plt entry PLT_ENTRY.
1282 mono_arch_get_plt_info_offset (guint8 *plt_entry, mgreg_t *regs, guint8 *code)
1284 #if defined(__native_client__) || defined(__native_client_codegen__)
1285 /* 18 = 3 (mov opcode) + 4 (disp) + 10 (nacljmp) + 1 (push opcode) */
1286 /* See aot-compiler.c arch_emit_plt_entry for details. */
1287 return *(guint32*)(plt_entry + 18);
1289 return *(guint32*)(plt_entry + 6);