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 (MonoLMF);
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 (MonoLMF);
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));
603 if (tramp_type == MONO_TRAMPOLINE_JIT || tramp_type == MONO_TRAMPOLINE_JUMP) {
604 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, arg_offset, sizeof(gpointer));
605 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), AMD64_R11, sizeof(gpointer));
607 amd64_mov_membase_imm (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), 0, sizeof(gpointer));
609 /* Save callee saved regs */
611 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rdi), AMD64_RDI, sizeof(mgreg_t));
612 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsi), AMD64_RSI, sizeof(mgreg_t));
614 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbx), AMD64_RBX, sizeof(mgreg_t));
615 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r12), AMD64_R12, sizeof(mgreg_t));
616 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r13), AMD64_R13, sizeof(mgreg_t));
617 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r14), AMD64_R14, sizeof(mgreg_t));
618 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r15), AMD64_R15, sizeof(mgreg_t));
621 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_get_lmf_addr");
623 amd64_mov_reg_imm (code, AMD64_R11, mono_get_lmf_addr);
625 amd64_call_reg (code, AMD64_R11);
628 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), AMD64_RAX, sizeof(gpointer));
629 /* Save previous_lmf */
630 /* Set the lowest bit to 1 to signal that this LMF has the ip field set */
631 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RAX, 0, sizeof(gpointer));
632 amd64_alu_reg_imm_size (code, X86_ADD, AMD64_R11, 1, sizeof(gpointer));
633 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_R11, sizeof(gpointer));
635 amd64_lea_membase (code, AMD64_R11, AMD64_RBP, lmf_offset);
636 amd64_mov_membase_reg (code, AMD64_RAX, 0, AMD64_R11, sizeof(gpointer));
640 /* Arg1 is the pointer to the saved registers */
641 amd64_lea_membase (code, AMD64_ARG_REG1, AMD64_RBP, saved_regs_offset);
643 /* Arg2 is the address of the calling code */
645 amd64_mov_reg_membase (code, AMD64_ARG_REG2, AMD64_RBP, 8, sizeof(gpointer));
647 amd64_mov_reg_imm (code, AMD64_ARG_REG2, 0);
649 /* Arg3 is the method/vtable ptr */
650 amd64_mov_reg_membase (code, AMD64_ARG_REG3, AMD64_RBP, arg_offset, sizeof(gpointer));
652 /* Arg4 is the trampoline address */
653 amd64_mov_reg_membase (code, AMD64_ARG_REG4, AMD64_RBP, tramp_offset, sizeof(gpointer));
656 char *icall_name = g_strdup_printf ("trampoline_func_%d", tramp_type);
657 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, icall_name);
659 tramp = (guint8*)mono_get_trampoline_func (tramp_type);
660 amd64_mov_reg_imm (code, AMD64_R11, tramp);
662 amd64_call_reg (code, AMD64_R11);
664 /* Check for thread interruption */
665 /* This is not perf critical code so no need to check the interrupt flag */
667 * Have to call the _force_ variant, since there could be a protected wrapper on the top of the stack.
669 amd64_mov_membase_reg (code, AMD64_RBP, res_offset, AMD64_RAX, sizeof(mgreg_t));
671 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_thread_force_interruption_checkpoint");
673 amd64_mov_reg_imm (code, AMD64_R11, (guint8*)mono_thread_force_interruption_checkpoint);
675 amd64_call_reg (code, AMD64_R11);
677 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, res_offset, sizeof(mgreg_t));
681 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), sizeof(gpointer));
682 amd64_alu_reg_imm_size (code, X86_SUB, AMD64_RCX, 1, sizeof(gpointer));
683 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), sizeof(gpointer));
684 amd64_mov_membase_reg (code, AMD64_R11, 0, AMD64_RCX, sizeof(gpointer));
687 * Save rax to the stack, after the leave instruction, this will become part of
690 amd64_mov_membase_reg (code, AMD64_RBP, rax_offset, AMD64_RAX, sizeof(mgreg_t));
692 /* Restore argument registers, r10 (imt method/rgxtx)
693 and rax (needed for direct calls to C vararg functions). */
694 for (i = 0; i < AMD64_NREG; ++i)
695 if (AMD64_IS_ARGUMENT_REG (i) || i == AMD64_R10 || i == AMD64_RAX)
696 amd64_mov_reg_membase (code, i, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), sizeof(mgreg_t));
698 for (i = 0; i < 8; ++i)
699 amd64_movsd_reg_membase (code, i, AMD64_RBP, saved_fpregs_offset + (i * sizeof(mgreg_t)));
704 if (MONO_TRAMPOLINE_TYPE_MUST_RETURN (tramp_type)) {
706 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, rax_offset - sizeof(mgreg_t), sizeof(mgreg_t));
709 /* call the compiled method using the saved rax */
710 amd64_jump_membase (code, AMD64_RSP, rax_offset - sizeof(mgreg_t));
713 g_assert ((code - buf) <= kMaxCodeSize);
715 nacl_global_codeman_validate (&buf, kMaxCodeSize, &code);
717 mono_arch_flush_icache (buf, code - buf);
720 tramp_name = mono_get_generic_trampoline_name (tramp_type);
721 *info = mono_tramp_info_create (tramp_name, buf, code - buf, ji, unwind_ops);
729 mono_arch_get_nullified_class_init_trampoline (MonoTrampInfo **info)
732 int size = NACL_SIZE (16, 32);
734 code = buf = mono_global_codeman_reserve (size);
737 nacl_global_codeman_validate(&buf, size, &code);
739 mono_arch_flush_icache (buf, code - buf);
742 *info = mono_tramp_info_create ("nullified_class_init_trampoline", buf, code - buf, NULL, NULL);
744 if (mono_jit_map_is_enabled ())
745 mono_emit_jit_tramp (buf, code - buf, "nullified_class_init_trampoline");
751 mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len)
753 guint8 *code, *buf, *tramp;
755 gboolean far_addr = FALSE;
757 tramp = mono_get_trampoline_code (tramp_type);
759 #if defined(__default_codegen__)
760 if ((((guint64)arg1) >> 32) == 0)
765 code = buf = mono_domain_code_reserve_align (domain, size, 1);
767 if (((gint64)tramp - (gint64)code) >> 31 != 0 && ((gint64)tramp - (gint64)code) >> 31 != -1) {
768 #ifndef MONO_ARCH_NOMAP32BIT
769 g_assert_not_reached ();
773 code = buf = mono_domain_code_reserve_align (domain, size, 1);
775 #elif defined(__native_client_codegen__)
777 /* Aligning the call site below could */
778 /* add up to kNaClAlignment-1 bytes */
779 size += (kNaClAlignment-1);
780 size = NACL_BUNDLE_ALIGN_UP (size);
781 buf = mono_domain_code_reserve_align (domain, size, kNaClAlignment);
786 amd64_mov_reg_imm (code, AMD64_R11, tramp);
787 amd64_call_reg (code, AMD64_R11);
789 amd64_call_code (code, tramp);
791 /* The trampoline code will obtain the argument from the instruction stream */
792 #if defined(__default_codegen__)
793 if ((((guint64)arg1) >> 32) == 0) {
795 *(guint32*)(code + 1) = (gint64)arg1;
799 *(guint64*)(code + 1) = (gint64)arg1;
802 #elif defined(__native_client_codegen__)
803 /* For NaCl, all tramp args are 32-bit because they're pointers */
804 *code = 0x68; /* push imm32 */
805 *(guint32*)(code + 1) = (gint32)arg1;
809 g_assert ((code - buf) <= size);
814 nacl_domain_code_validate(domain, &buf, size, &code);
816 mono_arch_flush_icache (buf, size);
822 mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot, MonoTrampInfo **info, gboolean aot)
826 guint8 **rgctx_null_jumps;
831 MonoJumpInfo *ji = NULL;
832 GSList *unwind_ops = NULL;
834 mrgctx = MONO_RGCTX_SLOT_IS_MRGCTX (slot);
835 index = MONO_RGCTX_SLOT_INDEX (slot);
837 index += MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT / sizeof (gpointer);
838 for (depth = 0; ; ++depth) {
839 int size = mono_class_rgctx_get_array_size (depth, mrgctx);
841 if (index < size - 1)
846 tramp_size = NACL_SIZE (64 + 8 * depth, 128 + 8 * depth);
848 code = buf = mono_global_codeman_reserve (tramp_size);
850 unwind_ops = mono_arch_get_cie_program ();
852 rgctx_null_jumps = g_malloc (sizeof (guint8*) * (depth + 2));
856 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
858 /* load rgctx ptr from vtable */
859 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_ARG_REG1, G_STRUCT_OFFSET (MonoVTable, runtime_generic_context), sizeof(gpointer));
860 /* is the rgctx ptr null? */
861 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
862 /* if yes, jump to actual trampoline */
863 rgctx_null_jumps [0] = code;
864 amd64_branch8 (code, X86_CC_Z, -1, 1);
867 for (i = 0; i < depth; ++i) {
868 /* load ptr to next array */
869 if (mrgctx && i == 0)
870 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT, sizeof(gpointer));
872 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, 0, sizeof(gpointer));
873 /* is the ptr null? */
874 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
875 /* if yes, jump to actual trampoline */
876 rgctx_null_jumps [i + 1] = code;
877 amd64_branch8 (code, X86_CC_Z, -1, 1);
881 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, sizeof (gpointer) * (index + 1), sizeof(gpointer));
882 /* is the slot null? */
883 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
884 /* if yes, jump to actual trampoline */
885 rgctx_null_jumps [depth + 1] = code;
886 amd64_branch8 (code, X86_CC_Z, -1, 1);
887 /* otherwise return */
890 for (i = mrgctx ? 1 : 0; i <= depth + 1; ++i)
891 mono_amd64_patch (rgctx_null_jumps [i], code);
893 g_free (rgctx_null_jumps);
895 /* move the rgctx pointer to the VTABLE register */
896 amd64_mov_reg_reg (code, MONO_ARCH_VTABLE_REG, AMD64_ARG_REG1, sizeof(gpointer));
899 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, g_strdup_printf ("specific_trampoline_lazy_fetch_%u", slot));
900 amd64_jump_reg (code, AMD64_R11);
902 tramp = mono_arch_create_specific_trampoline (GUINT_TO_POINTER (slot), MONO_TRAMPOLINE_RGCTX_LAZY_FETCH, mono_get_root_domain (), NULL);
904 /* jump to the actual trampoline */
905 amd64_jump_code (code, tramp);
908 nacl_global_codeman_validate (&buf, tramp_size, &code);
909 mono_arch_flush_icache (buf, code - buf);
911 g_assert (code - buf <= tramp_size);
914 char *name = mono_get_rgctx_fetch_trampoline_name (slot);
915 *info = mono_tramp_info_create (name, buf, code - buf, ji, unwind_ops);
923 mono_arch_create_generic_class_init_trampoline (MonoTrampInfo **info, gboolean aot)
927 static int byte_offset = -1;
928 static guint8 bitmask;
931 GSList *unwind_ops = NULL;
932 MonoJumpInfo *ji = NULL;
936 code = buf = mono_global_codeman_reserve (tramp_size);
939 mono_marshal_find_bitfield_offset (MonoVTable, initialized, &byte_offset, &bitmask);
941 amd64_test_membase_imm_size (code, MONO_AMD64_ARG_REG1, byte_offset, bitmask, 1);
943 amd64_branch8 (code, X86_CC_Z, -1, 1);
947 x86_patch (jump, code);
950 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_generic_class_init");
951 amd64_jump_reg (code, AMD64_R11);
953 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_GENERIC_CLASS_INIT, mono_get_root_domain (), NULL);
955 /* jump to the actual trampoline */
956 amd64_jump_code (code, tramp);
959 nacl_global_codeman_validate (&buf, tramp_size, &code);
961 mono_arch_flush_icache (buf, code - buf);
963 g_assert (code - buf <= tramp_size);
966 *info = mono_tramp_info_create ("generic_class_init_trampoline", buf, code - buf, ji, unwind_ops);
971 #ifdef MONO_ARCH_MONITOR_OBJECT_REG
974 mono_arch_create_monitor_enter_trampoline (MonoTrampInfo **info, gboolean aot)
978 guint8 *jump_obj_null, *jump_sync_null, *jump_cmpxchg_failed, *jump_other_owner, *jump_tid, *jump_sync_thin_hash = NULL;
980 int owner_offset, nest_offset, dummy;
981 MonoJumpInfo *ji = NULL;
982 GSList *unwind_ops = NULL;
984 g_assert (MONO_ARCH_MONITOR_OBJECT_REG == AMD64_RDI);
986 mono_monitor_threads_sync_members_offset (&owner_offset, &nest_offset, &dummy);
987 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (owner_offset) == sizeof (gpointer));
988 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
989 owner_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (owner_offset);
990 nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
994 code = buf = mono_global_codeman_reserve (tramp_size);
996 unwind_ops = mono_arch_get_cie_program ();
998 if (mono_thread_get_tls_offset () != -1) {
999 /* MonoObject* obj is in RDI */
1001 amd64_test_reg_reg (code, AMD64_RDI, AMD64_RDI);
1002 /* if yes, jump to actual trampoline */
1003 jump_obj_null = code;
1004 amd64_branch8 (code, X86_CC_Z, -1, 1);
1006 /* load obj->synchronization to RCX */
1007 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RDI, G_STRUCT_OFFSET (MonoObject, synchronisation), 8);
1009 if (mono_gc_is_moving ()) {
1010 /*if bit zero is set it's a thin hash*/
1011 /*FIXME use testb encoding*/
1012 amd64_test_reg_imm (code, AMD64_RCX, 0x01);
1013 jump_sync_thin_hash = code;
1014 amd64_branch8 (code, X86_CC_NE, -1, 1);
1016 /*clear bits used by the gc*/
1017 amd64_alu_reg_imm (code, X86_AND, AMD64_RCX, ~0x3);
1020 /* is synchronization null? */
1021 amd64_test_reg_reg (code, AMD64_RCX, AMD64_RCX);
1022 /* if yes, jump to actual trampoline */
1023 jump_sync_null = code;
1024 amd64_branch8 (code, X86_CC_Z, -1, 1);
1026 /* load MonoInternalThread* into RDX */
1027 code = mono_amd64_emit_tls_get (code, AMD64_RDX, mono_thread_get_tls_offset ());
1028 /* load TID into RDX */
1029 amd64_mov_reg_membase (code, AMD64_RDX, AMD64_RDX, G_STRUCT_OFFSET (MonoInternalThread, tid), 8);
1031 /* is synchronization->owner null? */
1032 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, owner_offset, 0, 8);
1033 /* if not, jump to next case */
1035 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1037 /* if yes, try a compare-exchange with the TID */
1039 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
1040 /* compare and exchange */
1041 amd64_prefix (code, X86_LOCK_PREFIX);
1042 amd64_cmpxchg_membase_reg_size (code, AMD64_RCX, owner_offset, AMD64_RDX, 8);
1043 /* if not successful, jump to actual trampoline */
1044 jump_cmpxchg_failed = code;
1045 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1046 /* if successful, return */
1049 /* next case: synchronization->owner is not null */
1050 x86_patch (jump_tid, code);
1051 /* is synchronization->owner == TID? */
1052 amd64_alu_membase_reg_size (code, X86_CMP, AMD64_RCX, owner_offset, AMD64_RDX, 8);
1053 /* if not, jump to actual trampoline */
1054 jump_other_owner = code;
1055 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1056 /* if yes, increment nest */
1057 amd64_inc_membase_size (code, AMD64_RCX, nest_offset, 4);
1061 x86_patch (jump_obj_null, code);
1062 if (jump_sync_thin_hash)
1063 x86_patch (jump_sync_thin_hash, code);
1064 x86_patch (jump_sync_null, code);
1065 x86_patch (jump_cmpxchg_failed, code);
1066 x86_patch (jump_other_owner, code);
1069 /* jump to the actual trampoline */
1070 #if MONO_AMD64_ARG_REG1 != AMD64_RDI
1071 amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RDI);
1075 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_enter");
1076 amd64_jump_reg (code, AMD64_R11);
1078 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_ENTER, mono_get_root_domain (), NULL);
1080 /* jump to the actual trampoline */
1081 amd64_jump_code (code, tramp);
1084 nacl_global_codeman_validate (&buf, tramp_size, &code);
1086 mono_arch_flush_icache (code, code - buf);
1087 g_assert (code - buf <= tramp_size);
1090 *info = mono_tramp_info_create ("monitor_enter_trampoline", buf, code - buf, ji, unwind_ops);
1096 mono_arch_create_monitor_exit_trampoline (MonoTrampInfo **info, gboolean aot)
1100 guint8 *jump_obj_null, *jump_have_waiters, *jump_sync_null, *jump_not_owned, *jump_sync_thin_hash = NULL;
1103 int owner_offset, nest_offset, entry_count_offset;
1104 MonoJumpInfo *ji = NULL;
1105 GSList *unwind_ops = NULL;
1107 g_assert (MONO_ARCH_MONITOR_OBJECT_REG == AMD64_RDI);
1109 mono_monitor_threads_sync_members_offset (&owner_offset, &nest_offset, &entry_count_offset);
1110 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (owner_offset) == sizeof (gpointer));
1111 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
1112 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (entry_count_offset) == sizeof (gint32));
1113 owner_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (owner_offset);
1114 nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
1115 entry_count_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (entry_count_offset);
1119 code = buf = mono_global_codeman_reserve (tramp_size);
1121 unwind_ops = mono_arch_get_cie_program ();
1123 if (mono_thread_get_tls_offset () != -1) {
1124 /* MonoObject* obj is in RDI */
1126 amd64_test_reg_reg (code, AMD64_RDI, AMD64_RDI);
1127 /* if yes, jump to actual trampoline */
1128 jump_obj_null = code;
1129 amd64_branch8 (code, X86_CC_Z, -1, 1);
1131 /* load obj->synchronization to RCX */
1132 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RDI, G_STRUCT_OFFSET (MonoObject, synchronisation), 8);
1134 if (mono_gc_is_moving ()) {
1135 /*if bit zero is set it's a thin hash*/
1136 /*FIXME use testb encoding*/
1137 amd64_test_reg_imm (code, AMD64_RCX, 0x01);
1138 jump_sync_thin_hash = code;
1139 amd64_branch8 (code, X86_CC_NE, -1, 1);
1141 /*clear bits used by the gc*/
1142 amd64_alu_reg_imm (code, X86_AND, AMD64_RCX, ~0x3);
1145 /* is synchronization null? */
1146 amd64_test_reg_reg (code, AMD64_RCX, AMD64_RCX);
1147 /* if yes, jump to actual trampoline */
1148 jump_sync_null = code;
1149 amd64_branch8 (code, X86_CC_Z, -1, 1);
1151 /* next case: synchronization is not null */
1152 /* load MonoInternalThread* into RDX */
1153 code = mono_amd64_emit_tls_get (code, AMD64_RDX, mono_thread_get_tls_offset ());
1154 /* load TID into RDX */
1155 amd64_mov_reg_membase (code, AMD64_RDX, AMD64_RDX, G_STRUCT_OFFSET (MonoInternalThread, tid), 8);
1156 /* is synchronization->owner == TID */
1157 amd64_alu_membase_reg_size (code, X86_CMP, AMD64_RCX, owner_offset, AMD64_RDX, 8);
1158 /* if no, jump to actual trampoline */
1159 jump_not_owned = code;
1160 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1162 /* next case: synchronization->owner == TID */
1163 /* is synchronization->nest == 1 */
1164 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, nest_offset, 1, 4);
1165 /* if not, jump to next case */
1167 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1168 /* if yes, is synchronization->entry_count zero? */
1169 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, entry_count_offset, 0, 4);
1170 /* if not, jump to actual trampoline */
1171 jump_have_waiters = code;
1172 amd64_branch8 (code, X86_CC_NZ, -1 , 1);
1173 /* if yes, set synchronization->owner to null and return */
1174 amd64_mov_membase_imm (code, AMD64_RCX, owner_offset, 0, 8);
1177 /* next case: synchronization->nest is not 1 */
1178 x86_patch (jump_next, code);
1179 /* decrease synchronization->nest and return */
1180 amd64_dec_membase_size (code, AMD64_RCX, nest_offset, 4);
1183 x86_patch (jump_obj_null, code);
1184 x86_patch (jump_have_waiters, code);
1185 x86_patch (jump_not_owned, code);
1186 x86_patch (jump_sync_null, code);
1189 /* jump to the actual trampoline */
1190 #if MONO_AMD64_ARG_REG1 != AMD64_RDI
1191 amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RDI);
1195 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_exit");
1196 amd64_jump_reg (code, AMD64_R11);
1198 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_EXIT, mono_get_root_domain (), NULL);
1199 amd64_jump_code (code, tramp);
1202 nacl_global_codeman_validate (&buf, tramp_size, &code);
1204 mono_arch_flush_icache (code, code - buf);
1205 g_assert (code - buf <= tramp_size);
1208 *info = mono_tramp_info_create ("monitor_exit_trampoline", buf, code - buf, ji, unwind_ops);
1215 mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
1217 /* FIXME: This is not thread safe */
1218 guint8 *code = ji->code_start;
1220 amd64_mov_reg_imm (code, AMD64_ARG_REG1, func_arg);
1221 amd64_mov_reg_imm (code, AMD64_R11, func);
1223 x86_push_imm (code, (guint64)func_arg);
1224 amd64_call_reg (code, AMD64_R11);
1229 handler_block_trampoline_helper (gpointer *ptr)
1231 MonoJitTlsData *jit_tls = mono_native_tls_get_value (mono_jit_tls_id);
1232 *ptr = jit_tls->handler_block_return_address;
1236 mono_arch_create_handler_block_trampoline (void)
1238 guint8 *tramp = mono_get_trampoline_code (MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD);
1240 int tramp_size = 64;
1241 code = buf = mono_global_codeman_reserve (tramp_size);
1244 This trampoline restore the call chain of the handler block then jumps into the code that deals with it.
1247 if (mono_get_jit_tls_offset () != -1) {
1248 code = mono_amd64_emit_tls_get (code, AMD64_RDI, mono_get_jit_tls_offset ());
1249 amd64_mov_reg_membase (code, AMD64_RDI, AMD64_RDI, G_STRUCT_OFFSET (MonoJitTlsData, handler_block_return_address), 8);
1250 /* Simulate a call */
1251 amd64_push_reg (code, AMD64_RAX);
1252 amd64_jump_code (code, tramp);
1254 /*Slow path uses a c helper*/
1255 amd64_mov_reg_reg (code, AMD64_RDI, AMD64_RSP, 8);
1256 amd64_mov_reg_imm (code, AMD64_RAX, tramp);
1257 amd64_push_reg (code, AMD64_RAX);
1258 amd64_jump_code (code, handler_block_trampoline_helper);
1261 mono_arch_flush_icache (buf, code - buf);
1262 g_assert (code - buf <= tramp_size);
1264 if (mono_jit_map_is_enabled ())
1265 mono_emit_jit_tramp (buf, code - buf, "handler_block_trampoline");
1271 * mono_arch_get_call_target:
1273 * Return the address called by the code before CODE if exists.
1276 mono_arch_get_call_target (guint8 *code)
1278 if (code [-5] == 0xe8) {
1279 guint32 disp = *(guint32*)(code - 4);
1280 guint8 *target = code + disp;
1289 * mono_arch_get_plt_info_offset:
1291 * Return the PLT info offset belonging to the plt entry PLT_ENTRY.
1294 mono_arch_get_plt_info_offset (guint8 *plt_entry, mgreg_t *regs, guint8 *code)
1296 #if defined(__native_client__) || defined(__native_client_codegen__)
1297 /* 18 = 3 (mov opcode) + 4 (disp) + 10 (nacljmp) + 1 (push opcode) */
1298 /* See aot-compiler.c arch_emit_plt_entry for details. */
1299 return *(guint32*)(plt_entry + 18);
1301 return *(guint32*)(plt_entry + 6);