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"
31 #if defined(__native_client_codegen__) && defined(__native_client__)
33 #include <nacl/nacl_dyncode.h>
36 #define IS_REX(inst) (((inst) >= 0x40) && ((inst) <= 0x4f))
39 * mono_arch_get_unbox_trampoline:
41 * @addr: pointer to native code for @m
43 * when value type methods are called through the vtable we need to unbox the
44 * this argument. This method returns a pointer to a trampoline which does
45 * unboxing before calling the method
48 mono_arch_get_unbox_trampoline (MonoMethod *m, gpointer addr)
51 int this_reg, size = NACL_SIZE (20, 32);
53 MonoDomain *domain = mono_domain_get ();
55 this_reg = mono_arch_get_this_arg_reg (NULL);
57 start = code = mono_domain_code_reserve (domain, size);
59 amd64_alu_reg_imm (code, X86_ADD, this_reg, sizeof (MonoObject));
60 /* FIXME: Optimize this */
61 amd64_mov_reg_imm (code, AMD64_RAX, addr);
62 amd64_jump_reg (code, AMD64_RAX);
63 g_assert ((code - start) < size);
65 nacl_domain_code_validate (domain, &start, size, &code);
67 mono_arch_flush_icache (start, code - start);
68 mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_UNBOX_TRAMPOLINE, m);
74 * mono_arch_get_static_rgctx_trampoline:
76 * Create a trampoline which sets RGCTX_REG to MRGCTX, then jumps to ADDR.
79 mono_arch_get_static_rgctx_trampoline (MonoMethod *m, MonoMethodRuntimeGenericContext *mrgctx, gpointer addr)
84 MonoDomain *domain = mono_domain_get ();
86 #ifdef MONO_ARCH_NOMAP32BIT
89 /* AOTed code could still have a non-32 bit address */
90 if ((((guint64)addr) >> 32) == 0)
91 buf_len = NACL_SIZE (16, 32);
93 buf_len = NACL_SIZE (30, 32);
96 start = code = mono_domain_code_reserve (domain, buf_len);
98 amd64_mov_reg_imm (code, MONO_ARCH_RGCTX_REG, mrgctx);
99 amd64_jump_code (code, addr);
100 g_assert ((code - start) < buf_len);
102 nacl_domain_code_validate (domain, &start, buf_len, &code);
103 mono_arch_flush_icache (start, code - start);
104 mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_GENERICS_TRAMPOLINE, NULL);
110 mono_arch_get_llvm_imt_trampoline (MonoDomain *domain, MonoMethod *m, int vt_offset)
112 guint8 *code, *start;
118 start = code = mono_domain_code_reserve (domain, buf_len);
120 this_reg = mono_arch_get_this_arg_reg (NULL);
123 amd64_mov_reg_imm (code, MONO_ARCH_IMT_REG, m);
124 /* Load vtable address */
125 amd64_mov_reg_membase (code, AMD64_RAX, this_reg, 0, 8);
126 amd64_jump_membase (code, AMD64_RAX, vt_offset);
129 g_assert ((code - start) < buf_len);
131 nacl_domain_code_validate (domain, &start, buf_len, &code);
133 mono_arch_flush_icache (start, code - start);
134 mono_profiler_code_buffer_new (start, code - start, MONO_PROFILER_CODE_BUFFER_IMT_TRAMPOLINE, NULL);
140 * mono_arch_patch_callsite:
142 * Patch the callsite whose address is given by ORIG_CODE so it calls ADDR. ORIG_CODE
143 * points to the pc right after the call.
146 mono_arch_patch_callsite (guint8 *method_start, guint8 *orig_code, guint8 *addr)
148 #if defined(__default_codegen__)
151 gboolean can_write = mono_breakpoint_clean_code (method_start, orig_code, 14, buf, sizeof (buf));
155 /* mov 64-bit imm into r11 (followed by call reg?) or direct call*/
156 if (((code [-13] == 0x49) && (code [-12] == 0xbb)) || (code [-5] == 0xe8)) {
157 if (code [-5] != 0xe8) {
159 InterlockedExchangePointer ((gpointer*)(orig_code - 11), addr);
160 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 11, sizeof (gpointer));
163 gboolean disp_32bit = ((((gint64)addr - (gint64)orig_code)) < (1 << 30)) && ((((gint64)addr - (gint64)orig_code)) > -(1 << 30));
165 if ((((guint64)(addr)) >> 32) != 0 && !disp_32bit) {
166 #ifdef MONO_ARCH_NOMAP32BIT
167 /* Print some diagnostics */
168 MonoJitInfo *ji = mono_jit_info_table_find (mono_domain_get (), (char*)orig_code);
170 fprintf (stderr, "At %s, offset 0x%zx\n", mono_method_full_name (jinfo_get_method (ji), TRUE), (guint8*)orig_code - (guint8*)ji->code_start);
171 fprintf (stderr, "Addr: %p\n", addr);
172 ji = mono_jit_info_table_find (mono_domain_get (), (char*)addr);
174 fprintf (stderr, "Callee: %s\n", mono_method_full_name (jinfo_get_method (ji), TRUE));
175 g_assert_not_reached ();
178 * This might happen when calling AOTed code. Create a thunk.
180 guint8 *thunk_start, *thunk_code;
182 thunk_start = thunk_code = mono_domain_code_reserve (mono_domain_get (), 32);
183 amd64_jump_membase (thunk_code, AMD64_RIP, 0);
184 *(guint64*)thunk_code = (guint64)addr;
186 g_assert ((((guint64)(addr)) >> 32) == 0);
187 mono_arch_flush_icache (thunk_start, thunk_code - thunk_start);
188 mono_profiler_code_buffer_new (thunk_start, thunk_code - thunk_start, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
192 InterlockedExchange ((gint32*)(orig_code - 4), ((gint64)addr - (gint64)orig_code));
193 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, 4);
197 else if ((code [-7] == 0x41) && (code [-6] == 0xff) && (code [-5] == 0x15)) {
198 /* call *<OFFSET>(%rip) */
199 gpointer *got_entry = (gpointer*)((guint8*)orig_code + (*(guint32*)(orig_code - 4)));
201 InterlockedExchangePointer (got_entry, addr);
202 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, sizeof (gpointer));
205 #elif defined(__native_client__)
206 /* These are essentially the same 2 cases as above, modified for NaCl*/
208 /* Target must be bundle-aligned */
209 g_assert (((guint32)addr & kNaClAlignmentMask) == 0);
210 /* Return target must be bundle-aligned */
211 g_assert (((guint32)orig_code & kNaClAlignmentMask) == 0);
213 if (orig_code[-5] == 0xe8) {
216 gint32 offset = (gint32)addr - (gint32)orig_code;
217 guint8 buf[sizeof(gint32)];
218 *((gint32*)(buf)) = offset;
219 ret = nacl_dyncode_modify (orig_code - sizeof(gint32), buf, sizeof(gint32));
223 else if (is_nacl_call_reg_sequence (orig_code - 10) && orig_code[-16] == 0x41 && orig_code[-15] == 0xbb) {
225 guint8 buf[sizeof(gint32)];
226 *((gint32 *)(buf)) = addr;
227 /* orig_code[-14] is the start of the immediate. */
228 ret = nacl_dyncode_modify (orig_code - 14, buf, sizeof(gint32));
232 g_assert_not_reached ();
240 mono_arch_create_llvm_native_thunk (MonoDomain *domain, guint8 *addr)
243 * The caller is LLVM code and the call displacement might exceed 32 bits. We can't determine the caller address, so
244 * we add a thunk every time.
245 * Since the caller is also allocated using the domain code manager, hopefully the displacement will fit into 32 bits.
246 * FIXME: Avoid this if possible if !MONO_ARCH_NOMAP32BIT and ADDR is 32 bits.
248 guint8 *thunk_start, *thunk_code;
250 thunk_start = thunk_code = mono_domain_code_reserve (mono_domain_get (), 32);
251 amd64_jump_membase (thunk_code, AMD64_RIP, 0);
252 *(guint64*)thunk_code = (guint64)addr;
254 mono_arch_flush_icache (thunk_start, thunk_code - thunk_start);
255 mono_profiler_code_buffer_new (thunk_start, thunk_code - thunk_start, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
260 mono_arch_patch_plt_entry (guint8 *code, gpointer *got, mgreg_t *regs, guint8 *addr)
263 gpointer *plt_jump_table_entry;
265 #if defined(__default_codegen__)
266 /* A PLT entry: jmp *<DISP>(%rip) */
267 g_assert (code [0] == 0xff);
268 g_assert (code [1] == 0x25);
270 disp = *(gint32*)(code + 2);
272 plt_jump_table_entry = (gpointer*)(code + 6 + disp);
273 #elif defined(__native_client_codegen__)
275 /* mov <DISP>(%rip), %r11d */
278 /* Verify the 'mov' */
279 g_assert (code [0] == 0x45);
280 g_assert (code [1] == 0x8b);
281 g_assert (code [2] == 0x1d);
283 disp = *(gint32*)(code + 3);
285 /* 7 = 3 (mov opcode) + 4 (disp) */
286 /* This needs to resolve to the target of the RIP-relative offset */
287 plt_jump_table_entry = (gpointer*)(code + 7 + disp);
289 #endif /* __native_client_codegen__ */
292 InterlockedExchangePointer (plt_jump_table_entry, addr);
296 get_vcall_slot (guint8 *code, mgreg_t *regs, int *displacement)
300 MonoJitInfo *ji = NULL;
303 /* code - 9 might be before the start of the method */
304 /* FIXME: Avoid this expensive call somehow */
305 ji = mono_jit_info_table_find (mono_domain_get (), (char*)code);
308 mono_breakpoint_clean_code (ji ? ji->code_start : NULL, code, 9, buf, sizeof (buf));
315 if ((code [0] == 0x41) && (code [1] == 0xff) && (code [2] == 0x15)) {
316 /* call OFFSET(%rip) */
317 g_assert_not_reached ();
318 *displacement = *(guint32*)(code + 3);
319 return (gpointer*)(code + disp + 7);
321 g_assert_not_reached ();
327 get_vcall_slot_addr (guint8* code, mgreg_t *regs)
331 vt = get_vcall_slot (code, regs, &displacement);
334 return (gpointer*)((char*)vt + displacement);
338 mono_arch_nullify_class_init_trampoline (guint8 *code, mgreg_t *regs)
341 MonoJitInfo *ji = NULL;
343 gpointer tramp = mini_get_nullified_class_init_trampoline ();
346 /* code - 7 might be before the start of the method */
347 /* FIXME: Avoid this expensive call somehow */
348 ji = mono_jit_info_table_find (mono_domain_get (), (char*)code);
351 can_write = mono_breakpoint_clean_code (ji ? ji->code_start : NULL, code, 7, buf, sizeof (buf));
357 * A given byte sequence can match more than case here, so we have to be
358 * really careful about the ordering of the cases. Longer sequences
361 if ((buf [0] == 0x41) && (buf [1] == 0xff) && (buf [2] == 0x15)) {
362 gpointer *vtable_slot;
364 /* call *<OFFSET>(%rip) */
365 vtable_slot = get_vcall_slot_addr (code, regs);
366 g_assert (vtable_slot);
368 *vtable_slot = tramp;
369 } else if (buf [2] == 0xe8) {
371 //guint8 *buf = code - 2;
374 * It would be better to replace the call with nops, but that doesn't seem
375 * to work on SMP machines even when the whole call is inside a cache line.
376 * Patching the call address seems to work.
386 mono_arch_patch_callsite (code - 5, code, tramp);
387 } else if ((buf [5] == 0xff) && x86_modrm_mod (buf [6]) == 3 && x86_modrm_reg (buf [6]) == 2) {
389 /* Generated by the LLVM JIT or on platforms without MAP_32BIT set */
390 mono_arch_patch_callsite (code - 13, code, tramp);
391 } else if (buf [4] == 0x90 || buf [5] == 0xeb || buf [6] == 0x66) {
392 /* Already changed by another thread */
395 printf ("Invalid trampoline sequence: %x %x %x %x %x %x %x\n", buf [0], buf [1], buf [2], buf [3],
396 buf [4], buf [5], buf [6]);
397 g_assert_not_reached ();
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, MONO_AMD64_ARG_REG1, 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 + MONO_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 + MONO_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 + MONO_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 + MONO_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 + MONO_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 + MONO_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 + MONO_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 + MONO_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);
704 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
707 tramp_name = mono_get_generic_trampoline_name (tramp_type);
708 *info = mono_tramp_info_create (tramp_name, buf, code - buf, ji, unwind_ops);
716 mono_arch_get_nullified_class_init_trampoline (MonoTrampInfo **info)
719 int size = NACL_SIZE (16, 32);
721 code = buf = mono_global_codeman_reserve (size);
724 nacl_global_codeman_validate(&buf, size, &code);
726 mono_arch_flush_icache (buf, code - buf);
727 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
730 *info = mono_tramp_info_create ("nullified_class_init_trampoline", buf, code - buf, NULL, NULL);
736 mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len)
738 guint8 *code, *buf, *tramp;
740 gboolean far_addr = FALSE;
742 tramp = mono_get_trampoline_code (tramp_type);
744 #if defined(__default_codegen__)
745 if ((((guint64)arg1) >> 32) == 0)
750 code = buf = mono_domain_code_reserve_align (domain, size, 1);
752 if (((gint64)tramp - (gint64)code) >> 31 != 0 && ((gint64)tramp - (gint64)code) >> 31 != -1) {
753 #ifndef MONO_ARCH_NOMAP32BIT
754 g_assert_not_reached ();
758 code = buf = mono_domain_code_reserve_align (domain, size, 1);
760 #elif defined(__native_client_codegen__)
762 /* Aligning the call site below could */
763 /* add up to kNaClAlignment-1 bytes */
764 size += (kNaClAlignment-1);
765 size = NACL_BUNDLE_ALIGN_UP (size);
766 buf = mono_domain_code_reserve_align (domain, size, kNaClAlignment);
771 amd64_mov_reg_imm (code, AMD64_R11, tramp);
772 amd64_call_reg (code, AMD64_R11);
774 amd64_call_code (code, tramp);
776 /* The trampoline code will obtain the argument from the instruction stream */
777 #if defined(__default_codegen__)
778 if ((((guint64)arg1) >> 32) == 0) {
780 *(guint32*)(code + 1) = (gint64)arg1;
784 *(guint64*)(code + 1) = (gint64)arg1;
787 #elif defined(__native_client_codegen__)
788 /* For NaCl, all tramp args are 32-bit because they're pointers */
789 *code = 0x68; /* push imm32 */
790 *(guint32*)(code + 1) = (gint32)arg1;
794 g_assert ((code - buf) <= size);
799 nacl_domain_code_validate(domain, &buf, size, &code);
801 mono_arch_flush_icache (buf, size);
802 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_SPECIFIC_TRAMPOLINE, mono_get_generic_trampoline_simple_name (tramp_type));
808 mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot, MonoTrampInfo **info, gboolean aot)
812 guint8 **rgctx_null_jumps;
817 MonoJumpInfo *ji = NULL;
818 GSList *unwind_ops = NULL;
820 mrgctx = MONO_RGCTX_SLOT_IS_MRGCTX (slot);
821 index = MONO_RGCTX_SLOT_INDEX (slot);
823 index += MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT / sizeof (gpointer);
824 for (depth = 0; ; ++depth) {
825 int size = mono_class_rgctx_get_array_size (depth, mrgctx);
827 if (index < size - 1)
832 tramp_size = NACL_SIZE (64 + 8 * depth, 128 + 8 * depth);
834 code = buf = mono_global_codeman_reserve (tramp_size);
836 unwind_ops = mono_arch_get_cie_program ();
838 rgctx_null_jumps = g_malloc (sizeof (guint8*) * (depth + 2));
842 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
844 /* load rgctx ptr from vtable */
845 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_ARG_REG1, MONO_STRUCT_OFFSET (MonoVTable, runtime_generic_context), sizeof(gpointer));
846 /* is the rgctx ptr null? */
847 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
848 /* if yes, jump to actual trampoline */
849 rgctx_null_jumps [0] = code;
850 amd64_branch8 (code, X86_CC_Z, -1, 1);
853 for (i = 0; i < depth; ++i) {
854 /* load ptr to next array */
855 if (mrgctx && i == 0)
856 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT, sizeof(gpointer));
858 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, 0, sizeof(gpointer));
859 /* is the ptr null? */
860 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
861 /* if yes, jump to actual trampoline */
862 rgctx_null_jumps [i + 1] = code;
863 amd64_branch8 (code, X86_CC_Z, -1, 1);
867 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, sizeof (gpointer) * (index + 1), sizeof(gpointer));
868 /* is the slot null? */
869 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
870 /* if yes, jump to actual trampoline */
871 rgctx_null_jumps [depth + 1] = code;
872 amd64_branch8 (code, X86_CC_Z, -1, 1);
873 /* otherwise return */
876 for (i = mrgctx ? 1 : 0; i <= depth + 1; ++i)
877 mono_amd64_patch (rgctx_null_jumps [i], code);
879 g_free (rgctx_null_jumps);
881 /* move the rgctx pointer to the VTABLE register */
882 amd64_mov_reg_reg (code, MONO_ARCH_VTABLE_REG, AMD64_ARG_REG1, sizeof(gpointer));
885 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, g_strdup_printf ("specific_trampoline_lazy_fetch_%u", slot));
886 amd64_jump_reg (code, AMD64_R11);
888 tramp = mono_arch_create_specific_trampoline (GUINT_TO_POINTER (slot), MONO_TRAMPOLINE_RGCTX_LAZY_FETCH, mono_get_root_domain (), NULL);
890 /* jump to the actual trampoline */
891 amd64_jump_code (code, tramp);
894 nacl_global_codeman_validate (&buf, tramp_size, &code);
895 mono_arch_flush_icache (buf, code - buf);
896 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_GENERICS_TRAMPOLINE, NULL);
898 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);
910 mono_arch_create_generic_class_init_trampoline (MonoTrampInfo **info, gboolean aot)
914 static int byte_offset = -1;
915 static guint8 bitmask;
918 GSList *unwind_ops = NULL;
919 MonoJumpInfo *ji = NULL;
923 code = buf = mono_global_codeman_reserve (tramp_size);
926 mono_marshal_find_bitfield_offset (MonoVTable, initialized, &byte_offset, &bitmask);
928 amd64_test_membase_imm_size (code, MONO_AMD64_ARG_REG1, byte_offset, bitmask, 1);
930 amd64_branch8 (code, X86_CC_Z, -1, 1);
934 x86_patch (jump, code);
937 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_generic_class_init");
938 amd64_jump_reg (code, AMD64_R11);
940 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_GENERIC_CLASS_INIT, mono_get_root_domain (), NULL);
942 /* jump to the actual trampoline */
943 amd64_jump_code (code, tramp);
946 nacl_global_codeman_validate (&buf, tramp_size, &code);
948 mono_arch_flush_icache (buf, code - buf);
949 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
951 g_assert (code - buf <= tramp_size);
954 *info = mono_tramp_info_create ("generic_class_init_trampoline", buf, code - buf, ji, unwind_ops);
960 mono_arch_create_monitor_enter_trampoline (MonoTrampInfo **info, gboolean aot)
964 guint8 *jump_obj_null, *jump_sync_null, *jump_cmpxchg_failed, *jump_other_owner, *jump_tid, *jump_sync_thin_hash = NULL;
966 int owner_offset, nest_offset, dummy;
967 MonoJumpInfo *ji = NULL;
968 GSList *unwind_ops = NULL;
969 int obj_reg = MONO_AMD64_ARG_REG1;
970 int sync_reg = MONO_AMD64_ARG_REG2;
971 int tid_reg = MONO_AMD64_ARG_REG3;
973 g_assert (MONO_ARCH_MONITOR_OBJECT_REG == obj_reg);
975 mono_monitor_threads_sync_members_offset (&owner_offset, &nest_offset, &dummy);
976 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (owner_offset) == sizeof (gpointer));
977 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
978 owner_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (owner_offset);
979 nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
983 code = buf = mono_global_codeman_reserve (tramp_size);
985 unwind_ops = mono_arch_get_cie_program ();
987 if (!aot && mono_thread_get_tls_offset () != -1) {
988 /* MonoObject* obj is in obj_reg */
990 amd64_test_reg_reg (code, obj_reg, obj_reg);
991 /* if yes, jump to actual trampoline */
992 jump_obj_null = code;
993 amd64_branch8 (code, X86_CC_Z, -1, 1);
995 /* load obj->synchronization to sync_reg */
996 amd64_mov_reg_membase (code, sync_reg, obj_reg, MONO_STRUCT_OFFSET (MonoObject, synchronisation), 8);
998 if (mono_gc_is_moving ()) {
999 /*if bit zero is set it's a thin hash*/
1000 /*FIXME use testb encoding*/
1001 amd64_test_reg_imm (code, sync_reg, 0x01);
1002 jump_sync_thin_hash = code;
1003 amd64_branch8 (code, X86_CC_NE, -1, 1);
1005 /*clear bits used by the gc*/
1006 amd64_alu_reg_imm (code, X86_AND, sync_reg, ~0x3);
1009 /* is synchronization null? */
1010 amd64_test_reg_reg (code, sync_reg, sync_reg);
1011 /* if yes, jump to actual trampoline */
1012 jump_sync_null = code;
1013 amd64_branch8 (code, X86_CC_Z, -1, 1);
1015 /* load MonoInternalThread* into tid_reg */
1016 code = mono_amd64_emit_tls_get (code, tid_reg, mono_thread_get_tls_offset ());
1017 /* load TID into tid_reg */
1018 amd64_mov_reg_membase (code, tid_reg, tid_reg, MONO_STRUCT_OFFSET (MonoInternalThread, tid), 8);
1020 /* is synchronization->owner null? */
1021 amd64_alu_membase_imm_size (code, X86_CMP, sync_reg, owner_offset, 0, 8);
1022 /* if not, jump to next case */
1024 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1026 /* if yes, try a compare-exchange with the TID */
1027 g_assert (tid_reg != X86_EAX);
1029 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
1030 /* compare and exchange */
1031 amd64_prefix (code, X86_LOCK_PREFIX);
1032 amd64_cmpxchg_membase_reg_size (code, sync_reg, owner_offset, tid_reg, 8);
1033 /* if not successful, jump to actual trampoline */
1034 jump_cmpxchg_failed = code;
1035 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1036 /* if successful, return */
1039 /* next case: synchronization->owner is not null */
1040 x86_patch (jump_tid, code);
1041 /* is synchronization->owner == TID? */
1042 amd64_alu_membase_reg_size (code, X86_CMP, sync_reg, owner_offset, tid_reg, 8);
1043 /* if not, jump to actual trampoline */
1044 jump_other_owner = code;
1045 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1046 /* if yes, increment nest */
1047 amd64_inc_membase_size (code, sync_reg, nest_offset, 4);
1051 x86_patch (jump_obj_null, code);
1052 if (jump_sync_thin_hash)
1053 x86_patch (jump_sync_thin_hash, code);
1054 x86_patch (jump_sync_null, code);
1055 x86_patch (jump_cmpxchg_failed, code);
1056 x86_patch (jump_other_owner, code);
1059 /* jump to the actual trampoline */
1060 if (MONO_AMD64_ARG_REG1 != obj_reg)
1061 amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, obj_reg, sizeof (mgreg_t));
1064 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_enter");
1065 amd64_jump_reg (code, AMD64_R11);
1067 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_ENTER, mono_get_root_domain (), NULL);
1069 /* jump to the actual trampoline */
1070 amd64_jump_code (code, tramp);
1073 nacl_global_codeman_validate (&buf, tramp_size, &code);
1075 mono_arch_flush_icache (code, code - buf);
1076 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_MONITOR, NULL);
1077 g_assert (code - buf <= tramp_size);
1080 *info = mono_tramp_info_create ("monitor_enter_trampoline", buf, code - buf, ji, unwind_ops);
1086 mono_arch_create_monitor_exit_trampoline (MonoTrampInfo **info, gboolean aot)
1090 guint8 *jump_obj_null, *jump_have_waiters, *jump_sync_null, *jump_not_owned, *jump_sync_thin_hash = NULL;
1093 int owner_offset, nest_offset, entry_count_offset;
1094 MonoJumpInfo *ji = NULL;
1095 GSList *unwind_ops = NULL;
1096 int obj_reg = MONO_AMD64_ARG_REG1;
1097 int sync_reg = MONO_AMD64_ARG_REG2;
1099 g_assert (obj_reg == MONO_ARCH_MONITOR_OBJECT_REG);
1101 mono_monitor_threads_sync_members_offset (&owner_offset, &nest_offset, &entry_count_offset);
1102 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (owner_offset) == sizeof (gpointer));
1103 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
1104 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (entry_count_offset) == sizeof (gint32));
1105 owner_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (owner_offset);
1106 nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
1107 entry_count_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (entry_count_offset);
1111 code = buf = mono_global_codeman_reserve (tramp_size);
1113 unwind_ops = mono_arch_get_cie_program ();
1115 if (!aot && mono_thread_get_tls_offset () != -1) {
1116 /* MonoObject* obj is in obj_reg */
1118 amd64_test_reg_reg (code, obj_reg, obj_reg);
1119 /* if yes, jump to actual trampoline */
1120 jump_obj_null = code;
1121 amd64_branch8 (code, X86_CC_Z, -1, 1);
1123 /* load obj->synchronization to RCX */
1124 amd64_mov_reg_membase (code, sync_reg, obj_reg, MONO_STRUCT_OFFSET (MonoObject, synchronisation), 8);
1126 if (mono_gc_is_moving ()) {
1127 /*if bit zero is set it's a thin hash*/
1128 /*FIXME use testb encoding*/
1129 amd64_test_reg_imm (code, sync_reg, 0x01);
1130 jump_sync_thin_hash = code;
1131 amd64_branch8 (code, X86_CC_NE, -1, 1);
1133 /*clear bits used by the gc*/
1134 amd64_alu_reg_imm (code, X86_AND, sync_reg, ~0x3);
1137 /* is synchronization null? */
1138 amd64_test_reg_reg (code, sync_reg, sync_reg);
1139 /* if yes, jump to actual trampoline */
1140 jump_sync_null = code;
1141 amd64_branch8 (code, X86_CC_Z, -1, 1);
1143 /* next case: synchronization is not null */
1144 /* load MonoInternalThread* into RAX */
1145 code = mono_amd64_emit_tls_get (code, AMD64_RAX, mono_thread_get_tls_offset ());
1146 /* load TID into RAX */
1147 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, MONO_STRUCT_OFFSET (MonoInternalThread, tid), 8);
1148 /* is synchronization->owner == TID */
1149 amd64_alu_membase_reg_size (code, X86_CMP, sync_reg, owner_offset, AMD64_RAX, 8);
1150 /* if no, jump to actual trampoline */
1151 jump_not_owned = code;
1152 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1154 /* next case: synchronization->owner == TID */
1155 /* is synchronization->nest == 1 */
1156 amd64_alu_membase_imm_size (code, X86_CMP, sync_reg, nest_offset, 1, 4);
1157 /* if not, jump to next case */
1159 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1160 /* if yes, is synchronization->entry_count zero? */
1161 amd64_alu_membase_imm_size (code, X86_CMP, sync_reg, entry_count_offset, 0, 4);
1162 /* if not, jump to actual trampoline */
1163 jump_have_waiters = code;
1164 amd64_branch8 (code, X86_CC_NZ, -1 , 1);
1165 /* if yes, set synchronization->owner to null and return */
1166 amd64_mov_membase_imm (code, sync_reg, owner_offset, 0, 8);
1169 /* next case: synchronization->nest is not 1 */
1170 x86_patch (jump_next, code);
1171 /* decrease synchronization->nest and return */
1172 amd64_dec_membase_size (code, sync_reg, nest_offset, 4);
1175 x86_patch (jump_obj_null, code);
1176 x86_patch (jump_have_waiters, code);
1177 x86_patch (jump_not_owned, code);
1178 x86_patch (jump_sync_null, code);
1181 /* jump to the actual trampoline */
1182 if (MONO_AMD64_ARG_REG1 != obj_reg)
1183 amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, obj_reg, sizeof (mgreg_t));
1186 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_exit");
1187 amd64_jump_reg (code, AMD64_R11);
1189 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_EXIT, mono_get_root_domain (), NULL);
1190 amd64_jump_code (code, tramp);
1193 nacl_global_codeman_validate (&buf, tramp_size, &code);
1195 mono_arch_flush_icache (code, code - buf);
1196 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_MONITOR, NULL);
1197 g_assert (code - buf <= tramp_size);
1200 *info = mono_tramp_info_create ("monitor_exit_trampoline", buf, code - buf, ji, unwind_ops);
1206 mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
1208 /* FIXME: This is not thread safe */
1209 guint8 *code = ji->code_start;
1211 amd64_mov_reg_imm (code, AMD64_ARG_REG1, func_arg);
1212 amd64_mov_reg_imm (code, AMD64_R11, func);
1214 x86_push_imm (code, (guint64)func_arg);
1215 amd64_call_reg (code, AMD64_R11);
1220 handler_block_trampoline_helper (gpointer *ptr)
1222 MonoJitTlsData *jit_tls = mono_native_tls_get_value (mono_jit_tls_id);
1223 *ptr = jit_tls->handler_block_return_address;
1227 mono_arch_create_handler_block_trampoline (MonoTrampInfo **info, gboolean aot)
1229 guint8 *tramp = mono_get_trampoline_code (MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD);
1231 int tramp_size = 64;
1232 MonoJumpInfo *ji = NULL;
1233 GSList *unwind_ops = NULL;
1237 code = buf = mono_global_codeman_reserve (tramp_size);
1240 This trampoline restore the call chain of the handler block then jumps into the code that deals with it.
1242 if (mono_get_jit_tls_offset () != -1) {
1243 code = mono_amd64_emit_tls_get (code, MONO_AMD64_ARG_REG1, mono_get_jit_tls_offset ());
1244 amd64_mov_reg_membase (code, MONO_AMD64_ARG_REG1, MONO_AMD64_ARG_REG1, MONO_STRUCT_OFFSET (MonoJitTlsData, handler_block_return_address), 8);
1245 /* Simulate a call */
1246 amd64_push_reg (code, AMD64_RAX);
1247 amd64_jump_code (code, tramp);
1249 /*Slow path uses a c helper*/
1250 amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RSP, 8);
1251 amd64_mov_reg_imm (code, AMD64_RAX, tramp);
1252 amd64_push_reg (code, AMD64_RAX);
1253 amd64_push_reg (code, AMD64_RAX);
1254 amd64_jump_code (code, handler_block_trampoline_helper);
1257 mono_arch_flush_icache (buf, code - buf);
1258 mono_profiler_code_buffer_new (buf, code - buf, MONO_PROFILER_CODE_BUFFER_HELPER, NULL);
1259 g_assert (code - buf <= tramp_size);
1262 *info = mono_tramp_info_create ("handler_block_trampoline", buf, code - buf, ji, unwind_ops);
1268 * mono_arch_get_call_target:
1270 * Return the address called by the code before CODE if exists.
1273 mono_arch_get_call_target (guint8 *code)
1275 if (code [-5] == 0xe8) {
1276 gint32 disp = *(gint32*)(code - 4);
1277 guint8 *target = code + disp;
1286 * mono_arch_get_plt_info_offset:
1288 * Return the PLT info offset belonging to the plt entry PLT_ENTRY.
1291 mono_arch_get_plt_info_offset (guint8 *plt_entry, mgreg_t *regs, guint8 *code)
1293 #if defined(__native_client__) || defined(__native_client_codegen__)
1294 /* 18 = 3 (mov opcode) + 4 (disp) + 10 (nacljmp) + 1 (push opcode) */
1295 /* See aot-compiler.c arch_emit_plt_entry for details. */
1296 return *(guint32*)(plt_entry + 18);
1298 return *(guint32*)(plt_entry + 6);