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 <sys/nacl_syscalls.h>
35 #define IS_REX(inst) (((inst) >= 0x40) && ((inst) <= 0x4f))
37 static guint8* nullified_class_init_trampoline;
40 * mono_arch_get_unbox_trampoline:
42 * @addr: pointer to native code for @m
44 * when value type methods are called through the vtable we need to unbox the
45 * this argument. This method returns a pointer to a trampoline which does
46 * unboxing before calling the method
49 mono_arch_get_unbox_trampoline (MonoMethod *m, gpointer addr)
54 MonoDomain *domain = mono_domain_get ();
56 this_reg = mono_arch_get_this_arg_reg (NULL);
58 start = code = mono_domain_code_reserve (domain, 20);
60 amd64_alu_reg_imm (code, X86_ADD, this_reg, sizeof (MonoObject));
61 /* FIXME: Optimize this */
62 amd64_mov_reg_imm (code, AMD64_RAX, addr);
63 amd64_jump_reg (code, AMD64_RAX);
64 g_assert ((code - start) < 20);
66 nacl_domain_code_validate (domain, &start, 20, &code);
68 mono_arch_flush_icache (start, code - start);
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)
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);
109 mono_arch_get_llvm_imt_trampoline (MonoDomain *domain, MonoMethod *m, int vt_offset)
111 guint8 *code, *start;
117 start = code = mono_domain_code_reserve (domain, buf_len);
119 this_reg = mono_arch_get_this_arg_reg (NULL);
122 amd64_mov_reg_imm (code, MONO_ARCH_IMT_REG, m);
123 /* Load vtable address */
124 amd64_mov_reg_membase (code, AMD64_RAX, this_reg, 0, 8);
125 amd64_jump_membase (code, AMD64_RAX, vt_offset);
128 g_assert ((code - start) < buf_len);
130 nacl_domain_code_validate (domain, &start, buf_len, &code);
132 mono_arch_flush_icache (start, code - start);
138 * mono_arch_patch_callsite:
140 * Patch the callsite whose address is given by ORIG_CODE so it calls ADDR. ORIG_CODE
141 * points to the pc right after the call.
144 mono_arch_patch_callsite (guint8 *method_start, guint8 *orig_code, guint8 *addr)
146 #if defined(__default_codegen__)
149 gboolean can_write = mono_breakpoint_clean_code (method_start, orig_code, 14, buf, sizeof (buf));
153 /* mov 64-bit imm into r11 (followed by call reg?) or direct call*/
154 if (((code [-13] == 0x49) && (code [-12] == 0xbb)) || (code [-5] == 0xe8)) {
155 if (code [-5] != 0xe8) {
157 InterlockedExchangePointer ((gpointer*)(orig_code - 11), addr);
158 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 11, sizeof (gpointer));
161 gboolean disp_32bit = ((((gint64)addr - (gint64)orig_code)) < (1 << 30)) && ((((gint64)addr - (gint64)orig_code)) > -(1 << 30));
163 if ((((guint64)(addr)) >> 32) != 0 && !disp_32bit) {
164 #ifdef MONO_ARCH_NOMAP32BIT
165 /* Print some diagnostics */
166 MonoJitInfo *ji = mono_jit_info_table_find (mono_domain_get (), (char*)orig_code);
168 fprintf (stderr, "At %s, offset 0x%zx\n", mono_method_full_name (ji->method, TRUE), (guint8*)orig_code - (guint8*)ji->code_start);
169 fprintf (stderr, "Addr: %p\n", addr);
170 ji = mono_jit_info_table_find (mono_domain_get (), (char*)addr);
172 fprintf (stderr, "Callee: %s\n", mono_method_full_name (ji->method, TRUE));
173 g_assert_not_reached ();
176 * This might happen when calling AOTed code. Create a thunk.
178 guint8 *thunk_start, *thunk_code;
180 thunk_start = thunk_code = mono_domain_code_reserve (mono_domain_get (), 32);
181 amd64_jump_membase (thunk_code, AMD64_RIP, 0);
182 *(guint64*)thunk_code = (guint64)addr;
184 g_assert ((((guint64)(addr)) >> 32) == 0);
185 mono_arch_flush_icache (thunk_start, thunk_code - thunk_start);
189 InterlockedExchange ((gint32*)(orig_code - 4), ((gint64)addr - (gint64)orig_code));
190 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, 4);
194 else if ((code [-7] == 0x41) && (code [-6] == 0xff) && (code [-5] == 0x15)) {
195 /* call *<OFFSET>(%rip) */
196 gpointer *got_entry = (gpointer*)((guint8*)orig_code + (*(guint32*)(orig_code - 4)));
198 InterlockedExchangePointer (got_entry, addr);
199 VALGRIND_DISCARD_TRANSLATIONS (orig_code - 5, sizeof (gpointer));
202 #elif defined(__native_client__)
203 /* These are essentially the same 2 cases as above, modified for NaCl*/
205 /* Target must be bundle-aligned */
206 g_assert (((guint32)addr & kNaClAlignmentMask) == 0);
207 /* Return target must be bundle-aligned */
208 g_assert (((guint32)orig_code & kNaClAlignmentMask) == 0);
210 if (orig_code[-5] == 0xe8) {
213 gint32 offset = (gint32)addr - (gint32)orig_code;
214 guint8 buf[sizeof(gint32)];
215 *((gint32*)(buf)) = offset;
216 ret = nacl_dyncode_modify (orig_code - sizeof(gint32), buf, sizeof(gint32));
220 else if (is_nacl_call_reg_sequence (orig_code - 10) && orig_code[-16] == 0x41 && orig_code[-15] == 0xbb) {
222 guint8 buf[sizeof(gint32)];
223 *((gint32 *)(buf)) = addr;
224 /* orig_code[-14] is the start of the immediate. */
225 ret = nacl_dyncode_modify (orig_code - 14, buf, sizeof(gint32));
229 g_assert_not_reached ();
237 mono_arch_patch_plt_entry (guint8 *code, gpointer *got, mgreg_t *regs, guint8 *addr)
240 gpointer *plt_jump_table_entry;
242 #if defined(__default_codegen__)
243 /* A PLT entry: jmp *<DISP>(%rip) */
244 g_assert (code [0] == 0xff);
245 g_assert (code [1] == 0x25);
247 disp = *(gint32*)(code + 2);
249 plt_jump_table_entry = (gpointer*)(code + 6 + disp);
250 #elif defined(__native_client_codegen__)
252 /* mov <DISP>(%rip), %r11d */
255 /* Verify the 'mov' */
256 g_assert (code [0] == 0x45);
257 g_assert (code [1] == 0x8b);
258 g_assert (code [2] == 0x1d);
260 disp = *(gint32*)(code + 3);
262 /* 7 = 3 (mov opcode) + 4 (disp) */
263 /* This needs to resolve to the target of the RIP-relative offset */
264 plt_jump_table_entry = (gpointer*)(code + 7 + disp);
266 #endif /* __native_client_codegen__ */
269 InterlockedExchangePointer (plt_jump_table_entry, addr);
273 get_vcall_slot (guint8 *code, mgreg_t *regs, int *displacement)
277 MonoJitInfo *ji = NULL;
280 /* code - 9 might be before the start of the method */
281 /* FIXME: Avoid this expensive call somehow */
282 ji = mono_jit_info_table_find (mono_domain_get (), (char*)code);
285 mono_breakpoint_clean_code (ji ? ji->code_start : NULL, code, 9, buf, sizeof (buf));
292 if ((code [0] == 0x41) && (code [1] == 0xff) && (code [2] == 0x15)) {
293 /* call OFFSET(%rip) */
294 g_assert_not_reached ();
295 *displacement = *(guint32*)(code + 3);
296 return (gpointer*)(code + disp + 7);
298 g_assert_not_reached ();
304 get_vcall_slot_addr (guint8* code, mgreg_t *regs)
308 vt = get_vcall_slot (code, regs, &displacement);
311 return (gpointer*)((char*)vt + displacement);
315 mono_arch_nullify_class_init_trampoline (guint8 *code, mgreg_t *regs)
318 MonoJitInfo *ji = NULL;
322 /* code - 7 might be before the start of the method */
323 /* FIXME: Avoid this expensive call somehow */
324 ji = mono_jit_info_table_find (mono_domain_get (), (char*)code);
327 can_write = mono_breakpoint_clean_code (ji ? ji->code_start : NULL, code, 7, buf, sizeof (buf));
333 * A given byte sequence can match more than case here, so we have to be
334 * really careful about the ordering of the cases. Longer sequences
337 if ((buf [0] == 0x41) && (buf [1] == 0xff) && (buf [2] == 0x15)) {
338 gpointer *vtable_slot;
340 /* call *<OFFSET>(%rip) */
341 vtable_slot = get_vcall_slot_addr (code, regs);
342 g_assert (vtable_slot);
344 *vtable_slot = nullified_class_init_trampoline;
345 } else if (buf [2] == 0xe8) {
347 //guint8 *buf = code - 2;
350 * It would be better to replace the call with nops, but that doesn't seem
351 * to work on SMP machines even when the whole call is inside a cache line.
352 * Patching the call address seems to work.
362 mono_arch_patch_callsite (code - 5, code, nullified_class_init_trampoline);
363 } else if ((buf [5] == 0xff) && x86_modrm_mod (buf [6]) == 3 && x86_modrm_reg (buf [6]) == 2) {
365 /* Generated by the LLVM JIT or on platforms without MAP_32BIT set */
366 mono_arch_patch_callsite (code - 13, code, nullified_class_init_trampoline);
367 } else if (buf [4] == 0x90 || buf [5] == 0xeb || buf [6] == 0x66) {
368 /* Already changed by another thread */
371 printf ("Invalid trampoline sequence: %x %x %x %x %x %x %x\n", buf [0], buf [1], buf [2], buf [3],
372 buf [4], buf [5], buf [6]);
373 g_assert_not_reached ();
378 mono_arch_nullify_plt_entry (guint8 *code, mgreg_t *regs)
380 if (mono_aot_only && !nullified_class_init_trampoline)
381 nullified_class_init_trampoline = mono_aot_get_trampoline ("nullified_class_init_trampoline");
383 mono_arch_patch_plt_entry (code, NULL, regs, nullified_class_init_trampoline);
387 stack_unaligned (MonoTrampolineType tramp_type)
389 printf ("%d\n", tramp_type);
390 g_assert_not_reached ();
394 mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInfo **info, gboolean aot)
396 guint8 *buf, *code, *tramp, *br [2], *r11_save_code, *after_r11_save_code;
397 int i, lmf_offset, offset, res_offset, arg_offset, rax_offset, tramp_offset, saved_regs_offset;
398 int saved_fpregs_offset, rbp_offset, framesize, orig_rsp_to_rbp_offset, cfa_offset;
400 GSList *unwind_ops = NULL;
401 MonoJumpInfo *ji = NULL;
402 const guint kMaxCodeSize = NACL_SIZE (600, 600*2);
404 #if defined(__native_client_codegen__)
405 const guint kNaClTrampOffset = 17;
408 if (tramp_type == MONO_TRAMPOLINE_JUMP || tramp_type == MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD)
413 code = buf = mono_global_codeman_reserve (kMaxCodeSize);
415 framesize = kMaxCodeSize + sizeof (MonoLMF);
416 framesize = (framesize + (MONO_ARCH_FRAME_ALIGNMENT - 1)) & ~ (MONO_ARCH_FRAME_ALIGNMENT - 1);
418 orig_rsp_to_rbp_offset = 0;
419 r11_save_code = code;
420 /* Reserve 5 bytes for the mov_membase_reg to save R11 */
422 after_r11_save_code = code;
424 // CFA = sp + 16 (the trampoline address is on the stack)
426 mono_add_unwind_op_def_cfa (unwind_ops, code, buf, AMD64_RSP, 16);
427 // IP saved at CFA - 8
428 mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RIP, -8);
430 /* Pop the return address off the stack */
431 amd64_pop_reg (code, AMD64_R11);
432 orig_rsp_to_rbp_offset += sizeof(mgreg_t);
434 cfa_offset -= sizeof(mgreg_t);
435 mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
438 * Allocate a new stack frame
440 amd64_push_reg (code, AMD64_RBP);
441 cfa_offset += sizeof(mgreg_t);
442 mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
443 mono_add_unwind_op_offset (unwind_ops, code, buf, AMD64_RBP, - cfa_offset);
445 orig_rsp_to_rbp_offset -= sizeof(mgreg_t);
446 amd64_mov_reg_reg (code, AMD64_RBP, AMD64_RSP, sizeof(mgreg_t));
447 mono_add_unwind_op_def_cfa_reg (unwind_ops, code, buf, AMD64_RBP);
448 amd64_alu_reg_imm (code, X86_SUB, AMD64_RSP, framesize);
451 rbp_offset = - offset;
453 offset += sizeof(mgreg_t);
454 rax_offset = - offset;
456 offset += sizeof(mgreg_t);
457 tramp_offset = - offset;
459 offset += sizeof(gpointer);
460 arg_offset = - offset;
462 /* Compute the trampoline address from the return address */
464 #if defined(__default_codegen__)
465 /* 7 = length of call *<offset>(rip) */
466 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 7);
467 #elif defined(__native_client_codegen__)
468 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, kNaClTrampOffset);
471 /* 5 = length of amd64_call_membase () */
472 amd64_alu_reg_imm (code, X86_SUB, AMD64_R11, 5);
474 amd64_mov_membase_reg (code, AMD64_RBP, tramp_offset, AMD64_R11, sizeof(gpointer));
476 offset += sizeof(mgreg_t);
477 res_offset = - offset;
479 /* Save all registers */
481 offset += AMD64_NREG * sizeof(mgreg_t);
482 saved_regs_offset = - offset;
483 for (i = 0; i < AMD64_NREG; ++i) {
484 if (i == AMD64_RBP) {
485 /* RAX is already saved */
486 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, rbp_offset, sizeof(mgreg_t));
487 amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), AMD64_RAX, sizeof(mgreg_t));
488 } else if (i != AMD64_R11) {
489 amd64_mov_membase_reg (code, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), i, sizeof(mgreg_t));
491 /* We have to save R11 right at the start of
492 the trampoline code because it's used as a
494 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));
495 g_assert (r11_save_code == after_r11_save_code);
498 offset += 8 * sizeof(mgreg_t);
499 saved_fpregs_offset = - offset;
500 for (i = 0; i < 8; ++i)
501 amd64_movsd_membase_reg (code, AMD64_RBP, saved_fpregs_offset + (i * sizeof(mgreg_t)), i);
503 /* Check that the stack is aligned */
504 #if defined(__default_codegen__)
505 amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, sizeof (mgreg_t));
506 amd64_alu_reg_imm (code, X86_AND, AMD64_R11, 15);
507 amd64_alu_reg_imm (code, X86_CMP, AMD64_R11, 0);
509 amd64_branch_disp (code, X86_CC_Z, 0, FALSE);
511 amd64_mov_reg_imm (code, AMD64_R11, 0);
512 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, 8);
514 amd64_mov_reg_imm (code, AMD64_RDI, tramp_type);
515 amd64_mov_reg_imm (code, AMD64_R11, stack_unaligned);
516 amd64_call_reg (code, AMD64_R11);
518 mono_amd64_patch (br [0], code);
519 //amd64_breakpoint (code);
522 if (tramp_type != MONO_TRAMPOLINE_GENERIC_CLASS_INIT &&
523 tramp_type != MONO_TRAMPOLINE_MONITOR_ENTER &&
524 tramp_type != MONO_TRAMPOLINE_MONITOR_EXIT &&
525 tramp_type != MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD) {
526 /* Obtain the trampoline argument which is encoded in the instruction stream */
528 /* Load the GOT offset */
529 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, sizeof(gpointer));
530 #if defined(__default_codegen__)
531 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 7, 4);
532 #elif defined(__native_client_codegen__)
533 /* The arg is hidden in a "push imm32" instruction, */
534 /* add one to skip the opcode. */
535 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, kNaClTrampOffset+1, 4);
537 /* Compute the address of the GOT slot */
538 amd64_alu_reg_reg_size (code, X86_ADD, AMD64_R11, AMD64_RAX, sizeof(gpointer));
540 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 0, sizeof(gpointer));
542 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, tramp_offset, sizeof(gpointer));
543 #if defined(__default_codegen__)
544 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_R11, 5, 1);
545 amd64_widen_reg (code, AMD64_RAX, AMD64_RAX, TRUE, FALSE);
546 amd64_alu_reg_imm_size (code, X86_CMP, AMD64_RAX, 4, 1);
548 x86_branch8 (code, X86_CC_NE, 6, FALSE);
549 /* 32 bit immediate */
550 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 4);
552 x86_jump8 (code, 10);
553 /* 64 bit immediate */
554 mono_amd64_patch (br [0], code);
555 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 8);
556 mono_amd64_patch (br [1], code);
557 #elif defined(__native_client_codegen__)
558 /* All args are 32-bit pointers in NaCl */
559 amd64_mov_reg_membase (code, AMD64_R11, AMD64_R11, 6, 4);
562 amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, sizeof(gpointer));
564 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, saved_regs_offset + (MONO_AMD64_ARG_REG1 * sizeof(mgreg_t)), sizeof(mgreg_t));
565 amd64_mov_membase_reg (code, AMD64_RBP, arg_offset, AMD64_R11, sizeof(gpointer));
570 offset += sizeof (MonoLMF);
571 lmf_offset = - offset;
575 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, 8, sizeof(gpointer));
577 amd64_mov_reg_imm (code, AMD64_R11, 0);
578 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rip), AMD64_R11, sizeof(mgreg_t));
580 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RSP, framesize, sizeof(mgreg_t));
581 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbp), AMD64_R11, sizeof(mgreg_t));
583 amd64_mov_reg_reg (code, AMD64_R11, AMD64_RSP, sizeof(mgreg_t));
584 amd64_alu_reg_imm (code, X86_ADD, AMD64_R11, framesize + 16);
585 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsp), AMD64_R11, sizeof(mgreg_t));
587 if (tramp_type == MONO_TRAMPOLINE_JIT || tramp_type == MONO_TRAMPOLINE_JUMP) {
588 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, arg_offset, sizeof(gpointer));
589 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), AMD64_R11, sizeof(gpointer));
591 amd64_mov_membase_imm (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), 0, sizeof(gpointer));
593 /* Save callee saved regs */
595 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rdi), AMD64_RDI, sizeof(mgreg_t));
596 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rsi), AMD64_RSI, sizeof(mgreg_t));
598 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, rbx), AMD64_RBX, sizeof(mgreg_t));
599 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r12), AMD64_R12, sizeof(mgreg_t));
600 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r13), AMD64_R13, sizeof(mgreg_t));
601 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r14), AMD64_R14, sizeof(mgreg_t));
602 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, r15), AMD64_R15, sizeof(mgreg_t));
605 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_get_lmf_addr");
607 amd64_mov_reg_imm (code, AMD64_R11, mono_get_lmf_addr);
609 amd64_call_reg (code, AMD64_R11);
612 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), AMD64_RAX, sizeof(gpointer));
613 /* Save previous_lmf */
614 /* Set the lowest bit to 1 to signal that this LMF has the ip field set */
615 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RAX, 0, sizeof(gpointer));
616 amd64_alu_reg_imm_size (code, X86_ADD, AMD64_R11, 1, sizeof(gpointer));
617 amd64_mov_membase_reg (code, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), AMD64_R11, sizeof(gpointer));
619 amd64_lea_membase (code, AMD64_R11, AMD64_RBP, lmf_offset);
620 amd64_mov_membase_reg (code, AMD64_RAX, 0, AMD64_R11, sizeof(gpointer));
624 /* Arg1 is the pointer to the saved registers */
625 amd64_lea_membase (code, AMD64_ARG_REG1, AMD64_RBP, saved_regs_offset);
627 /* Arg2 is the address of the calling code */
629 amd64_mov_reg_membase (code, AMD64_ARG_REG2, AMD64_RBP, 8, sizeof(gpointer));
631 amd64_mov_reg_imm (code, AMD64_ARG_REG2, 0);
633 /* Arg3 is the method/vtable ptr */
634 amd64_mov_reg_membase (code, AMD64_ARG_REG3, AMD64_RBP, arg_offset, sizeof(gpointer));
636 /* Arg4 is the trampoline address */
637 amd64_mov_reg_membase (code, AMD64_ARG_REG4, AMD64_RBP, tramp_offset, sizeof(gpointer));
640 char *icall_name = g_strdup_printf ("trampoline_func_%d", tramp_type);
641 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, icall_name);
643 tramp = (guint8*)mono_get_trampoline_func (tramp_type);
644 amd64_mov_reg_imm (code, AMD64_R11, tramp);
646 amd64_call_reg (code, AMD64_R11);
648 /* Check for thread interruption */
649 /* This is not perf critical code so no need to check the interrupt flag */
651 * Have to call the _force_ variant, since there could be a protected wrapper on the top of the stack.
653 amd64_mov_membase_reg (code, AMD64_RBP, res_offset, AMD64_RAX, sizeof(mgreg_t));
655 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_thread_force_interruption_checkpoint");
657 amd64_mov_reg_imm (code, AMD64_R11, (guint8*)mono_thread_force_interruption_checkpoint);
659 amd64_call_reg (code, AMD64_R11);
661 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RBP, res_offset, sizeof(mgreg_t));
665 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), sizeof(gpointer));
666 amd64_alu_reg_imm_size (code, X86_SUB, AMD64_RCX, 1, sizeof(gpointer));
667 amd64_mov_reg_membase (code, AMD64_R11, AMD64_RBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), sizeof(gpointer));
668 amd64_mov_membase_reg (code, AMD64_R11, 0, AMD64_RCX, sizeof(gpointer));
671 * Save rax to the stack, after the leave instruction, this will become part of
674 amd64_mov_membase_reg (code, AMD64_RBP, rax_offset, AMD64_RAX, sizeof(mgreg_t));
676 /* Restore argument registers, r10 (imt method/rgxtx)
677 and rax (needed for direct calls to C vararg functions). */
678 for (i = 0; i < AMD64_NREG; ++i)
679 if (AMD64_IS_ARGUMENT_REG (i) || i == AMD64_R10 || i == AMD64_RAX)
680 amd64_mov_reg_membase (code, i, AMD64_RBP, saved_regs_offset + (i * sizeof(mgreg_t)), sizeof(mgreg_t));
682 for (i = 0; i < 8; ++i)
683 amd64_movsd_reg_membase (code, i, AMD64_RBP, saved_fpregs_offset + (i * sizeof(mgreg_t)));
688 if (MONO_TRAMPOLINE_TYPE_MUST_RETURN (tramp_type)) {
690 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RSP, rax_offset - sizeof(mgreg_t), sizeof(mgreg_t));
693 /* call the compiled method using the saved rax */
694 amd64_jump_membase (code, AMD64_RSP, rax_offset - sizeof(mgreg_t));
697 g_assert ((code - buf) <= kMaxCodeSize);
699 nacl_global_codeman_validate (&buf, kMaxCodeSize, &code);
701 mono_arch_flush_icache (buf, code - buf);
703 if (tramp_type == MONO_TRAMPOLINE_CLASS_INIT) {
704 /* Initialize the nullified class init trampoline used in the AOT case */
705 nullified_class_init_trampoline = mono_arch_get_nullified_class_init_trampoline (NULL);
709 *info = mono_tramp_info_create (mono_get_generic_trampoline_name (tramp_type), buf, code - buf, ji, unwind_ops);
715 mono_arch_get_nullified_class_init_trampoline (MonoTrampInfo **info)
719 code = buf = mono_global_codeman_reserve (16);
722 nacl_global_codeman_validate(&buf, 16, &code);
724 mono_arch_flush_icache (buf, code - buf);
727 *info = mono_tramp_info_create (g_strdup_printf ("nullified_class_init_trampoline"), buf, code - buf, NULL, NULL);
729 if (mono_jit_map_is_enabled ())
730 mono_emit_jit_tramp (buf, code - buf, "nullified_class_init_trampoline");
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 buf = mono_domain_code_reserve_align (domain, size, kNaClAlignment);
770 amd64_mov_reg_imm (code, AMD64_R11, tramp);
771 amd64_call_reg (code, AMD64_R11);
773 amd64_call_code (code, tramp);
775 /* The trampoline code will obtain the argument from the instruction stream */
776 #if defined(__default_codegen__)
777 if ((((guint64)arg1) >> 32) == 0) {
779 *(guint32*)(code + 1) = (gint64)arg1;
783 *(guint64*)(code + 1) = (gint64)arg1;
786 #elif defined(__native_client_codegen__)
787 /* For NaCl, all tramp args are 32-bit because they're pointers */
788 *code = 0x68; /* push imm32 */
789 *(guint32*)(code + 1) = (gint32)arg1;
793 g_assert ((code - buf) <= size);
798 nacl_domain_code_validate(domain, &buf, size, &code);
800 mono_arch_flush_icache (buf, size);
806 mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot, MonoTrampInfo **info, gboolean aot)
810 guint8 **rgctx_null_jumps;
815 MonoJumpInfo *ji = NULL;
816 GSList *unwind_ops = NULL;
818 mrgctx = MONO_RGCTX_SLOT_IS_MRGCTX (slot);
819 index = MONO_RGCTX_SLOT_INDEX (slot);
821 index += MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT / sizeof (gpointer);
822 for (depth = 0; ; ++depth) {
823 int size = mono_class_rgctx_get_array_size (depth, mrgctx);
825 if (index < size - 1)
830 tramp_size = NACL_SIZE (64 + 8 * depth, 128 + 8 * depth);
832 code = buf = mono_global_codeman_reserve (tramp_size);
834 unwind_ops = mono_arch_get_cie_program ();
836 rgctx_null_jumps = g_malloc (sizeof (guint8*) * (depth + 2));
840 amd64_mov_reg_reg (code, AMD64_RAX, AMD64_ARG_REG1, 8);
842 /* load rgctx ptr from vtable */
843 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_ARG_REG1, G_STRUCT_OFFSET (MonoVTable, runtime_generic_context), sizeof(gpointer));
844 /* is the rgctx ptr null? */
845 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
846 /* if yes, jump to actual trampoline */
847 rgctx_null_jumps [0] = code;
848 amd64_branch8 (code, X86_CC_Z, -1, 1);
851 for (i = 0; i < depth; ++i) {
852 /* load ptr to next array */
853 if (mrgctx && i == 0)
854 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, MONO_SIZEOF_METHOD_RUNTIME_GENERIC_CONTEXT, sizeof(gpointer));
856 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, 0, sizeof(gpointer));
857 /* is the ptr null? */
858 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
859 /* if yes, jump to actual trampoline */
860 rgctx_null_jumps [i + 1] = code;
861 amd64_branch8 (code, X86_CC_Z, -1, 1);
865 amd64_mov_reg_membase (code, AMD64_RAX, AMD64_RAX, sizeof (gpointer) * (index + 1), sizeof(gpointer));
866 /* is the slot null? */
867 amd64_test_reg_reg (code, AMD64_RAX, AMD64_RAX);
868 /* if yes, jump to actual trampoline */
869 rgctx_null_jumps [depth + 1] = code;
870 amd64_branch8 (code, X86_CC_Z, -1, 1);
871 /* otherwise return */
874 for (i = mrgctx ? 1 : 0; i <= depth + 1; ++i)
875 mono_amd64_patch (rgctx_null_jumps [i], code);
877 g_free (rgctx_null_jumps);
879 /* move the rgctx pointer to the VTABLE register */
880 amd64_mov_reg_reg (code, MONO_ARCH_VTABLE_REG, AMD64_ARG_REG1, sizeof(gpointer));
883 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, g_strdup_printf ("specific_trampoline_lazy_fetch_%u", slot));
884 amd64_jump_reg (code, AMD64_R11);
886 tramp = mono_arch_create_specific_trampoline (GUINT_TO_POINTER (slot), MONO_TRAMPOLINE_RGCTX_LAZY_FETCH, mono_get_root_domain (), NULL);
888 /* jump to the actual trampoline */
889 amd64_jump_code (code, tramp);
892 nacl_global_codeman_validate (&buf, tramp_size, &code);
893 mono_arch_flush_icache (buf, code - buf);
895 g_assert (code - buf <= tramp_size);
898 *info = mono_tramp_info_create (mono_get_rgctx_fetch_trampoline_name (slot), buf, code - buf, ji, unwind_ops);
904 mono_arch_create_generic_class_init_trampoline (MonoTrampInfo **info, gboolean aot)
908 static int byte_offset = -1;
909 static guint8 bitmask;
912 GSList *unwind_ops = NULL;
913 MonoJumpInfo *ji = NULL;
917 code = buf = mono_global_codeman_reserve (tramp_size);
920 mono_marshal_find_bitfield_offset (MonoVTable, initialized, &byte_offset, &bitmask);
922 amd64_test_membase_imm_size (code, MONO_AMD64_ARG_REG1, byte_offset, bitmask, 1);
924 amd64_branch8 (code, X86_CC_Z, -1, 1);
928 x86_patch (jump, code);
931 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_generic_class_init");
932 amd64_jump_reg (code, AMD64_R11);
934 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_GENERIC_CLASS_INIT, mono_get_root_domain (), NULL);
936 /* jump to the actual trampoline */
937 amd64_jump_code (code, tramp);
940 nacl_global_codeman_validate (&buf, tramp_size, &code);
942 mono_arch_flush_icache (buf, code - buf);
944 g_assert (code - buf <= tramp_size);
947 *info = mono_tramp_info_create (g_strdup_printf ("generic_class_init_trampoline"), buf, code - buf, ji, unwind_ops);
952 #ifdef MONO_ARCH_MONITOR_OBJECT_REG
955 mono_arch_create_monitor_enter_trampoline (MonoTrampInfo **info, gboolean aot)
959 guint8 *jump_obj_null, *jump_sync_null, *jump_cmpxchg_failed, *jump_other_owner, *jump_tid, *jump_sync_thin_hash = NULL;
961 int owner_offset, nest_offset, dummy;
962 MonoJumpInfo *ji = NULL;
963 GSList *unwind_ops = NULL;
965 g_assert (MONO_ARCH_MONITOR_OBJECT_REG == AMD64_RDI);
967 mono_monitor_threads_sync_members_offset (&owner_offset, &nest_offset, &dummy);
968 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (owner_offset) == sizeof (gpointer));
969 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
970 owner_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (owner_offset);
971 nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
975 code = buf = mono_global_codeman_reserve (tramp_size);
977 unwind_ops = mono_arch_get_cie_program ();
979 if (mono_thread_get_tls_offset () != -1) {
980 /* MonoObject* obj is in RDI */
982 amd64_test_reg_reg (code, AMD64_RDI, AMD64_RDI);
983 /* if yes, jump to actual trampoline */
984 jump_obj_null = code;
985 amd64_branch8 (code, X86_CC_Z, -1, 1);
987 /* load obj->synchronization to RCX */
988 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RDI, G_STRUCT_OFFSET (MonoObject, synchronisation), 8);
990 if (mono_gc_is_moving ()) {
991 /*if bit zero is set it's a thin hash*/
992 /*FIXME use testb encoding*/
993 amd64_test_reg_imm (code, AMD64_RCX, 0x01);
994 jump_sync_thin_hash = code;
995 amd64_branch8 (code, X86_CC_NE, -1, 1);
997 /*clear bits used by the gc*/
998 amd64_alu_reg_imm (code, X86_AND, AMD64_RCX, ~0x3);
1001 /* is synchronization null? */
1002 amd64_test_reg_reg (code, AMD64_RCX, AMD64_RCX);
1003 /* if yes, jump to actual trampoline */
1004 jump_sync_null = code;
1005 amd64_branch8 (code, X86_CC_Z, -1, 1);
1007 /* load MonoInternalThread* into RDX */
1008 code = mono_amd64_emit_tls_get (code, AMD64_RDX, mono_thread_get_tls_offset ());
1009 /* load TID into RDX */
1010 amd64_mov_reg_membase (code, AMD64_RDX, AMD64_RDX, G_STRUCT_OFFSET (MonoInternalThread, tid), 8);
1012 /* is synchronization->owner null? */
1013 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, owner_offset, 0, 8);
1014 /* if not, jump to next case */
1016 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1018 /* if yes, try a compare-exchange with the TID */
1020 amd64_alu_reg_reg (code, X86_XOR, AMD64_RAX, AMD64_RAX);
1021 /* compare and exchange */
1022 amd64_prefix (code, X86_LOCK_PREFIX);
1023 amd64_cmpxchg_membase_reg_size (code, AMD64_RCX, owner_offset, AMD64_RDX, 8);
1024 /* if not successful, jump to actual trampoline */
1025 jump_cmpxchg_failed = code;
1026 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1027 /* if successful, return */
1030 /* next case: synchronization->owner is not null */
1031 x86_patch (jump_tid, code);
1032 /* is synchronization->owner == TID? */
1033 amd64_alu_membase_reg_size (code, X86_CMP, AMD64_RCX, owner_offset, AMD64_RDX, 8);
1034 /* if not, jump to actual trampoline */
1035 jump_other_owner = code;
1036 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1037 /* if yes, increment nest */
1038 amd64_inc_membase_size (code, AMD64_RCX, nest_offset, 4);
1042 x86_patch (jump_obj_null, code);
1043 if (jump_sync_thin_hash)
1044 x86_patch (jump_sync_thin_hash, code);
1045 x86_patch (jump_sync_null, code);
1046 x86_patch (jump_cmpxchg_failed, code);
1047 x86_patch (jump_other_owner, code);
1050 /* jump to the actual trampoline */
1051 #if MONO_AMD64_ARG_REG1 != AMD64_RDI
1052 amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RDI);
1056 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_enter");
1057 amd64_jump_reg (code, AMD64_R11);
1059 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_ENTER, mono_get_root_domain (), NULL);
1061 /* jump to the actual trampoline */
1062 amd64_jump_code (code, tramp);
1065 nacl_global_codeman_validate (&buf, tramp_size, &code);
1067 mono_arch_flush_icache (code, code - buf);
1068 g_assert (code - buf <= tramp_size);
1071 *info = mono_tramp_info_create (g_strdup_printf ("monitor_enter_trampoline"), buf, code - buf, ji, unwind_ops);
1077 mono_arch_create_monitor_exit_trampoline (MonoTrampInfo **info, gboolean aot)
1081 guint8 *jump_obj_null, *jump_have_waiters, *jump_sync_null, *jump_not_owned, *jump_sync_thin_hash = NULL;
1084 int owner_offset, nest_offset, entry_count_offset;
1085 MonoJumpInfo *ji = NULL;
1086 GSList *unwind_ops = NULL;
1088 g_assert (MONO_ARCH_MONITOR_OBJECT_REG == AMD64_RDI);
1090 mono_monitor_threads_sync_members_offset (&owner_offset, &nest_offset, &entry_count_offset);
1091 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (owner_offset) == sizeof (gpointer));
1092 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (nest_offset) == sizeof (guint32));
1093 g_assert (MONO_THREADS_SYNC_MEMBER_SIZE (entry_count_offset) == sizeof (gint32));
1094 owner_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (owner_offset);
1095 nest_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (nest_offset);
1096 entry_count_offset = MONO_THREADS_SYNC_MEMBER_OFFSET (entry_count_offset);
1100 code = buf = mono_global_codeman_reserve (tramp_size);
1102 unwind_ops = mono_arch_get_cie_program ();
1104 if (mono_thread_get_tls_offset () != -1) {
1105 /* MonoObject* obj is in RDI */
1107 amd64_test_reg_reg (code, AMD64_RDI, AMD64_RDI);
1108 /* if yes, jump to actual trampoline */
1109 jump_obj_null = code;
1110 amd64_branch8 (code, X86_CC_Z, -1, 1);
1112 /* load obj->synchronization to RCX */
1113 amd64_mov_reg_membase (code, AMD64_RCX, AMD64_RDI, G_STRUCT_OFFSET (MonoObject, synchronisation), 8);
1115 if (mono_gc_is_moving ()) {
1116 /*if bit zero is set it's a thin hash*/
1117 /*FIXME use testb encoding*/
1118 amd64_test_reg_imm (code, AMD64_RCX, 0x01);
1119 jump_sync_thin_hash = code;
1120 amd64_branch8 (code, X86_CC_NE, -1, 1);
1122 /*clear bits used by the gc*/
1123 amd64_alu_reg_imm (code, X86_AND, AMD64_RCX, ~0x3);
1126 /* is synchronization null? */
1127 amd64_test_reg_reg (code, AMD64_RCX, AMD64_RCX);
1128 /* if yes, jump to actual trampoline */
1129 jump_sync_null = code;
1130 amd64_branch8 (code, X86_CC_Z, -1, 1);
1132 /* next case: synchronization is not null */
1133 /* load MonoInternalThread* into RDX */
1134 code = mono_amd64_emit_tls_get (code, AMD64_RDX, mono_thread_get_tls_offset ());
1135 /* load TID into RDX */
1136 amd64_mov_reg_membase (code, AMD64_RDX, AMD64_RDX, G_STRUCT_OFFSET (MonoInternalThread, tid), 8);
1137 /* is synchronization->owner == TID */
1138 amd64_alu_membase_reg_size (code, X86_CMP, AMD64_RCX, owner_offset, AMD64_RDX, 8);
1139 /* if no, jump to actual trampoline */
1140 jump_not_owned = code;
1141 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1143 /* next case: synchronization->owner == TID */
1144 /* is synchronization->nest == 1 */
1145 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, nest_offset, 1, 4);
1146 /* if not, jump to next case */
1148 amd64_branch8 (code, X86_CC_NZ, -1, 1);
1149 /* if yes, is synchronization->entry_count zero? */
1150 amd64_alu_membase_imm_size (code, X86_CMP, AMD64_RCX, entry_count_offset, 0, 4);
1151 /* if not, jump to actual trampoline */
1152 jump_have_waiters = code;
1153 amd64_branch8 (code, X86_CC_NZ, -1 , 1);
1154 /* if yes, set synchronization->owner to null and return */
1155 amd64_mov_membase_imm (code, AMD64_RCX, owner_offset, 0, 8);
1158 /* next case: synchronization->nest is not 1 */
1159 x86_patch (jump_next, code);
1160 /* decrease synchronization->nest and return */
1161 amd64_dec_membase_size (code, AMD64_RCX, nest_offset, 4);
1164 x86_patch (jump_obj_null, code);
1165 x86_patch (jump_have_waiters, code);
1166 x86_patch (jump_not_owned, code);
1167 x86_patch (jump_sync_null, code);
1170 /* jump to the actual trampoline */
1171 #if MONO_AMD64_ARG_REG1 != AMD64_RDI
1172 amd64_mov_reg_reg (code, MONO_AMD64_ARG_REG1, AMD64_RDI);
1176 code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "specific_trampoline_monitor_exit");
1177 amd64_jump_reg (code, AMD64_R11);
1179 tramp = mono_arch_create_specific_trampoline (NULL, MONO_TRAMPOLINE_MONITOR_EXIT, mono_get_root_domain (), NULL);
1180 amd64_jump_code (code, tramp);
1183 nacl_global_codeman_validate (&buf, tramp_size, &code);
1185 mono_arch_flush_icache (code, code - buf);
1186 g_assert (code - buf <= tramp_size);
1189 *info = mono_tramp_info_create (g_strdup_printf ("monitor_exit_trampoline"), buf, code - buf, ji, unwind_ops);
1196 mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
1198 /* FIXME: This is not thread safe */
1199 guint8 *code = ji->code_start;
1201 amd64_mov_reg_imm (code, AMD64_ARG_REG1, func_arg);
1202 amd64_mov_reg_imm (code, AMD64_R11, func);
1204 x86_push_imm (code, (guint64)func_arg);
1205 amd64_call_reg (code, AMD64_R11);
1210 handler_block_trampoline_helper (gpointer *ptr)
1212 MonoJitTlsData *jit_tls = mono_native_tls_get_value (mono_jit_tls_id);
1213 *ptr = jit_tls->handler_block_return_address;
1217 mono_arch_create_handler_block_trampoline (void)
1219 guint8 *tramp = mono_get_trampoline_code (MONO_TRAMPOLINE_HANDLER_BLOCK_GUARD);
1221 int tramp_size = 64;
1222 code = buf = mono_global_codeman_reserve (tramp_size);
1225 This trampoline restore the call chain of the handler block then jumps into the code that deals with it.
1228 if (mono_get_jit_tls_offset () != -1) {
1229 code = mono_amd64_emit_tls_get (code, AMD64_RDI, mono_get_jit_tls_offset ());
1230 amd64_mov_reg_membase (code, AMD64_RDI, AMD64_RDI, G_STRUCT_OFFSET (MonoJitTlsData, handler_block_return_address), 8);
1231 /* Simulate a call */
1232 amd64_push_reg (code, AMD64_RAX);
1233 amd64_jump_code (code, tramp);
1235 /*Slow path uses a c helper*/
1236 amd64_mov_reg_reg (code, AMD64_RDI, AMD64_RSP, 8);
1237 amd64_mov_reg_imm (code, AMD64_RAX, tramp);
1238 amd64_push_reg (code, AMD64_RAX);
1239 amd64_jump_code (code, handler_block_trampoline_helper);
1242 mono_arch_flush_icache (buf, code - buf);
1243 g_assert (code - buf <= tramp_size);
1245 if (mono_jit_map_is_enabled ())
1246 mono_emit_jit_tramp (buf, code - buf, "handler_block_trampoline");
1252 * mono_arch_get_call_target:
1254 * Return the address called by the code before CODE if exists.
1257 mono_arch_get_call_target (guint8 *code)
1259 if (code [-5] == 0xe8) {
1260 guint32 disp = *(guint32*)(code - 4);
1261 guint8 *target = code + disp;
1270 * mono_arch_get_plt_info_offset:
1272 * Return the PLT info offset belonging to the plt entry PLT_ENTRY.
1275 mono_arch_get_plt_info_offset (guint8 *plt_entry, mgreg_t *regs, guint8 *code)
1277 #if defined(__native_client__) || defined(__native_client_codegen__)
1278 /* 18 = 3 (mov opcode) + 4 (disp) + 10 (nacljmp) + 1 (push opcode) */
1279 /* See aot-compiler.c arch_emit_plt_entry for details. */
1280 return *(guint32*)(plt_entry + 18);
1282 return *(guint32*)(plt_entry + 6);