2 * tramp-x86.c: JIT trampoline code for x86
5 * Dietmar Maurer (dietmar@ximian.com)
7 * (C) 2001 Ximian, Inc.
13 #include <mono/metadata/appdomain.h>
14 #include <mono/metadata/metadata-internals.h>
15 #include <mono/metadata/marshal.h>
16 #include <mono/metadata/tabledefs.h>
17 #include <mono/metadata/mono-debug.h>
18 #include <mono/metadata/mono-debug-debugger.h>
19 #include <mono/arch/x86/x86-codegen.h>
21 #ifdef HAVE_VALGRIND_MEMCHECK_H
22 #include <valgrind/memcheck.h>
28 static guint8* nullified_class_init_trampoline;
31 * mono_arch_get_unbox_trampoline:
32 * @gsctx: the generic sharing context
34 * @addr: pointer to native code for @m
36 * when value type methods are called through the vtable we need to unbox the
37 * this argument. This method returns a pointer to a trampoline which does
38 * unboxing before calling the method
41 mono_arch_get_unbox_trampoline (MonoGenericSharingContext *gsctx, MonoMethod *m, gpointer addr)
45 MonoDomain *domain = mono_domain_get ();
47 if (MONO_TYPE_ISSTRUCT (mono_method_signature (m)->ret))
50 mono_domain_lock (domain);
51 start = code = mono_code_manager_reserve (domain->code_mp, 16);
52 mono_domain_unlock (domain);
54 x86_alu_membase_imm (code, X86_ADD, X86_ESP, this_pos, sizeof (MonoObject));
55 x86_jump_code (code, addr);
56 g_assert ((code - start) < 16);
62 mono_arch_patch_callsite (guint8 *method_start, guint8 *orig_code, guint8 *addr)
66 gboolean can_write = mono_breakpoint_clean_code (method_start, orig_code, 8, buf, sizeof (buf));
69 if (mono_running_on_valgrind ())
72 /* go to the start of the call instruction
74 * address_byte = (m << 6) | (o << 3) | reg
75 * call opcode: 0xff address_byte displacement
81 if ((code [1] == 0xe8)) {
83 InterlockedExchange ((gint32*)(orig_code + 2), (guint)addr - ((guint)orig_code + 1) - 5);
85 #ifdef HAVE_VALGRIND_MEMCHECK_H
86 /* Tell valgrind to recompile the patched code */
87 //VALGRIND_DISCARD_TRANSLATIONS (code + 2, code + 6);
90 } else if (code [1] == 0xe9) {
91 /* A PLT entry: jmp <DISP> */
93 InterlockedExchange ((gint32*)(orig_code + 2), (guint)addr - ((guint)orig_code + 1) - 5);
95 printf ("Invalid trampoline sequence: %x %x %x %x %x %x %x\n", code [0], code [1], code [2], code [3],
96 code [4], code [5], code [6]);
97 g_assert_not_reached ();
102 mono_arch_patch_plt_entry (guint8 *code, guint8 *addr)
104 /* A PLT entry: jmp <DISP> */
105 g_assert (code [0] == 0xe9);
107 if (!mono_running_on_valgrind ())
108 InterlockedExchange ((gint32*)(code + 1), (guint)addr - (guint)code - 5);
112 mono_arch_nullify_class_init_trampoline (guint8 *code, gssize *regs)
115 gboolean can_write = mono_breakpoint_clean_code (NULL, code, 6, buf, sizeof (buf));
121 if (code [0] == 0xe8) {
122 if (!mono_running_on_valgrind ()) {
125 * Thread safe code patching using the algorithm from the paper
126 * 'Practicing JUDO: Java Under Dynamic Optimizations'
129 * First atomically change the the first 2 bytes of the call to a
133 InterlockedExchange ((gint32*)code, ops);
135 /* Then change the other bytes to a nop */
140 /* Then atomically change the first 4 bytes to a nop as well */
142 InterlockedExchange ((gint32*)code, ops);
143 #ifdef HAVE_VALGRIND_MEMCHECK_H
144 /* FIXME: the calltree skin trips on the self modifying code above */
146 /* Tell valgrind to recompile the patched code */
147 //VALGRIND_DISCARD_TRANSLATIONS (code, code + 8);
150 } else if (code [0] == 0x90 || code [0] == 0xeb) {
151 /* Already changed by another thread */
153 } else if ((code [-1] == 0xff) && (x86_modrm_reg (code [0]) == 0x2)) {
154 /* call *<OFFSET>(<REG>) -> Call made from AOT code */
155 gpointer *vtable_slot;
157 vtable_slot = mono_arch_get_vcall_slot_addr (code + 5, (gpointer*)regs);
158 g_assert (vtable_slot);
160 *vtable_slot = nullified_class_init_trampoline;
162 printf ("Invalid trampoline sequence: %x %x %x %x %x %x %x\n", code [0], code [1], code [2], code [3],
163 code [4], code [5], code [6]);
164 g_assert_not_reached ();
169 mono_arch_nullify_plt_entry (guint8 *code)
171 if (!mono_running_on_valgrind ()) {
175 InterlockedExchange ((gint32*)code, ops);
177 /* Then change the other bytes to a nop */
182 /* Change the first byte to a nop */
184 InterlockedExchange ((gint32*)code, ops);
189 mono_arch_create_trampoline_code (MonoTrampolineType tramp_type)
191 guint8 *buf, *code, *tramp;
192 int pushed_args, pushed_args_caller_saved;
194 code = buf = mono_global_codeman_reserve (256);
196 /* Note that there is a single argument to the trampoline
197 * and it is stored at: esp + pushed_args * sizeof (gpointer)
198 * the ret address is at: esp + (pushed_args + 1) * sizeof (gpointer)
201 /* Put all registers into an array on the stack
202 * If this code is changed, make sure to update the offset value in
203 * mono_arch_find_this_argument () in mini-x86.c.
205 x86_push_reg (buf, X86_EDI);
206 x86_push_reg (buf, X86_ESI);
207 x86_push_reg (buf, X86_EBP);
208 x86_push_reg (buf, X86_ESP);
209 x86_push_reg (buf, X86_EBX);
210 x86_push_reg (buf, X86_EDX);
211 x86_push_reg (buf, X86_ECX);
212 x86_push_reg (buf, X86_EAX);
214 pushed_args_caller_saved = pushed_args = 8;
216 /* Align stack on apple */
217 x86_alu_reg_imm (buf, X86_SUB, X86_ESP, 4);
223 /* save the IP (caller ip) */
224 if (tramp_type == MONO_TRAMPOLINE_JUMP)
225 x86_push_imm (buf, 0);
227 x86_push_membase (buf, X86_ESP, (pushed_args + 1) * sizeof (gpointer));
231 x86_push_reg (buf, X86_EBP);
232 x86_push_reg (buf, X86_ESI);
233 x86_push_reg (buf, X86_EDI);
234 x86_push_reg (buf, X86_EBX);
239 x86_push_reg (buf, X86_ESP);
240 /* Adjust ESP so it points to the previous frame */
241 x86_alu_membase_imm (buf, X86_ADD, X86_ESP, 0, (pushed_args + 2) * 4);
245 /* save method info */
246 if ((tramp_type == MONO_TRAMPOLINE_JIT) || (tramp_type == MONO_TRAMPOLINE_JUMP))
247 x86_push_membase (buf, X86_ESP, pushed_args * sizeof (gpointer));
249 x86_push_imm (buf, 0);
253 /* On apple, the stack is correctly aligned to 16 bytes because pushed_args is
254 * 16 and there is the extra trampoline arg + the return ip pushed by call
255 * FIXME: Note that if an exception happens while some args are pushed
256 * on the stack, the stack will be misaligned.
258 g_assert (pushed_args == 16);
260 /* get the address of lmf for the current thread */
261 x86_call_code (buf, mono_get_lmf_addr);
263 x86_push_reg (buf, X86_EAX);
264 /* push *lfm (previous_lmf) */
265 x86_push_membase (buf, X86_EAX, 0);
266 /* Signal to mono_arch_find_jit_info () that this is a trampoline frame */
267 x86_alu_membase_imm (buf, X86_ADD, X86_ESP, 0, 1);
269 x86_mov_membase_reg (buf, X86_EAX, 0, X86_ESP, 4);
274 /* starting the call sequence */
276 /* FIXME: Push the trampoline address */
277 x86_push_imm (buf, 0);
281 /* push the method info */
282 x86_push_membase (buf, X86_ESP, pushed_args * sizeof (gpointer));
286 /* push the return address onto the stack */
287 if (tramp_type == MONO_TRAMPOLINE_JUMP)
288 x86_push_imm (buf, 0);
290 x86_push_membase (buf, X86_ESP, (pushed_args + 1) * sizeof (gpointer));
292 /* push the address of the register array */
293 x86_lea_membase (buf, X86_EAX, X86_ESP, (pushed_args - 8) * sizeof (gpointer));
294 x86_push_reg (buf, X86_EAX);
299 /* check the stack is aligned after the ret ip is pushed */
300 /*x86_mov_reg_reg (buf, X86_EDX, X86_ESP, 4);
301 x86_alu_reg_imm (buf, X86_AND, X86_EDX, 15);
302 x86_alu_reg_imm (buf, X86_CMP, X86_EDX, 0);
303 x86_branch_disp (buf, X86_CC_Z, 3, FALSE);
304 x86_breakpoint (buf);*/
307 tramp = (guint8*)mono_get_trampoline_func (tramp_type);
308 x86_call_code (buf, tramp);
310 x86_alu_reg_imm (buf, X86_ADD, X86_ESP, 4*4);
314 /* Check for thread interruption */
315 /* This is not perf critical code so no need to check the interrupt flag */
316 x86_push_reg (buf, X86_EAX);
317 x86_call_code (buf, (guint8*)mono_thread_force_interruption_checkpoint);
318 x86_pop_reg (buf, X86_EAX);
322 /* ebx = previous_lmf */
323 x86_pop_reg (buf, X86_EBX);
325 x86_alu_reg_imm (buf, X86_SUB, X86_EBX, 1);
328 x86_pop_reg (buf, X86_EDI);
331 /* *(lmf) = previous_lmf */
332 x86_mov_membase_reg (buf, X86_EDI, 0, X86_EBX, 4);
334 /* discard method info */
335 x86_pop_reg (buf, X86_ESI);
339 x86_pop_reg (buf, X86_ESI);
342 /* restore caller saved regs */
343 x86_pop_reg (buf, X86_EBX);
344 x86_pop_reg (buf, X86_EDI);
345 x86_pop_reg (buf, X86_ESI);
346 x86_pop_reg (buf, X86_EBP);
350 /* discard save IP */
351 x86_alu_reg_imm (buf, X86_ADD, X86_ESP, 4);
354 /* restore LMF end */
356 /* Restore caller saved registers */
357 x86_mov_reg_membase (buf, X86_ECX, X86_ESP, (pushed_args - pushed_args_caller_saved + X86_ECX) * 4, 4);
358 x86_mov_reg_membase (buf, X86_EDX, X86_ESP, (pushed_args - pushed_args_caller_saved + X86_EDX) * 4, 4);
360 /* Pop saved reg array + stack align + method ptr */
361 x86_alu_reg_imm (buf, X86_ADD, X86_ESP, 10 * 4);
365 /* We've popped one more stack item than we've pushed (the
366 method ptr argument), so we must end up at -1. */
367 g_assert (pushed_args == -1);
369 if (tramp_type == MONO_TRAMPOLINE_CLASS_INIT ||
370 tramp_type == MONO_TRAMPOLINE_GENERIC_CLASS_INIT ||
371 tramp_type == MONO_TRAMPOLINE_RGCTX_LAZY_FETCH)
374 /* call the compiled method */
375 x86_jump_reg (buf, X86_EAX);
377 g_assert ((buf - code) <= 256);
379 if (tramp_type == MONO_TRAMPOLINE_CLASS_INIT) {
380 /* Initialize the nullified class init trampoline used in the AOT case */
381 nullified_class_init_trampoline = buf = mono_global_codeman_reserve (16);
388 #define TRAMPOLINE_SIZE 10
391 mono_arch_create_specific_trampoline (gpointer arg1, MonoTrampolineType tramp_type, MonoDomain *domain, guint32 *code_len)
393 guint8 *code, *buf, *tramp;
395 tramp = mono_get_trampoline_code (tramp_type);
397 mono_domain_lock (domain);
398 code = buf = mono_code_manager_reserve_align (domain->code_mp, TRAMPOLINE_SIZE, 4);
399 mono_domain_unlock (domain);
401 x86_push_imm (buf, arg1);
402 x86_jump_code (buf, tramp);
403 g_assert ((buf - code) <= TRAMPOLINE_SIZE);
405 mono_arch_flush_icache (code, buf - code);
408 *code_len = buf - code;
414 mono_arch_create_rgctx_lazy_fetch_trampoline (guint32 slot)
418 guint8 **rgctx_null_jumps;
424 mrgctx = MONO_RGCTX_SLOT_IS_MRGCTX (slot);
425 index = MONO_RGCTX_SLOT_INDEX (slot);
427 index += sizeof (MonoMethodRuntimeGenericContext) / sizeof (gpointer);
428 for (depth = 0; ; ++depth) {
429 int size = mono_class_rgctx_get_array_size (depth, mrgctx);
431 if (index < size - 1)
436 tramp_size = 36 + 6 * depth;
438 code = buf = mono_global_codeman_reserve (tramp_size);
440 rgctx_null_jumps = g_malloc (sizeof (guint8*) * (depth + 2));
442 /* load vtable/mrgctx ptr */
443 x86_mov_reg_membase (buf, X86_EAX, X86_ESP, 4, 4);
445 /* load rgctx ptr from vtable */
446 x86_mov_reg_membase (buf, X86_EAX, X86_EAX, G_STRUCT_OFFSET (MonoVTable, runtime_generic_context), 4);
447 /* is the rgctx ptr null? */
448 x86_test_reg_reg (buf, X86_EAX, X86_EAX);
449 /* if yes, jump to actual trampoline */
450 rgctx_null_jumps [0] = buf;
451 x86_branch8 (buf, X86_CC_Z, -1, 1);
454 for (i = 0; i < depth; ++i) {
455 /* load ptr to next array */
456 if (mrgctx && i == 0)
457 x86_mov_reg_membase (buf, X86_EAX, X86_EAX, sizeof (MonoMethodRuntimeGenericContext), 4);
459 x86_mov_reg_membase (buf, X86_EAX, X86_EAX, 0, 4);
460 /* is the ptr null? */
461 x86_test_reg_reg (buf, X86_EAX, X86_EAX);
462 /* if yes, jump to actual trampoline */
463 rgctx_null_jumps [i + 1] = buf;
464 x86_branch8 (buf, X86_CC_Z, -1, 1);
468 x86_mov_reg_membase (buf, X86_EAX, X86_EAX, sizeof (gpointer) * (index + 1), 4);
469 /* is the slot null? */
470 x86_test_reg_reg (buf, X86_EAX, X86_EAX);
471 /* if yes, jump to actual trampoline */
472 rgctx_null_jumps [depth + 1] = buf;
473 x86_branch8 (buf, X86_CC_Z, -1, 1);
474 /* otherwise return */
477 for (i = mrgctx ? 1 : 0; i <= depth + 1; ++i)
478 x86_patch (rgctx_null_jumps [i], buf);
480 g_free (rgctx_null_jumps);
482 x86_mov_reg_membase (buf, MONO_ARCH_VTABLE_REG, X86_ESP, 4, 4);
484 tramp = mono_arch_create_specific_trampoline (GUINT_TO_POINTER (slot), MONO_TRAMPOLINE_RGCTX_LAZY_FETCH, mono_get_root_domain (), NULL);
486 /* jump to the actual trampoline */
487 x86_jump_code (buf, tramp);
489 mono_arch_flush_icache (code, buf - code);
491 g_assert (buf - code <= tramp_size);
497 mono_arch_create_generic_class_init_trampoline (void)
501 static int byte_offset = -1;
502 static guint8 bitmask;
508 code = buf = mono_global_codeman_reserve (tramp_size);
511 mono_marshal_find_bitfield_offset (MonoVTable, initialized, &byte_offset, &bitmask);
513 x86_test_membase_imm (code, MONO_ARCH_VTABLE_REG, byte_offset, bitmask);
515 x86_branch8 (code, X86_CC_Z, -1, 1);
519 x86_patch (jump, code);
521 /* Push the vtable so the stack is the same as in a specific trampoline */
522 x86_push_reg (code, MONO_ARCH_VTABLE_REG);
524 tramp = mono_get_trampoline_code (MONO_TRAMPOLINE_GENERIC_CLASS_INIT);
526 /* jump to the actual trampoline */
527 x86_jump_code (code, tramp);
529 mono_arch_flush_icache (code, code - buf);
531 g_assert (code - buf <= tramp_size);
537 mono_arch_invalidate_method (MonoJitInfo *ji, void *func, gpointer func_arg)
539 /* FIXME: This is not thread safe */
540 guint8 *code = ji->code_start;
542 x86_push_imm (code, func_arg);
543 x86_call_code (code, (guint8*)func);