Exemple #1
0
gpointer
mono_arch_get_gsharedvt_trampoline (MonoTrampInfo **info, gboolean aot)
{
	guint8 *code, *buf;
	int buf_len, cfa_offset;
	GSList *unwind_ops = NULL;
	MonoJumpInfo *ji = NULL;
	guint8 *br_out, *br [16];
	int info_offset, mrgctx_offset;

	buf_len = 320;
	buf = code = mono_global_codeman_reserve (buf_len);

	/*
	 * This trampoline is responsible for marshalling calls between normal code and gsharedvt code. The
	 * caller is a normal or gshared method which uses the signature of the inflated method to make the call, while
	 * the callee is a gsharedvt method which has a signature which uses valuetypes in place of type parameters, i.e.
	 * caller:
	 * foo<bool> (bool b)
	 * callee:
	 * T=<type used to represent vtype type arguments, currently TypedByRef>
	 * foo<T> (T b)
	 * The trampoline is responsible for marshalling the arguments and marshalling the result back. To simplify
	 * things, we create our own stack frame, and do most of the work in a C function, which receives a
	 * GSharedVtCallInfo structure as an argument. The structure should contain information to execute the C function to
	 * be as fast as possible. The argument is received in EAX from a gsharedvt trampoline. So the real
	 * call sequence looks like this:
	 * caller -> gsharedvt trampoline -> gsharevt in trampoline -> start_gsharedvt_call
	 * FIXME: Optimize this.
	 */

	cfa_offset = sizeof (gpointer);
	mono_add_unwind_op_def_cfa (unwind_ops, code, buf, X86_ESP, cfa_offset);
	mono_add_unwind_op_offset (unwind_ops, code, buf, X86_NREG, -cfa_offset);
	x86_push_reg (code, X86_EBP);
	cfa_offset += sizeof (gpointer);
	mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
	mono_add_unwind_op_offset (unwind_ops, code, buf, X86_EBP, - cfa_offset);
	x86_mov_reg_reg (code, X86_EBP, X86_ESP, sizeof (gpointer));
	mono_add_unwind_op_def_cfa_reg (unwind_ops, code, buf, X86_EBP);
	/* Alloc stack frame/align stack */
	x86_alu_reg_imm (code, X86_SUB, X86_ESP, 8);
	info_offset = -4;
	mrgctx_offset = - 8;
	/* The info struct is put into EAX by the gsharedvt trampoline */
	/* Save info struct addr */
	x86_mov_membase_reg (code, X86_EBP, info_offset, X86_EAX, 4);
	/* Save rgctx */
	x86_mov_membase_reg (code, X86_EBP, mrgctx_offset, MONO_ARCH_RGCTX_REG, 4);

	/* Allocate stack area used to pass arguments to the method */
	x86_mov_reg_membase (code, X86_EAX, X86_EAX, MONO_STRUCT_OFFSET (GSharedVtCallInfo, stack_usage), sizeof (gpointer));
	x86_alu_reg_reg (code, X86_SUB, X86_ESP, X86_EAX);

#if 0
	/* Stack alignment check */
	x86_mov_reg_reg (code, X86_ECX, X86_ESP, 4);
	x86_alu_reg_imm (code, X86_AND, X86_ECX, MONO_ARCH_FRAME_ALIGNMENT - 1);
	x86_alu_reg_imm (code, X86_CMP, X86_ECX, 0);
	x86_branch_disp (code, X86_CC_EQ, 3, FALSE);
	x86_breakpoint (code);
#endif

	/* ecx = caller argument area */
	x86_mov_reg_reg (code, X86_ECX, X86_EBP, 4);
	x86_alu_reg_imm (code, X86_ADD, X86_ECX, 8);
	/* eax = callee argument area */
	x86_mov_reg_reg (code, X86_EAX, X86_ESP, 4);

	/* Call start_gsharedvt_call */
	/* Arg 4 */
	x86_push_membase (code, X86_EBP, mrgctx_offset);
	/* Arg3 */
	x86_push_reg (code, X86_EAX);
	/* Arg2 */
	x86_push_reg (code, X86_ECX);
	/* Arg1 */
	x86_push_membase (code, X86_EBP, info_offset);
	if (aot) {
		code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_x86_start_gsharedvt_call");
		x86_call_reg (code, X86_EAX);
	} else {
		x86_call_code (code, mono_x86_start_gsharedvt_call);
	}
	x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4 * 4);
	/* The address to call is in eax */
	/* The stack is now setup for the real call */
	/* Load info struct */
	x86_mov_reg_membase (code, X86_ECX, X86_EBP, info_offset, 4);
	/* Load rgctx */
	x86_mov_reg_membase (code, MONO_ARCH_RGCTX_REG, X86_EBP, mrgctx_offset, sizeof (gpointer));
	/* Make the call */
	x86_call_reg (code, X86_EAX);
	/* The return value is either in registers, or stored to an area beginning at sp [info->vret_slot] */
	/* EAX/EDX might contain the return value, only ECX is free */
	/* Load info struct */
	x86_mov_reg_membase (code, X86_ECX, X86_EBP, info_offset, 4);

	/* Branch to the in/out handling code */
	x86_alu_membase_imm (code, X86_CMP, X86_ECX, MONO_STRUCT_OFFSET (GSharedVtCallInfo, gsharedvt_in), 1);	
	br_out = code;
	x86_branch32 (code, X86_CC_NE, 0, TRUE);

	/*
	 * IN CASE
	 */

	/* Load ret marshal type */
	x86_mov_reg_membase (code, X86_ECX, X86_ECX, MONO_STRUCT_OFFSET (GSharedVtCallInfo, ret_marshal), 4);
	x86_alu_reg_imm (code, X86_CMP, X86_ECX, GSHAREDVT_RET_NONE);
	br [0] = code;
	x86_branch8 (code, X86_CC_NE, 0, TRUE);

	/* Normal return, no marshalling required */
	x86_leave (code);
	x86_ret (code);

	/* Return value marshalling */
	x86_patch (br [0], code);
	/* Load info struct */
	x86_mov_reg_membase (code, X86_EAX, X86_EBP, info_offset, 4);
	/* Load 'vret_slot' */
	x86_mov_reg_membase (code, X86_EAX, X86_EAX, MONO_STRUCT_OFFSET (GSharedVtCallInfo, vret_slot), 4);
	/* Compute ret area address */
	x86_shift_reg_imm (code, X86_SHL, X86_EAX, 2);
	x86_alu_reg_reg (code, X86_ADD, X86_EAX, X86_ESP);
	/* The callee does a ret $4, so sp is off by 4 */
	x86_alu_reg_imm (code, X86_SUB, X86_EAX, sizeof (gpointer));

	/* Branch to specific marshalling code */
	// FIXME: Move the I4 case to the top */
	x86_alu_reg_imm (code, X86_CMP, X86_ECX, GSHAREDVT_RET_DOUBLE_FPSTACK);
	br [1] = code;
	x86_branch8 (code, X86_CC_E, 0, TRUE);
	x86_alu_reg_imm (code, X86_CMP, X86_ECX, GSHAREDVT_RET_FLOAT_FPSTACK);
	br [2] = code;
	x86_branch8 (code, X86_CC_E, 0, TRUE);
	x86_alu_reg_imm (code, X86_CMP, X86_ECX, GSHAREDVT_RET_STACK_POP);
	br [3] = code;
	x86_branch8 (code, X86_CC_E, 0, TRUE);
	x86_alu_reg_imm (code, X86_CMP, X86_ECX, GSHAREDVT_RET_I1);
	br [4] = code;
	x86_branch8 (code, X86_CC_E, 0, TRUE);
	x86_alu_reg_imm (code, X86_CMP, X86_ECX, GSHAREDVT_RET_U1);
	br [5] = code;
	x86_branch8 (code, X86_CC_E, 0, TRUE);
	x86_alu_reg_imm (code, X86_CMP, X86_ECX, GSHAREDVT_RET_I2);
	br [6] = code;
	x86_branch8 (code, X86_CC_E, 0, TRUE);
	x86_alu_reg_imm (code, X86_CMP, X86_ECX, GSHAREDVT_RET_U2);
	br [7] = code;
	x86_branch8 (code, X86_CC_E, 0, TRUE);
	/* IREGS case */
	/* Load both eax and edx for simplicity */
	x86_mov_reg_membase (code, X86_EDX, X86_EAX, sizeof (gpointer), sizeof (gpointer));
	x86_mov_reg_membase (code, X86_EAX, X86_EAX, 0, sizeof (gpointer));
	x86_leave (code);
	x86_ret (code);
	/* DOUBLE_FPSTACK case */
	x86_patch (br [1], code);
	x86_fld_membase (code, X86_EAX, 0, TRUE);
	x86_jump8 (code, 0);
	x86_leave (code);
	x86_ret (code);
	/* FLOAT_FPSTACK case */
	x86_patch (br [2], code);
	x86_fld_membase (code, X86_EAX, 0, FALSE);
	x86_leave (code);
	x86_ret (code);
	/* STACK_POP case */
	x86_patch (br [3], code);
	x86_leave (code);
	x86_ret_imm (code, 4);
	/* I1 case */
	x86_patch (br [4], code);
	x86_widen_membase (code, X86_EAX, X86_EAX, 0, TRUE, FALSE);
	x86_leave (code);
	x86_ret (code);
	/* U1 case */
	x86_patch (br [5], code);
	x86_widen_membase (code, X86_EAX, X86_EAX, 0, FALSE, FALSE);
	x86_leave (code);
	x86_ret (code);
	/* I2 case */
	x86_patch (br [6], code);
	x86_widen_membase (code, X86_EAX, X86_EAX, 0, TRUE, TRUE);
	x86_leave (code);
	x86_ret (code);
	/* U2 case */
	x86_patch (br [7], code);
	x86_widen_membase (code, X86_EAX, X86_EAX, 0, FALSE, TRUE);
	x86_leave (code);
	x86_ret (code);

	/*
	 * OUT CASE
	 */

	x86_patch (br_out, code);
	/* Load ret marshal type into ECX */
	x86_mov_reg_membase (code, X86_ECX, X86_ECX, MONO_STRUCT_OFFSET (GSharedVtCallInfo, ret_marshal), 4);
	x86_alu_reg_imm (code, X86_CMP, X86_ECX, GSHAREDVT_RET_NONE);
	br [0] = code;
	x86_branch8 (code, X86_CC_NE, 0, TRUE);

	/* Normal return, no marshalling required */
	x86_leave (code);
	x86_ret (code);

	/* Return value marshalling */
	x86_patch (br [0], code);

	/* EAX might contain the return value */
	// FIXME: Use moves
	x86_push_reg (code, X86_EAX);

	/* Load info struct */
	x86_mov_reg_membase (code, X86_EAX, X86_EBP, info_offset, 4);
	/* Load 'vret_arg_slot' */
	x86_mov_reg_membase (code, X86_EAX, X86_EAX, MONO_STRUCT_OFFSET (GSharedVtCallInfo, vret_arg_slot), 4);
	/* Compute ret area address in the caller frame in EAX */
	x86_shift_reg_imm (code, X86_SHL, X86_EAX, 2);
	x86_alu_reg_reg (code, X86_ADD, X86_EAX, X86_EBP);
	x86_alu_reg_imm (code, X86_ADD, X86_EAX, 8);
	x86_mov_reg_membase (code, X86_EAX, X86_EAX, 0, sizeof (gpointer));

	/* Branch to specific marshalling code */
	x86_alu_reg_imm (code, X86_CMP, X86_ECX, GSHAREDVT_RET_DOUBLE_FPSTACK);
	br [1] = code;
	x86_branch8 (code, X86_CC_E, 0, TRUE);
	x86_alu_reg_imm (code, X86_CMP, X86_ECX, GSHAREDVT_RET_FLOAT_FPSTACK);
	br [2] = code;
	x86_branch8 (code, X86_CC_E, 0, TRUE);
	x86_alu_reg_imm (code, X86_CMP, X86_ECX, GSHAREDVT_RET_STACK_POP);
	br [3] = code;
	x86_branch8 (code, X86_CC_E, 0, TRUE);
	x86_alu_reg_imm (code, X86_CMP, X86_ECX, GSHAREDVT_RET_IREGS);
	br [4] = code;
	x86_branch8 (code, X86_CC_E, 0, TRUE);
	/* IREG case */
	x86_mov_reg_reg (code, X86_ECX, X86_EAX, sizeof (gpointer));
	x86_pop_reg (code, X86_EAX);
	x86_mov_membase_reg (code, X86_ECX, 0, X86_EAX, sizeof (gpointer));
	x86_leave (code);
	x86_ret_imm (code, 4);
	/* IREGS case */
	x86_patch (br [4], code);
	x86_mov_reg_reg (code, X86_ECX, X86_EAX, sizeof (gpointer));
	x86_pop_reg (code, X86_EAX);
	x86_mov_membase_reg (code, X86_ECX, sizeof (gpointer), X86_EDX, sizeof (gpointer));
	x86_mov_membase_reg (code, X86_ECX, 0, X86_EAX, sizeof (gpointer));
	x86_leave (code);
	x86_ret_imm (code, 4);
	/* DOUBLE_FPSTACK case */
	x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
	x86_patch (br [1], code);
	x86_fst_membase (code, X86_EAX, 0, TRUE, TRUE);
	x86_jump8 (code, 0);
	x86_leave (code);
	x86_ret_imm (code, 4);
	/* FLOAT_FPSTACK case */
	x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
	x86_patch (br [2], code);
	x86_fst_membase (code, X86_EAX, 0, FALSE, TRUE);
	x86_leave (code);
	x86_ret_imm (code, 4);
	/* STACK_POP case */
	x86_patch (br [3], code);
	x86_leave (code);
	x86_ret_imm (code, 4);

	g_assert ((code - buf) < buf_len);

	if (info)
		*info = mono_tramp_info_create ("gsharedvt_trampoline", buf, code - buf, ji, unwind_ops);

	mono_arch_flush_icache (buf, code - buf);
	return buf;
}
Exemple #2
0
guchar*
mono_arch_create_generic_trampoline (MonoTrampolineType tramp_type, MonoTrampInfo **info, gboolean aot)
{
	const char *tramp_name;
	guint8 *buf, *code, *tramp, *br_ex_check;
	GSList *unwind_ops = NULL;
	MonoJumpInfo *ji = NULL;
	int i, offset, frame_size, regarray_offset, lmf_offset, caller_ip_offset, arg_offset;
	int cfa_offset; /* cfa = cfa_reg + cfa_offset */

	code = buf = mono_global_codeman_reserve (256);

	/* Note that there is a single argument to the trampoline
	 * and it is stored at: esp + pushed_args * sizeof (target_mgreg_t)
	 * the ret address is at: esp + (pushed_args + 1) * sizeof (target_mgreg_t)
	 */

	/* Compute frame offsets relative to the frame pointer %ebp */
	arg_offset = sizeof (target_mgreg_t);
	caller_ip_offset = 2 * sizeof (target_mgreg_t);
	offset = 0;
	offset += sizeof (MonoLMF);
	lmf_offset = -offset;
	offset += X86_NREG * sizeof (target_mgreg_t);
	regarray_offset = -offset;
	/* Argument area */
	offset += 4 * sizeof (target_mgreg_t);
	frame_size = ALIGN_TO (offset, MONO_ARCH_FRAME_ALIGNMENT);

	/* ret addr and arg are on the stack */
	cfa_offset = 2 * sizeof (target_mgreg_t);
	mono_add_unwind_op_def_cfa (unwind_ops, code, buf, X86_ESP, cfa_offset);
	// IP saved at CFA - 4
	mono_add_unwind_op_offset (unwind_ops, code, buf, X86_NREG, -4);

	/* Allocate frame */
	x86_push_reg (code, X86_EBP);
	cfa_offset += sizeof (target_mgreg_t);
	mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
	mono_add_unwind_op_offset (unwind_ops, code, buf, X86_EBP, -cfa_offset);

	x86_mov_reg_reg (code, X86_EBP, X86_ESP);
	mono_add_unwind_op_def_cfa_reg (unwind_ops, code, buf, X86_EBP);

	/* There are three words on the stack, adding + 4 aligns the stack to 16, which is needed on osx */
	x86_alu_reg_imm (code, X86_SUB, X86_ESP, frame_size + sizeof (target_mgreg_t));

	/* Save all registers */
	for (i = X86_EAX; i <= X86_EDI; ++i) {
		int reg = i;

		if (i == X86_EBP) {
			/* Save original ebp */
			/* EAX is already saved */
			x86_mov_reg_membase (code, X86_EAX, X86_EBP, 0, sizeof (target_mgreg_t));
			reg = X86_EAX;
		} else if (i == X86_ESP) {
			/* Save original esp */
			/* EAX is already saved */
			x86_mov_reg_reg (code, X86_EAX, X86_EBP);
			/* Saved ebp + trampoline arg + return addr */
			x86_alu_reg_imm (code, X86_ADD, X86_EAX, 3 * sizeof (target_mgreg_t));
			reg = X86_EAX;
		}
		x86_mov_membase_reg (code, X86_EBP, regarray_offset + (i * sizeof (target_mgreg_t)), reg, sizeof (target_mgreg_t));
	}

	/* Setup LMF */
	/* eip */
	if (tramp_type == MONO_TRAMPOLINE_JUMP) {
		x86_mov_membase_imm (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, eip), 0, sizeof (target_mgreg_t));
	} else {
		x86_mov_reg_membase (code, X86_EAX, X86_EBP, caller_ip_offset, sizeof (target_mgreg_t));
		x86_mov_membase_reg (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, eip), X86_EAX, sizeof (target_mgreg_t));
	}
	/* method */
	if ((tramp_type == MONO_TRAMPOLINE_JIT) || (tramp_type == MONO_TRAMPOLINE_JUMP)) {
		x86_mov_reg_membase (code, X86_EAX, X86_EBP, arg_offset, sizeof (target_mgreg_t));
		x86_mov_membase_reg (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), X86_EAX, sizeof (target_mgreg_t));
	} else {
		x86_mov_membase_imm (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, method), 0, sizeof (target_mgreg_t));
	}
	/* esp */
	x86_mov_reg_membase (code, X86_EAX, X86_EBP, regarray_offset + (X86_ESP * sizeof (target_mgreg_t)), sizeof (target_mgreg_t));
	x86_mov_membase_reg (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, esp), X86_EAX, sizeof (target_mgreg_t));
	/* callee save registers */
	x86_mov_reg_membase (code, X86_EAX, X86_EBP, regarray_offset + (X86_EBX * sizeof (target_mgreg_t)), sizeof (target_mgreg_t));
	x86_mov_membase_reg (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, ebx), X86_EAX, sizeof (target_mgreg_t));
	x86_mov_reg_membase (code, X86_EAX, X86_EBP, regarray_offset + (X86_EDI * sizeof (target_mgreg_t)), sizeof (target_mgreg_t));
	x86_mov_membase_reg (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, edi), X86_EAX, sizeof (target_mgreg_t));
	x86_mov_reg_membase (code, X86_EAX, X86_EBP, regarray_offset + (X86_ESI * sizeof (target_mgreg_t)), sizeof (target_mgreg_t));
	x86_mov_membase_reg (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, esi), X86_EAX, sizeof (target_mgreg_t));
	x86_mov_reg_membase (code, X86_EAX, X86_EBP, regarray_offset + (X86_EBP * sizeof (target_mgreg_t)), sizeof (target_mgreg_t));
	x86_mov_membase_reg (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, ebp), X86_EAX, sizeof (target_mgreg_t));

	/* Push LMF */
	/* get the address of lmf for the current thread */
	if (aot) {
		code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_get_lmf_addr");
		x86_call_reg (code, X86_EAX);
	} else {
		x86_call_code (code, mono_get_lmf_addr);
	}
	/* lmf->lmf_addr = lmf_addr (%eax) */
	x86_mov_membase_reg (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), X86_EAX, sizeof (target_mgreg_t));
	/* lmf->previous_lmf = *(lmf_addr) */
	x86_mov_reg_membase (code, X86_ECX, X86_EAX, 0, sizeof (target_mgreg_t));
	/* Signal to mono_arch_unwind_frame () that this is a trampoline frame */
	x86_alu_reg_imm (code, X86_ADD, X86_ECX, 1);
	x86_mov_membase_reg (code, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), X86_ECX, sizeof (target_mgreg_t));
	/* *lmf_addr = lmf */
	x86_lea_membase (code, X86_ECX, X86_EBP, lmf_offset);
	x86_mov_membase_reg (code, X86_EAX, 0, X86_ECX, sizeof (target_mgreg_t));

	/* Call trampoline function */
	/* Arg 1 - registers */
	x86_lea_membase (code, X86_EAX, X86_EBP, regarray_offset);
	x86_mov_membase_reg (code, X86_ESP, (0 * sizeof (target_mgreg_t)), X86_EAX, sizeof (target_mgreg_t));
	/* Arg2 - calling code */
	if (tramp_type == MONO_TRAMPOLINE_JUMP) {
		x86_mov_membase_imm (code, X86_ESP, (1 * sizeof (target_mgreg_t)), 0, sizeof (target_mgreg_t));
	} else {
		x86_mov_reg_membase (code, X86_EAX, X86_EBP, caller_ip_offset, sizeof (target_mgreg_t));
		x86_mov_membase_reg (code, X86_ESP, (1 * sizeof (target_mgreg_t)), X86_EAX, sizeof (target_mgreg_t));
	}
	/* Arg3 - trampoline argument */
	x86_mov_reg_membase (code, X86_EAX, X86_EBP, arg_offset, sizeof (target_mgreg_t));
	x86_mov_membase_reg (code, X86_ESP, (2 * sizeof (target_mgreg_t)), X86_EAX, sizeof (target_mgreg_t));
	/* Arg4 - trampoline address */
	// FIXME:
	x86_mov_membase_imm (code, X86_ESP, (3 * sizeof (target_mgreg_t)), 0, sizeof (target_mgreg_t));

#ifdef __APPLE__
	/* check the stack is aligned after the ret ip is pushed */
	/*
	x86_mov_reg_reg (code, X86_EDX, X86_ESP);
	x86_alu_reg_imm (code, X86_AND, X86_EDX, 15);
	x86_alu_reg_imm (code, X86_CMP, X86_EDX, 0);
	x86_branch_disp (code, X86_CC_Z, 3, FALSE);
	x86_breakpoint (code);
	*/
#endif

	if (aot) {
		code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_TRAMPOLINE_FUNC_ADDR, GINT_TO_POINTER (tramp_type));
		x86_call_reg (code, X86_EAX);
	} else {
		tramp = (guint8*)mono_get_trampoline_func (tramp_type);
		x86_call_code (code, tramp);
	}

	/*
	 * Overwrite the trampoline argument with the address we need to jump to,
	 * to free %eax.
	 */
	x86_mov_membase_reg (code, X86_EBP, arg_offset, X86_EAX, 4);

	/* Restore LMF */
	x86_mov_reg_membase (code, X86_EAX, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, lmf_addr), sizeof (target_mgreg_t));
	x86_mov_reg_membase (code, X86_ECX, X86_EBP, lmf_offset + G_STRUCT_OFFSET (MonoLMF, previous_lmf), sizeof (target_mgreg_t));
	x86_alu_reg_imm (code, X86_SUB, X86_ECX, 1);
	x86_mov_membase_reg (code, X86_EAX, 0, X86_ECX, sizeof (target_mgreg_t));

	/* Check for interruptions */
	if (aot) {
		code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "mono_thread_force_interruption_checkpoint_noraise");
		x86_call_reg (code, X86_EAX);
	} else {
		x86_call_code (code, (guint8*)mono_thread_force_interruption_checkpoint_noraise);
	}

	x86_test_reg_reg (code, X86_EAX, X86_EAX);
	br_ex_check = code;
	x86_branch8 (code, X86_CC_Z, -1, 1);

	/*
	 * Exception case:
	 * We have an exception we want to throw in the caller's frame, so pop
	 * the trampoline frame and throw from the caller.
	 */
	x86_leave (code);
	/*
	 * The exception is in eax.
	 * We are calling the throw trampoline used by OP_THROW, so we have to setup the
	 * stack to look the same.
	 * The stack contains the ret addr, and the trampoline argument, the throw trampoline
	 * expects it to contain the ret addr and the exception. It also needs to be aligned
	 * after the exception is pushed.
	 */
	/* Align stack */
	x86_push_reg (code, X86_EAX);
	/* Push the exception */
	x86_push_reg (code, X86_EAX);
	//x86_breakpoint (code);
	/* Push the original return value */
	x86_push_membase (code, X86_ESP, 3 * 4);
	/*
	 * EH is initialized after trampolines, so get the address of the variable
	 * which contains throw_exception, and load it from there.
	 */
	if (aot) {
		/* Not really a jit icall */
		code = mono_arch_emit_load_aotconst (buf, code, &ji, MONO_PATCH_INFO_JIT_ICALL_ADDR, "rethrow_preserve_exception_addr");
	} else {
		x86_mov_reg_imm (code, X86_ECX, (guint8*)mono_get_rethrow_preserve_exception_addr ());
	}
	x86_mov_reg_membase (code, X86_ECX, X86_ECX, 0, sizeof (target_mgreg_t));
	x86_jump_reg (code, X86_ECX);

	/* Normal case */
	mono_x86_patch (br_ex_check, code);

	/* Restore registers */
	for (i = X86_EAX; i <= X86_EDI; ++i) {
		if (i == X86_ESP || i == X86_EBP)
			continue;
		if (i == X86_EAX && tramp_type != MONO_TRAMPOLINE_AOT_PLT)
			continue;
		x86_mov_reg_membase (code, i, X86_EBP, regarray_offset + (i * 4), 4);
	}

	/* Restore frame */
	x86_leave (code);
	cfa_offset -= sizeof (target_mgreg_t);
	mono_add_unwind_op_def_cfa (unwind_ops, code, buf, X86_ESP, cfa_offset);
	mono_add_unwind_op_same_value (unwind_ops, code, buf, X86_EBP);

	if (MONO_TRAMPOLINE_TYPE_MUST_RETURN (tramp_type)) {
		/* Load the value returned by the trampoline */
		x86_mov_reg_membase (code, X86_EAX, X86_ESP, 0, 4);
		/* The trampoline returns normally, pop the trampoline argument */
		x86_alu_reg_imm (code, X86_ADD, X86_ESP, 4);
		cfa_offset -= sizeof (target_mgreg_t);
		mono_add_unwind_op_def_cfa_offset (unwind_ops, code, buf, cfa_offset);
		x86_ret (code);
	} else {
		x86_ret (code);
	}

	g_assert ((code - buf) <= 256);
	MONO_PROFILER_RAISE (jit_code_buffer, (buf, code - buf, MONO_PROFILER_CODE_BUFFER_HELPER, NULL));

	tramp_name = mono_get_generic_trampoline_name (tramp_type);
	*info = mono_tramp_info_create (tramp_name, buf, code - buf, ji, unwind_ops);

	return buf;
}