amd64: vararg calls need the number of SSE registers used

ir/be/amd64/amd64_cconv.c

@@ -202,6 +202,7 @@ amd64_cconv_t *amd64_decide_calling_convention(ir_type *function_type,
 	cconv->parameters       = params;
 	cconv->param_stack_size = stack_offset;
 	cconv->n_param_regs     = n_param_regs_used;
+	cconv->n_xmm_regs       = float_param_regnum;
 	cconv->results          = results;
 	cconv->omit_fp          = omit_fp;
 	cconv->caller_saves     = caller_saves;

ir/be/amd64/amd64_cconv.h

@@ -40,7 +40,8 @@ typedef struct amd64_cconv_t
 	reg_or_stackslot_t *parameters;       /**< parameter info. */
 	unsigned            param_stack_size; /**< stack size for parameters */
 	unsigned            n_param_regs;     /**< number of values passed in a
-	                                           register */
+	                                           register (gp + xmm) */
+	unsigned            n_xmm_regs;       /**< number of xmm registers used */
 	reg_or_stackslot_t *results;          /**< result info. */
 	unsigned            n_reg_results;
 	unsigned           *caller_saves;     /**< bitset of caller saved registers */

ir/be/amd64/amd64_transform.c

@@ -1632,9 +1632,11 @@ static ir_node *gen_Call(ir_node *node)
 
 	/* vararg calls need the number of SSE registers used */
 	if (get_method_variadicity(type) == variadicity_variadic) {
-		ir_node *zero = new_bd_amd64_Xor0(NULL, block);
+		unsigned xmm_regs = cconv->n_xmm_regs;
+		ir_node *xmm_imm  = new_bd_amd64_MovImm(dbgi, block, INSN_MODE_32,
+		                                         xmm_regs, NULL);
 		in_req[in_arity] = amd64_registers[REG_RAX].single_req;
-		in[in_arity] = zero;
+		in[in_arity] = xmm_imm;
 		++in_arity;
 	}