在C的早期版本中，所有结构声明都共享一组公共字段。考虑到:

struct x {int x_mode; int q; float x_f};
struct y {int y_mode; int q; int y_l};
struct z {int z_mode; char name[20];};

a compiler would essentially produce a table of structures' sizes (and possibly alignments), and a separate table of structures' members' names, types, and offsets. The compiler didn't keep track of which members belonged to which structures, and would allow two structures to have a member with the same name only if the type and offset matched (as with member q of struct x and struct y). If p was a pointer to any structure type, p->q would add the offset of "q" to pointer p and fetch an "int" from the resulting address.

Given the above semantics, it was possible to write a function that could perform some useful operations on multiple kinds of structure interchangeably, provided that all the fields used by the function lined up with useful fields within the structures in question. This was a useful feature, and changing C to validate members used for structure access against the types of the structures in question would have meant losing it in the absence of a means of having a structure that can contain multiple named fields at the same address. Adding "union" types to C helped fill that gap somewhat (though not, IMHO, as well as it should have been).

An essential part of unions' ability to fill that gap was the fact that a pointer to a union member could be converted into a pointer to any union containing that member, and a pointer to any union could be converted to a pointer to any member. While the C89 Standard didn't expressly say that casting a T* directly to a U* was equivalent to casting it to a pointer to any union type containing both T and U, and then casting that to U*, no defined behavior of the latter cast sequence would be affected by the union type used, and the Standard didn't specify any contrary semantics for a direct cast from T to U. Further, in cases where a function received a pointer of unknown origin, the behavior of writing an object via T*, converting the T* to a U*, and then reading the object via U* would be equivalent to writing a union via member of type T and reading as type U, which would be standard-defined in a few cases (e.g. when accessing Common Initial Sequence members) and Implementation-Defined (rather than Undefined) for the rest. While it was rare for programs to exploit the CIS guarantees with actual objects of union type, it was far more common to exploit the fact that pointers to objects of unknown origin had to behave like pointers to union members and have the behavioral guarantees associated therewith.

COM接口中使用的VARIANT呢?它有两个字段——“type”和一个包含实际值的联合，该值根据“type”字段进行处理。

包含不同记录类型的文件。 包含不同请求类型的网络接口。

看一下这个:X.25缓冲区命令处理

许多可能的X.25命令中的一个被接收到缓冲区中，并通过使用所有可能结构的UNION进行适当的处理。

在C的早期版本中，所有结构声明都共享一组公共字段。考虑到:

struct x {int x_mode; int q; float x_f};
struct y {int y_mode; int q; int y_l};
struct z {int z_mode; char name[20];};

a compiler would essentially produce a table of structures' sizes (and possibly alignments), and a separate table of structures' members' names, types, and offsets. The compiler didn't keep track of which members belonged to which structures, and would allow two structures to have a member with the same name only if the type and offset matched (as with member q of struct x and struct y). If p was a pointer to any structure type, p->q would add the offset of "q" to pointer p and fetch an "int" from the resulting address.

Given the above semantics, it was possible to write a function that could perform some useful operations on multiple kinds of structure interchangeably, provided that all the fields used by the function lined up with useful fields within the structures in question. This was a useful feature, and changing C to validate members used for structure access against the types of the structures in question would have meant losing it in the absence of a means of having a structure that can contain multiple named fields at the same address. Adding "union" types to C helped fill that gap somewhat (though not, IMHO, as well as it should have been).

An essential part of unions' ability to fill that gap was the fact that a pointer to a union member could be converted into a pointer to any union containing that member, and a pointer to any union could be converted to a pointer to any member. While the C89 Standard didn't expressly say that casting a T* directly to a U* was equivalent to casting it to a pointer to any union type containing both T and U, and then casting that to U*, no defined behavior of the latter cast sequence would be affected by the union type used, and the Standard didn't specify any contrary semantics for a direct cast from T to U. Further, in cases where a function received a pointer of unknown origin, the behavior of writing an object via T*, converting the T* to a U*, and then reading the object via U* would be equivalent to writing a union via member of type T and reading as type U, which would be standard-defined in a few cases (e.g. when accessing Common Initial Sequence members) and Implementation-Defined (rather than Undefined) for the rest. While it was rare for programs to exploit the CIS guarantees with actual objects of union type, it was far more common to exploit the fact that pointers to objects of unknown origin had to behave like pointers to union members and have the behavioral guarantees associated therewith.

这里有一个来自我自己代码库的联合的例子(来自记忆和转述，所以可能不准确)。它被用来在我构建的解释器中存储语言元素。例如，以下代码:

set a to b times 7.

由以下语言元素组成:

[设置]符号 可变[a] 符号[到] 可变[b] 符号[时报] 康斯坦[7] 符号[。]

语言元素被定义为“#define”值，如下:

#define ELEM_SYM_SET        0
#define ELEM_SYM_TO         1
#define ELEM_SYM_TIMES      2
#define ELEM_SYM_FULLSTOP   3
#define ELEM_VARIABLE     100
#define ELEM_CONSTANT     101

下面的结构被用来存储每个元素:

typedef struct {
    int typ;
    union {
        char *str;
        int   val;
    }
} tElem;

然后，每个元素的大小是最大联合的大小(typ为4字节，联合为4字节，尽管这些是典型值，但实际大小取决于实现)。

为了创建一个“set”元素，你可以使用:

tElem e;
e.typ = ELEM_SYM_SET;

为了创建一个“variable[b]”元素，你可以使用:

tElem e;
e.typ = ELEM_VARIABLE;
e.str = strdup ("b");   // make sure you free this later

为了创建一个常量[7]元素，你可以使用:

tElem e;
e.typ = ELEM_CONSTANT;
e.val = 7;

你可以很容易地将其扩展为包含浮点数(float flt)或有理数(struct ratnl {int num;Int denom;})和其他类型。

基本前提是str和val在内存中不是连续的，它们实际上是重叠的，所以这是一种在同一块内存上获得不同视图的方法，如图所示，其中结构基于内存位置0x1010，整数和指针都是4字节:

       +-----------+
0x1010 |           |
0x1011 |    typ    |
0x1012 |           |
0x1013 |           |
       +-----+-----+
0x1014 |     |     |
0x1015 | str | val |
0x1016 |     |     |
0x1017 |     |     |
       +-----+-----+

如果只是在一个结构中，它看起来会是这样的:

       +-------+
0x1010 |       |
0x1011 |  typ  |
0x1012 |       |
0x1013 |       |
       +-------+
0x1014 |       |
0x1015 |  str  |
0x1016 |       |
0x1017 |       |
       +-------+
0x1018 |       |
0x1019 |  val  |
0x101A |       |
0x101B |       |
       +-------+

许多答案都涉及从一种类型转换到另一种类型。我从具有相同类型的联合中得到最多的使用(即在解析串行数据流时)。它们允许解析/构造一个有框架的包变得很简单。

typedef union
{
    UINT8 buffer[PACKET_SIZE]; // Where the packet size is large enough for
                               // the entire set of fields (including the payload)

    struct
    {
        UINT8 size;
        UINT8 cmd;
        UINT8 payload[PAYLOAD_SIZE];
        UINT8 crc;
    } fields;

}PACKET_T;

// This should be called every time a new byte of data is ready 
// and point to the packet's buffer:
// packet_builder(packet.buffer, new_data);

void packet_builder(UINT8* buffer, UINT8 data)
{
    static UINT8 received_bytes = 0;

    // All range checking etc removed for brevity

    buffer[received_bytes] = data;
    received_bytes++;

    // Using the struc only way adds lots of logic that relates "byte 0" to size
    // "byte 1" to cmd, etc...
}

void packet_handler(PACKET_T* packet)
{
    // Process the fields in a readable manner
    if(packet->fields.size > TOO_BIG)
    {
        // handle error...
    }

    if(packet->fields.cmd == CMD_X)
    {
        // do stuff..
    }
}

编辑 关于字节序和结构填充的评论是有效的，而且非常值得关注。我几乎完全在嵌入式软件中使用了这段代码，其中大部分我都可以控制管道的两端。