I am making a memory block copy routine and need to deal with blocks of raw memory in efficient chunks. My question is not about the specialized copy routine I'm making, but in how to correctly examine raw pointer alignment in C.
I have a raw pointer of memory, let's say it's already cast as a non-null char *.
In my architecture, I can very efficiently copy memory in 64 byte chunks WHEN IT IS ALIGNED TO A 64 BYTE chunk. So the (standard) trick is that I will do a simple copy of 0-63 bytes "manually" at the head and/or tail to transform the copy from an arbitrary char* of arbitrary length to a 64 byte aligned pointer with some multiple of 64 bytes in length.
Now the question is, how do you legally "examine" a pointer to determine (and manipulate) its alignment?
The obvious way is to cast it into an integer and just examine the bits:
char *pointer=something.
int p=(int)pointer;
char *alignedPointer=(char *)((p+63)&~63);
Note here I realize that alignedPointer doesn't point to the same memory as pointer... this is the "rounded up" pointer that I can call my efficient copy routine on, and I'll handle any other bytes at the beginning manually.
But compilers (justifiably) freak out at casting a pointer into an integer. But how else can I examine and manipulate the pointer's lower bits in LEGAL C? Ideally so that with different compilers I'd get no errors or warnings.
For integer types that are large enough to hold pointers, C99 stdint.h has:
uintptr_t
intptr_t
For data lengths there are:
size_t
ssize_t
which have been around since well before C99.
If your platform doesn't have these, you can maximise your code's portability by still using these type names, and making suitable typedefs for them.
I don't think that in the past people were as reluctant to do their own bit-banging, but maybe the current "don't touch that" mood would be conducive to someone creating some kind of standard library for aligning pointers. Lacking some kind of official api, you have no choice but to AND and OR your way through.
Instead of int, try a datatype that's guaranteed to be the same size as a pointer (INT_PTR on Win32/64). Maybe the compiler won't freak out too much. :) Or use a union, if 64-bit compatibility is not important.
Casting pointers to and from integers is valid, but the results are implementation-defined. See section 6.3.2.3 of the standard. The intention seems to be that the results are what anybody familiar with the system would expect, and indeed this appears to be routinely the case in practice.
If the architecture in question can efficiently manipulate pointers and integers interchangeably, and the issue is just whether it will work on all compilers for that system, then the answer is that it probably will anyway.
(Certainly, if I were writing this code, I would think it fine as-is until proven otherwise. My experience has been that compilers for a given system all behave in very similar ways at this sort of level; the assembly language just suggests a particular approach, that all then take.)
"Probably works" isn't very good general advice though, so my suggestion would be just write the code that works, surround it enough suitable #ifdefs that only the known compiler(s) will compile it, and defer to memcpy in other cases.
#ifdef is rarely ideal, but it's fairly lightweight compared to other possibilities. And if implementation-defined behaviour or compiler-specific tricks are needed then the options are pretty limited anyway.
Related
I have to implement an optimized version of malloc/realloc/free (tailored for my particular application). At the moment the code runs on a particular platform, but I would like to write it in a portable way, if possible (the platform may change in the future), or at least I would like to concentrate the possible platform differences in a single point (probably a .h). I am aware of some of the problems:
differences in memory alignment
differences in smallest memory blocks size suitable for "generic" allocation
differences in pointer size
(I'll ignore the differences in the basic system services for memory allocation here, since on some embedded systems they may be unavailable at all. Let's imagine that we work on a big preallocated memory block to be used as "heap").
The question(s):
Are there standard macros or functions in C for this kind of purpose?
What other issues may I face in this job?
The classic way to ensure that you maintain alignment suitable for all the basic types is to define a union:
union alloc_align {
void *dummy1;
long long dummy2;
long double dummy3;
};
...then ensure that the addresses you hand out are always offset by a multiple of sizeof (union alloc_align) from the aligned addresses you recieve from the system memory allocator.
I believe a method similar to this is described in K&R.
Alignment features are only handled in the new C standard, C11. It has keywords _Alignof, _Alignas and a function aligned_alloc. Theses features are not very difficult to emulate with most modern compilers (as indicated in other answers), so I'd suggest you write yourself small macros or wrappers that you'd use depending on __STDC_VERSION__.
aligned memory differs from compiler to compiler unfortunately (this is one issue), on MSVC, you have aligned_malloc, you also have POSIX memalign for Linux, and then there is also _mm_alloc which works under ICC, MSVC and GCC, IIRC, which should be the most portable.
The second issue is memory wastage from aligning it, it wouldn't be major, but on embedded systems, its something to take note of.
if you are stack allocating things that require alignment (like SIMD types), you also want to look into __attribute__((__aligned__(x))) and __declspec(align(x)).
in terms of portability of pointer arithmetic, you can use the types from stdint.h/pstdint.h to do it, but the standards may say something about UB when casting between uintptr_t and a pointer (unfortunately standards aren't my strong point :().
The main problem is that you only provide the total size of the memory block to malloc() and friends, without any information about the object granularity. If you view an allocation as an array of objects, then you have a size that is the sizeof of the basic object, and a number n that is the number of objects in the array, e.g.:
p = malloc(sizeof(*p) * n);
If you have only the total size, then you don't know if s=4 and n=10, or if s=2 and n=20, or s=1 and n=40, because all multiply to the total size of 40 bytes.
So the basic question is, do you want a direct substitute for the original functions, e.g. when you have thrown native calls all over your code base, or do you have a centralized and DRY modularity with wrapper functions. There you could use functions that provide s and n.
void *my_malloc (size_t s, size_t n)
Most of the time it should be a safe bet when the returned absolute memory address is a multiple of s to guarantee correct alignment.
Alternatively, when porting your implementation, you simply look at the alignment that the native malloc() uses for the target platform (e.g. multiples of 16), and use this for your own implementation.
If you have a look at #pragma pack, this may help you as it allows you to define structure packing and is implemented on most compilers.
C says malloc returns a pointer to memory aligned for any purpose. There is no portable way in C to achieve that with C features. This has the consequence that malloc is a function that if written in C cannot be written in a portable way.
(C99, 7.20.3p1) "The pointer returned if the allocation succeeds is suitably aligned so that it may be assigned to a pointer to any type of object and then used to access such an object or an array of such objects in the space allocated (until the space is explicitly deallocated)."
What forms of memory address spaces have been used?
Today, a large flat virtual address space is common. Historically, more complicated address spaces have been used, such as a pair of a base address and an offset, a pair of a segment number and an offset, a word address plus some index for a byte or other sub-object, and so on.
From time to time, various answers and comments assert that C (or C++) pointers are essentially integers. That is an incorrect model for C (or C++), since the variety of address spaces is undoubtedly the cause of some of the C (or C++) rules about pointer operations. For example, not defining pointer arithmetic beyond an array simplifies support for pointers in a base and offset model. Limits on pointer conversion simplify support for address-plus-extra-data models.
That recurring assertion motivates this question. I am looking for information about the variety of address spaces to illustrate that a C pointer is not necessarily a simple integer and that the C restrictions on pointer operations are sensible given the wide variety of machines to be supported.
Useful information may include:
Examples of computer architectures with various address spaces and descriptions of those spaces.
Examples of various address spaces still in use in machines currently being manufactured.
References to documentation or explanation, especially URLs.
Elaboration on how address spaces motivate C pointer rules.
This is a broad question, so I am open to suggestions on managing it. I would be happy to see collaborative editing on a single generally inclusive answer. However, that may fail to award reputation as deserved. I suggest up-voting multiple useful contributions.
Just about anything you can imagine has probably been used. The
first major division is between byte addressing (all modern
architectures) and word addressing (pre-IBM 360/PDP-11, but
I think modern Unisys mainframes are still word addressed). In
word addressing, char* and void* would often be bigger than
an int*; even if they were not bigger, the "byte selector"
would be in the high order bits, which were required to be 0, or
would be ignored for anything other than bytes. (On a PDP-10,
for example, if p was a char*, (int)p < (int)(p+1) would
often be false, even though int and char* had the same
size.)
Among byte addressed machines, the major variants are segmented
and non-segmented architectures. Both are still wide spread
today, although in the case of Intel 32bit (a segmented
architecture with 48 bit addresses), some of the more widely
used OSs (Windows and Linux) artificially restrict user
processes to a single segment, simulating a flat addressing.
Although I've no recent experience, I would expect even more
variety in embedded processors. In particular, in the past, it
was frequent for embedded processors to use a Harvard
architecture, where code and data were in independent address
spaces (so that a function pointer and a data pointer, cast to a
large enough integral type, could compare equal).
I would say you are asking the wrong question, except as historical curiosity.
Even if your system happens to use a flat address space -- indeed, even if every system from now until the end of time uses a flat address space -- you still cannot treat pointers as integers.
The C and C++ standards leave all sorts of pointer arithmetic "undefined". That can impact you right now, on any system, because compilers will assume you avoid undefined behavior and optimize accordingly.
For a concrete example, three months ago a very interesting bug turned up in Valgrind:
https://sourceforge.net/p/valgrind/mailman/message/29730736/
(Click "View entire thread", then search for "undefined behavior".)
Basically, Valgrind was using less-than and greater-than on pointers to try to determine if an automatic variable was within a certain range. Because comparisons between pointers in different aggregates is "undefined", Clang simply optimized away all of the comparisons to return a constant true (or false; I forget).
This bug itself spawned an interesting StackOverflow question.
So while the original pointer arithmetic definitions may have catered to real machines, and that might be interesting for its own sake, it is actually irrelevant to programming today. What is relevant today is that you simply cannot assume that pointers behave like integers, period, regardless of the system you happen to be using. "Undefined behavior" does not mean "something funny happens"; it means the compiler can assume you do not engage in it. When you do, you introduce a contradiction into the compiler's reasoning; and from a contradiction, anything follows... It only depends on how smart your compiler is.
And they get smarter all the time.
There are various forms of bank-switched memory.
I worked on an embedded system that had 128 KB of total memory: 64KB of RAM and 64KB of EPROM. Pointers were only 16-bit, so a pointer into the RAM could have the same value of a pointer in the EPROM, even though they referred to different memory locations.
The compiler kept track of the type of the pointer so that it could generate the instruction(s) to select the correct bank before dereferencing a pointer.
You could argue that this was like segment + offset, and at the hardware level, it essentially was. But the segment (or more correctly, the bank) was implicit from the pointer's type and not stored as the value of a pointer. If you inspected a pointer in the debugger, you'd just see a 16-bit value. To know whether it was an offset into the RAM or the ROM, you had to know the type.
For example, Foo * could only be in RAM and const Bar * could only be in ROM. If you had to copy a Bar into RAM, the copy would actually be a different type. (It wasn't as simple as const/non-const:
Everything in ROM was const, but not all consts were in ROM.)
This was all in C, and I know we used non-standard extensions to make this work. I suspect a 100% compliant C compiler probably couldn't cope with this.
From a C programmer's perspective, there are three main kinds of implementation to worry about:
Those which target machines with a linear memory model, and which are designed and/or configured to be usable as a "high-level assembler"--something the authors of the Standard have expressly said they did not wish to preclude. Most implementations behave in this way when optimizations are disabled.
Those which are usable as "high-level assemblers" for machines with unusual memory architectures.
Those which whose design and/or configuration make them suitable only for tasks that do not involve low-level programming, including clang and gcc when optimizations are enabled.
Memory-management code targeting the first type of implementation will often be compatible with all implementations of that type whose targets use the same representations for pointers and integers. Memory-management code for the second type of implementation will often need to be specifically tailored for the particular hardware architecture. Platforms that don't use linear addressing are sufficiently rare, and sufficiently varied, that unless one needs to write or maintain code for some particular piece of unusual hardware (e.g. because it drives an expensive piece of industrial equipment for which more modern controllers aren't available) knowledge of any particular architecture isn't likely to be of much use.
Implementations of the third type should be used only for programs that don't need to do any memory-management or systems-programming tasks. Because the Standard doesn't require that all implementations be capable of supporting such tasks, some compiler writers--even when targeting linear-address machines--make no attempt to support any of the useful semantics thereof. Even some principles like "an equality comparison between two valid pointers will--at worst--either yield 0 or 1 chosen in possibly-unspecified fashion don't apply to such implementations.
I have this StructType st = StructTypeSecondInstance->st; and it generates a segfault. The strange part is when the stack backtrace shows me:
0x1067d2cc: memcpy + 0x10 (0, 10000, 1, 1097a69c, 11db0720, bfe821c0) + 310
0x103cfddc: some_function + 0x60 (0, bfe823d8, bfe82418, 10b09b10, 0, 0) +
So, does struct assigment use memcpy?
One can't tell. Small structs may even be kept in registers. Whether memcpy is used is an implementation detail (it's not even implementation-defined, or unspecified -- it's just something the compiler writer choses and does not need to document.)
From a C Standard point of view, all that matters is that after the assigment, the struct members of the destination struct compare equal to the corresponding members of the source struct.
I would expect compiler writers to make a tradeoff between speed and simplicity, probably based on the size of the struct, the larger the more likely to use a memcpy. Some memcpy implementations are very sophisticated and use different algorithms depending on whether the length is some power of 2 or not, or the alignment of the src and dst pointers. Why reinvent the wheel or blow up the code with an inline version of memcpy?
It might, yes.
This shouldn't be surprising: the struct assignment needs to copy a bunch of bytes from one place to another as quickly as possible, which happens to be the exact thing memcpy() is supposed to be good at. Generating a call to it seems like a no-brainer if you're a compiler writer.
Note that this means that assigning structs with lots of padding might be less efficient than optimally, since memcpy() can't skip the padding.
The standard doesn't say anything at all about how assignment (or any other operator) is actually realized by the compiler. There's nothing stopping a compiler from (say) generating a function call for every operation in your source file.
The compiler has license to implement assignment as it thinks best. Most of the time, with most compilers on most platforms, this means that if the structure is reasonably small, the compiler will generate an inline sequence of move instructions; if the structure is large, calling memcpy is common.
It would be perfectly valid, however, for the compiler to loop over generating random bitfields and stop when one of them matches the source of the assignment (Let's call this algorithm bogocopy).
Compilers that support non-hosted operation usually give you a switch to turn off emitting such libcalls if you're targeting a platform without an available (or complete) libc.
It depends on the compiler and platform. Assignment of big objects can use memcpy. But it must not be the reason of segfault.
I have to implement an optimized version of malloc/realloc/free (tailored for my particular application). At the moment the code runs on a particular platform, but I would like to write it in a portable way, if possible (the platform may change in the future), or at least I would like to concentrate the possible platform differences in a single point (probably a .h). I am aware of some of the problems:
differences in memory alignment
differences in smallest memory blocks size suitable for "generic" allocation
differences in pointer size
(I'll ignore the differences in the basic system services for memory allocation here, since on some embedded systems they may be unavailable at all. Let's imagine that we work on a big preallocated memory block to be used as "heap").
The question(s):
Are there standard macros or functions in C for this kind of purpose?
What other issues may I face in this job?
The classic way to ensure that you maintain alignment suitable for all the basic types is to define a union:
union alloc_align {
void *dummy1;
long long dummy2;
long double dummy3;
};
...then ensure that the addresses you hand out are always offset by a multiple of sizeof (union alloc_align) from the aligned addresses you recieve from the system memory allocator.
I believe a method similar to this is described in K&R.
Alignment features are only handled in the new C standard, C11. It has keywords _Alignof, _Alignas and a function aligned_alloc. Theses features are not very difficult to emulate with most modern compilers (as indicated in other answers), so I'd suggest you write yourself small macros or wrappers that you'd use depending on __STDC_VERSION__.
aligned memory differs from compiler to compiler unfortunately (this is one issue), on MSVC, you have aligned_malloc, you also have POSIX memalign for Linux, and then there is also _mm_alloc which works under ICC, MSVC and GCC, IIRC, which should be the most portable.
The second issue is memory wastage from aligning it, it wouldn't be major, but on embedded systems, its something to take note of.
if you are stack allocating things that require alignment (like SIMD types), you also want to look into __attribute__((__aligned__(x))) and __declspec(align(x)).
in terms of portability of pointer arithmetic, you can use the types from stdint.h/pstdint.h to do it, but the standards may say something about UB when casting between uintptr_t and a pointer (unfortunately standards aren't my strong point :().
The main problem is that you only provide the total size of the memory block to malloc() and friends, without any information about the object granularity. If you view an allocation as an array of objects, then you have a size that is the sizeof of the basic object, and a number n that is the number of objects in the array, e.g.:
p = malloc(sizeof(*p) * n);
If you have only the total size, then you don't know if s=4 and n=10, or if s=2 and n=20, or s=1 and n=40, because all multiply to the total size of 40 bytes.
So the basic question is, do you want a direct substitute for the original functions, e.g. when you have thrown native calls all over your code base, or do you have a centralized and DRY modularity with wrapper functions. There you could use functions that provide s and n.
void *my_malloc (size_t s, size_t n)
Most of the time it should be a safe bet when the returned absolute memory address is a multiple of s to guarantee correct alignment.
Alternatively, when porting your implementation, you simply look at the alignment that the native malloc() uses for the target platform (e.g. multiples of 16), and use this for your own implementation.
If you have a look at #pragma pack, this may help you as it allows you to define structure packing and is implemented on most compilers.
C says malloc returns a pointer to memory aligned for any purpose. There is no portable way in C to achieve that with C features. This has the consequence that malloc is a function that if written in C cannot be written in a portable way.
(C99, 7.20.3p1) "The pointer returned if the allocation succeeds is suitably aligned so that it may be assigned to a pointer to any type of object and then used to access such an object or an array of such objects in the space allocated (until the space is explicitly deallocated)."
In embedded software domain for copying structure of same type people don't use direct assignment and do that by memcpy() function or each element copying.
lets have for example
struct tag
{
int a;
int b;
};
struct tag exmple1 = {10,20};
struct tag exmple2;
for copying exmple1 into exmple2..
instead of writing direct
exmple2=exmple1;
people use
memcpy(exmple2,exmple1,sizeof(struct tag));
or
exmple2.a=exmple1.a;
exmple2.b=exmple1.b;
why ????
One way or the other there is nothing specific about embedded systems that makes this dangerous, the language semantics are identical for all platforms.
C has been used in embedded systems for many years, and early C compilers, before ANSI/ISO standardisation did not support direct structure assignment. Many practitioners are either from that era, or have been taught by those that were, or are using legacy code written by such practitioners. This is probably the root of the doubt, but it is not a problem on an ISO compliant implementation. On some very resource constrained targets, the available compiler may not be fully ISO compliant for a number of reasons, but I doubt that this feature would be affected.
One issue (that applies to embedded and non-embedded alike), is that when assigning a structure, an implementation need not duplicate the value of any undefined padding bits, therefore if you performed a structure assignment, and then performed a memcmp() rather than member-by-member comparison to test for equality, there is no guarantee that they will be equal. However if you perform a memcpy(), any padding bits will be copied so that memcmp() and member-by-member comparison will yield equality.
So it is arguably safer to use memcpy() in all cases (not just embedded), but the improvement is marginal, and not conducive to readability. It would be a strange implementation that did not use the simplest method of structure assignment, and that is a simple memcpy(), so it is unlikely that the theoretical mismatch would occur.
In your given code there is no problem even if you write:
example2 = example1;
But just assume if in future, the struct definition changes to:
struct tag
{
int a[1000];
int b;
};
Now if you execute the assignment operator as above then (some of the) compiler might inline the code for byte by byte (or int by int) copying. i.e.
example1.a[0] = example.a[0];
example1.a[1] = example.a[1];
example1.a[2] = example.a[2];
...
which will result in code bloat in your code segment. Such kind of memory errors are not trivial to find. That's why people use memcpy.
[However, I have heard that modern compilers are capable enough to use memcpy internally when such instruction is encountered especially for PODs.]
Copying C-structures via memcpy() is often used by programmers who learned C decades ago and did not follow the standardization process since. They simple don't know that C supports assignment of structures (direct structure assignment was not available in all pre-ANSI-C89 compilers).
When they learn about this feature some still stick to the memcpy() way because it is their custom. There are also motivations that originate in cargo cult programming, e.g. it is claimed that memcpy is just faster - of course - without being able to back this up with a benchmark test case.
Structures are also memcpy()ied by some newbie programmers because they either confuse structure assignment with the assignment of a pointer of a structure - or they simply overuse memcpy() (they often also use memcpy() where strcpy() would be more appropriate).
There is also the memcmp() structure comparison anti-pattern that is sometimes cited by some programmers for using memcpy() instead of structure assignment. The reasoning behind this is the following: since C does not automatically generate a == operator for structures and writing a custom structure comparison function is tedious, memcmp() is used to compare structures. In the next step - to avoid differences in the padding bits of compared structures - memset(...,0,...) is used to initialize all structures (instead of using the C99 initializer syntax or initializing all fields separately) and memcpy() is used to copy the structures! Because memcpy() also copies the content of the padding bits ...
But note that this reasoning is flawed for several reasons:
the use of memcpy()/memcmp()/memset() introduce new error possibilities - e.g. supplying a wrong size
when the structure contains integer fields the ordering under memcmp() changes between big- and little-endian architectures
a char array field of size n that is 0-terminated at position x must also have all elements after position x zeroed out at any time - else 2 otherwise equal structs compare unequal
assignment from a register to a field may also set the neighbouring padding bits to values unequal 0, thus, following comparisons with otherwise equal structures yield an unequal result
The last point is best illustrated with a small example (assuming architecture X):
struct S {
int a; // on X: sizeof(int) == 4
char b; // on X: 24 padding bits are inserted after b
int c;
};
typedef struct S S;
S s1;
memset(&s1, 0, sizeof(S));
s1.a = 0;
s1.b = 'a';
s1.c = 0;
S s2;
memcpy(&s2, &s1, sizeof(S));
assert(memcmp(&s1, &s2, sizeof(S)==0); // assertion is always true
s2.b = 'x';
assert(memcmp(&s1, &s2, sizeof(S)!=0); // assertion is always true
// some computation
char x = 'x'; // on X: 'x' is stored in a 32 bit register
// as least significant byte
// the other bytes contain previous data
s1.b = x; // the complete register is copied
// i.e. the higher 3 register bytes are the new
// padding bits in s1
assert(memcmp(&s1, &s2, sizeof(S)==0); // assertion is not always true
The failure of the last assertion may depend on code reordering, change of the compiler, change of compiler options and stuff like that.
Conclusion
As a general rule: to increase code correctness and portability use direct struct assignment (instead of memcpy()), C99 struct initialization syntax (instead of memset) and a custom comparison function (instead of memcmp()).
In C people probably do that, because they think that memcpy would be faster. But I don't think that is true. Compiler optimizations would take care of that.
In C++ it may also have different semantics because of user defined assignment operator and copy constructors.
On top of what the others wrote some additional points:
Using memcpy instead of a simple assignment gives a hint to someone who maintains the code that the operation might be expensive. Using memcpy in these cases will improves the understanding of the code.
Embedded systems are often written with portability and performance in mind. Portability is important because you may want to re-use your code even if the CPU in the original design is not available or if a cheaper micro-controller can do the same job.
These days low-end micro-controllers come and go faster than the compiler developers can catch up, so it is not uncommon to work with compilers that use a simple byte-copy loop instead of something optimized for structure assignments. With the move to 32 bit ARM cores this is not true for a large part of embedded developers. There are however a lot of people out there who build products that target obscure 8 and 16 bit micro-controllers.
A memcpy tuned for a specific platform may be more optimal than what a compiler can generate. For example on embedded platforms having structures in flash memory is common. Reading from flash is not as slow as writing to it, but it is still a lot slower than a ordinary copy from RAM to RAM. A optimized memcpy function may use DMA or special features from the flash controller to speed up the copy process.
That is a complete nonsense. Use whichever way you prefer. The simplest is :
exmple2=exmple1;
Whatever you do, don't do this:
exmple2.a=exmple1.a;
exmple2.b=exmple1.b;
It poses a maintainability problem because any time that anyone adds a member to the structure, they have to add a line of code to do the copy of that member. Someone is going to forget to do that and it will cause a hard to find bug.
On some implementations, the way in which memcpy() is performed may differ from the way in which "normal" structure assignment would be performed, in a manner that may be important in some narrow contexts. For example, one or the other structure operand may be unaligned and the compiler might not know about it (e.g. one memory region might have external linkage and be defined in a module written in a different language that has no means of enforcing alignment). Use of a __packed declaration would be better if a compiler supported such, but not all compilers do.
Another reason for using something other than structure assignment could be that a particular implementation's memcpy might access its operands in a sequence that would work correctly with certain kinds of volatile source or destination, while that implementation's struct assignment might use a different sequence that wouldn't work. This is generally not a good reason to use memcpy, however, since aside from the alignment issue (which memcpy is required to handle correctly in any case) the specifications for memcpy don't promise much about how the operation will be performed. It would be better to use a specially-written routine which performed the operations exactly as required (for example, if the target is a piece of hardware which needs to have 4 bytes of structure data written using four 8-bit writes rather than one 32-bit writes, one should write a routine which does that, rather than hoping that no future version of memcpy decides to "optimize" the operation).
A third reason for using memcpy in some cases would be the fact that compilers will often perform small structure assignments using a direct sequence of loads and stores, rather than using a library routine. On some controllers, the amount of code this requires may vary depending upon where the structures are located in memory, to the point that the load/store sequence may end up being bigger than a memcpy call. For example, on a PICmicro controller with 1Kwords of code space and 192 bytes of RAM, coping a 4-byte structure from bank 1 to bank 0 would take 16 instructions. A memcpy call would take eight or nine (depending upon whether count is an unsigned char or int [with only 192 bytes of RAM total, unsigned char should be more than sufficient!] Note, however, that calling a memcpy-ish routine which assumed a hard-coded size and required both operands be in RAM rather than code space would only require five instructions to call, and that could be reduced to four with the use of a global variable.
first version is perfect.
second one may be used for speed (there is no reason for your size).
3rd one is used only if padding is different for target and source.