When to use restrict and when not to - c

I have a general understanding of restrict but I'm hoping to clarify some fine points. I have a function that reads a null-terminated string from one buffer and writes out a URL encoded version in another buffer. The function has this signature (currently without restrict):
char const *StringUrlEncode(char const *unencoded,
char *encoded,
char *encodedEnd);
unencoded is my null-terminated source string. The destination buffer is represented by encoded and encodedEnd, where encoded points to the first char in the buffer and encodedEnd points to the first char after the buffer, i.e. the function will write chars up to but not including the location pointed to by encodedEnd -- this is your basic begin/end iterator pair if you're familiar with C++ STL conventions.
If I add restrict to this function, should it only be applied to the first two parameters:
char const *StringUrlEncode(char const *restrict unencoded,
char *restrict encoded,
char *encodedEnd);
or is there some benefit I'm not understanding by adding it to all three parameters?
I can see that making the input and output buffers restrict helps the compiler know that they don't overlap. But since the last parameter, encodedEnd, is only used to mark the end of the output buffer, I'm thinking that restrict wouldn't really be any help to the compiler here (though I assume it wouldn't hurt, other than adding unnecessary noise to the function declaration).

Try Mike Acton's article here (old link). Restrict is frightening because of both the performance implications of not using it and the consequences of using it incorrectly.
In your case, it sounds like you could safely apply restrict to all three pointers as none alias the same memory area. However there is going to be little to no performance benefit from using it on the third pointer.

In this particular case it won't make a difference whether encodedEnd is restrict or not; you have promised the compiler that no one aliases unencoded and encoded, and so the reads and writes won't interfere with each other.
The real reason that restrict is important in this case is that without it the compiler can't know that writes through encoded won't affect reads through unencoded. For example, if
encoded == unencoded+1
then each write to encoded would affect each subsequent read from unencoded, so the compiler can't schedule the load until the write has completed. restrict promises the compiler that the two pointers don't affect the same memory, so it can schedule loads far enough ahead to avoid pipeline stalls.

I think you're right that it wouldn't hurt. Your loop pointer (call it p) will equal encodedEnd at the end of the loop. But nothing need be accessed after the loop (from either p or encodedEnd), so that shouldn't be a problem. I don't think it will help, either, because nothing is ever written or read from encodedEnd so there's nothing to optimize away.
But I agree with you having the first two restrict should really help.

Related

String array in C prints whole array instead of value at pointer [duplicate]

As much as I love C and C++, I can't help but scratch my head at the choice of null terminated strings:
Length prefixed (i.e. Pascal) strings existed before C
Length prefixed strings make several algorithms faster by allowing constant time length lookup.
Length prefixed strings make it more difficult to cause buffer overrun errors.
Even on a 32 bit machine, if you allow the string to be the size of available memory, a length prefixed string is only three bytes wider than a null terminated string. On 16 bit machines this is a single byte. On 64 bit machines, 4GB is a reasonable string length limit, but even if you want to expand it to the size of the machine word, 64 bit machines usually have ample memory making the extra seven bytes sort of a null argument. I know the original C standard was written for insanely poor machines (in terms of memory), but the efficiency argument doesn't sell me here.
Pretty much every other language (i.e. Perl, Pascal, Python, Java, C#, etc) use length prefixed strings. These languages usually beat C in string manipulation benchmarks because they are more efficient with strings.
C++ rectified this a bit with the std::basic_string template, but plain character arrays expecting null terminated strings are still pervasive. This is also imperfect because it requires heap allocation.
Null terminated strings have to reserve a character (namely, null), which cannot exist in the string, while length prefixed strings can contain embedded nulls.
Several of these things have come to light more recently than C, so it would make sense for C to not have known of them. However, several were plain well before C came to be. Why would null terminated strings have been chosen instead of the obviously superior length prefixing?
EDIT: Since some asked for facts (and didn't like the ones I already provided) on my efficiency point above, they stem from a few things:
Concat using null terminated strings requires O(n + m) time complexity. Length prefixing often require only O(m).
Length using null terminated strings requires O(n) time complexity. Length prefixing is O(1).
Length and concat are by far the most common string operations. There are several cases where null terminated strings can be more efficient, but these occur much less often.
From answers below, these are some cases where null terminated strings are more efficient:
When you need to cut off the start of a string and need to pass it to some method. You can't really do this in constant time with length prefixing even if you are allowed to destroy the original string, because the length prefix probably needs to follow alignment rules.
In some cases where you're just looping through the string character by character you might be able to save a CPU register. Note that this works only in the case that you haven't dynamically allocated the string (Because then you'd have to free it, necessitating using that CPU register you saved to hold the pointer you originally got from malloc and friends).
None of the above are nearly as common as length and concat.
There's one more asserted in the answers below:
You need to cut off the end of the string
but this one is incorrect -- it's the same amount of time for null terminated and length prefixed strings. (Null terminated strings just stick a null where you want the new end to be, length prefixers just subtract from the prefix.)
From the horse's mouth
None of BCPL, B, or C supports
character data strongly in the
language; each treats strings much
like vectors of integers and
supplements general rules by a few
conventions. In both BCPL and B a
string literal denotes the address of
a static area initialized with the
characters of the string, packed into
cells. In BCPL, the first packed byte
contains the number of characters in
the string; in B, there is no count
and strings are terminated by a
special character, which B spelled
*e. This change was made partially
to avoid the limitation on the length
of a string caused by holding the
count in an 8- or 9-bit slot, and
partly because maintaining the count
seemed, in our experience, less
convenient than using a terminator.
Dennis M Ritchie, Development of the C Language
C doesn't have a string as part of the language. A 'string' in C is just a pointer to char. So maybe you're asking the wrong question.
"What's the rationale for leaving out a string type" might be more relevant. To that I would point out that C is not an object oriented language and only has basic value types. A string is a higher level concept that has to be implemented by in some way combining values of other types. C is at a lower level of abstraction.
in light of the raging squall below:
I just want to point out that I'm not trying to say this is a stupid or bad question, or that the C way of representing strings is the best choice. I'm trying to clarify that the question would be more succinctly put if you take into account the fact that C has no mechanism for differentiating a string as a datatype from a byte array. Is this the best choice in light of the processing and memory power of todays computers? Probably not. But hindsight is always 20/20 and all that :)
The question is asked as a Length Prefixed Strings (LPS) vs zero terminated strings (SZ) thing, but mostly expose benefits of length prefixed strings. That may seem overwhelming, but to be honest we should also consider drawbacks of LPS and advantages of SZ.
As I understand it, the question may even be understood as a biased way to ask "what are the advantages of Zero Terminated Strings ?".
Advantages (I see) of Zero Terminated Strings:
very simple, no need to introduce new concepts in language, char
arrays/char pointers can do.
the core language just include minimal syntaxic sugar to convert
something between double quotes to a
bunch of chars (really a bunch of
bytes). In some cases it can be used
to initialize things completely
unrelated with text. For instance xpm
image file format is a valid C source
that contains image data encoded as a
string.
by the way, you can put a zero in a string literal, the compiler will
just also add another one at the end of the literal: "this\0is\0valid\0C".
Is it a string ? or four strings ? Or a bunch of bytes...
flat implementation, no hidden indirection, no hidden integer.
no hidden memory allocation involved (well, some infamous non
standard functions like strdup
perform allocation, but that's mostly
a source of problem).
no specific issue for small or large hardware (imagine the burden to
manage 32 bits prefix length on 8
bits microcontrollers, or the
restrictions of limiting string size
to less than 256 bytes, that was a problem I actually had with Turbo Pascal eons ago).
implementation of string manipulation is just a handful of
very simple library function
efficient for the main use of strings : constant text read
sequentially from a known start
(mostly messages to the user).
the terminating zero is not even mandatory, all necessary tools
to manipulate chars like a bunch of
bytes are available. When performing
array initialisation in C, you can
even avoid the NUL terminator. Just
set the right size. char a[3] =
"foo"; is valid C (not C++) and
won't put a final zero in a.
coherent with the unix point of view "everything is file", including
"files" that have no intrinsic length
like stdin, stdout. You should remember that open read and write primitives are implemented
at a very low level. They are not library calls, but system calls. And the same API is used
for binary or text files. File reading primitives get a buffer address and a size and return
the new size. And you can use strings as the buffer to write. Using another kind of string
representation would imply you can't easily use a literal string as the buffer to output, or
you would have to make it have a very strange behavior when casting it to char*. Namely
not to return the address of the string, but instead to return the actual data.
very easy to manipulate text data read from a file in-place, without useless copy of buffer,
just insert zeroes at the right places (well, not really with modern C as double quoted strings are const char arrays nowaday usually kept in non modifiable data segment).
prepending some int values of whatever size would implies alignment issues. The initial
length should be aligned, but there is no reason to do that for the characters datas (and
again, forcing alignment of strings would imply problems when treating them as a bunch of
bytes).
length is known at compile time for constant literal strings (sizeof). So why would
anyone want to store it in memory prepending it to actual data ?
in a way C is doing as (nearly) everyone else, strings are viewed as arrays of char. As array length is not managed by C, it is logical length is not managed either for strings. The only surprising thing is that 0 item added at the end, but that's just at core language level when typing a string between double quotes. Users can perfectly call string manipulation functions passing length, or even use plain memcopy instead. SZ are just a facility. In most other languages array length is managed, it's logical that is the same for strings.
in modern times anyway 1 byte character sets are not enough and you often have to deal with encoded unicode strings where the number of characters is very different of the number of bytes. It implies that users will probably want more than "just the size", but also other informations. Keeping length give use nothing (particularly no natural place to store them) regarding these other useful pieces of information.
That said, no need to complain in the rare case where standard C strings are indeed inefficient. Libs are available. If I followed that trend, I should complain that standard C does not include any regex support functions... but really everybody knows it's not a real problem as there is libraries available for that purpose. So when string manipulation efficiency is wanted, why not use a library like bstring ? Or even C++ strings ?
EDIT: I recently had a look to D strings. It is interesting enough to see that the solution choosed is neither a size prefix, nor zero termination. As in C, literal strings enclosed in double quotes are just short hand for immutable char arrays, and the language also has a string keyword meaning that (immutable char array).
But D arrays are much richer than C arrays. In the case of static arrays length is known at run-time so there is no need to store the length. Compiler has it at compile time. In the case of dynamic arrays, length is available but D documentation does not state where it is kept. For all we know, compiler could choose to keep it in some register, or in some variable stored far away from the characters data.
On normal char arrays or non literal strings there is no final zero, hence programmer has to put it itself if he wants to call some C function from D. In the particular case of literal strings, however the D compiler still put a zero at the end of each strings (to allow easy cast to C strings to make easier calling C function ?), but this zero is not part of the string (D does not count it in string size).
The only thing that disappointed me somewhat is that strings are supposed to be utf-8, but length apparently still returns a number of bytes (at least it's true on my compiler gdc) even when using multi-byte chars. It is unclear to me if it's a compiler bug or by purpose. (OK, I probably have found out what happened. To say to D compiler your source use utf-8 you have to put some stupid byte order mark at beginning. I write stupid because I know of not editor doing that, especially for UTF-8 that is supposed to be ASCII compatible).
I think, it has historical reasons and found this in wikipedia:
At the time C (and the languages that
it was derived from) were developed,
memory was extremely limited, so using
only one byte of overhead to store the
length of a string was attractive. The
only popular alternative at that time,
usually called a "Pascal string"
(though also used by early versions of
BASIC), used a leading byte to store
the length of the string. This allows
the string to contain NUL and made
finding the length need only one
memory access (O(1) (constant) time).
But one byte limits the length to 255.
This length limitation was far more
restrictive than the problems with the
C string, so the C string in general
won out.
Calavera is right, but as people don't seem to get his point, I'll provide some code examples.
First, let's consider what C is: a simple language, where all code has a pretty direct translation into machine language. All types fit into registers and on the stack, and it doesn't require an operating system or a big run-time library to run, since it were meant to write these things (a task to which is superbly well-suited, considering there isn't even a likely competitor to this day).
If C had a string type, like int or char, it would be a type which didn't fit in a register or in the stack, and would require memory allocation (with all its supporting infrastructure) to be handled in any way. All of which go against the basic tenets of C.
So, a string in C is:
char s*;
So, let's assume then that this were length-prefixed. Let's write the code to concatenate two strings:
char* concat(char* s1, char* s2)
{
/* What? What is the type of the length of the string? */
int l1 = *(int*) s1;
/* How much? How much must I skip? */
char *s1s = s1 + sizeof(int);
int l2 = *(int*) s2;
char *s2s = s2 + sizeof(int);
int l3 = l1 + l2;
char *s3 = (char*) malloc(l3 + sizeof(int));
char *s3s = s3 + sizeof(int);
memcpy(s3s, s1s, l1);
memcpy(s3s + l1, s2s, l2);
*(int*) s3 = l3;
return s3;
}
Another alternative would be using a struct to define a string:
struct {
int len; /* cannot be left implementation-defined */
char* buf;
}
At this point, all string manipulation would require two allocations to be made, which, in practice, means you'd go through a library to do any handling of it.
The funny thing is... structs like that do exist in C! They are just not used for your day-to-day displaying messages to the user handling.
So, here is the point Calavera is making: there is no string type in C. To do anything with it, you'd have to take a pointer and decode it as a pointer to two different types, and then it becomes very relevant what is the size of a string, and cannot just be left as "implementation defined".
Now, C can handle memory in anyway, and the mem functions in the library (in <string.h>, even!) provide all the tooling you need to handle memory as a pair of pointer and size. The so-called "strings" in C were created for just one purpose: showing messages in the context of writting an operating system intended for text terminals. And, for that, null termination is enough.
Obviously for performance and safety, you'll want to keep the length of a string while you're working with it rather than repeatedly performing strlen or the equivalent on it. However, storing the length in a fixed location just before the string contents is an incredibly bad design. As Jörgen pointed out in the comments on Sanjit's answer, it precludes treating the tail of a string as a string, which for example makes a lot of common operations like path_to_filename or filename_to_extension impossible without allocating new memory (and incurring the possibility of failure and error handling). And then of course there's the issue that nobody can agree how many bytes the string length field should occupy (plenty of bad "Pascal string" languages used 16-bit fields or even 24-bit fields which preclude processing of long strings).
C's design of letting the programmer choose if/where/how to store the length is much more flexible and powerful. But of course the programmer has to be smart. C punishes stupidity with programs that crash, grind to a halt, or give your enemies root.
Lazyness, register frugality and portability considering the assembly gut of any language, especially C which is one step above assembly (thus inheriting a lot of assembly legacy code).
You would agree as a null char would be useless in those ASCII days, it (and probably as good as an EOF control char ).
let's see in pseudo code
function readString(string) // 1 parameter: 1 register or 1 stact entries
pointer=addressOf(string)
while(string[pointer]!=CONTROL_CHAR) do
read(string[pointer])
increment pointer
total 1 register use
case 2
function readString(length,string) // 2 parameters: 2 register used or 2 stack entries
pointer=addressOf(string)
while(length>0) do
read(string[pointer])
increment pointer
decrement length
total 2 register used
That might seem shortsighted at that time, but considering the frugality in code and register ( which were PREMIUM at that time, the time when you know, they use punch card ). Thus being faster ( when processor speed could be counted in kHz), this "Hack" was pretty darn good and portable to register-less processor with ease.
For argument sake I will implement 2 common string operation
stringLength(string)
pointer=addressOf(string)
while(string[pointer]!=CONTROL_CHAR) do
increment pointer
return pointer-addressOf(string)
complexity O(n) where in most case PASCAL string is O(1) because the length of the string is pre-pended to the string structure (that would also mean that this operation would have to be carried in an earlier stage).
concatString(string1,string2)
length1=stringLength(string1)
length2=stringLength(string2)
string3=allocate(string1+string2)
pointer1=addressOf(string1)
pointer3=addressOf(string3)
while(string1[pointer1]!=CONTROL_CHAR) do
string3[pointer3]=string1[pointer1]
increment pointer3
increment pointer1
pointer2=addressOf(string2)
while(string2[pointer2]!=CONTROL_CHAR) do
string3[pointer3]=string2[pointer2]
increment pointer3
increment pointer1
return string3
complexity O(n) and prepending the string length wouldn't change the complexity of the operation, while I admit it would take 3 time less time.
On another hand, if you use PASCAL string you would have to redesign your API for taking in account register length and bit-endianness, PASCAL string got the well known limitation of 255 char (0xFF) beacause the length was stored in 1 byte (8bits), and it you wanted a longer string (16bits->anything) you would have to take in account the architecture in one layer of your code, that would mean in most case incompatible string APIs if you wanted longer string.
Example:
One file was written with your prepended string api on an 8 bit computer and then would have to be read on say a 32 bit computer, what would the lazy program do considers that your 4bytes are the length of the string then allocate that lot of memory then attempt to read that many bytes.
Another case would be PPC 32 byte string read(little endian) onto a x86 (big endian), of course if you don't know that one is written by the other there would be trouble.
1 byte length (0x00000001) would become 16777216 (0x0100000) that is 16 MB for reading a 1 byte string.
Of course you would say that people should agree on one standard but even 16bit unicode got little and big endianness.
Of course C would have its issues too but, would be very little affected by the issues raised here.
In many ways, C was primitive. And I loved it.
It was a step above assembly language, giving you nearly the same performance with a language that was much easier to write and maintain.
The null terminator is simple and requires no special support by the language.
Looking back, it doesn't seem that convenient. But I used assembly language back in the 80s and it seemed very convenient at the time. I just think software is continually evolving, and the platforms and tools continually get more and more sophisticated.
Assuming for a moment that C implemented strings the Pascal way, by prefixing them by length: is a 7 char long string the same DATA TYPE as a 3-char string? If the answer is yes, then what kind of code should the compiler generate when I assign the former to the latter? Should the string be truncated, or automatically resized? If resized, should that operation be protected by a lock as to make it thread safe? The C approach side stepped all these issues, like it or not :)
Somehow I understood the question to imply there's no compiler support for length-prefixed strings in C. The following example shows, at least you can start your own C string library, where string lengths are counted at compile time, with a construct like this:
#define PREFIX_STR(s) ((prefix_str_t){ sizeof(s)-1, (s) })
typedef struct { int n; char * p; } prefix_str_t;
int main() {
prefix_str_t string1, string2;
string1 = PREFIX_STR("Hello!");
string2 = PREFIX_STR("Allows \0 chars (even if printf directly doesn't)");
printf("%d %s\n", string1.n, string1.p); /* prints: "6 Hello!" */
printf("%d %s\n", string2.n, string2.p); /* prints: "48 Allows " */
return 0;
}
This won't, however, come with no issues as you need to be careful when to specifically free that string pointer and when it is statically allocated (literal char array).
Edit: As a more direct answer to the question, my view is this was the way C could support both having string length available (as a compile time constant), should you need it, but still with no memory overhead if you want to use only pointers and zero termination.
Of course it seems like working with zero-terminated strings was the recommended practice, since the standard library in general doesn't take string lengths as arguments, and since extracting the length isn't as straightforward code as char * s = "abc", as my example shows.
"Even on a 32 bit machine, if you allow the string to be the size of available memory, a length prefixed string is only three bytes wider than a null terminated string."
First, extra 3 bytes may be considerable overhead for short strings. In particular, a zero-length string now takes 4 times as much memory. Some of us are using 64-bit machines, so we either need 8 bytes to store a zero-length string, or the string format can't cope with the longest strings the platform supports.
There may also be alignment issues to deal with. Suppose I have a block of memory containing 7 strings, like "solo\0second\0\0four\0five\0\0seventh". The second string starts at offset 5. The hardware may require that 32-bit integers be aligned at an address that is a multiple of 4, so you have to add padding, increasing the overhead even further. The C representation is very memory-efficient in comparison. (Memory-efficiency is good; it helps cache performance, for example.)
One point not yet mentioned: when C was designed, there were many machines where a 'char' was not eight bits (even today there are DSP platforms where it isn't). If one decides that strings are to be length-prefixed, how many 'char's worth of length prefix should one use? Using two would impose an artificial limit on string length for machines with 8-bit char and 32-bit addressing space, while wasting space on machines with 16-bit char and 16-bit addressing space.
If one wanted to allow arbitrary-length strings to be stored efficiently, and if 'char' were always 8-bits, one could--for some expense in speed and code size--define a scheme were a string prefixed by an even number N would be N/2 bytes long, a string prefixed by an odd value N and an even value M (reading backward) could be ((N-1) + M*char_max)/2, etc. and require that any buffer which claims to offer a certain amount of space to hold a string must allow enough bytes preceding that space to handle the maximum length. The fact that 'char' isn't always 8 bits, however, would complicate such a scheme, since the number of 'char' required to hold a string's length would vary depending upon the CPU architecture.
The null termination allows for fast pointer based operations.
Not a Rationale necessarily but a counterpoint to length-encoded
Certain forms of dynamic length encoding are superior to static length encoding as far as memory is concerned, it all depends on usage. Just look at UTF-8 for proof. It's essentially an extensible character array for encoding a single character. This uses a single bit for each extended byte. NUL termination uses 8 bits. Length-prefix I think can be reasonably termed infinite length as well by using 64 bits. How often you hit the case of your extra bits is the deciding factor. Only 1 extremely large string? Who cares if you're using 8 or 64 bits? Many small strings (Ie Strings of English words)? Then your prefix costs are a large percentage.
Length-prefixed strings allowing time savings is not a real thing. Whether your supplied data is required to have length provided, you're counting at compile time, or you're truly being provided dynamic data that you must encode as a string. These sizes are computed at some point in the algorithm. A separate variable to store the size of a null terminated string can be provided. Which makes the comparison on time-savings moot. One just has an extra NUL at the end... but if the length encode doesn't include that NUL then there's literally no difference between the two. There's no algorithmic change required at all. Just a pre-pass you have to manually design yourself instead of having a compiler/runtime do it for you. C is mostly about doing things manually.
Length-prefix being optional is a selling point. I don't always need that extra info for an algorithm so being required to do it for a every string makes my precompute+compute time never able to drop below O(n). (Ie hardware random number generator 1-128. I can pull from an "infinite string". Let's say it only generates characters so fast. So our string length changes all the time. But my usage of the data probably doesn't care how many random bytes I have. It just wants the next available unused byte as soon as it can get it after a request. I could be waiting on the device. But I could also have a buffer of characters pre-read. A length comparison is a needless waste of computation. A null check is more efficient.)
Length-prefix is a good guard against buffer overflow? So is sane usage of library functions and implementation. What if I pass in malformed data? My buffer is 2 bytes long but I tell the function it's 7! Ex: If gets() was intended to be used on known data it could've had an internal buffer check that tested compiled buffers and malloc() calls and still follow spec. If it was meant to be used as a pipe for unknown STDIN to arrive at unknown buffer then clearly one can't know abut the buffer size which means a length arg is pointless, you need something else here like a canary check. For that matter, you can't length-prefix some streams and inputs, you just can't. Which means the length check has to be built into the algorithm and not a magic part of the typing system. TL;DR NUL-terminated never had to be unsafe, it just ended up that way via misuse.
counter-counter point: NUL-termination is annoying on binary. You either need to do length-prefix here or transform NUL bytes in some way: escape-codes, range remapping, etc... which of course means more-memory-usage/reduced-information/more-operations-per-byte. Length-prefix mostly wins the war here. The only upside to a transform is that no additional functions have to be written to cover the length-prefix strings. Which means on your more optimized sub-O(n) routines you can have them automatically act as their O(n) equivalents without adding more code. Downside is, of course, time/memory/compression waste when used on NUL heavy strings. Depending on how much of your library you end up duplicating to operate on binary data, it may make sense to work solely with length-prefix strings. That said one could also do the same with length-prefix strings... -1 length could mean NUL-terminated and you could use NUL-terminated strings inside length-terminated.
Concat: "O(n+m) vs O(m)" I'm assuming your referring to m as the total length of the string after concatenating because they both have to have that number of operations minimum (you can't just tack-on to string 1, what if you have to realloc?). And I'm assuming n is a mythical amount of operations you no longer have to do because of a pre-compute. If so, then the answer is simple: pre-compute. If you're insisting you'll always have enough memory to not need to realloc and that's the basis of the big-O notation then the answer is even more simple: do binary search on allocated memory for end of string 1, clearly there's a large swatch of infinite zeros after string 1 for us to not worry about realloc. There, easily got n to log(n) and I barely tried. Which if you recall log(n) is essentially only ever as large as 64 on a real computer, which is essentially like saying O(64+m), which is essentially O(m). (And yes that logic has been used in run-time analysis of real data structures in-use today. It's not bullshit off the top of my head.)
Concat()/Len() again: Memoize results. Easy. Turns all computes into pre-computes if possible/necessary. This is an algorithmic decision. It's not an enforced constraint of the language.
String suffix passing is easier/possible with NUL termination. Depending on how length-prefix is implemented it can be destructive on original string and can sometimes not even be possible. Requiring a copy and pass O(n) instead of O(1).
Argument-passing/de-referencing is less for NUL-terminated versus length-prefix. Obviously because you're passing less information. If you don't need length, then this saves a lot of footprint and allows optimizations.
You can cheat. It's really just a pointer. Who says you have to read it as a string? What if you want to read it as a single character or a float? What if you want to do the opposite and read a float as a string? If you're careful you can do this with NUL-termination. You can't do this with length-prefix, it's a data type distinctly different from a pointer typically. You'd most likely have to build a string byte-by-byte and get the length. Of course if you wanted something like an entire float (probably has a NUL inside it) you'd have to read byte-by-byte anyway, but the details are left to you to decide.
TL;DR Are you using binary data? If no, then NUL-termination allows more algorithmic freedom. If yes, then code quantity vs speed/memory/compression is your main concern. A blend of the two approaches or memoization might be best.
Many design decisions surrounding C stem from the fact that when it was originally implemented, parameter passing was somewhat expensive. Given a choice between e.g.
void add_element_to_next(arr, offset)
char[] arr;
int offset;
{
arr[offset] += arr[offset+1];
}
char array[40];
void test()
{
for (i=0; i<39; i++)
add_element_to_next(array, i);
}
versus
void add_element_to_next(ptr)
char *p;
{
p[0]+=p[1];
}
char array[40];
void test()
{
int i;
for (i=0; i<39; i++)
add_element_to_next(arr+i);
}
the latter would have been slightly cheaper (and thus preferred) since it only required passing one parameter rather than two. If the method being called didn't need to know the base address of the array nor the index within it, passing a single pointer combining the two would be cheaper than passing the values separately.
While there are many reasonable ways in which C could have encoded string lengths, the approaches that had been invented up to that time would have all required functions that should be able to work with part of a string to accept the base address of the string and the desired index as two separate parameters. Using zero-byte termination made it possible to avoid that requirement. Although other approaches would be better with today's machines (modern compilers often pass parameters in registers, and memcpy can be optimized in ways strcpy()-equivalents cannot) enough production code uses zero-byte terminated strings that it's hard to change to anything else.
PS--In exchange for a slight speed penalty on some operations, and a tiny bit of extra overhead on longer strings, it would have been possible to have methods that work with strings accept pointers directly to strings, bounds-checked string buffers, or data structures identifying substrings of another string. A function like "strcat" would have looked something like [modern syntax]
void strcat(unsigned char *dest, unsigned char *src)
{
struct STRING_INFO d,s;
str_size_t copy_length;
get_string_info(&d, dest);
get_string_info(&s, src);
if (d.si_buff_size > d.si_length) // Destination is resizable buffer
{
copy_length = d.si_buff_size - d.si_length;
if (s.src_length < copy_length)
copy_length = s.src_length;
memcpy(d.buff + d.si_length, s.buff, copy_length);
d.si_length += copy_length;
update_string_length(&d);
}
}
A little bigger than the K&R strcat method, but it would support bounds-checking, which the K&R method doesn't. Further, unlike the current method, it would be possible to easily concatenate an arbitrary substring, e.g.
/* Concatenate 10th through 24th characters from src to dest */
void catpart(unsigned char *dest, unsigned char *src)
{
struct SUBSTRING_INFO *inf;
src = temp_substring(&inf, src, 10, 24);
strcat(dest, src);
}
Note that the lifetime of the string returned by temp_substring would be limited by those of s and src, which ever was shorter (which is why the method requires inf to be passed in--if it was local, it would die when the method returned).
In terms of memory cost, strings and buffers up to 64 bytes would have one byte of overhead (same as zero-terminated strings); longer strings would have slightly more (whether one allowed amounts of overhead between two bytes and the maximum required would be a time/space tradeoff). A special value of the length/mode byte would be used to indicate that a string function was given a structure containing a flag byte, a pointer, and a buffer length (which could then index arbitrarily into any other string).
Of course, K&R didn't implement any such thing, but that's most likely because they didn't want to spend much effort on string handling--an area where even today many languages seem rather anemic.
According to Joel Spolsky in this blog post,
It's because the PDP-7 microprocessor, on which UNIX and the C programming language were invented, had an ASCIZ string type. ASCIZ meant "ASCII with a Z (zero) at the end."
After seeing all the other answers here, I'm convinced that even if this is true, it's only part of the reason for C having null-terminated "strings". That post is quite illuminating as to how simple things like strings can actually be quite hard.
I don't buy the "C has no string" answer. True, C does not support built-in higher-level types but you can still represent data-structures in C and that's what a string is. The fact a string is just a pointer in C does not mean the first N bytes cannot take on special meaning as a the length.
Windows/COM developers will be very familiar with the BSTR type which is exactly like this - a length-prefixed C string where the actual character data starts not at byte 0.
So it seems that the decision to use null-termination is simply what people preferred, not a necessity of the language.
One advantage of NUL-termination over length-prefixing, which I have not seen anyone mention, is the simplicity of string comparison. Consider the comparison standard which returns a signed result for less-than, equal, or greater-than. For length-prefixing the algorithm has to be something along the following lines:
Compare the two lengths; record the smaller, and note if they are equal (this last step might be deferred to step 3).
Scan the two character sequences, subtracting characters at matching indices (or use a dual pointer scan). Stop either when the difference is nonzero, returning the difference, or when the number of characters scanned is equal to the smaller length.
When the smaller length is reached, one string is a prefix of the other. Return negative or positive value according to which is shorter, or zero if of equal length.
Contrast this with the NUL-termination algorithm:
Scan the two character sequences, subtracting characters at matching indices [note that this is handled better with moving pointers]. Stop when the difference is nonzero, returning the difference. NOTE: If one string is a PROPER prefix of the other, one of the characters in the subtraction will be NUL, i.e zero, and the comparison will naturally stop there.
If the difference is zero, -only then- check if either character is NUL. If so, return zero, otherwise continue to next character.
The NUL-terminated case is simpler, and very easy to implement efficiently with a dual pointer scan. The length-prefixed case does at least as much work, nearly always more. If your algorithm has to do a lot of string comparisons [e.g a compiler!], the NUL-terminated case wins out. Nowadays that might not be as important, but back in the day, heck yeah.
gcc accept the codes below:
char s[4] = "abcd";
and it's ok if we treat is as an array of chars but not string. That is, we can access it with s[0], s[1], s[2], and s[3], or even with memcpy(dest, s, 4). But we'll get messy characters when we trying with puts(s), or worse with strcpy(dest, s).
I think the better question is why you think C owes you anything? C was designed to give you what you need, nothing more. You need to loose the mentality that the language must provide you with everything. Or just continue to use your higher level languages that will give you the luxary of String, Calendar, Containers; and in the case of Java you get one thing in tonnes of variety. Multiple types String, multiple types of unordered_map(s).
Too bad for you, this was not the purpose of C. C was not designed to be a bloated language that offers from a pin to an anchor. Instead you must rely on third party libraries or your own. And there is nothing easier than creating a simple struct that will contain a string and its size.
struct String
{
const char *s;
size_t len;
};
You know what the problem is with this though. It is not standard. Another language might decide to organize the len before the string. Another language might decide to use a pointer to end instead. Another might decide to use six pointers to make the String more efficient. However a null terminated string is the most standard format for a string; which you can use to interface with any language. Even Java JNI uses null terminated strings.
Lastly, it is a common saying; the right data structure for the task. If you find that need to know the size of a string more than anything else; well use a string structure that allows you to do that optimally. But don't make claims that that operation is used more than anything else for everybody. Like, why is knowing the size of a string more important than reading its contents. I find that reading the contents of a string is what I mostly do, so I use null terminated strings instead of std::string; which saves me 5 pointers on a GCC compiler. If I can even save 2 pointers that is good.

Sub-string representation — length or pointer to last byte?

Imagine parsing a string and wanting to extract a sub-string. To represent this sub-string, I see two ways:
// 1. represent it using a start pointer and a length
struct { char *start; size_t length; };
// 2. represent it using two pointers, start and end
struct { char *start; char *end; };
// or it could as well be returned by a function:
char *find_substring(char *s, size_t s_len, size_t *substring_len);
char *find_substring(char *s, size_t s_len, char **substring_end);
Is there a reason to prefer one form over the other? Is it only down to preferences? I don't see it affecting performances as one can be translated into the other using a simple addition/subtraction but I might be wrong.
The context is an HTTP request parser if that changes anything. I used the first one but I'm curious if the second one brings anything to the table as I have seen it used in picohttpparser.
Is there a reason to prefer one form over the other?
One could choose optimization and speed of execution as the measure of preference over on the other.
If more often you append data on the end, then *end++ would be faster over start[length++].
If more often you get the length of the string, then just length would be faster then end - start.
Remember about rules of optimization. The only real answer comes from profiling your code.
Is it only down to preferences?
I advise to prefer the more appropriate representation to the problem you are trying to model, based on how readable it is, how easy it is to use it and find bugs in it, which comes down to personal preference.
We could also inspect existing implementations. In C, all(all?) C standard functions and in POSIX like strbuf, aiocb, XSI messages queues, iovec use pointer+integer to represent a memory region. I think all C++ implementations of std::vector, like glibc std::vector or llvm vector, use pointers internally, but one can expect they be optimized for push_back() operations.
Generally I lean over to use pointers. When operating on size_t you have to handle overflow and underflow and negative/too big values or converting from pointer difference ptrdiff_t to size_t. Such problems kind-of disappear with pointers - a pointer is either valid, or not, you need only bound check using < > operators if you may in-/decrement it or not. However when writing an external api, I would use size_t, as C programmers are used to represent memory region using that convention.
In most cases this would be down to personal preference. I guess most people choose the first representation. But depending what you plan to do with that substring the second implementation may be better performance-wise.
With the second implementation you have to be specific about where end points to: is it the last character still in substring or the first character beyond substring.
The first way is preferred way. For example, consider that you have to deal with very large strings. Then it will not remain a simple allocation of bytes. In that case you have to represent it in a more complicated manner.
The second way leaks the information about internal representation of string while the first does not.

Two approaches to writing functions

I am asking this question in the context of the C language, though it applies really to any language supporting pointers or pass-by-reference functionality.
I come from a Java background, but have written enough low-level code (C and C++) to have observed this interesting phenomenon. Supposing we have some object X (not using "object" here in the strictest OOP sense of the word) that we want to fill with information by way of some other function, it seems there are two approaches to doing so:
Returning an instance of that object's type and assigning it, e.g. if X has type T, then we would have:
T func(){...}
X = func();
Passing in a pointer / reference to the object and modifying it inside the function, and returning either void or some other value (in C, for instance, a lot of functions return an int corresponding to the success/failure of the operation). An example of this here is:
int func(T* x){...x = 1;...}
func(&X);
My question is: in what situations makes one method better than the other? Are they equivalent approaches to accomplishing the same outcome? What are the restrictions of each?
Thanks!
There is a reason that you should always consider using the second method, rather than the first. If you look at the return values for the entirety of the C standard library, you'll notice that there's almost always an element of error handling involved in them. For example, you have to check the return value of the following functions before you assume they've succeeded:
calloc, malloc and realloc
getchar
fopen
scanf and family
strtok
There are other non-standard functions that follow this pattern:
pthread_create, etc.
socket, connect, etc.
open, read, write, etc.
Generally speaking, a return value conveys a number of items successfully read/written/converted or a flat-out boolean success/fail value, and in practice you'll almost always need such a return value, unless you're going to exit(EXIT_FAILURE); at any errors (in which case I would rather not use your modules, because they give me no opportunity to clean up within my own code).
There are functions that don't use this pattern in the standard C library, because they use no resources (e.g. allocations or files) and so there's no chance of any error. If your function is a basic translation function (e.g. like toupper, tolower and friends which translate single character values), for example, then you don't need a return value for error handling because there are no errors. I think you'll find this scenario quite rare indeed, but if that is your scenario, by all means use the first option!
In summary, you should always highly consider using option 2, reserving the return value for a similar use, for the sake of consistent with the rest of the world, and because you might later decide that you need the return value for communicating errors or number of items processed.
Method (1) passes the object by value, which requires that the object be copied. It's copied when you pass it in and copied again when it's returned. Method (2) passes only a pointer. When you're passing a primitive, (1) is just fine, but when you're passing an object, a struct, or an array, that's just wasted space and time.
In Java and many other languages, objects are always passed by reference. Behind the scenes, only a pointer is copied. This means that even though the syntax looks like (1), it actually works like (2).
I think I got you.
These to approach are very different.
The question you have to ask your self when ever you trying to decide which approach to take is :
Which class would have the responsibility?
In case you passing the reference to the object you are decapul the creation of the object to the caller and creating this functionality to be more serviceability and you would be able to create a util class that all of the functions inside will be stateless, they are getting object manipulate the input and returning it.
The other approach is more likely and API, you are requesting an opperation.
For an example, you are getting array of bytes and you would like to convert it to string, you would probably would chose the first approch.
And if you would like to do some opperation in DB you would chose the second one.
When ever you will have more than 1 function from the first approch that cover the same area you would encapsulate it into a util class, same applay to the second, you will encapsulate it into an API.
In method 2, we call x an output parameter. This is actually a very common design utilized in a lot of places...think some of the various built-in C functions that populate a text buffer, like snprintf.
This has the benefit of being fairly space-efficient, since you won't be copying structs/arrays/data onto the stack and returning brand new instances.
A really, really convenient quality of method 2 is that you can essentially have any number of "return values." You "return" data through the output parameters, but you can also return a success/error indicator from the function.
A good example of method 2 being used effectively is in the built-in C function strtol. This function converts a string to a long (basically, parses a number from a string). One of the parameters is a char **. When calling the function, you declare char * endptr locally, and pass in &endptr.
The function will return either:
the converted value if it was successful,
0 if it failed, or
LONG_MIN or LONG_MAX if it was out of range
as well as set the endptr to point to the first non-digit it found.
This is great for error reporting if your program depends on user input, because you can check for failure in so many ways and report different errors for each.
If endptr isn't null after the call to strtol, then you know precisely that the user entered a non-integer, and you can print straight away the character that the conversion failed on if you'd like.
Like Thom points out, Java makes implementing method 2 simpler by simulating pass-by-reference behavior, which is just pointers behind the scenes without the pointer syntax in the source code.
To answer your question: I think C lends itself well to the second method. Functions like realloc are there to give you more space when you need it. However, there isn't much stopping you from using the first method.
Maybe you're trying to implement some kind of immutable object. The first method will be the choice there. But in general, I opt for the second.
(Assuming we are talking about returning only one value from the function.)
In general, the first method is used when type T is relatively small. It is definitely preferable with scalar types. It can be used with larger types. What is considered "small enough" for these purposes depends on the platform and the expected performance impact. (The latter is caused by the fact that the returned object is copied.)
The second method is used when the object is relatively large, since this method does not perform any copying. And with non-copyable types, like arrays, you have no choice but to use the second method.
Of course, when performance is not an issue, the first method can be easily used to return large objects.
An interesting matter is optimization opportunities available to C compiler. In C++ language compilers are allowed to perform Return Value Optimizations (RVO, NRVO), which effectively turn the first method into the second one "under the hood" in situations when the second method offers better performance. To facilitate such optimizations C++ language relaxes some address-identity requirements imposed on the involved objects. AFAIK, C does not offer such relaxations, thus preventing (or at least impeding) any attempts at RVO/NRVO.
Short answer: take 2 if you don't have a necessary reason to take 1.
Long answer: In the world of C++ and its derived languages, Java, C#, exceptions help a lot. In C world, there is not very much you can do. Following is an sample API I take from CUDA library, which is a library I like and consider well designed:
cudaError_t cudaMalloc (void **devPtr, size_t size);
compare this API with malloc:
void *malloc(size_t size);
in old C interfaces, there are many such examples:
int open(const char *pathname, int flags);
FILE *fopen(const char *path, const char *mode);
I would argue to the end of the world, the interface CUDA is providing is much obvious and lead to proper result.
There are other set of interfaces that the valid return value space actually overlaps with the error code, so the designers of those interfaces scratched their heads and come up with not brilliant at all ideas, say:
ssize_t read(int fd, void *buf, size_t count);
a daily function like reading a file content is restricted by the definition of ssize_t. since the return value has to encode error code too, it has to provide negative number. in a 32bit system, the max of ssize_t is 2G, which is very much limited the number of bytes you can read from your file.
If your error designator is encoded inside of the function return value, I bet 10/10 programmers won't try to check it, though they really know they should; they just don't, or don't remember, because the form is not obvious.
And another reason, is human beings are very lazy and not good at dealing if's. The documentation of these functions will describe that:
if return value is NULL then ... blah.
if return value is 0 then ... blah.
yak.
In the first form, things changes. How do you judge if the value has been returned? No NULL or 0 any more. You have to use SUCCESS, FAILURE1, FAILURE2, or something similar. This interface forces users to code more safer and makes the code much robust.
With these macro, or enum, it's much easier for programmers to learn about the effect of the API and the cause of different exceptions too. With all these advantages, there actually is no extra runtime overhead for it too.
I will try to explain :)
Let say you have to load a giant rocket into semi,
Method 1)
Truck driver places a truck on a parking lot, and goes on to find a hookers, you are stack with putting the load onto forklift or some kind of trailer to bring it to the track.
Method 2)
Truck driver forgets hooker and backs truck up right to the rocket, then you need just to push it in.
That is the difference between those two :). What it boils down to in programming is:
Method 1)
Caller function reserves and address for called function to return its return value to, but how is calling function going to get that value does not matter, will it have to reserve another address or not does not matter, I need something returned, it is your job to get it to me :). So called function goes and reserves the address for its calculations and than stores the value in address then returns value to caller. So caller goes and say oh thank you let me just copy it to the address I reserved earlier.
Method 2)
Caller function says "Hey I will help you, I will give you the address that I have reserved, store what ever calculations you do in it", this way you save not only memory but you save in time.
And I think second is better, and here is why:
So let say that you have struct with 1000 ints inside of it, method 1 would be pointless, it will have to reserve 2*100*32 bits of memory, which is 6400 plus you have to copy it to first location than copy it to second one. So if each copy takes 1 millisecond you will need to way 6.4 seconds to store and copy variables. Where if you have address you only have to store it once.
They are equivalent to me but not in the implementation.
#include <stdio.h>
#include <stdlib.h>
int func(int a,int b){
return a+b;
}
int funn(int *x){
*x=1;
return 777;
}
int main(void){
int sx,*dx;
/* case static' */
sx=func(4,6); /* looks legit */
funn(&sx); /* looks wrong in this case */
/* case dynamic' */
dx=malloc(sizeof(int));
if(dx){
*dx=func(4,6); /* looks wrong in this case */
sx=funn(dx); /* looks legit */
free(dx);
}
return 0;
}
In a static' approach it is more comfortable to me doing your first method. Because I don't want to mess with the dynamic part (with legit pointers).
But in a dynamic' approach I'll use your second method. Because it is made for it.
So they are equivalent but not the same, the second approach is clearly made for pointers and so for the dynamic part.
And so far more clear ->
int main(void){
int sx,*dx;
sx=func(4,6);
dx=malloc(sizeof(int));
if(dx){
sx=funn(dx);
free(dx);
}
return 0;
}
than ->
int main(void){
int sx,*dx;
funn(&sx);
dx=malloc(sizeof(int));
if(dx){
*dx=func(4,6);
free(dx);
}
return 0;
}

Can a C implementation use length-prefixed-strings "under the hood"?

After reading this question: What are the problems of a zero-terminated string that length-prefixed strings overcome? I started to wonder, what exactly is stopping a C implementation from allocating a few extra bytes for any char or wchar_t array allocated on the stack or heap and using them as a "string prefix" to store the number N of its elements?
Then, if the N-th character is '\0', N - 1 would signify the string length.
I believe this could mightily boost performance of functions such as strlen or strcat.
This could potentially turn to extra memory consumption if a program uses non-0-terminated char arrays extensively, but that could be remedied by a compiler flag turning on or off the regular "count-until-you-reach-'\0'" routine for the compiled code.
What are possible obstacles for such an implementation? Does the C Standard allow for this? What problems can this technique cause that I haven't accounted for?
And... has this actually ever been done?
You can store the length of the allocation. And malloc implementations really do do that (or some do, at least).
You can't reasonably store the length of whatever string is stored in the allocation, though, because the user can change the contents to their whim; it would be unreasonable to keep the length up to date. Furthermore, users might start strings somewhere in the middle of the character array, or might not even be using the array to hold a string!
Then, if the N-th character is '\0', N - 1 would signify the string length.
Actually, no, and that's why this suggestion cannot work.
If I overwrite a character in a string with a 0, I have effectively truncated the string, and a subsequent call of strlen on the string must return the truncated length. (This is commonly done by application programs, including every scanner generated by (f)lex, as well as the strtok standard library function. Amongst others.)
Moreover, it is entirely legal to call strlen on an interior byte of the string.
For example (just for demonstration purposes, although I'll bet you can find code almost identical to this in common use.)
/* Split a string like 'key=value...' into key and value parts, and
* return the value, and optionally its length (if the second argument
* is not a NULL pointer).
* On success, returns the value part and modifieds the original string
* so that it is the key.
* If there is no '=' in the supplied string, neither it nor the value
* pointed to by plen are modified, and NULL is returned.
*/
char* keyval_split(char* keyval, int* plen) {
char* delim = strchr(keyval, '=');
if (delim) {
if (plen) *plen = strlen(delim + 1)
*delim = 0;
return delim + 1;
} else {
return NULL;
}
}
There's nothing fundamentally stopping you from doing this in your application, if that was useful (one of the comments noted this). There are two problems that would emerge, however:
You'd have to reimplement all the string-handling functions, and have my_strlen, my_strcpy, and so on, and add string-creating functions. That might be annoying, but it's a bounded problem.
You'd have to stop people, when writing for the system, deliberately or automatically treating the associated character arrays as ‘ordinary’ C strings, and using the usual functions on them. You might have to make sure that such usages broke promptly.
This means that it would, I think, be infeasible to smuggle a reimplemented ‘C string’ into an existing program.
Something like
typedef struct {
size_t len;
char* buf;
} String;
size_t my_strlen(String*);
...
might work, since type-checking would frustrate (2) (unless someone decided to hack things ‘for efficiency’, in which case there's not much you can do).
Of course, you wouldn't do this unless and until you'd proven that string management was the bottleneck in your code and that this approach provably improved things....
There are a couple of issues with this approach. First of all, you wouldn't be able to create arbitrarily long strings. If you only reserve 1 byte for length, then your string can only go up to 255 characters. You can certainly use more bytes to store the length, but how many? 2? 4?
What if you try to concatenate two strings that are both at the edge of their size limits (i.e., if you use 1 byte for length and try to concatenate two 250-character strings to each other, what happens)? Do you simply add more bytes to the length as necessary?
Secondly, where do you store this metadata? It somehow has to be associated with the string. This is similar to the problem Dennis Ritchie ran into when he was implementing arrays in C. Originally, array objects stored an explicit pointer to the first element of the array, but as he added struct types to the language, he realized that he didn't want that metadata cluttering up the representation of the struct object in memory, so he got rid of it and introduced the rule that array expressions get converted to pointer expressions in most circumstances.
You could create a new aggregate type like
struct string
{
char *data;
size_t len;
};
but then you wouldn't be able to use the C string library to manipulate objects of that type; an implementation would still have to support the existing interface.
You could store the length in the leading byte or bytes of the string, but how many do you reserve? You could use a variable number of bytes to store the length, but now you need a way to distinguish length bytes from content bytes, and you can't read the first character by simply dereferencing the pointer. Functions like strcat would have to know how to step around the length bytes, how to adjust the contents if the number of length bytes changes, etc.
The 0-terminated approach has its disadvantages, but it's also a helluva lot easier to implement and makes manipulating strings a lot easier.
The string methods in the standard library have defined semantics. If one generates an array of char that contains various values, and passes a pointer to the array or a portion thereof, the methods whose behavior is defined in terms of NUL bytes must search for NUL bytes in the same fashion as defined by the standard.
One could define one's own methods for string handling which use a better form of string storage, and simply pretend that the standard library string-related functions don't exist unless one must pass strings to things like fopen. The biggest difficulty with such an approach is that unless one uses non-portable compiler features it would not be possible to use in-line string literals. Instead of saying:
ns_output(my_file, "This is a test"); // ns -- new string
one would have to say something more like:
MAKE_NEW_STRING(this_is_a_test, "This is a test");
ns_output(my_file, this_is_a_test);
where the macro MAKE_NEW_STRING would create a union of an anonymous type, define an instance called this_is_a_test, and suitably initialize it. Since a lot of strings would be of different anonymous types, type-checking would require that strings be unions that include a member of a known array type, and code expecting strings should be given a pointer that member, likely using something like:
#define ns_output(f,s) (ns_output_func((f),(s).stringref))
It would be possible to define the types in such a way as to avoid the need for the stringref member and have code just accept void*, but the stringref member would essentially perform static duck-typing (only things with a stringref member could be given to such a macro) and could also allow type-checking on the type of stringref itself).
If one could accept those constraints, I think one could probably write code that was more efficient in almost every way that zero-terminated strings; the question would be whether the advantages would be worth the hassle.

Partially Prefixed Null Terminated Strings [duplicate]

As much as I love C and C++, I can't help but scratch my head at the choice of null terminated strings:
Length prefixed (i.e. Pascal) strings existed before C
Length prefixed strings make several algorithms faster by allowing constant time length lookup.
Length prefixed strings make it more difficult to cause buffer overrun errors.
Even on a 32 bit machine, if you allow the string to be the size of available memory, a length prefixed string is only three bytes wider than a null terminated string. On 16 bit machines this is a single byte. On 64 bit machines, 4GB is a reasonable string length limit, but even if you want to expand it to the size of the machine word, 64 bit machines usually have ample memory making the extra seven bytes sort of a null argument. I know the original C standard was written for insanely poor machines (in terms of memory), but the efficiency argument doesn't sell me here.
Pretty much every other language (i.e. Perl, Pascal, Python, Java, C#, etc) use length prefixed strings. These languages usually beat C in string manipulation benchmarks because they are more efficient with strings.
C++ rectified this a bit with the std::basic_string template, but plain character arrays expecting null terminated strings are still pervasive. This is also imperfect because it requires heap allocation.
Null terminated strings have to reserve a character (namely, null), which cannot exist in the string, while length prefixed strings can contain embedded nulls.
Several of these things have come to light more recently than C, so it would make sense for C to not have known of them. However, several were plain well before C came to be. Why would null terminated strings have been chosen instead of the obviously superior length prefixing?
EDIT: Since some asked for facts (and didn't like the ones I already provided) on my efficiency point above, they stem from a few things:
Concat using null terminated strings requires O(n + m) time complexity. Length prefixing often require only O(m).
Length using null terminated strings requires O(n) time complexity. Length prefixing is O(1).
Length and concat are by far the most common string operations. There are several cases where null terminated strings can be more efficient, but these occur much less often.
From answers below, these are some cases where null terminated strings are more efficient:
When you need to cut off the start of a string and need to pass it to some method. You can't really do this in constant time with length prefixing even if you are allowed to destroy the original string, because the length prefix probably needs to follow alignment rules.
In some cases where you're just looping through the string character by character you might be able to save a CPU register. Note that this works only in the case that you haven't dynamically allocated the string (Because then you'd have to free it, necessitating using that CPU register you saved to hold the pointer you originally got from malloc and friends).
None of the above are nearly as common as length and concat.
There's one more asserted in the answers below:
You need to cut off the end of the string
but this one is incorrect -- it's the same amount of time for null terminated and length prefixed strings. (Null terminated strings just stick a null where you want the new end to be, length prefixers just subtract from the prefix.)
From the horse's mouth
None of BCPL, B, or C supports
character data strongly in the
language; each treats strings much
like vectors of integers and
supplements general rules by a few
conventions. In both BCPL and B a
string literal denotes the address of
a static area initialized with the
characters of the string, packed into
cells. In BCPL, the first packed byte
contains the number of characters in
the string; in B, there is no count
and strings are terminated by a
special character, which B spelled
*e. This change was made partially
to avoid the limitation on the length
of a string caused by holding the
count in an 8- or 9-bit slot, and
partly because maintaining the count
seemed, in our experience, less
convenient than using a terminator.
Dennis M Ritchie, Development of the C Language
C doesn't have a string as part of the language. A 'string' in C is just a pointer to char. So maybe you're asking the wrong question.
"What's the rationale for leaving out a string type" might be more relevant. To that I would point out that C is not an object oriented language and only has basic value types. A string is a higher level concept that has to be implemented by in some way combining values of other types. C is at a lower level of abstraction.
in light of the raging squall below:
I just want to point out that I'm not trying to say this is a stupid or bad question, or that the C way of representing strings is the best choice. I'm trying to clarify that the question would be more succinctly put if you take into account the fact that C has no mechanism for differentiating a string as a datatype from a byte array. Is this the best choice in light of the processing and memory power of todays computers? Probably not. But hindsight is always 20/20 and all that :)
The question is asked as a Length Prefixed Strings (LPS) vs zero terminated strings (SZ) thing, but mostly expose benefits of length prefixed strings. That may seem overwhelming, but to be honest we should also consider drawbacks of LPS and advantages of SZ.
As I understand it, the question may even be understood as a biased way to ask "what are the advantages of Zero Terminated Strings ?".
Advantages (I see) of Zero Terminated Strings:
very simple, no need to introduce new concepts in language, char
arrays/char pointers can do.
the core language just include minimal syntaxic sugar to convert
something between double quotes to a
bunch of chars (really a bunch of
bytes). In some cases it can be used
to initialize things completely
unrelated with text. For instance xpm
image file format is a valid C source
that contains image data encoded as a
string.
by the way, you can put a zero in a string literal, the compiler will
just also add another one at the end of the literal: "this\0is\0valid\0C".
Is it a string ? or four strings ? Or a bunch of bytes...
flat implementation, no hidden indirection, no hidden integer.
no hidden memory allocation involved (well, some infamous non
standard functions like strdup
perform allocation, but that's mostly
a source of problem).
no specific issue for small or large hardware (imagine the burden to
manage 32 bits prefix length on 8
bits microcontrollers, or the
restrictions of limiting string size
to less than 256 bytes, that was a problem I actually had with Turbo Pascal eons ago).
implementation of string manipulation is just a handful of
very simple library function
efficient for the main use of strings : constant text read
sequentially from a known start
(mostly messages to the user).
the terminating zero is not even mandatory, all necessary tools
to manipulate chars like a bunch of
bytes are available. When performing
array initialisation in C, you can
even avoid the NUL terminator. Just
set the right size. char a[3] =
"foo"; is valid C (not C++) and
won't put a final zero in a.
coherent with the unix point of view "everything is file", including
"files" that have no intrinsic length
like stdin, stdout. You should remember that open read and write primitives are implemented
at a very low level. They are not library calls, but system calls. And the same API is used
for binary or text files. File reading primitives get a buffer address and a size and return
the new size. And you can use strings as the buffer to write. Using another kind of string
representation would imply you can't easily use a literal string as the buffer to output, or
you would have to make it have a very strange behavior when casting it to char*. Namely
not to return the address of the string, but instead to return the actual data.
very easy to manipulate text data read from a file in-place, without useless copy of buffer,
just insert zeroes at the right places (well, not really with modern C as double quoted strings are const char arrays nowaday usually kept in non modifiable data segment).
prepending some int values of whatever size would implies alignment issues. The initial
length should be aligned, but there is no reason to do that for the characters datas (and
again, forcing alignment of strings would imply problems when treating them as a bunch of
bytes).
length is known at compile time for constant literal strings (sizeof). So why would
anyone want to store it in memory prepending it to actual data ?
in a way C is doing as (nearly) everyone else, strings are viewed as arrays of char. As array length is not managed by C, it is logical length is not managed either for strings. The only surprising thing is that 0 item added at the end, but that's just at core language level when typing a string between double quotes. Users can perfectly call string manipulation functions passing length, or even use plain memcopy instead. SZ are just a facility. In most other languages array length is managed, it's logical that is the same for strings.
in modern times anyway 1 byte character sets are not enough and you often have to deal with encoded unicode strings where the number of characters is very different of the number of bytes. It implies that users will probably want more than "just the size", but also other informations. Keeping length give use nothing (particularly no natural place to store them) regarding these other useful pieces of information.
That said, no need to complain in the rare case where standard C strings are indeed inefficient. Libs are available. If I followed that trend, I should complain that standard C does not include any regex support functions... but really everybody knows it's not a real problem as there is libraries available for that purpose. So when string manipulation efficiency is wanted, why not use a library like bstring ? Or even C++ strings ?
EDIT: I recently had a look to D strings. It is interesting enough to see that the solution choosed is neither a size prefix, nor zero termination. As in C, literal strings enclosed in double quotes are just short hand for immutable char arrays, and the language also has a string keyword meaning that (immutable char array).
But D arrays are much richer than C arrays. In the case of static arrays length is known at run-time so there is no need to store the length. Compiler has it at compile time. In the case of dynamic arrays, length is available but D documentation does not state where it is kept. For all we know, compiler could choose to keep it in some register, or in some variable stored far away from the characters data.
On normal char arrays or non literal strings there is no final zero, hence programmer has to put it itself if he wants to call some C function from D. In the particular case of literal strings, however the D compiler still put a zero at the end of each strings (to allow easy cast to C strings to make easier calling C function ?), but this zero is not part of the string (D does not count it in string size).
The only thing that disappointed me somewhat is that strings are supposed to be utf-8, but length apparently still returns a number of bytes (at least it's true on my compiler gdc) even when using multi-byte chars. It is unclear to me if it's a compiler bug or by purpose. (OK, I probably have found out what happened. To say to D compiler your source use utf-8 you have to put some stupid byte order mark at beginning. I write stupid because I know of not editor doing that, especially for UTF-8 that is supposed to be ASCII compatible).
I think, it has historical reasons and found this in wikipedia:
At the time C (and the languages that
it was derived from) were developed,
memory was extremely limited, so using
only one byte of overhead to store the
length of a string was attractive. The
only popular alternative at that time,
usually called a "Pascal string"
(though also used by early versions of
BASIC), used a leading byte to store
the length of the string. This allows
the string to contain NUL and made
finding the length need only one
memory access (O(1) (constant) time).
But one byte limits the length to 255.
This length limitation was far more
restrictive than the problems with the
C string, so the C string in general
won out.
Calavera is right, but as people don't seem to get his point, I'll provide some code examples.
First, let's consider what C is: a simple language, where all code has a pretty direct translation into machine language. All types fit into registers and on the stack, and it doesn't require an operating system or a big run-time library to run, since it were meant to write these things (a task to which is superbly well-suited, considering there isn't even a likely competitor to this day).
If C had a string type, like int or char, it would be a type which didn't fit in a register or in the stack, and would require memory allocation (with all its supporting infrastructure) to be handled in any way. All of which go against the basic tenets of C.
So, a string in C is:
char s*;
So, let's assume then that this were length-prefixed. Let's write the code to concatenate two strings:
char* concat(char* s1, char* s2)
{
/* What? What is the type of the length of the string? */
int l1 = *(int*) s1;
/* How much? How much must I skip? */
char *s1s = s1 + sizeof(int);
int l2 = *(int*) s2;
char *s2s = s2 + sizeof(int);
int l3 = l1 + l2;
char *s3 = (char*) malloc(l3 + sizeof(int));
char *s3s = s3 + sizeof(int);
memcpy(s3s, s1s, l1);
memcpy(s3s + l1, s2s, l2);
*(int*) s3 = l3;
return s3;
}
Another alternative would be using a struct to define a string:
struct {
int len; /* cannot be left implementation-defined */
char* buf;
}
At this point, all string manipulation would require two allocations to be made, which, in practice, means you'd go through a library to do any handling of it.
The funny thing is... structs like that do exist in C! They are just not used for your day-to-day displaying messages to the user handling.
So, here is the point Calavera is making: there is no string type in C. To do anything with it, you'd have to take a pointer and decode it as a pointer to two different types, and then it becomes very relevant what is the size of a string, and cannot just be left as "implementation defined".
Now, C can handle memory in anyway, and the mem functions in the library (in <string.h>, even!) provide all the tooling you need to handle memory as a pair of pointer and size. The so-called "strings" in C were created for just one purpose: showing messages in the context of writting an operating system intended for text terminals. And, for that, null termination is enough.
Obviously for performance and safety, you'll want to keep the length of a string while you're working with it rather than repeatedly performing strlen or the equivalent on it. However, storing the length in a fixed location just before the string contents is an incredibly bad design. As Jörgen pointed out in the comments on Sanjit's answer, it precludes treating the tail of a string as a string, which for example makes a lot of common operations like path_to_filename or filename_to_extension impossible without allocating new memory (and incurring the possibility of failure and error handling). And then of course there's the issue that nobody can agree how many bytes the string length field should occupy (plenty of bad "Pascal string" languages used 16-bit fields or even 24-bit fields which preclude processing of long strings).
C's design of letting the programmer choose if/where/how to store the length is much more flexible and powerful. But of course the programmer has to be smart. C punishes stupidity with programs that crash, grind to a halt, or give your enemies root.
Lazyness, register frugality and portability considering the assembly gut of any language, especially C which is one step above assembly (thus inheriting a lot of assembly legacy code).
You would agree as a null char would be useless in those ASCII days, it (and probably as good as an EOF control char ).
let's see in pseudo code
function readString(string) // 1 parameter: 1 register or 1 stact entries
pointer=addressOf(string)
while(string[pointer]!=CONTROL_CHAR) do
read(string[pointer])
increment pointer
total 1 register use
case 2
function readString(length,string) // 2 parameters: 2 register used or 2 stack entries
pointer=addressOf(string)
while(length>0) do
read(string[pointer])
increment pointer
decrement length
total 2 register used
That might seem shortsighted at that time, but considering the frugality in code and register ( which were PREMIUM at that time, the time when you know, they use punch card ). Thus being faster ( when processor speed could be counted in kHz), this "Hack" was pretty darn good and portable to register-less processor with ease.
For argument sake I will implement 2 common string operation
stringLength(string)
pointer=addressOf(string)
while(string[pointer]!=CONTROL_CHAR) do
increment pointer
return pointer-addressOf(string)
complexity O(n) where in most case PASCAL string is O(1) because the length of the string is pre-pended to the string structure (that would also mean that this operation would have to be carried in an earlier stage).
concatString(string1,string2)
length1=stringLength(string1)
length2=stringLength(string2)
string3=allocate(string1+string2)
pointer1=addressOf(string1)
pointer3=addressOf(string3)
while(string1[pointer1]!=CONTROL_CHAR) do
string3[pointer3]=string1[pointer1]
increment pointer3
increment pointer1
pointer2=addressOf(string2)
while(string2[pointer2]!=CONTROL_CHAR) do
string3[pointer3]=string2[pointer2]
increment pointer3
increment pointer1
return string3
complexity O(n) and prepending the string length wouldn't change the complexity of the operation, while I admit it would take 3 time less time.
On another hand, if you use PASCAL string you would have to redesign your API for taking in account register length and bit-endianness, PASCAL string got the well known limitation of 255 char (0xFF) beacause the length was stored in 1 byte (8bits), and it you wanted a longer string (16bits->anything) you would have to take in account the architecture in one layer of your code, that would mean in most case incompatible string APIs if you wanted longer string.
Example:
One file was written with your prepended string api on an 8 bit computer and then would have to be read on say a 32 bit computer, what would the lazy program do considers that your 4bytes are the length of the string then allocate that lot of memory then attempt to read that many bytes.
Another case would be PPC 32 byte string read(little endian) onto a x86 (big endian), of course if you don't know that one is written by the other there would be trouble.
1 byte length (0x00000001) would become 16777216 (0x0100000) that is 16 MB for reading a 1 byte string.
Of course you would say that people should agree on one standard but even 16bit unicode got little and big endianness.
Of course C would have its issues too but, would be very little affected by the issues raised here.
In many ways, C was primitive. And I loved it.
It was a step above assembly language, giving you nearly the same performance with a language that was much easier to write and maintain.
The null terminator is simple and requires no special support by the language.
Looking back, it doesn't seem that convenient. But I used assembly language back in the 80s and it seemed very convenient at the time. I just think software is continually evolving, and the platforms and tools continually get more and more sophisticated.
Assuming for a moment that C implemented strings the Pascal way, by prefixing them by length: is a 7 char long string the same DATA TYPE as a 3-char string? If the answer is yes, then what kind of code should the compiler generate when I assign the former to the latter? Should the string be truncated, or automatically resized? If resized, should that operation be protected by a lock as to make it thread safe? The C approach side stepped all these issues, like it or not :)
Somehow I understood the question to imply there's no compiler support for length-prefixed strings in C. The following example shows, at least you can start your own C string library, where string lengths are counted at compile time, with a construct like this:
#define PREFIX_STR(s) ((prefix_str_t){ sizeof(s)-1, (s) })
typedef struct { int n; char * p; } prefix_str_t;
int main() {
prefix_str_t string1, string2;
string1 = PREFIX_STR("Hello!");
string2 = PREFIX_STR("Allows \0 chars (even if printf directly doesn't)");
printf("%d %s\n", string1.n, string1.p); /* prints: "6 Hello!" */
printf("%d %s\n", string2.n, string2.p); /* prints: "48 Allows " */
return 0;
}
This won't, however, come with no issues as you need to be careful when to specifically free that string pointer and when it is statically allocated (literal char array).
Edit: As a more direct answer to the question, my view is this was the way C could support both having string length available (as a compile time constant), should you need it, but still with no memory overhead if you want to use only pointers and zero termination.
Of course it seems like working with zero-terminated strings was the recommended practice, since the standard library in general doesn't take string lengths as arguments, and since extracting the length isn't as straightforward code as char * s = "abc", as my example shows.
"Even on a 32 bit machine, if you allow the string to be the size of available memory, a length prefixed string is only three bytes wider than a null terminated string."
First, extra 3 bytes may be considerable overhead for short strings. In particular, a zero-length string now takes 4 times as much memory. Some of us are using 64-bit machines, so we either need 8 bytes to store a zero-length string, or the string format can't cope with the longest strings the platform supports.
There may also be alignment issues to deal with. Suppose I have a block of memory containing 7 strings, like "solo\0second\0\0four\0five\0\0seventh". The second string starts at offset 5. The hardware may require that 32-bit integers be aligned at an address that is a multiple of 4, so you have to add padding, increasing the overhead even further. The C representation is very memory-efficient in comparison. (Memory-efficiency is good; it helps cache performance, for example.)
One point not yet mentioned: when C was designed, there were many machines where a 'char' was not eight bits (even today there are DSP platforms where it isn't). If one decides that strings are to be length-prefixed, how many 'char's worth of length prefix should one use? Using two would impose an artificial limit on string length for machines with 8-bit char and 32-bit addressing space, while wasting space on machines with 16-bit char and 16-bit addressing space.
If one wanted to allow arbitrary-length strings to be stored efficiently, and if 'char' were always 8-bits, one could--for some expense in speed and code size--define a scheme were a string prefixed by an even number N would be N/2 bytes long, a string prefixed by an odd value N and an even value M (reading backward) could be ((N-1) + M*char_max)/2, etc. and require that any buffer which claims to offer a certain amount of space to hold a string must allow enough bytes preceding that space to handle the maximum length. The fact that 'char' isn't always 8 bits, however, would complicate such a scheme, since the number of 'char' required to hold a string's length would vary depending upon the CPU architecture.
The null termination allows for fast pointer based operations.
Not a Rationale necessarily but a counterpoint to length-encoded
Certain forms of dynamic length encoding are superior to static length encoding as far as memory is concerned, it all depends on usage. Just look at UTF-8 for proof. It's essentially an extensible character array for encoding a single character. This uses a single bit for each extended byte. NUL termination uses 8 bits. Length-prefix I think can be reasonably termed infinite length as well by using 64 bits. How often you hit the case of your extra bits is the deciding factor. Only 1 extremely large string? Who cares if you're using 8 or 64 bits? Many small strings (Ie Strings of English words)? Then your prefix costs are a large percentage.
Length-prefixed strings allowing time savings is not a real thing. Whether your supplied data is required to have length provided, you're counting at compile time, or you're truly being provided dynamic data that you must encode as a string. These sizes are computed at some point in the algorithm. A separate variable to store the size of a null terminated string can be provided. Which makes the comparison on time-savings moot. One just has an extra NUL at the end... but if the length encode doesn't include that NUL then there's literally no difference between the two. There's no algorithmic change required at all. Just a pre-pass you have to manually design yourself instead of having a compiler/runtime do it for you. C is mostly about doing things manually.
Length-prefix being optional is a selling point. I don't always need that extra info for an algorithm so being required to do it for a every string makes my precompute+compute time never able to drop below O(n). (Ie hardware random number generator 1-128. I can pull from an "infinite string". Let's say it only generates characters so fast. So our string length changes all the time. But my usage of the data probably doesn't care how many random bytes I have. It just wants the next available unused byte as soon as it can get it after a request. I could be waiting on the device. But I could also have a buffer of characters pre-read. A length comparison is a needless waste of computation. A null check is more efficient.)
Length-prefix is a good guard against buffer overflow? So is sane usage of library functions and implementation. What if I pass in malformed data? My buffer is 2 bytes long but I tell the function it's 7! Ex: If gets() was intended to be used on known data it could've had an internal buffer check that tested compiled buffers and malloc() calls and still follow spec. If it was meant to be used as a pipe for unknown STDIN to arrive at unknown buffer then clearly one can't know abut the buffer size which means a length arg is pointless, you need something else here like a canary check. For that matter, you can't length-prefix some streams and inputs, you just can't. Which means the length check has to be built into the algorithm and not a magic part of the typing system. TL;DR NUL-terminated never had to be unsafe, it just ended up that way via misuse.
counter-counter point: NUL-termination is annoying on binary. You either need to do length-prefix here or transform NUL bytes in some way: escape-codes, range remapping, etc... which of course means more-memory-usage/reduced-information/more-operations-per-byte. Length-prefix mostly wins the war here. The only upside to a transform is that no additional functions have to be written to cover the length-prefix strings. Which means on your more optimized sub-O(n) routines you can have them automatically act as their O(n) equivalents without adding more code. Downside is, of course, time/memory/compression waste when used on NUL heavy strings. Depending on how much of your library you end up duplicating to operate on binary data, it may make sense to work solely with length-prefix strings. That said one could also do the same with length-prefix strings... -1 length could mean NUL-terminated and you could use NUL-terminated strings inside length-terminated.
Concat: "O(n+m) vs O(m)" I'm assuming your referring to m as the total length of the string after concatenating because they both have to have that number of operations minimum (you can't just tack-on to string 1, what if you have to realloc?). And I'm assuming n is a mythical amount of operations you no longer have to do because of a pre-compute. If so, then the answer is simple: pre-compute. If you're insisting you'll always have enough memory to not need to realloc and that's the basis of the big-O notation then the answer is even more simple: do binary search on allocated memory for end of string 1, clearly there's a large swatch of infinite zeros after string 1 for us to not worry about realloc. There, easily got n to log(n) and I barely tried. Which if you recall log(n) is essentially only ever as large as 64 on a real computer, which is essentially like saying O(64+m), which is essentially O(m). (And yes that logic has been used in run-time analysis of real data structures in-use today. It's not bullshit off the top of my head.)
Concat()/Len() again: Memoize results. Easy. Turns all computes into pre-computes if possible/necessary. This is an algorithmic decision. It's not an enforced constraint of the language.
String suffix passing is easier/possible with NUL termination. Depending on how length-prefix is implemented it can be destructive on original string and can sometimes not even be possible. Requiring a copy and pass O(n) instead of O(1).
Argument-passing/de-referencing is less for NUL-terminated versus length-prefix. Obviously because you're passing less information. If you don't need length, then this saves a lot of footprint and allows optimizations.
You can cheat. It's really just a pointer. Who says you have to read it as a string? What if you want to read it as a single character or a float? What if you want to do the opposite and read a float as a string? If you're careful you can do this with NUL-termination. You can't do this with length-prefix, it's a data type distinctly different from a pointer typically. You'd most likely have to build a string byte-by-byte and get the length. Of course if you wanted something like an entire float (probably has a NUL inside it) you'd have to read byte-by-byte anyway, but the details are left to you to decide.
TL;DR Are you using binary data? If no, then NUL-termination allows more algorithmic freedom. If yes, then code quantity vs speed/memory/compression is your main concern. A blend of the two approaches or memoization might be best.
Many design decisions surrounding C stem from the fact that when it was originally implemented, parameter passing was somewhat expensive. Given a choice between e.g.
void add_element_to_next(arr, offset)
char[] arr;
int offset;
{
arr[offset] += arr[offset+1];
}
char array[40];
void test()
{
for (i=0; i<39; i++)
add_element_to_next(array, i);
}
versus
void add_element_to_next(ptr)
char *p;
{
p[0]+=p[1];
}
char array[40];
void test()
{
int i;
for (i=0; i<39; i++)
add_element_to_next(arr+i);
}
the latter would have been slightly cheaper (and thus preferred) since it only required passing one parameter rather than two. If the method being called didn't need to know the base address of the array nor the index within it, passing a single pointer combining the two would be cheaper than passing the values separately.
While there are many reasonable ways in which C could have encoded string lengths, the approaches that had been invented up to that time would have all required functions that should be able to work with part of a string to accept the base address of the string and the desired index as two separate parameters. Using zero-byte termination made it possible to avoid that requirement. Although other approaches would be better with today's machines (modern compilers often pass parameters in registers, and memcpy can be optimized in ways strcpy()-equivalents cannot) enough production code uses zero-byte terminated strings that it's hard to change to anything else.
PS--In exchange for a slight speed penalty on some operations, and a tiny bit of extra overhead on longer strings, it would have been possible to have methods that work with strings accept pointers directly to strings, bounds-checked string buffers, or data structures identifying substrings of another string. A function like "strcat" would have looked something like [modern syntax]
void strcat(unsigned char *dest, unsigned char *src)
{
struct STRING_INFO d,s;
str_size_t copy_length;
get_string_info(&d, dest);
get_string_info(&s, src);
if (d.si_buff_size > d.si_length) // Destination is resizable buffer
{
copy_length = d.si_buff_size - d.si_length;
if (s.src_length < copy_length)
copy_length = s.src_length;
memcpy(d.buff + d.si_length, s.buff, copy_length);
d.si_length += copy_length;
update_string_length(&d);
}
}
A little bigger than the K&R strcat method, but it would support bounds-checking, which the K&R method doesn't. Further, unlike the current method, it would be possible to easily concatenate an arbitrary substring, e.g.
/* Concatenate 10th through 24th characters from src to dest */
void catpart(unsigned char *dest, unsigned char *src)
{
struct SUBSTRING_INFO *inf;
src = temp_substring(&inf, src, 10, 24);
strcat(dest, src);
}
Note that the lifetime of the string returned by temp_substring would be limited by those of s and src, which ever was shorter (which is why the method requires inf to be passed in--if it was local, it would die when the method returned).
In terms of memory cost, strings and buffers up to 64 bytes would have one byte of overhead (same as zero-terminated strings); longer strings would have slightly more (whether one allowed amounts of overhead between two bytes and the maximum required would be a time/space tradeoff). A special value of the length/mode byte would be used to indicate that a string function was given a structure containing a flag byte, a pointer, and a buffer length (which could then index arbitrarily into any other string).
Of course, K&R didn't implement any such thing, but that's most likely because they didn't want to spend much effort on string handling--an area where even today many languages seem rather anemic.
According to Joel Spolsky in this blog post,
It's because the PDP-7 microprocessor, on which UNIX and the C programming language were invented, had an ASCIZ string type. ASCIZ meant "ASCII with a Z (zero) at the end."
After seeing all the other answers here, I'm convinced that even if this is true, it's only part of the reason for C having null-terminated "strings". That post is quite illuminating as to how simple things like strings can actually be quite hard.
I don't buy the "C has no string" answer. True, C does not support built-in higher-level types but you can still represent data-structures in C and that's what a string is. The fact a string is just a pointer in C does not mean the first N bytes cannot take on special meaning as a the length.
Windows/COM developers will be very familiar with the BSTR type which is exactly like this - a length-prefixed C string where the actual character data starts not at byte 0.
So it seems that the decision to use null-termination is simply what people preferred, not a necessity of the language.
One advantage of NUL-termination over length-prefixing, which I have not seen anyone mention, is the simplicity of string comparison. Consider the comparison standard which returns a signed result for less-than, equal, or greater-than. For length-prefixing the algorithm has to be something along the following lines:
Compare the two lengths; record the smaller, and note if they are equal (this last step might be deferred to step 3).
Scan the two character sequences, subtracting characters at matching indices (or use a dual pointer scan). Stop either when the difference is nonzero, returning the difference, or when the number of characters scanned is equal to the smaller length.
When the smaller length is reached, one string is a prefix of the other. Return negative or positive value according to which is shorter, or zero if of equal length.
Contrast this with the NUL-termination algorithm:
Scan the two character sequences, subtracting characters at matching indices [note that this is handled better with moving pointers]. Stop when the difference is nonzero, returning the difference. NOTE: If one string is a PROPER prefix of the other, one of the characters in the subtraction will be NUL, i.e zero, and the comparison will naturally stop there.
If the difference is zero, -only then- check if either character is NUL. If so, return zero, otherwise continue to next character.
The NUL-terminated case is simpler, and very easy to implement efficiently with a dual pointer scan. The length-prefixed case does at least as much work, nearly always more. If your algorithm has to do a lot of string comparisons [e.g a compiler!], the NUL-terminated case wins out. Nowadays that might not be as important, but back in the day, heck yeah.
gcc accept the codes below:
char s[4] = "abcd";
and it's ok if we treat is as an array of chars but not string. That is, we can access it with s[0], s[1], s[2], and s[3], or even with memcpy(dest, s, 4). But we'll get messy characters when we trying with puts(s), or worse with strcpy(dest, s).
I think the better question is why you think C owes you anything? C was designed to give you what you need, nothing more. You need to loose the mentality that the language must provide you with everything. Or just continue to use your higher level languages that will give you the luxary of String, Calendar, Containers; and in the case of Java you get one thing in tonnes of variety. Multiple types String, multiple types of unordered_map(s).
Too bad for you, this was not the purpose of C. C was not designed to be a bloated language that offers from a pin to an anchor. Instead you must rely on third party libraries or your own. And there is nothing easier than creating a simple struct that will contain a string and its size.
struct String
{
const char *s;
size_t len;
};
You know what the problem is with this though. It is not standard. Another language might decide to organize the len before the string. Another language might decide to use a pointer to end instead. Another might decide to use six pointers to make the String more efficient. However a null terminated string is the most standard format for a string; which you can use to interface with any language. Even Java JNI uses null terminated strings.
Lastly, it is a common saying; the right data structure for the task. If you find that need to know the size of a string more than anything else; well use a string structure that allows you to do that optimally. But don't make claims that that operation is used more than anything else for everybody. Like, why is knowing the size of a string more important than reading its contents. I find that reading the contents of a string is what I mostly do, so I use null terminated strings instead of std::string; which saves me 5 pointers on a GCC compiler. If I can even save 2 pointers that is good.

Resources