Use string views instead of passing std:wstring by const&
giodicanio.com
giodicanio.com
There are then questions about the length prefix, with a simple solution: make this a platform-specific detail and use the machine word. 16-bit platforms get strings of length ~2^16, 32 b platforms get 2^32 (which is a 4 GB-long string, which is more than 1000× as long as the entire Lord of the Rings trilogy), 64 b platforms get 2^64 (which is ~10^19).
Edit: I think a lot of commenters are focusing on the 'Pascalness' of Pascal strings, which I was using as an umbrella terminology for length-prefixed strings.
And yes, today it would be easier to make a copy of the data... but remember we are talking about 90's, where RAM is measured in megabytes and your L1 cache may be only 8KB or so.
This is a special-case optimisation that I'm happy to lose in favour of the massive performance and security benefits otherwise.
Isn't length + pointer... Basically a Pascal string? Unless I am mistaken.
I think what was unsaid in your second point is that we really need to type-differentiate constant strings, dynamic strings, and string 'views', which Rust does in-language, and C++ does with the standard library. I prefer Rust's approach.
You cannot cheaply take an arbitrary view of the interior string - you can only truncate cheaply (and oob checks are easier to automate). That’s why pointer + length is important because it’s a generic view. For arrays it’s more complicated because you can have a stride which is important for multidimensional arrays.
Length + pointer is a record string, a pascal string has the length at the head of the buffer, behind the pointer.
The main difference is that if a string's length is by its data, you can't easily construct a pointer to part of that data without copying it into a new string, whereas if instead the length is by the data's address, you can cheaply construct pointers to any substring (by coming up with new length+address pairs) without having to construct entire new strings.
That’s not cache-friendly, though. I think the short string optimization (keeping short strings alongside the string length, but allocating a separate buffer for longer strings. See https://devblogs.microsoft.com/oldnewthing/20240510-00/?p=10... for how various C++ compilers implement that) may be the best option.
How so? The string implementations in that post are pretty much that:
struct string
{
char* ptr;
size_t size;
union {
size_t capacity;
char buf[16];
};
Only for small strings. Locally allocated and short-lived aren’t required for short string optimization to take an effect.
Also, I can’t find a good reference, but “only for small strings” in many programs means “for most strings”.
Plus, C-style strings allow a lot of optimizations - if you have a mutable buffer with data, you can make a string out of them with zero copy and zero allocations. strtok(3) is an example of such approach, but I've implemented plenty of similar parsers back in the day. INI, CSV, JSON, XML - query file size, allocate buffer once, read it into the buffer, drop some NULL's into strategic positions, maybe shuffle some bytes around for that rare escape case, and you have a whole bunch of C strings, ready to use, and with no length limits.
Compared to this, Pascal strings would be incredibly painful to use... So you query file size, allocate, read it, and then what? 1-byte length is too short, and for 2+ byte length, you need a secondary buffer to copy string to. And how big should this buffer be? Are you going to be dynamically resizing it or wasting some space?
And sure, _today_ I no longer write code like that, I don't mind dropping std::string into my code, it'd just a meg or so of libraries and 3x overhead for short strings - but that's nothing those days. But back when those conventions were established, it was really really important.
I have also done this, but I would argue that, even at the time, the design was very poor. A much much better solution would have been wise pointers — pass around the length of the string separately from the pointer, much like string_view or Rust’s &str. Then you could skip the NULL-writing part.
Maybe C strings made sense on even older machines which had severely limited registers —- if you have an accumulator and one resister usable as a pointer, you want to minimize the number of variables involved in a computation.
This is a red herring, because when you actually read the strings out, you still need to iterate through the length for each string—zero copy, zero allocation, but linear complexity.
> query file size, allocate buffer once, read it into the buffer, drop some NULL's into strategic positions, maybe shuffle some bytes around for that rare escape case, and you have a whole bunch of C strings, ready to use, and with no length limits.
I write parsers in a very different way—I keep the file buffer around as read-only until the end of the pipeline, prepare string views into the buffer, and pipe those along to the next step.
All I saw were 386's and 486's, and I am pretty sure every piece of software I ever used was either C or Turbo Pascal or direct assembly. In the mid-90s, Java appeared and I remember how horribly slow those Java apps were compared to C/Pascal code.
Remember, code takes up memory too.
So _today_ I write parsers in a very different way as well, copying strings is very cheap (today) and not worth it extra complexity.
But remember we are talking about the past, when those conventions are being established. And back in the 90's, zero copy and zero allocations were real advantage. Not in the theoretical CS sense, but in very practical - remember there was _no_ "dynamically resizing vector" in C's (or Pascal's) stdlib, it's just raw malloc() and realloc(), and it is up to you to assemble vector from it as needed. And free()/malloc() overhead was non-trivial, you had to re-use and grow the buffer as needed. And you want to store the parsed data, storing separate length would double your index size! So a parse-in-place + null-terminated strings approach would give you both smaller code and smaller runtime, at the expense of a few sharp corners. But we were all running with scissors back then.
We're just going to ignore Amigas, and any Unix workstations?
From strtok man page... "The first time that strtok() is called, str should be specified; subsequent calls, wishing to obtain further tokens from the same string, should pass a null pointer instead."
Really?? a null pointer.. This is valid code:
char str[] = "C is fucking weird, ok? I said it, sue me.";
char *result = strtok(str, ",");
char *res = strtok(NULL, ",");
The null pointer in place of the operand here just seemed like a really good quirk to point out
All of this is designed before C11, which means that hilariously it's actually always Undefined Behaviour to write multi-threaded code in C. There's no memory ordering rules yet in the language, and if you write a data race (how could you not in multi-threaded code) then the Sequentially Consistent if Data Race Free proof, SC/DRF does not apply and in C all bets are off if you lose Sequential Consistency† So in this world that's enough, absolute mastery and a single individual keeping track of everything. Does it work? Not very well but hey, it was cheap.
† This is common and you should assume you're fucked in any concurrent language which doesn't say otherwise. In safe Rust you can't write a data race so you're always SC, in Java losing SC is actually guaranteed safe (you probably no longer understand your program, but it does have a meaning and you could reason about it) but in many languages which say nothing it's game over because it was game over in C and they can't do better.
C++ strings had no choice but to copy to underlying string because of this unknown ownership and then added more ownership issues by letting you call the naked pointer within to pass it to C functions. In fact, that's an issue with pretty much every C++ container, including the smart pointers: you can just call get() an break out of the lifecycle management in unpredictable ways.
string_view came much later onto the scene and doesn't have ownership so you avoid a sometimes unnecessary copy but honestly it just makes things more complex.
I honestly think that as long as we continue to use C/C++ for crucial software and operating systems, we'll be dealing with buffer overflow CVEs until the end of time.
Today Bjarne will insist that the correct understanding of ownership and lifetimes was always inherent in his language and if you point out that it basically only warrants a brief mention in his early books about the language he'll say that the newer books have much more about this, as if that's not a confession...
Because &str (the non-owning reference to a slice of UTF-8 encoded text) in Rust is so ubiquitous, it's completely reasonable in Rust to use anything from the simple standard library owning String type, which is literally a growable array, Vec<u8> plus the promise about UTF-8 encoding, through to text which lives only on the stack as a re-interpreted array, or at the other extreme a raw pointer-sized short-string optimisation https://crates.io/crates/cold-string -- where unless it fits inline the string is length-prefixed on the heap like Pascal. All these approaches fit different niches and since a mere reference to the string is the same for all of them they're compatible. In C++ you will run into too many problems and regret trying.
Edited: Correct that String lives in the stdlib -- it isn't baked in to the core language because your tiny embedded system may not have an allocator to go around creating fancy growable arrays, but it still might want the non-owning string slice, &str which truly is in the core language.
[1]: https://learn.microsoft.com/en-gb/windows/apps/design/global...
It's possible to define your own string_view workalike that has a c_str() and binds to whatever is stringlike can has a c_str. It's a few hundred lines of code. You don't have to live with the double indirection.