Arrays of Arrays (2009)

Arrays of Arrays (2009)(ericlippert.com)

39 points by theawesomekhan 5 years ago | 27 comments

Animats 5 years ago |

Over a decade ago I was trying to get the C++ committee to support multidimensional arrays via overloading, so we could have 2D collections. In C++, you can overload "operator[]". But it can't take more than one argument. Why?

Because of the C "comma operator". In C, the comma operator evaluates both operands and discards the first result. It's an ancient hack used mostly with preprocessor defines.

In C and C++,

    foo(a,b)

is a call with two arguments. But

    foo[a,b]

invokes the comma operator and presents one value to "[]" You have to dig through the C++ specification to find this.

So I asked the question, could anyone find an example of the comma operator used in C or C++ inside square brackets in any production code. I couldn't. I got someone at IBM Almaden to search IBM's code archive and they couldn't. But there was the concern that somewhere, someone had used that misfeature for something.

So, no 2D array syntax for you.

geocar 5 years ago | |

> Over a decade ago I was trying to get the C++ committee to support multidimensional arrays via overloading, so we could have 2D collections.

You might've asked operator[] to take multiple arguments separated by semicolon, so:

    foo[a;b]

which is at least unambiguous. Maybe it would have been less offensive.

> So I asked the question, could anyone find an example of the comma operator used in C or C++ inside square brackets in any production code.

If you're still referring to operator overloading, there's things like Boost.Assign and Boost.Phoenix, but if you're referring to use as a sequence operator, i.e. where a[b,c] could potentially be b,a[c] except for the different sequence point, it doesn't surprise me you had a hard time finding an example: I've only ever seen people like Arthur do that (people writing APL in C)

tbrgmnn 5 years ago | |

With C++20, use of comma expressions as second argument of a subscript operator got deprecated.

gpderetta 5 years ago | | |

... with the intention of allowing multidimensional indexing in the next standard.

Animats 5 years ago | | |

Finally!

Multidimensional array discussions tended to dissolve into bikeshedding. Some people wanted the ability to store an array as either transpose, that is, by row or by column. Then that had to be extended to N dimensions. That added a lot of complexity to support a rare use case. Then the proposal got so complex it was shelved. All this is in old USENET comp.lang.c++, if that hasn't been lost yet.

thesz 5 years ago | | |

And it is awful. See my comment about yalsat's use of asserts() above.

thomasahle 5 years ago | |

> In C, the comma operator evaluates both operands and discards the first result.

Sounds like how the semicolon operator works in modern languages, like Rust.

anonymoushn 5 years ago | | |

It does not, because this code does not compile:

  fn main() {
      println!("ok"; "Hello, world!");
  }

thesz 5 years ago | |

Take a look at yalsat: http://fmv.jku.at/yalsat/

Unpack the "v" version, look at yals.c. You will see there the following:

  #define PUSH(S,E) \
  do { \
    if (FULL(S)) ENLARGE (S); \
    *(S).top++ = (E); \
  } while (0)
  
  #define POP(S) \
    (assert (!EMPTY (S)), *--(S).top)
  
  #define TOP(S) \
    (assert (!EMPTY (S)), (S).top[-1])
  
  #define PEEK(S,P) \
    (assert ((P) < COUNT(S)), (S).start[(P)])
  
  #define POKE(S,P,E) \
    do { assert ((P) < COUNT(S)); (S).start[(P)] = (E); } while (0)

These are definitions for typed dynamic arrays using C macros. Very useful.

As you can see, there's a couple of assert()'s before actual computation of the address to fetch.

I hope I proved you wrong.

dmurray 5 years ago | | |

> I hope I proved you wrong.

I don't see where you did - Animats said he couldn't find examples of the comma operator inside square brackets, not that he couldn't see it being used at all.

sesuximo 5 years ago | |

Imo foo[x][y] is better for many reasons:

- compiler can break it apart so foo[x] is evaluated once for many ys just like any other function

- you can make types that are unaware of their rank eg vector doesn’t know if it’s a 1D vector of int or a 2D vector of vectors

- “seamlessly” handles jagged arrays

jiggawatts 5 years ago | | |

Which reminds me of an insightful observation that the array indexing operator [] and the function call operator () are basically doing the same thing: evaluating a function and returning a value.

So what you've just described is basically Currying, or partial function evaluation. It's a bit like defining "al(i)" as array lookup for an index 'i' instead of "[i]". Then the following are equivalent:

    something[i][j]
    something.al(i).al(j)

What people also want are:

    something[i,j]
    something.al(i,j)

In languages like Haskell, every function can always be evaluated argument-by-argument, with no special effort by the programmer. The compiler takes care of generating the code for the various intermediate call forms. In languages like C++, it would be a significant hassle to create all of the various partially-evaluated wrapper types.

Athas 5 years ago | | |

It's not the same when you support slicing. foo[x:y][z] is not the same as foo[x:y,z]. I'm not sure anyone is proposing slicing to be supported with the indexing operator in C++, though, so it might not matter.

Bancakes 5 years ago | | |

It's semantically the same and the compiler should know that. But I do agree [][] makes more sense. The jth index of the ith array.

mhh__ 5 years ago | | |

They'd probably be lowered to exactly the same code, so the compiler argument isn't very good.

dls2016 5 years ago | |

I just learned this one while teaching an intro to C++ class... the student had no idea what they wrote and it took me a minute to realize why their code even compiled.

Per_Bothner 5 years ago |

Which is why the sane way to do it is to have the array-specifier come before the element type:

  [,][]int crazy;

It reads "2-dimensional array of array of int" in L-to-R order.

Similarly, the type "pointer-to-int" should be "*p" or equivalent.

Unfortunately, history/tradition as well parsing problems make the "sane" solution difficult. One solution is the Pascal approach putting the type after the variable being declared, with a colon between:

  crazy : [,][]int;

Athas 5 years ago | |

Strongly agreed! During my PhD I had to design array type syntax for a programming language, and we eventually came up with something similar to that: https://futhark-lang.org/blog/2016-12-09-two-syntax-design-p...

teleforce 5 years ago |

Take a look at D language library Mir for N-Dimensional array and its comparison to Numpy [1].

Its native performance inside GLAS implementation is something to shout about and even comparable to OpenBLAS that is currently being used by Julia and Matlab [2][3].

[1]http://docs.algorithm.dlang.io/latest/mir_ndslice.html

[2]http://blog.mir.dlang.io/glas/benchmark/openblas/2016/09/23/...

[3]http://docs.mir.dlang.io/latest/mir_glas.html

kqr 5 years ago |

This reasoning seems to me similar to how declaring a pointer in e.g. C works: many beginners write

    int* a, b;

And think that this constructs two pointers-to-integer. In fact, it constructs one pointer and one plain integer. A better (equivalent) way to write it is

    int *a, b;

You're not declaring variables of type

   int*

You're saying that b is an int and

*a

Is an int, where the asterisk means "the dereferenced a"!

pwdisswordfish4 5 years ago | |

Beginners and also C++ programmers, expert and beginner alike. Even though it’s just as misleading in C++. This:

    int a = 0;
    int& x = a, y = a;

declares x to be a reference, but y an independent variable. Similarly, a template parametrised by a variadic pack of function pointers is declared thus:

    template <void (*...x)(void)>

It’s a good thing I don’t work in C++ too often, because my forehead would have probably sustained noticeable injury from all the facepalming each time I see ‘int& x’ instead of ‘int &x’ and ‘typename... T’ instead of ‘typename ...T’.

xeromal 5 years ago |

That was a fun little exercise

#define PUSH(S,E) \ do { \ if (FULL(S)) ENLARGE (S); \ *(S).top++ = (E); \ } while (0) #define POP(S) \ (assert (!EMPTY (S)), *--(S).top) #define TOP(S) \ (assert (!EMPTY (S)), (S).top[-1]) #define PEEK(S,P) \ (assert ((P) < COUNT(S)), (S).start[(P)]) #define POKE(S,P,E) \ do { assert ((P) < COUNT(S)); (S).start[(P)] = (E); } while (0)