10

qsort_r() is the re-entrant version of qsort() which takes an additional 'thunk' argument and passes it into the compare function and I'd like to be able to use it in portable C code. qsort() is POSIX and everywhere but qsort_r() seems to be a BSD extension. As a specific question, does this exist or have an equivalent in the Windows C runtime?

dajobe
  • 4,938
  • 35
  • 41

3 Answers3

11

I've attempted to write a portable version of qsort_r / qsort_s (called sort_r) shown with an example. I've also put this code in a git repo (https://github.com/noporpoise/sort_r)

struct sort_r_data
{
  void *arg;
  int (*compar)(const void *a1, const void *a2, void *aarg);
};

int sort_r_arg_swap(void *s, const void *aa, const void *bb)
{
  struct sort_r_data *ss = (struct sort_r_data*)s;
  return (ss->compar)(aa, bb, ss->arg);
}

void sort_r(void *base, size_t nel, size_t width,
            int (*compar)(const void *a1, const void *a2, void *aarg), void *arg)
{
  #if (defined _GNU_SOURCE || defined __GNU__ || defined __linux__)

    qsort_r(base, nel, width, compar, arg);

  #elif (defined __APPLE__ || defined __MACH__ || defined __DARWIN__ || \
         defined __FREEBSD__ || defined __BSD__ || \
         defined OpenBSD3_1 || defined OpenBSD3_9)

    struct sort_r_data tmp;
    tmp.arg = arg;
    tmp.compar = compar;
    qsort_r(base, nel, width, &tmp, &sort_r_arg_swap);

  #elif (defined _WIN32 || defined _WIN64 || defined __WINDOWS__)

    struct sort_r_data tmp = {arg, compar};
    qsort_s(*base, nel, width, &sort_r_arg_swap, &tmp);

  #else
    #error Cannot detect operating system
  #endif
}

Example usage:

#include <stdio.h>

/* comparison function to sort an array of int, inverting a given region
   `arg` should be of type int[2], with the elements
   representing the start and end of the region to invert (inclusive) */
int sort_r_cmp(const void *aa, const void *bb, void *arg)
{
  const int *a = aa, *b = bb, *p = arg;
  int cmp = *a - *b;
  int inv_start = p[0], inv_end = p[1];
  char norm = (*a < inv_start || *a > inv_end || *b < inv_start || *b > inv_end);
  return norm ? cmp : -cmp;
}

int main()
{
  /* sort 1..19, 30..20, 30..100 */
  int arr[18] = {1, 5, 28, 4, 3, 2, 10, 20, 18, 25, 21, 29, 34, 35, 14, 100, 27, 19};
  /* Region to invert: 20-30 (inclusive) */
  int p[] = {20, 30};
  sort_r(arr, 18, sizeof(int), sort_r_cmp, p);

  int i;
  for(i = 0; i < 18; i++) printf(" %i", arr[i]);
  printf("\n");
}

Compile/run/output:

$ gcc -Wall -Wextra -pedantic -o sort_r sort_r.c
$ ./sort_r
 1 2 3 4 5 10 14 18 19 29 28 27 25 21 20 34 35 100

I've tested on mac & linux. Please update this code if you spot mistakes / improvement. You are free to use this code as you wish.

Isaac Turner
  • 2,673
  • 1
  • 26
  • 25
  • There is a bug: in my Clang 11 + macOS 10.15 environment, the `if (defined _GNU_SOURCE` is true and takes precedence over the `#elif (defined __APPLE__`, so `qsort_r` is called with the args in the wrong order. Reordering the `#if` branches so that Apple/BSD detection happens first fixes this. – chris Feb 24 '21 at 14:10
9

For Windows you would use qsort_s: http://msdn.microsoft.com/en-us/library/4xc60xas(VS.80).aspx

Apparently there is some controversy about BSD and GNU having incompatible versions of qsort_r, so be careful about using it in production code: http://sourceware.org/ml/libc-alpha/2008-12/msg00003.html

BTW, the _s stands for "secure" and the _r stands for "re-entrant", but both mean that there's an extra parameter.

Gabe
  • 84,912
  • 12
  • 139
  • 238
5

It's not specified in any portability standard. Also I think it's a mistake to call it a "thread-safe" version of qsort. The standard qsort is thread-safe, but qsort_r effectively allows you to pass a closure as your comparison function.

Obviously in a single-threaded environment, you can achieve the same result with a global variable and qsort, but this usage will not be thread-safe. A different approach to thread-safety would be to use thread-specific data and have your comparison function retrieve its parameter from the thread-specific data (pthread_getspecific with POSIX threads, or __thread variables in gcc and the upcoming C1x standard).

R.. GitHub STOP HELPING ICE
  • 208,859
  • 35
  • 376
  • 711
  • 2
    You're right. It's not thread-safe, it's re-entrant. That means you might still need it even in single-threaded environments. – Gabe Nov 29 '10 at 04:57
  • The `qsort` function itself is reentrant, or at least it should be in any sane implementation. The problem is making functions that want to call `qsort` with a comparison function requiring an argument reentrant. – R.. GitHub STOP HELPING ICE Nov 29 '10 at 05:11
  • Yes, obviously the `qsort` algorithm does not require global state, so it's de-facto re-entrant and thread-safe. I just meant that `qsort_r` is intended for use where re-entrancy might be required, and thus using thread-local storage does not always achieve the same result. – Gabe Nov 29 '10 at 05:23
  • You are right, it's re-entrant and in fact I don't care about threading, I just need a compare function with an extra user data parameter that I can use to access other state. – dajobe Nov 29 '10 at 17:40
  • And for Windows, you can also use the thread-local storage API: `TlsGetValue` and friends or the `__declspec(thread)` variable specifier. – Adam Rosenfield Nov 29 '10 at 17:50