Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

is is sometimes incorrect because == does not work for floating-point comparison in C #29

Open
noncombatant opened this issue Jun 9, 2023 · 2 comments

Comments

@noncombatant
Copy link

Thanks for publishing fe! It's very cool. Turning on more compiler warnings (I enjoy strange hobbies) I found this:

// src/fe.c:203:49: warning: comparing floating point with == or != is unsafe [-Wfloat-equal]
//  if (type(a) == FE_TNUMBER) { return number(a) == number(b); }
//                                      ~~~~~~~~~ ^  ~~~~~~~~~

I found that it is indeed a practical problem in the REPL: I have 2 expressions that print equal but for which is returns nil (correctly).

> (= x (/ 10.2 3))
nil
> x
3.4
> (= y (/ 30.6 9))
nil
> y
3.4
> (is x y)
nil

I propose 2 possible fixes in the C file below:

// src/fe.c:203:49: warning: comparing floating point with == or != is unsafe [-Wfloat-equal]
//  if (type(a) == FE_TNUMBER) { return number(a) == number(b); }
//                                      ~~~~~~~~~ ^  ~~~~~~~~~
//
// You can see with Python 3 that there is a tiny bit of floating point error:
//
// >>> 10.2 / 3
// 3.4
// >>> 30.6 / 9
// 3.4000000000000004
// >>> 3 * 3.4
// 10.2
// >>> 9 * 3.4
// 30.599999999999998
// >>> (10.2 / 3) == (30.6 / 9)
// False
//
// In Python, we can deal with this using `math.isclose`.
//
// fe prints them the same (format string `%.7g`), but does equality on their
// bitwise representations. So you end up with an error you can't see in the
// REPL, nor fix using library functions.
//
// % ./fe
// > (= x (/ 10.2 3))
// nil
// > x
// 3.4
// > (= y (/ 30.6 9))
// nil
// > y
// 3.4
// > (is x y)
// nil
//
// Changing fe to use `%.7f` at least makes the error visible in the REPL. I see
// 2 paths to solving this: (1) provide `isclose` (`double_nearly_equal` in the
// C below) in the standard library; or (2) make `is` use `double_nearly_equal`
// for number objects. With either option, using `%.7f` might also help people
// see the issue.

#include <float.h>
#include <math.h>
#include <stdbool.h>
#include <stdio.h>
#include <stdlib.h>

#define MIN(a, b) ((a) < (b) ? (a) : (b))

// Translated from [the original
// Java](https://floating-point-gui.de/errors/comparison/). If you want to use
// `float`, use `fabsf` and `FLT_*`. If you want to use `long double`, use
// `fabsl` and `LDBL_*`.
//
// See also
// https://randomascii.wordpress.com/2012/02/25/comparing-floating-point-numbers-2012-edition/.
static bool double_nearly_equal(double a, double b, double epsilon) {
  const double absA = fabs(a);
  const double absB = fabs(b);
  const double diff = fabs(a - b);
// It's OK to turn this warning off for this limited section of code, because
// it's in the context of handling tiny errors.
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wfloat-equal"
  if (a == b) {
    // Special case; handles infinities.
    return true;
  } else if (a == 0 || b == 0 || (absA + absB < DBL_MIN)) {
#pragma clang diagnostic pop
    // Either a or b is zero, or both are extremely close to it. Relative error
    // is less meaningful here.
    return diff < (epsilon * DBL_MIN);
  }
  // Use relative error.
  return diff / MIN((absA + absB), DBL_MAX) < epsilon;
}

int main(int count, char* arguments[]) {
  if (count != 5) {
    fprintf(stderr, "Usage: ./float_nearly_equal 10.2 3 30.6 9\n");
    return -1;
  }

  const double a = strtod(arguments[1], NULL);
  const double b = strtod(arguments[2], NULL);
  const double c = strtod(arguments[3], NULL);
  const double d = strtod(arguments[4], NULL);
  const double r1 = a / b;
  const double r2 = c / d;
  printf("%.20f / %.20f (%.20f) == %.20f / %.20f (%.20f) : %d\n", a, b, r1, c, d, r2, r1 == r2);
  printf("%.20f / %.20f double_nearly_equal %.20f / %.20f : %d\n", a, b, c, d,
         double_nearly_equal(a / b, c / d, DBL_EPSILON));
}
@ooichu
Copy link

ooichu commented Jun 10, 2023

The comparison of numbers is usually between integers, otherwise you have to account for float-point errors, which is expected behavior.
I think it's better to just implement a isclose like function and register it with fe_cfunc() and fe_set() if you really need it.

@noncombatant
Copy link
Author

Yes, I know that floating-point errors are expected. The trouble is that there is no way to detect or deal with them in terms of fe source code and in the REPL. I am proposing to resolve that in either of 2 ways.

(Regarding integers, the use of single-precision float in fe means that the usable range of integers is only 23 bits, which is fairly small. double provides 53 bits of integers.)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants