compare_double mishandles infinities: equal infinities compare unequal, opposite infinities may compare equal

[yayong3]

compare_double() does not handle non-finite values correctly.

In particular:

• two separately allocated numbers containing +INFINITY compare unequal;
• two separately allocated numbers containing -INFINITY compare unequal;
• +INFINITY and -INFINITY compare equal in cJSON_Compare().

The issue comes from applying the relative-error comparison to infinities without handling non-finite values first.

Affected code

In cJSON.c, compare_double() is currently equivalent to:

static cJSON_bool compare_double(double a, double b)
{
    double maxVal = fabs(a) > fabs(b) ? fabs(a) : fabs(b);
    return (fabs(a - b) <= maxVal * DBL_EPSILON) ? true : false;
}

For infinities this gives unexpected results.

For equal infinities:

a = +Inf
b = +Inf

a - b             = NaN
fabs(a - b)       = NaN
maxVal * epsilon  = Inf

NaN <= Inf        = false

so two distinct +Inf values compare unequal. The same happens for -Inf.

For opposite infinities:

a = +Inf
b = -Inf

a - b             = +Inf
fabs(a - b)       = +Inf
maxVal * epsilon  = +Inf

Inf <= Inf        = true

so +Inf and -Inf compare equal.

cJSON_Compare() uses this function for cJSON_Number, so these results are observable through the public API.

A similar compare_double() implementation also exists in cJSON_Utils.c.

Reproduction

#include <stdio.h>
#include <math.h>

#include "cJSON.h"

int main(void)
{
    cJSON *p1 = cJSON_CreateNumber(INFINITY);
    cJSON *p2 = cJSON_CreateNumber(INFINITY);
    cJSON *n1 = cJSON_CreateNumber(-INFINITY);
    cJSON *n2 = cJSON_CreateNumber(-INFINITY);

    if (!p1 || !p2 || !n1 || !n2)
    {
        return 1;
    }

    printf("+Inf vs +Inf: %d\n", cJSON_Compare(p1, p2, 1));
    printf("-Inf vs -Inf: %d\n", cJSON_Compare(n1, n2, 1));
    printf("+Inf vs -Inf: %d\n", cJSON_Compare(p1, n1, 1));
    printf("-Inf vs +Inf: %d\n", cJSON_Compare(n1, p1, 1));

    cJSON_Delete(p1);
    cJSON_Delete(p2);
    cJSON_Delete(n1);
    cJSON_Delete(n2);

    return 0;
}

Observed result:

+Inf vs +Inf: 0
-Inf vs -Inf: 0
+Inf vs -Inf: 1
-Inf vs +Inf: 1

Expected behavior would be:

+Inf vs +Inf: 1
-Inf vs -Inf: 1
+Inf vs -Inf: 0
-Inf vs +Inf: 0

Why this is reachable

Although JSON itself has no NaN or Infinity literals, non-finite values are reachable through the cJSON C API:

cJSON_CreateNumber(INFINITY);
cJSON_SetNumberValue(item, INFINITY);

Infinity can also arise while parsing a syntactically valid JSON number whose magnitude exceeds the finite range of double, depending on the strtod() implementation, for example:

cJSON_Parse("1e999");

So the comparison code should not assume that every stored double is finite.

Suggested fix

Handle non-finite values before applying the relative-error calculation.

For example, exact equality can be checked first:

static cJSON_bool compare_double(double a, double b)
{
    double maxVal;

    if (a == b)
    {
        return true;
    }

    if (!isfinite(a) || !isfinite(b))
    {
        return false;
    }

    maxVal = fabs(a) > fabs(b) ? fabs(a) : fabs(b);
    return (fabs(a - b) <= maxVal * DBL_EPSILON) ? true : false;
}

This gives the expected infinity behavior while preserving the existing approximate comparison for finite values.

The equivalent implementation in cJSON_Utils.c should be updated as well.

Related NaN-to-integer issue

While reviewing the same non-finite-number paths, cJSON_CreateNumber() / cJSON_SetNumberHelper() also reach:

valueint = (int)number;

for NaN, because both comparisons against INT_MAX and INT_MIN are false for NaN. Converting NaN to an integer type this way is undefined behavior when the value cannot be represented.

However, that issue is already tracked by #999 and addressed by PR #1000, so I am not proposing to duplicate it here.

This issue is specifically about the incorrect infinity comparison behavior in compare_double().