Make 0 divided by 0 results in NaN consistently

[itchyny]

0/0 is folded to nan on parsing but 0 as $x | $x/0 throws a zero-division error so 0/0 is not equal to 0 as $x | $x/0. This is not intuitive. I think 0 as $x | $x/0 should be nan as well as 0/0 is.

 % jq -n '0/0, (0 as $x | $x/0) | isnan'
true
jq: error (at <unknown>): number (0) and number (0) cannot be divided because the divisor is zero

[coveralls]

Coverage increased (+0.09%) to 84.228% when pulling 1199685 on itchyny:fix-zero-division-nan into 80052e5 on stedolan:master.

[wader]

I ran into a related issue today when using //. nan as JSON is null but is treated as true.

$ jq -n 'nan | ., if . then "true" else "false" end'
null
"true"
$ jq -n 'nan | ., (. // "alt")'
null
null

Treating it as false might be even more confusing? so maybe error is better?

[pkoppstein]

nan as JSON is null

That is incorrect. nan prints as null, but that’s it. The only falsey values in jq are null and false. But it is true that nan and infinite are a bit weird :-)

[wader]

nan as JSON is null

That is incorrect. nan prints as null, but that’s it. The only falsey values in jq are null and false. But it is true that nan and infinite are a bit weird :-)

Yes, sorry i was a bit unclear, that was what i meant, null is the current JSON representation of nan for jq. That was what made me confused while i was debugging some jq code, i saw null and it took quite a long while to understand that this was actually not of type-null null. But it's hard to imagine what other JSON value would be better so maybe throw error is best?

[pkoppstein]

@wader - Making a special case for just nan // E would be very strange. Anyway, gojq and jaq both have the same semantics regarding falseyness and nan etc, and so at this point, it’s clear that the cat is out of the bag.

[wader]

Now realized that i read the PR commit message a bit fast, I understood it as making it consistently throw error on division by zero. Yeah agree, it's probably too late to change any of this now.

[itchyny]

Thank you.

[emanuele6]

This code is actually problematic because division by 0 is unspecified for all values in C, so 0.0/0.0 will be undefined behaviour.
If we want to make 0/0 return nan, we should hardcode that case.

diff --git a/src/builtin.c b/src/builtin.c
index 09ffc04..1b4aec2 100644
--- a/src/builtin.c
+++ b/src/builtin.c
@@ -383,8 +383,13 @@ static jv f_multiply(jq_state *jq, jv input, jv a, jv b) {
 static jv f_divide(jq_state *jq, jv input, jv a, jv b) {
   jv_free(input);
   if (jv_get_kind(a) == JV_KIND_NUMBER && jv_get_kind(b) == JV_KIND_NUMBER) {
-    if (jv_number_value(b) == 0.0 && jv_number_value(a) != 0.0)
-      return type_error2(a, b, "cannot be divided because the divisor is zero");
+    if (jv_number_value(b) == 0.0) {
+      if (jv_number_value(a) != 0.0)
+        return type_error2(a, b, "cannot be divided because the divisor is zero");
+      jv_free(a);
+      jv_free(b);
+      return jv_number(NAN);
+    }
     jv r = jv_number(jv_number_value(a) / jv_number_value(b));
     jv_free(a);
     jv_free(b);