Array#sort() inconsistent implementations (support for booleans)

[Reinmar]

Hi,

I'm not sure that this is the right place to report this and I'm quite sure it must've been discussed in the past, but let's see.

I found out today that:

[ 1, 4, 2, 6, 0, 6, 2, 6 ].sort( ( a, b ) => a > b )

gives a sorted array in Chrome and Firefox and does nothing in Edge and Safari.

Obviously, sort()'s callback is meant to return a number, but, apparently, two engines also support booleans.

Does it make Chrome's and Firefox's behaviour a bug? Would it be worth (and, first of all, possible) to align Chrome and Firefox to the spec?

Or, since that would be a backwards incompatible change and taken that sort() doesn't need to implement a stable sorting algorithm anyway, a support for booleans could be standardised?

[mathiasbynens]

A comparefn that does not return numeric values is not a consistent comparison function. In that case, the sort order is implementation-defined.

From https://tc39.github.io/ecma262/#sec-array.prototype.sort:

If comparefn is not undefined and is not a consistent comparison function for the elements of this array (see below), the sort order is implementation-defined.

And below:

A function comparefn is a consistent comparison function for a set of values S if all of the requirements below are met […]
[…]
Furthermore, Type(v) is Number

You could file bugs trying to get the engines to agree on their behavior, after which the behavior could potentially be specified.

[mathiasbynens]

A comparefn that does not return numeric values is not a consistent comparison function. In that case, the sort order is implementation-defined.

From https://tc39.github.io/ecma262/#sec-array.prototype.sort:

If comparefn is not undefined and is not a consistent comparison function for the elements of this array (see below), the sort order is implementation-defined.

And below:

A function comparefn is a consistent comparison function for a set of values S if all of the requirements below are met […]
[…]
Furthermore, Type(v) is Number

You could file bugs trying to get the engines to agree on their behavior, after which the behavior could potentially be specified.

[Reinmar]

Thanks. I haven't even thought that Chrome and Firefox work within the spec's limit. I guess I could open tickets for Edge and Safari so they add support for booleans but I don't think that this has a big chances of passing.

[Reinmar]

Thanks. I haven't even thought that Chrome and Firefox work within the spec's limit. I guess I could open tickets for Edge and Safari so they add support for booleans but I don't think that this has a big chances of passing.

[claudepache]

The comparison function is supposed to return a number, and if it doesn’t, the result is coerced to a number: see step 4.a of SortCompare abstract operation. In particular, Boolean values will be coerced to 0 or 1. (Historical note: that coercion behaviour was added in ES6, because it was observed that it is what the implementations did.)

In the context of Array#sort, a consistent comparison function is a comparison function that models correctly an order relation according to which the set of values will be sorted, as explained in more details at the end of Section Array.prototype.sort. A comparison function that produces 0 or 1 but never -1 is certainly not consistent.

[claudepache]

The comparison function is supposed to return a number, and if it doesn’t, the result is coerced to a number: see step 4.a of SortCompare abstract operation. In particular, Boolean values will be coerced to 0 or 1. (Historical note: that coercion behaviour was added in ES6, because it was observed that it is what the implementations did.)

In the context of Array#sort, a consistent comparison function is a comparison function that models correctly an order relation according to which the set of values will be sorted, as explained in more details at the end of Section Array.prototype.sort. A comparison function that produces 0 or 1 but never -1 is certainly not consistent.

[ljharb]

@Reinmar what you probably want is [1, 4, 2, 6, 0, 6, 2, 6].sort((a, b) => a - b) (or b - a). For strings, a comparator can instead return a.localeCompare(b).

[ljharb]

@Reinmar what you probably want is [1, 4, 2, 6, 0, 6, 2, 6].sort((a, b) => a - b) (or b - a). For strings, a comparator can instead return a.localeCompare(b).

[erights]

What does the spec say should happen with: [1, NaN, 3].sort((a, b) => a - b) ?

…

On Thu, Apr 20, 2017 at 4:04 PM, Jordan Harband ***@***.***> wrote: @Reinmar <https://github.com/Reinmar> what you probably want is [1, 4, 2, 6, 0, 6, 2, 6].sort((a, b) => a - b) (or b - a). For strings, a comparator can instead return a.localeCompare(b).

[erights]

What does the spec say should happen with: [1, NaN, 3].sort((a, b) => a - b) ?

…

On Thu, Apr 20, 2017 at 4:04 PM, Jordan Harband ***@***.***> wrote: @Reinmar <https://github.com/Reinmar> what you probably want is [1, 4, 2, 6, 0, 6, 2, 6].sort((a, b) => a - b) (or b - a). For strings, a comparator can instead return a.localeCompare(b).

[allenwb]

Unspecified because the argument is not a "consistent comparison function"

…

On Apr 20, 2017, 4:26 PM, at 4:26 PM, "Mark S. Miller" ***@***.***> wrote: What does the spec say should happen with: [1, NaN, 3].sort((a, b) => a - b) ? On Thu, Apr 20, 2017 at 4:04 PM, Jordan Harband ***@***.***> wrote: > @Reinmar <https://github.com/Reinmar> what you probably want is [1, 4, 2, > 6, 0, 6, 2, 6].sort((a, b) => a - b) (or b - a). For strings, a > comparator can instead return a.localeCompare(b). > > — > You are receiving this because you are subscribed to this thread. > Reply to this email directly, view it on GitHub > <#902 (comment)>, or mute > the thread > <https://github.com/notifications/unsubscribe-auth/AAQtzCAHqZ2KokUJfEa5KfE88Qec0feRks5rx-SJgaJpZM4NC2o9> > . > -- Cheers, --MarkM -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: #902 (comment)

[allenwb]

Unspecified because the argument is not a "consistent comparison function"

…

On Apr 20, 2017, 4:26 PM, at 4:26 PM, "Mark S. Miller" ***@***.***> wrote: What does the spec say should happen with: [1, NaN, 3].sort((a, b) => a - b) ? On Thu, Apr 20, 2017 at 4:04 PM, Jordan Harband ***@***.***> wrote: > @Reinmar <https://github.com/Reinmar> what you probably want is [1, 4, 2, > 6, 0, 6, 2, 6].sort((a, b) => a - b) (or b - a). For strings, a > comparator can instead return a.localeCompare(b). > > — > You are receiving this because you are subscribed to this thread. > Reply to this email directly, view it on GitHub > <#902 (comment)>, or mute > the thread > <https://github.com/notifications/unsubscribe-auth/AAQtzCAHqZ2KokUJfEa5KfE88Qec0feRks5rx-SJgaJpZM4NC2o9> > . > -- Cheers, --MarkM -- You are receiving this because you are subscribed to this thread. Reply to this email directly or view it on GitHub: #902 (comment)

[ljharb]

Despite that, the steps for comparison functions (see https://tc39.github.io/ecma262/#sec-sortcompare step 4.b) treats it as +0, even though the consistency requirements require that it not return NaN.

[ljharb]

Despite that, the steps for comparison functions (see https://tc39.github.io/ecma262/#sec-sortcompare step 4.b) treats it as +0, even though the consistency requirements require that it not return NaN.

[erights]

Interesting. I didn't know about this conversion to +0. But even if we allowed comparefn to return NaN and treated it as if it had returned +0, (a, b) => a - b would still not be transitive, and so still not a consistent comparison function.

No action item here re NaN. I can't think of any realistic way to further improve the situation.

[erights]

Interesting. I didn't know about this conversion to +0. But even if we allowed comparefn to return NaN and treated it as if it had returned +0, (a, b) => a - b would still not be transitive, and so still not a consistent comparison function.

No action item here re NaN. I can't think of any realistic way to further improve the situation.

[claudepache]

Despite that, the steps for comparison functions (see https://tc39.github.io/ecma262/#sec-sortcompare step 4.b) treats it as +0, even though the consistency requirements require that it not return NaN.

It is probably a bug that the consistency requirements still require a non-NaN Number when the value is coerced to a non-NaN number in SortCompare anyway.

Or, in other words, the consistency conditions should apply to SortCompare(a, b) rather than comparefn(a, b).

[claudepache]

Despite that, the steps for comparison functions (see https://tc39.github.io/ecma262/#sec-sortcompare step 4.b) treats it as +0, even though the consistency requirements require that it not return NaN.

It is probably a bug that the consistency requirements still require a non-NaN Number when the value is coerced to a non-NaN number in SortCompare anyway.

Or, in other words, the consistency conditions should apply to SortCompare(a, b) rather than comparefn(a, b).

[ljharb]

I agree; it seems that either NaN should be allowed in consistent comparison functions, or SortCompare shouldn't need to mention it.

[ljharb]

I agree; it seems that either NaN should be allowed in consistent comparison functions, or SortCompare shouldn't need to mention it.

[Reinmar]

Webkit aligned its implementation to that of Firefox and Chrome :) https://bugs.webkit.org/show_bug.cgi?id=47825

[Reinmar]

Webkit aligned its implementation to that of Firefox and Chrome :) https://bugs.webkit.org/show_bug.cgi?id=47825

[claudepache]

Webkit aligned its implementation to that of Firefox and Chrome :) https://bugs.webkit.org/show_bug.cgi?id=47825

And what is the semantics of Firefox and Chrome?

Their patch seems to special-case boolean results of the comparison function, but I’m not sure it is what Firefox or Chrome do. Indeed, consider the following testcases, where the comparison function does not return a Boolean:

[1,3,2,5,4].sort((a,b) => Number(b<a)) // [ 1, 2, 3, 4, 5 ] in both FF in Chrome
[1,3,2,5,4].sort((a,b) => Number(a<b))  // [ 5, 4, 3, 2, 1 ] in both FF in Chrome

A reasonable semantics to me would be the following: If compareFn is the comparison function and if ToNumber(compareFn(a, b)) evaluates to 0 (or NaN), evaluates ToNumber(compareFn(b, a)) in order to decide how a and b should be ordered. But what they really do, I don’t know: I’ve not read the source code of FF or Chrome.

[claudepache]

Webkit aligned its implementation to that of Firefox and Chrome :) https://bugs.webkit.org/show_bug.cgi?id=47825

And what is the semantics of Firefox and Chrome?

Their patch seems to special-case boolean results of the comparison function, but I’m not sure it is what Firefox or Chrome do. Indeed, consider the following testcases, where the comparison function does not return a Boolean:

[1,3,2,5,4].sort((a,b) => Number(b<a)) // [ 1, 2, 3, 4, 5 ] in both FF in Chrome
[1,3,2,5,4].sort((a,b) => Number(a<b))  // [ 5, 4, 3, 2, 1 ] in both FF in Chrome

A reasonable semantics to me would be the following: If compareFn is the comparison function and if ToNumber(compareFn(a, b)) evaluates to 0 (or NaN), evaluates ToNumber(compareFn(b, a)) in order to decide how a and b should be ordered. But what they really do, I don’t know: I’ve not read the source code of FF or Chrome.

[claudepache]

Other testcases (same results in both FF and Chrome):

[1,3,2,5,4].sort((a,b) => 1.5*Number(b<a)) // [ 1, 2, 3, 4, 5 ]
[1,3,2,5,4].sort((a,b) => 1.5*Number(a<b))  // [ 5, 4, 3, 2, 1 ]

but (showing that they do not follow the “reasonable” semantics of my previous comment):

[1,3,2,5,4].sort((a,b) => -Number(b<a)) // [ 1, 3, 2, 5, 4 ] (not sorted)
[1,3,2,5,4].sort((a,b) => -Number(a<b)) // [ 1, 3, 2, 5, 4 ] (not sorted)

[claudepache]

Other testcases (same results in both FF and Chrome):

[1,3,2,5,4].sort((a,b) => 1.5*Number(b<a)) // [ 1, 2, 3, 4, 5 ]
[1,3,2,5,4].sort((a,b) => 1.5*Number(a<b))  // [ 5, 4, 3, 2, 1 ]

but (showing that they do not follow the “reasonable” semantics of my previous comment):

[1,3,2,5,4].sort((a,b) => -Number(b<a)) // [ 1, 3, 2, 5, 4 ] (not sorted)
[1,3,2,5,4].sort((a,b) => -Number(a<b)) // [ 1, 3, 2, 5, 4 ] (not sorted)

[Yaffle]

if some implementation uses
comparefn(a, b) < 0 internally, (a > b ? 1 : 0) < 0 gives false and probably the result will not be sorted.

Will it make implementations consistent if spec, that implementation should use > 0 for the result of comparefn(a, b) - comparefn(a, b) > 0 ?

[Yaffle]

if some implementation uses
comparefn(a, b) < 0 internally, (a > b ? 1 : 0) < 0 gives false and probably the result will not be sorted.

Will it make implementations consistent if spec, that implementation should use > 0 for the result of comparefn(a, b) - comparefn(a, b) > 0 ?

[claudepache]

Testcase:

function test(numValues, numDistinctValues, compareFn) {
    var arr = Array(numValues)
    for (let i = 0; i < numValues; i++) {
        arr[i] = Math.floor(Math.random()*numDistinctValues)
    }
    arr.sort(compareFn)
    for (let i = 1; i < numValues; i++) {
        if (arr[i] < arr[i - 1])
            return false
    }
    return true
}

Naturally, test(n, m, (a,b) => a - b) must return true for every n and m.

For test(n, m, (a,b) => a - b > 0), I get the following results:

• Firefox: always true;
• Chrome: true for n ≤ 10, usually false otherwise;
• Edge and Safari: usually false.

[claudepache]

Testcase:

function test(numValues, numDistinctValues, compareFn) {
    var arr = Array(numValues)
    for (let i = 0; i < numValues; i++) {
        arr[i] = Math.floor(Math.random()*numDistinctValues)
    }
    arr.sort(compareFn)
    for (let i = 1; i < numValues; i++) {
        if (arr[i] < arr[i - 1])
            return false
    }
    return true
}

Naturally, test(n, m, (a,b) => a - b) must return true for every n and m.

For test(n, m, (a,b) => a - b > 0), I get the following results:

• Firefox: always true;
• Chrome: true for n ≤ 10, usually false otherwise;
• Edge and Safari: usually false.

[claudepache]

While we’re at it, let’s test the stability of the sort algorithm (https://esdiscuss.org/topic/stable-sort-proposal):

function test(numValues, numDistinctValues, compareFn) {
    "use strict"
    
    let arr = Array(numValues)
    
    for (let i = 0; i < numValues; i++) {
        arr[i] = [ i, Math.floor(Math.random()*numDistinctValues) ]
    }
    
    arr.sort((a, b) => compareFn(a[1], b[1]))
    
    let correct = true
    let stable = true
    for (let i = 1; i < numValues; i++) {
        if (arr[i][1] < arr[i-1][1])
            correct = false
        if (arr[i][1] === arr[i-1][1] && arr[i][0] < arr[i-1][0])
            stable = false
    }
    return [ correct, stable ]
}

(To be tested with numDistinctValues small relatively to numValues for meaningful result about stability.)

Conclusion:

• Firefox and Safari always use a stable sort;
• Edge uses a stable sort for n ≤ 512 and an unstable sort for larger arrays;
• Chrome uses a stable sort for n ≤ 10 and an unstable sort for larger arrays.

[claudepache]

While we’re at it, let’s test the stability of the sort algorithm (https://esdiscuss.org/topic/stable-sort-proposal):

function test(numValues, numDistinctValues, compareFn) {
    "use strict"
    
    let arr = Array(numValues)
    
    for (let i = 0; i < numValues; i++) {
        arr[i] = [ i, Math.floor(Math.random()*numDistinctValues) ]
    }
    
    arr.sort((a, b) => compareFn(a[1], b[1]))
    
    let correct = true
    let stable = true
    for (let i = 1; i < numValues; i++) {
        if (arr[i][1] < arr[i-1][1])
            correct = false
        if (arr[i][1] === arr[i-1][1] && arr[i][0] < arr[i-1][0])
            stable = false
    }
    return [ correct, stable ]
}

(To be tested with numDistinctValues small relatively to numValues for meaningful result about stability.)

Conclusion:

• Firefox and Safari always use a stable sort;
• Edge uses a stable sort for n ≤ 512 and an unstable sort for larger arrays;
• Chrome uses a stable sort for n ≤ 10 and an unstable sort for larger arrays.

[ljharb]

Is the next step here removing step 4.b of SortCompare, since a consistent comparison function can't return NaN?

[ljharb]

Is the next step here removing step 4.b of SortCompare, since a consistent comparison function can't return NaN?