Dear runtime,

if I perform panic(nil) or panic(somethingThatCanBeNil), it may be a mistake. That's my problem. But I may also do that intentionally and with the magically changed value, I need to not forget about that and workaround the magic.

Less the magic, the less exceptional rules I have to think about. It makes me more productive and my code more comprehensible to my future self. Thanks.

edit: typo

An alternative solution would be to allow recover() to be used in ,ok assignments. For that to really solve the stated problem though, that would require all call sites to be updated.

My personal leaning at the moment is in favor of the proposal as stated.

Is any program out there using nil panics intentionally? A simple search for panic(nil) doesn't give anything on my entire GOPATH besides a go/ssa/interp test. But I'm more worried about panics with variables that could/would be nil.

In any case, I agree with the sentiment and the proposed solution. Perhaps runtime.NilPanic should be clarified that it's only for untyped nils, though. For example, this case has a nil value but wouldn't be equal to nil when recovered:

var err *myError
panic(err)

FTR: I admit this is a real problem. I just prefer explicitly handling it.

Is this a real problem, though? I doubt it. What if the implementation of panic with a nil argument instead just did, panic("nil value passed to panic")? Thereby fixing the problem and diagnosing it one one stroke.

Is this a real problem, though? I doubt it.

I've had to deal with it at least twice. net/http had hangs when people panicked with nil.

@mpvl was talking about error handling the other day and was showing some examples of how defensive code should ideally look like (and how it's hard to get right), and he forgot the nil panic case, showing it's even harder than it seems.

What if the implementation of panic with a nil argument instead just did, panic("nil value passed to panic")? Thereby fixing the problem and diagnosing it one one stroke.

That's what I'm proposing, except with a variable (which could have a String method with that text). But I'm fine (but less fine) with it being that string value exactly, as matching on strings is a little gross.

This proposal makes total sense--for a language which prides itself on its simplicity and obviousness, it is perplexing that the only way to check for a panic(nil) in a recover is by using some other variable. I can't possibly think up a situation when someone would call panic(nil) on purpose anyhow.

While I believe that this is a very good thing to be able to handle, I think that a , ok pattern would be a bit better. If I panic(nil) then in a language like Go, I would want nil to be the recover() value.

So then recover would look like:

defer func() {
    if e, ok := recover(); ok {
        // do recovery stuff
    }
}();

@deanveloper I like your idea for being backwards compatible.

The proper way is more like: […]

That turns out not to be correct either: runtime.Goexit can terminate the goroutine without a recoverable value, but that pattern (and the variant as written by @cznic) would incorrectly report a panic where no panic occurred.
(https://play.golang.org/p/yS1A1c5csrR)

As far as I can tell, there is no way for a defer call to distinguish between panic(nil) and runtime.Goexit(), which to me suggests that at least one of them should be removed.

As far as I can tell, there is no way for a defer call to distinguish between panic(nil) and runtime.Goexit(), which to me suggests that at least one of them should be removed.

As it turns out, we can use the fact that a runtime.Goexit() cannot be recover'd to distinguish it from panic(nil), via a “double defer sandwich”:
https://play.golang.org/p/nlYYWPRO720

Change https://golang.org/cl/134395 mentions this issue: errgroup: rethink concurrency patterns

I think we should treat panic(nil) as a runtime error, so it should panic with a value that implements the runtime.Error interface. This is less convenient for people who want to explicitly detect panic(nil), but I don't see why that matters; if you want to check for it, then, instead, don't do it.

if you want to check for it, then, instead, don't do it.

Perhaps you're using a badly-written library, and you don't have control over it. I've seen worse before 🤷‍♂️

Our team spent 3 hours today hunting down what we now know to be a panic(nil) call. The problem is that panic(nil) is valid yet undetectable since recover() only returns the value that was panic-ed.

This proposal as-is would work as the common if recover() != nil { pattern would see a non-nil value. Therefore, has my vote. (Some internal type would work too, as long as it is not nil)

Making recover() return two values (the panic-ed value and whether a panic occurred) as mentioned several times would make more sense, However, changing recover() will undoubtedly be rejected for being non-backward compatible. Or too complicated if made magical (like map and range that provide one or two values, depending how you use it).

Sometimes. people may call panic(nil) for success. See the case 3 in this article for an example.

[edit] in this example, the panic value is not nil, but it can be.

@go101, eliminating code like that would be an added bonus.

It seems unlikely that code both needs compatibility with an older toolchain and needs to catch panic(nil) in certain packages written for the newer version of Go. When you update from the older toolchain to the newer toolchain, by definition you don't have any packages written for the newer toolchain yet.

The thing I'm worried about there is that some dependency of a package may make use of panic(nil) internally. It's rare, sure, but if it happens and that dependency suddenly uses the new semantics without expecting it, it could lead to horrible subtle bugs for the client of that dependency. By tying it to the go.mod file of each module instead of a global one, the dependencies can update their code and the semantics simultaneously. By tying it to GODEBUG and making it a global flag, it breaks reproducibility and makes the main module dependent on probably undocumented behavior of its dependencies.

I feel like as much as possible semantics changes should be isolated to modules and specific go <version> directives, not global flags, or otherwise there's not much point to versioning them at all, since the main module changing will suddenly break stuff internal to other modules. It's smaller scale, but it's similar to the Python 3 debacle in that it makes code dependent on updates to their dependencies, something out of their direct control. Avoiding that was a big reason that the go <version> directive was made per-module in the first place. GODEBUG is better than Python 3's approach, but it's still a big step backwards from everything I've seen.

@DeedleFake, both approaches have the dependency problem: if we make the fix global then you are dependent on dependencies to correct any breakages, and if we make the fix per-module then you are dependent on dependencies to update their code to get fixed behavior. We have to weigh the two sides and decide which should be preferred.

For the specific case of panic(nil), the evidence is that there is a lot of code out there subtly broken today, which this change will fix, and there is hardly any code out there depending on panic(nil) showing up as a recover of nil, which this change will break. So the balance came out in favor of a global fix.

It is worth noting that making just about any change at all to Go at this point requires that same balance, so it will almost always also motivate a global fix like this.

There is also the problem of changing behavior based on where a function call is located, which in general is expensive and confusing. For this specific case an argument can be made that recover is already aware of its caller in a very specific way, and what's the harm in one additional way. But that argument would not apply to other cases, and I am not even convinced the situation warrants it here. It certainly wouldn't be the precedent we want in general, as I explained on #56986 (comment).

One minor change from the accepted proposal: writing the text in the spec it turns out to be simplest to say that panic(nil) causes a run-time error, in which case the type should be runtime.PanicNilError, not runtime.PanicNil.

For the specific case of panic(nil), the evidence is that there is a lot of code out there subtly broken today, which this change will fix, and there is hardly any code out there depending on panic(nil) showing up as a recover of nil, which this change will break. So the balance came out in favor of a global fix.

This is a good point. I still don't like the change being global, but it is definitely true that there is almost definitely more code out there failing to handle a recover from a panic(nil) properly than code that depends on it. I don't think there's a lot of code producing nil panics, though. It can only happen if you do purposefully call panic(nil) or pass it a nil interface. It's true that code that uses it might be relying on the recover() behavior of other modules that are then updated to use the new behavior, thus breaking the code creating the nil panics, but that's the whole point of module versions. If module A depends on a specific version of module B, then that version of module B will continue working as module A expects because it's literally the same code. Making the change global will break that version connection, effectively making it pointless.

Change https://go.dev/cl/461956 mentions this issue: runtime: replace panic(nil) with panic(new(runtime.PanicNilError))

Change https://go.dev/cl/462293 mentions this issue: cmd/go: update test for change in panic(nil) behavior

As a point of reference, libraries which already explicitly handled nil panics may need to make changes like go-check/check#136 to treat a nil panic prior to go1.21 and a *runtime.NilPanicError result in go1.21+ as equivalent.

I don't think we explicitly decided the question of whether to change net/http to continue to accept panic(nil) as an alias for panic(http.ErrAbortHandler) in Go 1.21 programs (that is, programs that set go 1.21 in their go.mod file).

That is, I don't think we decided whether to include a change like the one in this file from the draft CL:
https://go-review.googlesource.com/c/go/+/452555/1/src/net/http/server.go

Given that we have the GODEBUG and that http.ErrAbortHandler has existed since Go 1.8, I am inclined to leave that change out. The only impact will be that panic(nil) will cause new logging on servers running in Go 1.21 mode.

Change https://go.dev/cl/517036 mentions this issue: _content/doc/go1.21.html: add note about panic nil behavior

Change https://go.dev/cl/536955 mentions this issue: runtime: document GODEBU panicnil values