Go to declaration and definition

[bliet]

The language server protocol defines the endpoints textDocument/declaration and textDocument/definition.

nonsense sample code:

int methode(int a);

int ABC(SomeType *self)
{
    int HalloWelt;
    HalloWelt = self->in1;
    int a = methode(3);
    return HalloWelt;
}

int methode(int a) { return 4 + a; }

Line 1 is a declaration.
Line 7 is a usage.
Line 11 is a defintion.

In line 7, clangd returns the correct declaration or definition.
However, if I send a declaration request in line 1, the definition is returned.
And if I send a definition request on line 11, it returns the declaration.
I was expecting it to return itself.

Is there a reason for this behavior?

[HighCommander4]

This appears to be deliberate and introduced in https://reviews.llvm.org/D57580.

[bliet]

Thanks for the link.

[bliet]

One more quistion.

If I send a defintion request at ( | = caret position )
int method|e(int a) { return 4 + a; }
the declaration is returned.

If I send a definition request at
int methode|(int a) { return 4 + a; }
the defintion is returned.

That is confusing. Is this deliberate?

[HighCommander4]

One more quistion.

If I send a defintion request at ( | = caret position ) int method|e(int a) { return 4 + a; } the declaration is returned.

If I send a definition request at int methode|(int a) { return 4 + a; } the defintion is returned.

That is confusing. Is this deliberate?

I looked into this a bit. The way this happens is:

• The "toggle" logic, for returning the declaration instead of the definition, is conditioned on a check which amounts to "are you currently over the definition".
• This check fails when the cursor is before the (, because Range::contains() checks for Pos < end rather than Pos <= end (the range is methode).

I don't think the outcome is deliberate, but I'm having trouble at pointing to a place that's clearly wrong. (For example, the LSP spec says about ranges that "the end position is exclusive", so I think Range::contains() is correct to be using Pos < end).

I think what this ultimately boils down to is that the modeling of cursor positions in the LSP is imprecise. Some editors have cursors in between characters and some have cursors on characters, and the LSP doesn't really have a way of distinguishing those cases. The representation for "between two characters" is identical to the one for "on the character after the cursor", so when the cursor is before the (, it's as though it were on the (, and therefore outside the definition range.

Go-to-def still works when the the cursor is on the ( because we have some "bias towards identifier" logic to handle that, but that happens in a deeper layer of the code and isn't used for the Range::contains() check. With some effort things could probably be reorganized so it applies to the Range::contains() check as well. I'm skeptical about whether that's worth the effort, but feel free to file an issue if you feel it's important.