kadircet added inline comments.
================
Comment at: clang-tools-extra/clangd/FindSymbols.cpp:664
+ }
+ // NextPragma is after us but before the next symbol. Our reign is over.
+ break;
----------------
I suppose this is not correct in cases like.
```
#pragma mark - Outer
foo {
#pragma mark - Foo
}
bar {
# pragma mark - Bar
}
baz {
# pragma mark - Baz
}
```
We'll first move `foo` under `Outer`, then start processing `bar` we notice
that It comes after pragma `Foo`, and then also notice that pragma `Foo` is not
contained in `bar` hence break. but all of `foo,bar,baz` should've been moved
to be under `Outer` instead.
================
Comment at: clang-tools-extra/clangd/FindSymbols.cpp:535
+/// by range.
+std::vector<DocumentSymbol> mergePragmas(std::vector<DocumentSymbol> &Syms,
+ std::vector<PragmaMarkSymbol>
&Pragmas,
----------------
dgoldman wrote:
> kadircet wrote:
> > dgoldman wrote:
> > > kadircet wrote:
> > > > dgoldman wrote:
> > > > > kadircet wrote:
> > > > > > dgoldman wrote:
> > > > > > > sammccall wrote:
> > > > > > > > FWIW the flow control/how we make progress seem hard to follow
> > > > > > > > here to me.
> > > > > > > >
> > > > > > > > In particular I think I'm struggling with the statefulness of
> > > > > > > > "is there an open mark group".
> > > > > > > >
> > > > > > > > Possible simplifications:
> > > > > > > > - define a dummy root symbol, which seems clearer than the
> > > > > > > > vector<symbols> + range
> > > > > > > > - avoid reverse-sorting the list of pragma symbols, and just
> > > > > > > > consume from the front of an ArrayRef instead
> > > > > > > > - make the outer loop over pragmas, rather than symbols. It
> > > > > > > > would first check if the pragma belongs directly here or not,
> > > > > > > > and if so, loop over symbols to work out which should become
> > > > > > > > children. This seems very likely to be efficient enough in
> > > > > > > > practice (few pragmas, or most children are grouped into
> > > > > > > > pragmas)
> > > > > > > > define a dummy root symbol, which seems clearer than the
> > > > > > > > vector<symbols> + range
> > > > > > >
> > > > > > > I guess? Then we'd take in a `DocumentSymbol & and a
> > > > > > > ArrayRef<PragmaMarkSymbol> & (or just by value and then return it
> > > > > > > as well). The rest would be the same though
> > > > > > >
> > > > > > > > In particular I think I'm struggling with the statefulness of
> > > > > > > > "is there an open mark group".
> > > > > > >
> > > > > > > We need to track the current open group if there is one in order
> > > > > > > to move children to it.
> > > > > > >
> > > > > > > > make the outer loop over pragmas, rather than symbols. It would
> > > > > > > > first check if the pragma belongs directly here or not, and if
> > > > > > > > so, loop over symbols to work out which should become children.
> > > > > > > > This seems very likely to be efficient enough in practice (few
> > > > > > > > pragmas, or most children are grouped into pragmas)
> > > > > > >
> > > > > > > The important thing here is knowing where the pragma mark ends -
> > > > > > > if it doesn't, it actually gets all of the children. So we'd have
> > > > > > > to peak at the next pragma mark, add all symbols before it to us
> > > > > > > as children, and then potentially recurse to nest it inside of a
> > > > > > > symbol. I'll try it out and see if it's simpler.
> > > > > > >
> > > > > > >
> > > > > > ```
> > > > > > while(Pragmas) {
> > > > > > // We'll figure out where the Pragmas.front() should go.
> > > > > > Pragma P = Pragmas.front();
> > > > > > DocumentSymbol *Cur = Root;
> > > > > > while(Cur->contains(P)) {
> > > > > > auto *OldCur = Cur;
> > > > > > for(auto *C : Cur->children) {
> > > > > > // We assume at most 1 child can contain the pragma (as
> > > > > > pragmas are on a single line, and children have disjoint ranges)
> > > > > > if (C->contains(P)) {
> > > > > > Cur = C;
> > > > > > break;
> > > > > > }
> > > > > > }
> > > > > > // Cur is immediate parent of P
> > > > > > if (OldCur == Cur) {
> > > > > > // Just insert P into children if it is not a group and we are
> > > > > > done.
> > > > > > // Otherwise we need to figure out when current pragma is
> > > > > > terminated:
> > > > > > // if next pragma is not contained in Cur, or is contained in one
> > > > > > of the children, It is at the end of Cur, nest all the children
> > > > > > that appear after P under the symbol node for P.
> > > > > > // Otherwise nest all the children that appear after P but before
> > > > > > next pragma under the symbol node for P.
> > > > > > // Pop Pragmas and break
> > > > > > }
> > > > > > }
> > > > > > }
> > > > > > ```
> > > > > >
> > > > > > Does that make sense, i hope i am not missing something obvious?
> > > > > > Complexity-wise in the worst case we'll go all the way down to a
> > > > > > leaf once per pragma, since there will only be a handful of pragmas
> > > > > > most of the time it shouldn't be too bad.
> > > > > I've implemented your suggestion. I don't think it's simpler, but
> > > > > LMK, maybe it can be improved.
> > > > oops, i was looking into an older revision and missed mergepragmas2, i
> > > > think it looks quite similar to this one but we can probably get rid of
> > > > the recursion as well and simplify a couple more cases
> > > This makes sense, I think that works for the most part besides dropping
> > > the recursion, specifically for
> > >
> > > ```
> > > // Next pragma is contained in the Sym, it belongs there and doesn't
> > > // affect us at all.
> > > if (Sym.range.contains(NextPragma.DocSym.range)) {
> > > Sym.children = mergePragmas2(Sym.children, Pragmas, Sym.range);
> > > continue;
> > > }
> > > ```
> > >
> > > I guess we could explicitly forbid 3+ layers of nesting and handle it
> > > inline there? But I'm not sure it's worth the effort to rewrite all of
> > > this - the recursion shouldn't be deep and we avoid needing to shift
> > > vector elements over by recreating a new one.
> > Sorry I don't follow why we can't get rid of the recursion in this case.
> >
> > Two loop solution I described above literally tries to find the document
> > symbol node, such that the current pragma is contained in that node &&
> > current pragma isn't contained in any of that node's children. Afterwards
> > it inserts the pragma into that node and starts traversing the tree from
> > root again for the next pragma.
> >
> > Again I don't follow where the `3+ layers of nesting` constraint came from.
> > But I do feel like the iterative version is somewhat easier to reason about
> > (especially keeping track of what's happening with `pragmas.front()` and
> > the way it bails out via `parentrange` check). Shifting of the vector is
> > definitely unfortunate but I think it shouldn't imply big performance hits
> > in practice as we are only shifting the children of a single node.
> Yeah, I understand the first part, I think specifically handling the group
> case after you discover where it needs to be inserted is a bit more
> complicated, something like the following:
>
>
> ```
> // Pragma is a group, so we need to figure out where it terminates:
> // - If the next Pragma is not contained in Cur, P owns all of its
> // parent's children which occur after P.
> // - If the next pragma is contained in Cur but actually belongs to
> one
> // of the parent's children, we temporarily skip over it and look at
> // the next pragma to decide where we end.
> // - Otherwise nest all of its parent's children which occur after P
> but
> // before the next pragma.
> ```
>
> And yeah, shifting in the worst case is definitely worst (due to repeat
> shifting) although it shouldn't be too common in practice (things like a
> large @implementation block would probably have the most children).
> we temporarily skip over it and look at the next pragma to decide where we
> end.
Ah you are right I was missing this part. As you suggested we'll need to loop
over all the remaining pragmas, which will make the complexity `((quadratic in
terms of # of pragmas) * (depth of document symbol tree))` I think it is still
acceptable in practice. I don't think there'll be more than a dozen pragmas in
a file. (moreover i think this problem exists with the current implementation
you've provided, see my comment below)
Are you fine with reshaping the implementation in that way then? Because I
still feel like the code will end up looking a lot more straightforward, even
if not the most optimal and we can try to optimize the rest if it shows up in
the latency graphs.
Repository:
rG LLVM Github Monorepo
CHANGES SINCE LAST ACTION
https://reviews.llvm.org/D105904/new/
https://reviews.llvm.org/D105904
_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
