[cfe-commits] PATCH: Add support for C++ namespace-aware typo correcting

[Douglas Gregor]

On Mar 25, 2011, at 10:15 PM, Kaelyn Uhrain wrote:

> 
> 
> On Mon, Mar 14, 2011 at 10:28 AM, Douglas Gregor <dgregor at apple.com> wrote:
> 
>        1) Don't walk the namespaces at all; let the check against all of the identifiers in the translation unit winnow the list down to the best candidates based on name alone. We don't know which of these candidates is visible, or how to get there.
> 
> The problem with first checking by name alone is that you can mistakenly discard the best match--I had that problem in early iterations of my patch.  Take for example:
> 
> namespace foo {
>   namespace bar {
>      void tree_branches() { /* ... */ }
>   }
> }
> 
> void three_branches() { /* ... */ }
> 
> int main() {
>   tree_branches();
> }
> 
> Currently my patch will suggest three_branches as it has an edit distance of 1 versus an edit distance of 2 for foo::bar::tree_branches.  Checking against all identifiers first would cause the TypoCorrectionConsumer to discard three_branches in favor of tree_branches since it would think tree_branches has an edit distance of 0... but will later realize tree_branches is unreachable from the current context and discard it (as the code would do now), or possibly suggest a sub-optimal correction (if the code waited to compute the qualifier).  So the trick would be to figure out the qualifier needed for each identifier before making a judgment about edit distance without doing an unqualified lookup and potentially trying to figure out a valid qualified lookup for each identifier.  The trick is trying to get the right info at the right time without walking through the namespaces in the AST given that the IdentifierInfo objects stored in Context.Idents lack any reference to a declaration context.

Okay, that makes sense. Since we have to go through all of the identifiers anyway, we could certainly just keep several different buckets---best edit distance, best edit distance-1, best edit distance-2---so we can catch these cases. Then, if nothing in the first bucket matches, we go on to the next-best bucket, and so on.

>        2) When we're performing lookup for each candidate (e.g., the calls to LookupPotentialTypoResult), first try unqualified lookup. If that fails, save the candidate in a separate "try qualified name lookup" set.
> 
> With my above comments in mind, the unqualified lookup would have to be done for potentially every identifier in the translation unit (any subsequent identifier that has a greater edit distance than the closest correction so far that was successfully lookup can be skipped, but the worst-case scenario is for all identifiers to need to be looked up), followed by qualifier computation (involving walking the decls in the AST for child namespaces, unless I'm missing some handy method for building qualified identifiers) on every identifier that failed the unqualified lookup but does not have a greater edit distance than the closest valid unqualified identifier.

We can't look up all of the identifiers. It's much, much too expensive; that's why we have the system we have, so that we can quickly narrow down to a small set of entities that we perform lookups on.

It's better to guarantee acceptable performance and miss a few typo correction suggestions than it is to slow compilation down to a crawl due to, e.g., a missing header inclusion. 

>        3) If none of the candidates can be found by unqualified lookup, start looking in each of the namespaces we know about to see if qualified name lookup would work to find each of the candidates. You'll likely need a side structure containing all of the (unique, canonicalized) NamespaceDecls in the translation unit to make this efficient.
> 
> How would building the side structure containing all of those NamespaceDecls be more efficient than  ScanNamespaceForCorrections(), which stops walking through namespaces once the distance between a namespace and the current context is greater than the best edit distance known to the TypoCorrectionConsumer?  Or walk any less of the AST than ScanNamespaceForCorrections?

If we're only looking up a small number of identifiers, then we can perform targeted lookups of those identifiers, which (in the PCH case) boils down to a fairly quick search in an on-disk hash table. Since the number of namespaces is generally much fewer than the number of total declarations, we're doing less work.

> 
> Implemented this way, there's no additional penalty for typo correction of visible names. Moreover, we skip the AST walk entirely, and only perform targeted lookups for names that we know exist (because they were in the identifier table). This also permits a few more optimizations:
> 
>        - We could (roughly) order namespaces based on their depth, so that we'll tend to visit the more top-level (and, therefore, more likely to have short qualifiers) namespaces first. If the qualifier for a namespace would be longer than the shortest qualifier we've found so far, we don't have to look in that namespace.
> 
>  ScanNamespaceForCorrections already skips namespaces that would require overly-long qualifiers.

Sure, but in my experience, most large namespaces (std, boost, llvm, clang) are top-level namespaces, so pruning based on overly-long qualifiers is likely to only prune away the small namespaces. 

I still advocate doing that pruning, but I want quick hash-table lookups based on the candidate identifiers rather than a weak of all of the contents of these namespaces.

>        - For a given namespace, we could look for all of the candidates in that namespace at the same time, improving locality (which is particularly import when those lookups hit the disk in the PCH case).
> 
> ScanNamespaceForCorrections currently feeds candidates to the TypoCorrectionConsumer by namespace.  Admittedly, it could store child NamespaceDecls to be searched after searching the current DeclContext instead of searching them as they are found, so that the overall search for candidates is breadth-first instead depth-first.  (And now that I've thought of this, I'm going to be updating ScanNamespaceForCorrections to do so since it will further limit the number of namespaces searched).

As I see it, the main issue for scalability is to avoid having to walk through the declarations in the AST, because that's going to cause massive deserialization from the PCH file. By relying primarily on the walk we have to perform through all of the known identifiers, and then doing targeted lookups into namespaces we know about, we can bound the problem to O(# of candidate identifiers * # of namespaces that are close enough).

	- Doug
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20110326/9eebccf6/attachment.html>