There are several things I find particularly confusing about this. Fundamentally, Replacements refer to a specific version of the code. Say you have a Replacements object R and a Replacement r that you want to integrate into it. Further assume, that C0 is the initial version of the code whereas C1 is the version after applying R. If you *add* r to R, that means that r (specifically the offset and length in it) refer to C0. If you *merge* it, that means that r refers to C1. Correct replacements can always be *merged*, while *adding* might cause conflicts if they overlap with existing ones. Thus, a function "add or merge" fundamentally doesn't make sense to me. These are fundamentally different concepts.
What this function (and the different implementations we already have) is fundamentally trying to do is to have a version of *add* that resolves (some) conflicts. It does that by assuming that r is currently referring to C0, so it first shifts it and then calls *merge*. The fact that you can shortcut if there isn't actually a conflict (i.e. the "addOr" part) is just an optimization. I am fine with implementing that optimization, but it shouldn't be part of the name. Now, I don't (yet) have a good name as this has very intricate behavior. And I am not sure it is really a good idea to have this magically "resolve" all conflicts, because I think there are different classes. I think what we want here is to have a way to combine replacements that require a defined order, but that don't actually conflict. Lets look at a few cases. Assume R contains a single Replacement A and you trying to "addOrMerge" another Replacement B. 1. A and B are both inserts at the same code location. This seems relatively benign and we just want a defined order 2. A inserts at offset X and B replaces (X, X+N). Still quite benign, but we don't want B to replace anything that A inserted, it should replace the original text 3. B inserts at offset X and A replaces (X, X+N). Same thing, though interesting as we now actually have to switch the order 4. A and B actually replace overlapping code ranges. This is really bad and I think we should continue to report a conflict instead of doing something weird I'd think that your existing implementation gets #1 right but possibly only one of #2/#3. It will also do something very interesting for #4 and I really think what we want is to report an error. On Sat, Sep 10, 2016 at 5:53 PM, Eric Liu <ioe...@google.com> wrote: > This function has been implemented in several places, so I figure it might > be useful to have this function in the library. Although I'm not sure > whether this should be a class interface or just a standalone helper > function. > > @alexshap: thanks for the explanation. For your point B, I'm not sure > about the example, could you elaborate a bit? E.g. which use case to be > specific? What are the headers? > > Generally, the way users of `Replacements` class adding new replacements > also matters. For example, if you have multiple insertions at the same > offset, you could concatenate them and insert as one single replacement, if > you care about performance. > > For most use cases I've seen when converting from old std::set > implementation to the current implementation, performance can be improved > by changing the way replacements are added. And the most natural way is > often the most efficient one. > > > On Sat, Sep 10, 2016, 10:54 Alexander Shaposhnikov <shal1t...@gmail.com> > wrote: > >> alexshap added a comment. >> >> disclaimer: i don't know if this method is the right thing to add (to be >> honest i would prefer a better interface but don't have any good >> suggestions on my mind at the moment), i see several issues (IMO, i might >> be mistaken, apologize in advance) with the current interface of >> Replacements. I will not list all of them, but want to add some context: >> >> A. right now i see the same code (with minor changes) in several places: >> >> 1. >> >> llvm/tools/clang/tools/extra/include-fixer/IncludeFixer.cpp +382 >> >> auto R = tooling::Replacement( >> {FilePath, Info.Range.getOffset(), Info.Range.getLength(), >> Context.getHeaderInfos().front().QualifiedName}); >> auto Err = Replaces.add(R); >> if (Err) { >> llvm::consumeError(std::move(Err)); >> R = tooling::Replacement( >> R.getFilePath(), Replaces.getShiftedCodePosition(R.getOffset()), >> R.getLength(), R.getReplacementText()); >> Replaces = Replaces.merge(tooling::Replacements(R)); >> } >> >> 2. >> >> llvm/tools/clang/tools/extra/change-namespace/ChangeNamespace.cpp +126 >> (see https://reviews.llvm.org/D24183) >> >> >> void addOrMergeReplacement(const tooling::Replacement &R, >> tooling::Replacements *Replaces) { >> auto Err = Replaces->add(R); >> if (Err) { >> llvm::consumeError(std::move(Err)); >> auto Replace = getReplacementInChangedCode(*Replaces, R); >> *Replaces = Replaces->merge(tooling::Replacements(Replace)); >> } >> } >> >> B. For replacements in headers we can get into if(Err) branch quite often >> because the same replacement can be generated multiple times (if that >> header is included into several *.cpp files). >> >> >> https://reviews.llvm.org/D24383 >> >> >> >>
_______________________________________________ cfe-commits mailing list cfe-commits@lists.llvm.org http://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits