[cfe-dev] [llvm-dev] RFC: Modernizing our use of auto

Stephen Kelly via cfe-dev cfe-dev at lists.llvm.org
Sat Jan 5 11:36:46 PST 2019

On 31/12/2018 04:54, Chris Lattner wrote:
> On Dec 16, 2018, at 11:44 AM, Stephen Kelly via llvm-dev <llvm-dev at lists.llvm.org> wrote:
>> On 25/11/2018 14:43, Stephen Kelly via llvm-dev wrote:
>>> However this is a proposal for more modern thinking regarding the permissiveness of auto in LLVM codebases.
>>> Currently the rule on the use of auto is here:
>> Hi,
>> Thanks for the input on this topic, which has appeared here on the mailing list, on the phab review, and on IRC.
>> Almost all respondants generally expressed the claim "The type must be obvious if auto is used", though as I wrote before the guide uses auto in context that the type is not obvious:
>> for (const auto &Val : Container) { observe(Val); }
>> It seems that the respondants wish for 'almost never auto'. Fair enough, if the existing practice supports that.
>> There is one problem though, or rather ~56,000 problems. That's how many uses of auto there are in the entire codebase currently, according to someone on IRC.
> I find this to be a helpful line of argument.

Given what you wrote below, maybe you are missing a negation somewhere 
in this sentence?

> We should, as a community, decide what the right thing is regardless of the existing art.

The existing art is part of 'the community deciding what to do'.

And yes, I think it makes sense to 'standardize existing practice' where 

>> How did all of those uses get into the codebase? Does it indicate that the guide is not followed, or does it indicate that the guide is too subjective, or that maybe the 'the type must be obvios' guide does not reflect the 'reality at the coalface' anymore? Should those uses of auto be changed?
> My understanding is that there has been no widely understood or accepted policy, so different coders and reviewers are doing different things.

And different contexts within LLVM are fine with auto, but seem to get 
campaigning from other parts who have a different interpretation of the 


>> How is a new contributor to react to any of that? What are the real criteria that we can use to determine where auto will cause a patch to be rejected? Does it only depend on who you get as a reviewer?
> Right now, it is quite arbitrary.  This is a bug, not a feature.

Do we have some idea of who is interested in fixing the bug? It can't be 
just one person fixing it - this is a community issue. You've suggested 
that the guideline needs an update, and I've already suggested an 
update. Is it only the two of us? How can we proceed?

>> Here is a quote from this thread from Chris and supported by Chandler and Quentin at least:
>>> Perhaps the rule came be rewritten more generally, as
>>> in “auto should generally only be used when the inferred
>>> type is obvious from context, e.g. as a result of a cast
>>> like dyn_cast or as a result of constructing a value with
>>> a constructor containing a type name.”?
>> Is it reasonable to have that as a rule if there are ~12000 uses which break that rule?
> If you’d like to make progress on this, I think you should start by building consensus.

Well, I'm trying to find out what the positions people have are, but 
even though there are so many existing usages of auto, this thread is 
not getting responses from the people who put them there. So, the code 
says one thing, and the guideline says arguably the same thing, but 
people have alternative interpretations of the guideline. But at least 
the responses here are not really representative.

For me that means I'm not able to get my clang-query features 
upstream because I'm getting reviews which say "remove auto, here and 
everywhere in this file." in


That's a bit of a difficult review comment, given the ways it is already 
used throughout the code.

> It seems like there is widespread enough acknowledgement that the current state of things is broken, but there is no concrete proposal for a coding standards change.  Please prepare a patch so we can discuss it.

I made a proposal in my initial mail in this thread. See the end of the 
email: https://lists.llvm.org/pipermail/llvm-dev/2018-November/127953.html

Phab is not well-suited to discussion like this, so we should probably 
keep it on the mailing list for now.

But, here's some updated thinking:

New guidelines should

* Be easy to follow
* Have some consistency
* Be modern
* Be welcoming (or at least non-hostile) to newcomers
* Standardize existing practice in LLVM
* Achieve a consensus of support about the spirit of the guideline 
(consensus is not unanimity) and be acceptable to people who dislike auto

Can we agree on that much to start?

On the Phab review, some people expressed that they liked the examples 
of when auto is acceptable. Here is an updated attempt at guideline text 
for that section:

Some are advocating a policy of "almost always ``auto``" in C++11, 
however LLVM
uses a more moderate stance. Don't "almost always" use ``auto``, but it 
may be used where either the Concept or the type is obvious from the 
context. Here are
some cases where use of ``auto`` would make sense:

* Where the type appears elsewhere in the line (eg a dyn_cast)
* Where the variable name or initializing expression provides enough 
information (`auto SR = getSourceRange()`)
* Where the context makes the *Concept* obvious, even if the type is not 
obvious (eg, where the instance is only used as an iterator, or in an 
algorithm as a container-like concept, or only with a validity check, or 
an AST Matcher).

Exceptions may arise, but they should only arise in exceptional cases. 
If the case is not exceptional, apply the guidelines in review discussion.

The most important thing here is that it does not accept your proposal 
that 'the type must be obvious'. Instead, it recognizes that `auto` is 
really "an unspecified concept" - unspecified only because we can't 
specify the concept in C++ yet.

However, the point/concern seems to be that readers of code should know 
the instance may be used in its local scope.

That's why these guidelines would allow `auto Ctors` in

llvm::Optional<std::pair<std::string, MatcherCtor>>
  getNodeConstructorType(ASTNodeKind targetType) {
    const auto &Ctors = RegistryData->nodeConstructors();
    auto It = llvm::find_if(
        Ctors, [targetType](const NodeConstructorMap::value_type &Ctor) {
          return Ctor.first.isSame(targetType);
    if (It == Ctors.end())
      return llvm::None;
    return It->second;

because `Ctors` is obviously the Container *concept*, and knowing 
exactly what type it is is not necessary in the local context.

However, in the below code it would probably not be ok because we're 
calling methods on the instance which opens up more possibilities (is it 
a base interface? etc):

void SomeClass::foo(int input)
     auto Ctors = getCtors(input);

     m_widgets = Ctors->calculate();

Here, `Ctors` is definitely not a Container concept, we don't know what 
kind of concept it is, so we should know the type by seeing it typed in 
the code.

Another example from earlier in the thread:

  template <typename BaseT, typename DerivedT>
  void registerIfNodeMatcher(...) {
    auto NodeKind = ASTNodeKind::getFromNodeKind<DerivedT>();
    if (!NodeKind.isNone())
      NodeCtors[NodeKind] = std::make_pair(MatcherName, Descriptor);

Here, `NodeKind` is used as an Optional (or Maybe) *concept*. All we do 
is a validity check. So, `auto` should be allowed here.

This 'the concept must be obvious' guideline is also what allows the use 
of `auto` for iterators.

What do people think of "Either the Concept or the type should be 
obvious from the context" as a baseline guideline?



-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20190105/1ee4747b/attachment.html>

More information about the cfe-dev mailing list