This is an archive of the discontinued LLVM Phabricator instance.

Avoid -Wshadow warnings about constructor parameters named after fields
ClosedPublic

Authored by rnk on Mar 18 2016, 10:33 AM.

Diff Detail

Event Timeline

rnk updated this revision to Diff 51039.Mar 18 2016, 10:33 AM
rnk retitled this revision from to Avoid -Wshadow warnings about constructor parameters named after fields.
rnk updated this object.
rnk added reviewers: rsmith, rtrieu.
rnk added a subscriber: cfe-commits.

It's not just modifications of shadowed variables that are a problem - one of the one's I'm concerned we should catch is:

struct foo {
  std::unique_ptr<int> p;
  foo(std::unique_ptr<int> p) : p(std::move(p)) {
    f(*p);
  }
};

+Lang, because he was asking me recently about this improvement & thinking
of chipping in

rnk added a comment.Mar 18 2016, 11:08 AM

I'm not sure your example is in scope for -Wshadow, though. Any function call that takes a non-const reference to the parameter could modify it. I guess I'm thinking something like:

void trim_in_place(std::string &s);
struct A {
  std::string s;
  A(std::string s) : s(s) { trim_in_place(s); }
};

I think if we try to match that we'll have too many false positives.

I think your example would be better caught by something like -Wconsumed that looks for uses of objects that have been moved-from. That warning won't be confused by shadowing and will give a better diagnostic anyway.

alexfh added a subscriber: alexfh.Mar 23 2016, 11:03 AM

As a data point: I ran -Wshadow on our code base with a similar, but simpler patch (http://reviews.llvm.org/D18395, just disables the warning on ctor parameters named after fields). It removes more than half of the hits (from ~100k initially) and a random sample of ~100 removed removed hits didn't contain any issues that we'd like to still be warning about.

Ultimately, I think, this diagnostic should be made much less noisy to be actually useful. So maybe even starting with a removal of a larger subset of warnings (and then gradually adding back the cases that seem to be important to flag) would be better.

lib/Sema/SemaExpr.cpp
9858

Why this change?

... and then gradually adding back the cases that seem to be important to flag ...

Maybe as separate warnings.

It's not just modifications of shadowed variables that are a problem - one of the one's I'm concerned we should catch is:

struct foo {
  std::unique_ptr<int> p;
  foo(std::unique_ptr<int> p) : p(std::move(p)) {
    f(*p);
  }
};

This seems totally out of scope for -Wshadow, and seeing a -Wshadow warning on this wouldn't help (imagine a few tens of lines before the *p is used).

thakis added a subscriber: thakis.Mar 23 2016, 11:16 AM

FWIW we don't currently use this warning on Chromium because it's way to noisy. So something like this looks like a great change to me.

dblaikie, are you aware of any codebases that use this warning in its current form?

bcraig added a subscriber: bcraig.Mar 23 2016, 11:20 AM

FWIW we don't currently use this warning on Chromium because it's way to noisy. So something like this looks like a great change to me.

dblaikie, are you aware of any codebases that use this warning in its current form?

Not dblaikie, but I've used the GCC version of -Wshadow on reasonably large code bases before. The ctor pattern that this is trying to squelch was certainly a significant portion of the warnings, particularly when (old) boost headers got involved.

I believe that newer versions of boost (~1.50 and newer?) attempt to be -Wshadow clean in the headers.

FWIW we don't currently use this warning on Chromium because it's way to noisy. So something like this looks like a great change to me.

dblaikie, are you aware of any codebases that use this warning in its current form?

Not dblaikie, but I've used the GCC version of -Wshadow on reasonably large code bases before. The ctor pattern that this is trying to squelch was certainly a significant portion of the warnings, particularly when (old) boost headers got involved.

I believe that newer versions of boost (~1.50 and newer?) attempt to be -Wshadow clean in the headers.

Did squelching this ctor pattern find any bugs? Do you know if boost would miss the warning firing in this case?

rsmith edited edge metadata.Mar 23 2016, 11:31 AM

I think a reasonable approach would be: "do not warn on shadowing if the idiom of naming a constructor parameter after a class member is used, and the class member is initialized from that constructor parameter, and all uses of the parameter in the constructor body would still be valid if it were declared const".

In particular, I think we should probably warn if a non-const reference is bound to the parameter, or if the constructor's member initializer list does not contain field(field).

(I agree that catching A(X x) : x(std::move(x)) { use(x); } is outside the scope of this change.)

lib/Sema/SemaDecl.cpp
6418

Instead of redoing class member lookup every time we check to see if a variable is modifiable, perhaps you could populate a set of shadowed-but-not-diagnosed VarDecl*s here? That way, you can also suppress the duplicate diagnostics if the variable is modified multiple times by removing it from the set.

lib/Sema/SemaExpr.cpp
9663

This query is actually quite expensive (more so than some or all of the other checks below).

Not dblaikie, but I've used the GCC version of -Wshadow on reasonably large code bases before. The ctor pattern that this is trying to squelch was certainly a significant portion of the warnings, particularly when (old) boost headers got involved.

I believe that newer versions of boost (~1.50 and newer?) attempt to be -Wshadow clean in the headers.

Did squelching this ctor pattern find any bugs? Do you know if boost would miss the warning firing in this case?

I don't recall fixing any bugs with the ctor pattern. I can only recall two different classes of things that I have fixed that were flagged with -Wshadow.

  1. Long, horrible, deeply nested method, with a local that was trivially shadowing a different local in a larger scope.
  2. Mutex lock guards and similar "sentry" objects. "std::unique_lock<std::mutex>(name_of_a_member);" The user wanted that to lock 'name_of_member' and release it at end of scope, but instead it declared a default constructed unique_lock called 'name_of_a_member' that shadows the member variable. I've had to fix this kind of bug a number of times, and it is my main justification for turning on -Wshadow in the first place. On that note, I'm really looking forward to the day I can say "auto guard = std::unique_lock(name_of_member);" to effectively get rid of this problem.

I doubt boost would miss the warning, and probably wouldn't even notice (barring overlapping list membership). If you're really wondering if boost would care / notice, @mclow.lists would have a better idea.

rnk added a comment.Mar 23 2016, 4:24 PM

I think a reasonable approach would be: "do not warn on shadowing if the idiom of naming a constructor parameter after a class member is used, and the class member is initialized from that constructor parameter, and all uses of the parameter in the constructor body would still be valid if it were declared const".

That sounds like the right check, but I don't have a lot of time to work on this and I don't want to let the perfect be the enemy of the good. Do you think is a reasonable half-way point for us to leave it for the next year or so?

In particular, I think we should probably warn if a non-const reference is bound to the parameter, or if the constructor's member initializer list does not contain field(field).

Would the reference binding check be somewhere in SemaInit.cpp? I don't see an obvious spot to do this check at first glance.

(I agree that catching A(X x) : x(std::move(x)) { use(x); } is outside the scope of this change.)

Wouldn't the proposed check for non-const references find this example, though?

lib/Sema/SemaDecl.cpp
6418

That sounds like a good idea. I could even leave it empty if we're not doing -Wshadow.

lib/Sema/SemaExpr.cpp
9663

This check has to be faster than name lookup, right? Maybe do it before that?

In D18271#382082, @rnk wrote:

I think a reasonable approach would be: "do not warn on shadowing if the idiom of naming a constructor parameter after a class member is used, and the class member is initialized from that constructor parameter,

Depending on how strictly you define "initialized" here, this is either unhelpful or difficult to chpatterneck. Consider following examples:

class C1 {
  T field;
public:
  void set_field(T f) {
    field = f;
    // do something else...
  }
  C1(T field) { set_field(field); }
};

class C2 {
  T1 f1;
  T2 f2;
public:
  C2(T1 f1, T2 f2) {
    // Some form of initialization of `this->field` from `field` that is not just a copy construction/copy/move,
    // e.g. a method call or an assignment of a single member:
    this->f1.a = f1.a;
    this->f2.CopyFrom(f2);
  }
};

I've seen these patterns being used in the real code, and they don't seem to be dangerous or otherwise worth being flagged by this diagnostic.

and all uses of the parameter in the constructor body would still be valid if it were declared const".

That might also be overly restrictive. Consider a variant of set_field above that takes a T& (which is weird, but it doesn't seem like -Wshadow is the right warning in this case).

That sounds like the right check, but I don't have a lot of time to work on this and I don't want to let the perfect be the enemy of the good. Do you think is a reasonable half-way point for us to leave it for the next year or so?

In particular, I think we should probably warn if a non-const reference is bound to the parameter, or if the constructor's member initializer list does not contain field(field).

See above. There are many cases, when at least the latter pattern is used in a reasonable way.

Depending on how strictly you define "initialized" here, this is either unhelpful or difficult to chpatterneck.

s/chpatterneck/check/

rnk updated this revision to Diff 51697.Mar 25 2016, 4:12 PM
rnk edited edge metadata.
  • Address review comments, use a DenseMap instead of redoing lookup

Richard, is there anything else that blocks this patch?

lib/Sema/SemaDecl.cpp
6372

How about return isa<FieldDecl>(ShadowedDecl) ? SDK_Field : SDK_StaticMember;?

6376

No else after return, please.

rnk added a comment.Apr 7 2016, 12:59 PM

We had a conversation about this change around the office the other week, and people were concerned about false negatives like the trim_in_place one. Basically, I don't have time to discover all the ways you can modify your parameters:

struct B {
  A a;
  B(A a) : a(a) { // modifies parameter, not member
    a.a = 42;
    modify_in_place(&a);
    a.setMember(3);
  }
};

These kinds of bugs are uncommon, but the existing -Wshadow protects users against this today. I think it'll be OK if we give these users an extra -Wshadow-constructor preserves the more strict behavior.

rsmith added inline comments.Apr 7 2016, 1:17 PM
lib/Sema/SemaDecl.cpp
1615

This'd be a good place to produce a -Wshadow-constructor or similar warning if D is still in the map.

6483

Maybe map to the canonical decl before using the Decl* as a key? (This doesn't matter for now, but we'll likely want to extend this to other kinds of shadowing in the future.)

6489–6490

Could be useful to also note the location of the shadowing decl.

rnk updated this revision to Diff 52982.Apr 7 2016, 5:05 PM
rnk marked 3 inline comments as done.
  • Add -Wshadow-all and -Wshadow-field-in-constructor, also address review comments
rnk updated this revision to Diff 52983.Apr 7 2016, 5:09 PM
  • Add a test for -Wshadow-all
lib/Sema/SemaDecl.cpp
6489–6490

Sure, done. I reused and renamed a note from uninit variable diagnostics.

rsmith added inline comments.Apr 14 2016, 5:20 PM
lib/Sema/SemaDecl.cpp
6428–6436

I would prefer for us to produce the -Wshadow "modifying shadowed" warning where it applies, even when -Wshadow-all is enabled. The usual model is to act as if -W* flags are just a filter over the diagnostic output, and this diverges from that.

Is the goal to ensure that -Wshadow-field-in-constructor diagnoses all constructor parameter / field shadowing, even if -Wshadow is not enabled?

rnk added inline comments.Apr 15 2016, 9:35 AM
lib/Sema/SemaDecl.cpp
6428–6436

Well, if we do that we'll get two warnings about the shadowed field: one on the site of the parameter declaration and another on the modification. That felt a bit redundant.

I can remove the if/else and we'll get the behavior you're asking for. Most users will not use -Wshadow-all, and even if they do they'll only see the double diagnostic in very rare cases.

rnk updated this revision to Diff 54206.Apr 19 2016, 8:58 AM
  • Address review comments
  • Add -Wshadow-all and -Wshadow-field-in-constructor, also address review comments
  • Warn twice under -Wshadow-all if a shadowing parameter is modified
rsmith added inline comments.Apr 19 2016, 10:33 AM
lib/Sema/SemaDecl.cpp
6428–6436

Here's what I was thinking:

  • The new warning goes in its own group, say -Wshadow-field-in-constructor-modified
  • The new warning group is part of both -Wshadow and -Wshadow-field-in-constructor
  • We produce the -Wshadow-field-in-constructor warning from ActOnPopScope whenever we remove something from the map.

This means that our behavior is as if -W flags filter our warning output, and we only get one warning for each instance of shadowing, and -Wshadow-field-in-constructor can be used to control all constructor-parameter-shadows-field warnings (and even obscure cases like -Wshadow -Wno-shadow-field-in-constructor do what they look like they should do).

rnk updated this revision to Diff 55044.Apr 26 2016, 11:13 AM
  • Implement suggestions to avoid warning twice
rsmith accepted this revision.Apr 27 2016, 2:51 PM
rsmith edited edge metadata.
rsmith added inline comments.
include/clang/Basic/DiagnosticGroups.td
341–342 ↗(On Diff #55044)

Should ShadowFieldInConstructorModified be a subgroup of ShadowFieldInConstructor? I'd expect -Wshadow-field-in-constructor to enable both.

lib/Sema/SemaDecl.cpp
6427

existance -> existence

This revision is now accepted and ready to land.Apr 27 2016, 2:51 PM
rnk updated this revision to Diff 55359.Apr 27 2016, 4:48 PM
rnk edited edge metadata.
  • address comments
This revision was automatically updated to reflect the committed changes.