C++26 Annotation #166287

zebullax · 2025-11-04T02:20:38Z

⚠️ WIP to onboard C++26 annotation into Clang.
The ongoing work to support reflection done for ex. in #164692 is early yet, so the current PR does not add metafunctions that relate to annotations, since the eval mechanism isnt yet known.

The design is taken from the clang P2996 experimental compiler

annotation leverage attributes inner working
there is additional bookkeeping to be done on attribute being instantiated, for the AP value wrapped by this attribute to be updated, so tablegen for attribute is now allowing for some post instantiation code to run.

WIP

Mooooaar UT
Support unnamed parsed attributes
Address comment on post template instantiation re-eval of AP

TODO

Unwrap all other comments into todos

🙇 Credits shall go to @katzdm, blame to @zebullax

Recognize annotation starting token in attr parsing Signed-off-by: acassagnes <acassagnes@bloomberg.net>

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

Add parsing for annotation Signed-off-by: acassagnes <acassagnes@bloomberg.net>

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

github-actions · 2025-11-04T02:23:57Z

✅ With the latest revision this PR passed the C/C++ code formatter.

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

zebullax · 2025-11-04T03:33:45Z

@katzdm FYI 🙇

clang/lib/Parse/ParseDeclCXX.cpp

clang/lib/Sema/SemaDeclAttr.cpp

Signed-off-by: acassagnes <acassagnes@bloomberg.net> Undo dumb edit Signed-off-by: acassagnes <acassagnes@bloomberg.net>

clang/utils/TableGen/ClangAttrEmitter.cpp

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

Sirraide · 2025-11-05T12:08:48Z

CC @erichkeane

zebullax · 2025-11-11T02:16:52Z

AST representation for

struct [[=1, =2]] Foo4 {};

is

`-CXXRecordDecl 0x1300266d8 </main.cpp:1:1, col:25> col:19 struct Foo4 definition
  |-DefinitionData pass_in_registers empty aggregate standard_layout trivially_copyable pod trivial literal has_constexpr_non_copy_move_ctor can_const_default_init
  | |-DefaultConstructor exists trivial constexpr needs_implicit defaulted_is_constexpr
  | |-CopyConstructor simple trivial has_const_param needs_implicit implicit_has_const_param
  | |-MoveConstructor exists simple trivial needs_implicit
  | |-CopyAssignment simple trivial has_const_param needs_implicit implicit_has_const_param
  | |-MoveAssignment exists simple trivial needs_implicit
  | `-Destructor simple irrelevant trivial constexpr needs_implicit
  |-CXX26AnnotationAttr 0x1300267e8 <col:10>
  | `-IntegerLiteral 0x130026690 <col:11> 'int' 1
  |-CXX26AnnotationAttr 0x130026888 <col:14>
  | `-IntegerLiteral 0x1300266b0 <col:15> 'int' 2
  `-CXXRecordDecl 0x130026900 <col:1, col:19> col:19 implicit struct Foo4

Add smoke test Signed-off-by: acassagnes <acassagnes@bloomberg.net>

erichkeane · 2025-11-12T19:38:30Z

clang/include/clang/Basic/Attr.td

  code AdditionalMembers = [{}];
+    // Any additional text that should be included verbatim after instantiating
+  // an attribute on a template.
+  code PostInstantiationStmts =[{}];


This seems... particularly odd? Can you better explain what is going on here? That comment doesn't help very much. Also, needs a space after the =.

Still needs attention.

erichkeane · 2025-11-12T19:39:41Z

clang/include/clang/Basic/Attr.td

+  SourceLocation EqLoc;
+
+public:
+  static CXX26AnnotationAttr *Create(ASTContext &Ctx, \


These creates don't seem right, they should be auto-generated. Or at least help suppress the ones that are already created?

The fact that this is putting so much extra data into an attribute is concerning.

Those were not strictly needed, cleaned up.

erichkeane · 2025-11-12T19:39:52Z

clang/include/clang/Basic/Attr.td

+  }];
+
+  let PostInstantiationStmts = [{
+    Expr::EvalResult V;


Can you better explain why this is happening?

I believe there might be a more canonical way to achieve that via Sema::InstantiateAttrs , I'm giving this a try and that will remove this PostInstantiateStmts thingy

Gave it a try in 9ddf0a8
I assume (?) there are cases that will break and when I find those, we'll either fix it and keep this approach or revert to the more handcrafty PostInstantiationStmts

erichkeane · 2025-11-12T19:40:29Z

clang/include/clang/Basic/DiagnosticParseKinds.td

  "default scope specifier for attributes is incompatible with C++ standards "
  "before C++17">, InGroup<CXXPre17Compat>, DefaultIgnore;
+def warn_cxx26_compat_annotation : Warning<
+  "annotation is a C++26 extension">, InGroup<CXXPre26Compat>, DefaultIgnore;


Suggested change

"annotation is a C++26 extension">, InGroup<CXXPre26Compat>, DefaultIgnore;

"annotations are a C++26 extension">, InGroup<CXXPre26Compat>, DefaultIgnore;

clang/utils/TableGen/ClangAttrEmitter.cpp

erichkeane · 2025-11-12T19:47:03Z

clang/lib/Sema/SemaDeclAttr.cpp


+static void handleCxx26AnnotationAttr(Sema &S, Decl *D, const ParsedAttr &AL) {
+  Expr *CE = AL.getArgAsExpr(0);
+  if (CE->isLValue()) {


Can you include the standardeeze here as to what exactly is happening/the rules of what should happen on the RHS here?

Still would like this here, other comments not attended to in this func.

erichkeane · 2025-11-12T19:47:51Z

clang/lib/Sema/SemaDeclAttr.cpp

+
+      CE = CopyResult.get();
+    } else {
+      ExprResult RVExprResult = S.DefaultLvalueConversion(AL.getArgAsExpr(0));


Is there a reason the normal copy-init like above is incorrect here?

For this and the comment above, a sample like the following would not be properly handled by default lvalue conversion path

struct U { bool V; constexpr U(bool v) : V(v) {} U(const U&) = delete; // #del-U }; constexpr U u(true); struct [[ =u ]] h2{}; // expected-error {{call to deleted constructor of 'U'}} // expected-note@#del-U {{'U' has been explicitly marked deleted here}}

My understanding is that for record type we basically do not do much besides returning the expression itself in DefaultLvalueConversion

erichkeane · 2025-11-12T19:48:34Z

clang/lib/Parse/ParseDeclCXX.cpp

+    //  - We must not mix with an attribute
+    if (Tok.is(tok::equal)) {
+      if (!getLangOpts().CPlusPlus26) {
+        Diag(Tok.getLocation(), diag::warn_cxx26_compat_annotation);


This seems like it should be an error to me. Allowing annotations, even if ignored on earlier code, seems like a mistake.

Made this an error by default

clang/lib/Parse/ParseDeclCXX.cpp

erichkeane · 2025-11-12T19:53:14Z

clang/lib/Parse/ParseDeclCXX.cpp

+    return;
+
+  IdentifierTable &IT = Actions.PP.getIdentifierTable();
+  IdentifierInfo &Placeholder = IT.get("__annotation_placeholder");


Hmm, this is a little strange unfortunately. Can we try to modify ParsedAttr to accept one without a name? We already have special cases in there for other attribute kinds, and annotations not having one seems sensible.

I think it's doable, but I wonder if adding and handling that corner case is worth the wasted identifier ? Unless there is a more technical argument that I am missing

Its more than just the identifier, its that this is too much of a 'hack' here. ParsedAttr/etc are designed to be extended for things like this (See the large number of ctors/addNews anyway).

The 'addNew' being used here is for generic attributes, and annotations are special enough to be worth extra effort.

Thanks, I took a crack at it in 9109557 , is that what you had in mind ?

cor3ntin · 2025-11-13T01:41:24Z

I think we should have a minimum of reflection (parsing, infra for meta functions) before considering merging that change .
This will let us have fully tested features - at the moment there is nothing we can do except... parsing them.

Note that P3795R0 - approved by the C++ committee last week allow attributes on function parameters

cor3ntin

just making sure we don't merge that that before #164692

zebullax · 2025-11-13T02:32:19Z

Thanks @erichkeane @cor3ntin
I will address those but do give me a tad of time, the next 3 weeks are busy on my end :)

And I am ok with Corentin comment, we can keep this open until we get enough reflection merged, so we can onboard relevant metafunctions here as well.

erichkeane · 2025-11-13T14:15:19Z

I think we should have a minimum of reflection (parsing, infra for meta functions) before considering merging that change . This will let us have fully tested features - at the moment there is nothing we can do except... parsing them.

Note that P3795R0 - approved by the C++ committee last week allow attributes on function parameters

Eh, I disagree. Parsing + AST tests (to make sure they do what they're supposed to!) is a beneficial small-bite of this feature that I think is worth merging on its own. Enabling these ONLY under the -freflection flag though would be a good idea.

I think the meta-function querying them is a useful component to make sure these are fully usable, but IMO, there is enough functionality here (modulo a few small things) that we could merge this and simplify our review life from now on.

Improve documentation Emit error if annotation syntax found in tablegen Signed-off-by: acassagnes <acassagnes@bloomberg.net>

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

zebullax · 2025-11-14T08:21:49Z

I have addressed the quickest comments and the mishandled case on mixing attribute and annotation.
I will address the remaining questions as I free up some time on my end 🙇

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

tbaederr

You need to end your comments in a period (I'm just saying this before Aaron does).

tbaederr · 2025-11-14T09:09:57Z

clang/test/SemaCXX/cxx26-annotation.cpp

@@ -0,0 +1,35 @@
+// RUN: %clang_cc1 -std=c++26  -x c++ %s -verify


Does this also work when passing -fexperimental-new-constant-interpreter?

I think so unless I passed it wrong ?

tbaederr · 2025-11-14T09:11:56Z

clang/include/clang/Basic/Attr.td

+
+public:
+
+  APValue getValue() const { return Value; }


I think we should return a const reference here so we don't copy the APValue.

tbaederr · 2025-11-14T09:12:50Z

clang/lib/Sema/SemaDeclAttr.cpp

+      S.Diag(CE->getBeginLoc(), diag::err_attribute_argument_type)
+          << "C++26 annotation" << 5 << CE->getSourceRange();
+      return;
+    }


Does the order make sense here? The comment says "evaluate to", but the check has nothing to do the the evaluated value, just the ConstantExpr type.

tbaederr · 2025-11-14T09:14:19Z

clang/lib/Sema/SemaDeclAttr.cpp

+        S.Diag(P.first, P.second);
+
+      return;
+    }


I'm a bit confused by the usage of ConstantExpr here, we evaluate it, so it presumably doesn't have an APValue set yet. But we also don't set the value on it here. Is that just missing or am I misunderstanding something?

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

erichkeane

A release note might be a good idea here to show progress on reflection.

I think annotations should ALSO only work under the freflection flag for the short-term, but that isn't in place yet, so we might need to wait on that to merge this.

erichkeane · 2025-11-14T14:28:28Z

clang/include/clang/Basic/Attr.td

  code AdditionalMembers = [{}];
+    // Any additional text that should be included verbatim after instantiating
+  // an attribute on a template.
+  code PostInstantiationStmts =[{}];


Still needs attention.

erichkeane · 2025-11-14T14:30:57Z

clang/lib/Parse/ParseDeclCXX.cpp

+    return;
+
+  IdentifierTable &IT = Actions.PP.getIdentifierTable();
+  IdentifierInfo &Placeholder = IT.get("__annotation_placeholder");


Its more than just the identifier, its that this is too much of a 'hack' here. ParsedAttr/etc are designed to be extended for things like this (See the large number of ctors/addNews anyway).

The 'addNew' being used here is for generic attributes, and annotations are special enough to be worth extra effort.

erichkeane · 2025-11-14T14:33:15Z

clang/lib/Sema/SemaDeclAttr.cpp


+static void handleCxx26AnnotationAttr(Sema &S, Decl *D, const ParsedAttr &AL) {
+  Expr *CE = AL.getArgAsExpr(0);
+  if (CE->isLValue()) {


Still would like this here, other comments not attended to in this func.

erichkeane · 2025-11-14T14:34:22Z

clang/lib/Sema/SemaDeclAttr.cpp

+
+    if (!CE->EvaluateAsConstantExpr(Result, S.Context, CEKind)) {
+      S.Diag(CE->getBeginLoc(), diag::err_attribute_argument_type)
+          << "C++26 annotation" << 4 << CE->getSourceRange();


I don't see it addressed? Usually we'd do something like:

Suggested change

<< "C++26 annotation" << 4 << CE->getSourceRange();

<< "C++26 annotation" << /*template arg=*/4 << CE->getSourceRange();

Though better (and preferred) would be to convert the diagnostic to be a enum_select and use the actual names for these uses.

clang/test/SemaCXX/cxx26-annotation.cpp

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

github-actions · 2025-11-18T04:37:04Z

🐧 Linux x64 Test Results

111348 tests passed
4430 tests skipped

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

zebullax · 2025-11-20T09:06:40Z

@katzdm I suspect you may have tried the approach from 9ddf0a8 instead of having a PostInstantiate statements copied verbatim from attr.td.
If you recall having done so and why that didn't work out, please share 🙇

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

zebullax added 4 commits November 1, 2025 11:11

Add annotation to attribute tablegen

e2f13c0

Recognize annotation starting token in attr parsing Signed-off-by: acassagnes <acassagnes@bloomberg.net>

Build logic around parsing of annotation inside list

76337f8

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

Add sema handler for annotation

cc405d3

Add parsing for annotation Signed-off-by: acassagnes <acassagnes@bloomberg.net>

Fix the copy around of common info for attribute

0a8e125

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

frederick-vs-ja added clang:frontend Language frontend issues, e.g. anything involving "Sema" c++26 reflection Issues related to C++26 reflection labels Nov 4, 2025

Fill the doc for ParseAnnotationSpecifier

c579bd9

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

Apply formatter

7673cab

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

tbaederr reviewed Nov 4, 2025

View reviewed changes

clang/lib/Parse/ParseDeclCXX.cpp Outdated Show resolved Hide resolved

clang/lib/Sema/SemaDeclAttr.cpp Outdated Show resolved Hide resolved

Address review

c28d066

Signed-off-by: acassagnes <acassagnes@bloomberg.net> Undo dumb edit Signed-off-by: acassagnes <acassagnes@bloomberg.net>

zebullax force-pushed the annotation branch from 8e9b28f to c28d066 Compare November 4, 2025 06:22

frederick-vs-ja reviewed Nov 4, 2025

View reviewed changes

clang/utils/TableGen/ClangAttrEmitter.cpp Show resolved Hide resolved

Verify we are not emitting hasAttribute for Annotation spelling

4898a63

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

zebullax added 2 commits November 11, 2025 12:00

Fix checks to allow annotation on declaration

41511c0

Add smoke test Signed-off-by: acassagnes <acassagnes@bloomberg.net>

Merge branch 'main' into annotation

4b3218b

zebullax marked this pull request as ready for review November 11, 2025 04:42

erichkeane reviewed Nov 12, 2025

View reviewed changes

cor3ntin requested changes Nov 13, 2025

View reviewed changes

zebullax added 2 commits November 14, 2025 17:08

Fix logic on mixed attribute and annotation

dd45005

Improve documentation Emit error if annotation syntax found in tablegen Signed-off-by: acassagnes <acassagnes@bloomberg.net>

Apply formatter

0981555

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

Fix UT

0eaf4dc

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

tbaederr requested a review from AaronBallman November 14, 2025 09:09

tbaederr reviewed Nov 14, 2025

View reviewed changes

Close comments with .

8dd0e8b

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

erichkeane reviewed Nov 14, 2025

View reviewed changes

zebullax added 2 commits November 18, 2025 08:34

Add test covering deleted copy on record type for annotation

8559a6e

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

Add more test on structural type

3319fc6

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

zebullax added 2 commits November 19, 2025 08:47

Remove placeholder attr name for annotation

9109557

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

Leverage InstantiateAttrs to re-evaluate APValue

9ddf0a8

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

zebullax added 2 commits November 20, 2025 18:15

Add draft of release notes

a14918f

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

Return const ref AP Value from annotation attr

a440d3f

Signed-off-by: acassagnes <acassagnes@bloomberg.net>

	"annotation is a C++26 extension">, InGroup<CXXPre26Compat>, DefaultIgnore;
	"annotations are a C++26 extension">, InGroup<CXXPre26Compat>, DefaultIgnore;

		@@ -0,0 +1,35 @@
		// RUN: %clang_cc1 -std=c++26 -x c++ %s -verify

	<< "C++26 annotation" << 4 << CE->getSourceRange();
	<< "C++26 annotation" << /template arg=/4 << CE->getSourceRange();

C++26 Annotation #166287

Are you sure you want to change the base?

C++26 Annotation #166287

Uh oh!

Conversation

zebullax commented Nov 4, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

github-actions bot commented Nov 4, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

zebullax commented Nov 4, 2025

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Sirraide commented Nov 5, 2025

Uh oh!

zebullax commented Nov 11, 2025

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

cor3ntin commented Nov 13, 2025

Uh oh!

cor3ntin left a comment

Choose a reason for hiding this comment

Uh oh!

zebullax commented Nov 13, 2025

Uh oh!

erichkeane commented Nov 13, 2025

Uh oh!

zebullax commented Nov 14, 2025

Uh oh!

tbaederr left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

zebullax commented Nov 4, 2025 •

edited

Loading

github-actions bot commented Nov 4, 2025 •

edited

Loading

github-actions bot commented Nov 18, 2025 •

edited

Loading