[Clang] Instantiate local constexpr functions eagerly

[zyn0217]

We had a code path in Sema::MarkFunctionReferenced() that deferred local lambda instantiation even for constexprs. This resulted in any calls to them referring to function decls that lack bodies and hence failures at constant evaluation.

The issue doesn't occur when the lambda has no explicit return type because the return type deduction requires instantiation.

Fixes #35052
Fixes #94849

[llvmbot]

@llvm/pr-subscribers-clang

Author: Younan Zhang (zyn0217)

Changes

We had a code path in Sema::MarkFunctionReferenced() that deferred local lambda instantiation even for constexprs. This resulted in any calls to them referring to function decls that lack bodies and hence failures at constant evaluation.

The issue doesn't occur when the lambda has no explicit return type because the return type deduction requires instantiation.

Fixes #35052
Fixes #94849

---

Full diff: https://github.com/llvm/llvm-project/pull/95660.diff

3 Files Affected:

• (modified) clang/docs/ReleaseNotes.rst (+1)
• (modified) clang/lib/Sema/SemaExpr.cpp (+2-1)
• (modified) clang/test/SemaTemplate/instantiate-local-class.cpp (+24)

diff --git a/clang/docs/ReleaseNotes.rst b/clang/docs/ReleaseNotes.rst
index 8c2f737836a9d..a424554ed821d 100644
--- a/clang/docs/ReleaseNotes.rst
+++ b/clang/docs/ReleaseNotes.rst
@@ -847,6 +847,7 @@ Bug Fixes to C++ Support
 - Fixed several bugs in capturing variables within unevaluated contexts. (#GH63845), (#GH67260), (#GH69307),
   (#GH88081), (#GH89496), (#GH90669) and (#GH91633).
 - Fixed handling of brace ellison when building deduction guides. (#GH64625), (#GH83368).
+- Clang now instantiates local constexpr functions eagerly for constant evaluators. (#GH35052), (#GH94849)
 
 Bug Fixes to AST Handling
 ^^^^^^^^^^^^^^^^^^^^^^^^^
diff --git a/clang/lib/Sema/SemaExpr.cpp b/clang/lib/Sema/SemaExpr.cpp
index 99a8704298314..88e9b2b00f84d 100644
--- a/clang/lib/Sema/SemaExpr.cpp
+++ b/clang/lib/Sema/SemaExpr.cpp
@@ -18112,7 +18112,8 @@ void Sema::MarkFunctionReferenced(SourceLocation Loc, FunctionDecl *Func,
 
         if (FirstInstantiation || TSK != TSK_ImplicitInstantiation ||
             Func->isConstexpr()) {
-          if (isa<CXXRecordDecl>(Func->getDeclContext()) &&
+          if (!Func->isConstexpr() &&
+              isa<CXXRecordDecl>(Func->getDeclContext()) &&
               cast<CXXRecordDecl>(Func->getDeclContext())->isLocalClass() &&
               CodeSynthesisContexts.size())
             PendingLocalImplicitInstantiations.push_back(
diff --git a/clang/test/SemaTemplate/instantiate-local-class.cpp b/clang/test/SemaTemplate/instantiate-local-class.cpp
index 47591045fd26e..7eee131e28d60 100644
--- a/clang/test/SemaTemplate/instantiate-local-class.cpp
+++ b/clang/test/SemaTemplate/instantiate-local-class.cpp
@@ -1,5 +1,6 @@
 // RUN: %clang_cc1 -verify -std=c++11 %s
 // RUN: %clang_cc1 -verify -std=c++11 -fdelayed-template-parsing %s
+// RUN: %clang_cc1 -verify -std=c++20 -fsyntax-only %s
 
 template<typename T>
 void f0() {
@@ -509,3 +510,26 @@ namespace LambdaInDefaultMemberInitializer {
   }
   template void f<int>();
 }
+
+#if __cplusplus >= 201703L
+namespace GH35052 {
+
+template <typename F> constexpr int func(F f) {
+  if constexpr (f(1UL)) {
+    return 1;
+  }
+  return 0;
+}
+
+int main() {
+  auto predicate = [](auto v) /*implicit constexpr*/ -> bool {
+    return v == 1;
+  };
+
+  static_assert(predicate(1));
+  return func(predicate);
+}
+
+} // namespace GH35052
+
+#endif

[EricWF]

This looks good to me, but I'm not comfortable officially LGTMing it.

[mizvekov]

LGTM

[zyn0217]

Thanks for the review! I will commit when CI turns green.

[zyn0217]

Linux CI failed with an unrelated error.

TIMEOUT: lldb-api :: functionalities/fork/concurrent_vfork/TestConcurrentVFork.py (87750 of 87750)
  | ******************** TEST 'lldb-api :: functionalities/fork/concurrent_vfork/TestConcurrentVFork.py' FAILED ********************

[cor3ntin]

@zyn0217 What is the status of that?

[zyn0217]

It's still broken, and I haven’t figured out an approach to fix the regression.

I do recall that we were planning to wire up Sema and the constant evaluator, so this patch might end up being completely superseded if that gets implemented.