https://github.com/RexTechnology1 created 
https://github.com/llvm/llvm-project/pull/202958

Fixes #202956

`TemplateDeclInstantiator::VisitVarTemplatePartialSpecializationDecl` assumes 
the
primary member variable template has already been instantiated into the current
instantiation (`Owner`) — normally true, because the primary is declared before
its partial specializations and is therefore visited first while the enclosing
class is instantiated.

When the enclosing class template's pattern is deserialized from a **preamble
PCH built with compiler errors** (`-fallow-pch-with-compiler-errors`, which
clangd always uses for preambles), the primary may not have been materialized
into `Owner` by the time a partial specialization is visited. The lookup then
returns empty, tripping `assert(!Found.empty() && "Instantiation found
nothing?")` in assertions builds and dereferencing a null `Found.front()`
(SIGSEGV) in release builds.

This is hit in practice by clangd on code using `std::expected<T, E>` from
libstdc++ 15.x (its `__cons_from_expected` member variable template has a
`<…, bool>` partial specialization), whenever the preamble contains an
unresolved `#include` — a routine state in interactive editing (e.g.
not-yet-generated headers). A five-line reproducer (3-line header + 2-line main
file) is in the linked issue; it crashes deterministically, and removing only
the unresolved-include line makes it disappear. All released clang versions
tested (20, 21, 22) and current `main` are affected — the crashing function is
byte-identical at tip `53ae585a9` (2026-06-10).

**Fix:** when the lookup is empty, instantiate the primary member variable
template on demand and look it up again, restoring the invariant the assertions
expect. The original invariant assertions are kept (so a genuinely-missing
primary is still flagged in +Asserts builds), with a `nullptr` return as a
release-safe guard. The common (non-deserialized) path is unchanged.

**Verification:** on an assertions build of trunk, the reproducer no longer
crashes and the retained assertions do not fire. Instrumentation confirms the
on-demand instantiation succeeds, is taken exactly once per affected
instantiation, and creates no duplicate declarations (the member-instantiation
loop does not visit the primary again — counted via instrumentation on
`Owner->addDecl`). `SemaTemplate`, `PCH`, `Modules`, and `CXX/temp` show no
regressions (the only failures are identical on the unpatched baseline).

### Test
The reproducer is tiny (3-line header + 2-line main file) but currently fires
only through clangd's in-process `PrecompiledPreamble` with real libstdc++ 15
headers: a CLI `-emit-pch`/`-include-pch` split does not reproduce, even with
`-Xclang -fallow-pch-with-compiler-errors` and `-Xclang -preamble-bytes=N,1`,
and a self-contained mock of the `<expected>` structure does not trigger it. If
there is precedent for an errorful-preamble fixture (e.g. a clangd
`TestTU`-style unittest), I'm happy to add a regression test on top of it —
guidance welcome.

---
<sub>Developed with assistance from an AI coding tool (Claude);
reviewed by me, and I can speak to the analysis and trade-offs. The
commit carries `Assisted-by: Claude (Anthropic)`.</sub>


>From 64df0a761cb0a5849f514475d8241f13e53afeed Mon Sep 17 00:00:00 2001
From: rextechnology <[email protected]>
Date: Tue, 9 Jun 2026 21:32:36 +0300
Subject: [PATCH] [Clang][Sema] Fix crash instantiating a member variable
 template partial specialization when the enclosing class template is
 deserialized

VisitVarTemplatePartialSpecializationDecl assumes the primary member variable
template has already been instantiated into the current instantiation (Owner),
which is normally true because the primary is declared before its partial
specializations and is visited first while the enclosing class is instantiated.

When the enclosing class template's pattern is deserialized from a precompiled
preamble built with compiler errors (-fallow-pch-with-compiler-errors, which
clangd uses for preambles), the primary may not have been materialized into
Owner yet, leaving the lookup empty. This tripped the
assert(!Found.empty() && "Instantiation found nothing?") in assertions builds
and dereferenced a null Found.front() (SIGSEGV) in release builds.

This is hit in practice by clangd: completing std::expected<T, E> from
libstdc++ 15.x crashes whenever the preamble contains an unresolved #include,
a routine state in interactive editing (e.g. not-yet-generated headers). All
released clang versions tested (20, 21, 22) and current trunk are affected.

Fix: when the lookup is empty, instantiate the primary member variable template
on demand and look it up again, which restores the invariant the assertions
expect. The original invariant assertions are kept, so a genuinely-missing
primary is still flagged in +Asserts builds, with a null return as a
release-safe guard. The common (non-deserialized) path is unchanged.

Signed-off-by: rextechnology <[email protected]>
Assisted-by: Claude (Anthropic)
---
 .../lib/Sema/SemaTemplateInstantiateDecl.cpp  | 24 +++++++++++++++++++
 1 file changed, 24 insertions(+)

diff --git a/clang/lib/Sema/SemaTemplateInstantiateDecl.cpp 
b/clang/lib/Sema/SemaTemplateInstantiateDecl.cpp
index aa381f09138de..c11312843b18e 100644
--- a/clang/lib/Sema/SemaTemplateInstantiateDecl.cpp
+++ b/clang/lib/Sema/SemaTemplateInstantiateDecl.cpp
@@ -2558,10 +2558,34 @@ Decl 
*TemplateDeclInstantiator::VisitVarTemplatePartialSpecializationDecl(
 
   // Lookup the already-instantiated declaration and return that.
   DeclContext::lookup_result Found = Owner->lookup(VarTemplate->getDeclName());
+
+  // Normally the primary member variable template has already been 
instantiated
+  // into Owner, because it is declared before its partial specializations and
+  // so is visited first while instantiating the enclosing class. However, when
+  // the class template pattern is deserialized from a module or precompiled
+  // preamble, the primary may not have been materialized into Owner yet,
+  // leaving the lookup empty. Previously this asserted (and crashed release
+  // builds with a null dereference). Instantiate the primary on demand and 
look
+  // it up again.
+  if (Found.empty()) {
+    if (Decl *InstPrimary = Visit(VarTemplate))
+      if (auto *InstVTD = dyn_cast<VarTemplateDecl>(InstPrimary))
+        Found = Owner->lookup(InstVTD->getDeclName());
+  }
+
+  // After the on-demand instantiation above the primary must be present. Keep
+  // the original invariant assertions so a genuinely-missing primary is still
+  // flagged in +Asserts builds, with a null return as a release-safe guard so 
a
+  // stray case degrades gracefully instead of dereferencing an empty lookup
+  // result.
   assert(!Found.empty() && "Instantiation found nothing?");
+  if (Found.empty())
+    return nullptr;
 
   VarTemplateDecl *InstVarTemplate = dyn_cast<VarTemplateDecl>(Found.front());
   assert(InstVarTemplate && "Instantiation did not find a variable template?");
+  if (!InstVarTemplate)
+    return nullptr;
 
   if (VarTemplatePartialSpecializationDecl *Result =
           InstVarTemplate->findPartialSpecInstantiatedFromMember(D))

_______________________________________________
cfe-commits mailing list
[email protected]
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits

Reply via email to