CU-2083 Update scip-clang to newer Clang version #508

chrapkowski-sg · 2025-12-30T16:26:35Z

Update scip-clang to newer Clang 21 version

chrapkowski-sg · 2025-12-30T16:26:36Z

This change is part of the following stack:

CU-2083 Update scip-clang to newer Clang version #508 ◀

_{Change managed by git-spice.}

chrapkowski-sg · 2025-12-30T16:28:14Z

indexer/Indexer.cc

    case Kind::NamespaceAlias:
    case Kind::Global:
    case Kind::Super:
-    case Kind::TypeSpecWithTemplate:


TypeSpecWithTemplate was removed in LLVM 21 - the TypeSpec case now handles both template and non-template type specifiers

chrapkowski-sg · 2025-12-30T17:35:57Z

indexer/ApproximateNameResolver.cc

+        llvm::dyn_cast<clang::CXXRecordDecl>(baseType->getDecl());
+    if (!baseRecord || !baseRecord->hasDefinition())
+      continue;
+    auto baseResults =


CXXRecordDecl::lookupDependentName() was removed from the Clang API in LLVM 21. This is a local reimplementation with the same semantics.

chrapkowski-sg · 2025-12-30T17:35:57Z

indexer/Indexer.cc

  auto symbol = optSymbol.value();

-  if (functionDecl.isPure() || functionDecl.isThisDeclarationADefinition()) {
+  if (functionDecl.isPureVirtual() || functionDecl.isThisDeclarationADefinition()) {


isPure() was renamed to isPureVirtual() in LLVM 21 for clarity.

chrapkowski-sg · 2025-12-30T17:35:57Z

indexer/Indexer.h

  }
  DERIVE_CMP_ALL(FileLocalSourceRange)
-  DERIVE_EQ_ALL(FileLocalSourceRange)



DERIVE_CMP_ALL now generates operator== via C++20 operator<=>, making DERIVE_EQ_ALL redundant and causing a compile error.

chrapkowski-sg · 2025-12-30T17:35:57Z

indexer/Path.cc

  auto end = llvm::sys::path::end(this->value);
  for (auto it = start; it != end; ++it) {
-    if (it->equals(".") || it->equals("..")) {
+    if (*it == "." || *it == "..") {


StringRef::equals() was deprecated/removed in LLVM 21; use operator== instead.

chrapkowski-sg · 2025-12-30T17:35:57Z

indexer/Preprocessing.h

      clang::CharSourceRange fileNameRange,
      clang::OptionalFileEntryRef optFileEntry, clang::StringRef /*searchPath*/,
-      clang::StringRef /*relativePath*/, const clang::Module * /*importModule*/,
+      clang::StringRef /*relativePath*/,


LLVM 21 added moduleImported parameter to the PPCallbacks::InclusionDirective signature.

chrapkowski-sg · 2025-12-30T17:35:57Z

third_party/llvm-cuda-tooling.patch

@@ -0,0 +1,29 @@
+diff --git a/clang/lib/Tooling/Tooling.cpp b/clang/lib/Tooling/Tooling.cpp


Cherry-pick fix for CUDA assertion failure from llvm/llvm-project#173762. Can be removed once LLVM merges the fix.

chrapkowski-sg · 2025-12-30T17:35:57Z

setup_llvm.bzl

+    # LLVM 21 uses new naming convention: LLVM-VERSION-Platform.tar.xz
+    # with strip prefix LLVM-VERSION-Platform
    grailbio_llvm_toolchain(
        name = name,


LLVM 21 changed release archive naming from clang+llvm-VERSION-TRIPLE to LLVM-VERSION-Platform.

chrapkowski-sg · 2025-12-30T17:37:02Z

indexer/JsonIpcQueue.h

 #include "boost/interprocess/ipc/message_queue.hpp"
 #pragma clang diagnostic pop

-#include "llvm/ADT/Optional.h"


llvm::Optional was removed in favor of std::optional. This include is no longer needed.

chrapkowski-sg · 2025-12-30T18:26:52Z

test/index/types/inheritance.snapshot.cc


  template <>
  struct SpecializedBase<int> {};
-//       ^^^^^^^^^^^^^^^ reference [..] SpecializedBase#


Explicit template specializations (template <> struct SpecializedBase<int> and template <> struct DerivedFromSelf<int *>) no longer emit both a reference and a definition—now just a definition.

This is correct behavior. An explicit specialization defines a new complete type; it doesn't reference the primary template in the usage sense. A reference would be SpecializedBase<int> x; where you instantiate/use the type. LLVM 21 is more precise here.

chrapkowski-sg · 2025-12-31T12:01:45Z

test/index/cuda/kernelcall.snapshot.cu

    fp<<<1, 1>>>(42); // expected-error {{must have void return type}}
 //  ^^ reference local 5

    g1<<<undeclared, 1>>>(42); // expected-error {{use of undeclared identifier 'undeclared'}}


In the line g1<<<undeclared, 1>>>(42);, when the kernel configuration contains an undeclared identifier (undeclared), Clang 21 no longer generates a DeclRefExpr for the callee function g1

chrapkowski-sg · 2026-01-02T00:36:22Z

.bazelrc

 common --enable_platform_specific_config

+# Disable bzlmod for now - WORKSPACE dependencies not yet migrated
+common --enable_bzlmod=false


Update to bazel 7 but we still use WORKSPACE not modules

chrapkowski-sg · 2026-01-02T00:38:35Z

.bazelrc

-# Based on release notes from since we're on Bazel < 7
-# https://github.com/bazel-contrib/toolchains_llvm/releases/tag/0.10.3
-build --features=-supports_dynamic_linker
-build --features=-libtool


We use bazel 7+ so no longer needed

chrapkowski-sg · 2026-01-02T00:38:47Z

.bazelrc

 build --incompatible_enable_cc_toolchain_resolution

-build --repo_env BAZEL_DO_NOT_DETECT_CPP_TOOLCHAIN=1 
-build --@llvm_zstd//:llvm_enable_zstd=false


We use bazel 7+ so no longer needed

jupblb

The Release workflow triggered for this PR failed. This must be fixed before the PR can be merged. That said, the error doesn't seem to be caused by the code changes ("No space left on device," and I was able to build scip-clang manually on both Linux and macOS).

While not a hard requirement right now, in the future it would be nice to split such PRs into separate scopes (in this case, I imagine something like a Bazel update and a Clang update).

Overall, this looks good to me. Once we have the Release pipeline working, this should be easy to merge.

jupblb · 2026-01-08T19:38:50Z

.aspect/bazelrc/ci.bazelrc

@@ -1,3 +1,7 @@
+# Disable bzlmod since this project uses WORKSPACE
+# Docs: https://bazel.build/external/migration#disable-bzlmod
+common --noenable_bzlmod


I think this can be moved to the root .bazelrc. We probably don't want MODULE.bazel to be autogenerated in any case.

jupblb · 2026-01-08T20:43:45Z

.buildkite/pipeline.yaml

+      rosetta steps | buildkite-agent pipeline upload
+    agents:
+      queue: aspect-small-ubuntu-lts
+  # - label: "build and test"


Should we remove this?

jupblb · 2026-01-08T20:45:32Z

test/index/aliases/aliases.snapshot.cc

    static int zero_int = zero<int>;
 //             ^^^^^^^^ definition [..] i/zero_int.
-//                        ^^^^ reference [..] i/j/zero.
+//                        ^^^^ reference [..] i/zero.


This looks like a regression to me.

Amp-Thread-ID: https://ampcode.com/threads/T-019b7003-fbd1-73b0-84e1-30b1483d9aed Co-authored-by: Amp <[email protected]>

Amp-Thread-ID: https://ampcode.com/threads/T-019b702d-0fc2-77ed-83b3-ba3cb6095db7 Co-authored-by: Amp <[email protected]>

LLVM 21 now wraps template names in QualifiedTemplateName more often, even when there's no explicit qualification in the source. The indexer was not handling this case, causing template name references to be silently dropped in base class specifiers like 'struct Derived : Base<T>'. This commit unwraps QualifiedTemplateName (and DeducedTemplateStorage) at the start of saveTemplateSpecializationTypeLoc to get the underlying template before processing.

When looking up member names via this-> in template bodies, the previous code only handled RecordType base classes. LLVM 21 represents dependent base types like T0<T> as TemplateSpecializationType, which was being skipped. Add tryFindDeclForBaseType helper to handle: - Direct RecordType (non-dependent bases) - InjectedClassNameType (inside template definitions) - TemplateSpecializationType (dependent bases like T0<T>) This restores proper reference emission for inherited member calls.

LLVM 21 binaries require glibc 2.34+ which is not available on older CI agents (Ubuntu 20.04 has glibc 2.31). Fall back to system clang-format if the LLVM toolchain version is not compatible.

chrapkowski-sg commented Dec 30, 2025

View reviewed changes

chrapkowski-sg force-pushed the cu-2083-update-scip-clang-to branch from 1ca46e8 to 41926df Compare December 30, 2025 16:37

chrapkowski-sg commented Dec 30, 2025

View reviewed changes

chrapkowski-sg force-pushed the cu-2083-update-scip-clang-to branch 2 times, most recently from 88b71cb to 3480a7a Compare December 30, 2025 19:35

chrapkowski-sg commented Dec 31, 2025

View reviewed changes

chrapkowski-sg force-pushed the cu-2083-update-scip-clang-to branch from 3480a7a to 0fb5ccf Compare January 2, 2026 00:35

chrapkowski-sg commented Jan 2, 2026

View reviewed changes

chrapkowski-sg force-pushed the cu-2083-update-scip-clang-to branch from 0fb5ccf to 329783f Compare January 2, 2026 00:38

chrapkowski-sg commented Jan 2, 2026

View reviewed changes

chrapkowski-sg force-pushed the cu-2083-update-scip-clang-to branch 13 times, most recently from faff6c7 to 235fa63 Compare January 2, 2026 14:58

chrapkowski-sg marked this pull request as ready for review January 2, 2026 15:07

chrapkowski-sg force-pushed the cu-2083-update-scip-clang-to branch 2 times, most recently from ea1763e to 3333e18 Compare January 7, 2026 14:50

jupblb requested changes Jan 8, 2026

View reviewed changes

chrapkowski-sg and others added 11 commits January 9, 2026 16:42

CU-2083 Update scip-clang to newer Clang version

e8745f7

Amp-Thread-ID: https://ampcode.com/threads/T-019b7003-fbd1-73b0-84e1-30b1483d9aed Co-authored-by: Amp <[email protected]>

Cherry-pick CUDA tooling fix from LLVM PR #173762

ad9a099

Amp-Thread-ID: https://ampcode.com/threads/T-019b702d-0fc2-77ed-83b3-ba3cb6095db7 Co-authored-by: Amp <[email protected]>

Fix reformat.sh to fall back to system clang-format

a018ea1

LLVM 21 binaries require glibc 2.34+ which is not available on older CI agents (Ubuntu 20.04 has glibc 2.31). Fall back to system clang-format if the LLVM toolchain version is not compatible.

Updated build container for CI

3988af0

Test docker image

a5f5997

split step into lint and build and use diff agents

b30df85

define aspect config for aspect workflows

0cd29ff

Disabled bzmod for aspect

ee67b51

Re-introduced reformat

d4761e4

chrapkowski-sg force-pushed the cu-2083-update-scip-clang-to branch 4 times, most recently from ed33a2f to 5aab1f5 Compare January 9, 2026 17:57

Updated release action

2d95b5c

chrapkowski-sg force-pushed the cu-2083-update-scip-clang-to branch from 5aab1f5 to 2d95b5c Compare January 9, 2026 18:11

		@@ -0,0 +1,29 @@
		diff --git a/clang/lib/Tooling/Tooling.cpp b/clang/lib/Tooling/Tooling.cpp

CU-2083 Update scip-clang to newer Clang version #508

Are you sure you want to change the base?

CU-2083 Update scip-clang to newer Clang version #508

Conversation

chrapkowski-sg commented Dec 30, 2025

Uh oh!

chrapkowski-sg commented Dec 30, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

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!

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!

jupblb 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!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

chrapkowski-sg commented Dec 30, 2025 •

edited

Loading