[CIR] Implement static local variable initialization with guard variables #2046

lanza · 2025-12-03T09:48:28Z

Stack from ghstack (oldest at bottom):

-> [CIR] Implement static local variable initialization with guard variables #2046

This implements thread-safe initialization of function-local static
variables with dynamic initializers, following the Itanium C++ ABI.

The implementation:

Adds static_local attribute to GlobalOp and GetGlobalOp to mark
function-local statics requiring guarded initialization
Implements guarded initialization in LoweringPrepare pass using
__cxa_guard_acquire/__cxa_guard_release for thread safety
Adds helper methods createIsNull/createIsNotNull to CIRBaseBuilder
Implements emitGuardedInit in CIRGenItaniumCXXABI to emit the
ctor region and mark the global as static_local
Adds mangleStaticGuardVariable to ASTAttrInterfaces for mangling

The guard variable pattern follows Itanium ABI 3.3.2:

First byte of guard checked with acquire load for fast path
__cxa_guard_acquire called if uninitialized
__cxa_guard_release called after successful initialization

[ghstack-poisoned]

…bles This implements thread-safe initialization of function-local static variables with dynamic initializers, following the Itanium C++ ABI. The implementation: 1. Adds `static_local` attribute to GlobalOp and GetGlobalOp to mark function-local statics requiring guarded initialization 2. Implements guarded initialization in LoweringPrepare pass using __cxa_guard_acquire/__cxa_guard_release for thread safety 3. Adds helper methods createIsNull/createIsNotNull to CIRBaseBuilder 4. Implements emitGuardedInit in CIRGenItaniumCXXABI to emit the ctor region and mark the global as static_local 5. Adds mangleStaticGuardVariable to ASTAttrInterfaces for mangling The guard variable pattern follows Itanium ABI 3.3.2: - First byte of guard checked with acquire load for fast path - __cxa_guard_acquire called if uninitialized - __cxa_guard_release called after successful initialization ghstack-source-id: 6081423 Pull-Request: #2046

[ghstack-poisoned]

…bles This implements thread-safe initialization of function-local static variables with dynamic initializers, following the Itanium C++ ABI. The implementation: 1. Adds `static_local` attribute to GlobalOp and GetGlobalOp to mark function-local statics requiring guarded initialization 2. Implements guarded initialization in LoweringPrepare pass using __cxa_guard_acquire/__cxa_guard_release for thread safety 3. Adds helper methods createIsNull/createIsNotNull to CIRBaseBuilder 4. Implements emitGuardedInit in CIRGenItaniumCXXABI to emit the ctor region and mark the global as static_local 5. Adds mangleStaticGuardVariable to ASTAttrInterfaces for mangling The guard variable pattern follows Itanium ABI 3.3.2: - First byte of guard checked with acquire load for fast path - __cxa_guard_acquire called if uninitialized - __cxa_guard_release called after successful initialization ghstack-source-id: bf90e4f Pull-Request: #2046

[ghstack-poisoned]

…bles This implements thread-safe initialization of function-local static variables with dynamic initializers, following the Itanium C++ ABI. The implementation: 1. Adds `static_local` attribute to GlobalOp and GetGlobalOp to mark function-local statics requiring guarded initialization 2. Implements guarded initialization in LoweringPrepare pass using __cxa_guard_acquire/__cxa_guard_release for thread safety 3. Adds helper methods createIsNull/createIsNotNull to CIRBaseBuilder 4. Implements emitGuardedInit in CIRGenItaniumCXXABI to emit the ctor region and mark the global as static_local 5. Adds mangleStaticGuardVariable to ASTAttrInterfaces for mangling The guard variable pattern follows Itanium ABI 3.3.2: - First byte of guard checked with acquire load for fast path - __cxa_guard_acquire called if uninitialized - __cxa_guard_release called after successful initialization ghstack-source-id: 272ef4e Pull-Request: #2046

[ghstack-poisoned]

…bles This implements thread-safe initialization of function-local static variables with dynamic initializers, following the Itanium C++ ABI. The implementation: 1. Adds `static_local` attribute to GlobalOp and GetGlobalOp to mark function-local statics requiring guarded initialization 2. Implements guarded initialization in LoweringPrepare pass using __cxa_guard_acquire/__cxa_guard_release for thread safety 3. Adds helper methods createIsNull/createIsNotNull to CIRBaseBuilder 4. Implements emitGuardedInit in CIRGenItaniumCXXABI to emit the ctor region and mark the global as static_local 5. Adds mangleStaticGuardVariable to ASTAttrInterfaces for mangling The guard variable pattern follows Itanium ABI 3.3.2: - First byte of guard checked with acquire load for fast path - __cxa_guard_acquire called if uninitialized - __cxa_guard_release called after successful initialization ghstack-source-id: 29329f1 Pull-Request: #2046

bcardosolopes

Nice, this looks mostly good. Please add a .cir tests for parsing/printing. More comments inline

clang/include/clang/CIR/Interfaces/ASTAttrInterfaces.td

bcardosolopes · 2025-12-05T20:10:51Z

clang/lib/CIR/CodeGen/CIRGenDecl.cpp

+                             oldGlobalOp.getName(), typedInit.getType(),
+                             oldGlobalOp.getConstant(), globalOp.getLinkage());
+    // FIXME(cir): OG codegen inserts new GV before old one, we probably don't
+    // need that?


LLVM translation usually put them out in the proper order, because of the way MLIR handles this I don't think it should be a problem for us.

bcardosolopes · 2025-12-05T20:11:46Z

clang/lib/CIR/CodeGen/CIRGenFunction.h

+  void emitCXXGuardedInit(const VarDecl &varDecl, cir::GlobalOp globalOp,
+                          bool performInit);
+
+  /// EmitCXXGlobalVarDeclInit - Create the initializer for a C++


Remove the EmitCXXGlobalVarDeclInit - in the comment

bcardosolopes · 2025-12-05T20:12:57Z

clang/test/CIR/CodeGen/static-local.cpp

+//      CIR: cir.global "private" internal dso_local static_local @_ZZ3foovE3val = #cir.int<0> : !s32i {alignment = 4 : i64, ast = #cir.var.decl.ast}
+//      CIR: cir.global "private" internal dso_local @_ZGVZ3foovE3val = #cir.int<0> : !s64i {alignment = 8 : i64}
+//      CIR: cir.func {{.*}} @_Z3foov() extra(#fn_attr) {
+// CIR-NEXT:   %0 = cir.get_global static_local @_ZZ3foovE3val : !cir.ptr<!s32i>


None of our tests should have literal SSA values, regex it all please!

bcardosolopes · 2025-12-05T20:17:00Z

clang/test/CIR/CodeGen/static-local.cpp

+// CIRGEN-NEXT:   %1 = cir.call @_Z3fnAv() : () -> !s32i
+// CIRGEN-NEXT:   cir.store align(4) %1, %0 : !s32i, !cir.ptr<!s32i>
+// CIRGEN-NEXT: } {alignment = 4 : i64, ast = #cir.var.decl.ast}
+// CIRGEN-NEXT: cir.func


The check for cir.func should only be CIRGEN since it's a coincidence it shows up next but it isn't required to do so. Also add the name of the function too.

bcardosolopes · 2025-12-05T20:30:34Z

clang/test/CIR/CodeGen/static-local.cpp

+
+//      CIRGEN: cir.func private @_Z3fnAv() -> !s32i
+//      CIRGEN: cir.global "private" internal dso_local static_local @_ZZ3foovE3val = ctor : !s32i {
+// CIRGEN-NEXT:   %0 = cir.get_global @_ZZ3foovE3val : !cir.ptr<!s32i>


Having two instances of cir.get_global @_ZZ3foovE3val inside the initializer and another one outside with the specific keyword is a bit confusing, however still pretty handy when unwrapping so that lowering prepare could be more direct.

Since there's still more work in this area, better not over design too early, but we should revisit this in the future. Can you add a comment in the CIRGen code that generates this get_global so we can later take other things into consideration?

Update

0f8d66f

[ghstack-poisoned]

lanza requested review from andykaylor, bcardosolopes and xlauko as code owners December 3, 2025 09:48

Update

0e55e5d

[ghstack-poisoned]

Update

eb0c219

[ghstack-poisoned]

lanza marked this pull request as draft December 3, 2025 16:57

Update

e84a7f7

[ghstack-poisoned]

lanza marked this pull request as ready for review December 5, 2025 06:47

bcardosolopes requested changes Dec 5, 2025

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[CIR] Implement static local variable initialization with guard variables #2046

[CIR] Implement static local variable initialization with guard variables #2046

Uh oh!

lanza commented Dec 3, 2025 •

edited

Loading

Uh oh!

bcardosolopes left a comment

Uh oh!

Uh oh!

bcardosolopes Dec 5, 2025

Uh oh!

bcardosolopes Dec 5, 2025

Uh oh!

bcardosolopes Dec 5, 2025

Uh oh!

bcardosolopes Dec 5, 2025

Uh oh!

bcardosolopes Dec 5, 2025

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

[CIR] Implement static local variable initialization with guard variables #2046

Are you sure you want to change the base?

[CIR] Implement static local variable initialization with guard variables #2046

Uh oh!

Conversation

lanza commented Dec 3, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

bcardosolopes left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

bcardosolopes Dec 5, 2025

Choose a reason for hiding this comment

Uh oh!

bcardosolopes Dec 5, 2025

Choose a reason for hiding this comment

Uh oh!

bcardosolopes Dec 5, 2025

Choose a reason for hiding this comment

Uh oh!

bcardosolopes Dec 5, 2025

Choose a reason for hiding this comment

Uh oh!

bcardosolopes Dec 5, 2025

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

lanza commented Dec 3, 2025 •

edited

Loading