[clang] [OpenMP] Support capturing structured bindings in OpenMP regions. (PR #190832)
Alexey Bataev via cfe-commits
cfe-commits at lists.llvm.org
Tue Jun 16 06:14:40 PDT 2026
================
@@ -1496,8 +1540,41 @@ void CodeGenFunction::EmitOMPLastprivateClauseFinal(
auto ISrcRef = C->source_exprs().begin();
auto IDestRef = C->destination_exprs().begin();
for (const Expr *AssignOp : C->assignment_ops()) {
- const auto *PrivateVD =
- cast<VarDecl>(cast<DeclRefExpr>(*IRef)->getDecl());
+ const auto *PrivateDecl = cast<DeclRefExpr>(*IRef)->getDecl();
+ // Handle BindingDecls with the same level of support as VarDecls.
+ if (const auto *BD = dyn_cast<BindingDecl>(PrivateDecl)) {
+ QualType Type = BD->getType();
+ const auto *CanonicalBD = cast<ValueDecl>(BD->getCanonicalDecl());
+ if (AlreadyEmittedVars.insert(CanonicalBD).second) {
+ // Get the private address. The BindingDecl was registered in
+ // PrivateScope during initialization, so look it up in LocalDeclMap.
+ auto It = LocalDeclMap.find(CanonicalBD);
+ if (It == LocalDeclMap.end()) {
+ It = LocalDeclMap.find(BD);
+ }
+ assert(It != LocalDeclMap.end() &&
+ "lastprivate BindingDecl not found in LocalDeclMap");
----------------
alexey-bataev wrote:
Why need a double lookup here? This is fragile. Either explain the invariant (which key is guaranteed present and why) or use OMPPrivatizedBindings to retrieve the private address since that map was specifically designed for BindingDecl remapping. If neither map holds the private address at this point, the entire lastprivate-final path for BindingDecls is broken.
https://github.com/llvm/llvm-project/pull/190832
More information about the cfe-commits
mailing list