[flang-commits] [flang] [flang][cuda] Prefer valid intrinsics over host-only generics in device code (PR #205376)

[Zhen Wang via flang-commits]

================
@@ -3320,6 +3345,18 @@ auto ExpressionAnalyzer::GetCalleeAndArguments(const parser::Name &name,
     resolution = result.specific;
     dueToAmbiguity = result.failedDueToAmbiguity;
     tried = std::move(result.tried);
+    if (IsCudaDeviceIntrinsicShadowedByHostProcedure(
+            name.source, context_, resolution, isSubroutine)) {
+      ActualArguments localArguments{arguments};
+      if (std::optional<SpecificCall> specificCall{context_.intrinsics().Probe(
+              CallCharacteristics{name.source.ToString(), isSubroutine},
+              localArguments, GetFoldingContext())}) {
+        CheckBadExplicitType(*specificCall, *symbol);
+        return CalleeAndArguments{
+            ProcedureDesignator{std::move(specificCall->specificIntrinsic)},
+            std::move(specificCall->arguments)};
+      }
+    }
----------------
wangzpgi wrote:

I added a explicit CUDA feature guard.

https://github.com/llvm/llvm-project/pull/205376