[flang-commits] [flang] 68a0b93 - [flang][cuda] Emit error when a device actual argument is used in host intrinsic (#172914)

[via flang-commits]

Author: Valentin Clement (バレンタイン クレメン)
Date: 2025-12-30T08:49:20-08:00
New Revision: 68a0b93b1f582ac40fac512b8064af0f8771fe39

URL: https://github.com/llvm/llvm-project/commit/68a0b93b1f582ac40fac512b8064af0f8771fe39
DIFF: https://github.com/llvm/llvm-project/commit/68a0b93b1f582ac40fac512b8064af0f8771fe39.diff

LOG: [flang][cuda] Emit error when a device actual argument is used in host intrinsic (#172914)

This can lead to segfault if a device variable is used in a host
intrinsic. Emit an error in semantic.

Added: 
    

Modified: 
    flang/lib/Semantics/check-call.cpp
    flang/lib/Semantics/resolve-names.cpp
    flang/test/Lower/CUDA/cuda-data-transfer.cuf
    flang/test/Semantics/cuf23.cuf

Removed: 
    


################################################################################
diff  --git a/flang/lib/Semantics/check-call.cpp b/flang/lib/Semantics/check-call.cpp
index 5f276d9b9f4f8..b084746a3f787 100644
--- a/flang/lib/Semantics/check-call.cpp
+++ b/flang/lib/Semantics/check-call.cpp
@@ -18,6 +18,7 @@
 #include "flang/Parser/message.h"
 #include "flang/Semantics/scope.h"
 #include "flang/Semantics/tools.h"
+#include "llvm/ADT/StringSet.h"
 #include <map>
 #include <string>
 
@@ -340,6 +341,15 @@ static bool DefersSameTypeParameters(
   return true;
 }
 
+// List of intrinsics that are skipped when checking for device actual
+// arguments.
+static const llvm::StringSet<> cudaSkippedIntrinsics = {
+    "__builtin_c_f_pointer", "__builtin_c_loc"};
+// List of intrinsics that can have a device actual argument if it is an
+// allocatable or pointer.
+static const llvm::StringSet<> cudaAllowedIntrinsics = {"size", "lbound",
+    "ubound", "shape", "allocated", "associated", "kind", "present"};
+
 static void CheckExplicitDataArg(const characteristics::DummyDataObject &dummy,
     const std::string &dummyName, evaluate::Expr<evaluate::SomeType> &actual,
     characteristics::TypeAndShape &actualType, bool isElemental,
@@ -1140,6 +1150,32 @@ static void CheckExplicitDataArg(const characteristics::DummyDataObject &dummy,
           dummyName, toStr(dummyDataAttr), toStr(actualDataAttr));
     }
   }
+  // Emit an error message if an actual argument passed to a host intrinsic is
+  // on the device.
+  if (intrinsic && !FindCUDADeviceContext(scope) &&
+      !FindOpenACCConstructContaining(scope)) {
+    if (!cudaSkippedIntrinsics.contains(intrinsic->name)) {
+      std::optional<common::CUDADataAttr> actualDataAttr;
+      if (const auto *actualObject{actualLastSymbol
+                  ? actualLastSymbol->detailsIf<ObjectEntityDetails>()
+                  : nullptr}) {
+        actualDataAttr = actualObject->cudaDataAttr();
+      }
+      if (actualDataAttr && *actualDataAttr == common::CUDADataAttr::Device) {
+        // Allocatable or pointer with device attribute have their descriptor in
+        // managed memory. It is allowed to pass them to some inquiry
+        // intrinsics.
+        if (!actualLastSymbol || !IsAllocatableOrPointer(*actualLastSymbol) ||
+            (IsAllocatableOrPointer(*actualLastSymbol) &&
+                !cudaAllowedIntrinsics.contains(intrinsic->name))) {
+          messages.Say(
+              "Actual argument %s associated with host intrinsic %s is on the device"_err_en_US,
+              actualLastSymbol ? actualLastSymbol->name() : "",
+              intrinsic->name);
+        }
+      }
+    }
+  }
 
   // Warning for breaking F'2023 change with character allocatables
   if (intrinsic && dummy.intent != common::Intent::In) {

diff  --git a/flang/lib/Semantics/resolve-names.cpp b/flang/lib/Semantics/resolve-names.cpp
index 5cc887791048d..6b59eb94c1c9b 100644
--- a/flang/lib/Semantics/resolve-names.cpp
+++ b/flang/lib/Semantics/resolve-names.cpp
@@ -1081,6 +1081,8 @@ class DeclarationVisitor : public ArraySpecVisitor,
   bool Pre(const parser::SaveStmt &);
   bool Pre(const parser::BasedPointer &);
   void Post(const parser::BasedPointer &);
+  bool Pre(const parser::CUFKernelDoConstruct &);
+  void Post(const parser::CUFKernelDoConstruct &);
 
   void PointerInitialization(
       const parser::Name &, const parser::InitialDataTarget &);
@@ -9278,6 +9280,16 @@ void DeclarationVisitor::LegacyDataInitialization(const parser::Name &name,
   }
 }
 
+bool DeclarationVisitor::Pre(const parser::CUFKernelDoConstruct &x) {
+  // Treat CUDA kernel do construct as OpenACC construct.
+  PushScope(Scope::Kind::OpenACCConstruct, nullptr);
+  return true;
+}
+
+void DeclarationVisitor::Post(const parser::CUFKernelDoConstruct &x) {
+  PopScope();
+}
+
 void ResolveNamesVisitor::HandleCall(
     Symbol::Flag procFlag, const parser::Call &call) {
   common::visit(

diff  --git a/flang/test/Lower/CUDA/cuda-data-transfer.cuf b/flang/test/Lower/CUDA/cuda-data-transfer.cuf
index 015947430b07c..e0c23a1d5df8f 100644
--- a/flang/test/Lower/CUDA/cuda-data-transfer.cuf
+++ b/flang/test/Lower/CUDA/cuda-data-transfer.cuf
@@ -22,6 +22,15 @@ module mod1
     end function
   end interface
 
+  interface
+    function somefunction(d, n)
+      integer :: n
+      real(8), device :: d(n,n)
+      real(8) :: somefunction(n)
+    end function
+  end interface
+      
+
 contains
   function dev1(a)
     integer, device :: a(:)
@@ -447,9 +456,10 @@ end subroutine
 ! CHECK: cuf.data_transfer
 
 subroutine sub23(n)
+  use mod1
   integer :: n
   real(8), device :: d(n,n), x(n)
-  x = sum(d,dim=2) ! Was triggering Unsupported CUDA data transfer
+  x = somefunction(d,2) ! Was triggering Unsupported CUDA data transfer
 end subroutine
 
 ! CHECK-LABEL: func.func @_QPsub23

diff  --git a/flang/test/Semantics/cuf23.cuf b/flang/test/Semantics/cuf23.cuf
index 8c03c18d9b0db..73f13d1d46a17 100644
--- a/flang/test/Semantics/cuf23.cuf
+++ b/flang/test/Semantics/cuf23.cuf
@@ -53,3 +53,10 @@ attributes(global) subroutine global_with_block()
     print*, a ! ok
   end block
 end subroutine
+
+subroutine intrinsic_error(n)
+  integer :: n
+  real(8), device :: d(n,n), x(n)
+!ERROR: Actual argument d associated with host intrinsic sum is on the device
+  x = sum(d,dim=2)
+end subroutine