[flang-commits] [flang] [flang][cuda] Avoid false positive on multi device symbol with components (PR #192177)

[via flang-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-flang-semantics

Author: Valentin Clement (バレンタイン クレメン) (clementval)

<details>
<summary>Changes</summary>

Semantic was wrongly flagging derived-type components as two device resident object. Update how we collect symbols and count the number of device resident object. 

---
Full diff: https://github.com/llvm/llvm-project/pull/192177.diff


4 Files Affected:

- (modified) flang/include/flang/Evaluate/tools.h (+11) 
- (modified) flang/lib/Evaluate/tools.cpp (+34) 
- (modified) flang/lib/Semantics/check-cuda.cpp (+2-1) 
- (modified) flang/test/Lower/CUDA/cuda-data-transfer.cuf (+26) 


``````````diff
diff --git a/flang/include/flang/Evaluate/tools.h b/flang/include/flang/Evaluate/tools.h
index cab6ddbb6b27c..a1c66e8f50bd7 100644
--- a/flang/include/flang/Evaluate/tools.h
+++ b/flang/include/flang/Evaluate/tools.h
@@ -1299,6 +1299,13 @@ bool CheckForCoindexedObject(parser::ContextualMessages &,
     const std::optional<ActualArgument> &, const std::string &procName,
     const std::string &argName);
 
+// Get the symbol vectors of the expression where symbols are grouped together
+// if they are part of the same component expression.
+//
+// Example: a%b + c%d
+// Will be grouped as: [(a, b), (c, d)]
+std::vector<SymbolVector> GetSymbolVectors(const Expr<SomeType> &expr);
+
 bool IsCUDADeviceSymbol(const Symbol &sym);
 bool IsCUDADeviceOnlySymbol(const Symbol &sym);
 
@@ -1348,6 +1355,10 @@ template <typename A> inline int GetNbOfCUDADeviceSymbols(const A &expr) {
   return symbols.size();
 }
 
+// Get the number of unique symbols with CUDA device attribute.
+int GetNbOfUniqueCUDADeviceSymbols(
+    const std::vector<SymbolVector> &symbolVectors);
+
 // Get the number of distinct symbols with CUDA managed or unified
 // attribute in the expression.
 template <typename A>
diff --git a/flang/lib/Evaluate/tools.cpp b/flang/lib/Evaluate/tools.cpp
index 8aff09aa84c13..c0bd89782ab33 100644
--- a/flang/lib/Evaluate/tools.cpp
+++ b/flang/lib/Evaluate/tools.cpp
@@ -1130,6 +1130,40 @@ template semantics::UnorderedSymbolSet CollectCudaSymbols(
 template semantics::UnorderedSymbolSet CollectCudaSymbols(
     const Expr<SubscriptInteger> &);
 
+std::vector<SymbolVector> GetSymbolVectors(const Expr<SomeType> &expr) {
+  SymbolVector symbols{GetSymbolVector(expr)};
+  std::reverse(symbols.begin(), symbols.end());
+
+  std::vector<SymbolVector> symbolVectors;
+
+  SymbolVector crtSymbols;
+  for (const Symbol &sym : symbols) {
+    bool isComponent{sym.owner().IsDerivedType()};
+    if (isComponent) {
+      crtSymbols.push_back(sym);
+    } else {
+      crtSymbols.push_back(sym);
+      symbolVectors.push_back(crtSymbols);
+      crtSymbols.clear();
+    }
+  }
+  return symbolVectors;
+}
+
+int GetNbOfUniqueCUDADeviceSymbols(
+    const std::vector<SymbolVector> &symbolVectors) {
+  semantics::UnorderedSymbolSet symbols;
+  for (const auto &symbolVector : symbolVectors) {
+    for (const auto &sym : symbolVector) {
+      if (IsCUDADeviceSymbol(*sym)) {
+        symbols.insert(sym);
+        break;
+      }
+    }
+  }
+  return symbols.size();
+}
+
 bool HasCUDAImplicitTransfer(const Expr<SomeType> &expr) {
   semantics::UnorderedSymbolSet hostSymbols;
   semantics::UnorderedSymbolSet deviceSymbols;
diff --git a/flang/lib/Semantics/check-cuda.cpp b/flang/lib/Semantics/check-cuda.cpp
index 62ece9e57b5d6..1ec233abb29fa 100644
--- a/flang/lib/Semantics/check-cuda.cpp
+++ b/flang/lib/Semantics/check-cuda.cpp
@@ -795,7 +795,8 @@ void CUDAChecker::Enter(const parser::AssignmentStmt &x) {
   }
 
   int nbLhs{evaluate::GetNbOfCUDADeviceSymbols(assign->lhs)};
-  int nbRhs{evaluate::GetNbOfCUDADeviceSymbols(assign->rhs)};
+  int nbRhs{evaluate::GetNbOfUniqueCUDADeviceSymbols(
+      evaluate::GetSymbolVectors(assign->rhs))};
   int nbRhsManaged{evaluate::GetNbOfCUDAManagedOrUnifiedSymbols(assign->rhs)};
 
   // device to host transfer with more than one device object on the rhs is not
diff --git a/flang/test/Lower/CUDA/cuda-data-transfer.cuf b/flang/test/Lower/CUDA/cuda-data-transfer.cuf
index 1d0e510c110ee..e510ce244050b 100644
--- a/flang/test/Lower/CUDA/cuda-data-transfer.cuf
+++ b/flang/test/Lower/CUDA/cuda-data-transfer.cuf
@@ -673,3 +673,29 @@ end subroutine
 
 ! CHECK-LABEL: func.func @_QPsub37()
 ! CHECK: cuf.data_transfer
+
+subroutine sub38()
+  type :: bar
+    integer, device, allocatable :: m(:)
+  end type
+  type(bar), unified :: a
+  type(bar), managed :: ma
+  integer :: lm(5)
+
+
+  lm(1:5) = a%m(1:5)
+  lm(1:5) = ma%m(1:5)
+end subroutine
+
+subroutine sub39()
+  type :: foo
+    integer, device, allocatable :: m(:)
+  end type
+  type :: bar
+    type(foo) :: f
+  end type
+  type(bar) :: a
+  integer :: lm(5)
+
+  lm(1:5) = a%f%m(1:5)
+end subroutine

``````````

</details>


https://github.com/llvm/llvm-project/pull/192177