[CUDA][HIP] capture possible ODR-used var (llvm#136645)

yxsamliu · yxsamliu · commit 3a0e77d39c5c · 2025-08-05T09:56:02.000-04:00
In a lambda function, a call of a function may resolve to host and device functions with different signatures. Especially, a constexpr local variable may be passed by value by the device function and passed by reference by the host function, which will cause the constexpr variable captured by the lambda function in host compilation but not in the device compilation. The discrepancy in the lambda captures will violate ODR and causes UB for kernels using these lambdas. This PR fixes the issue by identifying discrepancy of ODR/non-ODR usages of constexpr local variables passed to host/device functions and conservatively capture them. Fixes: llvm#132068 c-p: llvm#136645 Fixes: SWDEV-522914
diff --git a/clang/include/clang/Sema/ScopeInfo.h b/clang/include/clang/Sema/ScopeInfo.h
@@ -949,6 +949,9 @@ class LambdaScopeInfo final :
 
   SourceLocation PotentialThisCaptureLocation;
 
+  /// Variables that are potentially ODR-used in CUDA/HIP.
+  llvm::SmallPtrSet<VarDecl *, 4> CUDAPotentialODRUsedVars;
+
   LambdaScopeInfo(DiagnosticsEngine &Diag)
       : CapturingScopeInfo(Diag, ImpCap_None) {
     Kind = SK_Lambda;
diff --git a/clang/include/clang/Sema/SemaCUDA.h b/clang/include/clang/Sema/SemaCUDA.h
@@ -274,6 +274,10 @@ class SemaCUDA : public SemaBase {
   /// parameters specified via <<<>>>.
   std::string getConfigureFuncName() const;
 
+  /// Record variables that are potentially ODR-used in CUDA/HIP.
+  void recordPotentialODRUsedVariable(MultiExprArg Args,
+                                      OverloadCandidateSet &CandidateSet);
+
 private:
   unsigned ForceHostDeviceDepth = 0;
 
diff --git a/clang/lib/Sema/SemaCUDA.cpp b/clang/lib/Sema/SemaCUDA.cpp
@@ -18,6 +18,7 @@
 #include "clang/Basic/TargetInfo.h"
 #include "clang/Lex/Preprocessor.h"
 #include "clang/Sema/Lookup.h"
+#include "clang/Sema/Overload.h"
 #include "clang/Sema/ScopeInfo.h"
 #include "clang/Sema/Sema.h"
 #include "clang/Sema/Template.h"
@@ -1081,3 +1082,49 @@ std::string SemaCUDA::getConfigureFuncName() const {
   // Legacy CUDA kernel configuration call
   return "cudaConfigureCall";
 }
+
+// Record any local constexpr variables that are passed one way on the host
+// and another on the device.
+void SemaCUDA::recordPotentialODRUsedVariable(
+    MultiExprArg Arguments, OverloadCandidateSet &Candidates) {
+  sema::LambdaScopeInfo *LambdaInfo = SemaRef.getCurLambda();
+  if (!LambdaInfo)
+    return;
+
+  for (unsigned I = 0; I < Arguments.size(); ++I) {
+    auto *DeclRef = dyn_cast<DeclRefExpr>(Arguments[I]);
+    if (!DeclRef)
+      continue;
+    auto *Variable = dyn_cast<VarDecl>(DeclRef->getDecl());
+    if (!Variable || !Variable->isLocalVarDecl() || !Variable->isConstexpr())
+      continue;
+
+    bool HostByValue = false, HostByRef = false;
+    bool DeviceByValue = false, DeviceByRef = false;
+
+    for (OverloadCandidate &Candidate : Candidates) {
+      FunctionDecl *Callee = Candidate.Function;
+      if (!Callee || I >= Callee->getNumParams())
+        continue;
+
+      CUDAFunctionTarget Target = IdentifyTarget(Callee);
+      if (Target == CUDAFunctionTarget::InvalidTarget ||
+          Target == CUDAFunctionTarget::Global)
+        continue;
+
+      bool CoversHost = (Target == CUDAFunctionTarget::Host ||
+                         Target == CUDAFunctionTarget::HostDevice);
+      bool CoversDevice = (Target == CUDAFunctionTarget::Device ||
+                           Target == CUDAFunctionTarget::HostDevice);
+
+      bool IsRef = Callee->getParamDecl(I)->getType()->isReferenceType();
+      HostByValue |= CoversHost && !IsRef;
+      HostByRef |= CoversHost && IsRef;
+      DeviceByValue |= CoversDevice && !IsRef;
+      DeviceByRef |= CoversDevice && IsRef;
+    }
+
+    if ((HostByValue && DeviceByRef) || (HostByRef && DeviceByValue))
+      LambdaInfo->CUDAPotentialODRUsedVars.insert(Variable);
+  }
+}
diff --git a/clang/lib/Sema/SemaExpr.cpp b/clang/lib/Sema/SemaExpr.cpp
@@ -19423,11 +19423,29 @@ static ExprResult rebuildPotentialResultsAsNonOdrUsed(Sema &S, Expr *E,
     return false;
   };
 
+  // Check whether this expression may be odr-used in CUDA/HIP.
+  auto MaybeCUDAODRUsed = [&]() -> bool {
+    if (!S.LangOpts.CUDA)
+      return false;
+    LambdaScopeInfo *LSI = S.getCurLambda();
+    if (!LSI)
+      return false;
+    auto *DRE = dyn_cast<DeclRefExpr>(E);
+    if (!DRE)
+      return false;
+    auto *VD = dyn_cast<VarDecl>(DRE->getDecl());
+    if (!VD)
+      return false;
+    return LSI->CUDAPotentialODRUsedVars.count(VD);
+  };
+
   // Mark that this expression does not constitute an odr-use.
   auto MarkNotOdrUsed = [&] {
-    S.MaybeODRUseExprs.remove(E);
-    if (LambdaScopeInfo *LSI = S.getCurLambda())
-      LSI->markVariableExprAsNonODRUsed(E);
+    if (!MaybeCUDAODRUsed()) {
+      S.MaybeODRUseExprs.remove(E);
+      if (LambdaScopeInfo *LSI = S.getCurLambda())
+        LSI->markVariableExprAsNonODRUsed(E);
+    }
   };
 
   // C++2a [basic.def.odr]p2:
diff --git a/clang/lib/Sema/SemaOverload.cpp b/clang/lib/Sema/SemaOverload.cpp
@@ -14270,6 +14270,8 @@ ExprResult Sema::BuildOverloadedCallExpr(Scope *S, Expr *Fn,
   // the UnresolvedLookupExpr was type-dependent.
   if (OverloadResult == OR_Success) {
     const FunctionDecl *FDecl = Best->Function;
+    if (LangOpts.CUDA)
+      CUDA().recordPotentialODRUsedVariable(Args, CandidateSet);
     if (FDecl && FDecl->isTemplateInstantiation() &&
         FDecl->getReturnType()->isUndeducedType()) {
       if (const auto *TP =
diff --git a/clang/test/CodeGenCUDA/lambda-constexpr-capture.cu b/clang/test/CodeGenCUDA/lambda-constexpr-capture.cu
@@ -0,0 +1,135 @@
+// RUN: %clang_cc1 -emit-llvm -x hip %s -o - -triple x86_64-linux-gnu \
+// RUN:   | FileCheck -check-prefixes=CHECK,HOST %s
+// RUN: %clang_cc1 -emit-llvm -x hip %s -o - -triple amdgcn-amd-amdhsa -fcuda-is-device \
+// RUN:   | FileCheck -check-prefixes=CHECK,DEV %s
+
+#include "Inputs/cuda.h"
+
+// CHECK: %class.anon = type { ptr, float, ptr, ptr }
+// CHECK: %class.anon.0 = type { ptr, float, ptr, ptr }
+// CHECK: %class.anon.1 = type { ptr, ptr, ptr }
+// CHECK: %class.anon.2 = type { ptr, float, ptr, ptr }
+
+// HOST: call void @_ZN8DevByVal21__device_stub__kernelIZNS_4testEPKfS2_PfEUljE_EEvT_(ptr noundef byval(%class.anon)
+// DEV: define amdgpu_kernel void @_ZN8DevByVal6kernelIZNS_4testEPKfS2_PfEUljE_EEvT_(ptr addrspace(4) noundef byref(%class.anon)
+
+// Only the device function passes arugments by value.
+namespace DevByVal {
+__device__ float fun(float x, float y) {
+  return x;
+}
+
+float fun(const float &x, const float &y) {
+  return x;
+}
+
+template<typename F>
+void __global__ kernel(F f)
+{
+  f(1);
+}
+
+void test(float const * fl, float const * A, float * Vf)
+{
+  float constexpr small(1.0e-25);
+
+  auto lambda = [=] __device__ __host__ (unsigned int n) {
+    float const value = fun(small, fl[0]);
+    Vf[0] = value * A[0];
+  };
+  kernel<<<1, 1>>>(lambda);
+}
+}
+
+// HOST: call void @_ZN9HostByVal21__device_stub__kernelIZNS_4testEPKfS2_PfEUljE_EEvT_(ptr noundef byval(%class.anon.0)
+// DEV: define amdgpu_kernel void @_ZN9HostByVal6kernelIZNS_4testEPKfS2_PfEUljE_EEvT_(ptr addrspace(4) noundef byref(%class.anon.0)
+
+// Only the host function passes arugments by value.
+namespace HostByVal {
+float fun(float x, float y) {
+  return x;
+}
+
+__device__ float fun(const float &x, const float &y) {
+  return x;
+}
+
+template<typename F>
+void __global__ kernel(F f)
+{
+  f(1);
+}
+
+void test(float const * fl, float const * A, float * Vf)
+{
+  float constexpr small(1.0e-25);
+
+  auto lambda = [=] __device__ __host__ (unsigned int n) {
+    float const value = fun(small, fl[0]);
+    Vf[0] = value * A[0];
+  };
+  kernel<<<1, 1>>>(lambda);
+}
+}
+
+// HOST: call void @_ZN9BothByVal21__device_stub__kernelIZNS_4testEPKfS2_PfEUljE_EEvT_(ptr noundef byval(%class.anon.1)
+// DEV: define amdgpu_kernel void @_ZN9BothByVal6kernelIZNS_4testEPKfS2_PfEUljE_EEvT_(ptr addrspace(4) noundef byref(%class.anon.1)
+
+// Both the host and device functions pass arugments by value.
+namespace BothByVal {
+float fun(float x, float y) {
+  return x;
+}
+
+__device__ float fun(float x, float y) {
+  return x;
+}
+
+template<typename F>
+void __global__ kernel(F f)
+{
+  f(1);
+}
+
+void test(float const * fl, float const * A, float * Vf)
+{
+  float constexpr small(1.0e-25);
+
+  auto lambda = [=] __device__ __host__ (unsigned int n) {
+    float const value = fun(small, fl[0]);
+    Vf[0] = value * A[0];
+  };
+  kernel<<<1, 1>>>(lambda);
+}
+}
+
+// HOST: call void @_ZN12NeitherByVal21__device_stub__kernelIZNS_4testEPKfS2_PfEUljE_EEvT_(ptr noundef byval(%class.anon.2)
+// DEV: define amdgpu_kernel void @_ZN12NeitherByVal6kernelIZNS_4testEPKfS2_PfEUljE_EEvT_(ptr addrspace(4) noundef byref(%class.anon.2)
+
+// Neither the host nor device function passes arugments by value.
+namespace NeitherByVal {
+float fun(const float& x, const float& y) {
+  return x;
+}
+
+__device__ float fun(const float& x, const float& y) {
+  return x;
+}
+
+template<typename F>
+void __global__ kernel(F f)
+{
+  f(1);
+}
+
+void test(float const * fl, float const * A, float * Vf)
+{
+  float constexpr small(1.0e-25);
+
+  auto lambda = [=] __device__ __host__ (unsigned int n) {
+    float const value = fun(small, fl[0]);
+    Vf[0] = value * A[0];
+  };
+  kernel<<<1, 1>>>(lambda);
+}
+}
