Reland [CUDA][HIP] Fix CTAD for host/device constructors (#168711)

Clang currently does not allow using CTAD in CUDA/HIP device functions
since deduction guides are treated as host-only. This patch fixes that
by treating deduction guides as host+device. The rationale is that
deduction guides do not actually generate code in IR, and there is an
existing check for device/host correctness for constructors.

The patch also suppresses duplicate implicit deduction guides from
host/device constructors with identical signatures and constraints
to prevent ambiguity.

For CUDA/HIP, deduction guides are now always implicitly enabled for
both host and device, which matches nvcc's effective behavior. Unlike
nvcc, which silently ignores explicit CUDA/HIP target attributes on
deduction guides, Clang diagnoses device- and host-only attributes
as errors to keep the syntax clean and avoid confusion. It emits a
deprecation warning for host+device attributes.

This ensures CTAD works correctly in CUDA/HIP for constructors with
different target attributes and provides clearer diagnostics when users
attempt to annotate deduction guides with CUDA/HIP target attributes.

Example:

```
  #include <tuple>

  __host__ __device__ void func()
  {
    std::tuple<int, int> t = std::tuple(1, 1);
  }
```

This compiles with nvcc but fails with clang for CUDA/HIP without this
fix.

Reference: https://godbolt.org/z/WhT1GrhWE

Fixes: ROCm/ROCm#5646

Fixes: #146646

Author: yxsamliu

clang/docs/HIPSupport.rst

@@ -412,6 +412,57 @@ Example Usage
    __host__ __device__ int Four(void) __attribute__((weak, alias("_Z6__Fourv")));
    __host__ __device__ float Four(float f) __attribute__((weak, alias("_Z6__Fourf")));
 
+C++17 Class Template Argument Deduction (CTAD) Support
+======================================================
+
+Clang supports C++17 Class Template Argument Deduction (CTAD) in both host and
+device code for HIP. This allows you to omit template arguments when creating
+class template instances, letting the compiler deduce them from constructor
+arguments.
+
+.. code-block:: c++
+
+   #include <tuple>
+
+   __host__ __device__ void func() {
+     std::tuple<int, int> t = std::tuple(1, 1);
+   }
+
+In the above example, ``std::tuple(1, 1)`` automatically deduces the type to be
+``std::tuple<int, int>``.
+
+Deduction Guides
+----------------
+
+User-defined deduction guides are also supported. Since deduction guides are not
+executable code and only participate in type deduction, they semantically behave
+as ``__host__ __device__``. This ensures they are available for deduction in both
+host and device contexts, and CTAD continues to respect any constraints on the
+corresponding constructors in the usual C++ way.
+
+.. code-block:: c++
+
+   template <typename T>
+   struct MyType {
+     T value;
+     __device__ MyType(T v) : value(v) {}
+   };
+
+   MyType(float) -> MyType<double>;
+
+   __device__ void deviceFunc() {
+     MyType m(1.0f); // Deduces MyType<double>
+   }
+
+.. note::
+
+   Explicit HIP target attributes such as ``__host__`` or ``__device__``
+   are currently only permitted on deduction guides when both are present
+   (``__host__ __device__``). This usage is deprecated and will be rejected
+   in a future version of Clang; prefer omitting HIP target attributes on
+   deduction guides entirely. Clang treats all deduction guides as if they
+   were ``__host__ __device__``, so ``__host__``-only, ``__device__``-only,
+   or ``__global__`` deduction guides are rejected as ill-formed.
 
 Host and Device Attributes of Default Destructors
 ===================================================

clang/docs/ReleaseNotes.rst

@@ -659,6 +659,23 @@ RISC-V Support
 CUDA/HIP Language Changes
 ^^^^^^^^^^^^^^^^^^^^^^^^^
 
+- Clang now supports C++17 Class Template Argument Deduction (CTAD) in CUDA/HIP
+  device code by treating deduction guides as if they were ``__host__ __device__``.
+
+- Clang avoids ambiguous CTAD in CUDA/HIP by not synthesizing duplicate implicit
+  deduction guides when ``__host__`` and ``__device__`` constructors differ only
+  in CUDA target attributes (same signature and constraints).
+
+- Clang diagnoses CUDA/HIP deduction guides that are annotated as host-only,
+  device-only, or ``__global__`` as errors. Explicit ``__host__ __device__``
+  deduction guides remain accepted for now but are deprecated and will be
+  rejected in a future version of Clang; deduction guides do not participate
+  in code generation and are treated as implicitly host+device.
+
+- Clang preserves distinct implicit deduction guides for constructors that differ
+  by constraints, so constraint-based CTAD works in CUDA/HIP device code as in
+  standard C++.
+
 CUDA Support
 ^^^^^^^^^^^^
 

clang/include/clang/Basic/DiagnosticSemaKinds.td

@@ -2769,6 +2769,14 @@ def err_deduction_guide_name_not_class_template : Error<
   "cannot specify deduction guide for "
   "%select{<error>|function template|variable template|alias template|"
   "template template parameter|concept|dependent template name}0 %1">;
+def err_deduction_guide_target_attr : Error<
+  "in CUDA/HIP, deduction guides may only be annotated with "
+  "'__host__ __device__'; '__host__'-only, '__device__'-only, or "
+  "'__global__' deduction guides are not allowed">;
+def warn_deduction_guide_target_attr_deprecated : Warning<
+  "use of CUDA/HIP target attributes on deduction guides is deprecated; "
+  "they will be rejected in a future version of Clang">,
+  InGroup<DeprecatedAttributes>;
 def err_deduction_guide_wrong_scope : Error<
   "deduction guide must be declared in the same scope as template %q0">;
 def err_deduction_guide_defines_function : Error<

clang/lib/Sema/SemaCUDA.cpp

@@ -215,6 +215,12 @@ CUDAFunctionTarget SemaCUDA::IdentifyTarget(const FunctionDecl *D,
   if (D == nullptr)
     return CurCUDATargetCtx.Target;
 
+  // C++ deduction guides are never codegen'ed and only participate in template
+  // argument deduction.  Treat them as if they were always host+device so that
+  // CUDA/HIP target checking never rejects their use based solely on target.
+  if (isa<CXXDeductionGuideDecl>(D))
+    return CUDAFunctionTarget::HostDevice;
+
   if (D->hasAttr<CUDAInvalidTargetAttr>())
     return CUDAFunctionTarget::InvalidTarget;
 
@@ -986,6 +992,12 @@ bool SemaCUDA::CheckCall(SourceLocation Loc, FunctionDecl *Callee) {
   if (ExprEvalCtx.isUnevaluated() || ExprEvalCtx.isConstantEvaluated())
     return true;
 
+  // C++ deduction guides participate in overload resolution but are not
+  // callable functions and are never codegen'ed.  Treat them as always
+  // allowed for CUDA/HIP compatibility checking.
+  if (isa<CXXDeductionGuideDecl>(Callee))
+    return true;
+
   // FIXME: Is bailing out early correct here?  Should we instead assume that
   // the caller is a global initializer?
   FunctionDecl *Caller = SemaRef.getCurFunctionDecl(/*AllowLambda=*/true);

clang/lib/Sema/SemaDeclAttr.cpp

@@ -8056,6 +8056,30 @@ void Sema::ProcessDeclAttributeList(
     }
   }
 
+  // CUDA/HIP: restrict explicit CUDA target attributes on deduction guides.
+  //
+  // Deduction guides are not callable functions and never participate in
+  // codegen; they are always treated as host+device for CUDA/HIP semantic
+  // checks.  We therefore allow either no CUDA target attributes or an explicit
+  // '__host__ __device__' annotation, but reject guides that are host-only,
+  // device-only, or marked '__global__'.  The use of explicit CUDA/HIP target
+  // attributes on deduction guides is deprecated and will be rejected in a
+  // future Clang version.
+  if (getLangOpts().CUDA)
+    if (auto *Guide = dyn_cast<CXXDeductionGuideDecl>(D)) {
+      bool HasHost = Guide->hasAttr<CUDAHostAttr>();
+      bool HasDevice = Guide->hasAttr<CUDADeviceAttr>();
+      bool HasGlobal = Guide->hasAttr<CUDAGlobalAttr>();
+
+      if (HasGlobal || HasHost != HasDevice) {
+        Diag(Guide->getLocation(), diag::err_deduction_guide_target_attr);
+        Guide->setInvalidDecl();
+      } else if (HasHost && HasDevice) {
+        Diag(Guide->getLocation(),
+             diag::warn_deduction_guide_target_attr_deprecated);
+      }
+    }
+
   // Do not permit 'constructor' or 'destructor' attributes on __device__ code.
   if (getLangOpts().CUDAIsDevice && D->hasAttr<CUDADeviceAttr>() &&
       (D->hasAttr<ConstructorAttr>() || D->hasAttr<DestructorAttr>()) &&

clang/lib/Sema/SemaTemplateDeductionGuide.cpp

@@ -54,6 +54,26 @@ using namespace clang;
 using namespace sema;
 
 namespace {
+
+/// Return true if two associated-constraint sets are semantically equal.
+static bool HaveSameAssociatedConstraints(
+    Sema &SemaRef, const NamedDecl *Old, ArrayRef<AssociatedConstraint> OldACs,
+    const NamedDecl *New, ArrayRef<AssociatedConstraint> NewACs) {
+  if (OldACs.size() != NewACs.size())
+    return false;
+  if (OldACs.empty())
+    return true;
+
+  // General case: pairwise compare each associated constraint expression.
+  Sema::TemplateCompareNewDeclInfo NewInfo(New);
+  for (size_t I = 0, E = OldACs.size(); I != E; ++I)
+    if (!SemaRef.AreConstraintExpressionsEqual(
+            Old, OldACs[I].ConstraintExpr, NewInfo, NewACs[I].ConstraintExpr))
+      return false;
+
+  return true;
+}
+
 /// Tree transform to "extract" a transformed type from a class template's
 /// constructor to a deduction guide.
 class ExtractTypeForDeductionGuide
@@ -218,9 +238,51 @@ buildDeductionGuide(Sema &SemaRef, TemplateDecl *OriginalTemplate,
       TInfo->getTypeLoc().castAs<FunctionProtoTypeLoc>().getParams();
 
   // Build the implicit deduction guide template.
+  QualType GuideType = TInfo->getType();
+
+  // In CUDA/HIP mode, avoid duplicate implicit guides that differ only in CUDA
+  // target attributes (same constructor signature and constraints).
+  if (IsImplicit && Ctor && SemaRef.getLangOpts().CUDA) {
+    SmallVector<AssociatedConstraint, 4> NewACs;
+    Ctor->getAssociatedConstraints(NewACs);
+
+    for (NamedDecl *Existing : DC->lookup(DeductionGuideName)) {
+      auto *ExistingFT = dyn_cast<FunctionTemplateDecl>(Existing);
+      auto *ExistingGuide =
+          ExistingFT
+              ? dyn_cast<CXXDeductionGuideDecl>(ExistingFT->getTemplatedDecl())
+              : dyn_cast<CXXDeductionGuideDecl>(Existing);
+      if (!ExistingGuide)
+        continue;
+
+      // Only consider guides that were also synthesized from a constructor.
+      auto *ExistingCtor = ExistingGuide->getCorrespondingConstructor();
+      if (!ExistingCtor)
+        continue;
+
+      // If the underlying constructors are overloads (different signatures once
+      // CUDA attributes are ignored), they should each get their own guides.
+      if (SemaRef.IsOverload(Ctor, ExistingCtor,
+                             /*UseMemberUsingDeclRules=*/false,
+                             /*ConsiderCudaAttrs=*/false))
+        continue;
+
+      // At this point, the constructors have the same signature ignoring CUDA
+      // attributes. Decide whether their associated constraints are also the
+      // same; only in that case do we treat one guide as a duplicate of the
+      // other.
+      SmallVector<AssociatedConstraint, 4> ExistingACs;
+      ExistingCtor->getAssociatedConstraints(ExistingACs);
+
+      if (HaveSameAssociatedConstraints(SemaRef, ExistingCtor, ExistingACs,
+                                        Ctor, NewACs))
+        return Existing;
+    }
+  }
+
   auto *Guide = CXXDeductionGuideDecl::Create(
-      SemaRef.Context, DC, LocStart, ES, Name, TInfo->getType(), TInfo, LocEnd,
-      Ctor, DeductionCandidate::Normal, FunctionTrailingRC);
+      SemaRef.Context, DC, LocStart, ES, Name, GuideType, TInfo, LocEnd, Ctor,
+      DeductionCandidate::Normal, FunctionTrailingRC);
   Guide->setImplicit(IsImplicit);
   Guide->setParams(Params);
 

clang/test/SemaCUDA/deduction-guide-attrs.cu

@@ -0,0 +1,32 @@
+// RUN: %clang_cc1 -std=c++17 -triple nvptx64-nvidia-cuda -fsyntax-only \
+// RUN:            -fcuda-is-device -verify %s
+// RUN: %clang_cc1 -std=c++17 -triple nvptx64-nvidia-cuda -fsyntax-only \
+// RUN:            -verify %s
+
+#include "Inputs/cuda.h"
+
+template <typename T>
+struct S {
+  __host__ __device__ S(T);
+};
+
+// A host+device deduction guide is allowed and participates in CTAD, but its
+// explicit target attributes are deprecated and will be rejected in a future
+// Clang version.
+template <typename T>
+__host__ __device__ S(T) -> S<T>; // expected-warning {{use of CUDA/HIP target attributes on deduction guides is deprecated; they will be rejected in a future version of Clang}}
+
+__host__ __device__ void use_hd_guide() {
+  S s(42); // uses the explicit __host__ __device__ deduction guide above
+}
+
+// CUDA/HIP target attributes on deduction guides are rejected when they make
+// the guide host-only, device-only, or a kernel.
+template <typename U>
+__host__ S(U) -> S<U>;   // expected-error {{in CUDA/HIP, deduction guides may only be annotated with '__host__ __device__'; '__host__'-only, '__device__'-only, or '__global__' deduction guides are not allowed}}
+
+template <typename V>
+__device__ S(V) -> S<V>; // expected-error {{in CUDA/HIP, deduction guides may only be annotated with '__host__ __device__'; '__host__'-only, '__device__'-only, or '__global__' deduction guides are not allowed}}
+
+template <typename W>
+__global__ S(W) -> S<W>; // expected-error {{in CUDA/HIP, deduction guides may only be annotated with '__host__ __device__'; '__host__'-only, '__device__'-only, or '__global__' deduction guides are not allowed}}

clang/test/SemaCUDA/deduction-guide-overload.cu

@@ -0,0 +1,111 @@
+// RUN: %clang_cc1 -std=c++20 -triple nvptx64-nvidia-cuda -fsyntax-only \
+// RUN:            -fcuda-is-device -verify %s
+// RUN: %clang_cc1 -std=c++20 -triple nvptx64-nvidia-cuda -fsyntax-only \
+// RUN:            -verify %s
+// expected-no-diagnostics
+
+#include "Inputs/cuda.h"
+
+// This test exercises class template argument deduction (CTAD) when there are
+// multiple constructors that differ only by constraints. In CUDA/HIP mode, the
+// implementation must *not* collapse implicit deduction guides that have the
+// same function type but different constraints; otherwise, CTAD can lose viable
+// candidates.
+
+template <typename T>
+concept Signed = __is_signed(T);
+
+template <typename T>
+concept NotSigned = !Signed<T>;
+
+// 1) Constrained ctors with different constraints: ensure we keep
+// deduction guides that differ only by constraints.
+
+template <typename T>
+struct OverloadCTAD {
+  __host__ __device__ OverloadCTAD(T) requires Signed<T>;
+  __host__ __device__ OverloadCTAD(T) requires NotSigned<T>;
+};
+
+__host__ __device__ void use_overload_ctad_hd() {
+  OverloadCTAD a(1);   // T = int, uses Signed-constrained guide
+  OverloadCTAD b(1u);  // T = unsigned int, uses NotSigned-constrained guide
+}
+
+__device__ void use_overload_ctad_dev() {
+  OverloadCTAD c(1);
+  OverloadCTAD d(1u);
+}
+
+__global__ void use_overload_ctad_global() {
+  OverloadCTAD e(1);
+  OverloadCTAD f(1u);
+}
+
+// 2) Add a pair of constructors that have the same signature and the same
+// constraint but differ only by CUDA target attributes. This exercises the
+// case where two implicit deduction guides would be identical except for
+// their originating constructor's CUDA target.
+
+template <typename T>
+struct OverloadCTADTargets {
+  __host__ OverloadCTADTargets(T) requires Signed<T>;
+  __device__ OverloadCTADTargets(T) requires Signed<T>;
+};
+
+__host__ void use_overload_ctad_targets_host() {
+  OverloadCTADTargets g(1);
+}
+
+__device__ void use_overload_ctad_targets_device() {
+  OverloadCTADTargets h(1);
+}
+
+// 3) Unconstrained host/device duplicates: identical signatures and no
+// constraints, differing only by CUDA target attributes.
+
+template <typename T>
+struct UnconstrainedHD {
+  __host__ UnconstrainedHD(T);
+  __device__ UnconstrainedHD(T);
+};
+
+__host__ __device__ void use_unconstrained_hd_hd() {
+  UnconstrainedHD u1(1);
+}
+
+__device__ void use_unconstrained_hd_dev() {
+  UnconstrainedHD u2(1);
+}
+
+__global__ void use_unconstrained_hd_global() {
+  UnconstrainedHD u3(1);
+}
+
+// 4) Constrained vs unconstrained ctors with the same signature: guides
+// must not be collapsed away when constraints differ.
+
+template <typename T>
+concept IsInt = __is_same(T, int);
+
+template <typename T>
+struct ConstrainedVsUnconstrained {
+  __host__ __device__ ConstrainedVsUnconstrained(T);
+  __host__ __device__ ConstrainedVsUnconstrained(T) requires IsInt<T>;
+};
+
+__host__ __device__ void use_constrained_vs_unconstrained_hd() {
+  ConstrainedVsUnconstrained a(1);    // T = int, constrained guide viable
+  ConstrainedVsUnconstrained b(1u);   // T = unsigned, only unconstrained guide
+}
+
+__device__ void use_constrained_vs_unconstrained_dev() {
+  ConstrainedVsUnconstrained c(1);
+  ConstrainedVsUnconstrained d(1u);
+}
+
+__global__ void use_constrained_vs_unconstrained_global() {
+  ConstrainedVsUnconstrained e(1);
+  ConstrainedVsUnconstrained f(1u);
+}
+