[HIP][SPIRV] Implicit new/delete should be cdecl on host (#152023)

Client apps can (and in the case of the MSVC STL do) set cdecl
explicitly on `new` / `delete` implementations. On the other hand, Clang
generates implicit decls with the target's default CC. This is
problematic for SPIR-V, since the default there is spir_function, which
is not compatible. Since we cannot change pre-existing headers /
implementations, this patch sets the CC to C for the implicit host-side
decls, when compiling for device, to prevent the conflict. This is fine
because the host-side overloards do not get emitted on device.

Author: AlexVlx

clang/lib/Sema/SemaExprCXX.cpp

@@ -3497,6 +3497,19 @@ void Sema::DeclareGlobalAllocationFunction(DeclarationName Name,
   }
 
   auto CreateAllocationFunctionDecl = [&](Attr *ExtraAttr) {
+    // The MSVC STL has explicit cdecl on its (host-side) allocation function
+    // specializations for the allocation, so in order to prevent a CC clash
+    // we use the host's CC, if available, or CC_C as a fallback, for the
+    // host-side implicit decls, knowing these do not get emitted when compiling
+    // for device.
+    if (getLangOpts().CUDAIsDevice && ExtraAttr &&
+        isa<CUDAHostAttr>(ExtraAttr) &&
+        Context.getTargetInfo().getTriple().isSPIRV()) {
+      if (auto *ATI = Context.getAuxTargetInfo())
+        EPI.ExtInfo = EPI.ExtInfo.withCallingConv(ATI->getDefaultCallingConv());
+      else
+        EPI.ExtInfo = EPI.ExtInfo.withCallingConv(CallingConv::CC_C);
+    }
     QualType FnType = Context.getFunctionType(Return, Params, EPI);
     FunctionDecl *Alloc = FunctionDecl::Create(
         Context, GlobalCtx, SourceLocation(), SourceLocation(), Name, FnType,

clang/test/SemaHIP/amdgcnspirv-implicit-alloc-function-calling-conv.hip

@@ -0,0 +1,32 @@
+// RUN: %clang_cc1 %s -fcuda-is-device -std=c++17 -triple spirv32 -verify
+// RUN: %clang_cc1 %s -fcuda-is-device -std=c++17 -triple spirv64 -verify
+// RUN: %clang_cc1 %s -fcuda-is-device -std=c++17 -triple spirv64-amd-amdhsa -verify
+// RUN: %clang_cc1 %s -fcuda-is-device -std=c++17 -triple spirv32 -aux-triple x86_64-unknown-linux-gnu -verify
+// RUN: %clang_cc1 %s -fcuda-is-device -std=c++17 -triple spirv64 -aux-triple x86_64-unknown-linux-gnu -verify
+// RUN: %clang_cc1 %s -fcuda-is-device -std=c++17 -triple spirv64-amd-amdhsa -aux-triple x86_64-unknown-linux-gnu -verify
+// RUN: %clang_cc1 %s -fcuda-is-device -std=c++17 -triple spirv32 -aux-triple x86_64-pc-windows-msvc -verify
+// RUN: %clang_cc1 %s -fcuda-is-device -std=c++17 -triple spirv64 -aux-triple x86_64-pc-windows-msvc -verify
+// RUN: %clang_cc1 %s -fcuda-is-device -std=c++17 -triple spirv64-amd-amdhsa -aux-triple x86_64-pc-windows-msvc -verify
+
+// expected-no-diagnostics
+
+namespace std
+{
+    enum class align_val_t : __SIZE_TYPE__ {};
+    struct nothrow_t { explicit nothrow_t() = default; };
+    extern nothrow_t const nothrow;
+}
+
+void* __attribute__((cdecl)) operator new(__SIZE_TYPE__);
+void* __attribute__((cdecl)) operator new[](__SIZE_TYPE__);
+void* __attribute__((cdecl)) operator new(__SIZE_TYPE__, ::std::align_val_t);
+void* __attribute__((cdecl)) operator new[](__SIZE_TYPE__, ::std::align_val_t);
+
+void __attribute__((cdecl)) operator delete(void*) noexcept;
+void __attribute__((cdecl)) operator delete[](void*) noexcept;
+void __attribute__((cdecl)) operator delete(void*, __SIZE_TYPE__) noexcept;
+void __attribute__((cdecl)) operator delete[](void*, __SIZE_TYPE__) noexcept;
+void __attribute__((cdecl)) operator delete(void*, ::std::align_val_t) noexcept;
+void __attribute__((cdecl)) operator delete[](void*, ::std::align_val_t) noexcept;
+void __attribute__((cdecl)) operator delete(void*, __SIZE_TYPE__, ::std::align_val_t) noexcept;
+void __attribute__((cdecl)) operator delete[](void*, __SIZE_TYPE__, ::std::align_val_t) noexcept;