Consider calls from implict host device functions as valid in SemaCUDA

[jpienaar]

Summary

In SemaCUDA all implicit functions were considered host device, this led to errors such as the following code snippet failing to compile:

struct Copyable {
  const Copyable& operator=(const Copyable& x) { return *this; }
};

struct Simple {
  Copyable b;
};

void foo() {
  Simple a, b;

  a = b;
}

Above the implicit copy assignment operator was inferred as host device but there was only a host assignment copy defined which is an error in device compilation mode.

[rnk]

lib/Sema/SemaCUDA.cpp

54

Builtins are for things like __builtin_popcount(), but your test case is looking at implicitly defined C++ special members, right? What was happening on the test case you have previously?

test/SemaCUDA/implicit-copy.cu

16

This seems like another interesting test case:

struct Copyable {
  __device__ const Copyable& operator=(const Copyable& x) { return *this; }
};
struct Simple {
  Copyable b;
};
void foo(Simple &a, Simple &b) {
  a = b;
}

Simple's implicit copy ctor should be rejected when the host is the target because it calls a device-only method from the host, right?

I'm also curious what happens when someone uses FP math builtins like __builtin_cos that might have a reasonable lowering on a GPU.

[jpienaar]

lib/Sema/SemaCUDA.cpp

54

Yes, although that might be too restrictive.

I was getting the following errors:

error: object of type 'Simple' cannot be assigned because its copy assignment operator is implicitly deleted
  a = b;
    ^
note: copy assignment operator of 'Simple' is implicitly deleted because field 'b' has no copy assignment operator
  Copyable b;
           ^

test/SemaCUDA/implicit-copy.cu

16

Yes, and it does that. The one where there is a problem is with

__device__ void foo(Simple &a, Simple &b) {
  a = b
}

There an error is flagged. In general, as CheckCUDATarget only looks at Caller and Callee, will we run into a problem in cases such as below

Host function calls a implicit function which calls a Device function

Now, a host function calling an implicit function is fine, an implicit function calling a device function is fine, but transitively we have a host function calling a device function. But this all seem to be just used for the method candidate set and errors are still reported if this happens. So it would seem that we could allow implicits in here while still catching such erroneous cases.

I have created a diff where implicits are considered always OK, I'll append it.

[jpienaar]

Changed the inference to treat all implicit functions uniformly. Extended test to include copy constructor.

[jpienaar]

Avoided adding an additional attribute.

[rnk]

lgtm

lib/Sema/SemaCUDA.cpp

     return CFT_Device;
   } else if (D->hasAttr<CUDAHostAttr>()) {
     return CFT_Host;
-  } else if (D->isImplicit()) {
-    // Some implicit declarations (like intrinsic functions) are not marked.
-    // Set the most lenient target on them for maximal flexibility.
+  } else if (D->isImplicit() && D->getBuiltinID()) {
+    // Intrinsic functions cannot (at present) be marked as __device__ so set
+    // the most lenient target on them for maximal flexibility.
     return CFT_HostDevice;
   }
 

test/SemaCUDA/implicit-copy.cu

+// RUN: %clang_cc1 -std=gnu++11 -triple nvptx64-unknown-unknown -fcuda-is-device -fsyntax-only -verify %s
+
+struct Copyable {
+  const Copyable& operator=(const Copyable& x) { return *this; }
+};
+
+struct Simple {
+  Copyable b;
+};
+
+// expected-no-diagnostics
+void foo() {
+  Simple a, b;
+
+  a = b;
+}