[PATCH] D123471: [CUDA] Create offloading entries when using the new driver

Fri May 6 11:38:50 PDT 2022

jhuber6 added inline comments.

================
Comment at: clang/lib/CodeGen/CGCUDARuntime.h:58-70
+      OffloadRegionKernelEntry = 0x0,
+    };
+
+    /// The kind flag of the global variable entry.
+    enum OffloadVarEntryKindFlag : uint32_t {
+      /// Mark the entry as a global variable.
+      OffloadGlobalVarEntry = 0x0,
----------------
tra wrote:
> I'm a bit puzzled by this arrangement. Are those actually flags (i.e. can be set independently) or are they enumerating specific offload kinds (i.e. only one of these values is intended to be set)?
> 
> I think we want the latter. If that's the case I'd propose to enumerate kernel and data together, so each kind gets a distinct value and is easy to tell when one needs to examine the offload table manually. Right now both kernels and global vars set the flags to 0.
> 
It probably should just be an enumeration. I was tentatively keeping them somewhat separate because OpenMP uses different values for these flags, but I think keeping this completely compatible is an impossible proposition. If we need them to use the same flag we should be able to configure that at some point. I will change it to just be a standard enum (I don't handle anything but kernels and regular globals in the linker wrapper right now anyway)

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D123471/new/

https://reviews.llvm.org/D123471