[clang] [clang][SYCL] Add SYCL spelling for AS attributes (PR #200849)

[via cfe-commits]

================
@@ -312,3 +313,26 @@ void clang::takeAndConcatenateAttrs(ParsedAttributes &First,
   if (Second.Range.getEnd().isValid())
     First.Range.setEnd(Second.Range.getEnd());
 }
+
+LangAS ParsedAttr::asLangAS() const {
+  switch (getParsedKind()) {
+  case ParsedAttr::AT_OffloadGlobalAddressSpace:
+    return OffloadGlobalAddressSpaceAttr::getLangAS(*this);
+  case ParsedAttr::AT_OffloadGlobalDeviceAddressSpace:
+    return OffloadGlobalDeviceAddressSpaceAttr::getLangAS(*this);
+  case ParsedAttr::AT_OffloadGlobalHostAddressSpace:
+    return OffloadGlobalHostAddressSpaceAttr::getLangAS(*this);
+  case ParsedAttr::AT_OffloadLocalAddressSpace:
+    return OffloadLocalAddressSpaceAttr::getLangAS(*this);
+  case ParsedAttr::AT_OffloadPrivateAddressSpace:
+    return OffloadPrivateAddressSpaceAttr::getLangAS(*this);
+  case ParsedAttr::AT_OffloadConstantAddressSpace:
+    return OffloadConstantAddressSpaceAttr::getLangAS(*this);
+  case ParsedAttr::AT_OffloadGenericAddressSpace:
+    return OffloadGenericAddressSpaceAttr::getLangAS(*this);
+  case ParsedAttr::AT_HLSLGroupSharedAddressSpace:
+    return LangAS::hlsl_groupshared;
+  default:
+    return LangAS::Default;
+  }
+}
----------------
elizabethandrews wrote:

>> Would you find it strange if I suggested that we just use the CUDA attributes for SYCL? If so, how would that be different from using the OpenCL spellings?

I understand where you are coming from  but I don't think I would find it strange if the attributes are semantically doing the same thing, especially if we are reusing CUDA/openCL attributes.  I would definitely add a SYCL spelling for the attribute for users to use in SYCL code going forward but I personally wouldn't disallow the use of the old spellings if we weren't changing behavior.

>>The mapping from (logical) SYCL address spaces to the (physical) address spaces of the device targets should vary by target device.
I agree. This was how it was prior to the second commit in this PR. It depended on LangOpts. I may have misunderstood your request but I modified this behavior to use spellings instead based on your request during offline conversations. 

>>My goal is 1) clarity of intent, and 2) flexibility to handle SYCL concerns separately from OpenCL concerns.
As long as the mapping is done based on LangOpts,  this is possible. 

https://github.com/llvm/llvm-project/pull/200849