[clang] [llvm] [TargetParser] Parse Intel GPU architecture names, and use the list in clang (PR #223199)

[Srividya Sundaram via llvm-commits]

================
@@ -73,18 +74,17 @@ OffloadArch StringToOffloadArch(llvm::StringRef S) {
     return OffloadArch::getAMDGCNSPIRV();
   if (S == "generic")
     return OffloadArch::getGeneric();
+  // Intel CPU offload has a single processor and no TargetParser list.
   if (S == "graniterapids")
-    return OffloadArch::getIntel(OffloadArch::TargetArch::IntelCPU,
-                                 OffloadArch::IntelArch::GRANITERAPIDS);
-  if (S == "bmg_g21")
-    return OffloadArch::getIntel(OffloadArch::TargetArch::IntelGPU,
-                                 OffloadArch::IntelArch::BMG_G21);
+    return OffloadArch::getIntelCPU();
 
   // Otherwise defer to the vendor TargetParser GPU lists.
   if (llvm::NVPTX::GPUKind NV = llvm::NVPTX::parseArch(S))
     return OffloadArch::getNVPTX(NV);
   if (llvm::AMDGPU::GPUKind AK = llvm::AMDGPU::parseArchAMDGCN(S))
     return OffloadArch::getAMDGPU(AK);
+  if (llvm::IntelGPU::GPUKind IK = llvm::IntelGPU::parseArch(S))
+    return OffloadArch::getIntelGPU(IK);
----------------
srividya-sundaram wrote:

`Map accepted Intel names to their ocloc device spelling at this boundary, retaining bmg_g21 support, and test the canonical-name invocation.`
Done.
The clang-sycl-linker now converts any accepted name into ocloc's spelling,` e.g. xe-bmg-g21 and xe_20.1.0 both become -device bmg_g21`, so AOT keeps working. Added tests. 
**Note**: ocloc is also adding the `xe-device` and `xe_version` names, so this mapping will go away later.

https://github.com/llvm/llvm-project/pull/223199