[clang] [llvm] [OffloadBundler] Rework the ctor of `OffloadTargetInfo` to support AMDGPU's generic target (PR #122629)

[Shilei Tian via llvm-commits]

================
@@ -84,31 +84,27 @@ OffloadTargetInfo::OffloadTargetInfo(const StringRef Target,
     : BundlerConfig(BC) {
 
   // TODO: Add error checking from ClangOffloadBundler.cpp
-  auto TargetFeatures = Target.split(':');
-  auto TripleOrGPU = TargetFeatures.first.rsplit('-');
-
-  if (clang::StringToOffloadArch(TripleOrGPU.second) !=
-      clang::OffloadArch::UNKNOWN) {
-    auto KindTriple = TripleOrGPU.first.split('-');
-    this->OffloadKind = KindTriple.first;
-
-    // Enforce optional env field to standardize bundles
-    llvm::Triple t = llvm::Triple(KindTriple.second);
-    this->Triple = llvm::Triple(t.getArchName(), t.getVendorName(),
-                                t.getOSName(), t.getEnvironmentName());
-
-    this->TargetID = Target.substr(Target.find(TripleOrGPU.second));
-  } else {
-    auto KindTriple = TargetFeatures.first.split('-');
-    this->OffloadKind = KindTriple.first;
-
-    // Enforce optional env field to standardize bundles
-    llvm::Triple t = llvm::Triple(KindTriple.second);
-    this->Triple = llvm::Triple(t.getArchName(), t.getVendorName(),
-                                t.getOSName(), t.getEnvironmentName());
-
+  // <kind>-<triple>[-<target id>[:target features]]
+  // <triple> := <arch>-<vendor>-<os>-<env>
+  SmallVector<StringRef, 6> Components;
+  Target.split(Components, '-', /*MaxSplit=*/5);
+  assert((Components.size() == 5 || Components.size() == 6) &&
----------------
shiltian wrote:

> We have users who executes clang-offload-bundler directly.

Not sure about people outside of ROCm stack. Every one in ROCm stack has already moved to 4-field triple.

This code is inside a constructor such that we can't emit an error here. If we really want to do that, we will need to duplicate the logic to somewhere else and emit an error there.

https://github.com/llvm/llvm-project/pull/122629