[llvm] [AMDGPU] Split struct kernel arguments (PR #133786)

[Yaxun Liu via llvm-commits]

================
@@ -357,17 +357,38 @@ void MetadataStreamerMsgPackV4::emitKernelArg(const Argument &Arg,
   Align ArgAlign;
   std::tie(ArgTy, ArgAlign) = getArgumentTypeAlign(Arg, DL);
 
+  // Assuming the argument is not split from struct-type argument by default,
+  // unless we find it in function attribute amdgpu-argument-mapping.
+  unsigned OriginalArgIndex = ~0U;
+  uint64_t OriginalArgOffset = 0;
+  Attribute Attr =
+      Func->getAttributes().getParamAttr(ArgNo, "amdgpu-original-arg");
+  if (Attr.isValid()) {
+    StringRef MappingStr = Attr.getValueAsString();
+    SmallVector<StringRef, 2> Elements;
+    MappingStr.split(Elements, ':');
+    if (Elements.size() == 2) {
+      if (Elements[0].getAsInteger(10, OriginalArgIndex))
+        report_fatal_error(
----------------
yxsamliu wrote:

Done. Replaced `report_fatal_error` with an assert plus graceful fallback. The assert catches malformed attributes during development (assertions-enabled builds), while release builds silently ignore invalid attributes and fall back to default values (no original arg mapping).

https://github.com/llvm/llvm-project/pull/133786