[llvm] Add the "initializes" attribute inference (PR #97373)

[Jan Voung via llvm-commits]

================
@@ -580,6 +582,205 @@ struct ArgumentUsesTracker : public CaptureTracker {
   const SCCNodeSet &SCCNodes;
 };
 
+struct ArgumentUse {
+  Use *U;
+  std::optional<int64_t> Offset;
+};
+
+// A struct of argument access info. "Unknown" accesses are the cases like
+// unrecognized instructions, instructions that have more than one use of
+// the argument, or volatile memory accesses. "Unknown" implies "IsClobber"
+// and an empty access range.
+// Write or Read accesses can be clobbers as well for example, a Load with
+// scalable type.
+struct ArgumentAccessInfo {
+  enum AccessType { Write, Read, Unknown };
+  AccessType ArgAccessType;
+  ConstantRangeList AccessRanges;
+  bool IsClobber = false;
+};
+
+struct UsesPerBlockInfo {
+  DenseMap<Instruction *, ArgumentAccessInfo> Insts;
+  bool HasWrites;
+  bool HasClobber;
+};
+
+ArgumentAccessInfo GetArgmentAccessInfo(const Instruction *I,
+                                        const ArgumentUse &IU,
----------------
jvoung wrote:

nit: Would it make more sense to name the variable "ArgUse"? Unclear to me why "I" in the name?

Similar below "ArgumentUse IU = ..." at line 756

https://github.com/llvm/llvm-project/pull/97373