[llvm] [DA] Add batch delinearization support for improved precision (PR #170519)

[Alexey Bataev via llvm-commits]

================
@@ -388,6 +396,24 @@ class DependenceInfo {
     unsigned char DirSet;
   };
 
+  /// DelinearizationCache - Cache for batch delinearization results.
+  /// When analyzing multiple memory accesses to the same base pointer,
+  /// caching allows us to compute array dimensions once using all accesses,
+  /// leading to better precision (similar to Polly's approach).
+  struct DelinearizationCache {
+    /// Map from base pointer to computed array dimension sizes.
+    DenseMap<const SCEVUnknown *, SmallVector<const SCEV *, 4>> ArraySizes;
+    /// Map from instruction to pre-computed subscripts.
+    DenseMap<const Instruction *, SmallVector<const SCEV *, 4>> Subscripts;
+    /// Element size for the array (used for validation).
+    DenseMap<const SCEVUnknown *, const SCEV *> ElementSizes;
----------------
alexey-bataev wrote:

Use SmallDenseMap, it improves compile time

https://github.com/llvm/llvm-project/pull/170519