[llvm] [SROA] Extend tree-structured merge to handle init + RMW pattern (PR #194441)

[Princeton Ferro via llvm-commits]

================
@@ -2856,12 +2856,14 @@ class AllocaSliceRewriter : public InstVisitor<AllocaSliceRewriter, bool> {
 
   /// Attempts to rewrite a partition using tree-structured merge optimization.
   ///
-  /// This function analyzes a partition to determine if it can be optimized
-  /// using a tree-structured merge pattern, where multiple non-overlapping
-  /// stores completely fill an alloca. And there is no load from the alloca in
-  /// the middle of the stores. Such patterns can be optimized by eliminating
-  /// the intermediate stores and directly constructing the final vector by
-  /// using shufflevectors.
+  /// This function handles two patterns. Both produce an O(log n) tree of
+  /// shufflevectors in place of the linear expand+blend chain that SROA would
+  /// otherwise emit for each partial store.
+  ///
+  /// Pattern 1 (stores-only):
+  ///   Multiple non-overlapping partial stores completely fill the alloca and
+  ///   there is exactly one full-width load. The stores are tree-merged into
----------------
Prince781 wrote:

nit?: "and there is exactly one full-width load coming after the stores"

https://github.com/llvm/llvm-project/pull/194441