[llvm] [AArch64] Add MATCH loops to LoopIdiomVectorizePass (PR #101976)

[Ricardo Jesus via llvm-commits]

================
@@ -24,19 +26,42 @@
 // boundaries. However, even with these checks it is still profitable to do the
 // transformation.
 //
-//===----------------------------------------------------------------------===//
-//
-// NOTE: This Pass matches a really specific loop pattern because it's only
-// supposed to be a temporary solution until our LoopVectorizer is powerful
-// enought to vectorize it automatically.
-//
 // TODO List:
 //
 // * Add support for the inverse case where we scan for a matching element.
 // * Permit 64-bit induction variable types.
 // * Recognize loops that increment the IV *after* comparing bytes.
 // * Allow 32-bit sign-extends of the IV used by the GEP.
 //
+// 2. A loop that finds the first matching character in an array among a set of
+// possible matches, e.g.:
+//
+//   for (; first != last; ++first)
+//     for (s_it = s_first; s_it != s_last; ++s_it)
+//       if (*first == *s_it)
+//         return first;
+//   return last;
+//
+// This corresponds to std::find_first_of (for arrays of bytes) from the C++
+// standard library. This function can be implemented efficiently for targets
+// that support @llvm.experimental.vector.match. For example, on AArch64 targets
+// that implement SVE2, this lower to a MATCH instruction, which enables us to
+// perform up to 16x16=256 comparisons in one go. This can lead to very
+// significant speedups.
+//
+// TODO:
+//
+// * Add support for `find_first_not_of' loops (i.e. with not-equal comparison).
+// * Make VF a configurable parameter (right now we assume 128-bit vectors).
+// * Potentially adjust the cost model to let the transformation kick-in even if
+//   @llvm.experimental.vector.match doesn't have direct support in hardware.
+//
+//===----------------------------------------------------------------------===//
+//
+// NOTE: This Pass matches really specific loop patterns because it's only
+// supposed to be a temporary solution until our LoopVectorizer is powerful
+// enought to vectorize them automatically.
----------------
rj-jesus wrote:

Done - thanks for catching this! 

https://github.com/llvm/llvm-project/pull/101976