[clang] [clang-tools-extra] [llvm] [AArch64] Add an AArch64 pass for loop idiom transformations (PR #72273)

[David Green via cfe-commits]

================
@@ -0,0 +1,839 @@
+//===- AArch64LoopIdiomTransform.cpp - Loop idiom recognition -------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This pass implements a pass that recognizes certain loop idioms and
+// transforms them into more optimized versions of the same loop. In cases
+// where this happens, it can be a significant performance win.
+//
+// We currently only recognize one loop that finds the first mismatched byte
+// in an array and returns the index, i.e. something like:
+//
+//  while (++i != n) {
+//    if (a[i] != b[i])
+//      break;
+//  }
+//
+// In this example we can actually vectorize the loop despite the early exit,
+// although the loop vectorizer does not support it. It requires some extra
+// checks to deal with the possibility of faulting loads when crossing page
+// boundaries. However, even with these checks it is still profitable to do the
+// transformation.
+//
+//===----------------------------------------------------------------------===//
+//
+// TODO List:
+//
+// * When optimizing for code size we may want to avoid some transformations.
----------------
davemgreen wrote:

Can we add a check for this now? It looks like it will be quite a bit bigger?

https://github.com/llvm/llvm-project/pull/72273