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

[Ricardo Jesus via llvm-commits]

================
@@ -914,6 +914,23 @@ AArch64TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
     }
     break;
   }
+  case Intrinsic::experimental_vector_match: {
+    EVT SearchVT = getTLI()->getValueType(DL, ICA.getArgTypes()[0]);
+    unsigned SearchSize =
+        cast<FixedVectorType>(ICA.getArgTypes()[1])->getNumElements();
+    // If we can't lower to MATCH, return an invalid cost.
+    if (getTLI()->shouldExpandVectorMatch(SearchVT, SearchSize))
+      return InstructionCost::getInvalid();
+    // Base cost for MATCH instructions. At least on the Neoverse V2 and
+    // Neoverse V3 these are cheap operations with the same latency as a vector
+    // ADD, though in most cases we also need to do an extra DUP.
+    InstructionCost Cost = 4;
+    // For fixed-length vectors we currently need an extra five--six
+    // instructions besides the MATCH.
+    if (isa<FixedVectorType>(RetTy))
----------------
rj-jesus wrote:

The intrinsic itself supports some types of fixed-length vectors, which is why I added this cost here. Nevertheless, since we don't use these in the transformation right now, I've marked their cost as invalid until we have a use case for them.

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