[llvm] [AArch64] Adjust costs for masked CompressStore and ExpandLoad (PR #226950)

[Benjamin Maxwell via llvm-commits]

================
@@ -5563,7 +5563,13 @@ AArch64TTIImpl::getMaskedMemoryOpCost(const MemIntrinsicCostAttributes &MICA,
       return InstructionCost::getInvalid();
 
     // Operation will be split into expand of masked.load
-    MemOpCost *= 2;
+    // Something like:
+    // cntp x8, p0, p0.s
+    // whilelo p1.s, xzr, x8
+    // ld1w { z1.s }, p1/z, [x0]
+    // expand z1.s, p0, z1.s
+    // mov z0.s, p0/m, z1.s1
+    MemOpCost *= 4;
----------------
MacDue wrote:

Should the cost of the extra instructions be added to the MemOpCost (rather than multiply it)? The multiplication is a little fishy, since there is still only one memory operation. 

https://github.com/llvm/llvm-project/pull/226950