[llvm] [AArch64] Lower factor-of-2 interleaved stores to STNP (PR #177938)

[Florian Hahn via llvm-commits]

================
@@ -18573,6 +18610,30 @@ bool AArch64TargetLowering::lowerInterleavedStore(Instruction *Store,
                             BaseAddr, DL)))
     return false;
 
+  // Conditionally skip nontemporal stores, because in that case we should
+  // prioritize emitting non-temporal store instructions, but AArch64 doesn't
+  // have non-temporal interleaved stores.
+  //
+  // Currently, STNP lowering can only either keep or increase code size,
+  // thus we predicate it to not apply when optimizing for code size.
+  //
+  // The check is conservative:
+  //
+  // - Don't skip if the interleaving factor is greater than 2, as the shuffling
+  // overhead becomes higher.
+  // - Don't skip if the store value types which are not directly legal. They
+  // may theoratically be split by legalization and lowered to STNPs, but they
+  // can also match only partially in the worst case and actually emit temporal
+  // stores.
----------------
fhahn wrote:

I think you could make this wording a bit more compact, removing the bit about splitting etc.

https://github.com/llvm/llvm-project/pull/177938