[llvm] [AArch64][CostModel] Add constraints on which partial reductions are (PR #163728)

[David Sherwood via llvm-commits]

================
@@ -5661,6 +5661,9 @@ InstructionCost AArch64TTIImpl::getPartialReductionCost(
       AccumType->getScalarSizeInBits() / InputTypeA->getScalarSizeInBits();
   if (VF.getKnownMinValue() <= Ratio)
     return Invalid;
+  // i32 -> i64 or i16 -> i32 is not natively supported on Neon and SVE.
----------------
david-arm wrote:

Also, we can in theory lower code for a ratio of 2 even for NEON or SVE. It just might not be optimal codegen. Perhaps the cost modelling below needs modifying to handle the case you are worried about?

https://github.com/llvm/llvm-project/pull/163728