[llvm] [AArch64] [CostModel] Fix cost modelling for saturating arithmetic intrinsics (PR #152333)

[Sander de Smalen via llvm-commits]

================
@@ -643,6 +643,13 @@ AArch64TTIImpl::getIntrinsicInstrCost(const IntrinsicCostAttributes &ICA,
         LT.second.getScalarSizeInBits() == RetTy->getScalarSizeInBits() ? 1 : 4;
     if (any_of(ValidSatTys, [&LT](MVT M) { return M == LT.second; }))
       return LT.first * Instrs;
+
+    TypeSize TS = getDataLayout().getTypeSizeInBits(RetTy);
+    uint64_t VectorSize = TS.getKnownMinValue();
+
+    if (ST->isSVEAvailable() && VectorSize >= 128 && isPowerOf2_64(VectorSize))
----------------
sdesmalen-arm wrote:

There's nothing wrong with the change, but I just want to leave a few observations;
(1) `getTypeLegalizationCost` will return a legal type, which must be a power of 2. So `isPowerOf2_64(VectorSize)` must be true.
(2) The existing code checks if the legal type is any of the legal NEON types, but I think the SVE and NEON case can be merged by changing it to look up the _element_ type in a list of "supported element types" and as long as the vector size of the legalised type is >= 64 bits, it can return `LT.first * Instrs`.
(3) just an observation that this could be rolled out for other intrinsics as well, such as the min/max above which are currently also limited to only (vscale x )128 types.

https://github.com/llvm/llvm-project/pull/152333