[llvm] [AArch64][CostModel] Increase the cost of illegal SVE int-to-fp converts (PR #130756)

[Graham Hunter via llvm-commits]

================
@@ -3144,6 +3144,21 @@ InstructionCost AArch64TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst,
       {ISD::SIGN_EXTEND, MVT::nxv8i32, MVT::nxv8i16, 2},
       {ISD::SIGN_EXTEND, MVT::nxv8i64, MVT::nxv8i16, 6},
       {ISD::SIGN_EXTEND, MVT::nxv4i64, MVT::nxv4i32, 2},
+
+      // Add cost for extending and converting to illegal -too wide- scalable
+      // Extending one size (e.g. i32 -> f64) takes 2 unpacks and 2 fcvts, while
----------------
huntergr-arm wrote:

So I've changed the approach slightly to use the pseudo-legalization from the base `getCastInstrCost` that NEON uses (note the code below the table is about using SVE for fixed-length, so doesn't always apply).

Using this approach, we'll still get some illegal types (e.g. mapping nxv2i16 -> nxv2f64, the input would be promoted to nxv2i64 but that's not done in the current code for NEON), but I'm covering the cases where the destination type is legal.

I've decided to back away from increasing the cost of direct fcvts here – even though they have less throughput than `add`, the NEON values are not written with that in mind so we might incorrectly decide to favour NEON (or scalar) code.

I'll rerun some benchmarking with these adjusted values to see whether there's any regression from doing this.

https://github.com/llvm/llvm-project/pull/130756