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

[Sander de Smalen 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
----------------
sdesmalen-arm wrote:

We just discussed this offline, but just sharing my thoughts here: IMO the table should represent the cost of casts of legal types. Illegal types should be handled by generic code that multiplies the cost by the 'type legalization cost'. This is actually what happens for fixed-length types (see the code just below the table), but not (yet) for scalable types. Otherwise, any other illegal types that are not in the table (which includes types that cannot be represented by MVTs because they're "too wide") will get some default cost, which may be far too low.

It also seems that `SINT_TO_FP` records are missing in the table for scalable vector types (only FP_TO_SINT is handled). This is probably just a historical omission because this table gets updated/botched on an ad-hoc basis when people find that the cost is wrong for some workload, for some type and operation. It would be nice to clean this up.

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