[llvm] [AArch64] Update costs for scalarizing i64->f32 int_to_fp. (PR #132366)

[David Green via llvm-commits]

================
@@ -3199,6 +3197,20 @@ InstructionCost AArch64TTIImpl::getCastInstrCost(unsigned Opcode, Type *Dst,
             FP16Tbl, ISD, DstTy.getSimpleVT(), SrcTy.getSimpleVT()))
       return AdjustCost(Entry->Cost);
 
+  // INT_TO_FP of i64->f32 will scalarize, which is required to avoid
+  // double-rounding issues.
+  if ((ISD == ISD::SINT_TO_FP || ISD == ISD::UINT_TO_FP) &&
+      DstTy.getScalarType() == MVT::f32 && SrcTy.getScalarSizeInBits() > 32 &&
----------------
davemgreen wrote:

Anything larger than a i32 (like an i48, if they ever come up) will be legalized to a i64 as we convert to legal types. I figure that anything larger than i64 will become an i128 which will be scalarized too, although the i64 is the important one.

https://github.com/llvm/llvm-project/pull/132366