[llvm] [AArch64] Don't try to vectorize fixed point to fp narrowing conversion (PR #130665)

[David Green via llvm-commits]

================
@@ -5106,6 +5106,30 @@ SDValue AArch64TargetLowering::LowerVectorINT_TO_FP(SDValue Op,
   uint64_t VTSize = VT.getFixedSizeInBits();
   uint64_t InVTSize = InVT.getFixedSizeInBits();
   if (VTSize < InVTSize) {
+    // AArch64 doesn't have a direct vector instruction to convert
+    // fixed point to floating point AND narrow it at the same time.
+    // Additional rounding when the target is f32/f64 causes subtle
+    // differences across different platforms (that do have such
+    // instructions). Conversion to f16 however is fine.
+    bool IsTargetf32Orf64 = VT.getVectorElementType() == MVT::f32 ||
+                            VT.getVectorElementType() == MVT::f64;
----------------
davemgreen wrote:

I think this can be `bool IsTargetf32 = VT.getScalarType() == MVT::f32;`, as if it was an f64 the integer type would need to be > 64bits, and those should already have been scalarized?

https://github.com/llvm/llvm-project/pull/130665