[llvm] [AMDGPU] Don't fold an i64 immediate value if it can't be replicated from its lower 32-bit (PR #168458)

[Stanislav Mekhanoshin via llvm-commits]

================
@@ -1420,6 +1420,13 @@ class GCNSubtarget final : public AMDGPUGenSubtargetInfo,
   /// \returns true if the target has instructions with xf32 format support.
   bool hasXF32Insts() const { return HasXF32Insts; }
 
+  /// \returns true if the target has packed f32 instructions that only read 32
+  /// bits from a scalar operand (SGPR or literal) and replicates the bits to
+  /// both channels.
+  bool hasPKF32InstsReplicatingLow32BitsOfScalarInput() const {
+    return getGeneration() == GFX12 && GFX1250Insts;
----------------
rampitec wrote:

GFX1250Insts is sufficient.

https://github.com/llvm/llvm-project/pull/168458