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

[Shilei Tian via llvm-commits]

================
@@ -1473,6 +1473,13 @@ def FeatureWaitsBeforeSystemScopeStores : SubtargetFeature<
   "Target requires waits for loads and atomics before system scope stores"
 >;
 
+def FeaturePKF32Insts : SubtargetFeature<"pk-f32-insts",
----------------
shiltian wrote:

Well, technically they are different. A packed fp32 instruction with `OPF_PK_F32` flag is the interesting one here.

> Then you probably only need a check in the subtarget anyway, not a feature itself.

I'm not sure if I'm following.

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