[llvm] [AMDGPU] Use subtarget feature for flat offset bit width instead of arch checks (PR #183742)

[Mariusz Sikora via llvm-commits]

================
@@ -59,6 +59,7 @@ class AMDGPUSubtarget {
   unsigned LocalMemorySize = 0;
   unsigned AddressableLocalMemorySize = 0;
   char WavefrontSizeLog2 = 0;
+  unsigned FlatOffsetBitWidth = 13;
----------------
mariusz-sikora-at-amd wrote:

Thanks for pointing this out.

In that case, I’m thinking of simply setting `FlatOffsetBitWidth` to `0` by default. We never access the field directly, it’s only read through `getNumFlatOffsetBits()`.

Alternatively, I could replace this with two boolean AMDGPUSubtargetFeatures instead of an unsigned value, if that would be preferable.

https://github.com/llvm/llvm-project/pull/183742