[llvm] [AMDGPU] Set gfx13 LDS to 192KB and model LDS size via subtarget features (PR #214513)

[Mariusz Sikora via llvm-commits]

================
@@ -42,8 +42,22 @@ class SubtargetFeatureAddressableLocalMemorySize <int Value> : SubtargetFeature<
 def FeatureAddressableLocalMemorySize32768 : SubtargetFeatureAddressableLocalMemorySize<32768>;
 def FeatureAddressableLocalMemorySize65536 : SubtargetFeatureAddressableLocalMemorySize<65536>;
 def FeatureAddressableLocalMemorySize163840 : SubtargetFeatureAddressableLocalMemorySize<163840>;
+def FeatureAddressableLocalMemorySize196608 : SubtargetFeatureAddressableLocalMemorySize<196608>;
 def FeatureAddressableLocalMemorySize327680 : SubtargetFeatureAddressableLocalMemorySize<327680>;
 
+// Total size of LDS on the whole block, in bytes. Set this only when it is
+// larger than the addressable size, i.e. when one workgroup cannot use the whole
+// block (e.g. gfx10/11/12: 128k physical block vs 64k addressable). If not set,
+// it defaults to the addressable size.
+class SubtargetFeatureLocalMemorySize <int Value> : SubtargetFeature<
----------------
mariusz-sikora-at-amd wrote:

I used `FeatureHalfAddressablePhysicalLocalMemory` as 'bool' feature for targets gfx10/11/12 where we can address half of physical LDS.

https://github.com/llvm/llvm-project/pull/214513