[llvm] [AMDGPU] Enable scalar subword loads for plain 16-bit loads (PR #225097)

[via llvm-commits]

================
@@ -1041,12 +1041,19 @@ multiclass ScalarBufferLoadIntrinsicPat <SDPatternOperator name, string Instr> {
 // selector to prefer those.
 let AddedComplexity = 100 in {
 
-defm : ScalarLoadWithExtensionPat <"S_LOAD_U8", smrd_extloadi8, i32>;
-defm : ScalarLoadWithExtensionPat <"S_LOAD_U8", smrd_zextloadi8, i32>;
-defm : ScalarLoadWithExtensionPat <"S_LOAD_I8", smrd_sextloadi8, i32>;
-defm : ScalarLoadWithExtensionPat <"S_LOAD_U16", smrd_extloadi16, i32>;
-defm : ScalarLoadWithExtensionPat <"S_LOAD_U16", smrd_zextloadi16, i32>;
-defm : ScalarLoadWithExtensionPat <"S_LOAD_I16", smrd_sextloadi16, i32>;
+// Extending subword loads
+defm : ScalarSubwordLoadPat <"S_LOAD_U8", smrd_extloadi8, i32>;
+defm : ScalarSubwordLoadPat <"S_LOAD_U8", smrd_zextloadi8, i32>;
----------------
LU-JOHN wrote:

`smrd_extloadi8` matches any extension.  

https://github.com/llvm/llvm-project/pull/225097