[llvm] [AMDGPU] Fix negative immediate offset for unbuffered smem loads (PR #79553)

[via llvm-commits]

================
@@ -116,6 +116,10 @@ def gi_smrd_sgpr_imm :
     GIComplexOperandMatcher<s64, "selectSmrdSgprImm">,
     GIComplexPatternEquiv<SMRDSgprImm>;
 
+def gi_smrd_prefetch_imm :
+    GIComplexOperandMatcher<s64, "selectSmrdPrefetchImm">,
+    GIComplexPatternEquiv<SMRDPrefetchImm>;
----------------
vangthao95 wrote:

This is splitting the handling of prefetch from general handling of smrd selection. There wasn't a distinction between prefetch and s_load instruction so I had to create a way to distinguish between the two. It should be a NFC change. Should this still be done separately?

https://github.com/llvm/llvm-project/pull/79553