[llvm] [AMDGPU] Require explicit immediate offsets for SGPR+IMM SMEM instructions. (PR #79131)

[Ivan Kosarev via llvm-commits]

================
@@ -201,6 +209,7 @@ multiclass SM_Pseudo_Probe<RegisterClass baseClass> {
   def _IMM  : SM_Probe_Pseudo <opName, baseClass, IMM_Offset>;
   def _SGPR : SM_Probe_Pseudo <opName, baseClass, SGPR_Offset>;
   def _SGPR_IMM : SM_Probe_Pseudo <opName, baseClass, SGPR_IMM_Offset>;
+  def _SGPR_OPT_IMM : SM_Probe_Pseudo <opName, baseClass, SGPR_IMM_OptOffset>;
----------------
kosarev wrote:

Having a whole new pseudo is probably not ideal, but maybe still better than hardcoding `InOperandList` for the `_SGPR_IMM_gfx12` real?

https://github.com/llvm/llvm-project/pull/79131