[llvm] [GlobalISel] Fix store merging incorrectly classifying an unknown index expr as 0. (PR #90375)

[Amara Emerson via llvm-commits]

================
@@ -84,21 +84,20 @@ BaseIndexOffset GISelAddressing::getPointerInfo(Register Ptr,
                                                 MachineRegisterInfo &MRI) {
   BaseIndexOffset Info;
   Register PtrAddRHS;
-  if (!mi_match(Ptr, MRI, m_GPtrAdd(m_Reg(Info.BaseReg), m_Reg(PtrAddRHS)))) {
-    Info.BaseReg = Ptr;
-    Info.IndexReg = Register();
-    Info.IsIndexSignExt = false;
+  Register BaseReg;
+  if (!mi_match(Ptr, MRI, m_GPtrAdd(m_Reg(BaseReg), m_Reg(PtrAddRHS)))) {
+    Info.setBase(Ptr);
+    Info.setOffset(0);
     return Info;
   }
-
+  Info.setBase(BaseReg);
   auto RHSCst = getIConstantVRegValWithLookThrough(PtrAddRHS, MRI);
   if (RHSCst)
-    Info.Offset = RHSCst->Value.getSExtValue();
+    Info.setOffset(RHSCst->Value.getSExtValue());
 
   // Just recognize a simple case for now. In future we'll need to match
   // indexing patterns for base + index + constant.
-  Info.IndexReg = PtrAddRHS;
-  Info.IsIndexSignExt = false;
+  Info.setIndex(PtrAddRHS);
----------------
aemerson wrote:

I think I'd rather keep the optional offset, since this is actually getting closer to mirroring the equivalent analysis in SelectionDAG, and one day I think it'd be nice to have some abstraction over LLT/SDValue for these things to share code.

https://github.com/llvm/llvm-project/pull/90375