[llvm] [AMDGPU] Fix negative immediate offset for unbuffered smem loads (PR #89165)

[Matt Arsenault via llvm-commits]

================
@@ -2072,13 +2084,22 @@ SDValue AMDGPUDAGToDAGISel::Expand32BitAddress(SDValue Addr) const {
 // true, match only 32-bit immediate offsets available on CI.
 bool AMDGPUDAGToDAGISel::SelectSMRDBaseOffset(SDValue Addr, SDValue &SBase,
                                               SDValue *SOffset, SDValue *Offset,
-                                              bool Imm32Only,
-                                              bool IsBuffer) const {
+                                              bool Imm32Only, bool IsBuffer,
+                                              bool HasSOffset,
+                                              int64_t ImmOffset) const {
   if (SOffset && Offset) {
     assert(!Imm32Only && !IsBuffer);
     SDValue B;
-    return SelectSMRDBaseOffset(Addr, B, nullptr, Offset) &&
-           SelectSMRDBaseOffset(B, SBase, SOffset, nullptr);
+
+    if (!SelectSMRDBaseOffset(Addr, B, nullptr, Offset, false, false, true))
+      return false;
+
+    int64_t ImmOff = 0;
+    if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(*Offset))
+      ImmOff = C->getSExtValue();
----------------
arsenm wrote:

Parsing out the constant offset here is harder to follow, SelectSMRDBaseOffset can do this just as well 

https://github.com/llvm/llvm-project/pull/89165