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

[via llvm-commits]

================
@@ -2016,7 +2018,14 @@ bool AMDGPUDAGToDAGISel::SelectSMRDOffset(SDValue ByteOffsetNode,
       AMDGPU::getSMRDEncodedOffset(*Subtarget, ByteOffset, IsBuffer);
   if (EncodedOffset && Offset && !Imm32Only) {
     *Offset = CurDAG->getTargetConstant(*EncodedOffset, SL, MVT::i32);
-    return true;
+    if (EncodedOffset >= 0 || IsBuffer || HasSOffset ||
+        !Subtarget->hasSignedSMRDImmOffset())
+      return true;
+    // For unbuffered smem loads, it is illegal for the Immediate Offset to be
+    // negative if the resulting (Offset + (M0 or SOffset or zero) is negative.
+    // Handle the case where the Immediate Offset is negative and there is no
+    // SOffset.
+    return false;
----------------
vangthao95 wrote:

I moved parts of the check inside getSMRDEncodedOffset but checking SOffset is a bit more complicated so I did not move that part inside the helper.

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