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

[via llvm-commits]

================
@@ -4257,6 +4258,27 @@ bool AMDGPUInstructionSelector::selectSmrdOffset(MachineOperand &Root,
   if (Offset && GEPI.SgprParts.size() == 1 && EncodedImm) {
     Base = GEPI.SgprParts[0];
     *Offset = *EncodedImm;
+    // For unbuffered smem loads, it is illegal and undefined 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.
+    //
+    // FIXME: Also handle M0 or SOffset case?
+    if (!IsPrefetch && *Offset < 0 &&
+        STI.getGeneration() >= AMDGPUSubtarget::GFX11) {
+      // Subtract the absolute value of the offset from the base register and
+      // set the immediate offset to 0.
+      Register SubtractReg =
+          MRI->createVirtualRegister(&AMDGPU::SReg_64RegClass);
+
+      BuildMI(*MBB, MI, MI->getDebugLoc(), TII.get(AMDGPU::S_SUB_U64),
----------------
vangthao95 wrote:

Thanks for catching this

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