[clang] [llvm] [BPF] Do atomic_fetch_*() pattern matching with memory ordering (PR #107343)

[via cfe-commits]

================
@@ -152,22 +152,93 @@ static bool hasLiveDefs(const MachineInstr &MI, const TargetRegisterInfo *TRI) {
   return false;
 }
 
-void BPFMIPreEmitChecking::processAtomicInsts() {
+bool BPFMIPreEmitChecking::processAtomicInsts() {
+  if (!MF->getSubtarget<BPFSubtarget>().getHasJmp32()) {
+    // Only check for cpu version 1 and 2.
+    for (MachineBasicBlock &MBB : *MF) {
+      for (MachineInstr &MI : MBB) {
+        if (MI.getOpcode() != BPF::XADDW && MI.getOpcode() != BPF::XADDD)
+          continue;
+
+        LLVM_DEBUG(MI.dump());
+        if (hasLiveDefs(MI, TRI)) {
+          DebugLoc Empty;
+          const DebugLoc &DL = MI.getDebugLoc();
+          const Function &F = MF->getFunction();
+          F.getContext().diagnose(DiagnosticInfoUnsupported{
+              F, "Invalid usage of the XADD return value", DL});
+        }
+      }
+    }
+  }
----------------
eddyz87 wrote:

Question: if errors were reported above, should we stop the compilation?
At the moment error is reported but codegen proceeds.

https://github.com/llvm/llvm-project/pull/107343