[llvm] [AtomicExpandPass][NVPTX] Preserve atomic and volatile nature of emulated operations (PR #188361)

[Matt Arsenault via llvm-commits]

================
@@ -1163,6 +1164,22 @@ bool AtomicExpandImpl::expandPartwordCmpXchg(AtomicCmpXchgInst *CI) {
   PHINode *Loaded_MaskOut = Builder.CreatePHI(PMV.WordType, 2);
   Loaded_MaskOut->addIncoming(InitLoaded_MaskOut, BB);
 
+  // The initial load must be atomic with the same synchronization scope
+  // to avoid a data race with concurrent stores. If the instruction being
+  // emulated is volatile, issue a volatile load.
+  // addIncoming is done first so that expandAtomicLoadToLibcall's
+  // replaceAllUsesWith correctly updates the PHI incoming value.
+  InitLoaded->setVolatile(CI->isVolatile());
+  if (TLI->issueAtomicInitLoadForAtomicEmulation()) {
+    // TODO: Get rid of the target hook once all backends start issuing
+    // atomic loads.
----------------
arsenm wrote:

Just do the mechanical updates. This is how these sorts of cross backend changes go 

https://github.com/llvm/llvm-project/pull/188361