[llvm] [NVPTX] Load/Store/Fence syncscope support (PR #106101)

[Artem Belevich via llvm-commits]

================
@@ -957,33 +1001,86 @@ static bool canLowerToLDG(MemSDNode *N, const NVPTXSubtarget &Subtarget,
   });
 }
 
-NVPTX::Ordering NVPTXDAGToDAGISel::insertMemoryInstructionFence(SDLoc DL,
-                                                                SDValue &Chain,
-                                                                MemSDNode *N) {
-  // Some memory instructions - loads, stores, atomics - need an extra fence
-  // instruction. Get the memory order of the instruction, and that of its
-  // fence, if any.
+static unsigned int getFenceOp(NVPTX::Ordering O, NVPTX::Scope S,
+                               NVPTXSubtarget const *T) {
+  if (S == NVPTX::Scope::Cluster)
+    T->failIfClustersUnsupported(".cluster scope fence");
+
+  switch (O) {
+  case NVPTX::Ordering::Acquire:
+  case NVPTX::Ordering::Release:
+  case NVPTX::Ordering::AcquireRelease: {
+    switch (S) {
+    case NVPTX::Scope::System:
+      return T->hasMemoryOrdering() ? NVPTX::atomic_thread_fence_acq_rel_sys
+                                    : NVPTX::INT_MEMBAR_SYS;
+    case NVPTX::Scope::Block:
+      return T->hasMemoryOrdering() ? NVPTX::atomic_thread_fence_acq_rel_cta
+                                    : NVPTX::INT_MEMBAR_CTA;
+    case NVPTX::Scope::Cluster:
+      return NVPTX::atomic_thread_fence_acq_rel_cluster;
+    case NVPTX::Scope::Device:
+      return T->hasMemoryOrdering() ? NVPTX::atomic_thread_fence_acq_rel_gpu
+                                    : NVPTX::INT_MEMBAR_GL;
+    case NVPTX::Scope::Thread:
+      report_fatal_error(
+          formatv("Unsupported scope \"{}\" for acquire/release/acq_rel fence.",
+                  ScopeToString(S)));
+    }
+  }
+  case NVPTX::Ordering::SequentiallyConsistent: {
+    switch (S) {
+    case NVPTX::Scope::System:
+      return T->hasMemoryOrdering() ? NVPTX::atomic_thread_fence_seq_cst_sys
+                                    : NVPTX::INT_MEMBAR_SYS;
+    case NVPTX::Scope::Block:
+      return T->hasMemoryOrdering() ? NVPTX::atomic_thread_fence_seq_cst_cta
+                                    : NVPTX::INT_MEMBAR_CTA;
+    case NVPTX::Scope::Cluster:
+      return NVPTX::atomic_thread_fence_seq_cst_cluster;
+    case NVPTX::Scope::Device:
+      return T->hasMemoryOrdering() ? NVPTX::atomic_thread_fence_seq_cst_gpu
+                                    : NVPTX::INT_MEMBAR_GL;
+    case NVPTX::Scope::Thread:
+      report_fatal_error(formatv("Unsupported scope \"{}\" for seq_cst fence.",
+                                 ScopeToString(S)));
+    }
+  }
+  case NVPTX::Ordering::NotAtomic:
+  case NVPTX::Ordering::Relaxed:
+  case NVPTX::Ordering::Volatile:
+  case NVPTX::Ordering::RelaxedMMIO:
+    report_fatal_error(
+        formatv("Unsupported \"{}\" ordering and \"{}\" scope for fence.",
+                OrderingToString(O), ScopeToString(S)));
+  }
+}
----------------
Artem-B wrote:

Please add `llvm_unreachable()` here, too.

https://github.com/llvm/llvm-project/pull/106101