[Mlir-commits] [mlir] [milr][gpu] Make barrier elimination address-space aware (PR #178101)

[Jakub Kuderski]

================
@@ -78,14 +78,132 @@ static void addAllValuelessEffects(
   effects.emplace_back(MemoryEffects::Effect::get<MemoryEffects::Free>());
 }
 
+/// Looks through known "view-like" ops to find the base memref.
+static Value getBase(Value v) {
+  while (true) {
+    Operation *definingOp = v.getDefiningOp();
+    if (!definingOp)
+      break;
+
+    bool shouldContinue = TypeSwitch<Operation *, bool>(v.getDefiningOp())
+                              .Case([&](ViewLikeOpInterface op) {
+                                v = op.getViewSource();
+                                return true;
+                              })
+                              .Case([&](memref::TransposeOp op) {
+                                v = op.getIn();
+                                return true;
+                              })
+                              .Default(false);
+    if (!shouldContinue)
+      break;
+  }
+  return v;
+}
+
+/// Returns `true` if accesses to the given memory space could potentially be
+/// fenced by a barrier synchoronizing on the given `fencedAddressSpaces`. If
+/// the set of address spaces is not given, it is equal to all possible address
+/// spaces. Memory spaces that are not `#gpu.address_space` are deemed to
+/// overlap with all GPU address spaces.
----------------
kuhar wrote:

```suggestion
/// Returns `true` if accesses to the given memory space could potentially be
/// fenced by a barrier synchronizing on the given `fencedAddressSpaces`. If
/// the set of address spaces is not given, it is equal to all possible address
/// spaces. Memory spaces that are not `#gpu.address_space` are deemed to
/// overlap with all GPU address spaces.
```

https://github.com/llvm/llvm-project/pull/178101