This change prevents convert_layout propagation of slice layouts across
broadcast/expand_dims.

It is generally a good idea, but it is especially important when running
triton-xla-squeeze-dims, which removes expand_dims ops. The expand_dims ops
remove the slice layouts, so we need to prevent them from propagating
further across broadcast ops which can cause large tensor materialization.

See also b/422133176

--- a/lib/Dialect/TritonGPU/Transforms/RemoveLayoutConversions.cpp
+++ b/lib/Dialect/TritonGPU/Transforms/RemoveLayoutConversions.cpp
@@ -343,6 +343,15 @@ SmallVector<Value> LayoutPropagation::propagateToUsers(Value value,
     if (auto reshapeOp = dyn_cast<ReshapeOp>(user);
         reshapeOp && reshapeOp.getEfficientLayout())
       continue;
+    // Heuristic: don't propagate slice layouts across broadcasts.
+    // This can cause massive register pressure. It's better to convert to
+    // blocked before the broadcast.
+    if (isa<BroadcastOp>(user) &&
+        llvm::any_of(info.encodings, [](Attribute encoding) {
+          return llvm::isa_and_nonnull<SliceEncodingAttr>(encoding);
+        })) {
+      continue;
+    }
     if (user->hasTrait<OpTrait::SameOperandsAndResultEncoding>() ||
         user->hasTrait<OpTrait::Elementwise>() ||
         isa<ReduceOp, ExpandDimsOp, ReshapeOp, TransOp, JoinOp, SplitOp,