[mlir][Vector] Improve vector.mask verifier

This PR improves the verifier for the `vector.mask` operation to
make sure it's not applying masking semantics to operations defined
outside of the `vector.mask` region. Documentation is updated to
emphasize that and make it clearer, even though it already stated that.

As part of this change, the logic that ensures that a terminator is
present in the region mask has been simplified to make it less surprising
to the user when a `vector.yield` is explicitly provided.

Author: dcaballe

mlir/include/mlir/Dialect/Vector/IR/VectorOps.td

@@ -2482,8 +2482,13 @@ def Vector_MaskOp : Vector_Op<"mask", [
     masked. Values used within the region are captured from above. Only one
     *maskable* operation can be masked with a `vector.mask` operation at a time.
     An operation is *maskable* if it implements the `MaskableOpInterface`. The
-    terminator yields all results of the maskable operation to the result of
-    this operation.
+    terminator yields all results from the maskable operation to the result of
+    this operation. No other values are allowed to be yielded.
+
+    An empty `vector.mask` operation is considered ill-formed but legal to
+    facilitate optimizations across the `vector.mask` operation. It is considered
+    a no-op regardless of its returned values and will be removed by the
+    canonicalizer.
 
     The vector mask argument holds a bit for each vector lane and determines
     which vector lanes should execute the maskable operation and which ones

mlir/lib/Dialect/Vector/IR/VectorOps.cpp

@@ -6550,29 +6550,31 @@ void mlir::vector::MaskOp::print(OpAsmPrinter &p) {
 }
 
 void MaskOp::ensureTerminator(Region &region, Builder &builder, Location loc) {
-  OpTrait::SingleBlockImplicitTerminator<vector::YieldOp>::Impl<
-      MaskOp>::ensureTerminator(region, builder, loc);
-  // Keep the default yield terminator if the number of masked operations is not
-  // the expected. This case will trigger a verification failure.
-  Block &block = region.front();
-  if (block.getOperations().size() != 2)
+  // Create default terminator if there are no ops to mask.
+  if (region.empty() || region.front().empty()) {
+    OpTrait::SingleBlockImplicitTerminator<vector::YieldOp>::Impl<
+        MaskOp>::ensureTerminator(region, builder, loc);
     return;
+  }
 
-  // Replace default yield terminator with a new one that returns the results
-  // from the masked operation.
-  OpBuilder opBuilder(builder.getContext());
-  Operation *maskedOp = &block.front();
-  Operation *oldYieldOp = &block.back();
-  assert(isa<vector::YieldOp>(oldYieldOp) && "Expected vector::YieldOp");
+  // If  region has an explicit terminator, we don't modify it.
+  Block &block = region.front();
+  if (isa<vector::YieldOp>(block.back()))
+    return;
 
-  // Empty vector.mask op.
-  if (maskedOp == oldYieldOp)
+  // Create default terminator if the number of masked operations is not
+  // one. This case will trigger a verification failure.
+  if (block.getOperations().size() != 1) {
+    OpTrait::SingleBlockImplicitTerminator<vector::YieldOp>::Impl<
+        MaskOp>::ensureTerminator(region, builder, loc);
     return;
+  }
 
-  opBuilder.setInsertionPoint(oldYieldOp);
+  // Create a terminator that yields the results from the masked operation.
+  OpBuilder opBuilder(builder.getContext());
+  Operation *maskedOp = &block.front();
+  opBuilder.setInsertionPointToEnd(&block);
   opBuilder.create<vector::YieldOp>(loc, maskedOp->getResults());
-  oldYieldOp->dropAllReferences();
-  oldYieldOp->erase();
 }
 
 LogicalResult MaskOp::verify() {
@@ -6607,6 +6609,11 @@ LogicalResult MaskOp::verify() {
     return emitOpError("expects number of results to match maskable operation "
                        "number of results");
 
+  if (!llvm::equal(maskableOp->getResults(), terminator.getOperands()))
+    return emitOpError(
+        "expects all the results from the MaskableOpInterface to "
+        "be returned by the terminator");
+
   if (!llvm::equal(maskableOp->getResultTypes(), getResultTypes()))
     return emitOpError(
         "expects result type to match maskable operation result type");

mlir/test/Dialect/Vector/invalid.mlir

@@ -1756,6 +1756,20 @@ func.func @vector_mask_empty_passthru_no_return_type(%mask : vector<8xi1>,
 
 // -----
 
+func.func @vector_mask_non_empty_external_return(%t0: tensor<?xf32>, %idx: index,
+                                                 %m0: vector<16xi1>, %ext: vector<16xf32>) -> vector<16xf32> {
+  %ft0 = arith.constant 0.0 : f32
+  // expected-error@+1 {{'vector.mask' op expects all the results from the MaskableOpInterface to be returned by the terminator}}
+  %0 = vector.mask %m0 {
+    %1 =vector.transfer_read %t0[%idx], %ft0 : tensor<?xf32>, vector<16xf32>
+    vector.yield %ext : vector<16xf32>
+  } : vector<16xi1> -> vector<16xf32>
+
+  return %0 : vector<16xf32>
+}
+
+// -----
+
 func.func @vector_mask_empty_passthru_empty_return_type(%mask : vector<8xi1>,
                                                         %passthru : vector<8xi32>) {
   // expected-error@+1 {{'vector.mask' expects a result if passthru operand is provided}}
@@ -1765,6 +1779,20 @@ func.func @vector_mask_empty_passthru_empty_return_type(%mask : vector<8xi1>,
 
 // -----
 
+func.func @vector_mask_non_empty_mixed_return(%t0: tensor<?xf32>, %idx: index,
+                                              %m0: vector<16xi1>, %ext: vector<16xf32>) -> (vector<16xf32>, vector<16xf32>) {
+  %ft0 = arith.constant 0.0 : f32
+  // expected-error@+1 {{'vector.mask' op expects number of results to match maskable operation number of results}}
+  %0:2 = vector.mask %m0 {
+    %1 =vector.transfer_read %t0[%idx], %ft0 : tensor<?xf32>, vector<16xf32>
+    vector.yield %1, %ext : vector<16xf32>, vector<16xf32>
+  } : vector<16xi1> -> (vector<16xf32>, vector<16xf32>)
+
+  return %0#0, %0#1 : vector<16xf32>, vector<16xf32>
+}
+
+// -----
+
 func.func @vector_scalable_insert_unaligned(%subv: vector<4xi32>, %vec: vector<[16]xi32>) {
   // expected-error@+1 {{op failed to verify that position is a multiple of the source length.}}
   %0 = vector.scalable.insert %subv, %vec[2] : vector<4xi32> into vector<[16]xi32>