diff --git a/include/graphblas/nonblocking/blas1.hpp b/include/graphblas/nonblocking/blas1.hpp
index c4853c7b2..1ac780da7 100644
--- a/include/graphblas/nonblocking/blas1.hpp
+++ b/include/graphblas/nonblocking/blas1.hpp
@@ -10398,6 +10398,9 @@ namespace grb {
 	// internal namespace for implementation of grb::dot
 	namespace internal {
 
+		/**
+		 * Invariant: x should have fewer nonzeroes than y
+		 */
 		template<
 			Descriptor descr,
 #ifdef GRB_BOOLEAN_DISPATCHER
@@ -10432,6 +10435,7 @@ namespace grb {
 #else
 			(void) upper_bound;
 #endif
+			assert( local_x.nonzeroes() <= local_y.nonzeroes() );
 
 			// get raw alias
 			const InputType1 * __restrict__ a = internal::getRaw( x );
@@ -10450,8 +10454,7 @@ namespace grb {
 
 					// prepare registers
 					for( size_t k = 0; k < AnyOp::blocksize; ++k, ++i ) {
-						mask[ k ] = already_dense_input_x ||
-							local_x.assigned( already_dense_input_y ? i : local_y.index( i ) );
+						mask[ k ] = already_dense_input_x || local_y.assigned(local_x.index( i ));
 					}
 
 					// rewind
@@ -10461,9 +10464,9 @@ namespace grb {
 					for( size_t k = 0; k < AnyOp::blocksize; ++k, ++i ) {
 						if( mask[ k ] ) {
 							xx[ k ] = static_cast< typename AnyOp::D1 >(
-								a[ ( already_dense_input_y ? i : local_y.index( i ) ) + lower_bound ] );
+								a[ ( already_dense_input_x ? i : local_x.index( i ) ) + lower_bound ] );
 							yy[ k ] = static_cast< typename AnyOp::D2 >(
-								b[ ( already_dense_input_y ? i : local_y.index( i ) ) + lower_bound ] );
+								b[ ( already_dense_input_x ? i : local_x.index( i ) ) + lower_bound ] );
 						}
 					}
 
@@ -10504,9 +10507,9 @@ namespace grb {
 				for( ; i < local_nz; ++i ) {
 					typename AddMonoid::D3 temp =
 						addMonoid.template getIdentity< typename AddMonoid::D3 >();
-					const size_t index = ( already_dense_input_y ? i : local_y.index( i ) ) +
+					const size_t index = ( already_dense_input_x ? i : local_x.index( i ) ) +
 						lower_bound;
-					if( already_dense_input_x || local_x.assigned( index - lower_bound ) ) {
+					if( already_dense_input_y || local_y.assigned( index - lower_bound ) ) {
 						apply( temp, a[ index ], b[ index ], anyOp );
 						foldr( temp, thread_local_output, addMonoid.getOperator() );
 					}
@@ -10658,7 +10661,9 @@ namespace grb {
 										already_dense_input_y, already_dense_input_x,
 										array_reduced[ thread_id ],
 										lower_bound, upper_bound,
-										local_y, local_x, x, y, local_y_nz,
+										local_y, local_x,
+										x, y,
+										local_y_nz,
 										addMonoid, anyOp
 									);
 							}
diff --git a/tests/unit/dot.cpp b/tests/unit/dot.cpp
index 4279109ce..326b4fe88 100644
--- a/tests/unit/dot.cpp
+++ b/tests/unit/dot.cpp
@@ -27,6 +27,7 @@ void grb_program( const size_t &n, grb::RC &rc ) {
 
 	// repeatedly used containers
 	grb::Vector< bool > even_mask( n );
+	grb::Vector< bool > odd_mask( n );
 	grb::Vector< size_t > temp( n );
 	grb::Vector< double > left( n );
 	grb::Vector< double > right( n );
@@ -42,7 +43,14 @@ void grb_program( const size_t &n, grb::RC &rc ) {
 		}, temp );
 	rc = rc ? rc : grb::set( even_mask, temp, true );
 	if( rc != grb::SUCCESS ) {
-		std::cerr << "\t initialisation of mask FAILED\n";
+		std::cerr << "\t initialisation of even mask FAILED\n";
+		return;
+	}
+
+	// create odd mask
+	rc = grb::set< grb::descriptors::invert_mask >( odd_mask, temp, true );
+	if( rc != grb::SUCCESS ) {
+		std::cerr << "\t initialisation of odd mask FAILED\n";
 		return;
 	}
 
@@ -160,6 +168,75 @@ void grb_program( const size_t &n, grb::RC &rc ) {
 		return;
 	}
 
+	// test 5, init
+	rc = rc ? rc :  grb::clear( x );
+	rc = rc ? rc :  grb::clear( y );
+	assert( rc == grb::SUCCESS );
+	int true_dot = 0;
+	alpha = 0;
+
+	rc = rc ? rc : grb::set< grb::descriptors::use_index >( y, 0 );
+	for( size_t i : {2,3,5,7,13,17,19,23,29} ){
+		if( i >= n ) break;
+		rc = rc ? rc :  grb::setElement( x, 1, i );
+		true_dot += i;
+	}
+	assert( rc == grb::SUCCESS );
+
+	// test 5, exec
+	rc = grb::dot( alpha, x, y, intRing );
+	if( rc != SUCCESS ) {
+		std::cerr << "\t test 5 (sparse non constant-value vectors) dot FAILED\n";
+		return;
+	}
+
+	// test 5, check
+	if( alpha != true_dot ) {
+		std::cerr << "\t test 5 (sparse non constant-value vectors) unexpected value "
+			<< alpha << ", expected " << true_dot << ".\n";
+		rc = FAILED;
+		return;
+	}
+
+	// test 6, init
+	rc = grb::set( y, x );
+	rc = rc ? rc : grb::set< grb::descriptors::use_index >( x, 0 );
+	alpha = 0;
+
+	// test 6, exec
+	rc = rc ? rc : grb::dot( alpha, x, y, intRing );
+	if( rc != SUCCESS ) {
+		std::cerr << "\t test 6 (swapped version of test 5) dot FAILED\n";
+		return;
+	}
+
+	// test 6, check
+	if( alpha != true_dot ) {
+		std::cerr << "\t test 6 (swapped version of test 5) unexpected value "
+			<< alpha << ", expected " << true_dot << ".\n";
+		rc = FAILED;
+		return;
+	}
+
+	// test 7, init
+	rc = grb::set< grb::descriptors::use_index >( x, odd_mask, 0 );
+	rc = rc ? rc : grb::set< grb::descriptors::use_index >( y, even_mask, 0 );
+	alpha = 0;
+
+	// test 7, exec
+	rc = rc ? rc : grb::dot( alpha, x, y, intRing );
+	if( rc != SUCCESS ) {
+		std::cerr << "\t test 7 (sparse dot no overlap) FAILED\n";
+		return;
+	}
+
+	// test 7, check
+	if( alpha != 0 ) {
+		std::cerr << "\t test 7 (sparse dot no overlap) unexpected value "
+			<< alpha << ", expected 0.\n";
+		rc = FAILED;
+		return;
+	}
 }
 
 int main( int argc, char ** argv ) {