Merge pull request #5205 from vespa-engine/arnej/ensure-more-optimization

Arnej/ensure more optimization

14 files changed, 106 insertions, 149 deletions

diff --git a/eval/src/tests/eval/tensor_function/tensor_function_test.cpp b/eval/src/tests/eval/tensor_function/tensor_function_test.cpp
index 0308b24e742..41037fa06ef 100644
--- a/eval/src/tests/eval/tensor_function/tensor_function_test.cpp
+++ b/eval/src/tests/eval/tensor_function/tensor_function_test.cpp
@@ -147,6 +147,7 @@ TEST("require that const_value works") {
     Value::UP my_const = ctx.make_tensor_matrix();
     Value::UP expect = ctx.make_tensor_matrix();
     const auto &fun = const_value(*my_const, ctx.stash);
+    EXPECT_TRUE(!fun.result_is_mutable());
     EXPECT_EQUAL(expect->type(), fun.result_type());
     const auto &prog = ctx.compile(fun);
     TEST_DO(verify_equal(*expect, ctx.eval(prog)));
@@ -157,6 +158,7 @@ TEST("require that tensor injection works") {
     size_t a_id = ctx.add_tensor(ctx.make_tensor_matrix());
     Value::UP expect = ctx.make_tensor_matrix();
     const auto &fun = inject(ValueType::from_spec("tensor(x[2],y[2])"), a_id, ctx.stash);
+    EXPECT_TRUE(!fun.result_is_mutable());
     EXPECT_EQUAL(expect->type(), fun.result_type());
     const auto &prog = ctx.compile(fun);
     TEST_DO(verify_equal(*expect, ctx.eval(prog)));
@@ -167,6 +169,7 @@ TEST("require that partial tensor reduction works") {
     size_t a_id = ctx.add_tensor(ctx.make_tensor_reduce_input());
     Value::UP expect = ctx.make_tensor_reduce_y_output();
     const auto &fun = reduce(inject(ValueType::from_spec("tensor(x[3],y[2])"), a_id, ctx.stash), Aggr::SUM, {"y"}, ctx.stash);
+    EXPECT_TRUE(fun.result_is_mutable());
     EXPECT_EQUAL(expect->type(), fun.result_type());
     const auto &prog = ctx.compile(fun);
     TEST_DO(verify_equal(*expect, ctx.eval(prog)));
@@ -176,6 +179,7 @@ TEST("require that full tensor reduction works") {
     EvalCtx ctx(SimpleTensorEngine::ref());
     size_t a_id = ctx.add_tensor(ctx.make_tensor_reduce_input());
     const auto &fun = reduce(inject(ValueType::from_spec("tensor(x[3],y[2])"), a_id, ctx.stash), Aggr::SUM, {}, ctx.stash);
+    EXPECT_TRUE(fun.result_is_mutable());
     EXPECT_EQUAL(ValueType::from_spec("double"), fun.result_type());
     const auto &prog = ctx.compile(fun);
     const Value &result = ctx.eval(prog);
@@ -188,6 +192,7 @@ TEST("require that tensor map works") {
     size_t a_id = ctx.add_tensor(ctx.make_tensor_map_input());
     Value::UP expect = ctx.make_tensor_map_output();
     const auto &fun = map(inject(ValueType::from_spec("tensor(x{},y{})"), a_id, ctx.stash), operation::Neg::f, ctx.stash);
+    EXPECT_TRUE(fun.result_is_mutable());
     EXPECT_EQUAL(expect->type(), fun.result_type());
     const auto &prog = ctx.compile(fun);
     TEST_DO(verify_equal(*expect, ctx.eval(prog)));
@@ -201,6 +206,7 @@ TEST("require that tensor join works") {
     const auto &fun = join(inject(ValueType::from_spec("tensor(x{},y{})"), a_id, ctx.stash),
                            inject(ValueType::from_spec("tensor(y{},z{})"), b_id, ctx.stash),
                            operation::Mul::f, ctx.stash);
+    EXPECT_TRUE(fun.result_is_mutable());
     EXPECT_EQUAL(expect->type(), fun.result_type());
     const auto &prog = ctx.compile(fun);
     TEST_DO(verify_equal(*expect, ctx.eval(prog)));
@@ -214,6 +220,7 @@ TEST("require that tensor concat works") {
     const auto &fun = concat(inject(ValueType::from_spec("tensor(x[2])"), a_id, ctx.stash),
                              inject(ValueType::from_spec("tensor(x[2])"), b_id, ctx.stash),
                              "y", ctx.stash);
+    EXPECT_TRUE(fun.result_is_mutable());
     EXPECT_EQUAL(expect->type(), fun.result_type());
     const auto &prog = ctx.compile(fun);
     TEST_DO(verify_equal(*expect, ctx.eval(prog)));
@@ -225,6 +232,7 @@ TEST("require that tensor rename works") {
     Value::UP expect = ctx.make_tensor_matrix_renamed();
     const auto &fun = rename(inject(ValueType::from_spec("tensor(x[2],y[2])"), a_id, ctx.stash),
                              {"x"}, {"z"}, ctx.stash);
+    EXPECT_TRUE(fun.result_is_mutable());
     EXPECT_EQUAL(expect->type(), fun.result_type());
     const auto &prog = ctx.compile(fun);
     TEST_DO(verify_equal(*expect, ctx.eval(prog)));
@@ -240,6 +248,7 @@ TEST("require that if_node works") {
     const auto &fun = if_node(inject(ValueType::double_type(), a_id, ctx.stash),
                               inject(ValueType::from_spec("tensor(x[2])"), b_id, ctx.stash),
                               inject(ValueType::from_spec("tensor(x[2])"), c_id, ctx.stash), ctx.stash);
+    EXPECT_TRUE(!fun.result_is_mutable());
     EXPECT_EQUAL(expect_true->type(), fun.result_type());
     const auto &prog = ctx.compile(fun);
     TEST_DO(verify_equal(*expect_true, ctx.eval(prog)));
@@ -247,6 +256,22 @@ TEST("require that if_node works") {
     TEST_DO(verify_equal(*expect_false, ctx.eval(prog)));
 }
 
+TEST("require that if_node result is mutable only when both children produce mutable results") {
+    Stash stash;
+    const Node &cond = inject(DoubleValue::double_type(), 0, stash);
+    const Node &a = inject(ValueType::from_spec("tensor(x[2])"), 0, stash);
+    const Node &b = inject(ValueType::from_spec("tensor(x[3])"), 0, stash);
+    const Node &tmp = concat(a, b, "x", stash); // will be mutable
+    const Node &if_con_con = if_node(cond, a, b, stash);
+    const Node &if_mut_con = if_node(cond, tmp, b, stash);
+    const Node &if_con_mut = if_node(cond, a, tmp, stash);
+    const Node &if_mut_mut = if_node(cond, tmp, tmp, stash);
+    EXPECT_TRUE(!if_con_con.result_is_mutable());
+    EXPECT_TRUE(!if_mut_con.result_is_mutable());
+    EXPECT_TRUE(!if_con_mut.result_is_mutable());
+    EXPECT_TRUE(if_mut_mut.result_is_mutable());
+}
+
 TEST("require that if_node gets expected result type") {
     Stash stash;
     const Node &a = inject(DoubleValue::double_type(), 0, stash);
diff --git a/eval/src/tests/tensor/dense_dot_product_function/dense_dot_product_function_test.cpp b/eval/src/tests/tensor/dense_dot_product_function/dense_dot_product_function_test.cpp
index fb48e445180..37f9602565d 100644
--- a/eval/src/tests/tensor/dense_dot_product_function/dense_dot_product_function_test.cpp
+++ b/eval/src/tests/tensor/dense_dot_product_function/dense_dot_product_function_test.cpp
@@ -131,7 +131,8 @@ void assertOptimized(const vespalib::string &expr) {
     EvalFixture fixture(prod_engine, expr, param_repo, true);
     EXPECT_EQUAL(fixture.result(), EvalFixture::ref(expr, param_repo));
     auto info = fixture.find_all<DenseDotProductFunction>();
-    EXPECT_EQUAL(info.size(), 1u);
+    ASSERT_EQUAL(info.size(), 1u);
+    EXPECT_TRUE(info[0]->result_is_mutable());
 }
 
 void assertNotOptimized(const vespalib::string &expr) {
diff --git a/eval/src/tests/tensor/dense_fast_rename_function/dense_fast_rename_function_test.cpp b/eval/src/tests/tensor/dense_fast_rename_function/dense_fast_rename_function_test.cpp
index fab16f1e276..45b38b48481 100644
--- a/eval/src/tests/tensor/dense_fast_rename_function/dense_fast_rename_function_test.cpp
+++ b/eval/src/tests/tensor/dense_fast_rename_function/dense_fast_rename_function_test.cpp
@@ -24,17 +24,19 @@ const TensorEngine &prod_engine = DefaultTensorEngine::ref();
 EvalFixture::ParamRepo make_params() {
     return EvalFixture::ParamRepo()
         .add("x5", spec({x(5)}, N()))
+        .add_mutable("mut_x5", spec({x(5)}, N()))
         .add("x5_u", spec({x(5)}, N()), "tensor(x[])")
         .add("x_m", spec({x({"a", "b", "c"})}, N()))
         .add("x5y3", spec({x(5),y(3)}, N()));
 }
 EvalFixture::ParamRepo param_repo = make_params();
 
-void verify_optimized(const vespalib::string &expr) {
-    EvalFixture fixture(prod_engine, expr, param_repo, true);
+void verify_optimized(const vespalib::string &expr, bool expect_mutable = false) {
+    EvalFixture fixture(prod_engine, expr, param_repo, true, true);
     EXPECT_EQUAL(fixture.result(), EvalFixture::ref(expr, param_repo));
     auto info = fixture.find_all<DenseFastRenameFunction>();
-    EXPECT_EQUAL(info.size(), 1u);
+    ASSERT_EQUAL(info.size(), 1u);
+    EXPECT_EQUAL(info[0]->result_is_mutable(), expect_mutable);
 }
 
 void verify_not_optimized(const vespalib::string &expr) {
@@ -71,4 +73,25 @@ TEST("require that non-dense renames are not optimized") {
     TEST_DO(verify_not_optimized("rename(x_m,x,y)"));
 }
 
+TEST("require that renaming a mutable result retains mutability") {
+    TEST_DO(verify_optimized("rename(mut_x5,x,y)", true));
+}
+
+TEST("require that child mutability changed under-the-hood is still reflected") {
+    Stash stash;
+    const Node &a = inject(ValueType::from_spec("tensor(x[2])"), 0, stash);
+    const Node &tmp = map(a, operation::Neg::f, stash); // will be mutable
+    DenseFastRenameFunction my_rename(ValueType::from_spec("tensor(y[2])"), a);
+    EXPECT_TRUE(!my_rename.result_is_mutable());
+    {
+        std::vector<TensorFunction::Child::CREF> list;
+        my_rename.push_children(list);
+        ASSERT_EQUAL(list.size(), 1u);
+        EXPECT_EQUAL(&(list[0].get().get()), &a);
+        const TensorFunction::Child &child = list[0];
+        child.set(tmp);
+    }
+    EXPECT_TRUE(my_rename.result_is_mutable());
+}
+
 TEST_MAIN() { TEST_RUN_ALL(); }
diff --git a/eval/src/tests/tensor/vector_from_doubles_function/vector_from_doubles_function_test.cpp b/eval/src/tests/tensor/vector_from_doubles_function/vector_from_doubles_function_test.cpp
index 0ba9871d672..41d98122e0f 100644
--- a/eval/src/tests/tensor/vector_from_doubles_function/vector_from_doubles_function_test.cpp
+++ b/eval/src/tests/tensor/vector_from_doubles_function/vector_from_doubles_function_test.cpp
@@ -1,164 +1,63 @@
 // Copyright 2017 Yahoo Holdings. Licensed under the terms of the Apache 2.0 license. See LICENSE in the project root.
 
-#include <vespa/log/log.h>
-LOG_SETUP("dense_dot_product_function_test");
-
 #include <vespa/vespalib/testkit/test_kit.h>
 #include <vespa/eval/eval/tensor_function.h>
-#include <vespa/eval/eval/operation.h>
 #include <vespa/eval/eval/simple_tensor.h>
 #include <vespa/eval/eval/simple_tensor_engine.h>
 #include <vespa/eval/tensor/default_tensor_engine.h>
 #include <vespa/eval/tensor/dense/vector_from_doubles_function.h>
 #include <vespa/eval/tensor/dense/dense_tensor.h>
-#include <vespa/eval/tensor/dense/dense_tensor_builder.h>
-#include <vespa/eval/tensor/dense/dense_tensor_view.h>
+#include <vespa/eval/eval/test/tensor_model.hpp>
+#include <vespa/eval/eval/test/eval_fixture.h>
 
 #include <vespa/vespalib/util/stringfmt.h>
 #include <vespa/vespalib/util/stash.h>
 
 using namespace vespalib;
 using namespace vespalib::eval;
+using namespace vespalib::eval::test;
 using namespace vespalib::tensor;
 using namespace vespalib::eval::tensor_function;
 
-const TensorEngine &ref_engine = SimpleTensorEngine::ref();
 const TensorEngine &prod_engine = DefaultTensorEngine::ref();
 
-//-----------------------------------------------------------------------------
-// verify that optimize() works as expected
-
-template<typename OPT>
-bool treeContains(const TensorFunction &expr) {
-    using Child = TensorFunction::Child;
-    Child root(expr);
-    std::vector<Child::CREF> nodes({root});
-    for (size_t i = 0; i < nodes.size(); ++i) {
-        nodes[i].get().get().push_children(nodes);
-    }
-    for (const Child &child : nodes) {
-        if (as<OPT>(child.get())) {
-            return true;
-        }
-    }
-    return false;
-}
-
-const TensorFunction &optimize_fun(const Function &fun, const NodeTypes &node_types, Stash &stash) {
-    const TensorFunction &plain_fun = make_tensor_function(prod_engine, fun.root(), node_types, stash);
-    return prod_engine.optimize(plain_fun, stash);
+EvalFixture::ParamRepo make_params() {
+    return EvalFixture::ParamRepo()
+        .add("a", spec(1.0))
+        .add("b", spec(2.0))
+        .add("c", spec(3.0))
+        .add("d", spec(4.0))
+        .add("x5", spec({x(5)}, N()));
 }
+EvalFixture::ParamRepo param_repo = make_params();
 
-std::vector<ValueType> extract_types(size_t n, const std::vector<TensorSpec> &input) {
-    std::vector<ValueType> vec;
-    for (const TensorSpec &spec : input) {
-        vec.push_back(ValueType::from_spec(spec.type()));
-    }
-    while (vec.size() < n) {
-        vec.push_back(ValueType::double_type());
+void verify(const vespalib::string &expr, size_t expect_optimized_cnt, size_t expect_not_optimized_cnt) {
+    EvalFixture fixture(prod_engine, expr, param_repo, true);
+    EXPECT_EQUAL(fixture.result(), EvalFixture::ref(expr, param_repo));
+    auto info = fixture.find_all<VectorFromDoublesFunction>();
+    EXPECT_EQUAL(info.size(), expect_optimized_cnt);
+    for (size_t i = 0; i < info.size(); ++i) {
+        EXPECT_TRUE(info[i]->result_is_mutable());
     }
-    return vec;
+    EXPECT_EQUAL(fixture.find_all<Concat>().size(), expect_not_optimized_cnt);
 }
 
-struct Context {
-    Stash stash;
-    Function function;
-    std::vector<TensorSpec> input;
-    std::vector<ValueType> input_types;
-    NodeTypes node_types;
-    const TensorFunction &optimized;
-
-    Context(const vespalib::string &expr, std::vector<TensorSpec> in)
-        : stash(),
-          function(Function::parse(expr)),
-          input(in),
-          input_types(extract_types(function.num_params(), input)),
-          node_types(function, input_types),
-          optimized(optimize_fun(function, node_types, stash))
-    {
-        EXPECT_EQUAL(actual(), expected());
-    }
-
-    ~Context() {}
-
-    struct Params : LazyParams {
-        std::vector<Value::UP> values;
-        Value &resolve(size_t idx, Stash &) const override {
-            return *values[idx];
-        }
-    };
-
-    Params gen_params(const TensorEngine &engine) {
-        Params p;
-        for (const TensorSpec &spec : input) {
-            p.values.emplace_back(engine.from_spec(spec));
-        }
-        while (p.values.size() < function.num_params()) {
-            double v = 1.0 + p.values.size();
-            p.values.emplace_back(std::make_unique<DoubleValue>(v));
-        }
-        return p;
-    }
-
-    TensorSpec actual() {
-        const LazyParams &params = gen_params(prod_engine);
-        InterpretedFunction prodIfun(prod_engine, optimized);
-        InterpretedFunction::Context prodIctx(prodIfun);
-        const Value &result = prodIfun.eval(prodIctx, params);
-        return prod_engine.to_spec(result);
-    }
-
-    TensorSpec expected() {
-        const LazyParams &params = gen_params(ref_engine);
-        InterpretedFunction refIfun(ref_engine, function, NodeTypes());
-        InterpretedFunction::Context refIctx(refIfun);
-        const Value &result = refIfun.eval(refIctx, params);
-        return ref_engine.to_spec(result);
-    }
-
-};
-
 //-----------------------------------------------------------------------------
 
-void verify_all_optimized(const vespalib::string &expr) {
-    Context context(expr, {});
-    EXPECT_TRUE(treeContains<VectorFromDoublesFunction>(context.optimized));
-    EXPECT_FALSE(treeContains<eval::tensor_function::Concat>(context.optimized));
-}
-
 TEST("require that multiple concats are optimized") {
-    TEST_DO(verify_all_optimized("concat(a,b,x)"));
-    TEST_DO(verify_all_optimized("concat(a,concat(b,concat(c,d,x),x),x)"));
-    TEST_DO(verify_all_optimized("concat(concat(concat(a,b,x),c,x),d,x)"));
-    TEST_DO(verify_all_optimized("concat(concat(a,b,x),concat(c,d,x),x)"));
-}
-
-//-----------------------------------------------------------------------------
-
-void verify_some_optimized(const vespalib::string &expr) {
-    Context context(expr, {});
-    EXPECT_TRUE(treeContains<VectorFromDoublesFunction>(context.optimized));
-    EXPECT_TRUE(treeContains<eval::tensor_function::Concat>(context.optimized));
+    TEST_DO(verify("concat(a,b,x)", 1, 0));
+    TEST_DO(verify("concat(a,concat(b,concat(c,d,x),x),x)", 1, 0));
+    TEST_DO(verify("concat(concat(concat(a,b,x),c,x),d,x)", 1, 0));
+    TEST_DO(verify("concat(concat(a,b,x),concat(c,d,x),x)", 1, 0));
 }
 
 TEST("require that concat along different dimension is not optimized") {
-    TEST_DO(verify_some_optimized("concat(concat(a,b,x),concat(c,d,x),y)"));
+    TEST_DO(verify("concat(concat(a,b,x),concat(c,d,x),y)", 2, 1));
 }
 
-//-----------------------------------------------------------------------------
-
 TEST("require that concat of vector and double is not optimized") {
-    TensorSpec vecspec = TensorSpec("tensor(x[3])")
-                         .add({{"x", 0}}, 7.0)
-                         .add({{"x", 1}}, 11.0)
-                         .add({{"x", 2}}, 13.0);
-    TensorSpec dblspec = TensorSpec("double")
-                         .add({}, 19.0);
-    Context context("concat(a,b,x)", {vecspec, dblspec});
-    EXPECT_TRUE(treeContains<eval::tensor_function::Concat>(context.optimized));
-    EXPECT_FALSE(treeContains<VectorFromDoublesFunction>(context.optimized));
+    TEST_DO(verify("concat(a,x5,x)", 0, 1));
+    TEST_DO(verify("concat(x5,b,x)", 0, 1));
 }
 
-//-----------------------------------------------------------------------------
-
 TEST_MAIN() { TEST_RUN_ALL(); }
diff --git a/eval/src/vespa/eval/eval/tensor_function.h b/eval/src/vespa/eval/eval/tensor_function.h
index 2c3e9d21e0b..4e989d251c1 100644
--- a/eval/src/vespa/eval/eval/tensor_function.h
+++ b/eval/src/vespa/eval/eval/tensor_function.h
@@ -69,7 +69,7 @@ struct TensorFunction
         void set(const TensorFunction &child) const { ptr = &child; }
     };
     virtual const ValueType &result_type() const = 0;
-    virtual bool result_is_mutable() const { return false; }
+    virtual bool result_is_mutable() const = 0;
 
     /**
      * Push references to all children (NB: implementation must use
@@ -162,6 +162,7 @@ private:
     const Value &_value;
 public:
     ConstValue(const Value &value_in) : Leaf(value_in.type()), _value(value_in) {}
+    bool result_is_mutable() const override { return false; }
     InterpretedFunction::Instruction compile_self(Stash &stash) const final override;
 };
 
@@ -175,6 +176,7 @@ public:
     Inject(const ValueType &result_type_in, size_t param_idx_in)
         : Leaf(result_type_in), _param_idx(param_idx_in) {}
     size_t param_idx() const { return _param_idx; }
+    bool result_is_mutable() const override { return false; }
     InterpretedFunction::Instruction compile_self(Stash &stash) const final override;
 };
 
@@ -193,6 +195,7 @@ public:
         : Op1(result_type_in, child_in), _aggr(aggr_in), _dimensions(dimensions_in) {}
     Aggr aggr() const { return _aggr; }
     const std::vector<vespalib::string> &dimensions() const { return _dimensions; }
+    bool result_is_mutable() const override { return true; }
     InterpretedFunction::Instruction compile_self(Stash &stash) const final override;
 };
 
@@ -208,7 +211,8 @@ public:
         map_fun_t function_in)
         : Op1(result_type_in, child_in), _function(function_in) {}
     map_fun_t function() const { return _function; }
-    InterpretedFunction::Instruction compile_self(Stash &stash) const final override;
+    bool result_is_mutable() const override { return true; }
+    InterpretedFunction::Instruction compile_self(Stash &stash) const override;
 };
 
 //-----------------------------------------------------------------------------
@@ -224,7 +228,8 @@ public:
          join_fun_t function_in)
         : Op2(result_type_in, lhs_in, rhs_in), _function(function_in) {}
     join_fun_t function() const { return _function; }
-    InterpretedFunction::Instruction compile_self(Stash &stash) const final override;
+    bool result_is_mutable() const override { return true; }
+    InterpretedFunction::Instruction compile_self(Stash &stash) const override;
 };
 
 //-----------------------------------------------------------------------------
@@ -240,6 +245,7 @@ public:
            const vespalib::string &dimension_in)
         : Op2(result_type_in, lhs_in, rhs_in), _dimension(dimension_in) {}
     const vespalib::string &dimension() const { return _dimension; }
+    bool result_is_mutable() const override { return true; }
     InterpretedFunction::Instruction compile_self(Stash &stash) const final override;
 };
 
@@ -258,6 +264,7 @@ public:
         : Op1(result_type_in, child_in), _from(from_in), _to(to_in) {}
     const std::vector<vespalib::string> &from() const { return _from; }
     const std::vector<vespalib::string> &to() const { return _to; }
+    bool result_is_mutable() const override { return true; }
     InterpretedFunction::Instruction compile_self(Stash &stash) const final override;
 };
 
@@ -279,6 +286,10 @@ public:
     const TensorFunction &true_child() const { return _true_child.get(); }
     const TensorFunction &false_child() const { return _false_child.get(); }
     void push_children(std::vector<Child::CREF> &children) const final override;    
+    bool result_is_mutable() const override {
+        return (true_child().result_is_mutable() &&
+                false_child().result_is_mutable());
+    }
     InterpretedFunction::Instruction compile_self(Stash &stash) const final override;
 };
 
diff --git a/eval/src/vespa/eval/tensor/dense/dense_dot_product_function.cpp b/eval/src/vespa/eval/tensor/dense/dense_dot_product_function.cpp
index 7985b59da56..ae217935fd9 100644
--- a/eval/src/vespa/eval/tensor/dense/dense_dot_product_function.cpp
+++ b/eval/src/vespa/eval/tensor/dense/dense_dot_product_function.cpp
@@ -24,7 +24,7 @@ CellsRef getCellsRef(const eval::Value &value) {
     return denseTensor.cellsRef();
 }
 
-void my_op(eval::InterpretedFunction::State &state, uint64_t param) {
+void my_dot_product_op(eval::InterpretedFunction::State &state, uint64_t param) {
     auto *hw_accelerator = (hwaccelrated::IAccelrated *)(param);
     DenseTensorView::CellsRef lhsCells = getCellsRef(state.peek(1));
     DenseTensorView::CellsRef rhsCells = getCellsRef(state.peek(0));
@@ -56,7 +56,7 @@ DenseDotProductFunction::DenseDotProductFunction(const eval::TensorFunction &lhs
 eval::InterpretedFunction::Instruction
 DenseDotProductFunction::compile_self(Stash &) const
 {
-    return eval::InterpretedFunction::Instruction(my_op, (uint64_t)(_hwAccelerator.get()));
+    return eval::InterpretedFunction::Instruction(my_dot_product_op, (uint64_t)(_hwAccelerator.get()));
 }
 
 const TensorFunction &
diff --git a/eval/src/vespa/eval/tensor/dense/dense_dot_product_function.h b/eval/src/vespa/eval/tensor/dense/dense_dot_product_function.h
index eec7448f041..46b04a446d4 100644
--- a/eval/src/vespa/eval/tensor/dense/dense_dot_product_function.h
+++ b/eval/src/vespa/eval/tensor/dense/dense_dot_product_function.h
@@ -19,6 +19,7 @@ public:
     DenseDotProductFunction(const eval::TensorFunction &lhs_in,
                             const eval::TensorFunction &rhs_in);
     eval::InterpretedFunction::Instruction compile_self(Stash &stash) const override;
+    bool result_is_mutable() const override { return true; }
     static const eval::TensorFunction &optimize(const eval::TensorFunction &expr, Stash &stash);
 };
 
diff --git a/eval/src/vespa/eval/tensor/dense/dense_fast_rename_function.h b/eval/src/vespa/eval/tensor/dense/dense_fast_rename_function.h
index e7de8e95ff0..1ca61d52915 100644
--- a/eval/src/vespa/eval/tensor/dense/dense_fast_rename_function.h
+++ b/eval/src/vespa/eval/tensor/dense/dense_fast_rename_function.h
@@ -17,6 +17,7 @@ public:
                             const eval::TensorFunction &child);
     ~DenseFastRenameFunction();
     eval::InterpretedFunction::Instruction compile_self(Stash &stash) const override;
+    bool result_is_mutable() const override { return child().result_is_mutable(); }
     static const eval::TensorFunction &optimize(const eval::TensorFunction &expr, Stash &stash);
 };
 
diff --git a/eval/src/vespa/eval/tensor/dense/dense_inplace_join_function.cpp b/eval/src/vespa/eval/tensor/dense/dense_inplace_join_function.cpp
index 53a5fe9bb27..78a407062c6 100644
--- a/eval/src/vespa/eval/tensor/dense/dense_inplace_join_function.cpp
+++ b/eval/src/vespa/eval/tensor/dense/dense_inplace_join_function.cpp
@@ -51,8 +51,7 @@ DenseInplaceJoinFunction::DenseInplaceJoinFunction(const ValueType &result_type,
                                                    const TensorFunction &rhs,
                                                    join_fun_t function_in,
                                                    bool write_left_in)
-    : eval::tensor_function::Op2(result_type, lhs, rhs),
-      _function(function_in),
+    : eval::tensor_function::Join(result_type, lhs, rhs, function_in),
       _write_left(write_left_in)
 {
 }
@@ -65,7 +64,7 @@ eval::InterpretedFunction::Instruction
 DenseInplaceJoinFunction::compile_self(Stash &) const
 {
     auto op = _write_left ? my_inplace_join_op<true> : my_inplace_join_op<false>;
-    return eval::InterpretedFunction::Instruction(op, (uint64_t)_function);
+    return eval::InterpretedFunction::Instruction(op, (uint64_t)function());
 }
 
 const TensorFunction &
diff --git a/eval/src/vespa/eval/tensor/dense/dense_inplace_join_function.h b/eval/src/vespa/eval/tensor/dense/dense_inplace_join_function.h
index de2cdae3778..d31ce99dda4 100644
--- a/eval/src/vespa/eval/tensor/dense/dense_inplace_join_function.h
+++ b/eval/src/vespa/eval/tensor/dense/dense_inplace_join_function.h
@@ -9,12 +9,11 @@ namespace vespalib::tensor {
 /**
  * Tensor function for inplace join operation on mutable dense tensors.
  **/
-class DenseInplaceJoinFunction : public eval::tensor_function::Op2
+class DenseInplaceJoinFunction : public eval::tensor_function::Join
 {
 public:
     using join_fun_t = ::vespalib::eval::tensor_function::join_fun_t;
 private:
-    join_fun_t _function;
     bool _write_left;
 public:
     DenseInplaceJoinFunction(const eval::ValueType &result_type,
@@ -23,7 +22,6 @@ public:
                              join_fun_t function_in,
                              bool write_left_in);
     ~DenseInplaceJoinFunction();
-    join_fun_t function() const { return _function; }
     bool write_left() const { return _write_left; }
     bool result_is_mutable() const override { return true; }
     eval::InterpretedFunction::Instruction compile_self(Stash &stash) const override;
diff --git a/eval/src/vespa/eval/tensor/dense/dense_inplace_map_function.cpp b/eval/src/vespa/eval/tensor/dense/dense_inplace_map_function.cpp
index 162bdb2ebfe..a0aba25f342 100644
--- a/eval/src/vespa/eval/tensor/dense/dense_inplace_map_function.cpp
+++ b/eval/src/vespa/eval/tensor/dense/dense_inplace_map_function.cpp
@@ -38,8 +38,7 @@ bool isConcreteDenseTensor(const ValueType &type) {
 DenseInplaceMapFunction::DenseInplaceMapFunction(const eval::ValueType &result_type,
                                                  const eval::TensorFunction &child,
                                                  map_fun_t function_in)
-    : eval::tensor_function::Op1(result_type, child),
-      _function(function_in)
+    : eval::tensor_function::Map(result_type, child, function_in)
 {
 }
 
@@ -50,7 +49,7 @@ DenseInplaceMapFunction::~DenseInplaceMapFunction()
 eval::InterpretedFunction::Instruction
 DenseInplaceMapFunction::compile_self(Stash &) const
 {
-    return eval::InterpretedFunction::Instruction(my_inplace_map_op, (uint64_t)_function);
+    return eval::InterpretedFunction::Instruction(my_inplace_map_op, (uint64_t)function());
 }
 
 const TensorFunction &
diff --git a/eval/src/vespa/eval/tensor/dense/dense_inplace_map_function.h b/eval/src/vespa/eval/tensor/dense/dense_inplace_map_function.h
index f02f83edae1..bfa5760aa3c 100644
--- a/eval/src/vespa/eval/tensor/dense/dense_inplace_map_function.h
+++ b/eval/src/vespa/eval/tensor/dense/dense_inplace_map_function.h
@@ -9,7 +9,7 @@ namespace vespalib::tensor {
 /**
  * Tensor function for inplace map operation on mutable dense tensors.
  **/
-class DenseInplaceMapFunction : public eval::tensor_function::Op1
+class DenseInplaceMapFunction : public eval::tensor_function::Map
 {
 public:
     using map_fun_t = ::vespalib::eval::tensor_function::map_fun_t;
@@ -20,7 +20,6 @@ public:
                             const eval::TensorFunction &child,
                             map_fun_t function_in);
     ~DenseInplaceMapFunction();
-    map_fun_t function() const { return _function; }
     bool result_is_mutable() const override { return true; }
     eval::InterpretedFunction::Instruction compile_self(Stash &stash) const override;
     static const eval::TensorFunction &optimize(const eval::TensorFunction &expr, Stash &stash);
diff --git a/eval/src/vespa/eval/tensor/dense/dense_xw_product_function.cpp b/eval/src/vespa/eval/tensor/dense/dense_xw_product_function.cpp
index fa07028ef27..24b1d8abfcf 100644
--- a/eval/src/vespa/eval/tensor/dense/dense_xw_product_function.cpp
+++ b/eval/src/vespa/eval/tensor/dense/dense_xw_product_function.cpp
@@ -58,7 +58,7 @@ void transposedProduct(const DenseXWProductFunction::Self &self,
 }
 
 template <bool commonDimensionInnermost>
-void my_op(eval::InterpretedFunction::State &state, uint64_t param) {
+void my_xw_product_op(eval::InterpretedFunction::State &state, uint64_t param) {
     DenseXWProductFunction::Self *self = (DenseXWProductFunction::Self *)(param);
 
     CellsRef vectorCells = getCellsRef(state.peek(1));
@@ -129,7 +129,7 @@ eval::InterpretedFunction::Instruction
 DenseXWProductFunction::compile_self(Stash &stash) const
 {
     Self &self = stash.create<Self>(result_type(), _vectorSize, _resultSize);
-    auto op = _commonDimensionInnermost ? my_op<true> : my_op<false>;
+    auto op = _commonDimensionInnermost ? my_xw_product_op<true> : my_xw_product_op<false>;
     return eval::InterpretedFunction::Instruction(op, (uint64_t)(&self));
 }
 
diff --git a/eval/src/vespa/eval/tensor/dense/vector_from_doubles_function.h b/eval/src/vespa/eval/tensor/dense/vector_from_doubles_function.h
index 417c60c2aca..378c9026f84 100644
--- a/eval/src/vespa/eval/tensor/dense/vector_from_doubles_function.h
+++ b/eval/src/vespa/eval/tensor/dense/vector_from_doubles_function.h
@@ -31,6 +31,7 @@ public:
     }
     size_t size() const { return _self.resultSize; }
     eval::InterpretedFunction::Instruction compile_self(Stash &stash) const override;
+    bool result_is_mutable() const override { return true; }
     static const eval::TensorFunction &optimize(const eval::TensorFunction &expr, Stash &stash);
 };