Add LateFunctionBinding declaration and fix constant folding

Adds the declaration for LateFunctionBindings, which can be used to
indicate to the Runtime and the Optimizers that the function will
be bound at runtime through the Activation.

This lets the constant folding optimizer know that the function
potentially has side effects and cannot be folded. Without this
the optimization will fail with an error for late bound functions
where all arguments are constants.

The implementation for late bound functions will be added in a
subsequent commit.

Author: zeitgeist87

cel/decls.go

@@ -330,6 +330,12 @@ func FunctionBinding(binding functions.FunctionOp) OverloadOpt {
 	return decls.FunctionBinding(binding)
 }
 
+// LateFunctionBinding indicates that the function has a binding which is not known at compile time.
+// This is useful for functions which have side-effects or are not deterministically computable.
+func LateFunctionBinding() OverloadOpt {
+	return decls.LateFunctionBinding()
+}
+
 // OverloadIsNonStrict enables the function to be called with error and unknown argument values.
 //
 // Note: do not use this option unless absoluately necessary as it should be an uncommon feature.

cel/folding.go

@@ -68,7 +68,8 @@ func (opt *constantFoldingOptimizer) Optimize(ctx *OptimizerContext, a *ast.AST)
 	// Walk the list of foldable expression and continue to fold until there are no more folds left.
 	// All of the fold candidates returned by the constantExprMatcher should succeed unless there's
 	// a logic bug with the selection of expressions.
-	foldableExprs := ast.MatchDescendants(root, constantExprMatcher)
+	constantExprMatcherCapture := func(e ast.NavigableExpr) bool { return constantExprMatcher(ctx, a, e) }
+	foldableExprs := ast.MatchDescendants(root, constantExprMatcherCapture)
 	foldCount := 0
 	for len(foldableExprs) != 0 && foldCount < opt.maxFoldIterations {
 		for _, fold := range foldableExprs {
@@ -77,6 +78,10 @@ func (opt *constantFoldingOptimizer) Optimize(ctx *OptimizerContext, a *ast.AST)
 			if fold.Kind() == ast.CallKind && maybePruneBranches(ctx, fold) {
 				continue
 			}
+			// Late-bound function calls cannot be folded.
+			if fold.Kind() == ast.CallKind && isLateBoundFunctionCall(ctx, a, fold) {
+				continue
+			}
 			// Otherwise, assume all context is needed to evaluate the expression.
 			err := tryFold(ctx, a, fold)
 			if err != nil {
@@ -85,7 +90,7 @@ func (opt *constantFoldingOptimizer) Optimize(ctx *OptimizerContext, a *ast.AST)
 			}
 		}
 		foldCount++
-		foldableExprs = ast.MatchDescendants(root, constantExprMatcher)
+		foldableExprs = ast.MatchDescendants(root, constantExprMatcherCapture)
 	}
 	// Once all of the constants have been folded, try to run through the remaining comprehensions
 	// one last time. In this case, there's no guarantee they'll run, so we only update the
@@ -139,6 +144,25 @@ func tryFold(ctx *OptimizerContext, a *ast.AST, expr ast.Expr) error {
 	return nil
 }
 
+func isLateBoundFunctionCall(ctx *OptimizerContext, a *ast.AST, expr ast.Expr) bool {
+	call := expr.AsCall()
+	overloadIds := a.GetOverloadIDs(expr.ID())
+	for _, f := range ctx.Functions() {
+		if f.Name() != call.FunctionName() {
+			continue
+		}
+
+		for _, o := range f.OverloadDecls() {
+			for _, id := range overloadIds {
+				if id == o.ID() && o.HasLateBinding() {
+					return true
+				}
+			}
+		}
+	}
+	return false
+}
+
 // maybePruneBranches inspects the non-strict call expression to determine whether
 // a branch can be removed. Evaluation will naturally prune logical and / or calls,
 // but conditional will not be pruned cleanly, so this is one small area where the
@@ -455,7 +479,7 @@ func adaptLiteral(ctx *OptimizerContext, val ref.Val) (ast.Expr, error) {
 // Only comprehensions which are not nested are included as possible constant folds, and only
 // if all variables referenced in the comprehension stack exist are only iteration or
 // accumulation variables.
-func constantExprMatcher(e ast.NavigableExpr) bool {
+func constantExprMatcher(ctx *OptimizerContext, a *ast.AST, e ast.NavigableExpr) bool {
 	switch e.Kind() {
 	case ast.CallKind:
 		return constantCallMatcher(e)
@@ -477,6 +501,10 @@ func constantExprMatcher(e ast.NavigableExpr) bool {
 			if e.Kind() == ast.IdentKind && !vars[e.AsIdent()] {
 				constantExprs = false
 			}
+			// Late-bound function calls cannot be folded.
+			if e.Kind() == ast.CallKind && isLateBoundFunctionCall(ctx, a, e) {
+				constantExprs = false
+			}
 		})
 		ast.PreOrderVisit(e, visitor)
 		return constantExprs

cel/folding_test.go

@@ -17,12 +17,14 @@ package cel
 import (
 	"reflect"
 	"sort"
+	"strings"
 	"testing"
 
 	"google.golang.org/protobuf/encoding/prototext"
 	"google.golang.org/protobuf/proto"
 
 	"github.com/google/cel-go/common/ast"
+	"github.com/google/cel-go/common/types/ref"
 
 	proto3pb "github.com/google/cel-go/test/proto3pb"
 	exprpb "google.golang.org/genproto/googleapis/api/expr/v1alpha1"
@@ -313,6 +315,89 @@ func TestConstantFoldingOptimizer(t *testing.T) {
 	}
 }
 
+func TestConstantFoldingCallsWithSideEffects(t *testing.T) {
+	tests := []struct {
+		expr   string
+		folded string
+		error  string
+	}{
+		{
+			expr:   `noSideEffect(3)`,
+			folded: `3`,
+		},
+		{
+			expr:   `withSideEffect(3)`,
+			folded: `withSideEffect(3)`,
+		},
+		{
+			expr:   `[{}, {"a": 1}, {"b": 2}].exists(i, has(i.b) && withSideEffect(i.b) == 1)`,
+			folded: `[{}, {"a": 1}, {"b": 2}].exists(i, has(i.b) && withSideEffect(i.b) == 1)`,
+		},
+		{
+			expr:   `[{}, {"a": 1}, {"b": 2}].exists(i, has(i.b) && noSideEffect(i.b) == 2)`,
+			folded: `true`,
+		},
+		{
+			expr:  `noImpl(3)`,
+			error: `constant-folding evaluation failed: no such overload: noImpl`,
+		},
+	}
+	e, err := NewEnv(
+		OptionalTypes(),
+		EnableMacroCallTracking(),
+		Function("noSideEffect",
+			Overload("noSideEffect_int_int",
+				[]*Type{IntType},
+				IntType, FunctionBinding(func(args ...ref.Val) ref.Val {
+					return args[0]
+				}))),
+		Function("withSideEffect",
+			Overload("withSideEffect_int_int",
+				[]*Type{IntType},
+				IntType, LateFunctionBinding())),
+		Function("noImpl",
+			Overload("noImpl_int_int",
+				[]*Type{IntType},
+				IntType)),
+	)
+	if err != nil {
+		t.Fatalf("NewEnv() failed: %v", err)
+	}
+	for _, tst := range tests {
+		tc := tst
+		t.Run(tc.expr, func(t *testing.T) {
+			checked, iss := e.Compile(tc.expr)
+			if iss.Err() != nil {
+				t.Fatalf("Compile() failed: %v", iss.Err())
+			}
+			folder, err := NewConstantFoldingOptimizer()
+			if err != nil {
+				t.Fatalf("NewConstantFoldingOptimizer() failed: %v", err)
+			}
+			opt := NewStaticOptimizer(folder)
+			optimized, iss := opt.Optimize(e, checked)
+			if tc.error != "" {
+				if iss.Err() == nil {
+					t.Errorf("got nil, wanted error containing %q", tc.error)
+				} else if !strings.Contains(iss.Err().Error(), tc.error) {
+					t.Errorf("got %q, wanted error containing %q", iss.Err().Error(), tc.error)
+				}
+				return
+			}
+			if iss.Err() != nil {
+				t.Fatalf("Optimize() generated an invalid AST: %v", iss.Err())
+			}
+			folded, err := AstToString(optimized)
+			if err != nil {
+				t.Fatalf("AstToString() failed: %v", err)
+			}
+			if folded != tc.folded {
+				t.Errorf("got %q, wanted %q", folded, tc.folded)
+			}
+		})
+	}
+}
+
 func TestConstantFoldingOptimizerMacroElimination(t *testing.T) {
 	tests := []struct {
 		expr       string

common/decls/decls.go

@@ -232,6 +232,9 @@ func (f *FunctionDecl) AddOverload(overload *OverloadDecl) error {
 			}
 			return fmt.Errorf("overload redefinition in function. %s: %s has multiple definitions", f.Name(), oID)
 		}
+		if overload.HasLateBinding() != o.HasLateBinding() {
+			return fmt.Errorf("overload with late binding cannot be added to function %s: cannot mix late and non-late bindings", f.Name())
+		}
 	}
 	f.overloadOrdinals = append(f.overloadOrdinals, overload.ID())
 	f.overloads[overload.ID()] = overload
@@ -257,8 +260,10 @@ func (f *FunctionDecl) Bindings() ([]*functions.Overload, error) {
 	}
 	overloads := []*functions.Overload{}
 	nonStrict := false
+	hasLateBinding := false
 	for _, oID := range f.overloadOrdinals {
 		o := f.overloads[oID]
+		hasLateBinding = hasLateBinding || o.HasLateBinding()
 		if o.hasBinding() {
 			overload := &functions.Overload{
 				Operator:     o.ID(),
@@ -276,6 +281,9 @@ func (f *FunctionDecl) Bindings() ([]*functions.Overload, error) {
 		if len(overloads) != 0 {
 			return nil, fmt.Errorf("singleton function incompatible with specialized overloads: %s", f.Name())
 		}
+		if hasLateBinding {
+			return nil, fmt.Errorf("singleton function incompatible with late bindings: %s", f.Name())
+		}
 		overloads = []*functions.Overload{
 			{
 				Operator:     f.Name(),
@@ -516,6 +524,9 @@ type OverloadDecl struct {
 	argTypes         []*types.Type
 	resultType       *types.Type
 	isMemberFunction bool
+	// hasLateBinding indicates that the function has a binding which is not known at compile time.
+	// This is useful for functions which have side-effects or are not deterministically computable.
+	hasLateBinding bool
 	// nonStrict indicates that the function will accept error and unknown arguments as inputs.
 	nonStrict bool
 	// operandTrait indicates whether the member argument should have a specific type-trait.
@@ -571,6 +582,14 @@ func (o *OverloadDecl) IsNonStrict() bool {
 	return o.nonStrict
 }
 
+// HasLateBinding returns whether the overload has a binding which is not known at compile time.
+func (o *OverloadDecl) HasLateBinding() bool {
+	if o == nil {
+		return false
+	}
+	return o.hasLateBinding
+}
+
 // OperandTrait returns the trait mask of the first operand to the overload call, e.g.
 // `traits.Indexer`
 func (o *OverloadDecl) OperandTrait() int {
@@ -739,6 +758,9 @@ func UnaryBinding(binding functions.UnaryOp) OverloadOpt {
 		if len(o.ArgTypes()) != 1 {
 			return nil, fmt.Errorf("unary function bound to non-unary overload: %s", o.ID())
 		}
+		if o.hasLateBinding {
+			return nil, fmt.Errorf("overload already has a late binding: %s", o.ID())
+		}
 		o.unaryOp = binding
 		return o, nil
 	}
@@ -754,6 +776,9 @@ func BinaryBinding(binding functions.BinaryOp) OverloadOpt {
 		if len(o.ArgTypes()) != 2 {
 			return nil, fmt.Errorf("binary function bound to non-binary overload: %s", o.ID())
 		}
+		if o.hasLateBinding {
+			return nil, fmt.Errorf("overload already has a late binding: %s", o.ID())
+		}
 		o.binaryOp = binding
 		return o, nil
 	}
@@ -766,11 +791,26 @@ func FunctionBinding(binding functions.FunctionOp) OverloadOpt {
 		if o.hasBinding() {
 			return nil, fmt.Errorf("overload already has a binding: %s", o.ID())
 		}
+		if o.hasLateBinding {
+			return nil, fmt.Errorf("overload already has a late binding: %s", o.ID())
+		}
 		o.functionOp = binding
 		return o, nil
 	}
 }
 
+// LateFunctionBinding indicates that the function has a binding which is not known at compile time.
+// This is useful for functions which have side-effects or are not deterministically computable.
+func LateFunctionBinding() OverloadOpt {
+	return func(o *OverloadDecl) (*OverloadDecl, error) {
+		if o.hasBinding() {
+			return nil, fmt.Errorf("overload already has a binding: %s", o.ID())
+		}
+		o.hasLateBinding = true
+		return o, nil
+	}
+}
+
 // OverloadIsNonStrict enables the function to be called with error and unknown argument values.
 //
 // Note: do not use this option unless absoluately necessary as it should be an uncommon feature.

common/decls/decls_test.go

@@ -473,6 +473,24 @@ func TestSingletonOverloadCollision(t *testing.T) {
 	}
 }
 
+func TestSingletonOverloadLateBindingCollision(t *testing.T) {
+	fn, err := NewFunction("id",
+		Overload("id_any", []*types.Type{types.AnyType}, types.AnyType,
+			LateFunctionBinding(),
+		),
+		SingletonUnaryBinding(func(arg ref.Val) ref.Val {
+			return arg
+		}),
+	)
+	if err != nil {
+		t.Fatalf("NewFunction() failed: %v", err)
+	}
+	_, err = fn.Bindings()
+	if err == nil || !strings.Contains(err.Error(), "incompatible with late bindings") {
+		t.Errorf("NewFunction() got %v, wanted incompatible with late bindings", err)
+	}
+}
+
 func TestSingletonUnaryBindingRedefinition(t *testing.T) {
 	_, err := NewFunction("id",
 		Overload("id_any", []*types.Type{types.AnyType}, types.AnyType),
@@ -592,6 +610,74 @@ func TestOverloadFunctionBindingRedefinition(t *testing.T) {
 	}
 }
 
+func TestOverloadFunctionLateBinding(t *testing.T) {
+	function, err := NewFunction("id",
+		Overload("id_bool", []*types.Type{types.BoolType}, types.AnyType, LateFunctionBinding(), LateFunctionBinding()),
+	)
+	if err != nil {
+		t.Fatalf("NewFunction() failed: %v", err)
+	}
+	if len(function.OverloadDecls()) != 1 {
+		t.Fatalf("NewFunction() got %v, wanted 1 overload", function.OverloadDecls())
+	}
+	if !function.OverloadDecls()[0].HasLateBinding() {
+		t.Errorf("overload %v did not have a late binding", function.OverloadDecls()[0])
+	}
+}
+
+func TestOverloadFunctionMixLateAndNonLateBinding(t *testing.T) {
+	_, err := NewFunction("id",
+		Overload("id_bool", []*types.Type{types.BoolType}, types.AnyType, LateFunctionBinding()),
+		Overload("id_int", []*types.Type{types.IntType}, types.AnyType),
+	)
+	if err == nil || !strings.Contains(err.Error(), "cannot mix late and non-late bindings") {
+		t.Errorf("NewCustomEnv() got %v, wanted cannot mix late and non-late bindings", err)
+	}
+}
+
+func TestOverloadFunctionBindingWithLateBinding(t *testing.T) {
+	_, err := NewFunction("id",
+		Overload("id_bool", []*types.Type{types.BoolType}, types.AnyType, FunctionBinding(func(args ...ref.Val) ref.Val {
+			return args[0]
+		}), LateFunctionBinding()),
+	)
+	if err == nil || !strings.Contains(err.Error(), "already has a binding") {
+		t.Errorf("NewCustomEnv() got %v, wanted already has a binding", err)
+	}
+}
+
+func TestOverloadFunctionLateBindingWithBinding(t *testing.T) {
+	_, err := NewFunction("id",
+		Overload("id_bool", []*types.Type{types.BoolType}, types.AnyType, LateFunctionBinding(),
+			FunctionBinding(func(args ...ref.Val) ref.Val {
+				return args[0]
+			})),
+	)
+	if err == nil || !strings.Contains(err.Error(), "already has a late binding") {
+		t.Errorf("NewCustomEnv() got %v, wanted already has a late binding", err)
+	}
+
+	_, err = NewFunction("id",
+		Overload("id_bool", []*types.Type{types.BoolType}, types.AnyType, LateFunctionBinding(),
+			UnaryBinding(func(arg ref.Val) ref.Val {
+				return arg
+			})),
+	)
+	if err == nil || !strings.Contains(err.Error(), "already has a late binding") {
+		t.Errorf("NewCustomEnv() got %v, wanted already has a late binding", err)
+	}
+
+	_, err = NewFunction("id",
+		Overload("id_bool", []*types.Type{types.BoolType, types.BoolType}, types.AnyType, LateFunctionBinding(),
+			BinaryBinding(func(arg1 ref.Val, arg2 ref.Val) ref.Val {
+				return arg1
+			})),
+	)
+	if err == nil || !strings.Contains(err.Error(), "already has a late binding") {
+		t.Errorf("NewCustomEnv() got %v, wanted already has a late binding", err)
+	}
+}
+
 func TestOverloadIsNonStrict(t *testing.T) {
 	fn, err := NewFunction("getOrDefault",
 		MemberOverload("get",