cuelang.org/go@v0.13.0/internal/core/adt/typocheck.go

// Copyright 2025 CUE Authors
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//     https://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

package adt

// This file holds CUE's algorithm to detect misspelled field names.
//
// ## Outline
//
// Typo checking MAY be enabled whenever a node is unified with a definition or
// a struct returned from the close builtin. Each distinct unification of such a
// struct triggers a separate "typo check", which is associated with a unique
// identifier. This identifier is passed along all values that result from
// unification with this struct.
//
// Once the processing of a node is complete, it is checked that for all its
// fields there is supportive evidence that the field is allowed for all structs
// for which typo checking is enabled.
//
// ## Selecting which Structs to Typo Check
//
// The algorithm is quite general and allows many types of heuristics to be
// applied. The initial goal is to mimic V2 semantics as closely as possible to
// facilitate migration to V3.
//
// ### Embeddings
//
// Embeddings provide evidence allowing fields for their enclosing scopes. Even
// when an embedding references a definition, it will not start its own typo
// check.
//
// ### Definitions
//
// By default, all references to non-embedded definitions are typo checked. The
// following situations qualify.
//
//      #A: b: {a: int}
//      a: #A
//
//      // Trigger typo check
//      r1: #A
//      r2: a
//
//      // Do NOT trigger typo check
//      r3: a.b
//
// In the case of r3, no typo check is triggered as the inserted value does not
// reference a definition directly. This choice is somewhat arbitrary. The main
// reason to pick this semantics is to be compatible with V2.
//
// ### Inline structs
//
// Like in V2, inline structs are generally not typo checked. We can add this
// later if necessary.
//
// ## Tracking Evidence
//
// The basic principle of this algorithm is that each (dynamic) reference is
// associated with a unique identifier. When a node is unified with a definition
// all its descendant nodes are tagged with this number if any of the conjuncts
// of this definition end up being unified with such nodes. Once a node finished
// processing, it is checked, recursively, that all its fields adhere to this
// schema by checking that there is supportive evidence that this schema allows
// such fields.
//
// Consider, for instance, the following CUE. Here, the reference to #Schema
// will be assigned the unique identifier 7.
//
// 	foo: #Schema & {
// 		a: 1  // marked with 7 as #Schema has a field `a`.
// 		b: 1  // marked with 7 as the pattern of #Schema allows `b`.
// 		c: 1  // not marked with 7 as #Schema does not have a field `c`.
// 	}
//
// 	#Schema: {
// 		a?:      int
// 		[<="b"]: int
// 	}
//
// Details of this algorithm are included below.
//
// ### Evidence sets and multiple insertions
//
// The same struct may be referred to within a node multiple times. If it is not
// itself a typo-checked definition, it may provide evidence for multiple other
// typo-checked definitions. To avoid having to process the conjuncts of such
// structs multiple times, we will still assign a unique identifier to such
// structs, and will annotate each typo-checked definition to which this applies
// that it is satisfied by this struct.
//
// In other words, each typo-checked struct may be associated with multiple
// unique identifiers that provide evidence for an allowed field. See the
// section below for a more detailed example.
//
// ### Embeddings
//
// Embeddings are a special case of structs that are not typo checked, but may
// provide evidence for other typo-checked definitions. As a struct associated
// with an embedding may be referred to multiple times, we will also assign a
// unique identifier to embeddings. This identifier is then also added to the
// evidence set of the typo-checked definition.
//
// Note that if a definition is embedded within a struct, that struct is
// considered closed after unifying all embeddings. We need to track this
// separately. This is discussed in the code below where appropriate. Ideally,
// we would get rid of this behavior by only allowing "open" or "closed"
// unifications, without considering any of the other context.
//
// Consider the following example:
//
// 	a: #A & { #A, b: int, c: int }
//  #A: #B
//  #B: { b: int }
//
// In this case, for `a`, we need find evidence that all fields of `#A` are
// allowed. If `a` were to be referenced by another field, we would also
// need to check that the struct with the embedded definition, which is
// closed after unification, is valid.
//
// So, for `a` we track two requirements, say `1`, which corresponds to the
// reference to `#A` and `2`, which corresponds to the literal struct.
// During evaluation, we additionally assign `3` to `#B`.
//
// The algorithm now proceeds as follows. We start with the requirement sets
// `{1}` and `{2}`. For the first, while evaluating definition `#A`, we find
// that this redirects to `#B`. In this case, as `#B` is at least as strict
// as `#A`, we can rewrite the set as `{3}`. This means we need to find
// evidence that each field of `a` is marked with a `3`.
// For the second case, `{2}`, we find that `#A` is embedded. As this is not
// a further restriction, we add the ID for `#A` to the set, resulting in
// `{2, 1}`. Subsequently, `#A` maps to `#B` again, but in this case, as `#A`
// is embedded, it is also additive, resulting in `{2, 1, 3}`, where each field
// in `a` needs to be marked with at least one of these values.
//
// After `a` is fully unified, field `b` will be marked with `3` and thus has
// evidence for both requirements. Field `c`, however, is marked with `2`, which
// is only supports the second requirement, and thus will result in a typo
// error.
//
// NOTE 1: that the second set is strictly more permissive then the first
// requirement and could be elided.
// NOTE 2: a reference to the same node within one field is only assigned a
// single ID and only needs to be processed once per node.
//
// ### Pruning of sub nodes
//
// For definitions, typo checking proceeds recursively. However, typo checking
// is disabled for certain fields, even when a node still has subfields. Typo
// checking is disabled if a field has:
//
// - a "naked" top `_` (so not, for instance `{_}`), - an ellipsis `...`, or -
// it has a "non-concrete" validator, like matchn.
//
// In the latter case, the builtin takes over the responsibility of checking the
// type.
//
import (
	"math"
	"slices"

	"cuelang.org/go/cue/ast"
)

type defID uint32

const deleteID defID = math.MaxUint32

func (c *OpContext) getNextDefID() defID {
	c.stats.NumCloseIDs++
	c.nextDefID++
	return c.nextDefID
}

type refInfo struct {
	v  *Vertex
	id defID

	// exclude defines a subtree of CloseInfo.def that should not be
	// transitively added to the set of allowed fields.
	//
	// It would probably be more efficient to track a separate allow and deny
	// list, rather than tracking the entire tree except for the excluded.
	exclude defID

	// ignore defines whether we should not do typo checking for this defID.
	ignore bool

	// isOuterStruct indicates that this struct is marked as "outerID" in a
	// CloseInfo. The debug visualization in ./debug.go uses this to show a
	// mark ("S") next to the defID to visualize that this fact.
	isOuterStruct bool
}

type conjunctFlags uint8

const (
	cHasEllipsis conjunctFlags = 1 << iota
	cHasTop
	cHasStruct
	cHasOpenValidator
)

type conjunctInfo struct {
	id    defID
	kind  Kind
	flags conjunctFlags
}

func (c conjunctFlags) hasTop() bool {
	return c&(cHasTop) != 0
}

func (c conjunctFlags) hasStruct() bool {
	return c&(cHasStruct) != 0
}

func (c conjunctFlags) forceOpen() bool {
	return c&(cHasEllipsis|cHasOpenValidator) != 0
}

type replaceID struct {
	from defID
	to   defID
	add  bool // If true, add to the set. If false, replace from with to.
}

func (n *nodeContext) addReplacement(x replaceID) {
	if x.from == x.to {
		return
	}
	// TODO: we currently may compute n.reqSets too early in some rare
	// circumstances. We clear the set if it needs to be recomputed.
	n.computedCloseInfo = false
	n.reqSets = n.reqSets[:0]

	n.replaceIDs = append(n.replaceIDs, x)
}

func (n *nodeContext) updateConjunctInfo(k Kind, id CloseInfo, flags conjunctFlags) {
	if n.ctx.OpenDef {
		return
	}

	for i, c := range n.conjunctInfo {
		if c.id == id.defID {
			n.conjunctInfo[i].kind &= k
			n.conjunctInfo[i].flags |= flags
			return
		}
	}
	n.conjunctInfo = append(n.conjunctInfo, conjunctInfo{
		id:    id.defID,
		kind:  k,
		flags: flags,
	})
}

// addResolver adds a resolver to typo checking. Both definitions and
// non-definitions should typically be added: non-definitions may be added
// multiple times to a single node. As we only want to insert each conjunct
// once, we need to ensure that within all contexts a single ID assigned to such
// a resolver is tracked.
func (n *nodeContext) addResolver(v *Vertex, id CloseInfo, forceIgnore bool) CloseInfo {
	if n.ctx.OpenDef {
		return id
	}

	isClosed := id.FromDef || v.ClosedNonRecursive

	if isClosed && !forceIgnore {
		for i, x := range n.reqDefIDs {
			if x.id == id.outerID && id.outerID != 0 {
				n.reqDefIDs[i].ignore = false
				break
			}
		}
	}

	var ignore bool
	switch {
	case forceIgnore:
		// Special mode to always ignore the outer enclosing group.
		// This is the case, for instance, if a resolver resolves to a
		// non-definition.
		ignore = true
		// TODO: Consider resetting FromDef.
		// id.FromDef = false
	case id.enclosingEmbed != 0 || id.outerID == 0:
		// We have a reference within an inner embedding group. If this is
		// a definition, or otherwise typo checked struct, we need to track
		// the embedding for mutual compatibility.
		// is a definition:
		// 		a: {
		// 			// Even though #A and #B do not constraint `a` individually,
		// 			// they need to be checked for mutual consistency within
		// 			// the embedding.
		// 			#A & #B
		// 		}
		ignore = !isClosed
	default:
		// In the default case we can disable typo checking this type if it is
		// an embedding.
		ignore = id.FromEmbed
	}

	srcID := id.defID
	dstID := defID(0)
	for _, x := range n.reqDefIDs {
		if x.v == v {
			dstID = x.id
			break
		}
	}

	if dstID == 0 || id.enclosingEmbed != 0 {
		next := n.ctx.getNextDefID()
		if dstID != 0 {
			// If we need to activate an enclosing embed group, and the added
			// resolver was already before, we need to allocate a new ID and
			// add the original ID to the set of the new one.
			n.addReplacement(replaceID{from: next, to: dstID, add: true})
		}
		dstID = next

		n.reqDefIDs = append(n.reqDefIDs, refInfo{
			v:       v,
			id:      dstID,
			exclude: id.enclosingEmbed,
			ignore:  ignore,
		})
	}
	id.defID = dstID

	// TODO: consider using add: !isClosed
	n.addReplacement(replaceID{from: srcID, to: dstID, add: true})
	if id.enclosingEmbed != 0 && !ignore {
		ph := id.outerID
		n.addReplacement(replaceID{from: dstID, to: ph, add: true})
		_, dstID = n.newGroup(id, false)
		id.enclosingEmbed = dstID
	}

	return id
}

// subField updates a CloseInfo for subfields of a struct.
func (c *OpContext) subField(ci CloseInfo) CloseInfo {
	ci.outerID = 0
	ci.enclosingEmbed = 0
	return ci
}

func (n *nodeContext) newGroup(id CloseInfo, placeholder bool) (CloseInfo, defID) {
	srcID := id.defID
	dstID := n.ctx.getNextDefID()
	// TODO: consider only adding when record || OpenGraph
	n.reqDefIDs = append(n.reqDefIDs, refInfo{
		v:             emptyNode,
		id:            dstID,
		ignore:        true,
		isOuterStruct: placeholder,
	})
	id.defID = dstID
	n.addReplacement(replaceID{from: srcID, to: dstID, add: true})
	return id, dstID
}

// AddOpenConjunct adds w as a conjunct of v and disables typo checking for w,
// even if it is a definition.
func (v *Vertex) AddOpenConjunct(ctx *OpContext, w *Vertex) {
	n := v.getBareState(ctx)
	ci := n.injectEmbedNode(w, CloseInfo{})
	c := MakeConjunct(nil, w, ci)
	v.AddConjunct(c)
}

// injectEmbedNode is used to track typo checking within an embedding.
// Consider, for instance:
//
//	#A: {a: int}
//	#B: {b: int}
//	#C: {
//		#A & #B // fails
//		c: int
//	}
//
// In this case, even though #A and #B are both embedded, they are intended
// to be mutually exclusive. We track this by introducing a separate defID
// for the embedding. Suppose that the embedding #A&#B is assigned defID 2,
// where its parent is defID 1. Then #A is assigned 3 and #B is assigned 4.
//
// We can then say that requirement 3 (node A) holds if all fields contain
// either label 3, or any field within 1 that is not 2.
func (n *nodeContext) injectEmbedNode(x Decl, id CloseInfo) CloseInfo {
	id.FromEmbed = true

	// Filter cases where we do not need to track the definition.
	switch x := x.(type) {
	case *DisjunctionExpr, *Disjunction, *Comprehension:
		return id
	case *BinaryExpr:
		if x.Op != AndOp {
			return id
		}
	}

	id, dstID := n.newGroup(id, false)
	id.enclosingEmbed = dstID

	return id
}

// splitStruct is used to mark the outer struct of a field in which embeddings
// occur. The significance is that a reference to this node expects a node
// to be closed, even if it only has embeddings. Consider for instance:
//
//	// A is closed  and allows the fields of #B plus c.
//	A: { { #B }, c: int }
//
// TODO(flatclose): this is a temporary solution to handle the case where a
// definition is embedded within a struct. It can be removed if we implement
// the #A vs #A... semantics.
func (n *nodeContext) splitStruct(s *StructLit, id CloseInfo) CloseInfo {
	if n.ctx.OpenDef {
		return id
	}

	if _, ok := s.Src.(*ast.File); ok {
		// If this is not a file, the struct indicates the scope/
		// boundary at which closedness should apply. This is not true
		// for files.
		// We should also not spawn if this is a nested Comprehension,
		// where the spawn is already done as it may lead to spurious
		// field not allowed errors. We can detect this with a nil s.Src.
		// TODO(evalv3): use a more principled detection mechanism.
		// TODO: set this as a flag in StructLit so as to not have to
		// do the somewhat dangerous cast here.
		return id
	}

	return n.splitScope(id)
}

func (n *nodeContext) splitScope(id CloseInfo) CloseInfo {
	id, dstID := n.newGroup(id, true)

	if id.outerID == 0 {
		id.outerID = dstID
	}

	return id
}

func (n *nodeContext) checkTypos() {
	ctx := n.ctx
	if ctx.OpenDef {
		return
	}
	noDeref := n.node // noDeref retained for debugging purposes.
	v := noDeref.DerefValue()

	// Stop early, avoiding the work in appendRequired below, if we have no arcs to check.
	if len(v.Arcs) == 0 {
		return
	}

	// Avoid unnecessary errors.
	if b, ok := v.BaseValue.(*Bottom); ok && !b.CloseCheck {
		return
	}

	required := getReqSets(n)
	if len(required) == 0 {
		return
	}

	// TODO(perf): reuse buffers via OpContext
	requiredCopy := make(reqSets, 0, len(required))
	var replacements []replaceID

	var err *Bottom
	// outer:
	for _, a := range v.Arcs {
		f := a.Label
		if a.IsFromDisjunction() {
			continue // Already checked in disjuncts.
		}

		// TODO(mem): child states of uncompleted nodes must have a state.
		na := a.state

		replacements = na.getReplacements(replacements[:0])
		required := append(requiredCopy[:0], required...)
		// do the right thing in appendRequired either way.
		required.replaceIDs(n.ctx, replacements...)

		a = a.DerefDisjunct()
		// TODO(perf): somehow prevent error generation of recursive structures,
		// or at least make it cheap. Right now if this field is a typo, likely
		// all descendents will be regarded as typos.
		if b, ok := a.BaseValue.(*Bottom); ok {
			if !b.CloseCheck {
				continue
			}
		}

		if allowedInClosed(f) {
			continue
		}

		if n.hasEvidenceForAll(required, na.conjunctInfo) {
			continue
		}

		// TODO: do not descend on optional?

		// openDebugGraph(ctx, a, "NOT ALLOWED") // Uncomment for debugging.
		if b := ctx.notAllowedError(a); b != nil && a.ArcType <= ArcRequired {
			err = CombineErrors(nil, err, b)
		}
	}

	if err != nil {
		n.AddChildError(err) // TODO: should not be necessary.
	}
}

// hasEvidenceForAll reports whether there is evidence in a set of
// conjuncts for each of the typo-checked structs represented by
// the reqSets.
func (n *nodeContext) hasEvidenceForAll(a reqSets, conjuncts []conjunctInfo) bool {
	for i := uint32(0); int(i) < len(a); i += a[i].size {
		if a[i].size == 0 {
			panic("unexpected set length")
		}
		if !hasEvidenceForOne(a[i:i+a[i].size], conjuncts) {
			return false
		}
	}
	return true
}

// hasEvidenceForOne reports whether a single typo-checked set has evidence for
// any of its defIDs.
func hasEvidenceForOne(a reqSets, conjuncts []conjunctInfo) bool {
	for _, c := range conjuncts {
		for _, ri := range a {
			if c.id == ri.id {
				return true
			}
		}
	}
	return false
}

// reqSets defines a set of sets of defIDs that can each satisfy a required id.
//
// A reqSet holds a sequence of a defID "representative", or "head" for a
// requirement, followed by all defIDs that satisfy this requirement. For head
// elements, size indicates the number of entries in the set, including the
// head. For non-head elements, size is 0.
type reqSets []reqSet

// A single reqID might be satisfied by multiple defIDs, if the definition
// associated with the reqID embeds other definitions, for instance. In this
// case we keep a list of defIDs that may also be satisfied.
//
// This type is used in [nodeContext.equivalences] as follows:
//   - if an embedding is used by a definition, it is inserted in the list
//     pointed to by its refInfo. Similarly,
//     refInfo is added to the list
type reqSet struct {
	id defID
	// size is the number of elements in the set. This is only set for the head.
	// Entries with equivalence IDs have size set to 0.
	size uint32
	del  defID // TODO(flatclose): can be removed later.
	once bool
}

// assert checks the invariants of a reqSets. It can be used for debugging.
func (a reqSets) assert() {
	for i := 0; i < len(a); {
		e := a[i]
		if e.size == 0 {
			panic("head element with 0 size")
		}
		if i+int(e.size) > len(a) {
			panic("set extends beyond end of slice")
		}
		for j := 1; j < int(e.size); j++ {
			if a[i+j].size != 0 {
				panic("non-head element with non-zero size")
			}
		}

		i += int(e.size)
	}
}

// replaceIDs replaces defIDs mappings in the receiving reqSets in place.
//
// The following rules apply:
//
//   - Mapping a typo-checked definition to a new typo-checked definition replaces
//     the set from the "from" definition in its entirety. It is assumed that the
//     set is already added to the requirements.
//
//   - A mapping from a typo-checked definition to 0 removes the set from the
//     requirements. This typically comes from _ or ....
//
//   - A mapping from an equivalence (non-head) value to 0 removes the
//     equivalence. This does not change the outcome of typo checking, but
//     reduces the size of the equivalence list, which helps performance.
//
//   - A mapping from an equivalence (non-head) value to a new value widens the
//     allowed values for the respective set by adding the new value to the
//     equivalence list.
//
// In words;
// - Definition: if not in embed, create new group
// - If in active definition, replace old definition
// - If in embed, replace embed in respective sets. definition starts new group
// - child definition replaces parent definition
func (a *reqSets) replaceIDs(ctx *OpContext, b ...replaceID) {
	temp := *a
	temp = temp[:0]
	buf := ctx.reqSetsBuf[:0]
outer:
	for i := 0; i < len(*a); {
		e := (*a)[i]
		if e.size != 0 {
			// If the head is dropped, the entire group is deleted.
			for _, x := range b {
				if e.id == x.from && !x.add {
					i += int(e.size)
					continue outer
				}
			}
			if len(buf) > 0 {
				buf[0].size = uint32(len(buf))
				if len(temp)+len(buf) > i {
					*a = slices.Replace(*a, len(temp), i, buf...)
					i = len(temp) + len(buf)
					temp = (*a)[:i]
				} else {
					temp = append(temp, buf...)
				}
				buf = buf[:0]
			}
		}

		buf = transitiveMapping(ctx, buf, e, b)

		i++
	}
	if len(buf) > 0 {
		buf[0].size = uint32(len(buf))
		temp = append(temp, buf...)
	}
	ctx.reqSetsBuf = buf[:0] // to be reused later on
	*a = temp
}

// TODO: this is a polynomial algorithm. At some point, if this turns out to be
// a bottleneck, we should consider filtering unnecessary entries and/or using a
// more efficient algorithm.
func transitiveMapping(ctx *OpContext, buf reqSets, x reqSet, b []replaceID) reqSets {
	// Trim subtree for embedded conjunctions.
	if len(buf) > 0 && buf[0].del == x.id {
		return buf
	}

	// do not add duplicates
	for _, y := range buf {
		if x.id == y.id {
			return buf
		}
	}

	buf = append(buf, x)
	ctx.stats.CloseIDElems++

	for _, y := range b {
		if x.id == y.from {
			if y.to == deleteID { // do we need this?
				buf = buf[:len(buf)-1]
				return buf
			}
			buf = transitiveMapping(ctx, buf, reqSet{
				id:   y.to,
				once: x.once,
			}, b)
		}
	}
	return buf
}

// mergeCloseInfo merges the conjunctInfo of nw that is missing from nv into nv.
//
// This is used to merge conjunctions from a disjunct to the Vertex that
// originated the disjunct.
// TODO: consider whether we can do without. We usually aim to not check
// such nodes, but sometimes we do.
func mergeCloseInfo(nv, nw *nodeContext) {
	v := nv.node
	w := nw.node
	if w == nil {
		return
	}
	// Merge missing closeInfos
outer:
	for _, wci := range nw.conjunctInfo {
		for _, vci := range nv.conjunctInfo {
			if wci.id == vci.id {
				continue outer
			}
		}
		nv.conjunctInfo = append(nv.conjunctInfo, wci)
	}

outer2:
	for _, d := range nw.replaceIDs {
		for _, vd := range nv.replaceIDs {
			if d == vd {
				continue outer2
			}
		}
		nv.replaceIDs = append(nv.replaceIDs, d)
	}

	for _, wa := range w.Arcs {
		for _, va := range v.Arcs {
			if va.Label == wa.Label {
				mergeCloseInfo(va.state, wa.state)
				break
			}
		}
	}
}

func (n *nodeContext) getReplacements(a []replaceID) []replaceID {
	for p := n.node; p != nil && p.state != nil; p = p.Parent {
		a = append(a, p.state.replaceIDs...)
	}
	return a
}

// getReqSets initializes, if necessary, and returns the reqSets for n.
func getReqSets(n *nodeContext) reqSets {
	if n == nil {
		return nil
	}

	if n.computedCloseInfo {
		return n.reqSets
	}
	n.reqSets = n.reqSets[:0]
	n.computedCloseInfo = true

	a := n.reqSets
	v := n.node

	if p := v.Parent; p != nil {
		aReq := getReqSets(p.state)
		if !n.dropParentRequirements {
			a = append(a, aReq...)
		}
	}
	a.filterNonRecursive()

outer:
	for _, y := range n.reqDefIDs {
		if y.ignore {
			continue
		}
		for _, x := range a {
			if x.id == y.id {
				continue outer
			}
		}
		once := false
		if y.v != nil {
			once = !y.v.ClosedRecursive
		}

		a = append(a, reqSet{
			id:   y.id,
			once: once,
			del:  y.exclude,
			size: 1,
		})
	}

	a.replaceIDs(n.ctx, n.replaceIDs...)

	// If 'v' is a hidden field, then all reqSets in 'a' for which there is no
	// corresponding entry in conjunctInfo should be removed from 'a'.
	if allowedInClosed(v.Label) {
		a.filterSets(func(a []reqSet) bool {
			for _, e := range a {
				for _, c := range n.conjunctInfo {
					if c.id == e.id {
						return true // keep the set
					}
				}
			}
			return false // discard the set
		})
	}

	a.filterTop(n.conjunctInfo)

	n.reqSets = a
	return a
}

// If there is a top or ellipsis for all supported conjuncts, we have
// evidence that this node can be dropped.
func (a *reqSets) filterTop(conjuncts []conjunctInfo) {
	a.filterSets(func(a []reqSet) bool {
		var f conjunctFlags
		for _, e := range a {
			for _, c := range conjuncts {
				if e.id != c.id {
					continue
				}
				flags := c.flags
				if c.id < a[0].id {
					flags &^= cHasStruct
				}
				f |= flags
			}
		}
		if (f.hasTop() && !f.hasStruct()) || f.forceOpen() {
			return false
		}
		return true
	})
}

func (a *reqSets) filterNonRecursive() {
	a.filterSets(func(e []reqSet) bool {
		x := e[0]
		if x.once { //  || x.id == 0
			return false // discard the entry
		}
		return true // keep the entry
	})
}

// filter keeps all reqSets e in a for which f(e) and removes the rest.
func (a *reqSets) filterSets(f func(e []reqSet) bool) {
	temp := (*a)[:0]
	for i := 0; i < len(*a); {
		e := (*a)[i]
		set := (*a)[i : i+int(e.size)]

		if f(set) {
			temp = append(temp, set...)
		}

		i += int(e.size)
	}
	*a = temp
}