Hold a calc operand's draw for the expansion

A field the format rendered and a calc read was drawn twice, so the operand
shown could differ from the operand computed — the same disagreement the
dotted-path rule already fences, in its plainest spelling. The README
carried it as a caveat, which is what a rule like this exists to remove.

A calc operand now joins the names an expansion holds, so it is drawn once
and every later read of it — the calc, and any bare token spelling it —
sees that draw. The hold stays per expansion: each repeat iteration and each
nested template draws its own.

expand reads a calc's operands before the call and hands over their values,
so a builtin takes (emitted, operands) rather than the sibling fields, and
the evaluator indexes that slice instead of walking the node tree. That is
what keeps the draws a local. Handing them to a builtin instead lets them
escape through an indirect call, which put two maps on the heap for every
held format, calc or not: BenchmarkBound went 56 B/5 allocs -> 744 B/10.

Measured against main: Bound 359 -> 349 ns at 56 B/5 allocs, unchanged;
Calc 416 -> 577 ns and one more alloc, which is what the correlation costs.

Validation is unchanged: t.bound still carries only the levels a dotted
token reads, so the overlap fences fence exactly what they did.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
M
2026-08-30 23:09:58 +02:00
committed by lilleman-tw
parent 4f87a4dd9e
commit 9ed0d3bf55
5 changed files with 155 additions and 86 deletions
+10 -10
View File
@@ -31,35 +31,35 @@ var builtins = map[string]builtin{
"ulid": {arity: 0, prep: generate(ulid)},
"nanoid": {arity: 1, check: posIntArg, prep: func(a []string) callFn {
n := atoi(a[0])
return func(s *session, _ string, _ map[string]node) string { return nanoid(s, n) }
return func(s *session, _ string, _ []string) string { return nanoid(s, n) }
}},
"hex": {arity: 1, check: posIntArg, prep: func(a []string) callFn {
n := atoi(a[0])
return func(s *session, _ string, _ map[string]node) string { return randHex(s, n) }
return func(s *session, _ string, _ []string) string { return randHex(s, n) }
}},
"base64": {arity: 1, check: posIntArg, prep: func(a []string) callFn {
n := atoi(a[0])
return func(s *session, _ string, _ map[string]node) string {
return func(s *session, _ string, _ []string) string {
return base64.StdEncoding.EncodeToString(randBytes(s, n))
}
}},
"int": {arity: 2, check: intRangeArgs, prep: func(a []string) callFn {
lo, span := atoi(a[0]), atoi(a[1])-atoi(a[0])+1
return func(s *session, _ string, _ map[string]node) string { return strconv.Itoa(lo + s.IntN(span)) }
return func(s *session, _ string, _ []string) string { return strconv.Itoa(lo + s.IntN(span)) }
}},
"float": {arity: 3, check: floatArgs, prep: func(a []string) callFn {
lo, _ := strconv.ParseFloat(a[0], 64)
hi, _ := strconv.ParseFloat(a[1], 64)
dp := atoi(a[2])
return func(s *session, _ string, _ map[string]node) string {
return func(s *session, _ string, _ []string) string {
return strconv.FormatFloat(lo+s.Float64()*(hi-lo), 'f', dp, 64)
}
}},
"iban": {arity: 1, check: ibanArg, prep: func(a []string) callFn {
cc := a[0]
return func(s *session, _ string, _ map[string]node) string { return iban(s, cc) }
return func(s *session, _ string, _ []string) string { return iban(s, cc) }
}},
// calc is the one builtin that reads the sibling fields (to render its operands);
// calc is the one builtin that names operands (expand reads them for it);
// every other ignores them. 1 or 2 args: the expression and an optional decimals.
"calc": {arity: -1, check: checkCalc, prep: calcPrep},
// seq is the one stateful builtin: a per-session counter from 1, advancing on
@@ -70,7 +70,7 @@ var builtins = map[string]builtin{
if len(a) == 1 {
key = a[0]
}
return func(s *session, _ string, _ map[string]node) string {
return func(s *session, _ string, _ []string) string {
return strconv.FormatUint(s.next(key), 10)
}
}},
@@ -81,13 +81,13 @@ var builtins = map[string]builtin{
// lifts that one function into the prep every registry entry supplies.
func derive(f func(emitted string) string) func([]string) callFn {
return func([]string) callFn {
return func(_ *session, emitted string, _ map[string]node) string { return f(emitted) }
return func(_ *session, emitted string, _ []string) string { return f(emitted) }
}
}
func generate(f func(rng) string) func([]string) callFn {
return func([]string) callFn {
return func(s *session, _ string, _ map[string]node) string { return f(s) }
return func(s *session, _ string, _ []string) string { return f(s) }
}
}
+68 -32
View File
@@ -9,42 +9,48 @@ import (
)
// calc is the {calc(expr[, dp])} token: an arithmetic expression over number
// literals and sibling-field names (rendered, then parsed as numbers), with
// + - * /, unary minus and parentheses. It is a registry builtin like any other
// {name(args)} function — the one whose check and prep read the sibling fields (to
// render its operands). The value prints in minimal decimal form, or rounded to dp
// when given. A field that doesn't render to a number becomes NaN, which
// propagates and prints as "NaN" — visible, never a render error. The expression is
// parsed once at New into an AST every render shares and none mutates, so render
// cannot fail and must not carry per-render state.
// literals and sibling-field names, with + - * /, unary minus and parentheses. It is
// a registry builtin like any other {name(args)} function — the one that names
// operands, which expand reads for it (once per expansion, so the value computed is
// the value the format showed) and hands over as strings. The value prints in
// minimal decimal form, or rounded to dp when given. An operand that isn't a number
// becomes NaN, which propagates and prints as "NaN" — visible, never a render error.
// The expression is parsed once at New into an AST every render shares and none
// mutates, so render cannot fail and must not carry per-render state.
// calcNode is a parsed expression node.
// calcNode is a parsed expression node. It evaluates over the operand values expand
// read, so the evaluator touches neither the rng nor the node tree.
type calcNode interface {
eval(s *session, fields map[string]node) float64
eval(operands []string) float64
}
type calcNum float64 // a number literal
type calcVar string // a sibling-field name
type calcVar string // a sibling-field name, before indexVars places it
type calcIdx int // an operand, by its position in the values expand read
type calcNeg struct{ x calcNode }
type calcBin struct { // a + - * / b
op byte
l, r calcNode
}
func (n calcNum) eval(*session, map[string]node) float64 { return float64(n) }
func (n calcNum) eval([]string) float64 { return float64(n) }
func (n calcVar) eval(s *session, fields map[string]node) float64 {
v, err := strconv.ParseFloat(strings.TrimSpace(render(s, fields[string(n)])), 64)
func (n calcIdx) eval(operands []string) float64 {
v, err := strconv.ParseFloat(strings.TrimSpace(operands[n]), 64)
if err != nil {
return math.NaN() // a non-numeric operand stays visible, never an error
}
return v
}
func (n calcNeg) eval(s *session, fields map[string]node) float64 { return -n.x.eval(s, fields) }
// eval on an unplaced name cannot happen: calcPrep runs indexVars over every
// expression it compiles, so only a calcIdx reaches a render.
func (n calcVar) eval([]string) float64 { return math.NaN() }
func (n calcBin) eval(s *session, fields map[string]node) float64 {
l, r := n.l.eval(s, fields), n.r.eval(s, fields)
func (n calcNeg) eval(operands []string) float64 { return -n.x.eval(operands) }
func (n calcBin) eval(operands []string) float64 {
l, r := n.l.eval(operands), n.r.eval(operands)
switch n.op {
case '+':
return l + r
@@ -82,19 +88,39 @@ func checkCalc(fields map[string]node, args []string) error {
return nil
}
// calcPrep parses the expression and decimals once, at compile time. checkCalc
// proved both valid, so neither step here can fail; dp -1 prints the minimal form.
// calcPrep parses the expression and decimals once, at compile time, and places each
// operand name at the position expand will read it into. checkCalc proved both args
// valid, so no step here can fail; dp -1 prints the minimal form.
func calcPrep(args []string) callFn {
expr, _ := parseCalc(args[0])
at := make(map[string]int)
for i, name := range calcVars(expr) {
at[name] = i
}
placed := indexVars(expr, at)
dp := -1
if len(args) == 2 {
dp = atoi(args[1])
}
return func(s *session, _ string, fields map[string]node) string {
return strconv.FormatFloat(expr.eval(s, fields), 'f', dp, 64)
return func(_ *session, _ string, operands []string) string {
return strconv.FormatFloat(placed.eval(operands), 'f', dp, 64)
}
}
// indexVars replaces each operand name with its position in the values expand reads.
// Both sides take that order from calcVars, so they cannot drift.
func indexVars(n calcNode, at map[string]int) calcNode {
switch n := n.(type) {
case calcVar:
return calcIdx(at[string(n)])
case calcNeg:
return calcNeg{indexVars(n.x, at)}
case calcBin:
return calcBin{n.op, indexVars(n.l, at), indexVars(n.r, at)}
}
return n
}
// calcOperands lists the sibling-field names every {calc(...)} token in a format
// reads, so cycle detection sees the field edges calc renders through (a function
// token otherwise carries no field edge).
@@ -124,19 +150,29 @@ func calcTokenOperands(body string) []string {
return calcVars(expr)
}
// calcVars lists the field names an expression references, for the existence
// check in checkCalc.
// calcVars lists the distinct field names an expression reads, in the order it first
// names each. That order is the contract between expand, which reads the operands
// into a slice, and indexVars, which places each name at its position in it.
func calcVars(n calcNode) []string {
switch n := n.(type) {
case calcVar:
return []string{string(n)}
case calcNeg:
return calcVars(n.x)
case calcBin:
return append(calcVars(n.l), calcVars(n.r)...)
default:
return nil
var out []string
seen := map[string]bool{}
var walk func(calcNode)
walk = func(n calcNode) {
switch n := n.(type) {
case calcVar:
if name := string(n); !seen[name] {
seen[name] = true
out = append(out, name)
}
case calcNeg:
walk(n.x)
case calcBin:
walk(n.l)
walk(n.r)
}
}
walk(n)
return out
}
// calcParser is a recursive-descent parser over the expression runes, threading
+6 -3
View File
@@ -47,9 +47,12 @@ type template struct {
ops []op // format compiled once (see compileOps); what expand walks
grow int // minimum output size, to size the render buffer
// bound maps each field the format addresses by dotted path to one path token
// reading it: drawn once per expansion (see compileOps and expand), and the
// token names the other half of an overlap. nil when the format takes no path.
// reading it, which is the half of an overlap the fences name. nil when the
// format takes no path.
bound map[string]string
// held is every name drawn once per expansion: the bound levels above, plus the
// siblings a {calc()} reads. nil when the format holds nothing (see expand).
held map[string]bool
}
func (*template) isNode() {}
@@ -160,7 +163,7 @@ func compileTemplate(m map[string]any) (node, error) {
if err := checkTokens(format, t.fields); err != nil {
return nil, err
}
t.ops, t.grow, t.bound = compileOps(format)
t.ops, t.grow, t.bound, t.held = compileOps(format)
if err := checkNoOverlap(format, t.bound); err != nil {
return nil, err
}
+39 -29
View File
@@ -137,13 +137,14 @@ func classChar(s *session, c rune) byte {
func expand(s *session, t *template) string {
var b strings.Builder
b.Grow(t.grow)
// One draw per bound head, held for this expansion only: a nested template and
// each repeat iteration get their own, since each is its own expansion.
var bound *draws
if len(t.bound) > 0 {
bound = &draws{
variant: make(map[string]node, len(t.bound)),
value: make(map[string]string, len(t.bound)),
// One draw per held name, for this expansion only: a nested template and each
// repeat iteration get their own, since each is its own expansion. held stays a
// local: nothing hands it to a builtin, so it and its maps never leave the stack.
var held *draws
if len(t.held) > 0 {
held = &draws{
variant: make(map[string]node, len(t.held)),
value: make(map[string]string, len(t.held)),
}
}
for i := range t.ops {
@@ -154,57 +155,66 @@ func expand(s *session, t *template) string {
case 'c':
b.WriteByte(classChar(s, o.r))
case 'f':
b.WriteString(resolve(s, o.arms, t, bound))
b.WriteString(readField(s, t, held, o.arms[s.IntN(len(o.arms))]))
case 'b':
b.WriteString(o.call(s, b.String(), t.fields)) // b.String() is the output so far
// A calc's operands are read here, in the order its expression first
// names them, so the value it computes is the value the format showed.
var operands []string
if len(o.operands) > 0 {
operands = make([]string, len(o.operands))
for j, name := range o.operands {
operands[j] = readField(s, t, held, arm{name: name, key: name})
}
}
b.WriteString(o.call(s, b.String(), operands)) // b.String() is the output so far
}
}
return b.String()
}
// draws is what an expansion has already drawn for its bound heads: the variant
// each head was drawn as, so every path under it reads one row, and the value each
// path read, so the same path read twice reads one value.
// draws is what an expansion has already drawn for its held names: the variant each
// was drawn as, so every path under it reads one row, and the value each read, so
// the same name read twice reads one value.
type draws struct {
variant map[string]node
value map[string]string
}
// resolve renders one field alternation: the '|' alternatives, one picked at
// random. An arm's key is a sibling field or a {..path} reference, which linkRefs
// bound into fields too. A field the format addresses by dotted path is bound: it
// is drawn once for the expansion, so {place.postal-code} and {place.locality}
// read one row and either read twice gives one value. checkTokens, checkPath and
// linkRefs prove every step, so this cannot fail.
func resolve(s *session, arms []arm, t *template, bound *draws) string {
a := arms[s.IntN(len(arms))]
if _, isBound := t.bound[a.key]; !isBound {
// readField renders one arm of a token. An arm's key is a sibling field or a
// {..path} reference, which linkRefs bound into fields too. A name the expansion
// holds — a level some token addresses by dotted path, or a sibling a {calc()}
// reads — is drawn once and kept, so {place.postal-code} and {place.locality} read
// one row, either read twice gives one value, and a shown operand is the operand
// computed. Every other name is drawn afresh, so {word} {word} still draws twice.
// checkTokens, checkPath and linkRefs prove every step, so this cannot fail.
func readField(s *session, t *template, held *draws, a arm) string {
if !t.held[a.key] {
return render(s, t.fields[a.key])
}
if v, read := bound.value[a.name]; read {
if v, read := held.value[a.name]; read {
return v
}
n, drew := bound.variant[a.key]
n, drew := held.variant[a.key]
if !drew {
n = drawn(s, t.fields[a.key])
bound.variant[a.key] = n
held.variant[a.key] = n
}
// Hold the draw at every level passed through, so two paths sharing a prefix
// share it.
for i, seg := range a.tail {
if i < len(a.steps) {
held, drew := bound.variant[a.steps[i]]
step, drew := held.variant[a.steps[i]]
if !drew {
held = drawn(s, child(n, seg))
bound.variant[a.steps[i]] = held
step = drawn(s, child(n, seg))
held.variant[a.steps[i]] = step
}
n = held
n = step
continue
}
n = child(n, seg)
}
v := render(s, n)
bound.value[a.name] = v
held.value[a.name] = v
return v
}
+32 -12
View File
@@ -60,8 +60,8 @@ func eachToken(format string, fn func(ftoken) error) error {
// builtin is a format-string function invoked as {name(args)}. It receives the
// session (its rng, and the {seq()} counters), the output emitted so far in the
// current expansion (for derivations such as a checksum over preceding digits),
// the sibling fields (only calc reads them, to render its operands), and its args.
// current expansion (for derivations such as a checksum over preceding digits), and
// the values of the operands it named (only calc names any).
// Almost all are pure over (rng, emitted, args) — no wall-clock, no crypto/rand —
// so seeding stays reproducible; a time-based id derives its time from the rng.
// seq is the one exception: it advances per-session counter state, which is itself
@@ -183,7 +183,7 @@ func checkNoRepeatedArm(body string, names []string) error {
// fieldTokens returns the field and reference names a format renders via {name}
// or {a|..b} tokens (function tokens, which carry no field edges, are excluded).
// These are exactly the child nodes expand's resolve recurses into.
// These are exactly the child nodes expand recurses into, through readField.
func fieldTokens(format string) []string {
var names []string
_ = eachToken(format, func(t ftoken) error {
@@ -303,8 +303,10 @@ func splitArms(body string) []arm {
return arms
}
// callFn is a builtin bound to one call site: its args already parsed.
type callFn func(s *session, emitted string, fields map[string]node) string
// callFn is a builtin bound to one call site: its args already parsed. It reads the
// output emitted so far in the current expansion (a derivation's payload) and the
// values of the operands it named, which expand read for it.
type callFn func(s *session, emitted string, operands []string) string
// op is one compiled unit of a format string: a literal run, a class char, a field
// alternation, or a builtin already bound to its args. compile builds these so
@@ -315,19 +317,32 @@ type op struct {
r rune // kind 'c'
arms []arm // kind 'f': the '|' alternatives, split into key and path once
call callFn
// operands names the sibling fields a {calc()} reads, in the order its
// expression first names them. expand reads them before the call, so the
// evaluator never touches the node tree. nil for every other builtin.
operands []string
}
// compileOps turns a format string into ops, and returns the smallest output it can
// produce (literals plus one byte per class char) to size the render buffer, plus
// the fields the format addresses by dotted path each drawn once per expansion,
// so every token reading one sees the same row (see expand). bound is nil when the
// format takes no path, so data that uses none carries no render-time cost.
// Call checkTokens first: it is what proves the scan and every token are valid.
func compileOps(format string) ([]op, int, map[string]string) {
// two sets. bound is the levels the format addresses by dotted path, each mapped to
// the first path reading it, which is what the overlap fences name. held is every
// name drawn once per expansion — those levels, plus the siblings a {calc()} reads,
// so an operand shown is the operand computed. Both are nil when the format needs
// neither, so data that uses neither carries no render-time cost. Call checkTokens
// first: it is what proves the scan and every token are valid.
func compileOps(format string) ([]op, int, map[string]string, map[string]bool) {
var ops []op
var lit strings.Builder
var bound map[string]string
var held map[string]bool
grow := 0
hold := func(name string) {
if held == nil {
held = map[string]bool{}
}
held[name] = true
}
flush := func() {
if lit.Len() > 0 {
grow += lit.Len()
@@ -346,7 +361,11 @@ func compileOps(format string) ([]op, int, map[string]string) {
case 'b':
flush()
if name, args, ok := funcCall(t.body); ok {
ops = append(ops, op{kind: 'b', call: builtins[name].prep(args)})
operands := calcTokenOperands(t.body)
for _, operand := range operands {
hold(operand) // a calc renders its operand, so the expansion holds that draw
}
ops = append(ops, op{kind: 'b', call: builtins[name].prep(args), operands: operands})
} else {
arms := splitArms(t.body)
for _, a := range arms {
@@ -357,6 +376,7 @@ func compileOps(format string) ([]op, int, map[string]string) {
if _, named := bound[a.key]; !named {
bound[a.key] = a.name // the first path reading it, for error messages
}
hold(a.key)
}
}
ops = append(ops, op{kind: 'f', arms: arms})
@@ -365,5 +385,5 @@ func compileOps(format string) ([]op, int, map[string]string) {
return nil
})
flush()
return ops, grow, bound
return ops, grow, bound, held
}