* plans/lintfixes.md
* plans/lintfixes.md
* Fix remaining govet lint findings
- Rename MarshalJSON -> FormatAsJSON on Mlrval and Mlrmap (govet
stdmethods): the methods shadowed json.Marshaler with an
incompatible signature.
- Remove unreachable return after exhaustive if-else in
pkg/mlrval/mlrval_collections.go (govet unreachable).
- Update plans/lintfixes.md with current status: 84 findings remain
(50 errcheck, 34 staticcheck).
Part of #2109.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Return nil on successful single-index array unset
removeIndexedOnArray removed the element on the in-bounds path but
then fell through to return an "array index out of bounds for unset"
error, so the success path never returned nil. Callers currently
ignore the error, which masked this; return nil on success so that
upcoming errcheck fixes can propagate the error meaningfully. This
matches removeIndexedOnMap, which returns nil on success.
Add unit tests for RemoveIndexed on arrays.
Part of #2109.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Found with golang.org/x/tools/cmd/deadcode (rooted at cmd/mlr + tests)
and staticcheck U1000; each finding verified by hand before deletion.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
* Update snapcore/action-build to node24 SHA, drop unused matrix
snapcore/action-build@v1 targets Node.js 20, which is deprecated on
GitHub Actions runners (forced to run on Node.js 24 with a warning).
Upstream has an open PR to update (https://github.com/snapcore/action-build/pull/1)
but it has been stale for a while, so we pin directly to its HEAD SHA
(edf78ca) until upstream merges and cuts a new tag.
Also removes the `strategy.matrix.node-version: [20.x]` block, which
was dead config — there was no `actions/setup-node` step consuming it.
Fixes the warning seen in:
https://github.com/johnkerl/miller/actions/runs/28449033140
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
* comment
---------
Co-authored-by: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
Replaces 100+ if/else-if chains on a single variable with tagged switch
statements across 72 files. The bulk are transformer option-parsing loops
(switch on opt string), plus a handful of value-dispatch sites in mlrval,
dsl/cst, repl, lib, auxents, and bifs. One case (surv.go) required a
labeled break to preserve the loop-exit behavior of the original else branch.
Fixes staticcheck QF1003 findings.
Co-authored-by: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
* Add JSON index and mlr which capability router (#2098)
PR 2 of the AI-friendly roadmap (plans/plan-2098-llm.md):
- `mlr help --as-json --index` emits a lightweight [{kind,name,summary}]
index across all 651 catalog items (verbs, functions, flags, keywords),
sorted by kind then name. Agents use this as a cheap first call to pick
a verb before fetching its full entry.
- `mlr which "<query>"` is a new terminal that tokenizes a natural-language
query, scores every catalog item (name match +20/token, body match +5/token),
and returns ranked JSON [{kind,name,score,summary}]. Exit code 0 means a
confident match (at least one token hit the item name); exit code 2 means
low confidence. Agents branch on the exit code rather than parsing prose.
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
* Move mlr which into pkg/terminals/help, deduplicate firstLine
pkg/terminals/which/ was a misplaced package: it imported the same four
catalog registries as pkg/terminals/help/ and duplicated the firstLine
helper. Moving the logic into help/entry_which.go fixes both issues:
- WhichMain and all which helpers now live alongside the other --as-json
catalog machinery in pkg/terminals/help/
- indexFirstLine (entry_json.go) and firstLine (which/entry.go) collapse
into a single firstLine shared by both files
- pkg/terminals/terminals.go calls help.WhichMain directly; the
pkg/terminals/which/ package is deleted
No behavior change.
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
* Fix five Go-purity issues found in code review
1. Exit-code false positive: whichScore now returns (int, bool) where the
bool records whether any token hit the item name. WhichMain uses
results[0].nameHit for exit-code 0, so 4 body-only token hits (4×5=20)
no longer incorrectly signal a confident match.
2. Flag Summary inconsistency: whichSearch was setting Summary: fl.Help
directly for flags while using firstLine(...) for functions and keywords.
Changed to firstLine(fl.Help) so all four kinds behave consistently.
3+5. kindOrder/whichKindRank duplication: the verb<function<flag<keyword
ordering was encoded twice — as a local map[string]int in buildIndex and
as a switch in whichKindRank. Replaced both with a single package-level
kindRank() function. The map lookup also silently returned 0 (= verb rank)
for unknown kinds; the switch correctly returns 4 (sorts last).
4. extractIndexFlag/extractAsJSONFlag duplication: both had identical loop
bodies differing only in the sentinel string. Introduced a generic
extractFlag(args, flag) helper; both are now one-liners.
Also promoted whichStopwords to a package-level var so whichTokenize does
not allocate a new map on every call.
Co-Authored-By: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
Emit Miller's existing help catalog (verbs, functions, flags, keywords)
as structured JSON so AI agents and tooling can model Miller's surface
without scraping prose. The --json token may appear anywhere on a
`mlr help ...` command line; plain text help is unchanged.
mlr help --json # full catalog
mlr help verb cat --json # one or more verbs
mlr help function splitax --json # one or more functions
mlr help flag --ifs --json # one or more flags
mlr help keyword ENV --json # one or more keywords
Functions and flags serialize fully (name/class/arity/help/examples;
section/name/alt_names/arg/help). Verbs carry a summary, ignores_input,
and captured raw usage_text as a Tier-1 fallback, since per-verb options
are prose-only today (each verb hand-writes its UsageFunc). Structured
verb options are a planned follow-on (see #2098).
This is a serialization layer over the existing registries -- no
refactor of the text-help path.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* Add roadmap doc for making Miller AI-friendly (#2098)
Living roadmap derived from issue #2098 and @aborruso's comment:
PR-by-PR arc from a machine-readable help catalog (mlr help --as-json)
through structured verb options, structured errors, DSL validate,
mlr describe, and an MCP server + agent skill.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Clarify PR3 enums are static codelists, not data-dependent constraints
Distinguish @aborruso's codelist (binary-fixed values, PR3) from
constraint (input-dependent values, PR6 mlr describe).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* ci: add golangci-lint workflow
- Adds a GitHub Actions workflow that runs golangci-lint on push and PR
against the main branch, complementing the existing build/test matrix
in go.yml.
- Pins Go 1.25 to match go.mod and golangci-lint v1.61.0 for reproducible
linting; the job is run on ubuntu-latest with a 5-minute timeout.
- Uses concurrency cancellation per ref to keep the CI queue short and
read-only contents permissions per least-privilege guidance.
* ci: use valid golangci-lint action release
Signed-off-by: dashitongzhi <civilization.cn@outlook.com>
* ci: scope golangci-lint to production packages
* ci: make golangci-lint job advisory (continue-on-error)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---------
Signed-off-by: dashitongzhi <civilization.cn@outlook.com>
Co-authored-by: John Kerl <kerl.john.r@gmail.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* initial attempt
* fix bash
* fix zsh
* Add shell-completion docs page
Documents the new 'mlr completion {bash,zsh}' feature: the then-chain
context model, install instructions for bash and zsh (including the macOS
bash-3.2 'eval' caveat and zsh compinit self-init), and examples of
context-aware completion. Added to the nav under "Miller in more detail".
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Add enum value completion for format and separator flags
Completes the argument value for arg-taking main flags whose values are a
known set: file-format names for -i/-o/--io, separator aliases for
--ifs/--ofs/--ips/etc., and regex-separator aliases for --ifs-regex/--ips-regex.
Other arg-taking flags continue to fall back to filename completion.
Candidate sets come from new cli getters (GetFileFormatNames,
GetSeparatorAliasNames, GetSeparatorRegexAliasNames) that read the same maps
Miller uses at runtime, so there is no separate list to keep in sync. The
command-line walk now records which flag a value position belongs to.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Include format-conversion keystroke-savers in bare-dash completion
Reverts the suppression of --c2j/--x2y-style flags from 'mlr -<TAB>'. The full
set of main flags (297) is now offered, matching what is valid on the command
line. GetFlagNames no longer takes an includeSuppressed argument.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Complete terminal subcommands and top-level help/version flags
'mlr <TAB>' now offers subcommand names (help, version, repl, regtest, script,
completion, terminal-list) alongside verb names, and 'mlr -<TAB>' offers the
top-level terminal flags (-h, --help, --version, --bare-version, and the help
shorthands -g/-l/-L/-f/-F/-k/-K). Subcommand names are offered only as the
first non-flag token, where they are valid.
To let the completion engine know these names without an import cycle
(pkg/terminals imports pkg/terminals/completion), the canonical terminal names
and version-flag spellings are factored into a new leaf package
pkg/terminals/registry, imported by pkg/terminals, pkg/climain, and completion.
The help-flag spellings come from a new help.GetTerminalFlagNames derived from
the existing shorthand table, so nothing drifts.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Complete 'mlr help' topics and topic arguments
'mlr help <TAB>' now completes help topics (flags, verb, function, keyword,
list-verbs, ...), and topics that take a name argument complete it too:
'mlr help verb <TAB>' -> verb names, 'mlr help function <TAB>' -> function
names, 'mlr help keyword <TAB>' -> keyword names, 'mlr help flag <TAB>' ->
flag names. 'mlr completion <TAB>' completes bash/zsh.
A terminal subcommand consumes the rest of the command line, so the walk now
returns a ctxTerminalArgs context carrying the terminal name and the words
typed after it. New getters supply the candidate names without drift:
help.GetTopicNames, help.GetFunctionNames/GetKeywordNames (wrapping new
cst.BuiltinFunctionManager.GetBuiltinFunctionNames and cst.GetKeywordNames).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* docs-neaten
* Move flag-value-candidate logic into pkg/cli; fix verb-flag collision
The mapping of which flags take a format/separator/regex-separator argument is
flag metadata, so it now lives with the flags in pkg/cli as
cli.FlagValueCandidates, alongside the existing GetFileFormatNames /
GetSeparatorAliasNames getters, replacing the maps that were in
pkg/terminals/completion/value_completion.go (now removed).
This also fixes a bug: value completion now applies only to main flags, not to
identically-spelled verb flags. Previously 'mlr uniq -o <TAB>' offered file
formats because uniq's -o (an output field name) collided with the main -o
format flag; it now correctly falls back to filename completion.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Inspired by GNU head & tail, they match their behavior while supporting
the usual grouping operations.
Co-authored-by: John Kerl <kerl.john.r@gmail.com>
seqgen.Transform was generating its full sequence on every call,
including once per upstream record. When chaining two seqgens of
N records each, the second produced N+1 copies of N records in
memory (O(N²)) before writing only the first N. Fix: return
immediately for non-EOS input; generate the sequence only when
the upstream end-of-stream arrives, which is the correct
semantics for a verb that discards its input record stream.
Co-authored-by: Claude Sonnet 4.6 (1M context) <noreply@anthropic.com>
* Batch-allocate per-record objects; reuse CSV writer field buffer
After batch-arena field allocation, profiling cat over 1M-record CSV showed
the remaining ~5M allocations were almost entirely per-record (one each):
the Mlrmap struct, the RecordAndContext wrapper, the CSV writer's []string,
and the go-csv parser's own buffers.
Address the first three:
- mlrval.RecordArena gains NewRecord(), vending the Mlrmap struct itself from
a per-batch slab (respecting --no-hash-records). Rolled out to every
line-based reader (CSV, CSV-lite, TSV, DKVP, NIDX, PPRINT, XTAB, DKVPX) in
place of NewMlrmapAsRecord.
- The CSV reader batch-allocates RecordAndContext wrappers from a per-batch
slab instead of one heap object per record (comment/output-string entries
still allocate individually, but they are rare).
- RecordWriterCSV reuses a single fieldsBuffer []string across records instead
of allocating one per Write; WriteCSVRecordMaybeColorized consumes it
synchronously and the writer is single-goroutine, so this is safe.
Effect (big.*, 1M records, cat, best of 5):
csv 0.26 -> 0.22
dkvp 0.51 -> 0.45 (Mlrmap slab)
For CSV, cat's allocation-object count drops ~5.0M -> ~2.1M. The remaining
~2M are the go-csv parser's per-record backing string and field slice, which
are intrinsic to parsing and would require a zero-copy/batch-slab parser
rework. A CPU profile of cat now shows it is I/O-bound (syscall ~56%, bufio
read+flush), with allocation/GC down to ~10% -- i.e. further allocation
trimming no longer moves cat's wall-clock. GOGC=off confirms (no change).
Verified: go test ./pkg/... and full regression suite pass; output is
byte-identical across all formats including record-retaining verbs (tac),
hashed and --no-hash-records.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Pool DSL stack frames across records (~8-9% on put)
A StackFrameSet lives on the persistent runtime.State and is reused across
all records, but every block entry (StatementBlockNode.Execute does
PushStackFrame/PopStackFrame, which runs once per record for the main block,
plus once per if/for/etc.) allocated a fresh StackFrame -- a []*var slice and
a map[string]int -- and discarded it on exit. For `put`/`filter` that is
millions of throwaway allocations.
Since push/pop is strictly LIFO, retain popped frames in a per-frameset free
list and clear-and-reuse them on the next push. After the first record
establishes the max block-nesting depth, per-record block execution is
allocation-free for frames. len(stackFrames) remains the logical depth, so
get/set/defineTyped/unset/etc. are unchanged.
Measured (big.csv, 1M rows, best of 4):
put chain-1 0.78 -> 0.72 (~8%)
put chain-4 0.96 -> 0.87 (~9%)
Allocation objects for put chain-1 drop ~23.1M -> ~20.0M (the per-record
newStackFrame churn, ~2.86M, is eliminated). UDF calls still allocate a fresh
frameset per call (PushStackFrameSet); pooling those is a separate change.
The dominant remaining DSL allocator is FromFloat (~6.8M, interior arithmetic
temporaries); eliminating it needs node-owned result slots + in-place bif
variants, a much larger and aliasing-sensitive change, left for follow-up.
Verified: go test ./pkg/... and full regression suite pass; put output is
byte-identical, including UDFs with locals/loops/blocks.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Pool DSL stack-frame *sets* across UDF/subr calls (~31% on function-heavy put)
Companion to the per-block frame pooling: that left PushStackFrameSet /
PopStackFrameSet (entered once per user-defined function or subroutine call)
allocating. Each call did newStackFrameSet() -- a StackFrameSet plus its
initial StackFrame (a slice and a map) -- AND, worse, prepended it with
append([]*StackFrameSet{head}, sets...), allocating a fresh backing slice and
copying the whole save-stack every call.
Two changes:
- Treat the frameset save-stack as a tail stack (append to push, truncate to
pop) instead of prepending at index 0. get/set only ever touch the cached
head, so list order is irrelevant; this removes the per-call slice
realloc + O(depth) copy.
- Pool popped framesets (LIFO) and reset-and-reuse them on the next push,
mirroring the per-frameset frame free list. A reset trims back to one
cleared base frame (extras go to the frame pool). After warmup, repeated
calls allocate no framesets or frames.
Measured (big.csv, 1M rows, best of 5):
put, 2 nested func calls/record: 2.73 -> 1.87 (~31%)
GC cycles 25 -> 16; newStackFrameSet/newStackFrame fall out of the allocation
profile entirely. (chain-1 etc. have no UDFs and are unaffected.)
Verified: go test ./pkg/... and full regression suite pass; recursion
(fact/fib), local-scope isolation, and subroutine+oosvar all correct.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Drop redundant deep-copy of UDF return values (~3-16% on UDF put)
A user-defined function's return value was deep-copied twice on the way out:
once in ReturnNode.Execute (returnValue.Copy() when building the block-exit
payload) and again in UDFCallsite.EvaluateWithArguments
(blockReturnValue.Copy() at the end).
The ReturnNode copy is the necessary one: it detaches the value from the
callee's frame so it survives the frame being popped (and, since perf-try-7,
pooled/reused). By the time EvaluateWithArguments returns, blockReturnValue is
therefore already an independent deep copy, so the second copy is pure waste --
and callers that retain the result copy again anyway (field/oosvar/local
assignment all PutCopy/Copy). The other return paths (implicit-absent, error)
don't use blockReturnValue, so this only affects the BLOCK_EXIT_RETURN_VALUE
path.
Return blockReturnValue directly.
Measured (big.csv, 1M rows, best of 5):
put, 2 nested scalar-returning calls/record: 1.89 -> 1.83 (~3%)
put, map-returning func per record: 2.34 -> 1.97 (~16%)
Win scales with return-value size (the avoided copy is deep). All UDF/HOF
callsites (apply/reduce/sort/select/fold/...) go through this path.
Verified: go test ./pkg/... and full regression suite pass; recursion, HOFs,
and returned-map isolation (mutating a returned map does not affect a
subsequent call) all correct.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Bind scalar locals/params by reference, not by copy (~4-9% on DSL)
NewTypeGatedMlrvalVariable and TypeGatedMlrvalVariable.Assign deep-copied every
value bound to a local variable or function parameter -- ~6.9M allocations on a
UDF-heavy put. For scalars that copy is unnecessary:
Aliasing audit. Assignment everywhere REPLACES pointers rather than mutating
Mlrvals in place: Mlrmap.PutCopy reassigns pe.Value, Assign reassigns
tvar.value. The only in-place mutation a scalar undergoes is idempotent
type-inference caching (printrep -> typed). So a local/param bound by reference
to a scalar source can never observe its source change, and reassigning the
local replaces its own pointer without touching the source -- capture-by-value
semantics are preserved. Maps and arrays, by contrast, ARE mutated in place by
indexed assignment (m[k]=v), so an aliased collection would corrupt its source;
those must still be deep-copied.
So copyForBind copies only arrays/maps and binds scalars by reference. (Return
values are independently safe: ReturnNode.Execute still deep-copies them.)
Measured (big.csv, 1M rows, best of 5):
UDF-heavy put (scalar args/locals): 1.84 -> 1.68 (~9%)
x = $a+$b; $s = x*2 (no UDF): 0.50 -> 0.48 (~4%)
Verified: go test ./pkg/... and full regression suite pass, plus targeted
alias-then-mutate tests: scalar locals capture-by-value (source change after
bind not observed; reassigning one of two aliases leaves the other intact;
mutating a scalar param leaves the caller field intact), and collections stay
independent (local/param/oosvar-element map copies isolate in-place mutation).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* docs: add "Allocation/GC optimizations: June 2026" perf section
Documents the mid-2026 allocation-reduction work on the performance page:
before/after wall-clock and peak-RSS tables (best-of-five, M1, the same
million-record files used elsewhere on the page), plus notes on why streaming
I/O verbs and DSL/UDF workloads benefit most and why sort's time improves
while its memory does not. Added to both performance.md.in and the generated
performance.md (the page has no live-code GENMD blocks).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* Batch-allocate per-record objects; reuse CSV writer field buffer
After batch-arena field allocation, profiling cat over 1M-record CSV showed
the remaining ~5M allocations were almost entirely per-record (one each):
the Mlrmap struct, the RecordAndContext wrapper, the CSV writer's []string,
and the go-csv parser's own buffers.
Address the first three:
- mlrval.RecordArena gains NewRecord(), vending the Mlrmap struct itself from
a per-batch slab (respecting --no-hash-records). Rolled out to every
line-based reader (CSV, CSV-lite, TSV, DKVP, NIDX, PPRINT, XTAB, DKVPX) in
place of NewMlrmapAsRecord.
- The CSV reader batch-allocates RecordAndContext wrappers from a per-batch
slab instead of one heap object per record (comment/output-string entries
still allocate individually, but they are rare).
- RecordWriterCSV reuses a single fieldsBuffer []string across records instead
of allocating one per Write; WriteCSVRecordMaybeColorized consumes it
synchronously and the writer is single-goroutine, so this is safe.
Effect (big.*, 1M records, cat, best of 5):
csv 0.26 -> 0.22
dkvp 0.51 -> 0.45 (Mlrmap slab)
For CSV, cat's allocation-object count drops ~5.0M -> ~2.1M. The remaining
~2M are the go-csv parser's per-record backing string and field slice, which
are intrinsic to parsing and would require a zero-copy/batch-slab parser
rework. A CPU profile of cat now shows it is I/O-bound (syscall ~56%, bufio
read+flush), with allocation/GC down to ~10% -- i.e. further allocation
trimming no longer moves cat's wall-clock. GOGC=off confirms (no change).
Verified: go test ./pkg/... and full regression suite pass; output is
byte-identical across all formats including record-retaining verbs (tac),
hashed and --no-hash-records.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Pool DSL stack frames across records (~8-9% on put)
A StackFrameSet lives on the persistent runtime.State and is reused across
all records, but every block entry (StatementBlockNode.Execute does
PushStackFrame/PopStackFrame, which runs once per record for the main block,
plus once per if/for/etc.) allocated a fresh StackFrame -- a []*var slice and
a map[string]int -- and discarded it on exit. For `put`/`filter` that is
millions of throwaway allocations.
Since push/pop is strictly LIFO, retain popped frames in a per-frameset free
list and clear-and-reuse them on the next push. After the first record
establishes the max block-nesting depth, per-record block execution is
allocation-free for frames. len(stackFrames) remains the logical depth, so
get/set/defineTyped/unset/etc. are unchanged.
Measured (big.csv, 1M rows, best of 4):
put chain-1 0.78 -> 0.72 (~8%)
put chain-4 0.96 -> 0.87 (~9%)
Allocation objects for put chain-1 drop ~23.1M -> ~20.0M (the per-record
newStackFrame churn, ~2.86M, is eliminated). UDF calls still allocate a fresh
frameset per call (PushStackFrameSet); pooling those is a separate change.
The dominant remaining DSL allocator is FromFloat (~6.8M, interior arithmetic
temporaries); eliminating it needs node-owned result slots + in-place bif
variants, a much larger and aliasing-sensitive change, left for follow-up.
Verified: go test ./pkg/... and full regression suite pass; put output is
byte-identical, including UDFs with locals/loops/blocks.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Pool DSL stack-frame *sets* across UDF/subr calls (~31% on function-heavy put)
Companion to the per-block frame pooling: that left PushStackFrameSet /
PopStackFrameSet (entered once per user-defined function or subroutine call)
allocating. Each call did newStackFrameSet() -- a StackFrameSet plus its
initial StackFrame (a slice and a map) -- AND, worse, prepended it with
append([]*StackFrameSet{head}, sets...), allocating a fresh backing slice and
copying the whole save-stack every call.
Two changes:
- Treat the frameset save-stack as a tail stack (append to push, truncate to
pop) instead of prepending at index 0. get/set only ever touch the cached
head, so list order is irrelevant; this removes the per-call slice
realloc + O(depth) copy.
- Pool popped framesets (LIFO) and reset-and-reuse them on the next push,
mirroring the per-frameset frame free list. A reset trims back to one
cleared base frame (extras go to the frame pool). After warmup, repeated
calls allocate no framesets or frames.
Measured (big.csv, 1M rows, best of 5):
put, 2 nested func calls/record: 2.73 -> 1.87 (~31%)
GC cycles 25 -> 16; newStackFrameSet/newStackFrame fall out of the allocation
profile entirely. (chain-1 etc. have no UDFs and are unaffected.)
Verified: go test ./pkg/... and full regression suite pass; recursion
(fact/fib), local-scope isolation, and subroutine+oosvar all correct.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Drop redundant deep-copy of UDF return values (~3-16% on UDF put)
A user-defined function's return value was deep-copied twice on the way out:
once in ReturnNode.Execute (returnValue.Copy() when building the block-exit
payload) and again in UDFCallsite.EvaluateWithArguments
(blockReturnValue.Copy() at the end).
The ReturnNode copy is the necessary one: it detaches the value from the
callee's frame so it survives the frame being popped (and, since perf-try-7,
pooled/reused). By the time EvaluateWithArguments returns, blockReturnValue is
therefore already an independent deep copy, so the second copy is pure waste --
and callers that retain the result copy again anyway (field/oosvar/local
assignment all PutCopy/Copy). The other return paths (implicit-absent, error)
don't use blockReturnValue, so this only affects the BLOCK_EXIT_RETURN_VALUE
path.
Return blockReturnValue directly.
Measured (big.csv, 1M rows, best of 5):
put, 2 nested scalar-returning calls/record: 1.89 -> 1.83 (~3%)
put, map-returning func per record: 2.34 -> 1.97 (~16%)
Win scales with return-value size (the avoided copy is deep). All UDF/HOF
callsites (apply/reduce/sort/select/fold/...) go through this path.
Verified: go test ./pkg/... and full regression suite pass; recursion, HOFs,
and returned-map isolation (mutating a returned map does not affect a
subsequent call) all correct.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Bind scalar locals/params by reference, not by copy (~4-9% on DSL)
NewTypeGatedMlrvalVariable and TypeGatedMlrvalVariable.Assign deep-copied every
value bound to a local variable or function parameter -- ~6.9M allocations on a
UDF-heavy put. For scalars that copy is unnecessary:
Aliasing audit. Assignment everywhere REPLACES pointers rather than mutating
Mlrvals in place: Mlrmap.PutCopy reassigns pe.Value, Assign reassigns
tvar.value. The only in-place mutation a scalar undergoes is idempotent
type-inference caching (printrep -> typed). So a local/param bound by reference
to a scalar source can never observe its source change, and reassigning the
local replaces its own pointer without touching the source -- capture-by-value
semantics are preserved. Maps and arrays, by contrast, ARE mutated in place by
indexed assignment (m[k]=v), so an aliased collection would corrupt its source;
those must still be deep-copied.
So copyForBind copies only arrays/maps and binds scalars by reference. (Return
values are independently safe: ReturnNode.Execute still deep-copies them.)
Measured (big.csv, 1M rows, best of 5):
UDF-heavy put (scalar args/locals): 1.84 -> 1.68 (~9%)
x = $a+$b; $s = x*2 (no UDF): 0.50 -> 0.48 (~4%)
Verified: go test ./pkg/... and full regression suite pass, plus targeted
alias-then-mutate tests: scalar locals capture-by-value (source change after
bind not observed; reassigning one of two aliases leaves the other intact;
mutating a scalar param leaves the caller field intact), and collections stay
independent (local/param/oosvar-element map copies isolate in-place mutation).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* Batch-allocate per-record objects; reuse CSV writer field buffer
After batch-arena field allocation, profiling cat over 1M-record CSV showed
the remaining ~5M allocations were almost entirely per-record (one each):
the Mlrmap struct, the RecordAndContext wrapper, the CSV writer's []string,
and the go-csv parser's own buffers.
Address the first three:
- mlrval.RecordArena gains NewRecord(), vending the Mlrmap struct itself from
a per-batch slab (respecting --no-hash-records). Rolled out to every
line-based reader (CSV, CSV-lite, TSV, DKVP, NIDX, PPRINT, XTAB, DKVPX) in
place of NewMlrmapAsRecord.
- The CSV reader batch-allocates RecordAndContext wrappers from a per-batch
slab instead of one heap object per record (comment/output-string entries
still allocate individually, but they are rare).
- RecordWriterCSV reuses a single fieldsBuffer []string across records instead
of allocating one per Write; WriteCSVRecordMaybeColorized consumes it
synchronously and the writer is single-goroutine, so this is safe.
Effect (big.*, 1M records, cat, best of 5):
csv 0.26 -> 0.22
dkvp 0.51 -> 0.45 (Mlrmap slab)
For CSV, cat's allocation-object count drops ~5.0M -> ~2.1M. The remaining
~2M are the go-csv parser's per-record backing string and field slice, which
are intrinsic to parsing and would require a zero-copy/batch-slab parser
rework. A CPU profile of cat now shows it is I/O-bound (syscall ~56%, bufio
read+flush), with allocation/GC down to ~10% -- i.e. further allocation
trimming no longer moves cat's wall-clock. GOGC=off confirms (no change).
Verified: go test ./pkg/... and full regression suite pass; output is
byte-identical across all formats including record-retaining verbs (tac),
hashed and --no-hash-records.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Pool DSL stack frames across records (~8-9% on put)
A StackFrameSet lives on the persistent runtime.State and is reused across
all records, but every block entry (StatementBlockNode.Execute does
PushStackFrame/PopStackFrame, which runs once per record for the main block,
plus once per if/for/etc.) allocated a fresh StackFrame -- a []*var slice and
a map[string]int -- and discarded it on exit. For `put`/`filter` that is
millions of throwaway allocations.
Since push/pop is strictly LIFO, retain popped frames in a per-frameset free
list and clear-and-reuse them on the next push. After the first record
establishes the max block-nesting depth, per-record block execution is
allocation-free for frames. len(stackFrames) remains the logical depth, so
get/set/defineTyped/unset/etc. are unchanged.
Measured (big.csv, 1M rows, best of 4):
put chain-1 0.78 -> 0.72 (~8%)
put chain-4 0.96 -> 0.87 (~9%)
Allocation objects for put chain-1 drop ~23.1M -> ~20.0M (the per-record
newStackFrame churn, ~2.86M, is eliminated). UDF calls still allocate a fresh
frameset per call (PushStackFrameSet); pooling those is a separate change.
The dominant remaining DSL allocator is FromFloat (~6.8M, interior arithmetic
temporaries); eliminating it needs node-owned result slots + in-place bif
variants, a much larger and aliasing-sensitive change, left for follow-up.
Verified: go test ./pkg/... and full regression suite pass; put output is
byte-identical, including UDFs with locals/loops/blocks.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Pool DSL stack-frame *sets* across UDF/subr calls (~31% on function-heavy put)
Companion to the per-block frame pooling: that left PushStackFrameSet /
PopStackFrameSet (entered once per user-defined function or subroutine call)
allocating. Each call did newStackFrameSet() -- a StackFrameSet plus its
initial StackFrame (a slice and a map) -- AND, worse, prepended it with
append([]*StackFrameSet{head}, sets...), allocating a fresh backing slice and
copying the whole save-stack every call.
Two changes:
- Treat the frameset save-stack as a tail stack (append to push, truncate to
pop) instead of prepending at index 0. get/set only ever touch the cached
head, so list order is irrelevant; this removes the per-call slice
realloc + O(depth) copy.
- Pool popped framesets (LIFO) and reset-and-reuse them on the next push,
mirroring the per-frameset frame free list. A reset trims back to one
cleared base frame (extras go to the frame pool). After warmup, repeated
calls allocate no framesets or frames.
Measured (big.csv, 1M rows, best of 5):
put, 2 nested func calls/record: 2.73 -> 1.87 (~31%)
GC cycles 25 -> 16; newStackFrameSet/newStackFrame fall out of the allocation
profile entirely. (chain-1 etc. have no UDFs and are unaffected.)
Verified: go test ./pkg/... and full regression suite pass; recursion
(fact/fib), local-scope isolation, and subroutine+oosvar all correct.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Drop redundant deep-copy of UDF return values (~3-16% on UDF put)
A user-defined function's return value was deep-copied twice on the way out:
once in ReturnNode.Execute (returnValue.Copy() when building the block-exit
payload) and again in UDFCallsite.EvaluateWithArguments
(blockReturnValue.Copy() at the end).
The ReturnNode copy is the necessary one: it detaches the value from the
callee's frame so it survives the frame being popped (and, since perf-try-7,
pooled/reused). By the time EvaluateWithArguments returns, blockReturnValue is
therefore already an independent deep copy, so the second copy is pure waste --
and callers that retain the result copy again anyway (field/oosvar/local
assignment all PutCopy/Copy). The other return paths (implicit-absent, error)
don't use blockReturnValue, so this only affects the BLOCK_EXIT_RETURN_VALUE
path.
Return blockReturnValue directly.
Measured (big.csv, 1M rows, best of 5):
put, 2 nested scalar-returning calls/record: 1.89 -> 1.83 (~3%)
put, map-returning func per record: 2.34 -> 1.97 (~16%)
Win scales with return-value size (the avoided copy is deep). All UDF/HOF
callsites (apply/reduce/sort/select/fold/...) go through this path.
Verified: go test ./pkg/... and full regression suite pass; recursion, HOFs,
and returned-map isolation (mutating a returned map does not affect a
subsequent call) all correct.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* Batch-allocate per-record objects; reuse CSV writer field buffer
After batch-arena field allocation, profiling cat over 1M-record CSV showed
the remaining ~5M allocations were almost entirely per-record (one each):
the Mlrmap struct, the RecordAndContext wrapper, the CSV writer's []string,
and the go-csv parser's own buffers.
Address the first three:
- mlrval.RecordArena gains NewRecord(), vending the Mlrmap struct itself from
a per-batch slab (respecting --no-hash-records). Rolled out to every
line-based reader (CSV, CSV-lite, TSV, DKVP, NIDX, PPRINT, XTAB, DKVPX) in
place of NewMlrmapAsRecord.
- The CSV reader batch-allocates RecordAndContext wrappers from a per-batch
slab instead of one heap object per record (comment/output-string entries
still allocate individually, but they are rare).
- RecordWriterCSV reuses a single fieldsBuffer []string across records instead
of allocating one per Write; WriteCSVRecordMaybeColorized consumes it
synchronously and the writer is single-goroutine, so this is safe.
Effect (big.*, 1M records, cat, best of 5):
csv 0.26 -> 0.22
dkvp 0.51 -> 0.45 (Mlrmap slab)
For CSV, cat's allocation-object count drops ~5.0M -> ~2.1M. The remaining
~2M are the go-csv parser's per-record backing string and field slice, which
are intrinsic to parsing and would require a zero-copy/batch-slab parser
rework. A CPU profile of cat now shows it is I/O-bound (syscall ~56%, bufio
read+flush), with allocation/GC down to ~10% -- i.e. further allocation
trimming no longer moves cat's wall-clock. GOGC=off confirms (no change).
Verified: go test ./pkg/... and full regression suite pass; output is
byte-identical across all formats including record-retaining verbs (tac),
hashed and --no-hash-records.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Pool DSL stack frames across records (~8-9% on put)
A StackFrameSet lives on the persistent runtime.State and is reused across
all records, but every block entry (StatementBlockNode.Execute does
PushStackFrame/PopStackFrame, which runs once per record for the main block,
plus once per if/for/etc.) allocated a fresh StackFrame -- a []*var slice and
a map[string]int -- and discarded it on exit. For `put`/`filter` that is
millions of throwaway allocations.
Since push/pop is strictly LIFO, retain popped frames in a per-frameset free
list and clear-and-reuse them on the next push. After the first record
establishes the max block-nesting depth, per-record block execution is
allocation-free for frames. len(stackFrames) remains the logical depth, so
get/set/defineTyped/unset/etc. are unchanged.
Measured (big.csv, 1M rows, best of 4):
put chain-1 0.78 -> 0.72 (~8%)
put chain-4 0.96 -> 0.87 (~9%)
Allocation objects for put chain-1 drop ~23.1M -> ~20.0M (the per-record
newStackFrame churn, ~2.86M, is eliminated). UDF calls still allocate a fresh
frameset per call (PushStackFrameSet); pooling those is a separate change.
The dominant remaining DSL allocator is FromFloat (~6.8M, interior arithmetic
temporaries); eliminating it needs node-owned result slots + in-place bif
variants, a much larger and aliasing-sensitive change, left for follow-up.
Verified: go test ./pkg/... and full regression suite pass; put output is
byte-identical, including UDFs with locals/loops/blocks.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Pool DSL stack-frame *sets* across UDF/subr calls (~31% on function-heavy put)
Companion to the per-block frame pooling: that left PushStackFrameSet /
PopStackFrameSet (entered once per user-defined function or subroutine call)
allocating. Each call did newStackFrameSet() -- a StackFrameSet plus its
initial StackFrame (a slice and a map) -- AND, worse, prepended it with
append([]*StackFrameSet{head}, sets...), allocating a fresh backing slice and
copying the whole save-stack every call.
Two changes:
- Treat the frameset save-stack as a tail stack (append to push, truncate to
pop) instead of prepending at index 0. get/set only ever touch the cached
head, so list order is irrelevant; this removes the per-call slice
realloc + O(depth) copy.
- Pool popped framesets (LIFO) and reset-and-reuse them on the next push,
mirroring the per-frameset frame free list. A reset trims back to one
cleared base frame (extras go to the frame pool). After warmup, repeated
calls allocate no framesets or frames.
Measured (big.csv, 1M rows, best of 5):
put, 2 nested func calls/record: 2.73 -> 1.87 (~31%)
GC cycles 25 -> 16; newStackFrameSet/newStackFrame fall out of the allocation
profile entirely. (chain-1 etc. have no UDFs and are unaffected.)
Verified: go test ./pkg/... and full regression suite pass; recursion
(fact/fib), local-scope isolation, and subroutine+oosvar all correct.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* Batch-allocate per-record objects; reuse CSV writer field buffer
After batch-arena field allocation, profiling cat over 1M-record CSV showed
the remaining ~5M allocations were almost entirely per-record (one each):
the Mlrmap struct, the RecordAndContext wrapper, the CSV writer's []string,
and the go-csv parser's own buffers.
Address the first three:
- mlrval.RecordArena gains NewRecord(), vending the Mlrmap struct itself from
a per-batch slab (respecting --no-hash-records). Rolled out to every
line-based reader (CSV, CSV-lite, TSV, DKVP, NIDX, PPRINT, XTAB, DKVPX) in
place of NewMlrmapAsRecord.
- The CSV reader batch-allocates RecordAndContext wrappers from a per-batch
slab instead of one heap object per record (comment/output-string entries
still allocate individually, but they are rare).
- RecordWriterCSV reuses a single fieldsBuffer []string across records instead
of allocating one per Write; WriteCSVRecordMaybeColorized consumes it
synchronously and the writer is single-goroutine, so this is safe.
Effect (big.*, 1M records, cat, best of 5):
csv 0.26 -> 0.22
dkvp 0.51 -> 0.45 (Mlrmap slab)
For CSV, cat's allocation-object count drops ~5.0M -> ~2.1M. The remaining
~2M are the go-csv parser's per-record backing string and field slice, which
are intrinsic to parsing and would require a zero-copy/batch-slab parser
rework. A CPU profile of cat now shows it is I/O-bound (syscall ~56%, bufio
read+flush), with allocation/GC down to ~10% -- i.e. further allocation
trimming no longer moves cat's wall-clock. GOGC=off confirms (no change).
Verified: go test ./pkg/... and full regression suite pass; output is
byte-identical across all formats including record-retaining verbs (tac),
hashed and --no-hash-records.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Pool DSL stack frames across records (~8-9% on put)
A StackFrameSet lives on the persistent runtime.State and is reused across
all records, but every block entry (StatementBlockNode.Execute does
PushStackFrame/PopStackFrame, which runs once per record for the main block,
plus once per if/for/etc.) allocated a fresh StackFrame -- a []*var slice and
a map[string]int -- and discarded it on exit. For `put`/`filter` that is
millions of throwaway allocations.
Since push/pop is strictly LIFO, retain popped frames in a per-frameset free
list and clear-and-reuse them on the next push. After the first record
establishes the max block-nesting depth, per-record block execution is
allocation-free for frames. len(stackFrames) remains the logical depth, so
get/set/defineTyped/unset/etc. are unchanged.
Measured (big.csv, 1M rows, best of 4):
put chain-1 0.78 -> 0.72 (~8%)
put chain-4 0.96 -> 0.87 (~9%)
Allocation objects for put chain-1 drop ~23.1M -> ~20.0M (the per-record
newStackFrame churn, ~2.86M, is eliminated). UDF calls still allocate a fresh
frameset per call (PushStackFrameSet); pooling those is a separate change.
The dominant remaining DSL allocator is FromFloat (~6.8M, interior arithmetic
temporaries); eliminating it needs node-owned result slots + in-place bif
variants, a much larger and aliasing-sensitive change, left for follow-up.
Verified: go test ./pkg/... and full regression suite pass; put output is
byte-identical, including UDFs with locals/loops/blocks.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
After batch-arena field allocation, profiling cat over 1M-record CSV showed
the remaining ~5M allocations were almost entirely per-record (one each):
the Mlrmap struct, the RecordAndContext wrapper, the CSV writer's []string,
and the go-csv parser's own buffers.
Address the first three:
- mlrval.RecordArena gains NewRecord(), vending the Mlrmap struct itself from
a per-batch slab (respecting --no-hash-records). Rolled out to every
line-based reader (CSV, CSV-lite, TSV, DKVP, NIDX, PPRINT, XTAB, DKVPX) in
place of NewMlrmapAsRecord.
- The CSV reader batch-allocates RecordAndContext wrappers from a per-batch
slab instead of one heap object per record (comment/output-string entries
still allocate individually, but they are rare).
- RecordWriterCSV reuses a single fieldsBuffer []string across records instead
of allocating one per Write; WriteCSVRecordMaybeColorized consumes it
synchronously and the writer is single-goroutine, so this is safe.
Effect (big.*, 1M records, cat, best of 5):
csv 0.26 -> 0.22
dkvp 0.51 -> 0.45 (Mlrmap slab)
For CSV, cat's allocation-object count drops ~5.0M -> ~2.1M. The remaining
~2M are the go-csv parser's per-record backing string and field slice, which
are intrinsic to parsing and would require a zero-copy/batch-slab parser
rework. A CPU profile of cat now shows it is I/O-bound (syscall ~56%, bufio
read+flush), with allocation/GC down to ~10% -- i.e. further allocation
trimming no longer moves cat's wall-clock. GOGC=off confirms (no change).
Verified: go test ./pkg/... and full regression suite pass; output is
byte-identical across all formats including record-retaining verbs (tac),
hashed and --no-hash-records.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* Lazy per-record hashing: ~15-30% faster on common workloads
Records (NewMlrmapAsRecord) eagerly allocated and populated a
map[string]*MlrmapEntry on construction whenever hashRecords was true
(the default). For streaming verbs that never look records up by key
(e.g. `mlr cat`) that map is pure overhead: a heap allocation plus N
map-inserts per record, and N more pointer-heavy objects for the GC to
scan. Profiling 1M-record CSV shows runtime allocation/GC machinery
dominating every workload, and `--no-hash-records` was 25-30% faster --
but that flag makes wide-record lookups O(n), the regression that
motivated hashing in #1506.
Make record hashing lazy instead: allocate no index up front; build it
in findEntry on the first lookup, and only when the record is wide
enough (FieldCount >= mlrmapHashThreshold) that linear search would
hurt. Narrow records and never-looked-up records never pay for a map;
wide records that are actually queried still get hash-accelerated
lookups, matching the old eager-hash default. DSL maps (NewMlrmap) keep
eager hashing to limit the behavioral surface.
This is transparent: findEntry already fell back to linear scan when
keysToEntries was nil, and every mutator already guarded on
keysToEntries != nil.
Measured (big.csv, 1M x 7 cols, default flags, best of 3):
cat 0.62 -> 0.47 (~24%)
put 1.08 -> 0.82 (~24%)
stats1 0.66 -> 0.57 (~14%)
sort 2.9 -> 2.0 (~30%)
Wide-column case protected: 60-col file with field lookups, lazy (1.42s)
matches old eager default (1.40s) and beats pure linear (1.55s).
Verified: go test ./pkg/... and full regression suite pass; output is
byte-identical to forced --hash-records for sort, stats1, cut,
wide-column put, and duplicate-key dedupe.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Batch-arena field allocation for line-based readers (approach B)
Following the lazy-hashing commit, profiling showed the dominant remaining
cost in read/write-bound workloads is allocation *operations* (not bytes):
each input field allocated two heap objects -- an Mlrval (FromDeferredType)
and an MlrmapEntry. For 1M x 7-field CSV that is ~13.4M of ~18.6M total
allocations.
Introduce mlrval.RecordArena, a per-batch slab allocator: a reader draws
each field's entry and value from contiguous []MlrmapEntry / []Mlrval slabs,
turning two allocations per field into roughly two per slab. The arena grows
on demand, so the size hint need not be exact; on duplicate keys it mirrors
PutReferenceMaybeDedupe semantics. findEntry/linkNewEntry already supported
externally-constructed entries, so this is transparent.
Wired into every line-based reader that builds records from deferred-type
strings: CSV, CSV-lite, TSV, DKVP, NIDX, PPRINT, XTAB, DKVPX. (JSON values
arrive already typed and are unaffected.) Readers with inline batch loops use
a local arena; those that build records via a helper (DKVP/NIDX line
splitter, XTAB stanza) hold the arena on the reader struct, reset per batch
and also initialized in the constructor so direct/test callers never see nil.
Measured (big.*, 1M records, default flags, cat, best of 3):
csv 0.46 -> 0.27 (~41%)
dkvp 0.75 -> 0.46 (~39%)
nidx 1.92 -> 1.58 (~18%)
(xtab ~flat: dominated by stanza parse/emit, not field allocation)
For cat the allocation-object count drops from ~18.6M to ~4.85M and peak RSS
from ~402MB to ~237MB (slabs are compact and freed as units). Alloc *bytes*
are essentially unchanged -- confirming the cost was per-allocation overhead,
not volume. Streaming and accumulating verbs (put/sort) are unchanged: their
bottleneck is DSL-side allocation / heap scanning, not field construction.
Verified: go test ./pkg/... and full regression suite pass; output is
byte-identical across all formats (hashed and --no-hash-records).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Records (NewMlrmapAsRecord) eagerly allocated and populated a
map[string]*MlrmapEntry on construction whenever hashRecords was true
(the default). For streaming verbs that never look records up by key
(e.g. `mlr cat`) that map is pure overhead: a heap allocation plus N
map-inserts per record, and N more pointer-heavy objects for the GC to
scan. Profiling 1M-record CSV shows runtime allocation/GC machinery
dominating every workload, and `--no-hash-records` was 25-30% faster --
but that flag makes wide-record lookups O(n), the regression that
motivated hashing in #1506.
Make record hashing lazy instead: allocate no index up front; build it
in findEntry on the first lookup, and only when the record is wide
enough (FieldCount >= mlrmapHashThreshold) that linear search would
hurt. Narrow records and never-looked-up records never pay for a map;
wide records that are actually queried still get hash-accelerated
lookups, matching the old eager-hash default. DSL maps (NewMlrmap) keep
eager hashing to limit the behavioral surface.
This is transparent: findEntry already fell back to linear scan when
keysToEntries was nil, and every mutator already guarded on
keysToEntries != nil.
Measured (big.csv, 1M x 7 cols, default flags, best of 3):
cat 0.62 -> 0.47 (~24%)
put 1.08 -> 0.82 (~24%)
stats1 0.66 -> 0.57 (~14%)
sort 2.9 -> 2.0 (~30%)
Wide-column case protected: 60-col file with field lookups, lazy (1.42s)
matches old eager default (1.40s) and beats pure linear (1.55s).
Verified: go test ./pkg/... and full regression suite pass; output is
byte-identical to forced --hash-records for sort, stats1, cut,
wide-column put, and duplicate-key dedupe.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
