* Reduce marshal and unmarshal overhead
Targeted optimizations to reduce performance overhead introduced by
recent feature additions and the unsafe removal.
Unmarshal:
- parseKeyval: access the node directly in the builder's slice to set
Raw, bypassing NodeAt which triggers a GC write barrier for the
nodes-pointer on every key-value expression.
- Iterator.Next: cache the *nodes slice dereference in a local variable
to avoid repeated pointer-to-slice indirection in the hot loop.
Marshal:
- Guard shouldOmitZero calls with an inlineable options.omitzero check.
shouldOmitZero has inlining cost 1145 (budget 80), so avoiding the
function call when omitzero is not set removes per-field overhead.
- Inline the isNil check in encodeMap. isNil has inlining cost 93
(budget 80), so expanding it at the single hot call site avoids
per-map-entry function call overhead.
Update README benchmarks.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Fix benchmark script replacing internal package imports
The sed command in bench() was replacing all occurrences of the go-toml
module path, including sub-package imports like internal/assert. This
caused the BurntSushi/toml benchmark to fail because it tried to import
github.com/BurntSushi/toml/internal/assert which doesn't exist.
Fix by anchoring the sed pattern to only match the import path when
followed by a closing quote, preserving internal package imports.
Also add a guard in the benchstathtml Python script to give a clear
error instead of an IndexError when no benchmark results are available.
https://claude.ai/code/session_016JGASo49PeFSfCaDxvrGFE
* Update benchmark results in README
https://claude.ai/code/session_016JGASo49PeFSfCaDxvrGFE
---------
Co-authored-by: Claude <noreply@anthropic.com>
When parsing a key without '=' at EOF (e.g., "a = 1\nb = 2\nc"), the
error highlight was an empty slice, causing subsliceOffset to return 0
and the error to point at line 1 instead of line 3. Pass the consumed
key bytes as the highlight instead of the empty remainder.
Fixes#1032https://claude.ai/code/session_01UWv8pyc8P1ktAPfHpveixj
Co-authored-by: Claude <noreply@anthropic.com>
Fixes#873
Extend the unstable.Unmarshaler interface support to work with tables
and array tables, not just single values.
When a type implementing unstable.Unmarshaler is the target of a table
(e.g., [table] or [[array]]), the UnmarshalTOML method receives a
synthetic InlineTable node containing all the key-value pairs belonging
to that table.
Key changes:
- Add handleKeyValuesUnmarshaler to collect and process table content
- Add copyExpressionNodes to deep-copy AST nodes for synthetic tables
- Add helper functions in unstable/ast.go for node manipulation
- Update documentation for EnableUnmarshalerInterface
- Add comprehensive tests for table and array table unmarshaling
* Implement bytes-based Unmarshaler interface for tables and arrays (#873)
This change brings back support for the unstable.Unmarshaler interface
for tables and array tables, addressing issue #873.
Key changes:
- Changed UnmarshalTOML signature from (*Node) to ([]byte) to provide
raw TOML bytes instead of AST nodes
- Added RawMessage type (similar to json.RawMessage) for capturing raw
TOML bytes for later processing
- Updated handleKeyValuesUnmarshaler to reconstruct key-value lines
from the parsed keys and raw value bytes
- Added support for slice types implementing Unmarshaler (e.g., RawMessage)
- Removed unused AST helper functions from unstable/ast.go
The bytes-based interface allows users to:
- Get raw TOML bytes for custom parsing
- Delay TOML decoding using RawMessage
- Implement custom unmarshaling logic for complex types
Tests added for:
- Table unmarshaler with various scenarios
- Array table unmarshaler
- Split tables (same parent defined in multiple places)
- RawMessage usage
- Nested tables and mixed regular fields
* Fix lint issues and improve test coverage for Unmarshaler interface
- Apply De Morgan's law in keyNeedsQuoting to satisfy staticcheck QF1001
- Remove unused splitTableUnmarshaler type from test
- Fix unused parameter lint warning in errorUnmarshaler873
- Add test for quoted keys that need special handling
- Add test for error propagation from UnmarshalTOML
- Update customTable873 parser to handle quoted keys properly
Coverage improved:
- handleKeyValuesUnmarshaler: 80.0% -> 93.3%
- keyNeedsQuoting: 66.7% -> 83.3%
- Overall main package: 97.2% -> 97.5%
* Add test for dotted keys to improve coverage
Add TestIssue873_DottedKeys to test dotted key handling (e.g., sub.key = value)
in the Unmarshaler interface. This improves coverage for handleKeyValuesUnmarshaler
from 93.3% to 96.7%.
* Add double pointer test to achieve 100% coverage for handleKeyValues
Add TestIssue873_DoublePointerUnmarshaler to test pointer-to-pointer
to Unmarshaler types. This covers the pointer dereferencing loop in
handleKeyValues, bringing its coverage from 88% to 100%.
Total coverage: 97.4%
* Add Example tests and fix raw value extraction for boolean types
Add two godoc Example tests:
- ExampleDecoder_EnableUnmarshalerInterface_dynamicConfig: shows dynamic
unmarshaling based on a type field
- ExampleDecoder_EnableUnmarshalerInterface_rawMessage: demonstrates
RawMessage usage for deferred parsing
Fix handleKeyValuesUnmarshaler to handle values where Raw.Length == 0
(like boolean types) by using value.Data as fallback.
* Preserve original formatting in Unmarshaler by using raw byte ranges
Instead of reconstructing key-value lines from parsed components, now
uses the original raw bytes from the document. This preserves:
- Whitespace around '=' (e.g., "key = value")
- String quoting style (basic vs literal)
- Number formats (hex, octal, binary)
- Inline table formatting
Changes:
- Add Raw range tracking to KeyValue expressions in parseKeyval
- Update handleKeyValuesUnmarshaler to use expr.Raw directly
- Remove keyNeedsQuoting helper (no longer needed)
- Add TestIssue873_FormattingPreservation test
- Update expected output in ExampleParser_comments
* Prevent test matrix from canceling on first failure
Add fail-fast: false to the test workflow strategy so that all
OS/Go version combinations continue running even if one fails.
This provides better visibility into which specific combinations
have issues.
---------
Co-authored-by: Claude <noreply@anthropic.com>
Return a type mismatch error instead of panicking when datetime values
(DateTime, LocalDate, LocalTime, LocalDateTime) are unmarshaled into
incompatible Go types. This makes the decoder safer for processing
untrusted TOML input.
https://claude.ai/code/session_011jwvtDS5M2KncLrqJpgMr5
Co-authored-by: Claude <noreply@anthropic.com>
When a type implements encoding.TextUnmarshaler, the unmarshaler now
skips calling UnmarshalText for Array and InlineTable TOML values.
This allows types to support both:
- Simple string values via UnmarshalText
- Structured table values via field-by-field unmarshaling
Previously, UnmarshalText was called unconditionally, which prevented
proper struct unmarshaling when the TOML value was a table or array
of tables.
Co-authored-by: Claude <noreply@anthropic.com>
When marshaling a map with nil pointer values, the keys were being
silently dropped, breaking round-trip fidelity. For example:
map[string]*struct{}{"foo": nil}
Would produce an empty TOML document instead of "[foo]".
This change converts nil pointer values in maps to their zero values
(consistent with how nil pointers in slices are handled), allowing the
keys to be preserved as empty tables.
Nil interface values (map[string]any{"foo": nil}) are still skipped
since there's no type information to derive a zero value.
Fixes#975
Also, pin golangci-lint version to v2.8.0 in CI and document in AGENTS.md
- Explicitly set golangci-lint version in lint.yml to ensure consistent
behavior across CI runs
- Update AGENTS.md with instructions to use the same linter version locally
---------
Co-authored-by: Claude <noreply@anthropic.com>
The omitzero tag now respects custom IsZero() methods on types,
similar to how encoding/json handles this. Previously, only
reflect.Value.IsZero() was used, which ignores user-defined
implementations.
Fixes#1003
Co-authored-by: Claude <noreply@anthropic.com>
Go's time.Date() normalizes leap seconds (second=60) by adding 1 minute.
When parsing the maximum valid TOML date 9999-12-31 23:59:60z, this causes
the year to overflow to 10000, which exceeds the valid TOML year range
(0000-9999) and breaks round-trip serialization.
The fix rejects leap seconds (second > 59) during parsing. This is
consistent with the resolution of issue #913 which determined that
emitting an error is less surprising than silently normalizing leap
seconds.
Fixes#1015
Co-authored-by: Claude <noreply@anthropic.com>
Removes all unsafe operations from go-toml, making the codebase
fully safe Go code. The internal/danger package that contained
unsafe operations has been deleted.
Changes:
- Replace pointer-based node navigation with index-based navigation
- Node.next and Node.child now store absolute indices into the
backing nodes slice instead of relative offsets
- Add nodes pointer to Node and Iterator for safe navigation
- Replace danger.TypeID with reflect.Type for cache keys
- Delete internal/danger package entirely
Performance overhead is under 10% compared to the unsafe version,
which is acceptable for the safety and maintainability benefits.
[Cursor][claude-sonnet-4-20250514]
This file provides a concise summary of the contribution guidelines
from CONTRIBUTING.md, specifically tailored for AI agents working on
the codebase. It covers testing requirements, backward compatibility,
performance considerations, and code style expectations.
Co-authored-by: Claude <noreply@anthropic.com>
Recent versions of Go object to the use of non-constant variables a
format strings. This commit fixes errors like this:
cli.go:26:47: non-constant format string in call to fmt.Fprintf
Signed-off-by: W. Michael Petullo <mike@flyn.org>
Only 3 consecutive quotation marks need to be quoted. We choose to quote
all quotation marks in a sequence if there are 3 or more consecutive
present.
Fixes#990
---------
Co-authored-by: Thomas Pelletier <thomas@pelletier.dev>
As recommended, an `internal/assert` package was added with a reduced set of assertions. All tests were then refactored to use the internal assertions. When more complex assertions were used, they have been rewritten using logic and the simplified assertions.
Fancy formatting for failures was omitted. The `internal/assert/assertions.diff` function could be overwritten for better formatting. That is where diff libraries are used in other test suites.
Refs: #872
Co-authored-by: Alex Mikitik <alex.mikitik@oracle.com>
This way, calls to Unmarshal or Decoder.Decode allocate once
at the start rather than twice.
│ old │ new │
│ allocs/op │ allocs/op vs base │
Unmarshal/HugoFrontMatter-8 141.0 ± 0% 140.0 ± 0% -0.71% (p=0.002 n=6)
* go.mod: bump minimum and language to 1.21
CI only tests Go 1.21 and 1.22, and older versions of Go are no longer
getting any bug or security fixes, so advertise that we only support
Go 1.21 or later via go.mod.
While here, ensure the module is tidy and resolve deprecation warnings,
and remove now-unnecessary Go version build tags.
* replace sort.Slice with slices.SortFunc
The latter is more efficient, and allocates less, since sort.Slice
needs to go through sort.Interface which causes allocations.
goos: linux
goarch: amd64
pkg: github.com/pelletier/go-toml/v2/benchmark
cpu: AMD Ryzen 7 PRO 5850U with Radeon Graphics
│ old │ new │
│ sec/op │ sec/op vs base │
Marshal/HugoFrontMatter-8 7.612µ ± 1% 6.730µ ± 1% -11.59% (p=0.002 n=6)
│ old │ new │
│ B/s │ B/s vs base │
Marshal/HugoFrontMatter-8 65.52Mi ± 1% 74.11Mi ± 1% +13.11% (p=0.002 n=6)
│ old │ new │
│ B/op │ B/op vs base │
Marshal/HugoFrontMatter-8 5.672Ki ± 0% 5.266Ki ± 0% -7.16% (p=0.002 n=6)
│ old │ new │
│ allocs/op │ allocs/op vs base │
Marshal/HugoFrontMatter-8 85.00 ± 0% 73.00 ± 0% -14.12% (p=0.002 n=6)
Claude