Skip to content

ROADMAP.md

Latest commit

 

History

History
136 lines (110 loc) · 5.82 KB

ROADMAP.md

File metadata and controls

136 lines (110 loc) · 5.82 KB

Goro Roadmap

Current state (April 2026): 11,838 / 12,121 PHP 8.5.5 tests passing (97.7%). 196 failures, 87 skipped in CI. All 20 vendored extensions implemented in pure Go; CGO_ENABLED=0 go build ./... succeeds.

With the high-level features in place, remaining work is mostly a long tail of targeted fixes: error-message format parity, a handful of reflection/attribute edge cases, DST/DatePeriod subtleties, and a few tests gated on language features we haven't implemented (goto, full exec() subprocess support).

Remaining failures by area

From the last CI run (commit 852b33af). The Other category is the long tail — ~40 distinct areas, at most three failures each.

Area Failures Notes
ext/date 51 DatePeriod serialization format, date_parse edge cases, DST fallback transitions
attributes 20 Reflection __toString formatting, delayed target validation, AST printing
exceptions 9 __toString error location, variance autoload, stream wrappers
closures 9 Closure const expressions, binding edge cases
clone 6 AST printing, clone-with edge cases
exit 5 Tests rely on exec() subprocess spawning, disabling exit
constexpr 5 New-in-defaults, array unpack in const, stack trace format
asymmetric_visibility 5 Static props, nested variations, indirect modification
assert 5 Callback exceptions, ??= in assert, AST pretty-printer
ext/mbstring 4 Encoding conversion edge cases
constants 4 Constant evaluation edge cases
ext/hash 3 PHP serialization format edge cases
ext/gmp 3 GMP unserialize with references
Other ~67 Scattered across ~40 areas

Known hard corners

Goro mirrors PHP's reference-counted object model: every ZObject carries an int32 refcount (see core/phpobj/zobject.go, IncRef/DecRef), object IDs are recycled via a free list in core/phpctx/global.go (NextObjectID/ReleaseObjectID), and destructors fire synchronously when refcount hits zero. We do NOT rely on Go's tracing GC for observable PHP semantics.

That said, a few engine-level mismatches still bite:

  • Function-argument lifetime differs — Go temporaries holding a callee's arguments stay alive until the enclosing Go statement finishes, which shifts object IDs by one step in tests that interleave new with function calls. (~3 ext/gmp tests and a scattering elsewhere — search the skip list for "object ID timing".)
  • goto — not implemented. (~5 tests.)
  • debug_zval_dump refcount display — PHP prints the exact refcount; some of ours are off by one because we don't reliably decrement for short-lived intermediates. (~handful of tests.)
  • gmp_random_* seeding — mt_rand/rand use real MT19937 via github.com/goark/mt/mt19937 and produce byte-identical output to PHP for a given seed. But ext/gmp/misc.go seeds Go's legacy math/rand (LCG-style) for its random helpers, which disagrees with libgmp's own MT state. One skipped test (gmp_random_seed.phpt). Switching gmp over to the same MT source would at least make it internally consistent, but still wouldn't byte-match PHP's libgmp.
  • SensitiveParameter backtrace redaction — not implemented.

Test suite coverage

The vendored PHP 8.5.5 test set is 12,121 tests. Ten extensions have their tests imported (ctype, date, gmp, hash, json, mbstring, pcre, reflection, spl, standard), alongside Zend/tests (core language) and tests/ (general).

Area Tests Pass rate
ext/standard 3,808 high
ext/spl 781 ~82%
ext/date 688 ~93%
ext/reflection 493 ~75%
ext/mbstring 416 ~97%
ext/pcre 163 67% (PCRE2 via gopcre2)
ext/gmp 99 97%
ext/json 88 98%
ext/hash 80 94%
ext/ctype 49 100%

Not yet vendored: ext/{bz2, curl, gd, mbstring (some subdirs), mysqli, openssl, session, sockets, sqlite3, xml, zlib} tests. The extensions themselves are implemented (see README), but the upstream test suites aren't imported into the repo yet.

Planned work

Short-term (individual test fixes)

  • ext/date, attributes, exceptions, closures — these four buckets account for ~50% of remaining failures. Each is a grab-bag of small fixes (error-message format, reflection output, AST printing).
  • AST pretty-printer — multiple test areas (assert, clone, attributes, constexpr) depend on it matching PHP's exact format. A focused pass on the printer would clear a dozen tests.
  • Reflection __toString — the remaining attribute/reflection failures largely come from slight format differences.

Medium-term (features)

  • PDO — via database/sql; the MySQL and SQLite drivers are already in the tree (go-sql-driver/mysql, glebarez/go-sqlite).
  • iconv — maps cleanly onto golang.org/x/text/transform.
  • exec() subprocess support — would unlock the exit/ tests.

Longer-term / won't-do

  • intl — full ICU is substantial; leave unless a concrete user needs it.
  • Phar — PHP archive format; rarely needed outside legacy code.
  • Deterministic object-id / destructor semantics — would require moving off Go's GC onto refcounting throughout the value system. Test impact is ~50 scattered failures; not worth the cost.

Recent milestones

  • April 2026 — PHP test suite refreshed to 8.5.5 (12,110 → 12,121 tests). All cgo removed; pure-Go bcrypt/SHA-crypt/MD5-crypt and a POSIX "C" locale emulation replaced the libc wrappers.
  • March 2026 — 20th extension (gd, pure Go via gogd) added; pass rate crossed 97% on the full PHP suite.
  • Earlier 2026 — New extensions (session, xml, curl, sockets, zlib, mysqli, sqlite3, bz2); library integrations (gopcre2, gotz, gobzip2, go-sql-driver/mysql, glebarez/go-sqlite).