| Age | Commit message (Collapse) | Author |
|---|
|
CompilingMatcher::match formats `<operator><constraint>;<version>` into a
fresh String on every call, including the cache hits the key exists to serve.
Profiling put the formatting at 2.9 % of self time, more than the match it
guards. Build the key into a reused thread-local buffer and copy it only when
there is a miss to record.
Taking the version as `&str` rather than `String` removes the copy the callers
made for the same reason: `optimize_by_identical_dependencies` allocated one
per candidate package in its innermost loop.
laravel/framework require --no-install (warm cache, network disabled):
instructions:u 9142799608 -> 8895247586 (-2.7 %)
cycles:u 4626440169 -> 4496152041 (-2.8 %)
wall (hyperfine, 25 runs)
1.239 s +- 0.014 s -> 1.227 s +- 0.016 s (-1.0 %)
monolog/monolog is unchanged (102.1 ms -> 101.6 ms, within noise).
composer.lock is byte-identical for both packages.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
startCachedAsyncDownload's cached branch decoded the provider JSON into an
owned value and then rebuilt it entry by entry, recursively cloning every
package definition only to drop the original. Match the decoded value by
value instead. PHP hands the decoded array to the closure by COW, so no copy
happens there either.
laravel/framework require --no-install (warm cache, network disabled):
instructions:u 10276126917 -> 9884102191 (-3.8 %)
cycles:u 5048262624 -> 4777881212 (-5.4 %)
wall (hyperfine, 20 runs)
1.345 s +- 0.015 s -> 1.316 s +- 0.018 s (-2.2 %)
monolog/monolog is unchanged (100.7 ms -> 101.7 ms, within noise).
composer.lock is byte-identical for both packages.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
create_packages handed ArrayLoader::load_packages a clone of the package
array it had just built and then dropped its own copy, so every package
definition in the response was deep-copied once for nothing.
Measured on `require laravel/laravel` (77 packages) against a warm cache
with the network disabled: the create_packages span drops from 420 ms to
275 ms, and the run from 2.333 s to 2.192 s. Peak RSS is unchanged -- the
copy was transient. `composer.lock` is byte-identical.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
MetadataMinifier::expand now returns ExpandedVersions, which holds the
minified input plus, for each expanded version, a table of references to
where its fields live. A version is copied only when materialize() asks
for it. ComposerRepository::load_async_packages runs its constraint and
stability filters straight off that view through the new VersionFields
trait, so the versions it rejects are never copied at all.
Benchmarks are new under crates/shirabe/benches. load_packages against
real packagist p2 metadata, before -> after:
symfony/console (768 versions)
0 accepted 9.01 ms -> 4.40 ms -51%
50 accepted 10.70 ms -> 6.30 ms -41%
147 accepted 13.17 ms -> 9.62 ms -27%
329 accepted 18.58 ms -> 16.89 ms -9%
663 accepted 27.27 ms -> 27.89 ms +2%
laravel/framework (1277 versions)
0 accepted 39.44 ms -> 11.22 ms -72%
81 accepted 44.76 ms -> 18.25 ms -59%
840 accepted 125.44 ms -> 120.12 ms -4%
1266 accepted 163.18 ms -> 182.01 ms +12%
The crossover sits near 80% acceptance. Past it the view loses, because
materialize rebuilds a map where the old code cloned one, and the
minified input stays alive alongside the copies; the 1266-of-1277 case
measured between +5% and +12% across runs. Loads with a real constraint
sit far below the crossover.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Cache::read()/write() rewrite every character outside the allowlist to
"-", so a package name containing "_" was stored and looked up under a
file name Composer never writes: the repository metadata cache was never
hit for those packages, and "vendor/foo_bar" collided with
"vendor/foo-bar". With no cached copy there is no last-modified date to
send either, so a degraded repository resolved such packages as not
found instead of serving the cached metadata.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
The capture groups were discarded at 162 of the preg_match call sites,
which only tested the Option. They now call preg_is_match, which lets the
regex engine skip capture tracking.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Every call site but one passed offset 0. The remaining one, the UTF-8
chunking loop in Application, slices the subject instead: its pattern has
no anchor or lookaround, so matching a suffix is equivalent to starting
the search at that offset.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Preg had shed everything it owned: after the last few rounds its methods
were one-line forwards to the shim's preg_*(), differing only in a
default argument or a wrapper the caller unwrapped anyway. The 460 call
sites now name the shim function, and shirabe-pcre is gone from the
workspace along with its LICENSE entry.
The forwards expand as they read: isMatch becomes
preg_match2(.., 0).is_some() (is_none() where PHP negates it), isMatch3
and match3 drop the .is_some(), matchAll counts through
preg_match_all2(..).occurrence_count(), and replace4/replace5 spell out
the limit and count arguments preg_replace2 takes. Callbacks are the one
place the shapes differ: preg_replace_callback carries an error out of
the callback, so the fourteen infallible closures wrap their result in
Ok() and expect() it back.
Config::process() is the fifteenth, and it drops the `error` cell it
captured to smuggle a failure past a closure that could only return a
String. The `?` in the closure now carries it, which is what the PHP
does -- a throw from the callback leaves preg_replace_callback at the
failing match rather than running the remaining replacements and
reporting the last error.
The module doc that explained why composer/pcre's exceptions and
*StrictGroups() variants have no counterpart moves to the shim's preg
module, where the functions it describes live.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
PregMatches keyed both forms of a capture group through CaptureKey, so
every read built one: a usize wrapped in an enum, or worse, a String
allocated to name a group that regex::Captures can look up from a &str.
It now mirrors regex::Captures instead -- get() takes the group number,
name() the group name -- and the enum drops out of the type entirely.
That is 285 call sites across 59 files, and the named ones carry most of
the win: `matches.get(&CaptureKey::ByName("host".to_string()))` reads as
`matches.name("host")`. ProcessExecutor loses a `user_key` binding that
existed only to build the key once.
CaptureKey stays as the key type of PregMatchesAll and
PregMatchesAllWithOffsets, where numbered and named entries share one
IndexMap and a key type is the point. Five files still name it.
Also retargets the two preg_match_all comments that described the
occurrence count through `matches[&CaptureKey::ByIndex(0)].len()`, an
Index impl these types no longer carry.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Preg::match4 and Preg::replace_callback gave callers a
PregMatchedGroups: an IndexMap rebuilt from the match with an owned
String per group, plus a second String for a named group's name key.
That is the copy PregMatches shed when it started wrapping
regex::Captures, reinstated one layer up -- and nearly every regex call
in the tree goes through Preg rather than the shim's preg_* directly, so
almost nothing saw the borrow.
PregMatchedGroups existed only to drop the null (unmatched) groups the
old PregMatches held as Option<String> values. PregMatches::get reports
a non-participating group as None on its own, so the two read alike and
the type collapses into it. Call sites still reach groups through
get(&CaptureKey::ByIndex(N)); what changes is that the value arrives as
a &str borrowed from the subject, which the signatures now carry as a
lifetime.
Three places needed the borrow reckoned with rather than a mechanical
rewrite: PhpFileCleaner::clean and Problem::get_messages read their
groups out before mutating what the match borrows, and
Git::get_authentication_failure names the lifetime of its url argument,
which the result borrows instead of self.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
`Composer\Pcre\Preg` fills `$matches` through a by-ref parameter, and the
port mirrored that with a `&mut` (or `Option<&mut>`) out-param plus a
bool or count return. Callers had to declare an empty map one line ahead
of the call, and the type never said the map is only meaningful when the
call matched. Return the matches instead:
- match3/match4/is_match3/is_match4 -> Option<PregMatchedGroups>
- is_match_named -> Option<PregNamedGroups>
- match_all2/is_match_all -> PregMatchesAll
- is_match_all_with_offsets3 -> PregMatchesAllWithOffsets
Nothing is lost: the bool is `Option::is_some()`, and the occurrence
count is the length of any one column of a PREG_PATTERN_ORDER map, now
spelled `PregMatchesAll::occurrence_count()`. is_match() still answers
the bool question directly for callers that want no groups.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
The five IndexMap shapes that the preg_* functions and Preg fill in are
now distinct types generated by preg_match_map!, so a matches map no
longer interchanges with any other map of the same key and value type.
Index<usize> is kept alongside Index<&Q> because call sites such as
config_command and event_dispatcher reach for a group by its position in
the map rather than by its capture key.
|
|
preg_split2()'s limit was always -1 and its flags were always either 0 or
PREG_SPLIT_DELIM_CAPTURE alone, so both arguments are gone: the shim now
exposes preg_split() and preg_split_delim_capture() over a shared
preg_split_impl(). That leaves Preg::split()/split4() as bare
pass-throughs, so callers use the shim functions directly and the wrappers
are dropped along with the now-unreferenced PREG_SPLIT_* constants.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Callers had to build a temporary Vec<&str> at every call site to satisfy
the &[&str] parameter. Taking IntoIterator and yielding the matched items
lets them pass owned or borrowed strings directly.
The flags variant and PREG_GREP_INVERT go away with it: no caller passes
flags, and Composer's Preg::grep() has no such parameter either.
|
|
The assoc flag was always a literal at every call site, so the boolean
carried no information the function name could not. json_decode_assoc
and json_decode_obj make the resulting PhpMixed shape visible at the
call site.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
parse_url now returns Option<UrlComponents> instead of a PhpMixed array,
and the component-selecting overload with the PHP_URL_* constants is gone:
callers read the field they want. Two call sites change behaviour as a
result, both towards PHP:
* CurlDownloader::handle_redirect tested scheme and host with is_null(), so
an unparsable Location header (PhpMixed::Bool(false)) counted as an
absolute URL. PHP's truthiness test sends it to the relative-path branch.
* Url::get_origin appended a literal port 0, which PHP treats as falsy and
leaves off.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Porting mapped every PHP `protected` member onto `pub(crate)`, which is
wider than nearly all of them need. Each item demoted here is reached
only from the module that defines it, so the crate-wide visibility
conveyed nothing.
Every `pub(crate)` that survives has at least one reader in another
module of the same crate.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
The binary called itself Composer everywhere: the application name, the
logo, --version, about, and every warning that talks about the running
program. Prompts to file a bug also pointed at Composer's issue tracker.
Add SHIRABE_VERSION and SHIRABE_RELEASE_DATE next to the Composer version
constants and report those, naming the Composer version this port tracks
alongside them. Composer::VERSION and getVersion() are untouched, so the
composer platform package, composer-runtime-api and the HTTP User-Agent
keep the value plugins and package repositories expect.
build.rs stamps the release date with the UTC date of the HEAD commit,
the way Composer's Compiler fills in @release_date@ when building the
phar. It now also fails the build when git cannot be read, instead of
letting COMPOSER_DEV_WARNING_TIME fall back to the tagged-release value
and suppress the outdated-build warning forever.
Messages about the Composer ecosystem keep their wording. Two of them are
pinned by upstream installer fixtures (Rule's "cannot be modified by
Composer" and SolverProblemsException's "you can run Composer with") and
stay as they are so those fixtures can keep being used verbatim.
The e2e list comparison against upstream Composer now skips the banner,
which cannot match by design, and compares everything below it as before.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
`http_build_query(['packages' => array_keys($packageConstraintMap)])` maps a
single array value, so the body PHP sends is `packages%5B0%5D=a&packages%5B1%5D=b`.
The port flattened the array into repeated `packages=<name>` pairs, losing the
indices, and passed `"&"` as the numeric prefix and `"="` as the argument
separator, joining every pair with `=`.
Build the query from a `packages` key holding the name list and hand it to
`http_build_query_mixed`, which walks the nested value.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Move `Composer\Pcre` out of shirabe-external-packages and into its own
crate, so the path is `shirabe_pcre::preg::Preg` instead of
`shirabe_external_packages::composer::pcre::preg::Preg`.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Composer hands the exception it caught to the one it throws in its place,
so `getPrevious()` reaches the cause and Application's renderer prints the
whole chain. Every ported site dropped it, because the flat exception
structs had nowhere to put one. `AnyThrowable::into_previous` turns the
caught error into that argument, and the 13 sites now pass it.
`getCode()` came along for the ride at the four sites that derive the new
exception's code from the caught one (PharArchiver, ArrayLoader x2), and
ComposerRepository's message now names the caught exception's class
instead of the literal "Exception".
GitHubDriver::attemptCloneFallback took the previous exception's message
and appended it to its own, which no `\RuntimeException('Fallback to git
driver disabled')` in Composer ever says; it now chains it instead.
Git::syncMirror restores what PHP's `finally` does to an exception in
flight: the `git remote set-url` that scrubs credentials back out of the
URL runs in a `finally`, and when it fails PHP propagates *its* exception
over the one already leaving, chaining the displaced one as previous. The
port discarded the finally's result, so a failure to scrub the URL was
reported as a successful mirror sync. `AnyThrowable::set_previous`
models the engine-level chaining.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Ported exceptions were flat structs reached with `downcast_ref`, so
Composer's `catch (\RuntimeException $e)` only matched the exact leaf
type and `get_class($e)` had nothing to report. Each exception now
embeds an instance of the class it extends and travels inside an
`AnyThrowable`; `Catch::catch`/`catch_mut` walk that chain, and
`PhpClass::php_class_name` yields the PHP FQCN.
Dropping the `std::error::Error` impls from the exception types leaves
`AnyThrowable` as the only route into an `anyhow::Error`, so the walk
cannot be bypassed. A `no_exception_downcast` linter catches the
`downcast::<X>()` calls that would now silently answer `None`.
Three sites change behavior as a result: the `TransportException`
exit-code override reaches `MaxFileSizeExceededException`, the
`catch (\LogicException)` in findSimilar() reaches its subclasses, and
rendered exception titles carry the real class name rather than a
guess. `get_class_err()` is no longer a `todo!()`, which re-enables
FilesystemRepositoryTest::testCorruptedRepositoryFile.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
PHP's version_compare takes its operator as a string, so the shim's port
did too, and Constraint carried two families of operator constants plus
translation tables to convert between the string form and its own int
codes. Five copies of those tables had accumulated across Constraint,
CompilingMatcher and the plugin value bridge.
version_compare now takes a CmpOp, which makes an invalid operator
unrepresentable and removes the tables' reason to exist. Constraint
stores a CmpOp and keeps only the string parsing its constructor needs;
getOperator, compile and CompilingMatcher::match speak CmpOp as well.
PHP's OP_* numbering stays observable: a plugin reads the raw integer off
the Constraint object over RPC, so get_operator_constant and its new
inverse hold that 0..5 mapping.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Call sites whose haystack was an inline array of literals (or a local
built solely to feed one) had to wrap both sides in PhpMixed just to
compare, allocating a String per element on every call. matches! does
the same test against the underlying &str/i64/Option directly, so the
PhpMixed round trip and its .to_string()/.clone()/.iter().map()
conversions are gone.
Sites whose haystack is a runtime value or a named constant array are
left on in_array_strict: inlining a named constant would duplicate its
contents at the call site.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
|
|
Three fixes for the ComposerRepository/FilesystemRepository/PlatformRepository
hazards where inner-composition delegation skipped PHP's late-bound virtual
dispatch:
- ComposerRepository::has_package now builds its packageMap through the
late-bound getPackages() equivalent, so lazy-providers repos surface the
LogicException and available-packages repos load their package list, as in
PHP, instead of silently answering false from the raw array.
- RepositoryInterface::get_repo_name returns anyhow::Result<String>: PHP's
getRepoName() counts through the late-bound initialize(), which is fallible
in file-reading subclasses. FilesystemRepository and PackageRepository now
run that initialization instead of freezing the inner array repository to an
empty state (which also made a later write() truncate installed.json).
Supporting changes keep the initialization chain callable from &self:
JsonFile::read takes &self (indent moved into a RefCell), FilesystemRepository
dev_mode became a Cell, and WritableArrayRepository dev_package_names a
RefCell.
- PlatformRepository::new routes constructor packages through its own
add_package so the override handling and full platform initialization run
as they do via PHP's parent constructor; the inner find_package/add_package
delegations inside add_package (and ComposerRepository::add_package) gained
the same is_initialized guard, since the constructor path would otherwise
freeze the repository.
Same defect class as 7db937af, 97b5211a and 3e367f78.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
fallthroughs
findPackage()/findPackages()/search() delegate their non-lazy, non-provider
fallthrough to the inner ArrayRepository, whose self-initialization froze the
packages array before ComposerRepository::initialize could read the root file,
so a plain v1-style repo (inline "packages" in packages.json) always answered
empty. Guard the delegations with the same is_initialized() check used by
count()/hasPackage().
getProviders() had the inverse defect: PHP reads the raw $this->packages
property, but the port went through count(), whose initialize poisoned the
initialization flag so the root file would never load afterwards. Check the
raw field for non-empty instead, matching PHP's truthiness test.
Same defect class as 97b5211a and the 3e367f78 downloader fix.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
PHP's JsonFile::encode throws a RuntimeException when json_encode fails; the
port swallowed that into an .unwrap() marked TODO(phase-c). Return
anyhow::Result from encode/encode_with_options and propagate at every call
site (print_table and list_repositories become Result-returning to carry it).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
Retag every Shirabe-authored TODO comment to one of the fixed tags:
phase-c, phase-d, plugin, php-runtime, phase-e.
Upstream-authored TODO comments from Composer/Symfony are left
untouched to preserve the ported code shape.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
PHP's ArrayRepository::count()/hasPackage() call $this->initialize(),
which late-binds to the concrete repository class and lazily loads its
packages. The Rust pass-throughs skipped that: they ran ArrayRepository's
stub initialize instead, returning 0/false and marking the repository
initialized with an empty package list, which made ensure_initialized()
skip the real initialization forever after.
Take &mut self in RepositoryInterface::count/has_package so the lazy
repositories (Filesystem, Platform, Composer) can guard with their real
initialize, and return Result from has_package since that initialization
can fail (PHP propagates the exception). InstallerInterface::is_installed
and InstallationManager::is_package_installed/mark_alias_installed
propagate the same way, which also resolves the TODO(phase-d) markers on
Package/Path/Artifact/Vcs repositories about initialization errors being
swallowed.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
PHP's SecurityAdvisoryPoolFilter stores advisory *object references* in
$securityRemovedVersions, and PoolOptimizer::applyRemovalsToPool hands
that array to the new Pool by copy-on-write. Porting AnySecurityAdvisory
as a value type turned both of those into deep copies.
Measured on `require laravel/framework` (offline, warm cache): 113
distinct advisories were duplicated into 314,309 copies of ~1.08 KiB,
retaining 331.9 MiB in the filter loop and another 331.9 MiB when
apply_removals_to_pool cloned the whole map.
3.54s -> 2.62s (-26%), peak RSS 975 MB -> 343 MB, which matches the upper
bound measured by ablation. Composer runs the same workload in 1.49s.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
|
|
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
For a simple, non-lazy composer repository (no providers-url/
metadata-url/providers-lazy-url), load_packages() delegates to
self.inner.load_packages() (PHP: parent::loadPackages()). PHP's
ArrayRepository::loadPackages() calls $this->getPackages(), which
virtual-dispatches through ComposerRepository::getPackages() down to
ComposerRepository::initialize() (the real HTTP/file fetch).
Composition doesn't get that dispatch for free: self.inner is a bare
ArrayRepository, so self.inner.load_packages() ends up calling
ArrayRepository::initialize() — a no-op stub that just sets an empty
package list — instead of ComposerRepository::initialize(). The
repository's packages.json fetched and parsed successfully, but the
result was silently discarded, so the repo always looked empty.
get_packages() already carries the identical guard with the same
diagnosis in its own comment; load_packages() was just missing it.
Documented the same latent gap in find_package/find_packages/count/
has_package, which call into ArrayRepository the same unguarded way
but aren't known to be exercised by any test yet.
|
|
advisories
get_security_advisories cloned each package's full metadata blob
(hundreds of versions for popular packages) just to peek at its
"security-advisories" field. Same pattern already fixed in
load_async_packages by 2862d2da; move ownership through pattern
matches and IndexMap::shift_remove instead.
Measured with laravel/laravel create-project: response processing in
the security-advisories metadata branch drops ~80% (236ms -> 48ms),
and the whole run_security_advisory_filter phase drops from ~420-500ms
to ~250-270ms. End-to-end, shirabe now matches or edges out upstream
Composer (6.57s vs 6.85s in this run) instead of trailing by ~1.2s.
|
|
metadata
load_async_packages cloned each response's version metadata up to
three times (spec_list/response, response_arr, versions_mixed, then
every per-version field) before building the versions Vec. Move
ownership through pattern matches and IndexMap::shift_remove instead.
Measured with laravel/laravel create-project: per-batch response
processing time in load_async_packages drops ~39% (1.30s -> 0.80s
cumulative), narrowing the E2E gap vs upstream Composer from 1.37s to
1.14s on average.
|
|
abandoned propagation
The json format branch ignored search results entirely and always wrote
null. Encode results into the same name/description/abandoned/url shape
Composer's array-backed repositories produce. Also fix
ComposerRepository's RepositoryInterface::search adapter, which dropped
the abandoned field from raw API results even when present.
|
|
regex::Regex::clone() does not share the underlying meta engine's
search-cache pool, so every fresh clone pays a ~10us warmup cost on
its first use. Two changes together eliminate this across nearly all
preg_* call sites:
- A php_regex! macro resolves PHP-style patterns to a per-call-site
&'static regex::Regex (via regex-macro's LazyLock), applied at the
majority of call sites throughout the codebase.
- Call sites still passing dynamic pattern strings go through
PATTERN_CACHE, which now stores Arc<(Regex, bool)> and hands out
Arc::clone()s instead of cloning the Regex itself.
PregPattern::resolve() returns a ResolvedPattern enum (Arc or
'static reference) rather than an owned Regex, so neither path ever
clones the Regex proper.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
|
|
security-advisories in a p2 provider response is a JSON array, which
decodes to PhpMixed::List rather than PhpMixed::Array. The check only
matched Array, so it always missed and silently fell through to the
api-url fallback instead of using the already-cached provider data.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
|
|
Replace sync_executor::block_on's reactor-less busy-spin poller with
tokio::task::block_in_place + Handle::current().block_on(), riding a
single tokio Runtime entered once in main.rs (falling back to a
disposable one when no ambient runtime exists, e.g. in tests). This
lets HttpDownloader::dispatch await CurlDownloader::download directly
instead of bouncing through the separate curl_runtime() bridge, which
is now deleted.
Manual create-project verification against the real network caught a
concurrency bug this exposed: async_fetch_file held http_downloader's
RefMut across the await on add(), which only panics once downloads
genuinely overlap. add() only needs &self, so borrow() fixes it.
With everything now sharing one real reactor, the FuturesOrdered
fan-out added for ComposerRepository::get_security_advisories/
load_async_packages finally overlaps for real: fetching 8 packages'
metadata dropped from ~7-40s to a consistent ~3-4s in a before/after
comparison, with identical resulting lock files.
sync_executor::block_on's call sites are still synchronous rather
than async fn propagated up to Command::execute, which remains the
end goal (see the TODO(phase-e) in sync_executor.rs) - nested block_on
calls elsewhere don't get this same overlap, only prevented panics.
|
|
get_security_advisories/load_async_packages serialized every
start_cached_async_download call through sync_executor::block_on
per name, matching PHP's structure but not its promise-based
concurrency. Wrap the mutable state those downloads touch (cache,
fresh_metadata_urls, packages_not_found_cache, degraded_mode) in
RefCell/Cell so start_cached_async_download and its async_fetch_file
helpers can drop to &self, then fan out all downloads for a batch via
FuturesOrdered before processing responses sequentially in original
order, mirroring PHP's build-all-promises-then-wait shape.
Real I/O overlap still awaits item 7 (HttpDownloader::add currently
resolves through the curl_runtime()/sync_executor::block_on bridge
either way), so this is architectural groundwork, not a perf win yet.
|
|
load_async_packages (the v2 metadata-url/packagist protocol path)
called a separate create_packages_static helper instead of the
instance method create_packages. PHP has a single createPackages
method used everywhere, so this duplicate silently skipped the
notification-url injection (and dist-mirror/transport-options setup)
that create_packages performs. Every package resolved via the
lazy provider path ended up missing notification-url in
composer.lock/installed.json. Removed the now-dead duplicate.
|
|
The CurlDownloader owned a tokio runtime and block_on'd reqwest from its
sync tick(), while the repository/installer/downloader sync bridges each
created another Runtime and block_on'd async fns that reach that leaf.
Driving one Runtime::block_on from within another panics with "Cannot
start a runtime from within a runtime", hit by `require` when fetching
p2 metadata.
Switch CurlDownloader to a blocking reqwest client (its own internal
thread, never nested) and replace the per-call Runtime::new().block_on
bridges with a no-reactor sync_executor::block_on helper. No awaited
future parks on a reactor once the only async I/O is blocking, so the
helper can be nested freely.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
|
|
|
|
|
Port perforce (36), locker (10), composer_repository (7), installation_manager
(6), file_downloader (5), and event_dispatcher (6) tests via the mock infra.
Fix production porting bugs surfaced en route: BufferIO::get_output look-behind
regex, ComposerRepository list-form package iteration and initialize dispatch,
gethostname and spl_autoload_functions shims; add EventDispatcher get_listeners
test seam.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
|
composer::semver stubs
Flatten shirabe-semver's modules into glob re-exports at the crate
root and route all consumers through the short paths. Remove the
duplicate composer::semver stubs from shirabe-external-packages in
favor of the shirabe-semver types.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
|
Move MetadataMinifier and SpdxLicenses out of shirabe-external-packages
into dedicated shirabe-metadata-minifier and shirabe-spdx-licenses
crates, updating all import sites accordingly.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
|
The List and Array variants of PhpMixed boxed their elements
unnecessarily. Store PhpMixed values directly and update all callers
accordingly.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
|
