Skip to content

EGCL v0.0.10

Choose a tag to compare

@github-actions github-actions released this 20 Oct 00:56
· 446 commits to main since this release
a242963

0.0.10 - 2026-10-19

Added

  • install-egcl-forks also installs the EGCL forks of Ironclad, pure-tls and chipz, which HTTPS through pure-tls and archive extraction with chipz need (#429).

  • EGCL-MOP:FORWARD-REFERENCED-CLASS, the metaclass of a class named as a superclass before it is defined. CLASS-OF answers it and such a class prints as #<FORWARD-REFERENCED-CLASS FOO>, as on SBCL; it is a CLASS but not a STANDARD-CLASS, since nothing has yet said what its instances look like, and it becomes an ordinary class -- the same object, so references already taken to it stay EQ -- when the definition arrives (#417).

  • EGCL-MOP, the metaobject-protocol package docs/egcl-lisp-api.md specified but the bootstrap never installed. It provides CLASS-DIRECT-SUPERCLASSES, CLASS-DIRECT-SUBCLASSES, GENERIC-FUNCTION-METHODS, CLASS-FINALIZED-P, FINALIZE-INHERITANCE, INTERN-EQL-SPECIALIZER and VALIDATE-SUPERCLASS, and is the home of CLASS-SLOTS, CLASS-DIRECT-SLOTS and the SLOT-DEFINITION-* readers. COMMON-LISP-USER uses it, so those names stay reachable unqualified where a person is typing (#380; issue #371).

  • DEFCLASS's :METACLASS takes effect: CLASS-OF answers the named metaclass instead of STANDARD-CLASS. The option was parsed and discarded, so a kernel that tells its own classes from plain ones with (eql (class-of (find-class name)) (find-class 'its-metaclass)) -- Gendl's does -- saw every class as plain (#380; issue #371).

  • Package-local nicknames (CDR 10). DEFPACKAGE accepted :LOCAL-NICKNAMES and dropped it, so a library's (:local-nicknames (:a :alexandria)) left a:define-constant unreadable: "package not found: A". The option now takes effect, the reader, FIND-PACKAGE, INTERN and FIND-SYMBOL resolve a prefix through the current package's local nicknames before the global names, EGCL-EXT exports ADD-PACKAGE-LOCAL-NICKNAME, REMOVE-PACKAGE-LOCAL-NICKNAME, PACKAGE-LOCAL-NICKNAMES and PACKAGE-LOCALLY-NICKNAMED-BY-LIST as the API document already listed them, :PACKAGE-LOCAL-NICKNAMES is on *FEATURES*, and UIOP's DEFINE-PACKAGE finds the functions in EGCL-EXT (#372).

  • Socket endpoint queries: (egcl::%socket-listener-local-name listener), (egcl::%socket-local-name stream) and (egcl::%socket-peer-name stream) report an endpoint as an octet vector and a port together, four elements for IPv4 and sixteen for IPv6. A library adapter could previously learn a listener's port but never any address, so usocket's get-local-address, get-peer-address, get-peer-port and get-local-name had no primitive behind them and AllegroServe could not log or print a socket (#352, #366; issue #350).

  • egcl-ext:chdir and egcl-ext:delete-directory, and through them uiop:chdir, uiop:delete-empty-directory and uiop:delete-directory-tree, which signalled "Not (currently) implemented on EGCL" (#392).

  • The manual now documents every EGCL-EXT function, including SETENV, UNSETENV, GET-PRECISE-TIME, REAL-TIME-NANOSECONDS, the package-local nickname functions, HASH-TABLE-WEAKNESS, the funcallable-instance functions, declaration introspection and the embedded-file functions. Each one has an anchored entry and a symbol-index row (#409).

Fixed

  • A thread waiting in SLEEP no longer holds up a garbage collection. A collection needed by another thread used to wait on the sleeper and give up, so a server whose main thread parks in (loop (sleep 60)) could abort a worker with "memory allocation of 16 bytes failed" (#434).
  • LOAD and COMPILE-FILE bind *PACKAGE* for the extent of the file, so a file that sets *PACKAGE* and then starts a thread stays in its own package. The file's package used to live in the global value every thread shares, and the loader could read its next form in another package, then report "undefined function" for a function the file had just defined (#433).
  • A MAKE-PACKAGE name clash or a DELETE-PACKAGE of a missing package signals a SIMPLE-PACKAGE-ERROR whose report says what failed, such as "a package named FOO already exists". Both are still correctable with CONTINUE; until now they signalled a bare PACKAGE-ERROR with no message (#436).
  • SLEEP inside a fiber suspends only that fiber, so other fibers on the same carrier keep running. It used to block the carrier thread until it woke. In a pinned fiber it follows EGCL-FIBER:*PINNED-BLOCKING-ACTION*, like the other blocking operations (#437).
  • A macro defined inside a LET keeps its captured binding across a saved image, as a DEFUN or DEFMETHOD in the same LET already did. Only a macro's name, lambda list and body reached the image, and every restored expander was rebound to the fresh global environment, so an expansion that mentioned a captured variable signalled UNBOUND-VARIABLE. anaphora defines SYMBOLIC and ANAPHORIC inside (with-unique-names ...), so in any saved core every anaphoric macro failed with "the variable ANAPHORA::S-INDICATOR is unbound". A macro now records which captured frame it closed over, and restored definitions share one frame rather than each getting a copy, so a binding several of them see stays shared (#431; issue #426).
  • A failed socket call carries its operating system error number, so usocket reports a port in use as ADDRESS-IN-USE-ERROR and a refused connection as CONNECTION-REFUSED-ERROR. Both used to arrive as UNKNOWN-ERROR with errno 0, and a port probe that handled only ADDRESS-IN-USE-ERROR retried the same busy port forever. Socket failures signal the new EGCL-EXT:SOCKET-OS-ERROR, which is still a FILE-ERROR, and EGCL-EXT:SOCKET-OS-ERROR-ERRNO reads the number. The usocket fork also accepts listeners with the default character element type (#432; issue #427).
  • A generic function dispatches an interpreted lambda or closure on the FUNCTION class. CLASS-OF, FUNCTIONP and TYPEP all answered FUNCTION, but dispatch disagreed with them in both tiers, so (brt (lambda (y) y)) chose a LIST method over a FUNCTION one and a LIST-only generic accepted a lambda outright. A compiled lambda and #'car were already correct (#423).
  • SETQ of a parameter whose name is special reaches the callees that read it. Such a parameter is bound in the frame and then made dynamic, and the assignment path tried the frame before the value cell, so it updated a stale frame copy: after (defun f (&key ((:indent *ind*) *ind*)) (setq *ind* 0) (reader)), (f :indent t) answered the old value instead of 0 (#425).
  • A class metaobject prints its own name: #<STANDARD-CLASS UC> and #<STRUCTURE-CLASS ST>, as on SBCL, CCL and ECL. Once a class became an object it printed through the generic instance path, so every standard class in the image printed as the identical #<STANDARD-CLASS> -- useless in a backtrace or a slot dump (#402).
  • A class restored from an image keeps the slot shadowing its :allocation :class declaration implies. The names a class declares class allocated were not written to the core, so a class finalized for the first time after a restore -- which is what subclassing a restored class does -- rebuilt its layout without them, and an inherited instance slot of the same name wrongly entered that layout. CLST format version 4 records the field (#393).
  • INTERRUPT-THREAD interrupts a running thread. It posted a value nothing ever observed, read only after the entry function returned, so rather than interrupting a thread it overrode the return value of one already finishing. It now raises the pending signal taken at every tier's safepoints, and sends a directed SIGUSR1 so a thread parked in a syscall wakes; a wait with no deadline no longer parks forever through the signal (#373).
  • FILE-POSITION works on a Gray stream. It signalled a bare STREAM-ERROR, so cl-tar could not extract any archive and ocicl could not install a package. The Gray protocol gains STREAM-FILE-POSITION, as in SBCL, which FILE-POSITION calls and EGCL-GRAY-STREAMS exports (#428).
  • A compiled method shows up as (METHOD <generic-function>), with its qualifier, in EGCL_SPROF profiles, the tiering log and backtraces, instead of an anonymous METHOD shared by every method (#421).
  • A form whose value is only tested no longer leaks its other values into the result. (if (values x 99) x (list x)) returned X and 99 in compiled code, and a compiled UNLESS could even return the test's value instead of NIL; IF, WHEN, UNLESS, COND and non-final AND/OR forms now pass on only the primary value, as in SBCL (#419).
  • A fill-pointer or adjustable string made with :element-type 'base-char is a BASE-STRING: ARRAY-ELEMENT-TYPE answers BASE-CHAR, it is stored 8-bit, and VECTOR-PUSH, VECTOR-PUSH-EXTEND and ADJUST-ARRAY signal TYPE-ERROR for a wider character instead of storing it, as in SBCL (#418).
  • A class redefinition or CHANGE-CLASS no longer moves the instance it updates, so an object keeps its identity: two references to it stay EQ, an EQ hash table still finds it, and SLOT-MAKUNBOUND returns the object the caller passed in. Instances hold their slots indirectly, as in SBCL, and an update replaces that storage rather than the object. A metaclass that gains a slot now gives its existing classes that slot without disturbing them, which previously corrupted the class registry (#414).
  • BASE-CHAR, STANDARD-CHAR and EXTENDED-CHAR are proper subtypes of CHARACTER at more doors: (coerce #\λ 'base-char) and (coerce #\a 'extended-char) signal TYPE-ERROR instead of returning the character, the interpreter's (the base-char #\λ) signals one, and an EXTENDED-CHAR array is a CHARACTER string. COERCE to a string signals TYPE-ERROR for an element that is not a character instead of dropping it, and MAKE-ARRAY rejects :initial-contents whose length is not the vector's instead of padding it (#416).
  • String literals are CHARACTER strings, as in SBCL, so a copy of one can hold any character: (setf (char (copy-seq "abc") 0) #\λ) works, and a literal is the same kind of string whether its file was compiled or not. EGCL-EXT:READTABLE-BASE-CHAR-PREFERENCE (:SYMBOLS, :STRINGS, :BOTH or NIL, as in SBCL) can make a readtable read compact 8-bit literals instead (#411).
  • Several threads may read one hash table at once, as in SBCL; only writes need exclusive access. Concurrent lookups in a table keyed by conses or other objects hashed by address used to crash once a garbage collection moved its keys (#400).
  • Threads running interpreted code no longer crash while another thread collects garbage. A :synchronized t hash table now takes its lock on every operation as documented; it never did, and the interpreter's shared macro-expansion cache is such a table (#398).
  • A class is an object rather than its own name symbol, so (symbolp (find-class 'foo)) is false and DEFCLASS returns the class as CLHS requires. A class with a user :METACLASS is now an instance of it: TYPEP holds, a method specialized on the metaclass applies and is preferred over one on STANDARD-CLASS, and the class carries the metaclass's slots (issue #384, #399).
  • A slot added by redefining a class takes its :INITFORM on instances that already exist, as CLHS 4.3.6 requires: (defclass c () ()), an instance, then (defclass c () ((s :initform 42))) left S unbound where it now answers 42. UPDATE-INSTANCE-FOR-REDEFINED-CLASS is called for the update, so a method specialized on it participates and receives the added slots, the discarded slots and their values (#412, #408).
  • A class that is only a forward reference -- named as a superclass and not yet defined -- can no longer be instantiated, as in SBCL: MAKE-INSTANCE, ALLOCATE-INSTANCE and CHANGE-CLASS each used to return an object with no slots, built against a layout the eventual DEFCLASS then contradicts. EGCL-MOP:CLASS-FINALIZED-P also stops claiming T for such a class, or for a class that inherits from one, and EGCL-MOP:FINALIZE-INHERITANCE signals for them instead of silently doing nothing (#406).
  • A class named as a superclass before it is defined, and then defined with a :METACLASS that has slots, is an instance of that metaclass like any other. CLASS-OF answered the metaclass while TYPEP, SLOT-EXISTS-P, SLOT-VALUE, the metaclass's slot accessors and method dispatch all still saw a plain STANDARD-CLASS. A kernel that appends a late-defined mixin to every class -- Gendl's does -- hit this on exactly the classes that mattered (#404; issue #384).
  • An inherited slot is reachable when the superclass is defined after the subclass: (defclass sub (super) ...) followed by (defclass super () ((a))) left A unusable on an instance of SUB. Redefining a class now updates its subclasses' instance layouts, and no longer discards its record of them (#399).
  • TYPE-OF answers the class itself when the class has no proper name, rather than a name that no longer finds it (#399).
  • A built-in class is an object rather than an integer. (find-class 'integer) was the fixnum -14, so (eql -14 (find-class 'integer)) and (integerp (find-class 'standard-class)) were both true, (class-name -14) answered INTEGER, a method specialized on BUILT-IN-CLASS applied to the integer -14, and a class printed as -40 (#396).
  • A readtable built from the standard syntax with GET-MACRO-CHARACTER and SET-MACRO-CHARACTER, as named-readtables' (:merge :standard) does, reads the forms inside a list with that readtable. It used the standard one, so a pythonic """…""" string inside a form broke apart (40ants-doc). LOAD also binds *READTABLE*, so a loaded file's IN-READTABLE no longer leaks into the code that loaded it (#392).
  • A constant one library defines while it compiles no longer becomes the symbol another package reads for that name. ironclad's and flexi-streams' +BUFFER-SIZE+ turned up in drakma and salza2, and salza2's DEFPARAMETER of it failed as an assignment to a constant (#392).
  • COMPILE-FILE no longer panics on data shaped like a DEFCONSTANT form, such as the CASE key ((defconstant) …). A literal string with a fill pointer, such as cl-interpol's #?"…" reads, compiles into the fasl instead of making the whole file source-only, and the read-time #. forms of a file that is still source-only see its source as *COMPILE-FILE-PATHNAME* (#392).
  • A saved image keeps every compiled function. One whose code referred to a constant the compiled-unit format could not hold, such as a package, a hash table, a host-built closure or a HANDLER-BIND handler, was left out with a warning and ran interpreted after a restart; an image of ocicl lost several hundred. Compiled files now hold package and hash-table constants too (#392).
  • BASE-CHAR is the element type of an 8-bit base string, consistently. TYPEP called every character a BASE-CHAR while BASE-STRING matched only 8-bit storage, MAKE-STRING ignored :element-type, and MAP, COERCE and MAKE-SEQUENCE built 32-bit strings for BASE-STRING, so (make-array 5 :element-type 'base-char) and (map 'base-string ...) answered strings TYPEP refused as base strings, and TYPE-OF called every string SIMPLE-BASE-STRING. BASE-CHAR is now the code points below 256 and EXTENDED-CHAR the rest, UPGRADED-ARRAY-ELEMENT-TYPE and ARRAY-ELEMENT-TYPE answer BASE-CHAR for 8-bit strings (a string displaced onto one included), the constructors honor the requested width (a wide character is a TYPE-ERROR), COPY-SEQ, SUBSEQ, REVERSE and the REMOVE, SUBSTITUTE and REMOVE-DUPLICATES families keep a string's width, and TYPE-OF answers (SIMPLE-ARRAY CHARACTER (n)) for a 32-bit string (ansi sequences MAP.SPECIALIZED-STRING.3, LENGTH.STRING.6; the sequences chapter now passes completely and joins the gate) (#403).
  • Hash tables find every key their test calls equal. An EQUALP table missed 1d0 for 1, 0.5 for 1/2, an adjustable or fill-pointer vector or string for a simple one, a string for a vector of its characters, and equal structures and hash tables; EQUAL tables missed fill-pointer strings and lost an adjustable-vector key once an element changed. EQL (and EQL/EQUAL tables) no longer treat -0d0 as 0d0, nor EQUAL -0.0 as 0.0, and SXHASH no longer changes when the garbage collector moves an object (#388).
  • A generic function dispatched on a class object no longer answers whatever the first such call decided. With methods on STANDARD-CLASS and STRUCTURE-CLASS, whichever class was passed first fixed the answer for every class passed after it, and a method on BUILT-IN-CLASS traded decisions with one on INTEGER the same way (#389).
  • A saved image no longer lets newly compiled code redefine the local functions of code it saved. Loading the ocicl runtime in the make image egcl replaced one inside PARSE-INTEGER, after which (parse-integer "1") signalled "not an integer" and ASDF could not load babel (#386).
  • RESTART is consistently a standard class. Its class precedence list has always been (RESTART STANDARD-OBJECT T), but (class-of (find-class 'restart)) answered BUILT-IN-CLASS and (subtypep 'restart 'standard-object) answered false. All three agree now (#387).
  • MAKE-PATHNAME no longer builds a pathname with a wrong component when a garbage collection happens while it runs: with a :directory list or a namestring :defaults, a freshly made :name or :type string could be replaced by unrelated data, such as the directory's namestring (#382).
  • A library's own DEFUN of CLASS-PRECEDENCE-LIST no longer redefines the implementation's. EGCL-MOP owns and exports the name now, and the interpreter dispatches that spelling, so a bare read of it inside a package that does not inherit EGCL-MOP interns that package's own symbol instead of the implementation's. With CLASS-SLOTS in the previous release this completes #370 (#383).
  • FORMAT's ~R prints bignums in English: (format nil "~R" (expt 10 20)) signalled a TYPE-ERROR saying its argument was not an integer. It also names every short-scale group through vigintillion, so 10^9 is "one billion", not "one thousand million" (#381).
  • ~D, ~B, ~O, ~X and ~R keep a pad or comma character outside Latin-1: it was cut to its low byte, so #\★ padded with a control character (#381).
  • NTH-VALUE and DECODE-UNIVERSAL-TIME signal a TYPE-ERROR for an argument that is not a non-negative integer instead of converting it through a float: a compiled (nth-value -1 ...) answered the first value, and (decode-universal-time 1.5) decoded a time (#381).
  • A library's own DEFUN of CLASS-SLOTS no longer redefines the implementation's. The non-ANSI MOP readers were reachable as a bare name from any package, so an internal helper of that name in a package that exports nothing hijacked the implementation's function, process-wide and silently. They live in EGCL-MOP now. CLASS-PRECEDENCE-LIST and METHOD-SPECIALIZERS are unchanged: they are name-dispatched builtins whose dispatch ignores the package, so moving the symbol does not stop the override (#380; issue #370).
  • A package designator that names no package now signals PACKAGE-ERROR instead of creating the package. INTERN, EXPORT, IMPORT, SHADOW and USE-PACKAGE created it; FIND-SYMBOL, UNUSE-PACKAGE and PACKAGE-USE-LIST answered as though the lookup had simply found nothing. So a typo in a package name, or a library loaded before the package it names, left a live package of that name: FIND-PACKAGE then answered it, a later DEFPACKAGE of the real one merged into the mistake, and the fault surfaced far from where it was made. DEFPACKAGE still resolves a forward :use reference, which is what the create-on-miss path exists for (#379; issue #375).
  • OPEN and WITH-OPEN-FILE honor :external-format :latin-1 (also :iso-8859-1) and :ascii, alone or as the first element of a list such as (:latin-1 :eol-style :lf); :default and :utf-8 remain UTF-8. The keyword was accepted and ignored, so a Latin-1 file was decoded as UTF-8 and the first accented letter was a STREAM-ERROR: cl-pdf opens its font metrics and cl-typesetting its French hyphenation patterns that way, and cl-typesetting could not load. Reading decodes one octet per character, writing refuses a character outside the format's range, and STREAM-EXTERNAL-FORMAT answers the format a file stream was opened with instead of :default. A format the stream layer cannot decode (:utf-16, :utf-32) is refused at OPEN rather than read as UTF-8 (#374).
  • LISTEN on a freshly opened octet file stream answers T while bytes remain and NIL at the end of the file. With nothing buffered yet it asked the handle as if it were a socket, and a regular file always answered NIL, so a (unsigned-byte 8) file stream had to be read once before LISTEN saw it (#369). The file-stream case joins the socket test of #360.
  • A thread waiting for input on a stream no longer locks every other thread out of it. Input parked while holding the stream's state mutex, so for as long as one thread waited for a peer, any other operation on that stream blocked -- CLOSE, a write on the other half of a socket, even INPUT-STREAM-P, which performs no I/O. A server could not close a worker's idle keep-alive connection, and on a bidirectional socket it could not write a reply while another thread read. Readers are now serialized on a separate lock and wait with the state mutex released (#353; issue #350).
  • The same fix now covers composite streams. A TWO-WAY, ECHO, CONCATENATED or SYNONYM stream reached its component while still holding its own lock, so a read parked on the component locked every other thread out of the composite -- and *TERMINAL-IO* is bidirectional (#354; issue #350).
  • Character reads and process-pipe reads use the same non-parking path as byte reads, so a read that waits for input holds no stream mutex. Regular file I/O is deliberately excluded: a readiness poll always reports a regular descriptor ready, so polling cannot make a disk read cooperative, which is also where the JVM leaves file I/O for its virtual threads (#365; issue #350).
  • READ-SEQUENCE on a socket no longer deadlocks other threads using that stream. It was the last read that still waited while holding the stream's state mutex, and the reachable one: socket streams are (unsigned-byte 8), so READ-SEQUENCE is how a library reads a socket. Any other thread touching the stream blocked -- CLOSE, a write on the other half, even STREAM-ELEMENT-TYPE, which performs no I/O -- and because the thread that would supply the bytes was usually the blocked one, the program deadlocked rather than merely stalling. READ-SEQUENCE now waits with the mutex released and resumes where it left off, so a request split across several arrivals still fills completely and a short count still means end of file (#367; issue #350).
  • COMPILE-FILE now gives the body of a load-time-only top-level EVAL-WHEN the top-level processing CLHS 3.2.3.1 requires. Without :COMPILE-TOPLEVEL but with :LOAD-TOPLEVEL, the body forms are still top-level forms, so a nested (eval-when (:compile-toplevel) ...), a DEFMACRO or a DEFPACKAGE among them keeps its compile-time effect. The whole form was evaluated instead, which answered NIL for the outer form and never reached the inner one: cl-store's DEFBACKEND expands to exactly that nest, so every later use in the file compiled as a function call and the fasl died loading with "STREAM is unbound" (#355).
  • GET-DISPATCH-MACRO-CHARACTER now reports a callable function for each standard # sub-character (CLHS 2.4.8) rather than NIL for all but #\. A reader that wraps the standard syntax -- zacl rewrites Allegro's (version>= ...) features and hands the rest back to the standard #+ reader -- died calling NIL (#356).
  • A control transfer out of a reader macro now passes through COMPILE-FILE's pre-read. A HANDLER-CASE clause around COMPILE-FILE taking a condition a reader macro signalled, or an invoked restart, was reported as a read failure and silently downgraded the output to a source-only artifact, so the clause never ran (#357).
  • The reader now accepts the character names SBCL and Clozure CL share: Ff, Formfeed, Lf, Cr, Bs, Ht, Sp, Del and Altmode. CHAR-NAME still answers the standard names, so printing round-trips are unchanged (#358).
  • The reader now closes the list when a Lisp reader-macro function returns no values immediately before the closing parenthesis, as it already did for
    #|...|#. The list previously lost its terminator
    (#359).
  • LISTEN no longer consumes the pending byte of an octet stream. It was
    READ-CHAR-NO-HANG followed by UNREAD-CHAR for every stream, which on an
    octet socket parked the byte in the character unread slot where READ-BYTE
    never looks: LISTEN answered T and the READ-BYTE after it hung or lost an
    octet outright. usocket's WAIT-FOR-INPUT calls LISTEN on every stream
    socket before its backend runs, so AllegroServe hung on its first read after
    every wait (#360).
  • FUNDAMENTAL-STREAM now sits under STREAM, so a method on STREAM applies
    to a Gray stream and (typep gray-stream 'stream) is true
    (#362).
  • DEFINE-SYMBOL-MACRO definitions now live in one process-wide table, so a
    symbol macro defined on one thread is visible on every other. Each thread
    previously kept its own, and a definition made before a thread started was
    invisible inside it
    (#363).

Performance

  • A function with &optional parameters binds them by position, filling in defaults in a short prologue, instead of through the general lambda-list binder on every call. Calls to such functions in pure-tls's certificate parsing and in SEARCH took about a third less time (#392).
  • A compiled function whose IF returns a call's result in one arm, such as a recursive (if (< n 2) (leaf n) (+ (f (- n 1)) (f (- n 2)))), now uses the direct native calling convention instead of going through Rust for every call: about 5 ns per call instead of about 700 ns (#420).
  • Generic-function calls do less repeated work. A DEFCLASS accessor reads or writes its slot directly once dispatch selects it (user methods on the accessor still take over), slot writes skip the class-slot search for ordinary slots, and per-generic and per-class facts are cached. From compiled callers an accessor read takes about 3 us instead of 11, a write 8 instead of 23, and READ-BYTE on a Gray stream 40 instead of about 110; a repeated HTTPS request through drakma and pure-tls is about twice as fast (#415).
  • A method body that loops now reaches optimized native code (T2) like the same body in a DEFUN. It used to stop at baseline native code (T1) until its 4096th call: a 1000-iteration loop took about 4 ms per call as a method and 30 us as a function, and now takes about 50 us as a method. Methods restored from a saved image are covered too (#413).
  • AES, GCM's GHASH multiply and SHA-256 compression run in the runtime, as EGCL-EXT:CRYPTO-* block primitives that EGCL's Ironclad fork calls in place of its portable Lisp. With the fork, encrypting 64 KB with AES-GCM takes seconds instead of 193 s, hashing it with SHA-256 takes under 0.3 s instead of 18 s, and a repeated HTTPS request through pure-tls is about 10x faster. EGCL-EXT:SHA256 and EGCL-EXT:CRC32 are now documented (#410).
  • Calls made from compiled (native) code into other compiled functions are cheaper: a call no longer looks FUNCALL/APPLY up by name, rebuilds the callee's per-call buffers, rescans its bytecode or registers a global root for it, and finding the current execution's state is quicker. A recursive native caller spends about 27% less per call (about 38% when it establishes a HANDLER-CASE), and a cached (asdf:load-system :babel) is about 5% faster (#390).
  • Compiled code finds its execution-local state with fewer checks per access, so a native call that establishes a HANDLER-CASE costs about 7% less again (#397).