X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=NEWS;h=283e97971f1cf96ee2870b2fd07865cb8e704842;hb=02313fc9ceecedacefc2e2a13c5c67a2ca9a57e8;hp=b69cc1611e07f621d2ce2c80db977c24d3e5521f;hpb=d6e6a3798dc767045cd7495f483bd4d0b6250d00;p=sbcl.git diff --git a/NEWS b/NEWS index b69cc16..283e979 100644 --- a/NEWS +++ b/NEWS @@ -1,20 +1,215 @@ ;;;; -*- coding: utf-8; fill-column: 78 -*- -changes relative to sbcl-1.0.53: +changes relative to sbcl-1.0.57: + * enchancement: implicit generic function warnings now specify the package + in which the new generic function is being created. + * enhancement: SB-EXT:ATOMIC-UPDATE makes it easy to perform non-destructive + updates of CAS-able places (similar to Clojure's swap!). + * bug fix: potential for infinite recursion during compilation of CLOS slot + typechecks when dependency graph had loops. (lp#1001799) + * bug fix: error forms reported with some program-errors were not escaped + properly. + * bug fix: functions from EVAL are now on more equal footing with functions + from COMPILE. (lp#1000783, lp#851170, lp#922408) + * bug fix: ENSURE-GENERIC-METHOD-COMBINATION accepts method combination + objects as its :METHOD-COMBINATION argument, not just lists designating + method combinations. (lp#936513) + +changes in sbcl-1.0.57 relative to sbcl-1.0.56: + * RANDOM enhancements and bug fixes: + ** bug fix: the range and distribution of random integers could be + catastrophically wrong when the compiler derived the type of its + argument as a disjoint set of small integers. + ** bug fix: the distribution of random integers is now completely + uniform even when the specified limit is not a power of two. + (Previously some values could be about 0.1 % more probable than + others in the worst case.) + ** RANDOM on large integer arguments is generally faster and conses + less than before; this is visible for fixnums above a length of + about 24 bits, but extremely so for bignums: the old implementation + used time and space quadratical in the size of the argument there, + the new one is linear. + * enhancement: redesigned protocol for quitting SBCL. SB-EXT:EXIT is the new + main entry point, SB-EXT:QUIT is deprecated. + * enhancement: additions to the SB-THREAD API: RETURN-FROM-THREAD, + ABORT-THREAD, MAIN-THREAD-P, and MAIN-THREAD. + * enhancement: FASL loading no longer grabs the world-lock. + * enhancement: GENCGC reclaims space more aggressively when objects being + allocated are a large fraction of the total available heap space. + (lp#936304) + * enhancement: backtraces show the correct number of arguments for frames + called with too many arguments. + * enhancement: support for abort(3), exit(3), and _exit(2) has been added to + SB-POSIX. + * enhancement: ASDF has been updated 2.21. + * optimization: fewer uses of full calls to signed modular functions. + (lp#903821) + * optimization: typechecking alien values is typically 5 x faster. + * optimization: FDEFINITION, SYMBOL-FUNCTION, MACRO-FUNCTION, and FBOUNDP + are 20% faster. + * bug fix: file compilation performance issues when dumping subtypes + of CHARACTER (lp#994487) + * bug fix: fixed disassembly of some SSE instructions on x86-64. + * bug fix: SB-SIMPLE-STREAMS signals an error for bogus :CLASS arguments in + OPEN. (lp#969352, thanks to Kambiz Darabi) + * bug fix: CASE normal-clauses do not allow T and OTHERWISE as keys. + (lp#959687) + * bug fix: (SETF (FIND-CLASS X) NIL) removed proper name of the underlying + classoid even if X was not the proper name of the class. (lp#941102) + * bug fix: declaration leakage between lexical environments due to careless + use of NCONC in MAKE-LEXENV. (lp#924276) + * bug fix: ENSURE-DIRECTORIES-EXIST now works when + *default-pathname-defaults* contains NAME or TYPE components. + * bug fix: PPRINT couldn't print improper lists with CARs being some symbols + from CL package, e.g. (loop . 10). + * bug fix: run-program with existent or non-existent files for :output or + :input when :if-output-exists or :if-input-does-not-exist are NIL properly + returns NIL instead of signalling an obscure error. + * bug fix: fix miscompilation of some logand forms with large constant + arguments. (lp#974406) + * bug fix: account for funcallable-instance objects properly in ROOM. + * bug fix: incorrect octets reported for c-string decoding errors. + (lp#985505) + * bug fix: miscompilation of LDB on the PowerPC platform. (thanks to Bruce + O'Neel) + * bug fix: better input error reporting for COMPILE-FILE. (lp#493380) + * bug fix: default size of non-nursery generations has been shrunk on GENCGC, + allowing faster release of memory back to the OS. (lp#991293) + * bug fix: WITH-DEADLINE (:SECONDS NIL :OVERRIDE T) now drops any + existing deadline for the dynamic scope of its body. + * bug fix: compiler-internal interval arithmetic needed to be more + conservative about open intervals when operated on by monotonic but not + strictly-monotonic functions. (lp#975528) + * bug fix: copy-tree caused stack exhaustion on long linear lists, and now + it's also slightly faster. (lp#998926) + * bug fix: better error messages for malformed declarations. + (lp#1000239) + * bug fix: define-condition didn't return the name of the defined condition. + * documentation: + ** improved docstrings: REPLACE (lp#965592) + +changes in sbcl-1.0.56 relative to sbcl-1.0.55: + * bug fix: fix copy-structure. When copying from stack to heap, garbage + could end up in the heap making GC unhappy. + (Thanks to James Knight, lp#911027) + * enhancements + ** SBCL can now be built using Clang. + ** ASDF has been updated 2.20. + * bug fix: compiler errors when weakening hairy integer types. (lp#913232) + * bug fix: don't complain about a too-hairy lexical environment for inlining + when the function has never been requested for inlining. (lp#963530) + +changes in sbcl-1.0.55 relative to sbcl-1.0.54: + * enhancements to building SBCL using make.sh: + ** --fancy can be specified to enable all supported feature enhancements. + ** --with- and --without- can be used to specify + which features to build with. + ** --arch option can be used to specify the architecture to build for. + (Mainly useful for building 32-bit SBCL's on x86-64 hosts, not + full-blows cross-compilation.) + * enhancement: extended package prefix syntax :: + which allows specifying name of the default interning package for the + whole form. + * enhancement: when *READ-EVAL* is true, arrays with element type other than + T can be printed readably using #.-based syntax. (Thanks to Robert Brown) + * enhancement: MAKE-ALIEN signals a storage-condition instead of returning a + null alien when malloc() fails. (lp#891268) + * enhancement: SB-EXT:PRINT-UNREADABLY restart for PRINT-NOT-READABLE + conditions can be conveniently accessed through function with the same + name, analogously to CONTINUE. + * enhancement: SB-EXT:*SUPPRESS-PRINT-ERRORS* can be used to suppress errors + from the printer by type, causing an error marker to be printed instead. + (Thanks to Attila Lendvai) + * enhancement: BACKTRACE and DESCRIBE now bind *PRINT-CIRCLE* to T, and + generally behave better when errors occur during printing. + * enhancement: the test runner now takes a --report-skipped-tests argument + to report the individual tests skipped as well as the number of skipped + tests. + * enhancement: undefined functions now appear in backtraces as ("undefined + function") instead of ("bogus stack frame") on x86oids. + * enhancement: detected deadlocks no longer cause stderr to be spammed, and + deadlock errors are reported in an easier-to-decipher manner. + * enhancement: DESCRIBE on type designators reports the expansion in more + cases. + * enhancement: SBCL now provides either an explicit :BIG-ENDIAN or + :LITTLE-ENDIAN in *FEATURES*, instead of :BIG-ENDIAN being implied by lack + of the :LITTLE-ENDIAN feature. (Thanks to Luís Oliveira, lp#901661) + * enhancement: better disassembly of segment-prefixes on x86 and other + instruction prefixes (e.g. LOCK) on x86 and x86-64. + * optimization: FIND and POSITION on bit-vectors are orders of magnitude + faster (assuming KEY and TEST are not used, or are sufficiently trivial.) + * optimization: SUBSEQ on vectors of unknown element type is substantially + faster. (lp#902537) + * optimization: specialized arrays with non-zero :INITIAL-ELEMENT can + be stack-allocated. (lp#902351) + * optimization: the compiler is smarter about representation selection for + floating point constants used in full calls. + * optimization: the compiler no longer refuses to coerce large fixnums to + single floats inline, except on x86 where this limitation is still + necessary. + * optimization: truncation operations on integers with constant divisor + arguments 1 and -1 are optimized away. + * bug fix: deadlock detection could report the same deadlock twice, for + two different threads. Now a single deadlock is reported exactly once. + * bug fix: interval-arithmetic division during type derivation did not + account for signed zeros. + * bug fix: compiler error when typechecking a call to a function with + non-constant keyword arguments. + * bug fix: misoptimization of TRUNCATE causing erratic behaviour. + * bug fix: condition slot accessors no longer cause undefined function + style-warnings when used in the :REPORT clause of the DEFINE-CONDITION + form that defines them. (lp#896379) + * bug fix: DEFGENERIC warns about unsupported declarations, as specified + by ANSI. (lp#894202) + * bug fix: SUBTYPEP tests involving forward-referenced classes no longer + bogusly report NIL, T. + * bug fix: bogus style-warnings for DEFMETHOD forms that both declared some + required arguments ignored and performed assignments to others. + (lp#898331) + * bug fix: *EVALUATOR-MODE* :COMPILE treated (LET () ...) identically + to (LOCALLY ...) leading to internally inconsistent toplevel-formness. + * bug fix: non-toplevel DEFSTRUCT signaled a style warning for unknown type. + * bug fix: redefining a function whose previous definition contained an + unknown type no longer causes a style-warning. (lp#806243) + * bug fix: undefined functions now appear in backtraces as ("undefined + function") instead of ("bogus stack frame") on non-x86oids. + * bug fix: backtraces are no longer cut off at ("undefined function") when + called under certain circumstances (involving a caller-allocated stack + frame) on PPC. + * bug fix: RUN-PROGRAM leaked a file-descriptor per call on non-Windows + systems. (regression since 1.0.53) + * bug fix: GC deadlocks from dladdr() on certain platforms. + * bug fix: broken standard streams no longer automatically cause recursive + errors on debugger entry. + * bug fix: build ignored --dynamic-space-size= argument to make.sh + (regression since 1.0.53) + * bug fix: attempts to stack allocate a required argument to a function with + an external entry point caused compiler-errors. + * bug fix: compiler notes for failed stack allocation for a function argument + no longer claim to be unable to stack allocate the function. + * bug fix: COERCE now signals a type-error on several coercions to + subtypes of CHARACTER that are forbidden according to ANSI. (lp#841312) + * bug fix: missing failure-to-stack-allocate compiler notes for some + forms of MAKE-ARRAY with dynamic-extent. (lp#902351) + * bug fix: some of the compile-time side-effects of DEFCLASS were not caught + by package locks. + * bug fix: Proper handling of --dynamic-space-size option on 32 bit platforms. + +changes in sbcl-1.0.54 relative to sbcl-1.0.53: * minor incompatible changes: - ** RENAME-FILE on a symbolic links used to rename the linked-to file instead - of the link. + ** RENAME-FILE on a symbolic links used to rename the linked-to file + instead of the link. ** DELETE-DIRECTORY on symbolic link to a directory used to delete the - directory, but now signal an error instead. Use TRUENAME to resolve - the pathname if you wish to delete the linked directory, and DELETE-FILE - if you wish to delete the + directory, but now signal an error instead. Use TRUENAME to resolve the + pathname if you wish to delete the linked directory, and DELETE-FILE if + you wish to delete the ** The internal SB-THREAD::SPINLOCK API has been deprecated, and using symbols associated with it will trigger a compile-time warning. * thread-related enhancements: (This work has been funded by the SBCL Threading 2011 IndieGoGo campaign. Many thanks to generous donors!) - ** Threading is now more reliable on non-Linux platforms. We still - don't consider threads on non-Linux platforms good enough to - enable them by default, but they're in a clearly better shape now. + ** Threading is now more reliable on non-Linux platforms. We still don't + consider threads on non-Linux platforms good enough to enable them by + default, but they're in a clearly better shape now. ** Deadlines supported now on all platforms. ** All blocking functions in the threading API now have a :TIMEOUT argument. @@ -27,7 +222,7 @@ changes relative to sbcl-1.0.53: * GC-related enhancements and bug fixes: ** --dynamic-space-size and --control-stack-size now understand Kb, Mb, and Gb suffixes. Default is megabytes as before. - ** Default dynamic space size on all GENCGC platforms is now 512Mb for + ** on GENCGC targets, the default dynamic space size is now 512Mb for 32-bit systems, and 1Gb for 64-bit systems. (OpenBSD/x86-64 is the only exception, defaulting to mere 444Mb to fit under default ulimits.) The new defaults are in place to prevent hitting swap on low-end systems. @@ -35,23 +230,27 @@ changes relative to sbcl-1.0.53: another default, or the runtime option to adjust the size at startup: a good size is at most equal to the amount of physical memory the system has. - ** on GENCGC systems nursery and generation sizes now default to 5% of + ** on GENCGC targets, nursery and generation sizes now default to 5% of dynamic-space size. - ** on 64-bit systems setting the nursery size above 4Gb now - works. (lp#870868) - ** SB-KERNEL:MAKE-LISP-OBJ on GENCGC no longer categorically refuses to - create SIMPLE-FUN objects. + ** on GENCGC targets, SB-KERNEL:MAKE-LISP-OBJ no longer categorically + refuses to create SIMPLE-FUN objects. + ** on 64-bit GENCGC targets, setting the nursery size above 4Gb now works. + (lp#870868) + ** on CHENEYGC targets, SB-KERNEL:MAKE-LISP-OBJ now does the same + validation of pointer objects as GENCGC does, instead of a + comparatively weak bounds-check against the heap spaces. * SB-BSD-SOCKETS bug fixes: ** GET-PROTOCOL-BY-NAME had a significant memory leak. ** GET-HOST-BY-NAME and GET-HOST-BY-ADDRESS small amounts of memory on systems with getaddrinfo(). ** GET-HOST-BY-NAME and GET-HOST-BY-ADDRESS weren't thread or interrupt safe outside systems with getaddrinfo(). + * enhancement: ASDF has been updated 2.019. + * enhancement: special-case TCO prevention for functions which never return + extended to untrusted types, keeping one more frame's worth of debug + information around in many cases. * enhancement: debug-names of anonymous and local function are more descriptive. Affects backtraces and SB-SPROF results. (lp#805100) - * enhancement: on CHENEYGC targets, SB-KERNEL:MAKE-LISP-OBJ now does - the same validation of pointer objects as GENCGC does, instead of a - comparatively weak bounds-check against the heap spaces. * enhancement: on win32, ABS of complex floats guards better against overflows. (lp#888410) * enhancement: RUN-PROGRAM now distinguishes exec() failing from child @@ -59,13 +258,15 @@ changes relative to sbcl-1.0.53: * enhancement: convenience function SET-SBCL-SOURCE-LOCATION for informing the system where on the filesystem the SBCL sources themselves are located. (Thanks to Zach Beane) + * enhancement: the compiler is now able to derive tighter bounds for + floating point numbers in some cases. (Thanks to Lutz Euler, lp#894498) * bug fix: on 64-bit targets, atomic-incf/aref does index computation correctly, even on wide-fixnum builds. (lp#887220) - * bug fix: (directory "foo/*/*.*") did not follow symlinks in foo/ that + * bug fix: (DIRECTORY "foo/*/*.*") did not follow symlinks in foo/ that resolved to directories. * bug fix: type mismatch when assigning to lexical variables no longer result in fasl-dumping internal type objects. (lp#890750) - * bug fix: type mismatch on (setf aref) and function return values no + * bug fix: type mismatch on (SETF AREF) and function return values no longer result in fasl-dumping internal type objects. * bug fix: With several combinations of argument types, for example (EXPT <(complex double)>), EXPT now uses double-precision throughout