changes in sbcl-1.1.8 relative to sbcl-1.1.7:
* notice: The implementation of MAP-ALLOCATED-OBJECTS (the heart of
ROOM, plus a few SB-INTROSPECT functions) has been completely
rewritten; it may now exhibit different bugs than before.
* new feature: minimal runtime/compiler intrastructure support for SSE
SIMD values, as SB-EXT:SIMD-PACK. Enabled on x86-64 via the build-time
feature :sb-simd-pack. This should enable intrinsics extensions as
libraries, without patching SBCL itself. Thanks to Alexander Gavrilov
for maintaining a branch for so long.
* enhancement: RUN-PROGRAM supports a :DIRECTORY argument to set
the working directory of the spawned process.
(lp#791800) (patch by Matthias Benkard)
* enhancement: boxed vectors (vectors of T or of FIXNUM) can now be
stack-allocated on PPC.
* enhancement: "fixed objects" can now be stack-allocated on PPC.
* enhancement: WITH-PINNED-OBJECTS no longer conses on PPC/GENCGC.
* enhancement: (sb-introspect:find-definition-sources-by-name x :vop) now
also returns VOPs which do not translate any functions, and finds
defoptimizer types ir2convert and stack-allocate-result.
* enhancement: better type derivation for APPEND, NCONC, LIST.
(lp#538957)
* enhancement: MAP-ALLOCATED-OBJECTS (the heart of ROOM) now walks the
heap in a manner similar to the GC rather than its previous ad-hoc
scheme, and is therefore no less and possibly more accurate.
* fixes and enhancements related to Unicode:
** the character database information has been updated to Unicode 6.2;
** support for canonical and compatibility decomposition of characters has
been added, along with support for primary composition;
** support for Unicode normalization forms of strings (NFC, NFKC, NFD and
NFKD) has been included;
** querying the character database for code points not defined by Unicode
gives less wrong answers (lp#1178038, reported by Ken Harris)
* enhancement: print intermediate evaluation results for some ASSERTed
expressions. (lp#789497) (patch by Alexandra Barchunova)
* enhancement: x86-64 disassemblies are annotated with unboxed constant
values when there are references to (RIP-relative) unboxed constants.
* bug fix: type derivation for LOG{AND,IOR,XOR} scales linearly instead
of quadratically with the size of the input in the worst case.
(lp#1096444)
* bug fix: handle errors when initializing *default-pathname-defaults*,
sb-ext:*runtime-pathname*, sb-ext:*posix-argv* on startup, like character
decoding errors, or directories being deleted.
* bug fix: Loading a core with a discontiguous dynamic space now correctly
computes the amount of dynamic space used.
* bug fix: disassembler missing ",8" on SHLD
* bug fix: backtracing through INTERNAL-ERROR signal handlers on systems
that provide an "invalid" stack frame link for the signal handler
itself now use the saved-fp-and-pc mechanism to pick up from the stack
frame in the interrupt (signal) context. This is known to affect
threaded FreeBSD/x86-64.
* bug fix: some LOOP statements couldn't be compiled.
(lp#1178989)
* bug fix: sb-sequence:dosequence works on literal vectors.
* bug fix: errors in generic arithmetic show the assembly routine's
caller on x86 and x86-64. (lp#800343)
* bug fix: Compile-time type errors should never result in COMPILE-FILE
failure. (lp#943953)
* bug fix: Known bad local calls do not cause strange failures when
emitting or dumping code. (lp#504121)
* bug fix: Multiply-inlined structure constructor don't cause IR2-time
codegen errors: type checks are inserted as necessary. (lp#1177703)
* bug fix: Unsigned modular arithmetic is correctly converted into signed
modular arithemtic (mostly to exploit fixnum-width VOPs). (lp#1026634)
* bug fix: a combination of inlined local function with &optional and
recursion no longer causes undescriptive compiler errors. (lp#1180992)
* bug fix: sub-word BOOLEAN alien types now disregard higher order bits
when testing for non-zero-ness.
* bug fix: (CONCATENATE 'null ...) no longer fails for generic sequences.
(lp#1162301)
* bug fix: Type mismatch for the value of conditional expressions are
correctly reported when detected at compile-time, instead of complaining
about a constant NIL (similar for non-EQ-comparable catch tags).
* bug fix: Referring to INLINE global functions as values should not result
in a compilation failure. (lp#1035721)
* optimization: faster ISQRT on fixnums and small bignums
* optimization: faster and smaller INTEGER-LENGTH on fixnums on x86-64.
* optimization: On x86-64, the number of multi-byte NOP instructions used
for code alignment is now always minimal.
* optimization: On 64-bit targets, the compiler generates much faster
code for type checks for types known at compile time that are smaller
than (SIGNED-BYTE 64) or (UNSIGNED-BYTE 64) and larger than FIXNUM, and
their COMPLEX variants.
* optimization: On x86 targets, more uses of ALIEN-FUNCALL during cross
compilation now inline the INVOKE-WITH-SAVED-FP-AND-PC dance.
* optimization: ROOM no longer conses so egregiously on non-x86oid
systems.
* optimization: associative bitwise operations reassociate patterns like
(f (f x k1) k2) to expose (f x (f k1 k2)). Same for + and * of
rational values.
* optimization: quasiquote expressions now perform more constant folding,
instead of consing equal lists at runtime. (lp#1026439)
* optimization: local call analysis of inlined higher-order function
should converge more quickly, resulting in better code for complex
functions.
* optimization: On x86-64, medium (word-sized but wider than 32 bits)
integer constants are handled more cleverly, especially when they
can be represented as sign-extended (signed-byte 32). (Based on a
patch by Douglas Katzman)
* optimization: IF/IF conversion should reliably result in sane code
when (some of) the results of the inner IF are always false or
always true.
* optimization: On x86oids, variable right shifts of machine words are
compiled into straight SAR/SHR, instead of going through the generic
VOP. (lp#1066204)