- update to 110.99.3:
* Implemented Basis Library proposal 2021-001 (Add getWindowSz
function to Posix.TTY structure).
* Implemented Basis Library proposal 2022-001 (Add tau to MATH signature).
* Fix for makeml and installml scripts do not like locations with spaces.
* Fix for Last branch of case expression omitted by PPAst.ppAst.
* Fix for Pretty printing of Absyn in error message does not respect fixity.
* Fix for Unable to interrupt execution by using SIGINT.
* Fixed some interfaces that did not agree with the SML Basis
Library specification. Bugs #318 (IEEEReal.decimal_approx
does not match the Basis Library) and #319 (Type of
Real.fromDecimal does not match the Basis Library).
* Fixed bug #316 (Real.fromManExp does not return expected value if man = 0.0).
* Fixed bug #317 (Conversion from string to real does not accept non-finite values).
* Fixed bug #314 (IEEEReal.float_class does not match the Basis
Library). For some reason, the NAN constructor took an
argument in our implementation. This code was probably an
early design of the API that was changed in the Basis Library
specification, but not in our code.
* Fixed bug #313 (Real.fromLargeInt crashes on large integer
input). Changed the assembly code for the AMD64 to return
infinity when the scaled exponent exceeds the maximum allowed
(instead of generating an overflow).
* Adding support for generating a SIG_GC signal when there is a garbage collection.
* Fixed bug #314 (IEEEReal.setRoundingMode is a no-op on Linux).
* Fixed bug #312 (CM.make is unable to handle filenames that
contain a backslash). We have changed the semantics of paths
given to the functions in the CM structure to be interpreted
using the native pathname syntax (instead of CM's generic syntax).
* Fixed bug #284 (Compiler bug: Contract: UsageMap on 132). The
problem was the the CPSTrans.cpstrans function was generating
code for loading spilled parameters in reverse order. In
addition to fixing the bug, added some detailed documentation of the code.
* Fixed bug #310 (Error when REPL tries to print value of type
Posix.FileSys.ST.stat).
* Fixed bug #306 (Word8VectorSlice: mapping a subslice produces
wrong result or crashes SML/NJ).
* Fix a module compilation performance bug by removing packStr
and packFct from Elaborator/modules/sigmatch.sml (reducing
the size of that file by about 25%), and removing the call of
packStr (in function constrStr) in Elaborator/elaborate/elabmod.
sml, replacing it with a call to Instantiate.instAbstr.
Goodbye at last to packStr!
* Fix for MSI installer for SMLNJ 110.99.2 reports as version 0.0.0.0.
* Fix for ml-build fails on Windows with error
* support for macOS 13 (Ventura)
- includes 110.99.2:
* Fix a benign bug where the size of a floating-point spill
record was twice as large as necessary on 64-bit systems.
* Split out the Real.toLargeInt implementation into target-word-
size versions (the Real64ToIntInf module). For 64-bit targets,
the new version uses the bit representation of the real
number to compute the result. The 32-bit version is the old code
that uses floating-point operations. This change fixes bug #279 (
Real.toLargeInt returns zero for anything in range [-512,512]).
* In the translation from Absyn to PLambda, there was a
function (inlops) that was used to build the primop and type
data structures for numeric types. This function was being
called for every primitive operator, even though its results
only depended on the numeric type. I added a hash table to cache
the results indexed numeric kind. This change speeds up the
compiler by about 3% (e.g., compiling the compiler went from 58s
to 55s on a MacBookPro with a 2.4GHz Intel i9 processor).
* support installing and running SML/NJ on macOS 12 (Monterey)
- includes 110.99.1:
* Changed the AMD64 frame layout to include a word to hold the
Overflow exception. This value is used by the LLVM backend to
generate the exception for checked arithmetic operations.
* Some minor restructuring of the logic in the generic installer.
* Fixed a bug with how FLINT numeric types were being
translated to CPS types. Specifically, types that were smaller
than the default integer size (e.g., word8) should have been
marked as having a tagged representation.
* Fixed a serious performance bug in the implementation of the
CharBuffer and MonoBuffer structures. Essentially, if one did
not reserve sufficient space for the contents, it could take
quadratic time to fill the buffer. We now grow the buffer by a
factor of 1.5 of its current size, with an upper bound on the extra growth of 256K.
* Fix for bug #280 (110.99 config/install.sh -64 fails on macOS
10.15.7). I was unable to reproduce this problem, but after
some investigation, it appears that the problem was inconsistent
build tools being picked up from the user’s path. To protect
against this issue, I made the paths to the ar and ranlib tools absolute.
* Added support for running SML/NJ on M1 Macs via the Rosetta2
emulator. The change is to identify the arm processor as amd64
in the config/_arch-n-opsys script. Note that while the system
basically seems to work okay under rosetta, trying to run the
makeml command after having compiled the complier caused a crash.