Ana Guerrero
bf09586486
- update to 110.95.5:
* Changed Unsafe.Real64.toBits to Unsafe.Real64.castToWord and
Unsafe.Real64.fromBits to Unsafe.Real64.castFromWord. These
new names match MLton’s names for these operations.
* Reimplement the PackReal64Big and PackReal64Big structures to
use the bit-casts between 64-bit reals and 64-bit words. This
change fixes Issue #300 (PackReal64{Big,Little}.update is not implemented).
* Added ALIGN_STACK_16 preprocessor flag to x86 assembly code (
mach-dep/X86.prim.asm) in runtime system. This flag controls
whether the SML stack frame is a multiple of 16 bytes (which
allows foreign function calls to ensure 16-byte stack alignment).
* Expose the machine property functions as Unsafe.isBigEndian and
Unsafe.wordSize. These functions are inlined by the compiler,
which allows one to expect tests of the form if Unsafe.
isBigEndian() to be evaluated at compile time. These primops are
also available to user code via the Unsafe structure.
* Improved the Unsafe.Real64.fromBits implementation to use a
cast and a reference (instead of pack/unpack from a byte array).
* We now use Word_t to represent lengths in the various
allocation routines in gc/ml-objects.c. This change fixes
Issue #283 (TextIO.inputAll segfaults when reading 980M file).
* Added implementation of Real64.nextAfter.
* Rewrote the Real64.split function to use the bit-
representation of reals in its implementation. This new
version fixes Issue #269 (Real.realMod and Real.split produce
incorrect result for values close to zero).
* Did a complete rewrite of the mechanisms used to implement
conversions between strings and reals. As part of this rewrite,
I implemented the support for the StringCvt.EXACT formatting
mode and implemented the missing Real.toDecimal/fromDecimal
functions. These changes also fix Issue #194 (Incorrect
formatting of real number). The new implementation is based on
the Ryũ library and related PLDI 2018 paper by Ulf Adams.
* Fixed Issue #298 (Incorrect printing of source code in error message).
* Added Unsafe.Real64 structure to hold conversions between
Real64.real and Word64.word values. The fromBits function can be
further improved, but it needs the new host_big_endian primop to
select the best code path based on endianess.
* Refactoring the Basis implementation directory structure to
move code related to real numbers into its own directory.
Modules that support sequences of reals are not included in this
refactoring. This is the first step to a reimplementation of the
support for printing reals.
* Added host_word_size and host_big_endian primops. These will
expand into constant functions (e.g., fn () ⇒ 64 for
host_word_size on a 64-bit system) that will be visible to the optimizer.
* The implementations of PackWord64Little.update and PackWord64Big
.update were swapped on 64-bit platforms (resulting in byte-
order reversal). These have been fixed. Note that the 32-bit versions are correct.
* Modified the implementation of OS.Path to allow for systems (
like Windows) that support more than one valid arc separator
character. This change addresses Issue #280 (Support forward-
slash ("/") as a separator in the Windows implementation of OS.Path).
* Modified the CPS contraction phase to address Issue #292 (
Word8.toLargeInt is broken in 32-bit version). The change is to
disable fusing of conversions to IntInf.int that use 64-bit
numbers as the intermediate type on 32-bit platforms.
* There was an inconsistency in the way that datatypes were
printed in the REPL depending on if they were a top-level
definition or defined in a structure. It was also the case that
the printing of top-level declarations (TopLevel/print/ppdec.sml)
was using relative indentation, while the printing of
declarations in modules (ElabData/modules/ppmod.sml) was
using absolute indentation. I changed the code in TopLevel/
print/ppdec.sml to use absolute indentation.
* Fix bug in new-literals.sml, where the integer value 231 was
being stored as a 32-bit signed integer in the literal table,
instead of as a 64-bit signed integer. This change fixes Issue
#287 (The word literal 0wx80000000 is incorrectly converted to 0wx7FFFFFFF80000000).
* Add contraction rule for when two calls to SETHDLR appear in
sequence. This situation arises when the body of a handler is optimized away.
OBS-URL: https://build.opensuse.org/request/show/1172026
OBS-URL: https://build.opensuse.org/package/show/openSUSE:Factory/smlnj?expand=0&rev=27