More updates in preparation for release.
Use recursive-lambda icon as default on windows.
Force working directory to user's home-dir when started as mac application bundle.
Add new recursive-lambda icon for use in mac and windows.
Restrict *.ico rule.
Update documentation to reflect changes concerning eq hash tables.
Document MAKE-WEAK-EQ-HASH-TABLE and MAKE-WEAK-EQV-HASH-TABLE.
Warn that MAKE-EQ-HASH-TABLE and MAKE-EQV-HASH-TABLE may become
aliases for MAKE-STRONG-EQ-HASH-TABLE and MAKE-STRONG-EQV-HASH-TABLE
instead. Suggest using MAKE-STRONG-EQ-HASH-TABLE for symbol-keyed
tables, rather than MAKE-EQ-HASH-TABLE.
Eradicate use of MAKE-EQ-HASH-TABLE.
Replace each use by a constructor with a more specific name.
Full analysis:
* Use of MAKE-EQ-HASH-TABLE in MIT Scheme, 2009-12-21 -*- outline -*-
In MIT Scheme, MAKE-EQ-HASH-TABLE yields a hash table whose keys are
held only weakly, so that they will be garbage-collected if there are
no strong references to them. To make a similar hash table whose keys
are held strongly, one must use MAKE-STRONG-EQ-HASH-TABLE explicitly.
Of the sixty-two uses of MAKE-EQ-HASH-TABLE throughout the MIT Scheme
source code, only four appear to need weak references to their keys,
and for only two more is there an obvious reason to use weak
references. This list categorizes most uses of MAKE-EQ-HASH-TABLE.
Each entry is marked by the decision that was made for it of whether
it should be strong or weak.
** Hash table must be key-weak
*** [weak] edwin/curren.scm, screen-buffer-layouts
*** [weak] edwin/eystep.scm, stepper-buffers
*** [weak] edwin/xterm.scm, display/cached-atoms-table
*** [weak] edwin/xterm.scm, display/selection-records
** Hash table should be key-weak, but will work as key-strong
*** [weak] edwin/comman.scm, permanent-local-variables
As long as Edwin variables are strongly interned, it is safe to use a
key-strong hash table for properties on Edwin variables such as the
permanent-local property, but if Edwin variables were weakly interned
then this would be a space leak.
*** [strong] sos/class.scm, built-in-class-table
Dispatch tags that are garbage-collected shouldn't accumulate space
here. These might arise, for instance, from re-evaluating
DEFINE-RECORD-TYPE forms causing the old record type's dispatch tags
to be garbage-collected. That requires key- and datum-weak hash
tables, though. Using MAKE-WEAK-EQ-HASH-TABLE here breaks the
bootstrap from the
20090107 snapshot, which doesn't have a binding for
that name. So just use a strong eq hash table for now.
** Hash table should be key-strong, but will work as key-weak
*** [strong] compiler/machines/i386/lapopt.scm, *rules*
If the compiler never generates instructions with certain symbols in
them, and hence the compiler's code has no strong references to those
symbols, then the rules for those symbols may as well be discarded.
But that's pretty sketchy.
*** [strong] compiler/machines/svm/assembler-runtime.scm, symbolic-operators
*** [strong] compiler/machines/svm/assembler-runtime.scm, pvar-type-table
*** [strong] compiler/machines/svm/lapopt.scm, *rules* (not really used)
*** [strong] compiler/machines/x86-64/lapopt.scm, *rules* (not really used)
*** [strong] edwin/nntp.scm, equivalences in build-equivalence-classes
The only reason that this works as a key-weak hash table is that every
key is also strongly referenced by the hash table's data.
*** [strong] edwin/nntp.scm, tables in convert-header-graphs-to-trees
In the only caller of CONVERT-HEADER-GRAPHS-TO-TREES, the (strong)
list of headers is still strongly referenced, so the keys of the two
hash tables in TABLES will not be garbage-collected.
*** [strong] edwin/xterm.scm, built-in-atoms-table
If the binding for BUILT-IN-ATOMS were collected while that for
BUILT-IN-ATOMS-TABLE were not, then the latter would be in trouble.
This generally doesn't happen currently.
*** [strong] microcode/os2pm.scm, type-abbreviations
*** [strong] microcode/os2pm.scm, id-external-roots
This code is probably defunct, but if it weren't, and if the
presentation manager procedure abstraction were used outside this
file, it would probably be necessary to make these two hash tables
key-strong. Another hash table, PM-PROCEDURES, is incorrectly
key-weak.
*** [strong] runtime/genio.scm, {en,de}coder/sizer/{,de}normalizer maps
Since there are maps in both directions, each hash table's keys also
have strong references in the data positions of the other hash table.
But this is pretty fragile, and in any case there is no need to use
key-weak hash tables.
*** [strong] runtime/syntax-output.scm, unmappings
*** [strong] runtime/syntax-output.scm, rename-databases' mapping-tables
*** [strong] runtime/syntax-output.scm, rename-databases' id-tables
*** [weak] ssp/xmlrpc.scm, methods in get-xmlrpc-method-handler
Since the hash table is used only in one place, and only one key is
fetched out of it, that key will be strongly referenced until it is
fetched, and the other keys don't matter. (For that matter, why use a
hash table at all?)
*** [strong] xml/turtle.scm, table in write-prefixes
** Hash table must be key-strong
*** [strong] compiler/machines/svm/assembler-runtime.scm, symbol tables
These are probably meant to be treated like the symbol tables
implemented in compiler/back/symtab.scm.
*** [strong] imail/imail-file.scm, file-folder-types
*** [strong] imail/imail-mime.scm, mime-encodings
*** [strong] microcode/os2pm.scm, pm-procedures
*** [strong] runtime/http-syntax.scm, header-value-defns
*** [strong] ssp/mod-lisp.scm, mime-handlers
*** [strong] ssp/xhtml-expander.scm, *sabbr-table*
The processing instructions are processed incrementally as the file is
parsed, so keys in the sabbr table may be garbage-collected and then
re-interned, between which times the associations would be destroyed.
*** [strong] star-parser/matcher.scm, matcher-preprocessors
*** [strong] star-parser/matcher.scm, matcher-compilers
*** [strong] star-parser/parser.scm, parser-preprocessors
*** [strong] star-parser/parser.scm, parser-compilers
*** [strong] star-parser/shared.scm, make-parser-macros
*** [strong] star-parser/shared.scm, *global-parser-macros*
*** [strong] xdoc/validate-xdoc.scm, element-checkers
*** [strong] xdoc/xdoc.scm, html-generators
*** [strong] xdoc/xdoc.scm, when-conditions
*** [strong] xdoc/xdoc.scm, xdoc-content-types
*** [strong] xdoc/xdoc.scm, xdoc-element-types
*** [strong] xdoc/xdoc.scm, xdoc-input-canonicalizers
*** [strong] xdoc/xdoc.scm, xdoc-output-definitions
*** [strong] xml/xhtml.scm, element-context-map
** Other
*** [strong] compiler/machines/C/stackify.scm, stackify count tables
It is not immediately clear to me whether these need to be strong, but
they probably should be.
*** [strong] compiler/rtlbase/rtlobj.scm, label->object maps
It's not immediately obvious to me whether LABEL->OBJECT will ever be
used after the last references to the keys of the hash tables involved
(which cause strong references to those keys to be dropped, if
COMPILER:PRESERVE-DATA-STRUCTURES? is false). A little further
analysis is required. A conservative guess would be that these should
to be strong.
*** [weak] edwin/eystep.scm, ynode-regions
What are the keys to these hash tables? I don't know how long they
persist.
Chris's answer: The keys are stepper nodes; they are stored in a text
property and consequently will be held appropriately. This should be
a weak table.
*** [weak] edwin/prompt.scm, prompt-histories
Making this key-strong is not likely to be a space leak, but if a
command is garbage-collected, then so may the symbol for its history
be garbage-collected, and its history is presumably no longer needed.
*** [strong-eqv] edwin/win32.scm, event-handlers
The keys in EVENT-HANDLERS are integers. I think this should be a
key-strong eqv hash table rather than a key-weak eq hash table.
*** [strong] edwin/xterm.scm, selection->record table in display/selection-records
I'm not sure what the domain of possible keys to this hash table is --
it may be just the symbols PRIMARY and CLIPBOARD, which will probably
be strongly referenced by the rest of the Edwin code, but on the other
hand I think this should probably be a key-strong hash table.
*** [strong] edwin/xterm.scm, symbol->x-atom table in display/cached-atoms-table
Whether this should be key-strong or key-weak depends on whether the
set of atoms that it will map is arbitrarily large.
*** [weak] imail/imail-core.scm, memoized-resources
As long as URLs are interned strongly, it doesn't matter whether
MEMOIZED-RESOURCES is key-weak or key-strong. This requires further
analysis, but leaving it weak for now (i.e. not changing it) is safe
until internment of URLs changes.
*** [strong with comment] runtime/sfile.scm, interned-mime-types
This should really be a key-strong, datum-weak hash table, which the
hash table abstraction does not presently support. (The same goes for
UNUSUAL-INTERNED-MIME-TYPES.) Barring that, INTERNED-MIME-TYPES
should be key-strong, although it happens to work as a key-weak hash
table because its data have strong references to its keys anyway.
*** [weak] xdoc/xdoc.scm, *xdoc-element-properties*
*** [weak] xdoc/xdoc.scm, *xdoc-inputs*
*** [weak] xdoc/xdoc.scm, *xdoc-outputs*
As maps from XML elements by identity to properties, these three
should probably be key-weak, but since new ones are created for each
xdoc expansion, they could probably safely be key-strong without
badly leaking memory.
*** [strong] xdoc/xdoc.scm, *xdoc-id-map*
This should be key-strong, although it doesn't matter until XML names
become weakly interned.
*** [strong-eqv] xml/xhtml-entities.scm, table in html-char->name-map
The keys in the table inside HTML-CHAR->NAME-MAP is keyed by
characters. Should this be a key-strong eqv hash table rather than a
key-weak eq hash table?
*** [strong with comment] xml/xml-names.scm
If we had datum-weak hash tables, then EXPANDED-NAMES should be
key-weak and the other two should be datum-weak. But we don't, so
instead all three should be strong, since in each one the data all
have strong references to their corresponding keys.
Reflect specification bug in definition of HASH-TABLE-UPDATE!.
We can't use HASH-TABLE/MODIFY! because the specification tacitly
permits the procedure to edit the hash table. This was a silly
oversight in the SRFI document.
Add names for hash table types with the word `weak' in them.
New global bindings:
MAKE-WEAK-EQ-HASH-TABLE
MAKE-WEAK-EQV-HASH-TABLE
WEAK-EQ-HASH-TABLE-TYPE
WEAK-EQV-HASH-TABLE-TYPE
{EQ,EQV}-HASH-TABLE-TYPE and MAKE-{EQ,EQV}-HASH-TABLE are now aliases
for WEAK-{EQ,EQV}-HASH-TABLE-TYPE and MAKE-WEAK-{EQ,EQV}-HASH-TABLE,
as are MAKE-{SYMBOL,OBJECT}-HASH-TABLE.
Eliminate dist/windows, moving its contents to other places.
Adjust scripts to account for change to default band.
Tweak rule used for building weak keys, and update comment to match.
Make sure all necessary load-time dirs and debugging info is present.
Allow uninterned symbols to be weak keys.
Ensure weak symbols aren't removed from hash table keys.
Fix: W32 file-system flags can be a bignum.
Tweak CUSTOM-TABLE-TYPE in hashtb.scm further.
Use existing names for table types where possible, rather than
repeating the calls to MAKE-*-TYPE.
Reduces the size of hashtb.{bin,com,bci} by 10-12% on i386, and
probably measurably reduces the time to compile them too.
For SRFI 69's MAKE-HASH-TABLE, always yield strong hash tables.
SRFI 69 (intentionally) says nothing about weak references, and other
implementations always give strong hash tables.
Fix thinko.
Add "lib" to list.
Tweak command-line options.
Write windows script to copy binaries into "c:\local".
Change default library path on windows.
Change make-lib to work more like "make install" on unix.
Tweak build scripts; add new script to build lib directory.
Fix windows .bat files to work without depending on defaults.
Must have fully lower-case makefile for windows.
Add more mostly windows-build patterns.
Restore mistakenly-deleted patterns.
Move MacOS X file into macosx subdir.
Add files to build and distribute Windows binaries.
Convert to use winsock2.h instead of winsock.h.
Add missing utabmd.*.
Change types to match Open Watcom 1.8.
Add declarations for external references.
Add better name for MacOS X app rule.
Add rule to make MacOS X application.
Wrap over-long line.
Wrap over-long line.
Restructure to allow caller to override various environment variables.
Use autoconf macro to capture code that computes native-code value.
Make sure AC_CHECK_DECL is run *after* AC_PROG_CC.
Include icon in app.
Use much larger default heap size on 64-bit machines.
Use much larger default heap size on 64-bit machines.
When compiler supports x86-64, use x86-64 in preference to x86.
Find structure field default value procedures by index, not by name.
Otherwise, every constructor generated by DEFINE-STRUCTURE performs
lookups by name for every field with a default initializer, every
time you call the constructor.
This is a backward-compatible change, in the sense that .bin and .com
files generated with the old definition of DEFINE-STRUCTURE can still
be loaded into images with the new definition of DEFINE-STRUCTURE and
its corresponding run-time support. However, .bin and .com files
generated with the new definition cannot be loaded into images with
the old run-time support, which lacks procedures needed by the new
definition for lookup by index.
Note: For the full benefit of this change, run at least a two-stage
build so that the runtime will include the new definition and be
compiled with the new definition.
Bum case canonicalization and character input in the parser.
This avoids an ASSQ for every character read and some interpreter
environment lookups for every character dispatched on and atom read,
and replaces a general unknown procedure call by some open-coded
string and character operations for every character read in an atom
to canonicalize its case.
Use PRIMITIVE-OBJECT-SET-TYPE, not OBJECT-SET-TYPE, in EOF-OBJECT.
Neither is constant-folded, but PRIMITIVE-OBJECT-SET-TYPE is
open-coded while OBJECT-SET-TYPE is not. (Is there a reason for
this? An artefact of Multischeme?)
Export ERROR:NOT-INPUT-PORT and ERROR:NOT-OUTPUT-PORT.
These should be exported anyway, but in particular they caused
INTEGRATE-EXTERNAL in files other than port.scm to generate
references to them out of their scope.
Reduce (<= a b ...) to (and (<= a b) ...) in genio column tracking.
Neither SF nor LIAR currently performs that reduction, so every
character written to a generic I/O port was causing a rest list to be
consed up and control to go through a generic microcode primitive
trampoline to do the reduction at run-time.
Also use FIX:<= rather than <=, since the numbers in question are all
fixnums.
Fix ERROR:ILLEGAL-STREAM-ELEMENT in runtime/stream.scm.
The first argument to what CONDITION-CONSTRUCTOR returns must be a
continuation, so (ERROR (MAKE-ILLEGAL-STREAM-ELEMENT ...)) doesn't
work. Instead use CONDITION-SIGNALLER.
Fix bug in recent change to runtime/uenvir.scm's WALK-GLOBAL.
MAP-ENTRY takes two arguments, not just one.
Use weak pairs for some entries in obarray buckets.
Entries for symbols with global bindings still use ordinary pairs.
This enables interned symbols without global bindings to be garbage-
collected. Operations in lookup.c that create or destroy global
bindings update the relevant obarray bucket entry to strengthen or
weaken it.
In the process, fix intern.c's string_hash to accept a long string
length, rather than a uint32_t string length, so that strings longer
than 4 GB will not lose on 64-bit systems.
Fix invocation of out-of-line FIXNUM-LSH hook for i386 and x86-64.
Add comment explaining why it is unnecessary to clear the register
map before invoking the hook.
Clean up texinfo auxiliary files (e.g. imail.aux).
* doc/make-common.in (clean): Use new TEXINFO_ROOT variable to delete
the texinfo auxiliary files. Also used it in place of the GNU make
specific "word" function.
* doc/imail/Makefile.in (clean): Set TEXINFO_ROOT.
* doc/sos/Makefile.in (clean): ditto
* doc/ref-manual/ref-manual (clean): ditto
* doc/user-manual/Makefile.in (clean): ditto
Use port/%state in genio ports.
Use port/%state in stringio ports.
Export %char-set-member? to parser.
Use faster write-char when transcribing chars.
Use faster i/o in parse.
Ensure input and output are syntaxed early on.
Export fast input and output operations.
Ensure port is syntaxed before input and output.
Add fast write-char and discretionary-flush.
Add fast read-char and peek-char.
Expose fast port operators.
Don't run-boot-inits! if saved-boot-inits is unbound.
Add an assembly hook for SET-INTERRUPT-ENABLES! on i386.
Mark OVERFLOW? ignored in FIXNUM-LSH methods on i386 and x86-64.
Show only current year in start-up copyright.
Fix references to flonum_zero and flonum_one.
Make x86 and x86-64 assembly hooks for out-of-line FIXNUM-LSH.
I believe this now covers all cases, and is more reasonable than
generating for each FIXNUM-LSH invocation the nearly two dozen
instructions composing the assembly hooks.
Fixes Savannah bug 27384.
Fix bug in new double_round: flip order of arguments to copysign.
Use ja/jb, not jg/jl, in floating-point predicates on x86-64.
Fix flonum case of ROUND to use IEEE round-to-even, per the R5RS.
Fix i386 and x86-64 generic multiplication hooks on fixnums.
Before multiplying, both tags must be cleared, and one of the
operands must be shifted to include a factor of 2^6. Shifting both
operands, so that there is an extra factor of 2^12, or not clearing
the tag on both operands, is wrong. However, this didn't affect any
real code, since the fixnum case of generic multiplication is always
open-coded by the compiler unless you futz with compiler switches.
Cache PORT/OPERATION/READ-CHAR in INPUT-PORT/READ-LINE and INPUT-PORT/READ-STRING.
Remove redundant PORT? test.
Fix bug in x86-64 instruction encoder: REX prefix must be last.
(Thanks, AMD, for making it seem like the 66, F2, and F3 bytes are
part of the `opcode' all throughout the media instruction set
manual.)
Also change ([U]COMISF ...) to be ([U]COMIF S ...) for consistency.
Fix headings on x86-64's instr1.scm, instr2.scm, and lapgen.scm.
Open-code floating-point primitives on AMD x86-64.
Only lightly tested. Use with caution. Slippery when wet. Keep out
of reach of children.
Test stack overflow in LIARC interrupt checks.
Provide default binding for MIT_SCHEME_EXE.
Fix typo.
Merge git pull.
Fix thinko: the boot-root executable is _not_ $(MIT_SCHEME_EXE).
Fix typo.
Fix various minor problems introduced in the configure scripts.
Also kludge up a fix to etc/run-three-stages.
Simplify x86-64 utility interface, and nuke all x87 code from it.
The calling convention for microcode utilities now more closely
matches the standard calling convention for C on the AMD64, which
reduces the amount of code needed in scheme_to_interface. The i386
hack to split the assembly hook addresses between negative addresses
and positive addresses is no more; most of what little benefit it
added on the i386, it ceased to add on x86-64, since no hooks fit in
positive byte offsets from the register block start any longer, and
only sixteen would fit in negative byte offsets. Perhaps later it
would be worth deciding which sixteen deserve byte offsets (e.g.,
scheme-to-interface certainly would), but this is simpler for now.
Note that there is still a space advantage to invoking a hook (jump
to an offset from Registers) over invoking an interface (move its
number to AL, jump to scheme-to-interface), since invoking an
interface requires going through a hook anyway.
These change serves mainly to simplify the code, not to improve
performance, although shuffling the calling convention may improve
performance as a bonus.
The x87 code appears to have been causing problems, and will later
be replaced by the AMD64's 128-bit media (SSEn) instructions anyway.
The compiler still has some x87 vestiges left in it, but not in any
code that is hit, and nearly all of it is commented out now.
*** NOTE: Since this changes the calling convention of utilities
from compiled code, it breaks all existing x86-64 .com files, which
you must recompile with the new compiler. This means that you can't
straightforwardly rebuild Scheme the usual way from a prior x86-64
installation. The easiest way to proceed is to start again from a
LIARC snapshot and rebootstrap the x86-64 code. Expect more of this
to come; in particular, since we now have a reasonable number of
machine registers, it may be worthwhile to assign one to be the
return value register.
projects / mit-scheme.git / log