* etc/NEWS: Document this.
* src/bytecode.c (exec_byte_code):
* src/data.c (arithcompare):
Do not lose information when comparing floats to integers.
* test/src/data-tests.el (data-tests-=, data-tests-<)
(data-tests->, data-tests-<=, data-tests->=):
Test this.
* src/bytecode.c: (exec_byte_code) While linear searching the jump
table, compare the value's hash table first to avoid calling
h->test.cmpfn every time.
* src/bytecode.c (exec_byte_code): Don't check that the hash code is
not nil when linear scanning the jump table. Hash tables for are
declared with :size as the exact number of cases, so each entry i
should have a hash code. When BYTE_CODE_SAFE, do it as a sanity
check.
* lisp/emacs-lisp/byte-opt.el,
lisp/emacs-lisp/bytecomp.el (byte-decompile-bytecode-1),
(byte-compile-lapcode): Calculate the actual jump address while
compiling, store it in the jump table.
* src/bytecode.c: Jump to the looked up value directly, do a linear
search when the number of elements is <= 5.
This fixes some infinite loops that cannot be quitted out of,
e.g., (defun foo () (nth most-positive-fixnum '#1=(1 . #1#)))
when byte-compiled and when run under X. See:
http://lists.gnu.org/archive/html/emacs-devel/2017-01/msg00577.html
This also attempts to keep the performance improvements I recently
added, as much as possible under the constraint that the infloops
must be caught. In some cases this fixes infloop bugs recently
introduced when I removed immediate_quit.
* src/alloc.c (Fmake_list):
Use rarely_quit, not maybe_quit, for speed in the usual case.
* src/bytecode.c (exec_byte_code):
* src/editfns.c (Fcompare_buffer_substrings):
* src/fns.c (Fnthcdr):
* src/syntax.c (scan_words, skip_chars, skip_syntaxes)
(Fbackward_prefix_chars):
Use rarely_quit so that users can C-g out of long loops.
* src/callproc.c (call_process_cleanup, call_process):
* src/fileio.c (read_non_regular, Finsert_file_contents):
* src/indent.c (compute_motion):
* src/syntax.c (scan_words, Fforward_comment):
Remove now-unnecessary maybe_quit calls.
* src/callproc.c (call_process):
* src/doc.c (get_doc_string, Fsnarf_documentation):
* src/fileio.c (Fcopy_file, read_non_regular, Finsert_file_contents):
* src/lread.c (safe_to_load_version):
* src/sysdep.c (system_process_attributes) [GNU_LINUX]:
Use emacs_read_quit instead of emacs_read in places where
C-g handling is safe.
* src/eval.c (maybe_quit): Move comment here from lisp.h.
* src/fileio.c (Fcopy_file, e_write):
Use emacs_write_quit instead of emacs_write_sig in places where
C-g handling is safe.
* src/filelock.c (create_lock_file): Use emacs_write, not
plain write, as emacs_write no longer has a problem.
(read_lock_data): Use emacs_read, not read, as emacs_read
no longer has a problem.
* src/fns.c (rarely_quit): Move to lisp.h and rename to
incr_rarely_quit. All uses changed..
* src/fns.c (Fmemq, Fmemql, Fassq, Frassq, Fplist_put, Fplist_member):
* src/indent.c (compute_motion):
* src/syntax.c (find_defun_start, back_comment, forw_comment)
(Fforward_comment, scan_lists, scan_sexps_forward):
Use incr_rarely_quit so that users can C-g out of long loops.
* src/fns.c (Fnconc): Move incr_rarely_quit call to within
inner loop, so that it catches C-g there too.
* src/keyboard.c (tty_read_avail_input): Remove commented-out
and now-obsolete code dealing with interrupts.
* src/lisp.h (rarely_quit, incr_rarely_quit): New functions,
the latter moved here from fns.c and renamed from rarely_quit.
(emacs_read_quit, emacs_write_quit): New decls.
* src/search.c (find_newline, search_buffer, find_newline1):
Add maybe_quit to catch C-g.
* src/sysdep.c (get_child_status): Always invoke maybe_quit
if interruptible, so that the caller need not bother.
(emacs_nointr_read, emacs_read_quit, emacs_write_quit):
New functions.
(emacs_read): Rewrite in terms of emacs_nointr_read.
Do not handle C-g or signals; that is now for emacs_read_quit.
(emacs_full_write): Replace PROCESS_SIGNALS two-way arg
with INTERRUPTIBLE three-way arg. All uses changed.
The old code that sets and clears immediate_quit was
ineffective except when Emacs is running in terminal mode, and
has problematic race conditions anyway, so remove it. This
will introduce some hangs when Emacs runs in terminal mode,
and these hangs should be fixed in followup patches.
* src/keyboard.c (immediate_quit): Remove. All uses removed.
There’s no longer need to have QUIT stand for a slug of C statements.
Use the more-obvious function-call syntax instead.
Also, use true and false when setting immediate_quit.
These changes should not affect the generated machine code.
* src/lisp.h (QUIT): Remove. All uses replaced by maybe_quit.
'switch' takes two arguments from the stack: the variable to test, and
a jump table (implemented as a hash-table with the appropriate :test
function). By looking up the value of the variable in the hash table,
the interpreter can jump to the label pointed to by the value, if any.
This implementation can only be used for `cond' forms of the type
`(cond ((test x 'foo) 'bar) ...)`, such that the function `test` and
variable `x` is same for all clauses.
* lisp/emacs-lisp/bytecomp.el:
* Add (byte-compile-cond-valid-obj2-p), (byte-compile-cond-vars),
(byte-compile-cond-jump-table-info), (byte-compile-jump-table-add-tag),
(byte-compile-cond-jump-table), byte-compile-jump-tables.
* Add defcustom `byte-compile-cond-use-jump-table'.
* (byte-compile-cond): Use them.
* (byte-compile-lapcode): Patch tags present in jump tables, if any.
* lisp/emacs-lisp//byte-opt.el: (byte-optimize-lapcode): Add checks to
some peephole optimizations to prevent them from messing up any code
involving `byte-switch`.
* src/bytecode.c: (exec_byte_code): Add bytecode Bswitch.
Inspired by a suggestion from Chris Gregory in:
http://lists.gnu.org/archive/html/emacs-devel/2016-12/msg00965.html
On my platform, this generates exactly the same machine insns.
* src/eval.c (prog_ignore): Rename from unwind_body, since
it’s more general than that. All callers changed.
(Fprog1): Simplify by using prog_ignore.
(Fwhile): Clarify by using prog_ignore.
This improves performance overall on my benchmark on x86-64,
since the interpreted program-counter resides in a machine
register rather than in RAM.
* etc/DEBUG, src/.gdbinit: Remove xbytecode GDB command, as there
is no longer a byte stack to decode.
* src/bytecode.c (struct byte_stack, byte_stack_list)
(relocate_byte_stack): Remove. All uses removed.
(FETCH): Simplify now that pc is now local (typically, in a
register) and no longer needs to be relocated.
(CHECK_RANGE): Remove. All uses now done inline, in a different way.
(BYTE_CODE_QUIT): Remove; now done by op_relative_branch.
(exec_byte_code): Allocate a copy of the function’s bytecode,
so that there is no problem if GC moves it.
* src/lisp.h (struct handler): Remove byte_stack member.
All uses removed.
* src/thread.c (unmark_threads): Remove. All uses removed.
* src/thread.h (struct thread_state): Remove m_byte_stack_list member.
All uses removed. m_stack_bottom is now the first non-Lisp field.
* src/bytecode.c (BYTE_MAINTAIN_TOP): Remove; no longer needed.
(struct byte_stack) [BYTE_MAINTAIN_TOP]:
Remove unused members ‘top’ and ‘bottom’.
(exec_byte_code): Nest inside { } to avoid GCC warning about
jumping over declaration when compiled with -DBYTE_CODE_SAFE.
* src/bytecode.c (exec_byte_code):
* src/image.c (parse_image_spec):
Prefer FUNCTIONP (x) to !NILP (Ffunctionp (x)).
* src/eval.c (FUNCTIONP): Move here ...
* src/lisp.h: ... from here. No longer inline, as that
bloats the text and does not help speed (at least on my platform).
(functionp): Remove this name, since callers use FUNCTIONP.
This does not let Emacs run, just build.
* lib-src/etags.c (main):
* lib-src/profile.c (main):
Use return, not exit.
* src/bytecode.c (BYTE_CODE_THREADED) [__CHKP__]:
Do not define, as -fcheck-pointer-bounds is incompatible with taking
addresses of labels.
* src/menu.c (Fx_popup_dialog): Use eassume, not eassert,
to pacify gcc -fcheck-pointer-bounds -Wnull-dereference.
* src/bytecode.c (exec_byte_code): Simplify, mostly by moving
initializers into decls, and by omitting some unnecessary
changes to ‘top’. This improves performance a bit on x86-64,
as it happens.
This improves performance overall on my benchmark on x86-64,
since the interpreted program-counter resides in a machine
register rather than in RAM.
* etc/DEBUG, src/.gdbinit: Remove xbytecode GDB command, as there
is no longer a byte stack to decode.
* src/bytecode.c (struct byte_stack, byte_stack_list)
(relocate_byte_stack): Remove. All uses removed.
(FETCH): Simplify now that pc is now local (typically, in a
register) and no longer needs to be relocated.
(CHECK_RANGE): Remove. All uses now done inline, in a different way.
(BYTE_CODE_QUIT): Remove; now done by op_relative_branch.
(exec_byte_code): Allocate a copy of the function’s bytecode,
so that there is no problem if GC moves it.
* src/lisp.h (struct handler): Remove byte_stack member.
All uses removed.
(SAFE_ALLOCA_LISP_EXTRA): New macro, a generalization of
SAFE_ALLOCA_LISP.
(SAFE_ALLOCA_LISP): Use it.
* src/bytecode.c (exec_byte_code): Remove MAX_ALLOCA-based limit
on bytecode maxdepth, by using SAFE_ALLOCA_LISP instead of alloca.
pipeline is fuller.
* src/bytecode.c (BYTE_CODE_QUIT): Check for GC, too. Do the
check only once every 256 times. This should be good enough, and
improves performance significantly on x86-64 as branch-prediction
typically assumes checking will not be done so the instruction
pipeline stays fuller.
(exec_byte_code): Set up the quit counter. Don’t call maybe_gc
directly, as BYTE_CODE_QUIT does that now.
* src/bytecode.c (BYTE_CODE_SAFE): Default to false, so that it
can be used outside #if. All uses of ‘defined BYTE_CODE_SAFE’
changed to ‘BYTE_CODE_SAFE’. Use BYTE_CODE_SAFE in plain
expressions instead of #if expressions when this is easy.
(struct byte_stack) [BYTE_CODE_SAFE]: Remove member ‘bottom’,
as it is no longer needed.
(exec_byte_code): Omit #if BYTE_CODE_SAFE when this is easy.
Simplify stack-overflow checking when BYTE_CODE_SAFE.
BYTE_CODE_METER hasn’t worked since 2013, and nobody seems to have
noticed. Perhaps we should remove it?
* src/bytecode.c (METER_2) [BYTE_CODE_METER]:
Use *aref_addr instead of AREF, since it needs to be an lvalue.
Although for decades ‘signal’ has been documented to not return,
a corner case in the Lisp debugger causes ‘signal’ to return.
Remove the corner case and adjust Emacs internals accordingly.
An alternative would be to document the corner case, but this
would complicate the Lisp API unnecessarily. (Bug#24047)
* src/eval.c (signal_or_quit): New function, with most of the
old contents of Fsignal.
(quit): New function, which uses signal_or_quit and which
might return. All keyboard-based callers of Fsignal (Qquit,
Qnil) changed to use this new function instead.
(Fsignal): Use signal_or_quit. Now _Noreturn. All callers
changed.
(xsignal): Move to lisp.h.
* src/lisp.h (xsignal): Now an inline function, as it's now
just an alias for Fsignal.
The new approach is less selective, but also less intrusive on C code.
* src/bytecode.c, src/image.c, src/keyboard.c: Ignore -Wclobbered.
* src/conf_post.h (NONVOLATILE): Remove. All uses removed.
It is no longer needed now that we assume GC_MARK_STACK ==
GC_MAKE_GCPROS_NOOPS.
* src/bytecode.c (BYTE_MAINTAIN_TOP): Remove.
All uses removed, and code simplified accordingly.
(BEFORE_POTENTIAL_GC, AFTER_POTENTIAL_GC): Remove, since they
are always no-ops now. All uses removed.
(MAYBE_GC): Remove. All uses replaced by maybe_gc,
since it is now equivalent.
* src/conf_post.h (GNUC_PREREQ): New macro.
Change uses of __GNUC_MINOR__ and __GNUC_PATCHLEVEL__ to use this
macro instead, for clarity and consistency.
(PRINTF_ARCHETYPE): New macro. Define it to __gnu_printf__ only
if glibc, since non-GNU platforms don’t necessarily support GNU
printf formats.
(ATTRIBUTE_FORMAT_PRINTF): Use it.
* src/bytecode.c (get_byte_code_arity): Omit unnecessary
runtime test for integer argument, unless debugging.
Use EMACS_INT for Emacs integers.
* src/eval.c (Ffunc_arity): Omit unused locals.
Avoid side effects in ‘if’ expr.
(lambda_arity): Use bool for boolean, and EMACS_INT for Emacs ints.
