* doc/lispref/internals.texi (Stack-allocated Objects):
Adjust to match the revised, less error-prone macros.
* src/frame.h (AUTO_FRAME_ARG): Rename from FRAME_PARAMETER.
* src/lisp.h (AUTO_CONS): Rename from scoped_cons.
(AUTO_LIST1): Rename from scoped_list1.
(AUTO_LIST2): Rename from scoped_list2.
(AUTO_LIST3): Rename from scoped_list3.
(AUTO_LIST4): Rename from scoped_list4.
(AUTO_STRING): Rename from SCOPED_STRING.
* src/frame.h (AUTO_FRAME_ARG):
* src/lisp.h (AUTO_CONS, AUTO_LIST1, AUTO_LIST2, AUTO_LIST3)
(AUTO_LIST4, AUTO_STRING):
Prepend a new argument 'name'.
Declare a variable instead of yielding a value.
All uses changed.
* src/lisp.h (STACK_CONS, AUTO_CONS_EXPR): New internal macros.
The build_local_string macro was used in two ways: (1) string
literals for which scoped allocation suffices, and (2) file name
components, where it's not safe in general to assume bounded-size
ASCII data. Simplify by defining a new macro SCOPED_STRING that
allocates a block-scope string, and by using SCOPED_STRING for (1)
and build_string for (2). Furthermore, actually use stack
allocation only for objects known to have sufficient alignment.
This simpler implementation means Emacs can make
USE_STACK_LISP_OBJECTS the default unless GC_MARK_STACK !=
GC_MAKE_GCPROS_NOOPS.
* lisp.h (GCALIGNED): Align even if !USE_STACK_LISP_OBJECTS,
for fewer differences among implementations.
(struct Lisp_String): Now GCALIGNED.
(USE_STACK_LISP_OBJECTS): Default to true, since the
implementation no longer insists on a nonempty GCALIGNED.
But make it false if GC_MARK_STACK != GC_MAKE_GCPROS_NOOPS.
(SCOPED_CONS_INITIALIZER): Remove, since it's no longer needed
separately. Move definiens to scoped_cons. The old definition
was incorrect when GCALIGNED was defined to be empty.
(union Aligned_String): New type.
(USE_STACK_CONS, USE_STACK_STRING): New constants, so that the
implementation ports to compilers that don't align strictly enough.
Don't worry about the union sizes; it's not worth bothering about.
(scoped_cons, scoped_list1, scoped_list3, scoped_list4):
Rewrite using USE_STACK_CONS.
(scoped_cons): Assume the use of union Aligned_Cons.
(lisp_string_size, make_local_string, build_local_string): Remove.
Unless otherwise specified, all callers of build_local_string
changed to use SCOPED_STRING.
(SCOPED_STRING): New macro.
* data.c (wrong_choice):
* menu.c (single_menu_item):
* process.c (Fformat_network_address):
Hoist use of SCOPED_STRING out of a scope, so that its returned
object lives long enough.
* fileio.c (Fexpand_file_name): Use build_string, not SCOPED_STRING,
as the string might be long or might not be ASCII.
* configure.ac (HAVE_STATEMENT_EXPRESSIONS): Remove.
For USE_STACK_LISP_OBJECTS, we always assume __GNUC__.
* lisp.h (union Aligned_Cons) [!GCALIGNED]: Define as such.
(SCOPED_CONS_INITIALIZER): New macro.
(scoped_cons) [USE_STACK_LISP_OBJECTS]: Use it.
(USE_LOCAL_ALLOCA): Remove.
(local_cons, local_list1, local_list2, local_list3, local_list4):
Remove. Stack overflow checking makes them too slow.
(make_local_vector): Likewise. Also we just don't have enough
users for it.
(enum LISP_STRING_OVERHEAD): Remove.
(local_string_init, local_vector_init): Remove prototypes.
(make_local_string, build_local_string): Redesign to target short
compile-time string constants, fall back to regular string allocation
where appropriate.
(lisp_string_size): New function.
(verify_ascii) [ENABLE_CHECKING]: Add prototype.
* alloc.c (local_string_init, local_vector_init): Remove.
(verify_ascii) [ENABLE_CHECKING]: New function.
* buffer.c, charset.c, chartab.c, data.c, editfns.c, emacs.c, fileio.c:
* fns.c, font.c, fontset.c, frame.c, keyboard.c, keymap.c, lread.c:
* menu.c, minibuf.c, process.c, textprop.c, xdisp.c, xfns.c, xfont.c:
* xselect.c, xterm.c: All related users changed.
src/data.c (set_internal): Use assq_no_quit, not Fassq, to find an
existing binding of a variable, to avoid silently aborting
commands that use specbind.
* src/fileio.c (CHECK_LENGTH): Remove.
Rewrite callers so that they don't need it.
(Fexpand_file_name) [DOS_NT]: Fix a case where directory length
variable wasn't set.
There is certainly nothing wrong with writing code like 'lo <= i
&& i <= hi', even if LO happens to a constant. There isn't even
anything wrong in general with writing 'a <= b' if A happens to
be a constant. At any rate stylistic changes shouldn't
be done like this without discussion.
* keyboard.c (Qvertical_scroll_bar):
* frame.c (Qleft, Qright): Move to ...
* buffer.c (Qleft, Qright, Qvertical_scroll_bar): ... here.
* buffer.c (Qchoice, Qrange, Qoverwrite_mode, Qfraction): New symbols.
(syms_of_buffer): DEFSYM all of the above, attach special properties.
Use special symbols to DEFVAR_PER_BUFFER overwrite-mode,
vertical-scroll-bar, scroll-up-aggressively
and scroll-down-aggressively.
* buffer.h (Qchoice, Qrange, Qleft, Qright, Qvertical_scroll_bar):
Add declarations.
* nsfns.m, frame.h (Qleft, Qright):
* nsterm.m (Qleft): Remove declarations.
* gtkutil.c (toplevel): Include buffer.h.
* data.c (wrong_choice, wrong_range): New functions.
(store_symval_forwarding): Handle special properties of buffer-local
variables and use functions from the above to signal error, if any.
Likewise for xd_signature in dbusbind.c.
This is more consistent with the usual practice in Emacs, which is
that any non-nil value counts as true.
* doc/lispref/sequences.texi (Bool-Vectors): Coalesce discussion of how to
print them. bool-vector's args need not be t or nil.
* src/alloc.c (Fbool_vector): Don't require args to be t or nil.
* src/dbusbind.c (xd_signature): Likewise, for booleans.
* src/data.c, lisp.h (Qbooleanp):
* src/lisp.h (CHECK_BOOLEAN): Remove. All uses removed.
This ports to GCC 4.9.0 with -fsanitize=undefined.
* alloc.c (bool_vector_fill, SETMARKBIT, UNSETMARKBIT):
* data.c (Fash):
* regex.c (extract_number):
* lisp.h (make_number, XINT):
Do not shift a 1 bit left into a sign bit.
* alloc.c (struct cons_block, struct float_block): Use unsigned,
not int, for gcmarkbits. All uses changed.
This port requires IRIX cc, as I did not have time to get
undump working with the old GCC on the system I had access to,
but that's better than nothing.
* configure.ac (gl_GCC_VERSION_IFELSE): Remove unused macro
that wouldn't have worked anyway, with IRIX cc.
(emacs_cv_clang, emacs_cv_sanitize_address)
(ns_osx_have_104, ns_osx_have_105):
Don't assume '#error' makes the compiler fail,
as this doesn't work with IRIX cc.
(CFLAGS, LIBS): Don't let the GnuTLS results infect later 'configure'
checks. This runs afoul of an IRIX configuration where GnuTLS is
in an optional library that also contains getdelim, and causes
a later 'configure' to incorrectly think getdelim is supported.
* src/alloc.c (TAGGABLE_NULL): New constant,
for porting to hosts with nontrivial DATA_SEG_BITS settings.
(next_vector, set_next_vector): Use it.
* src/conf_post.h (INET6) [IRIX6_5]: Define.
(HAVE_GETADDRINFO) [IRIX6_5]: Undef.
* src/data.c (BITS_PER_ULL): Don't assume ULLONG_MAX is defined.
* src/lisp.h (lisp_h_XPNTR): Don't OR in bits that aren't masked out,
for consistency with how TAGGABLE_NULL is computed.
Fixes: debbugs:9684
* doc/lispref/searching.texi (String Search): Incremental word search fixes.
* lisp/sort.el (delete-duplicate-lines): Remove `:weakness 'key'
from `make-hash-table'.
* lisp/textmodes/ispell.el (ispell-init-process): Change message format
to be consistent with other messages.
* data.c (arithcompare_driver):
* fileio.c (Fcar_less_than_car):
* fns.c (internal_equal):
* frame.c (delete_frame):
* lisp.h (enum More_Lisp_Bits):
* lread.c (read1):
Avoid C99 constructs that don't work in C89.
* data.c (ULL_MAX, count_trailing_zeros_ll): New macros,
to port to C89, which doesn't have 'long long'.
(count_trailing_zero_bits): Use them.
This was due to an optimization that went awry.
Reported by Glenn Morris in
<http://lists.gnu.org/archive/html/emacs-devel/2013-11/msg00622.html>.
* alloc.c (make_uninit_bool_vector): Don't allocate a dummy word
for empty vectors, undoing the 2013-11-18 change.
* data.c (bool_vector_binop_driver): Rely on this.
Fix bug that occasionally overran the destination.
* lisp.h (struct Lisp_Bool_vector): Document this.
The old API had (bool-vector-count-matches A B)
and (bool-vector-count-matches-at A B I), which gave the
misleading impression that the two functions were variants, one
with a location I. The new API has (bool-vector-count-population A)
and (bool-vector-count-consecutive A B I) to make the distinction
clearer. The first function no longer has a B argument, since the
caller can easily determine the number of nils if the length and
number of ts is known.
* src/data.c (Fbool_vector_count_population): Rename from
bool_vector_count_matches, and accept just 1 argument.
(Fbool_vector_count_consecutive): Rename from
Fbool_vector_count_matches_at.
* test/automated/data-tests.el: Adjust to API changes.
Fixes: debbugs:15912
See Daniel Colascione in:
http://lists.gnu.org/archive/html/emacs-devel/2013-11/msg00518.html
* alloc.c (make_uninit_bool_vector): Always allocate at least one word.
* data.c (bool_vector_binop_driver): Rely on this. Tune.
* lisp.h (struct Lisp_Bool_vector): Document this.
* configure.ac (BITSIZEOF_SIZE_T, SIZEOF_SIZE_T): Remove.
* src/alloc.c (bool_vector_exact_payload_bytes)
(bool_vector_payload_bytes): Remove.
(bool_vector_fill): Return its argument.
* src/alloc.c (bool_vector_fill):
* src/lread.c (read1):
* src/print.c (print_object):
Simplify by using bool_vector_bytes.
* src/alloc.c (make_uninit_bool_vector):
New function, broken out from Fmake_bool_vector.
(Fmake_bool_vector): Use it. Use tail call.
(make_uninit_bool_vector, vector_nbytes): Simplify size calculations.
* src/data.c (BITS_PER_ULL): New constant.
(ULLONG_MAX, count_one_bits_ll): Fall back on long counterparts
if long long versions don't exist.
(shift_right_ull): New function.
(count_one_bits_word): New function, replacing popcount_bits_word
macro. Don't assume that bits_word is no wider than long long.
(count_one_bits_word, count_trailing_zero_bits):
Don't assume that bits_word is no wider than long long.
* src/data.c (bool_vector_binop_driver, bool_vector_not):
* src/fns.c (Fcopy_sequence):
* src/lread.c (read1):
Create an uninitialized destination, to avoid needless work.
(internal_equal): Simplify.
(Ffillarray): Prefer tail call.
* src/data.c (bool_vector_binop_driver): Don't assume bit vectors always
contain at least one word.
(bits_word_to_host_endian): Prefer if to #if. Don't assume
chars are narrower than ints.
* src/data.c (Fbool_vector_count_matches, Fbool_vector_count_matches_at):
* src/fns.c (Fcopy_sequence):
Simplify and tune.
* src/lisp.h (bits_word, BITS_WORD_MAX, BITS_PER_BITS_WORD):
Don't try to port to hosts where bits_word values have holes; the
code wouldn't work there anyway. Verify this assumption, though.
(bool_vector_bytes): New function.
(make_uninit_bool_vector): New decl.
(bool_vector_fill): Now returns Lisp_Object.
* configure.ac (BITSIZEOF_SIZE_T, SIZEOF_SIZE_T):
New symbols to configure.
* src/alloc.c (ROUNDUP): Move here from lisp.h, since it's now used
only in this file. Use a more-efficient implementation if the
second argument is a power of 2.
(ALIGN): Rewrite in terms of ROUNDUP. Make it a function.
Remove no-longer-necessary compile-time checks.
(bool_vector_exact_payload_bytes): New function.
(bool_vector_payload_bytes): Remove 2nd arg; callers that need
exact payload changed to call the new function. Do not assume
that the arg or result fits in ptrdiff_t.
(bool_vector_fill): New function.
(Fmake_bool_vector): Use it. Don't assume bit counts fit
in ptrdiff_t.
(vroundup_ct): Don't assume arg fits in size_t.
* src/category.c (SET_CATEGORY_SET): Remove. All callers now just
invoke set_category_set.
(set_category_set): 2nd arg is now EMACS_INT and 3rd is now bool.
All callers changed. Use bool_vector_set.
* src/category.h (XCATEGORY_SET): Remove; no longer needed.
(CATEGORY_MEMBER): Now a function. Rewrite in terms of
bool_vector_bitref.
* src/data.c (Faref): Use bool_vector_ref.
(Faset): Use bool_vector_set.
(bits_word_to_host_endian): Don't assume you can shift by CHAR_BIT.
(Fbool_vector_not, Fbool_vector_count_matches)
(Fbool_vector_count_matches_at): Don't assume CHAR_BIT == 8.
* src/fns.c (concat): Use bool_vector_ref.
(Ffillarray): Use bool_vector_fill.
(mapcar1): Use bool_vector_ref.
(sxhash_bool_vector): Hash words, not bytes.
* src/lisp.h (BOOL_VECTOR_BITS_PER_CHAR): Now a macro as well as
a constant, since it's now used in #if.
(bits_word, BITS_WORD_MAX, BITS_PER_BITS_WORD): Fall back on
unsigned char on unusual architectures, so that we no longer
assume that the number of bits per bits_word is a power of two or
is a multiple of 8 or of CHAR_BIT.
(Qt): Add forward decl.
(struct Lisp_Bool_Vector): Don't assume EMACS_INT is aligned
at least as strictly as bits_word.
(bool_vector_data, bool_vector_uchar_data): New accessors.
All data structure accesses changed to use them.
(bool_vector_words, bool_vector_bitref, bool_vector_ref)
(bool_vector_set): New functions.
(bool_vector_fill): New decl.
(ROUNDUP): Move to alloc.c as described above.
* data.c (bool_vector_spare_mask, Fbool_vector_count_matches)
(Fbool_vector_count_matches_at):
Use EMACS_INT, not ptrdiff_t, to record bit counts, as a bit count
can exceed PTRDIFF_MAX, at least in theory.
(Fbool_vector_count_matches_at):
Use int, not ptrdiff_t, to record a value that can't exceed INT_MAX.
The old behavior left garbage in the result vector sometimes,
and didn't seem to be useful.
* data.c (Qwrong_length_argument): New static var.
(wrong_length_argument): New function.
(bool_vector_binop_driver): Check that args agree in length.
All uses of XBOOL_VECTOR (x)->size changed to bool_vector_size (x).
* data.c (bool_vector_spare_mask, bool_vector_binop_driver)
(Fbool_vector_not, Fbool_vector_count_matches_at):
Remove uses of 'eassume' that should no longer be needed,
because they are subsumed by the 'eassume' in bool_vector_size.
Also, include <verify.h>, for 'assume'.
* alloc.c (bool_vector_payload_bytes, Fmake_bool_vector)
(vroundup, vector_nbytes):
* data.c (bool_vector_spare_mask, bool_vector_binop_driver)
(Fbool_vector_not, Fbool_vector_count_matches)
(Fbool_vector_count_matches_at):
Use eassume, not eassert.
* casetab.c (set_identity, shuffle):
* composite.c (composition_gstring_put_cache):
* dispnew.c (update_frame_1):
* ftfont.c (ftfont_shape_by_flt):
* image.c (gif_load):
* intervals.c (offset_intervals):
* macfont.m (macfont_shape):
Remove calls to 'assume' that are no longer needed, because
--enable-gcc-warnings no longer generates bogus warnings
when these calls are removed.
All uses of 'size_t' and 'SIZE_MAX' changed to use them, when
they're talking about words in Lisp bool vectors.
(BITS_PER_BITS_WORD): Rename from BITS_PER_SIZE_T. All uses changed.
Do this by using the Gnulib modules for this.
This should generate faster code on non-GCC, non-MSC platforms,
and make the code a bit more portable, at least in theory.
* admin/merge-gnulib (GNULIB_MODULES): Add count-one-bits
and count-trailing-zeros.
* lib/count-one-bits.c, lib/count-one-bits.h:
* lib/count-trailing-zeros.c, lib/count-trailing-zeros.h:
* m4/count-one-bits.m4, m4/count-trailing-zeros.m4:
New files, copied from gnulib.
* lib/gnulib.mk, m4/gnulib-comp.m4: Regenerate.
* nt/gnulib.mk: Merge changes from lib/gnulib.mk.
* src/data.c: Include <count-one-bits.h>, <count-trailing-zeros.h>.
(USE_MSC_POPCOUNT, POPCOUNT_STATIC_INLINE)
(NEED_GENERIC_POPCOUNT, popcount_size_t_generic)
(popcount_size_t_msc, popcount_size_t_gcc):
Remove; now done by Gnulib.
(popcount_size_t): Now a macro that defers to Gnulib.
(count_trailing_zero_bits): Return int, for consistency with
Gnulib and because Emacs prefers signed to unsigned int.
Don't assume that size_t is either unsigned int or unsigned long
or unsigned long long.
(size_t_to_host_endian): Do not assume that size_t is either
exactly 32 or exactly 64 bits wide.
* src/lisp.h (BITS_PER_SIZE_T): Define consistently with BITS_PER_LONG
etc., so that it's now an enum constant, not a macro.
No need to assume that it's either 32 or 64.
Fixes: debbugs:15550
In other words, take on the behavior of eassert_and_assume.
This makes Emacs 0.2% smaller on my platform (Fedora 19, x86-64).
(eassert_and_assume): Remove. All uses replaced by eassert.
This is ugly, but it should fix the performance problem for older
GCC versions in the short run. I'll look into integrating the
Gnulib module for popcount, as a better fix.
See the thread starting in:
http://lists.gnu.org/archive/html/emacs-devel/2013-09/msg00474.html
(popcount_size_t_generic) [NEED_GENERIC_POPCOUNT]:
(popcount_size_t_msc) [USE_MSC_POPCOUNT]:
(popcount_size_t_gcc) [USE_GCC_POPCOUNT]:
(popcount_size_t): Use it.
* conf_post.h (assume): Always return void. Use lint version
only if GCC and MSC versions don't apply.
* conf_post.h (assume):
* data.c (USC_MSC_POPCOUNT, count_trailing_zero_bits):
Depend on _MSC_VER, not __MSC_VER, for consistency with
the rest of Emacs.
* data.c (bool_vector_spare_mask, popcount_size_t_generic)
(popcount_size_t_msc, popcount_size_t_gcc, popcount_size_t)
(bool_vector_binop_driver, count_trailing_zero_bits)
(size_t_to_host_endian): Now static, not static inline;
the latter isn't needed with modern compilers and doesn't
work with older compilers anyway.
