These bugs were introduced after bignums were added.
* src/data.c (cons_to_unsigned, cons_to_signed):
* src/xdisp.c (calc_line_height_property):
Invoke integer_to_intmax and integer_to_uintmax only on integers.
* src/data.c (emacs_mpz_mul_2exp): 2nd arg is now a
nonnegative EMACS_INT not mp_bitcnt_t, to simplify checking.
(Fash): Support COUNT values that are bignums or that
exceed mp_bitcnt_t range.
* test/src/data-tests.el (data-tests-ash-lsh): Test this.
* src/bignum.c (make_neg_biguint): New function.
* src/lread.c (read1): Do not mishandle an unquoted symbol
with name equal to something like "1\0x", i.e., a string
of numeric form followed by a NUL byte.
Formerly these symbols were misread as numbers.
(string_to_number): Change last argument from an integer flag
to a pointer to the length. This lets the caller figure out
how much of the prefix was used. All callers changed.
Add a fast path if the integer (sans sign) fits in uintmax_t.
Update comments and simplify now that bignums are present.
* src/print.c (print_object): Fix quoting of symbols that look
like numbers, by relying on string_to_number for the tricky
cases rather than trying to redo its logic, incorrectly. For
example, (read (prin1-to-string '\1e+NaN)) formerly returned
"1e+NaN", which was wrong: a backslash is needed in the output
to prevent it from being read as a NaN. Escape NO_BREAK_SPACE
too, since lread.c treats it like SPACE.
* test/src/print-tests.el (print-read-roundtrip):
Add tests illustrating the abovementioned bugs.
* src/bignum.c (mpz_bufsize): New function.
(bignum_bufsize): Use it.
(mpz_get_d_rounded): New function.
(bignum_to_double): Use it.
* src/bignum.c (bignum_to_double):
* src/data.c (bignum_arith_driver):
When converting bignums to double, round instead of
truncating, to be consistent with what happens with fixnums.
* test/src/floatfns-tests.el (bignum-to-float): Test rounding.
This better corresponds to what emacs-26 did in the
rare cases where it checked for integer overflow.
* src/alloc.c (range_error): Remove.
All uses changed to overflow_error.
* src/eval.c (overflow_error): New function.
Improve arithmetic performance by avoiding bignums until needed.
Also, simplify bignum memory management, fixing some unlikely leaks.
This patch improved the performance of (+ 2 2) by a factor of ten
on a simple microbenchmark computing (+ x 2), byte-compiled,
with x a local variable initialized to 2 via means the byte
compiler could not predict: performance improved from 135 to 13 ns.
The platform was Fedora 28 x86-64, AMD Phenom II X4 910e.
Performance also improved 0.6% on ‘make compile-always’.
* src/bignum.c (init_bignum_once): New function.
* src/emacs.c (main): Use it.
* src/bignum.c (mpz): New global var.
(make_integer_mpz): Rename from make_integer. All uses changed.
* src/bignum.c (double_to_bignum, make_bignum_bits)
(make_bignum, make_bigint, make_biguint, make_integer_mpz):
* src/data.c (bignum_arith_driver, Frem, Flogcount, Fash)
(expt_integer, Fadd1, Fsub1, Flognot):
* src/floatfns.c (Fabs, rounding_driver, rounddiv_q):
* src/fns.c (Fnthcdr):
Use mpz rather than mpz_initting and mpz_clearing private
temporaries.
* src/bignum.h (bignum_integer): New function.
* src/data.c (Frem, Fmod, Fash, expt_integer):
* src/floatfns.c (rounding_driver):
Use it to simplify code.
* src/data.c (FIXNUMS_FIT_IN_LONG, free_mpz_value):
Remove. All uses removed.
(floating_point_op): New function.
(floatop_arith_driver): New function, with much of the guts
of the old float_arith_driver.
(float_arith_driver): Use it.
(floatop_arith_driver, arith_driver):
Simplify by assuming NARGS is at least 2.
All callers changed.
(float_arith_driver):
New arg, containing the partly converted value of the next arg.
Reorder args for consistency. All uses changed.
(bignum_arith_driver): New function.
(arith_driver): Use it. Do fixnum-only integer calculations
in intmax_t instead of mpz_t, when they fit.
Break out mpz_t calculations into bignum_arith_driver.
(Fquo): Use floatop_arith_driver instead of float_arith_driver,
since the op is known to be valid.
(Flogcount, Fash): Simplify by coalescing bignum and fixnum code.
(Fadd1, Fsub1): Simplify by using make_int.
* src/bignum.c, src/bignum.h: New files. Only modules that
need to know how bignums are implemented should include
bignum.h. Currently these are alloc.c, bignum.c (of course),
data.c, emacs.c, emacs-module.c, floatfns.c, fns.c, print.c.
* src/Makefile.in (base_obj): Add bignum.o.
* src/alloc.c (make_bignum_str): Move to bignum.c.
(make_number): Remove; replaced by bignum.c’s make_integer.
All callers changed.
* src/conf_post.h (ARG_NONNULL): New macro.
* src/json.c (json_to_lisp): Use it.
* src/data.c (Fnatnump):
Move NATNUMP’s implementation here from lisp.h.
* src/data.c (Fnumber_to_string):
* src/editfns.c (styled_format):
Move conversion of string to bignum to bignum_to_string, and
call it here.
* src/emacs-module.c (module_make_integer):
* src/floatfns.c (Fabs):
Simplify by using make_int.
* src/emacs.c: Include bignum.h, to expand its inline fns.
* src/floatfns.c (Ffloat): Simplify by using XFLOATINT.
(rounding_driver): Simplify by using double_to_bignum.
(rounddiv_q): Clarify use of temporaries.
* src/lisp.h: Move decls that need to know bignum internals to
bignum.h. Do not include gmp.h or mini-gmp.h; that is now
bignum.h’s job.
(GMP_NUM_BITS, struct Lisp_Bignum, XBIGNUM, mpz_set_intmax):
Move to bignum.h.
(make_int): New function.
(NATNUMP): Remove; all callers changed to use Fnatnump.
(XFLOATINT): If arg is a bignum, use bignum_to_double, so that
bignum internals are not exposed here.
* src/print.c (print_vectorlike): Use SAFE_ALLOCA to avoid the
need for a record_unwind_protect_ptr.
Also, since Emacs historically reported a range error when
rounding operations overflowed, do that consistently for all
bignum overflows.
* doc/lispref/errors.texi (Standard Errors):
* doc/lispref/numbers.texi (Integer Basics): Document range errors.
* src/alloc.c (range_error): Rename from integer_overflow.
All uses changed.
* src/floatfns.c (rounding_driver): When the result of a floating
point rounding operation does not fit into a fixnum, put it
into a bignum instead of always signaling an range error.
* test/src/floatfns-tests.el (divide-extreme-sign):
These tests now return the mathematically-correct answer
instead of signaling an error.
(bignum-round): Check that integers round to themselves.
* doc/lispref/objects.texi (Integer Type): Mention
most-negative-fixnum and most-positive-fixnum as alternatives
to fixnump and bignump.
* lisp/subr.el (fixnump, bignump): Now written in Lisp.
* src/data.c (Ffixnump, Fbignump): No longer written in C,
as these new functions are not crucial for performance.
libgmp calls ‘abort’ when given numbers too big for its
internal data structures. The numeric limit is large and
platform-dependent; with 64-bit GMP 6.1.2 it is around
2**2**37. Work around the problem by refusing to call libgmp
functions with arguments that would cause an abort. With luck
libgmp will have a better way to do this in the future.
Also, introduce a variable integer-width that lets the user
control how large bignums can be. This currently defaults
to 2**16, i.e., it allows bignums up to 2**2**16. This
should be enough for ordinary computation, and should
help Emacs to avoid thrashing or hanging.
Problem noted by Pip Cet (Bug#32463#71).
* doc/lispref/numbers.texi, etc/NEWS:
Document recent bignum changes, including this one.
Improve documentation for bitwise operations, in the light
of bignums.
* src/alloc.c (make_number): Enforce integer-width.
(integer_overflow): New function.
(xrealloc_for_gmp, xfree_for_gmp):
Move here from emacs.c, as it's memory allocation.
(init_alloc): Initialize GMP here, rather than in emacs.c.
(integer_width): New var.
* src/data.c (GMP_NLIMBS_MAX, NLIMBS_LIMIT): New constants.
(emacs_mpz_size, emacs_mpz_mul)
(emacs_mpz_mul_2exp, emacs_mpz_pow_ui): New functions.
(arith_driver, Fash, expt_integer): Use them.
(expt_integer): New function, containing integer code
that was out of place in floatfns.c.
(check_bignum_size, xmalloc_for_gmp): Remove.
* src/emacs.c (main): Do not initialize GMP here.
* src/floatfns.c (Fexpt): Use expt_integer, which
now contains integer code moved from here.
* src/lisp.h (GMP_NUMB_BITS): Define if gmp.h doesn’t.
* src/alloc.c (mpz_set_intmax_slow): Tighten assertion.
Work even in the unlikely case where libgmp uses nails.
* src/data.c (FIXNUMS_FIT_IN_LONG): New constant.
(arith_driver): Use it to tighten compile-time checks.
* src/lisp.h (mpz_set_intmax): Do not assume that converting
an out-of-range value to ‘long’ is harmless, as it might raise
a signal. Use simpler expression; compiler can optimize.
* src/data.c (isnan): Remove, as we can assume C99.
(bignumcompare): Remove, folding its functionality
into arithcompare.
(arithcompare): Compare bignums directly here.
Fix bugs when comparing NaNs to bignums.
When comparing a bignum to a fixnum, just look at the
bignum’s sign, as that’s all that is needed.
Decrease scope of locals when this is easy.
* test/src/data-tests.el (data-tests-bignum): Test bignum vs NaN.
* doc/lispref/numbers.texi (Bitwise Operations): Document that
the traditional (lsh A B) behavior is for fixnums, and that it
is an error if A and B are both negative and A is a bignum.
See Bug#32463.
* lisp/subr.el (lsh): New function, moved here from src/data.c.
* src/data.c (ash_lsh_impl): Remove, moving body into Fash
since it’s the only caller now.
(Fash): Check for out-of-range counts. If COUNT is zero,
return first argument instead of going through libgmp. Omit
lsh code since lsh is now done in Lisp. Add code for shifting
fixnums right, to avoid a round trip through libgmp.
(Flsh): Remove; moved to lisp/subr.el.
* test/lisp/international/ccl-tests.el (shift):
Test for traditional lsh behavior, instead of assuming
lsh is like ash when bignums are present.
* test/src/data-tests.el (data-tests-logand)
(data-tests-logior, data-tests-logxor, data-tests-ash-lsh):
New tests.
These operations incorrectly treated negative fixnums as
bignums greater than most-positive-fixnum.
* src/alloc.c (mpz_set_intmax_slow): Avoid undefined
behavior if signed unary negation overflows, while
we’re in the neighborhood.
(mpz_set_uintmax_slow): Remove. All uses removed.
* src/data.c (arith_driver): Treat fixnums as signed, not
unsigned, even for logical operations.
* src/lisp.h (mpz_set_uintmax): Remove. All uses removed.
* test/src/data-tests.el (data-tests-logand)
(data-tests-logior, data-tests-logxor): New tests.
Problem reported by Eli Zaretskii in:
https://lists.gnu.org/r/emacs-devel/2018-08/msg00380.html
* src/data.c (arithcompare): Work around incompatibility
between gcc -fexcess-precision=fast and the C standard on x86,
by capturing the results of floating-point comparisons before
the excess precision spontaneously decays. Although this fix
might not work in general, it does work here and is probably
good enough for the platforms we care about.
(cherry picked from commit a84cef9095)
Problem reported by Eli Zaretskii in:
https://lists.gnu.org/r/emacs-devel/2018-08/msg00380.html
* src/data.c (arithcompare): Work around incompatibility
between gcc -fexcess-precision=fast and the C standard on x86,
by capturing the results of floating-point comparisons before
the excess precision spontaneously decays. Although this fix
might not work in general, it does work here and is probably
good enough for the platforms we care about.
Eliminate the category of miscellaneous objects, and turn all
such objects into pseudovectors. The immediate motivation
for this change is to free up an enum Lisp_Type tag value, a
scarce resource that can be better used elsewhere. However,
this change is worthwhile in its own right, as it improves
performance slightly on my platform, 0.3% faster for 'make
compile-always' on Fedora 28, and it simplifies the garbage
collector and interpreter (Bug#32405).
* doc/lispref/internals.texi (Garbage Collection):
* etc/NEWS:
Document change to garbage-collect return value.
* src/alloc.c (total_markers, total_free_markers):
(union aligned_Lisp_Misc, MARKER_BLOCK_SIZE)
(struct marker_block, marker_block, marker_block_index)
(misc_free_list, allocate_misc, live_misc_holding)
(live_misc_p, sweep_misc):
* src/lisp.h (lisp_h_MARKERP, lisp_h_MISCP, MARKERP, MISCP)
(Lisp_Misc, enum Lisp_Misc_Type, Lisp_Misc_Free)
(Lisp_Misc_Marker, Lisp_Misc_Overlay, Lisp_Misc_Finalizer)
(Lisp_Misc_Ptr, Lisp_Misc_User_Ptr, Lisp_Misc_Limit)
(Lisp_Misc_Bignum)
(XSETMISC, struct Lisp_Misc_Any, XMISCANY, XMISCTYPE)
(struct Lisp_Free, union Lisp_Misc, XMISC):
Remove. All uses removed.
(cleanup_vector): Clean up objects that were formerly misc
and are now pseudovectors.
(make_misc_ptr, build_overlay, Fmake_marker, build_marker)
(make_bignum_str, make_number, make_pure_bignum)
(make_user_ptr, Fmake_finalizer):
Build as pseudovectors, not as misc objects.
(mark_finalizer_list, queue_doomed_finalizers)
(compact_undo_list, mark_overlay, mark_object)
(unchain_dead_markers):
Mark as vector-like objects, not as misc objects.
(mark_maybe_object, mark_maybe_pointer, valid_lisp_object_p)
(total_bytes_of_live_objects, survives_gc_p):
* src/fns.c (sxhash):
No need to worry about misc objects.
(garbage_collect_1): Do not generate a 'misc' component.
(syms_of_alloc): No need for 'misc' symbol.
* src/buffer.c (overlays_at, overlays_in, overlay_touches_p)
(overlay_strings, recenter_overlay_lists)
(fix_start_end_in_overlays, fix_overlays_before)
(Foverlay_lists, report_overlay_modification)
(evaporate_overlays):
* src/editfns.c (overlays_around):
* src/data.c (Ftype_of):
* src/fns.c (internal_equal):
* src/lisp.h (mint_ptrp, xmint_pointer, FINALIZERP)
(XFINALIZER, MARKERP, XMARKER, OVERLAYP, XOVERLAY, USER_PTRP)
(XUSER_PTR, BIGNUMP, XBIGNUM):
* src/print.c (print_vectorlike, print_object):
* src/undo.c (record_marker_adjustments):
* src/xdisp.c (load_overlay_strings):
Formerly misc objects are now pseudovectors.
* src/lisp.h (PVEC_MARKER, PVEC_OVERLAY, PVEC_FINALIZER)
(PVEC_BIGNUM, PVEC_MISC_PTR, PVEC_USER_PTR):
New constants, replacing their misc versions. All uses changed.
(struct Lisp_Marker, struct Lisp_Overlay, struct Lisp_Misc_Ptr)
(struct Lisp_Bignum, struct Lisp_User_Ptr, struct Lisp_Finalizer):
Make usable as a pseudovector by using a pseudovector header,
replacing any DIY components, and putting Lisp_Object members
first. All uses changed.
* lisp/international/ccl.el (ccl-fixnum): New function.
(ccl-embed-data, ccl-embed-current-address, ccl-dump): Use it.
* src/alloc.c (make_number): Handle case where EMACS_INT is
larger than long.
* src/data.c (bignumcompare): Handle case where EMACS_INT is
larger than long.
(arith_driver): Likewise. Coerce markers.
(float_arith_driver): Coerce markers.
(Flogcount): Use mpz_sgn.
(ash_lsh_impl): Fix bugs.
(Fsub1): Fix underflow check.
* src/lisp.h (NUMBERP): Don't check BIGNUMP.
(CHECK_FIXNUM_OR_FLOAT_COERCE_MARKER): Fix indentation.
* test/lisp/international/ccl-tests.el: New file.
* src/data.c (set_default_internal): Use FOR_EACH_LIVE_BUFFER
when binding variables that don't nominally have a local
value, to avoid slowing down due to a large number of dead
buffers. (Bug#18522) (Bug#31853)
