to get rid of the bias introduced by making the hash never zero
… he also pointed out a memory (heap) usage optimisation… which
may impact code size a bit though as I’d need to pass an additional
argument on hashtable function calls… or, forgo the benefit of not
having to pointer-align the key in the structure, which can be as
much as 3/7 octets per item, heap storage… OTOH the saved space is
4/8 octets per not-allocated item, possibly some code (use of an
multiply-add opcode), but the function call overhead/cost would
possibly be quite a bit… I guess I’ll have to measure…
• promote SCO OpenServer and UnixWare to !oswarn
• omit trying -O2/-O on OpenServer 5 and USL C
• cast mksh_ari_t to int, mksh_uari_t to unsigned int for printf
• skip ulimit-1 on syllable (which is still too broken)
• write ((mksh_ari_t)-2147483648) ipv UB ((mksh_ari_t)1 << 31)
and add a comment that that is actually meant
• rewrite functions returning !void ending in NOTREACHED
so they’ve got a jump target returning an error at the
end, to aid older compilers and just to be safe
• cast struct stat.st_size to off_t or size_t explicitly when needed
• shorten struct env by two bytes and an alignment, at least
also, optimise control flow and fix more paren matching cases
is not found, from a suggestion by RT (LP: #912691)
• try harder (in a loop) to acquire a file lock if the locking mechanism
documents EINTR is a possibility (fcntl always, flock on Linux not .Ox)
• use -std=c99 not -std=gnu99 if it must be at all
deep changes into R40-stable branch
• Version accordingly: HEAD gets 2012/03/24 (hi Eddy/Chris) so we backdate
$KSH_VERSION for R40-stable to 2012/03/20 (hi David) as that comes before
even though it includes today’s latest fixes
• Also, sync clog (including MFC indicators)
(cid 100487B467E068A55D6 and 10048949D196A7C1390) discovered
by Jb_boin: time with a not-TCOM subtree would now trash its
string argument (which is the loop variable for TFOR); amend
regression testsuite
prerequisites are actually fulfilled, i.e. evrn further down than with
the last commit doing this, and move some prerequisites of stuff that
has wandered outside the !INCLUDES_ONLY block with the compile-time
assert changes to the outside, too
fixes FTBFS on MSYS which has neither <stdint.h> nor uint32_t
reported by RT
handle any more, octal 010 style constants, as promised
• overhaul the manpage re. arithmetic expressions, make the guarantees
mksh code has explicitly, precisely, clear
• to reduce burden of the compiler, getint() now operates on mksh_uari_t
internally; it already applied the sign after operation, anyway (C99
guarantees wraparound on unsigned types, but for signed types we need
specific compiler support; apparently, this comes from hardware limits)
• use const and shuffle order of locals around while here
• integer base flat (10, 010, 0x10)
‣ posix, right variants (heh, just like timezones)
⇒ both enabled, which means one test always fails right now,
until the code changes are in…
☹ mksh R21 and up, AT&T ksh93, posh, GNU bash, busybox sh, dash
all use the posux variant
☺ mksh R20b and below, pdksh, zsh, python, perl, php all do it right
‣ posix is stupid to use C (strongly typed, compiled) semantics
for an untyped _scripting_ language
• integer bases 1 (mksh specific), 2 to 36, and that 37 errors out
‣ both directions
• integer arithmetic, signed and unsigned, wraparound and value span
in the cases where they are defined unambiguously; bug reported by
Jilles Tjoelker in <20111129232526.GC14357@stack.nl> due to a report
by Stefano Lattarini on bug-autoconf
in the ambiguous case, I stick to traditional pdksh behaviour, which means
test ! a = b vs. test a = b
and
test ! a -o b vs. test a -o b
behave different from each other (in the second case, the NOT operator
binds strong; POSIX demands a reduction to 3 arguments and negating
that result in the first case), so we're at two known not-ok in the
FreeBSD® testsuite. (81 and 82 in regress.sh,v 1.3)
tg