Nick Bowler [Sat, 6 Jan 2024 21:51:25 +0000 (16:51 -0500)]
copysym: Improve compatibility with old compilers.
Some C89 compilers with 32-bit longs won't automatically widen integer
constants to (unsigned) long long in order to represent values that
don't fit in an (unsigned) long, even if those compilers otherwise
support 64-bit long long.
The "ll" and "ull" suffixes seem to get things working.
Nick Bowler [Sat, 6 Jan 2024 20:36:59 +0000 (15:36 -0500)]
test-tap.at: Use PATH instead of hardcoding "$builddir/t"
Hardcoding "$builddir/t" doesn't work at all if VPATH builds need to
run programs in "$srcdir", which we do now need. Generally PATH
lookups are more flexible in Autotest and work out much better.
Technically this is a change in behaviour so some downstream projects
may need to be updated.
Nick Bowler [Sat, 6 Jan 2024 07:34:10 +0000 (02:34 -0500)]
test-tap.at: Fix regression passing program arguments.
Moving the test_run_tap() function definition into a macro failed to
properly quote $@ so it isn't eaten by m4. So instead of calling
the test program with "$@" as the argument, it is changed to "".
Fix that up, and add a very simple test case to check this behaviour.
Nick Bowler [Wed, 3 Jan 2024 03:07:52 +0000 (22:07 -0500)]
DX_PROG_FLEX: redirect lex output to the log file.
AIX 7.2 includes a "lex" program which prints error messages to standard
output. So ensure that is also redirected to the log file, otherwise
we get a bunch of garbage printed to the terminal when configuring.
Nick Bowler [Mon, 1 Jan 2024 21:37:55 +0000 (16:37 -0500)]
gen-strtab.awk: Work around AIX substitution bug.
On AIX 7.2 awk, the sub and gsub functions have a bug when the
replacement string contains a "\1" character. The substituted
strings have \1 characters replaced with amperands, for example:
Very strange. The problem is internal to (g)sub, there is no
problem using "\1" characters in other contexts -- including
with the other arguments to (g)sub. It should be fine to use
"\2" instead which avoids this particular bug.
Nick Bowler [Mon, 1 Jan 2024 20:21:12 +0000 (15:21 -0500)]
DX_GNULIB_SYMFILES: Don't use "awk -f -".
Even though a filename of "-" is specified by POSIX to read the program
from standard input and works with almost every awk implementation out
there, it seems it does not work on AIX 7.2 awk (which tries to read the
program from a file named "-" instead). How wonderful.
In this case, it is easy enough to use a quoted string for the program
instead of a here-document.
Nick Bowler [Fri, 22 Dec 2023 05:28:08 +0000 (00:28 -0500)]
at-compat.at: Add patch to avoid DJGPP date issue.
DJGPP's "date +%s" does not actually work like GNU, although it probably
was supposed to. It can return a string which begins with zero. The
testsuite only checks that the output is all digits, so accepts this as
being a working GNU-like date output.
This is then used in an arithmetic expansion, and the shell interprets
this string as octal digits, crapping out if any of those digits are 8
or 9. If this happens (perhaps depending on the phase of the moon)
then the testsuite crashes before printing the final summary.
Nick Bowler [Thu, 7 Dec 2023 01:56:24 +0000 (20:56 -0500)]
DX_AUTOMAKE_COMPAT: Fix compile with newer Automake.
Current versions of Automake's compile script depend on shell functions
which are not supported by ULTRIX 4.5 /bin/sh.
Now ksh does support these and configure does find and use this shell
on ULTRIX, but compile starts with #!/bin/sh so when it is executed
directly, it is run with /bin/sh instead of the shell we want.
This is a bit tricky to work around since Automake updates $CC (if
needed) to refer to the compile script, which is used throughout
the configure run. So we can't just patch it up whenever, it has
to be done early on. AC_BEFORE can be used to at least warn if the
patches are ineffective.
Nick Bowler [Wed, 6 Dec 2023 04:00:42 +0000 (23:00 -0500)]
fix-gnulib.pl: Convert rules to dirstamp prerequisites.
ULTRIX 4.5 make fails during rule lookup when it tries to build a
target in a nonexistent directory that might have inference rules
associated with it (i.e., ends with a defined suffix), for example:
We can work around this problem by using a prerequisite to create
the output directory instead of creating it directly in the rule.
Automake itself actually does this already using dirstamps, so
let's try to patch things to use the same mechanism. As a bonus,
this should reduce the number of mkdir invocations.
This breaks some of the relevant tests which are looking for specific
output. Drop the %reldir% test completely as it appears every usage
in gnulib is for this exact scenario, and patch the subdirectory test
to expect the new output.
Nick Bowler [Wed, 6 Dec 2023 02:11:35 +0000 (21:11 -0500)]
glconfig.mk: Remove some redundant prerequisites.
Since the glsym rules are only invoked recursively by a rule that
already has all required dependencies, it is not necessary to have
prerequisites on the individual glsym files to the gnulib headers.
As it turns out these redundant prerequisites trip up "make" on
ULTRIX 4.5 due to a bug expanding $(gnulib_symfiles) (which I do
not fully understand), dropping it basically makes things work
just fine, so no real reason to hang onto it.
Nick Bowler [Wed, 6 Dec 2023 01:46:26 +0000 (20:46 -0500)]
DX_AUTOMAKE_COMPAT: Fix install-sh with newer Automake.
Current versions of Automake's install-sh depends on substitutions like
${var:-foo}, which are not supported by ULTRIX 4.5 /bin/sh.
Now, ksh does support these and configure does find and use this shell
on ULTRIX, but install-sh starts with #!/bin/sh so when it is executed
directly, it is run with /bin/sh instead of the shell we want.
We can work around the problem by patching up the MKDIR_P and INSTALL
assignments at configure time to include the known-good shell, which
seems to work well.
Nick Bowler [Wed, 6 Dec 2023 01:10:08 +0000 (20:10 -0500)]
help_print_optstring: Better output on some old systems.
Some very old printf implementations return 0 on success instead of the
number of bytes written. We should never see a return of 0 normally,
so we can improve the output to be less of a garbled mess by printing a
newline if that happens (same as the error case). This is a simple
tweak that should have virtually no impact on modern systems.
The tests are adapted to verify the new behaviour, although we'll mark
the result as "skipped" (assuming it is correct) instead of "passed"
since it's only sortof working.
Nick Bowler [Tue, 5 Dec 2023 02:37:43 +0000 (21:37 -0500)]
help: Perform some suffix merging.
Change the format strings to change the two-space indent to just one
space. This looks perfectly fine, and for the normal case enables half
the strings to be dropped as they are now suffixes of other strings.
We also move the strings in help_print_desc into the table, as one of
these is a suffix of the other, but out of the box GCC does not fully
perform this optimization.
Nick Bowler [Tue, 5 Dec 2023 02:16:36 +0000 (21:16 -0500)]
help_print_optstring: New table-based implementation.
By moving the format strings into a single table and a table-based
lookup to select the matching string, we can drastically simplify
the control flow in this function.
This results in about a ~35% code size reduction compiled with gcc.
The tables are generated using gen-strtab.awk but we manually copy
it into the source file, as downstream users just pick out help.c
and help.h by themselves.
Nick Bowler [Tue, 5 Dec 2023 01:55:43 +0000 (20:55 -0500)]
help_print_desc: Simplify implementation.
Simplify the way this function works by basing the loop around strcspn.
There is a slight change of behaviour when called with i==0 and an empty
description. Previously, no newline was printed. Now, a newline is
printed. I think the original behaviour was a mistake as it does not
make much sense (and this function serves no real purpose unless i is
nonzero). The tests are updated to expect this behaviour change.
As a bonus, we no longer tickle a problem with some ancient systems
(e.g., ULTRIX 4.5) which have a printf that doesn't support a negative
precision argument to mean "default precision". This problem actually
has no coverage in the test suite, so fix that up too.
Nick Bowler [Tue, 5 Dec 2023 01:36:17 +0000 (20:36 -0500)]
gen-strtab.awk: Allow repeated whitespace in one-line strings.
The documentation says that, for a one-line string, all whitespace
immediately following the identifier is deleted.
It does not say that anything happens to other withespace on the line,
but it turns out that consecutive sequences of whitespace are replaced
by a single space. This is unexpected.
Fix up the script so that doesn't happen, and add a test for this.
Nick Bowler [Sun, 3 Dec 2023 06:04:21 +0000 (01:04 -0500)]
gen-strtab.awk: Work around weird ULTRIX nawk bug.
ULTRIX 4.5 nawk has a weird bug where it seems to sometimes "forget" the
length of $0 after it has been modified, when assigning the modified
value directly to a named variable. For example:
% echo x | nawk '{ $0 = "hello"; x = $0; print x "rld"; }'
hrld
% echo xx | nawk '{ $0 = "hello"; x = $0; print x "rld"; }'
herld
The bug does not occur if there is an intervening assignment to any of
the field variables, or if the assignment is made slightly more complex,
or if any other variable is used:
% echo x | nawk '{ $0 = "hello"; $1 = $1; x = $0; print x "rld"; }'
hellorld
% echo x | nawk '{ $0 = "hello"; x = "" $0; print x "rld"; }'
hellorld
% nawk 'BEGIN { z = "hello"; x = z; print x "rld"; }' z=x
hellorld
Nick Bowler [Sun, 3 Dec 2023 05:52:53 +0000 (00:52 -0500)]
gen-options.awk: Don't pass ERE to split().
Strange behaviour is observed on ULTRIX 4.5 nawk when passing a
literal ERE as the fs argument to the split function. There is
no problem passing an equivalent string instead.
Nick Bowler [Sun, 3 Dec 2023 05:36:52 +0000 (00:36 -0500)]
Explicitly test for empty strings in awk scripts.
ULTRIX 4.5 nawk does not treat arbitrary nonempty strings as "true"
in boolean contexts; they appear to be converted to integers and
compared against zero, which means most nonempty strings are false.
Easy enough to work around.
Nick Bowler [Sun, 3 Dec 2023 05:21:49 +0000 (00:21 -0500)]
Avoid local array parameters in awk scripts.
ULTRIX 4.5 nawk appears to not support local array parameters on user-
defined functions (it is OK for array arguments passed by the caller).
Things seems to work on the first call but on the second call the array
is not empty and other weird behaviours occur.
Using a global array instead seems OK and avoids the problem.
Nick Bowler [Sun, 3 Dec 2023 04:13:24 +0000 (23:13 -0500)]
Rework backslash substitutions in awk scripts.
It appears that even using plain "\\" in a replacement string for sub
and gsub is still problematic. Most awks handle this OK (returning a
literal backslash) provided the next character is not a backslash, but
on ULTRIX 4.5 nawk eats the next character regardless and interprets
it as a backslash-escape sequence, oops.
So let's try something a bit different. We can ask awk what it does
with "\\\\", and use either "\\\\" or "\\" depending on the answer.
Nick Bowler [Sun, 3 Dec 2023 03:47:29 +0000 (22:47 -0500)]
Include AT_INIT bugfix monkey patch in the testsuite.
I keep running into this stupid bug where AT_CHECK uses redirections on
: and this fails on certain shells[1]. Maybe one day it will be fixed
upstream. We already have a workaround but the dxcommon test suite is
not using it... easily solved.
Nick Bowler [Sun, 3 Dec 2023 03:33:49 +0000 (22:33 -0500)]
Add a script to embed config.h into installed headers.
When installing a library, sometimes it is difficult to avoid that
certain aspects of the interface depend on the build configuration.
While it is possible use configuration header templates to create
reasonably installable header files, this script enables a different
approach: just write code (mostly) normally and then patch up any
configuration macro references at install time.
Nick Bowler [Fri, 1 Dec 2023 02:31:06 +0000 (21:31 -0500)]
Add a helper macro for gnulib monkey patches.
Certain gnulib patches can be applied directly at m4 time, which
makes it pretty simple to share between projects, and should be a
bit more robust than e.g., patching conditional dependencies in
bootstrap scripts.
Nick Bowler [Thu, 30 Nov 2023 01:33:29 +0000 (20:33 -0500)]
Add a configure test for C99-ish "for" declarations.
Several C89-era compilers support for loop declarations (although
I'm not sure if they implement C99 block scoping rules exactly and
this new macro doesn't check for that).
It is generally easy to simply not use this feature, so conditional
compilation based on its presence is of limited utility. But it can
sometimes be handy to know when (optionally) including third-party
code that depends on this syntax.
The test is short-circuited based on Autoconf's C version probes, which
requires an annoying amount of complexity because Autoconf 2.70 has
totally broken C99 detection and Autoconf 2.71 includes incompatible
changes to previously-documented behaviour.
Nick Bowler [Tue, 21 Nov 2023 05:20:26 +0000 (00:20 -0500)]
gettext.mk: Rewrite rules with less forking.
Use a pure shell solution to parse the .mo filenames instead
of forking out to expr, and don't use subshells with "set -x"
to print commands when we can almost as easily use plain echo.
Nick Bowler [Tue, 21 Nov 2023 03:57:23 +0000 (22:57 -0500)]
gettext.mk: Fix installation rules with some shells.
The following syntax is not portable to certain shells:
for f in ; do stuff; done
When no translations are installed, a command just like this is
generated in the installation rules. Instead of doing nothing
as desired, HP-UX 11 /bin/sh, heirloom-sh, and presumably also
old Solaris /bin/sh will exit with an error.
To work around this problem, we can stuff the list into a shell
variable and then use that in the for loop.
Nick Bowler [Tue, 21 Nov 2023 03:42:33 +0000 (22:42 -0500)]
tests: Add some coverage of the gettext.mk snippet.
There is zero test coverage of this snippet, add a new test case to
executes these rules in various ways. As with other test cases, one
can set TEST_SHELL in the environment to test the behaviour with a
particular shell. The astute reader may notice that the rules do
not work with some shells...
Nick Bowler [Sat, 18 Nov 2023 20:28:10 +0000 (15:28 -0500)]
pack: Improve 64-bit detection.
HP-UX 11.11 does not provide (U)LLONG_MAX in <limits.h>, but does provide
(U)LONG_LONG_MAX. Tweak the check to find this, and also accept the
Autoconf-provided HAVE_LONG_LONG_INT so things just work if that is
used.
Nick Bowler [Sat, 18 Nov 2023 07:40:51 +0000 (02:40 -0500)]
tests: Improve portability for perl script tests.
Older versions of perl do not recognize the "-f" option, which I think
was just used by mistake anyway, so we can just remove it.
Sed commands like /foo/{g;p} are not portable; POSIX requires it to be
written like:
/foo/{
g
p
}
HP-UX 11 sed does not interpret "-f -" to mean "read the script from
standard input", as intended; it looks for a file called - instead.
There is little reason to do this, we can just quote the script
normally.
Fixing all this allows the tests to run on HP-UX 11.
Nick Bowler [Fri, 17 Nov 2023 08:03:42 +0000 (03:03 -0500)]
gen-strtab.awk: Work around HP-UX shell bug in test case.
On HP-UX /bin/sh, redirections on "read" in a subshell can drop some
input on the floor. Since Autotest runs many things in subshells, we
hit this when trying to construct the program to dump the string table.
Rearranging the code to use only a single redirection works around
the problem.
Nick Bowler [Fri, 17 Nov 2023 05:28:11 +0000 (00:28 -0500)]
help: Fix NLS test on HP-UX 11.
This system has a "C.utf8" locale which appears disfunctional, all
other utf8 locales are OK. So change the locale probe to avoid this
one if possible.
The HP-UX shell also seems to have some trouble with UTF-8 (especially
if the user set this C.utf8 locale where it just barfs on everything),
so rework the test to not involve the shell trying to decode strings.
Nick Bowler [Fri, 17 Nov 2023 04:07:34 +0000 (23:07 -0500)]
copysym: Avoid dependency on <stdint.h>
The <inttypes.h> header is older and a bit more portable. But even
then, it is not really essential for this code. So adjust things to
use <inttypes.h> only if the Autoconf-provided HAVE_INTTYPES_H is
defined; otherwise a C89 <limits.h>-based fallback is used.
Nick Bowler [Fri, 17 Nov 2023 02:29:27 +0000 (21:29 -0500)]
help: Allow tests to build/run without <getopt.h>
We only need a suitable definition of struct option in order to build
the tests; this can be provided by a stub header if the system does not
provide it.
Nick Bowler [Thu, 16 Nov 2023 01:33:31 +0000 (20:33 -0500)]
help_print_optstring: Don't depend on snprintf.
The current code calls snprintf unconditionally, even if NLS is disabled
(although the function is not called, the code is still compiled), so it
won't build in environments which lack this function.
For the NLS-enabled case we can rely on the GNU libintl fallback being
available as required.
Otherwise we can just provide a stub to ensure the dead code compiles,
and for the NLS test programs hack in sprintf as a replacement which
should work just fine.
Nick Bowler [Thu, 16 Nov 2023 01:04:16 +0000 (20:04 -0500)]
help_print_optstring: Test fullwidth/halfwidth character output.
This function is supposed to support proper alignment with translated
strings possibly containing fullwidth characters, but we have no
test coverage of that aspect whatsoever.
Let's try to have at least some basic tests, which is made a bit
tricky since Autotest/m4sh busts the locale (but we can work around
this problem).
Nick Bowler [Sat, 1 Jul 2023 21:32:57 +0000 (17:32 -0400)]
test-tap: Fix Solaris /bin/sh "prove: not found" on stderr.
The prove command is intended to be optional, just to display a summary
of the test result if it is available. However, on old Solaris /bin/sh,
our command
prove ... 2>&1
doesn't redirect the error message from the shell when prove is not
found. We can use a command like like
Nick Bowler [Sat, 1 Jul 2023 20:38:06 +0000 (16:38 -0400)]
Add helper to work around Solaris /bin/sh bug in Autotest.
This hackjob edits the definition of AT_INIT to fix the problem
described in this patch[1], even for versions of Autoconf which
do not include this fix (which, as of this writing, is all of
them).
Specifically, older Solaris /bin/sh does not correctly handle
redirections on : within a shell function, but Autotest test
suites expect this to work when preparing the environment
for AT_CHECK.
Nick Bowler [Fri, 16 Jun 2023 04:01:35 +0000 (00:01 -0400)]
copysym: Fix pointer qualifier mismatch.
The "const void *" in the search function can't be directly assigned to
the "const char (*)[5]" variable, due to the differently-qualified
pointer target types. Let's just cast off that qualifier, shall we?
Nick Bowler [Tue, 13 Jun 2023 01:01:27 +0000 (21:01 -0400)]
tap: Include <config.h> as needed.
When using a config header in Autoconf, it is necessary to #include
that header before including anything else that might be influenced
by configure tests, which includes standard C headers.
In particular, things can go completely off the rails if Gnulib is
involved and generates replacements for the standard headers used
by this file.
Nick Bowler [Tue, 30 May 2023 00:35:15 +0000 (20:35 -0400)]
copysym: Avoid including xtra.h.
Using the xtra.h header just for the arraysize macro is just added
complexity for downstream projects (who then need to ensure the xtra.h
header is distributed even if they don't otherwise need it).
Nick Bowler [Sat, 27 May 2023 00:56:45 +0000 (20:56 -0400)]
Add a dedicated function to emit the copyright symbol.
This is intended to replace gnulib's str_iconv function in programs
when the only use of that function is to generate a copyright symbol
for output. This implementation should be much more compact.
Nick Bowler [Thu, 13 Apr 2023 01:11:38 +0000 (21:11 -0400)]
help_print_desc: Ensure newline is printed for empty description.
Callers expect help_print_desc to always end on a new line, so that the
next option can be printed at the start of a new line. Currently,
however, this function does nothing for an empty description string,
leading to incorrect output unless the option name was long enough
that help_print_optstring took care of it.
Add a special case to fix that up, and a new test case.
Nick Bowler [Thu, 26 Jan 2023 02:06:21 +0000 (21:06 -0500)]
gen-strtab.awk: Add an option to suppress object definitions.
Defining objects in the generated header can make some usage more
difficult. Add an option which defines a macro instead, allowing
the user to control definitions.
Nick Bowler [Mon, 23 Jan 2023 04:34:16 +0000 (23:34 -0500)]
xtra: Avoid undefined ## usage.
According to the C spec, every use of the ## operator has to produce a
valid preprocessing token, otherwise the behaviour is undefined. It is
not enough to merely ensure that a sequence of ## operators results in
a valid token at the end.
This matters in practice, as at least some versions of the Sun Studio
compiler will turn:
#define PASTE(a, b) a ## b
#define PASTE2(a, b) PASTE(a, b)
PASTE2(uint_least, PASTE(8, _t))
into two tokens:
uint_least8 _t
instead of uint_least8_t as desired. The results are not good for
the XTRA_PACKED_LOPTS macro. Fortunately, it is straightforward to
rearrange the expansions to avoid this problem.
Nick Bowler [Mon, 23 Jan 2023 04:31:24 +0000 (23:31 -0500)]
tests: Don't use AT_TESTED for optional build tools.
Turns out that AT_TESTED causes a hard failure if a listed tool is not
installed. This makes it useless to get version numbers recorded for
build tools that may or may not be available, so I guess we just have
to code that up explicitly.
In the main test group for gen-options.awk, various combinations of
sed and shell line noise are used to create the expected output.
It turns out that at least some versions of Solaris /bin/sed don't
accept patterns like [^\n] as meaning "any character other than
newline", which causes this line noise to fail. We could perhaps
use /usr/xpg4/bin/sed instead which seems to not have this specific
problem, but let's just do everything in awk which is much simpler.
Nick Bowler [Sun, 8 Jan 2023 05:32:22 +0000 (00:32 -0500)]
Add a helper macro to use the new gen-options packed format.
Abstract some of the boilerplate into a new macro. The gen-options.awk
is augmented to also give the packed format width as a power of two
(8/16/32/64). Other than that we can use all the existing definitions
to make this common code in a new "xtra.h" header file.
Nick Bowler [Sun, 8 Jan 2023 04:19:18 +0000 (23:19 -0500)]
gen-options.awk: Work around busybox regex bug.
It appears that busybox awk does not recognize \] in a character class,
interpreting the backslash literally and taking the ] as the end of the
character class:
Busybox does accept []] as meaning "] is part of the character class",
but this syntax is not portable to heirloom (and presumably also Solaris
10) awk, which only works with [\]]. GNU awk accepts either form.
So instead, let's avoid this problematic construct entirely by writing
the test in a different way.
Nick Bowler [Sun, 8 Jan 2023 04:14:56 +0000 (23:14 -0500)]
gen-options.awk: Fix generated help text under mawk.
It appears that the following construction is not portable:
% gawk 'BEGIN { a[0] = 0 in a ? 54 : 42; print a[0]; }'
42
% mawk 'BEGIN { a[0] = 0 in a ? 54 : 42; print a[0]; }'
54
In gen-options.awk, this results in results in bogus leading newlines
added to the help text. It is easy enough to restructure the code
to not do that, separating the test from the assignment.
Nick Bowler [Sun, 8 Jan 2023 02:14:10 +0000 (21:14 -0500)]
gen-options.awk: Add a more compact data representation.
Add a new set of macros that can be used to encode the long options
array as integers, which is possible provided that the "action"
feature is not used. In addition to reducing the size of the
static array, it should be helpful to avoid relocations in
position-independent executables and improve data sharing.
This causes errors when the script tries to use gsub to restore
the double backslashes on these implementations. There seems to
be no trouble substiting "\\" to get a single backslash, so we
can work around the problem by doubling up the characters being
replaced instead (which additionally requires an adjustment to
the length calculations to reverse this).
Nick Bowler [Wed, 4 Jan 2023 02:19:09 +0000 (21:19 -0500)]
gen-strtab.awk: Work around parse issue on HP-UX 11.
HP-UX 11 awk seems to have some issues parsing ! in expressions.
For example:
% awk 'BEGIN { print 1 + !0 }'
syntax error The source line is 1.
The error context is
BEGIN { print 1 + >>> ! <<< 0 }
awk: The statement cannot be correctly parsed.
The source line is 1.
Adding parentheses appers sufficient to avoid the problem:
Nick Bowler [Tue, 3 Jan 2023 04:08:00 +0000 (23:08 -0500)]
gen-strtab.awk: Add a feature to disable l10n markings.
Preceding an identifier with two ampersands now causes the output to
not include the N_ macro for that string, so xgettext won't pick it up.
This enables creating a mixed string table with some strings marked as
translatable and others not.
Nick Bowler [Fri, 18 Nov 2022 02:55:34 +0000 (21:55 -0500)]
Add a macro to probe -mwindows on MinGW.
MinGW is typically configured to build console applications by default.
So in order to build a normal Windows GUI application, this flag is
needed when linking. Otherwise, a console window will be opened when
the program is run.
Other Windows compilers probably have similar options which can added to
the probe in the future.
Nick Bowler [Thu, 21 Apr 2022 05:29:22 +0000 (01:29 -0400)]
exported.sh: Work around DJGPP shell redirection bug.
In DJGPP's bash port, it seems that redirections of the form
/absolute/path/to/program 3>&-
result in the file descriptor being closed for the whole shell,
rather than just for the one command as expected. Since some of
the configure substitutions into exported.sh can be absolute paths,
this causes failures.
Using a different command syntax avoids this problem. For example,
{ /absolute/path/to/program; } 3>&-
appears to work as expected. Adjust the exported.sh.in snippet to
do just that. There is still one redirection error message printed
to stderr by DJGPP bash but it appears spurious at first glance.
All the redirections now appear to be working properly and the
script output looks good.
Nick Bowler [Thu, 21 Apr 2022 05:23:05 +0000 (01:23 -0400)]
exported.sh: Restructure argument processing a bit.
Adjust the script to iterate over arguments using a for loop instead of
repeatedly checkind $# and shifting. More significantly, avoid calling
expr on each filename where a case pattern can do the job.
Additionally, add a new test case to verify the basic sanity of this
script and its corresponding Autoconf macro.
Nick Bowler [Sat, 26 Mar 2022 03:58:29 +0000 (23:58 -0400)]
Fix autotest snippet with --disable-dependency-tracking.
When configuring with --disable-dependency-tracking, the .deps
directories are not created by config.status so unconditionally
creating files in there will not work.
We can avoid this problem by adjusting the rule to skip updating
testsuite.P in this configuration. Add a new test case to expose
the issue.
Nick Bowler [Thu, 24 Mar 2022 02:22:50 +0000 (22:22 -0400)]
Fix glconfig with --disable-dependency-tracking.
When configuring with --disable-dependency-tracking, the .deps
directories are not created by config.status so unconditionally
creating files in there will not work.
Making the dependency file generation conditional will avoid this
problem. Add a new test case to cover this issue.
Normally there is no problem, but when performing a VPATH builds
with --disable-dependency-tracking the "lib" directory does not get
created by configure and some of the rules that create header files
will fail.
It seems recent versions of Gnulib have a potential solution for this
particular problem but is easy enough to patch it up here too.
Nick Bowler [Sun, 27 Feb 2022 01:21:42 +0000 (20:21 -0500)]
fix-gnulib: Drop silly libgnu warning options.
We really don't care about upstream's choice of build warning flags.
All pushing this on downstream users does is massively inflate generated
makefiles for no actual benefit.
Nick Bowler [Sat, 26 Feb 2022 22:55:30 +0000 (17:55 -0500)]
fix-ltdl: Fix LIBOBJ mangling to work with automake-1.16.
Apparently automake 1.16 changed how object files in a subdirectory are
named when using subdir-objects. As the generated rules to copy compiled
objects to the ltdl_LTLIBOBJ substituted names, these rules are now failing
on current automake.
Fix this by explicitly setting a _SHORTNAME for the dummy libobj library,
which explicitly sets the name to use and thus it will be consistent
between different Automake versions.
Nick Bowler [Sat, 26 Feb 2022 20:26:58 +0000 (15:26 -0500)]
fix-ltdl: Fix order-only hack with HP-UX make.
HP-UX make gets confused by the space in $(findstring ...,$(...))
syntax: instead of expanding to the empty string as desired, it
mismatches the parentheses and we are left with a single ). E.g.:
Since $(x) and ${x} are equivalent in Make, a neat solution is to
just replace one set of parentheses with braces; now there are no
nested parentheses to mismatch in the first place.
Nick Bowler [Thu, 24 Feb 2022 01:47:55 +0000 (20:47 -0500)]
help_print_optstring: Fix handling of 'flag' options.
When getopt_long options use the 'flag' feature, the 'val' member is
not meaningful as a short option. So we shouldn't try to format it
as if there is a short option, as this leads to printing garbage.
Nick Bowler [Thu, 24 Feb 2022 01:41:35 +0000 (20:41 -0500)]
fix-gnulib: Patch %reldir% too.
Gnulib has recently started producing rules using %reldir%. This
is cool but since we are postprocessing the automake fragment anyway,
we can easily substitute %reldir% ourselves which avoids requiring
newer Automake just for this.
Nick Bowler [Thu, 24 Feb 2022 01:29:23 +0000 (20:29 -0500)]
fix-gnulib: Fix pattern for variable mangling.
Recently, gnulib outputs variables named like SED_HEADER_xxx which do
not contain filenames but are matched by the (unnecessarily broad)
pattern for which variables need substitution.
Fix that up and include a test case to reproduce the issue.
Nick Bowler [Sat, 19 Feb 2022 22:56:28 +0000 (17:56 -0500)]
Add object prefixing option for DX_GNULIB_SYMFILES.
This is needed to avoid an Automake limitation when mixing libtool
and non-libtool builds of the same source files. This prefix should
match the object renaming done by Automake when the subdir-objects
option is used.
Nick Bowler [Wed, 16 Feb 2022 03:22:59 +0000 (22:22 -0500)]
fix-gnulib: Reduce build-time impact of symbol renaming.
Often a package will use more gnulib modules than actually needed by a
library, with included utilities also making use of portability features
provided by gnulib.
With symbol renaming computed over all objects, as well as linking
everything into both libraries and programs, this leads to compiling
every required gnulib source 3 times, ouch! And due to how libtool
convenience libraries work, the unused modules are probably wasting
space in the library to boot.
We can do better.
We can partition the gnulib objects into two groups: those that are
needed by the library (and thus require symbol renaming) and the rest
which are only needed by programs. Only the first group needs any
special treatment. The others can be put into an ordinary static
library which is not installed.
A new macro, DX_GNULIB_SYMFILES, provides the machinery to do this.
Nick Bowler [Sun, 30 Jan 2022 03:02:47 +0000 (22:02 -0500)]
Check for stray configure output.
Instead of ignoring all configure output in tests, we can use the
--quiet option to suppress informational messages and then ensure
the output is empty. This should catch stray output from macros.
Nick Bowler [Sun, 30 Jan 2022 01:48:17 +0000 (20:48 -0500)]
Add "join" detection macro.
It seems that unconditional use of "join" is not portable. Alpine
Linux, for example, does not include it. Add a macro which detects
whether join exists and works, with a fallback to an (incomplete)
awk implementation of the tool.
Nick Bowler [Wed, 24 Nov 2021 01:45:02 +0000 (20:45 -0500)]
help: Include <locale.h> if NLS is enabled.
It seems gettext.h, at least the version in gnulib, expects LC_MESSAGES
to be available in some scenarios but does not include <locale.h> which
provides it.
Nick Bowler [Thu, 30 Sep 2021 01:58:46 +0000 (21:58 -0400)]
help: Add a single-dash long option mode.
For programs using getopt_long_only with single-dash long options, add a
compile time flag to select a mode that produces more appropriate help
output. This is typical of command-line options for X applications.