Hello, Akim and Tom!
> Pavel> The problem happens quite often because install cannot
> Pavel> overwrite running executable files and happily goes ahead and
> Pavel> installs data files. I have been fooled by this two or three
> Pavel> times.
>
> I can't reproduce it with ash. Could you write some documentation for
> autoconf.texi? Thanks!
Sorry, it took me a while to reply.
The problem _is_ generic, not Cygwin specific. It's just that on Cygwin
you are much more likely to be unable to replace one file, but not the
others, in the same directory.
The problem can been reproduced on Red Hat Linux 7.3 (bash 2.05a, GNU make
3.79.1, Autoconf 2.53, Automake 1.6.3).
/usr/local/bin/mc is a directory with permissions 000. make tries to
install a file with that name, ignores the failure and finishes
"successfully":
[proski@marabou src]$ make install
make[1]: Entering directory `/usr/local/src/mc.v0/src'
/bin/sh ../config/mkinstalldirs /usr/local/bin
/usr/bin/install -c mc /usr/local/bin/mc
/usr/bin/install: cannot stat `/usr/local/bin/mc/mc': Permission denied
/usr/bin/install -c mcmfmt /usr/local/bin/mcmfmt
make install-exec-hook
make[2]: Entering directory `/usr/local/src/mc.v0/src'
make install_mcview
make[3]: Entering directory `/usr/local/src/mc.v0/src'
cd /usr/local/bin/ && rm -f mcview && ln -s mc mcview
make[3]: Leaving directory `/usr/local/src/mc.v0/src'
make install_mcedit
make[3]: Entering directory `/usr/local/src/mc.v0/src'
cd /usr/local/bin/ && rm -f mcedit && ln -s mc mcedit
make[3]: Leaving directory `/usr/local/src/mc.v0/src'
make[2]: Leaving directory `/usr/local/src/mc.v0/src'
/bin/sh ../config/mkinstalldirs /usr/local/lib/mc/bin
/usr/bin/install -c cons.saver /usr/local/lib/mc/bin/cons.saver
make[1]: Leaving directory `/usr/local/src/mc.v0/src'
[proski@marabou src]$
I don't understand the reference to "set -e". Take this makefile:
default:
for i in 0 1; do \
echo testing $$i; \
test $$i = 1; \
done
Run make on the same pure GNU/Linux system:
[proski@marabou shell-test]$ make
for i in 0 1; do \
echo testing $i; \
test $i = 1; \
done
testing 0
testing 1
[proski@marabou shell-test]$
There's no error. Now add "set -e", so that the makefile is now:
default:
set -e; \
for i in 0 1; do \
echo testing $$i; \
test $$i = 1; \
done
[proski@marabou shell-test]$ make
set -e; \
for i in 0 1; do \
echo testing $i; \
test $i = 1; \
done
testing 0
make: *** [default] Error 1
[proski@marabou shell-test]$
Feel the difference. Sorry, I don't know if I was supposed to try it on
BSD. Two important results (sorry if they are obvious):
1) make (even GNU make) executes scripts without implicit "set -e" - it
just executes separate commandes separately and checks the exit code of
each of them. "set -e" is not in effect for the shell. make does its
best to emulate "set -e", but doesn't split commands on the same line.
2) Explicit exit should be used inside separate commands (separate from
the point of view of make) if errors may not affect the exit code of the
whole command.
By the way, what if the user runs "make -k"? An "exit" would terminate
the loop, even though the user wants make to do as much as possible.
This would be especially bad if the loop is for subdirectories - a failure
in the first of them would cause all other subdirectories to be skipped,
and only local commands outside the loop would be executed.
You can consider this as a low-priority feature request for Automake.
Perhaps it can be done by examining $(MAKE) for GNU make.
--
Regards,
Pavel Roskin
