cygwin build_alias issue (and a possible workaround...)

Smith, Gene gene.smith at
Tue Sep 14 15:40:37 UTC 2004

Bogdan Vacaliuc wrote, On 9/14/2004 5:36 AM:

> Thanks Chris,
> Hmm.  Well.  Interesting.  For me, ccjfail (without --srcdir's)
> fails, but all the rest don't (within the time I'm willing to wait 
> for).
> Let's settle on ccjfail as the fail function (see if removing the
> --srcdir items (3 of them) still makes it fail for you), then our 
> systems are in common.  In the attached script I made the default
> ccjfail
>> I have and I attach the script for you plus the fail.log (appended
>>  together). Note, all the failing tests on WinME pass on WindowsXP
>> with MSYS. More below ...
> Hmm.
> There was some activity on the cygwin list recently.
> We're beginning to narrow down the issue after some straces were
> captured.  There is a M$ bug I referenced:
> "BUG: Registry access from multiple threads might fail" (WinNT,
> Win2K);en-us;176906
> Which applies only to WinNT and Win2K, but not (apparently) to WinXP?
> Would everyone who is watching this thread, please let us know if the
> attached script fails on their machine within a small number of
> iterations, and also what OS they are running?
> $ ./test-configure
> I'm particularly curious now if people with WinXP are
> testing/building OK, but those with Win2K and otherwise are not.
> This would lend some credence to M$ KB176906 and suggest a possible
> workaround in either the bash or the cygwin.dll code...
> Thanks,
> -bogdan

After updating to latest cygwin, 1.5.11-1, it ran 19 times before 
failing with this message:

*** TEST 19 FAILS ***
configure: error: invalid feature name: multilib

Then I ran it again and it failed quickly like this:

*** TEST 1 FAILS ***
configure: error: invalid variable name: build_alias

Windows 2000

Then I checked and the upgrade had put ash back in /bin/sh. Made /bin/sh 
bash again and got:

$ ./test-configure.txt
Running using function ccjfail()...
./test-configure.txt: line 152: [: too many arguments
./test-configure.txt: line 156: [: too many arguments
*** TEST 0 FAILS ***
configure: error: invalid variable name: build_alias

Made /bin/sh back to the default ash and it currently up to TEST 50 with 
no errors yet. So it appears the default /bin/sh works a bit better than 
when you make it bash.


More information about the users mailing list