retitle 17399 Problems building guile-2.0.11 with libunistring 0.9.0
thanks
Hi Ed,
I finally looked into this test failure that you experienced when
compiling guile-2.0.11 from source code on Ubuntu 12.04. I see that
you've since updated to Ubuntu 14.04, which includes guile-2.0.11, so
your problems are happily solved, but I'm still curious to learn what
happened here.
To summarize, your problems all stemmed from a very old GNU libunistring
that you had installed in /usr/local. It was quite a bit older than the
minimum required version (0.9.3) which was packaged in Ubuntu 12.04.
The configure script identified the installed version as 0.9.0.
You should probably remove that old version of libunistring, as it will
likely cause problems with other software you build from source.
The rest of this email is primarily aimed at other Guile developers.
* * *
First, we should improve our configure script to bail out if the
libunistring is too old.
After Ed had manually edited the old "unistr.h" to fix the compilation
problem, he encountered this test failure:
ERROR: ports.test: unicode byte-order marks (BOMs):
Don't read from the port unless user asks to - arguments:
((decoding-error "scm_from_stringn" "input locale conversion error"
84 #vu8(254 255)))
Here's the relevant section of test code in ports.test:
--8<---------------cut here---------------start------------->8---
(pass-if "Don't read from the port unless user asks to"
(let* ((p (make-soft-port
(vector
(lambda (c) #f) ; write char
(lambda (s) #f) ; write string
(lambda () #f) ; flush
(lambda () (throw 'fail)) ; read char
(lambda () #f))
"rw")))
(set-port-encoding! p "UTF-16")
(display "abc" p)
(set-port-encoding! p "UTF-32")
(display "def" p)
#t))
--8<---------------cut here---------------end--------------->8---
The error occurred within 'sf_write' in vports.c while writing the BOM
to the soft port (before writing "abc"). The problem is that soft ports
are fundamentally based on strings, and anything written to them is
first converted to a string (using the locale encoding) within
'sf_write', in order to pass to the string to the user-provided "write"
procedure. Attempting to convert the UTF-16-encoded-BOM (0xFE 0xFF) to
the locale encoding failed, unsurprisingly.
I believe the locale should have been "C", because of the call
(setlocale LC_ALL "C") in test-suite/guile-test.
I'm actually surprised that this has ever worked, and it warrants
further investigation.
Mark
