Skip to contentSkip to conversation

configure.ac: Properly check for libintl

PATCH In review Vadim Kochan 2 participants 6 messages Apr 18, 2019 – Apr 19, 2019

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 configure.ac: Properly check for libintl

Some libc implementations like uclibc or musl provides gettext stubs via libintl library but this case is not checked by AC_CHECK_LIB(c, gettext ...) because gcc has gettext as builtin which passess the check.

So check it with included libintl.h where gettext may unfold into libintl_gettext which will cause check to fail if libintl_gettext are needed to be linked with -lintl.

---
 configure.ac | 16 +++++++++++++---
 1 file changed, 13 insertions(+), 3 deletions(-)
configure.ac +13 -3
index e0d0da3c0c..be3b55f1cc 100644
@@ -763,9 +763,19 @@ AC_CHECK_LIB([c], [basename],
 GIT_CONF_SUBST([NEEDS_LIBGEN])
 test -n "$NEEDS_LIBGEN" && LIBS="$LIBS -lgen"
 
-AC_CHECK_LIB([c], [gettext],
-[LIBC_CONTAINS_LIBINTL=YesPlease],
-[LIBC_CONTAINS_LIBINTL=])
+AC_DEFUN([LIBINTL_SRC], [
+AC_LANG_PROGRAM([[
+#include <libintl.h>
+]],[[
+char *msg = gettext("test");
+]])])
+
+AC_MSG_CHECKING([if libc contains libintl])
+AC_LINK_IFELSE([LIBINTL_SRC],
+	[AC_MSG_RESULT([yes])
+	LIBC_CONTAINS_LIBINTL=YesPlease],
+	[AC_MSG_RESULT([no])
+	LIBC_CONTAINS_LIBINTL=])
 GIT_CONF_SUBST([LIBC_CONTAINS_LIBINTL])
 
 #
Signature · 1 line
2.14.1
#1 lore raw
Quoted 8 lines · Vadim Kochan

Some libc implementations like uclibc or musl provides gettext stubs via libintl library but this case is not checked by AC_CHECK_LIB(c, gettext ...) because gcc has gettext as builtin which passess the check.

So check it with included libintl.h where gettext may unfold into libintl_gettext which will cause check to fail if libintl_gettext are needed to be linked with -lintl.

Let me make sure I can understand the above correctly (otherwise, that is a sign that the proposed log message is lacking) by trying to rephrase (iow, this is not me saying "your log message should be rewritten like so"; it is "if I read what you wrote above correctly, I think what I am going to write here is also correct"):

Some libc implementations have function called gettext() that can be linked via -lc without -lintl, but these are mere stubs and do not do useful i18n. On these systems, if a program that calls gettext() is built _with_ "#include <libintl.h>", the linker calls for the real version (e.g. libintl_gettext()) and that can be satisfied only by linking with -lintl.

The current check to see if -lc provides with gettext() is sufficient for libc implementations like GNU libc that actually has full fledged gettext(); to detect libc with stub gettext() and libintl with real gettext(), aliased via <libintl.h>, the check to see if -lintl is necessary must be done with a sample source that #include's the header file.

Is that what is going on and why this patch is needed?

I think the only possibile kind of system this change could break that currently is working is the one that has a usable gettext() in -lc, but does not offer <libintl.h>, as the new test program added by this patch will fail to compile, but I do not think that is possible in practice---our own gettext.c #include's <libintl.h> so there is no way such a hypothetical system that would be broken by this change could possibly have built Git successfully.

Assuming that the way I read your log message is in line with what you wanted to say, I think the patch looks good.

Thanks.

Quoted 32 lines
---
 configure.ac | 16 +++++++++++++---
 1 file changed, 13 insertions(+), 3 deletions(-)
configure.ac +13 -3
index e0d0da3c0c..be3b55f1cc 100644
@@ -763,9 +763,19 @@ AC_CHECK_LIB([c], [basename],
 GIT_CONF_SUBST([NEEDS_LIBGEN])
 test -n "$NEEDS_LIBGEN" && LIBS="$LIBS -lgen"
 
-AC_CHECK_LIB([c], [gettext],
-[LIBC_CONTAINS_LIBINTL=YesPlease],
-[LIBC_CONTAINS_LIBINTL=])
+AC_DEFUN([LIBINTL_SRC], [
+AC_LANG_PROGRAM([[
+#include <libintl.h>
+]],[[
+char *msg = gettext("test");
+]])])
+
+AC_MSG_CHECKING([if libc contains libintl])
+AC_LINK_IFELSE([LIBINTL_SRC],
+	[AC_MSG_RESULT([yes])
+	LIBC_CONTAINS_LIBINTL=YesPlease],
+	[AC_MSG_RESULT([no])
+	LIBC_CONTAINS_LIBINTL=])
 GIT_CONF_SUBST([LIBC_CONTAINS_LIBINTL])
 
 #
#2 lore raw

Hi Junio,

Quoted 41 lines · Junio C Hamano
Quoted 8 lines · Vadim Kochan

Some libc implementations like uclibc or musl provides gettext stubs via libintl library but this case is not checked by AC_CHECK_LIB(c, gettext ...) because gcc has gettext as builtin which passess the check.

So check it with included libintl.h where gettext may unfold into libintl_gettext which will cause check to fail if libintl_gettext are needed to be linked with -lintl.

Let me make sure I can understand the above correctly (otherwise, that is a sign that the proposed log message is lacking) by trying to rephrase (iow, this is not me saying "your log message should be rewritten like so"; it is "if I read what you wrote above correctly, I think what I am going to write here is also correct"):

Some libc implementations have function called gettext() that can be linked via -lc without -lintl, but these are mere stubs and do not do useful i18n. On these systems, if a program that calls gettext() is built _with_ "#include <libintl.h>", the linker calls for the real version (e.g. libintl_gettext()) and that can be satisfied only by linking with -lintl.

The current check to see if -lc provides with gettext() is sufficient for libc implementations like GNU libc that actually has full fledged gettext(); to detect libc with stub gettext() and libintl with real gettext(), aliased via <libintl.h>, the check to see if -lintl is necessary must be done with a sample source that #include's the header file.

Is that what is going on and why this patch is needed?

I think the only possibile kind of system this change could break that currently is working is the one that has a usable gettext() in -lc, but does not offer <libintl.h>, as the new test program added by this patch will fail to compile, but I do not think that is possible in practice---our own gettext.c #include's <libintl.h> so there is no way such a hypothetical system that would be broken by this change could possibly have built Git successfully.

Quoted 4 lines

Assuming that the way I read your log message is in line with what you wanted to say, I think the patch looks good.

Thanks.

Yes you are correct. 'gettext' even might be defined as libintl_gettext. When I got build error I checked config.log and I saw that gcc claims that gettext builtin is used, so I think that was the reason why test passed for non-glibc. So, if to use <libintl.h> then gettext might be preprocessed to libintl_gettext in the test. The original error which I was fixing is:

	http://autobuild.buildroot.net/results/8eeac7f7ddd97576eaeb87311bf0988d59d8b132/build-end.log
Quoted 34 lines
Quoted 34 lines
---
 configure.ac | 16 +++++++++++++---
 1 file changed, 13 insertions(+), 3 deletions(-)
configure.ac +13 -3
index e0d0da3c0c..be3b55f1cc 100644
@@ -763,9 +763,19 @@ AC_CHECK_LIB([c], [basename],
 GIT_CONF_SUBST([NEEDS_LIBGEN])
 test -n "$NEEDS_LIBGEN" && LIBS="$LIBS -lgen"
 
-AC_CHECK_LIB([c], [gettext],
-[LIBC_CONTAINS_LIBINTL=YesPlease],
-[LIBC_CONTAINS_LIBINTL=])
+AC_DEFUN([LIBINTL_SRC], [
+AC_LANG_PROGRAM([[
+#include <libintl.h>
+]],[[
+char *msg = gettext("test");
+]])])
+
+AC_MSG_CHECKING([if libc contains libintl])
+AC_LINK_IFELSE([LIBINTL_SRC],
+	[AC_MSG_RESULT([yes])
+	LIBC_CONTAINS_LIBINTL=YesPlease],
+	[AC_MSG_RESULT([no])
+	LIBC_CONTAINS_LIBINTL=])
 GIT_CONF_SUBST([LIBC_CONTAINS_LIBINTL])
 
 #
#3 lore raw
Quoted 14 lines · Vadym Kochan
Quoted 14 lines

Some libc implementations have function called gettext() that can be linked via -lc without -lintl, but these are mere stubs and do not do useful i18n. On these systems, if a program that calls gettext() is built _with_ "#include <libintl.h>", the linker calls for the real version (e.g. libintl_gettext()) and that can be satisfied only by linking with -lintl.

The current check to see if -lc provides with gettext() is sufficient for libc implementations like GNU libc that actually has full fledged gettext(); to detect libc with stub gettext() and libintl with real gettext(), aliased via <libintl.h>, the check to see if -lintl is necessary must be done with a sample source that #include's the header file.

Quoted 3 lines
Quoted 2 lines

Is that what is going on and why this patch is needed?

Yes you are correct. 'gettext' even might be defined as libintl_gettext.

With this exchange, I was aiming for extracting a more useful title for this patch out of you ;-), and I think I accomplished my goal.

"Properly" is fairly a useless adverb in the context of a patch title, as it does not tell us why we thought the way in which the updated code works is more "proper". In addition, because no code is perfect, future developers are bound to find something inproperly done in checking for libintl after this patch gets applied. It is better to say the most important thing the change does concisely and concretely.

I think

autoconf: #include <libintl.h> when checking for gettext()

is probably a better title.

Together with your originally proposed log message, which we now know explains why this inclusion makes a difference sufficiently to be understandable by an average Git developer, the resulting commit will communicate to our future developers the reason why we thought this was a good change clearly.

Thanks.

#4 lore raw

Hi Junio,

Quoted 37 lines · Junio C Hamano
Quoted 14 lines · Vadym Kochan
Quoted 14 lines

Some libc implementations have function called gettext() that can be linked via -lc without -lintl, but these are mere stubs and do not do useful i18n. On these systems, if a program that calls gettext() is built _with_ "#include <libintl.h>", the linker calls for the real version (e.g. libintl_gettext()) and that can be satisfied only by linking with -lintl.

The current check to see if -lc provides with gettext() is sufficient for libc implementations like GNU libc that actually has full fledged gettext(); to detect libc with stub gettext() and libintl with real gettext(), aliased via <libintl.h>, the check to see if -lintl is necessary must be done with a sample source that #include's the header file.

Quoted 3 lines
Quoted 2 lines

Is that what is going on and why this patch is needed?

Yes you are correct. 'gettext' even might be defined as libintl_gettext.

With this exchange, I was aiming for extracting a more useful title for this patch out of you ;-), and I think I accomplished my goal.

"Properly" is fairly a useless adverb in the context of a patch title, as it does not tell us why we thought the way in which the updated code works is more "proper". In addition, because no code is perfect, future developers are bound to find something inproperly done in checking for libintl after this patch gets applied. It is better to say the most important thing the change does concisely and concretely.

I think

autoconf: #include <libintl.h> when checking for gettext()

is probably a better title.

Quoted 7 lines

Together with your originally proposed log message, which we now know explains why this inclusion makes a difference sufficiently to be understandable by an average Git developer, the resulting commit will communicate to our future developers the reason why we thought this was a good change clearly.

Thanks.

Thanks! Should I re-submit patch with title updated v2 ?

Regards,
Vadim Kochan

#5 lore raw
Quoted 12 lines · Vadim Kochan
Quoted 12 lines

I think

autoconf: #include <libintl.h> when checking for gettext()

is probably a better title.

Together with your originally proposed log message, which we now know explains why this inclusion makes a difference sufficiently to be understandable by an average Git developer, the resulting commit will communicate to our future developers the reason why we thought this was a good change clearly.

Quoted 3 lines
Quoted 1 line

Thanks.

Thanks! Should I re-submit patch with title updated v2 ?

You could if you feel like it, but in a case like this, responding with "Yeah, I agree with your suggestion" and telling me to retitle is sufficient, if you have nothing else to change in the v2.

I've already did the retitling while queuing it, so no need to resend.

Thanks.