Re: Patch to fix regenerating libiberty/config.in

["Kaveh R. Ghazi" <ghazi at caip dot rutgers dot edu>]

 > From: Eric Botcazou <ebotcazou at libertysurf dot fr>
 > 
 > > > 2003-02-12  Kaveh R. Ghazi  <ghazi at caip dot rutgers dot edu>
 > > >
 > > > 	* acconfig.h: New file.  Add uintptr_t.
 > > > 	* config.in: Regenerated.
 > >
 > > This patch breaks bootstrap on sparc-sun-solaris2.5.1:
 > >
 > > "../src/libiberty/regex.c", line 305: invalid type combination
 > > "../src/libiberty/regex.c", line 305: warning: useless declaration
 > > "../src/libiberty/regex.c", line 305: warning: typedef declares no type
 > > name
 > 
 > The problem is pretty obvious: config.h contains these lines
 > /* Define to `unsigned long' if <sys/types.h> doesn't define.  */
 > #define uintptr_t unsigned long
 > /* Define if you have the `uintptr_t' type. */
 > /* #undef HAVE_UINTPTR_T */
 > and regex.c contains these lines
 > # if !defined _LIBC && !defined HAVE_UINTPTR_T
 > typedef unsigned long int uintptr_t;
 > # endif
 > so there is a conflict. Reverting the patch or removing the typedef appears 
 > to be sufficient to restore bootstrap, but I don't know enough libiberty to 
 > decide which solution is the best/safest. Thanks in advance for your help.
 > Eric Botcazou

I had a chance to look more closely this morning.

While reverting the original patch may restore bootstrap, I think that
would be wrong.  See the original motivation for it here:
http://gcc.gnu.org/ml/gcc-patches/2003-02/msg00938.html

I'd also prefer not to remove the typedef in regex.c since we want to
avoid gratuitous diffs with the upstream glibc copy.

I believe the following patch is the proper solution.  What we had was
a genuine bug in configure.in.  In there we check for uintptr_t and
provide a fallback define if that type doesn't exist.  Then we'd check
the shell variable result of the test and define HAVE_UINTPTR_T only
if the test succeeded.  That's bogus.

Since the fallback macro ensures we always have uintptr_t in some
form, we should unconditionally define HAVE_UINTPTR_T for the benefit
of regex.c and anyone else in libiberty who might rely on it.
(Currently no other files in 3.2.x do.)

I bootstrapped this on sparc-sun-solaris2.7, but that's not much of a
test since solaris2.7 has a uintptr_t and never saw your problem.

Can you please try the patch with 3.2.x on solaris2.5.1 and see if it
cures your bootstrap breakage?

If it does, ok everywhere?

		Thanks,
		--Kaveh


2003-04-13  Kaveh R. Ghazi  <ghazi at caip dot rutgers dot edu>

	* configure.in (HAVE_UINTPTR_T): Always define.

diff -rup orig/egcc-3.2-CVS20030413/libiberty/configure.in egcc-3.2-CVS20030413/libiberty/configure.in
--- orig/egcc-3.2-CVS20030413/libiberty/configure.in	Sun Feb  2 23:05:46 2003
+++ egcc-3.2-CVS20030413/libiberty/configure.in	Sun Apr 13 12:01:58 2003
@@ -134,11 +134,10 @@ AC_HEADER_TIME
 libiberty_AC_DECLARE_ERRNO
 
 AC_CHECK_TYPE(uintptr_t, unsigned long)
-
-if test $ac_cv_type_uintptr_t = yes
-then
-  AC_DEFINE(HAVE_UINTPTR_T, 1, [Define if you have the \`uintptr_t' type.])
-fi
+# Given the above check, we always have uintptr_t or a fallback
+# definition.  So define HAVE_UINTPTR_T in case any imported code
+# relies on it.
+AC_DEFINE(HAVE_UINTPTR_T, 1, [Define if you have the \`uintptr_t' type.])
 
 # This is the list of functions which libiberty will provide if they
 # are not available on the host.