https://gcc.gnu.org/bugzilla/show_bug.cgi?id=86377
--- Comment #6 from U.Mutlu <um at mutluit dot com> --- The configure scripts in the subdirs/libs (like libiberty) do compile some test codes to populate the config.h of the subdir. And there the difference happens if -Wuninitialized is treated as error, ie. if user sets -Werror=uninitialized either explicitly or implicitly thru -Werror or a similar option: Example for the libiberty subdir: once w/o -Werror and once with -Werror: < conftest.c:53:7: warning: 't' is used uninitialized in this function [-Wuninitialized] --- > conftest.c:53:7: error: 't' is used uninitialized in this function > [-Werror=uninitialized] 361c361 < conftest.c:75:13: warning: 'cs[0]' is used uninitialized in this function [-Wuninitialized] --- > conftest.c:75:13: error: 'cs[0]' is used uninitialized in this function > [-Werror=uninitialized] 364,365c364,447 < configure:4686: $? = 0 < configure:4693: result: yes ... > configure:4693: result: no And in the 2nd case (>) from there on this causes many things in the config.h's to be set differently, incl. the said NEED_DECLARATION_ERRNO ... So, I think if the warnings in the stanza-codes could be eliminated then the compilation would go further. It is the following code that generates the above warnings: the variables t and cs should be initialized: > | /* end confdefs.h. */ > | > | int > | main () > | { > | /* FIXME: Include the comments suggested by Paul. */ > | #ifndef __cplusplus > | /* Ultrix mips cc rejects this. */ > | typedef int charset[2]; > | const charset cs; > | /* SunOS 4.1.1 cc rejects this. */ > | char const *const *pcpcc; > | char **ppc; > | /* NEC SVR4.0.2 mips cc rejects this. */ > | struct point {int x, y;}; > | static struct point const zero = {0,0}; > | /* AIX XL C 1.02.0.0 rejects this. > | It does not let you subtract one const X* pointer from another in > | an arm of an if-expression whose if-part is not a constant > | expression */ > | const char *g = "string"; > | pcpcc = &g + (g ? g-g : 0); > | /* HPUX 7.0 cc rejects these. */ > | ++pcpcc; > | ppc = (char**) pcpcc; > | pcpcc = (char const *const *) ppc; > | { /* SCO 3.2v4 cc rejects this. */ > | char *t; > | char const *s = 0 ? (char *) 0 : (char const *) 0; > | > | *t++ = 0; > | if (s) return 0; > | } > | { /* Someone thinks the Sun supposedly-ANSI compiler will reject this. */ > | int x[] = {25, 17}; > | const int *foo = &x[0]; > | ++foo; > | } > | { /* Sun SC1.0 ANSI compiler rejects this -- but not the above. */ > | typedef const int *iptr; > | iptr p = 0; > | ++p; > | } > | { /* AIX XL C 1.02.0.0 rejects this saying > | "k.c", line 2.27: 1506-025 (S) Operand must be a modifiable lvalue. > */ > | struct s { int j; const int *ap[3]; }; > | struct s *b; b->j = 5; > | } > | { /* ULTRIX-32 V3.1 (Rev 9) vcc rejects this */ > | const int foo = 10; > | if (!foo) return 0; > | } > | return !cs[0] && !zero.x; > | #endif > | > | ; > | return 0; > | } As this code is just for test-compilation by the configure script, then it makes sense to fix the warnings. I will try this for all occurrences of that code and similar ones, and if the result is satisfactory, then post a patch to the mailing list for review.