Re: [PATCH v4] Check global symbols in static library against shared library

[Florian Weimer <fweimer@redhat.com>]

* H. J. Lu:

>> But with our glibc 2.34 variant, I get a link failure:
>>
>> /usr/bin/ld: /tmp/ccOU5PJG.o: in function `my_scanf':
>> t.c:(.text+0x5c): undefined reference to `__nldbl__IO_vfscanf'
>> collect2: error: ld returned 1 exit status
>>
>> So your tests have found another genuine bug, I think.
>>
>
> Can you open a glibc bug?  We should fix it with a testcase.
> What should happen to __nldbl__IO_vfscanf in libc.a and libc.so?

Digging somewhat further, I believe this was an intentional change as
part of this commit:

commit 349718d4d7841df46bcc36df9bc2baef4c40d6f5
Author: Zack Weinberg <zackw@panix.com>
Date:   Wed Mar 7 14:31:58 2018 -0500

    Add __vfscanf_internal and __vfwscanf_internal with flags arguments.

(The commit message has a longer explanation of the change.)

It is this part which I expect causes the change to be dropped from the
static build:

diff --git a/sysdeps/ieee754/ldbl-opt/nldbl-compat.c b/sysdeps/ieee754/ldbl-opt/nldbl-compat.c
index 7a1e89c1a3..91ea27a423 100644
--- a/sysdeps/ieee754/ldbl-opt/nldbl-compat.c
+++ b/sysdeps/ieee754/ldbl-opt/nldbl-compat.c
@@ -330,16 +330,20 @@ __nldbl_wprintf (const wchar_t *fmt, ...)
   return done;
 }
 
+#if SHLIB_COMPAT (libc, GLIBC_2_0, GLIBC_2_29)
 int
 attribute_compat_text_section
 __nldbl__IO_vfscanf (FILE *s, const char *fmt, va_list ap, int *errp)
 {
   int res;
   set_no_long_double ();
-  res = _IO_vfscanf (s, fmt, ap, errp);
+  res = __vfscanf_internal (s, fmt, ap, 0);
   clear_no_long_double ();
+  if (__glibc_unlikely (errp != 0))
+    *errp = (res == -1);
   return res;
 }
+#endif
 
 int
 attribute_compat_text_section

But __nldbl__IO_vfscanf is not marked as a compat symbol:

# eu-readelf --symbols=.dynsym /lib64/libc.so.6 | grep __nldbl__IO_vfscanf
 1203: 00000000001c2240    132 FUNC    GLOBAL DEFAULT       11 __nldbl__IO_vfscanf@@GLIBC_2.17

And this must be what trips your test case.  We should turn it into a
compat symbol, rather than fixing the test case.

Thanks,
Florian