Re: [Qemu-devel] [PATCH v2 05/39] fdsets: use weak aliases instead of qemu-tool.c/qemu-user.c

[Paolo Bonzini <pbonzini@redhat.com>]

Il 15/11/2012 23:18, Stefan Weil ha scritto:
> Am 15.11.2012 21:52, schrieb Paolo Bonzini:
>> Il 15/11/2012 19:01, Stefan Weil ha scritto:
>>> Hi Paolo,
>>>
>>> this patch breaks QEMU on 32 and 64 bit hosts, native and with Wine.
>>> It's easy to reproduce the SIGSEGV crash: just add a -snapshot option.
>>> Obviously the critical code is executed only when this option was used.
>>
>> I cannot reproduce this, so it must be an assembler or linker bug.
>>
>> Can you try the alternative code that is used for Mac OS X?
> 
> The code which is used for Mac OS X also compiles and
> results in the same run-time bug with Wine:

Ok, I reproduced the original binutils bug, and found a typo in the
weakrefs implementation.  Does this work for you?

diff --git a/compiler.h b/compiler.h
index 55d7d74..d552757 100644
--- a/compiler.h
+++ b/compiler.h
@@ -50,11 +50,12 @@
 #   define __printf__ __gnu_printf__
 #  endif
 # endif
-# if defined(__APPLE__)
+# if defined(__APPLE__) || defined(_WIN32)
 #  define QEMU_WEAK_ALIAS(newname, oldname) \
-        static typeof(oldname) weak_##newname __attribute__((unused, weakref(#oldname)))
+        static typeof(oldname) weak_##newname __attribute__((unused, weakref(#newname)))
 #  define QEMU_WEAK_REF(newname, oldname) (weak_##newname ? weak_##newname : oldname)
 # else
+#error
 #  define QEMU_WEAK_ALIAS(newname, oldname) \
         typeof(oldname) newname __attribute__((weak, alias (#oldname)))
 #  define QEMU_WEAK_REF(newname, oldname) newname


If it still doesn't work, let's make sure that this reduced testcase works
for you:

g1.c:
#include <stdio.h>
int f() { printf("strong"); return 82; }
int g() { printf("strong"); return 83; }

g2.c:
#include <stdio.h>
static int weak_f() { return 42; }
static int weak_g() { return 43; }
typeof(weak_f) f __attribute__((__weak__, __alias__("weak_f")));
typeof(weak_g) g __attribute__((__weak__, __alias__("weak_g")));
int main() { printf("%d/%d\n", f(), g()); }

g3.c:
#include <stdio.h>
static int default_f() { return 42; }
static int default_g() { return 43; }
static typeof(default_f) weak_f __attribute__((__weakref__("f")));
static typeof(default_g) weak_g __attribute__((__weakref__("g")));
int main() { printf("%d/%d\n", (weak_f?:default_f)(), (weak_g?:default_g)()); }

Output should be:
- 42/43 for "gcc g2.c"
- 42/43 for "gcc g3.c"
- strongstrong82/83 for "gcc g1.c g2.c"
- strongstrong82/83 for "gcc g1.c g3.c"

Output on Windows is:
- 42/42 for "gcc g2.c"
- 42/43 for "gcc g3.c"
- segfault for "gcc g1.c g2.c"
- strongstrong82/83 for "gcc g1.c g3.c"

So, indeed "normal" weak symbols are broken, but weakrefs seem to work.

If the above patch doesn't work, and/or the reduced testcase fails,
please send to me:

- the .exe for "gcc g1.c g3.c"
- the .s file for g3.s
- the .o file for osdep.o
- the linked .exe (I assume both the windows and console versions fail)
- perhaps the .s file for osdep too; add --save-temps to the compiler command line to get it

Paolo