> During "make test" of latest CVS revision OpenSSL_0_9_8-stable, the
> sha512 test fails on a segmentation fault in sha512_block_data_order():
>
> | ../util/shlib_wrap.sh ./sha512t
> | make[1]: *** [test_sha] Segmentation fault (core dumped)
> | make[1]: Leaving directory `/tmp/openssl-0.9.8-cvs/test'
> | make: *** [tests] Error 2
>
> Note OpenSSL release 0.9.8g works OK. Output of "make report" comes
> attached as testlog.gz
The only difference between sha512 in 098g and 098cvs is byte swapping
inline assembler. As your compiler is rather old (2.7.2 according to
testlog), I'd guess it's a compiler bug. Could you collect disassembler
output for offending code? I.e. run 'gdb test/sha512t core' (or rather
util/shlib_wrap.sh gdb test/sha512t as you have build for shared), run
'where', note address, run 'disassemble' and advance to just noted
address. Submit *complete* disassemble output from function start to
offending instruction. Test attached patch. Drop optimization level... A.
Index: crypto/sha/sha512.c
===================================================================
RCS file: /e/openssl/cvs/openssl/crypto/sha/sha512.c,v
retrieving revision 1.8.2.3
diff -u -r1.8.2.3 sha512.c
--- crypto/sha/sha512.c 11 Nov 2007 13:34:07 -0000 1.8.2.3
+++ crypto/sha/sha512.c 22 Apr 2008 13:58:22 -0000
@@ -327,10 +327,10 @@
((SHA_LONG64)hi)<<32|lo; })
# else
# define PULL64(x) ({ const unsigned int *p=(const unsigned int *)(&(x));\
- unsigned int hi,lo; \
+ unsigned int hi=p[0],lo=p[1]; \
asm ("bswapl %0; bswapl %1;" \
: "=r"(lo),"=r"(hi) \
- : "0"(p[1]),"1"(p[0])); \
+ : "0"(lo),"1"(hi)); \
((SHA_LONG64)hi)<<32|lo; })
# endif
# elif (defined(_ARCH_PPC) && defined(__64BIT__)) || defined(_ARCH_PPC64)
