[PATCH] s390: Fix missing early-clobber in __to_generic_vx [BZ #34952]

Stefan Liebler stli@linux.ibm.com
Thu Nov 5 12:12:45 GMT 2026


On 11/5/26 13:04, Stefan Liebler wrote:
> On 11/5/26 01:13, Florian Fuchs wrote:
>> R_TMP is declared as a plain output ("=d") without an early-clobber, so
>> the compiler may give it the same register as one of the inputs.
>>
>> At -O1, GCC does so for R_BUF: the buffer address is zeroed and the
>> following "vst %v16,0(%[R_I],%[R_BUF])" stores to address 0. Any
>> conversion of 16 or more characters to IBM037, IBM1047 or a similar
>> 8-bit charset then crashes. -O2 and -Os happen to allocate different
>> registers.
>>
>> Mark R_TMP as early-clobber ("=&d"), like R_I in the same asm.
>>
>> Reproducer, when glibc built with -O1:
>>
>> 	head -c16 /dev/zero | tr '\0' A | iconv -f UTF-8 -t IBM1047
>>
>> Tested on s390x with glibc built with -O1: the reproducer above
>> crashes without this change and works with it.
>>
>> Signed-off-by: Florian Fuchs <fuchsfl@gmail.com>
>> ---
>> Issue was hit, while building s390-tools with an -O1 glibc: GCC with
>> -fexec-charset=IBM1047 crashed with an internal compiler error in
>> iconv. cmdline that reproduces the gcc ice:
>>
>> 	echo 'const char *s = "Start subchannel failed";' | \
>> 		gcc -fexec-charset=IBM1047 -x c -c - -o /dev/null
>>
>>  sysdeps/s390/multiarch/8bit-generic.c | 2 +-
>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/sysdeps/s390/multiarch/8bit-generic.c b/sysdeps/s390/multiarch/8bit-generic.c
>> index 8224cec557..18671a8775 100644
>> --- a/sysdeps/s390/multiarch/8bit-generic.c
>> +++ b/sysdeps/s390/multiarch/8bit-generic.c
>> @@ -266,7 +266,7 @@
>>  			  ".machine pop"				\
>>  			  : /* outputs */ [R_IN] "+a" (inptr)		\
>>  			    , [R_I] "=&a" (index)			\
>> -			    , [R_TMP] "=d" (tmp)			\
>> +			    , [R_TMP] "=&d" (tmp)			\
>>  			    , [R_LI] "+d" (remaining_loop_count)	\
>>  			  : /* inputs */ [R_BUF] "a" (buf)		\
>>  			    , [R_MAX] "d" (max)				\
> Hi Florian,
> 
> many thanks for the bug-report and the fix. You are absolute right, I
> also see the described register usages with -O1. With marking R_TMP as
> early clobber, different registers are used and the reproducer "head
> -c16 /dev/zero | tr '\0' A | iconv -f UTF-8 -t IBM1047" is fixed with an
> -O1 build.
> 
> I've already double checked the remaining inline-assemblies in
> 8bit-generic.c for similar issues and will check the remaining
> s390x-inline-assemblies.
> 
> Reviewed-by: Stefan Liebler <stli@linux.ibm.com>
Committed and also closed the bug (If there are still issues. Please
reopen again).

Thanks,
Stefan


More information about the Libc-alpha mailing list