[Bug c/96399] New: Arithmetic shift with vector extension becomes logical right shift on s390x

[Bug c/96399] New: Arithmetic shift with vector extension becomes logical right shift on s390x

[yohei at jp dot ibm.com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96399

            Bug ID: 96399
           Summary: Arithmetic shift with vector extension becomes logical
                    right shift on s390x
           Product: gcc
           Version: 10.2.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: c
          Assignee: unassigned at gcc dot gnu.org
          Reporter: yohei at jp dot ibm.com
  Target Milestone: ---

When I applied arithmetic shift (>>) to a 16-element signed char vector
variable, I observed behavior of logical right shift on s390x.

-----
#include <stdio.h>

typedef char vec16char __attribute__ ((vector_size (16)));

void test(vec16char *p, vec16char *q)
{
    *q = *p >> 7;
}

int main(int argc, char *argv[])
{
    vec16char in = {0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80,
                    0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80, 0x80};
    vec16char out;

    test(&in, &out);

    printf("out[0] = %d\n", out[0]);
}
-----

I compiled the above code with gcc -Wall -march=z14 -O3 -o vectest vectest.c

Then, I got the following output on s390x:

out[0] = 1

I expected the following output:

out[0] = -1

At least on x86_64, I got the expected output.

On s390x, vesrlb is generated, but I think vesrab is appropriate.

[Bug c/96399] Arithmetic shift with vector extension becomes logical right shift on s390x

[yohei at jp dot ibm.com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96399

--- Comment #1 from Yohei Ueda <yohei at jp dot ibm.com> ---
Output of gcc -v:

Using built-in specs.
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/usr/local/libexec/gcc/s390x-linux-gnu/10.2.0/lto-wrapper
Target: s390x-linux-gnu
Configured with: /usr/src/gcc/configure --build=s390x-linux-gnu
--disable-multilib --enable-languages=c,c++,fortran,go
Thread model: posix
Supported LTO compression algorithms: zlib
gcc version 10.2.0 (GCC)

[Bug c/96399] Arithmetic shift with vector extension becomes logical right shift on s390x

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=96399

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |jakub at gcc dot gnu.org
             Status|UNCONFIRMED                 |RESOLVED
         Resolution|---                         |INVALID

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
s390x defaults to -funsigned-char, while x86_64 defaults to -fsigned-char, so
what you get is exactly what should happen.  If you want portable code, you
should use signed char or unsigned char and use char only if you don't care if
it is signed or unsigned.