Bug#267206: gcc-3.4: vector op code generation regression

[Matthew Dempsky <jivera@flame.org>]

Package: gcc-3.4
Version: 3.4.1-4sarge1
Severity: normal

I'm compiling the following code example under both gcc 3.3.4-2 and
3.4.1-4sarge1:

    #include <stdio.h>
    
    typedef int v4si __attribute__ ((mode(V4SI)));
    
    static void
    print_v4si (const char * name, v4si val)
    {
      int x;
    
      printf ("%s:\n", name);
      for (x = 0; x < 4; ++x)
        printf ("  vals[%d] = %d\n", x, ((int *)&val)[x]);
      printf ("\n");
    }
    
    int
    main ()
    {
      v4si a = { 1, 2, 3, 4 };
      v4si b = { 2, 3, 4, 5 };
    
      print_v4si ("a", a);
      print_v4si ("b", b);
      print_v4si ("a+b", a + b);
    
      return 0;
    }

And here are some shell outputs compiling with 3.3 and 3.4 with and
without sse (along with running the results and commentary):

    $ gcc-3.3 foo.c -o foo-3.3
    foo.c: In function `print_v4si':
    foo.c:14: internal compiler error: in ix86_function_arg_boundary, at config/i386/i386.c:2476
    Please submit a full bug report,
    with preprocessed source if appropriate.
    See <URL:http://gcc.gnu.org/bugs.html> for instructions.
    For Debian GNU/Linux specific bug reporting instructions, see
    <URL:file:///usr/share/doc/gcc-3.3/README.Bugs>.

This is wrong, but it seems fixed in 3.4:

    $ gcc-3.4 foo.c -o foo-3.4
    foo.c: In function `print_v4si':
    foo.c:7: warning: SSE vector argument without SSE enabled changes the ABI
    foo.c: In function `main':
    foo.c:22: warning: SSE vector argument without SSE enabled changes the ABI
    foo.c:23: warning: SSE vector argument without SSE enabled changes the ABI
    foo.c:24: warning: SSE vector argument without SSE enabled changes the ABI
    $ ./foo-3.4
    a:
      vals[0] = 1
      vals[1] = 2
      vals[2] = 3
      vals[3] = 4
    
    b:
      vals[0] = 2
      vals[1] = 3
      vals[2] = 4
      vals[3] = 5
    
    a+b:
      vals[0] = 3
      vals[1] = 5
      vals[2] = 7
      vals[3] = 9

Only nuissance is that the warnings seems superfluous (others may
disagree).

    $ gcc-3.3 foo.c -o foo-3.3-sse -msse
    $ ./foo-3.3-sse
    a:
      vals[0] = 1
      vals[1] = 2
      vals[2] = 3
      vals[3] = 4
    
    b:
      vals[0] = 2
      vals[1] = 3
      vals[2] = 4
      vals[3] = 5
    
    a+b:
      vals[0] = 3
      vals[1] = 5
      vals[2] = 7
      vals[3] = 9

This is correct.

    $ gcc-3.4 foo.c -o foo-3.4-sse -msse
    $ ./foo-3.4-sse
    a:
      vals[0] = 1
      vals[1] = 2
      vals[2] = 1
      vals[3] = -1073743580
    
    b:
      vals[0] = 2
      vals[1] = 3
      vals[2] = 1
      vals[3] = -1073743580
    
    a+b:
      vals[0] = 3
      vals[1] = 5
      vals[2] = 1
      vals[3] = -1073743580

No warning, but the generated code seems incorrect (or at least a
regression from 3.3) unless ((int *)&val)[x] isn't the correct
portable way to access a vector element, but there doesn't seem to be
an alternative that I've been able to deduce (and the documentation
doesn't list any).

-- System Information:
Debian Release: 3.1
  APT prefers testing
  APT policy: (500, 'testing')
Architecture: i386 (i686)
Kernel: Linux 2.6.7
Locale: LANG=en_US.UTF-8, LC_CTYPE=en_US.UTF-8

Versions of packages gcc-3.4 depends on:
ii  binutils                 2.14.90.0.7-8   The GNU assembler, linker and bina
ii  cpp-3.4                  3.4.1-4sarge1   The GNU C preprocessor
ii  gcc-3.4-base             3.4.1-4sarge1   The GNU Compiler Collection (base 
ii  libc6                    2.3.2.ds1-13    GNU C Library: Shared libraries an
ii  libgcc1                  1:3.4.1-4sarge1 GCC support library

-- no debconf information