c++/3299: __attribute__((aligned(16))) seems to be ignored on x86

[ryants@home.com]

>Number:         3299
>Category:       c++
>Synopsis:       __attribute__((aligned(16))) seems to be ignored on x86
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    unassigned
>State:          open
>Class:          sw-bug
>Submitter-Id:   net
>Arrival-Date:   Wed Jun 20 10:56:03 PDT 2001
>Closed-Date:
>Last-Modified:
>Originator:     Ryan T. Sammartino
>Release:        gcc version 2.95.3 19991030 (prerelease)
>Organization:
>Environment:
Linux 2.4.5-6mdksmp #1 SMP Mon Jun 18 10:19:22 PDT 2001 i686 unknown
>Description:
The following code (which doesn't include anything, so it's the same as the .ii that is generated) seg faults because none of v1, v2, or ret are aligned to a 16 byte boundary, as required for the movaps instruction:

---cut here---
class vector
{
public:
    float x;
    float y;
    float z;
    float w;

    vector(float _x, float _y, float _z, float _w)
            : x(_x), y(_y), z(_z), w(_w)
        {}
    vector()
        {}

    vector operator+(const vector &v1)
        {
            vector ret __attribute__((aligned(16)));

            asm("movl   %1,      %%esi   \n"
                "movl   %2,      %%edi   \n"
                "movaps (%%esi), %%xmm0  \n"
                "addps  (%%edi), %%xmm0  \n"
                "movaps %%xmm0,  %0      \n"
                : "=g" (ret) : "r" (this), "r" (&v1));

            return ret;
        }
} __attribute__((aligned(16)));

int main(int argc, char *argv[])
{
    vector v1(3.0f, 4.0f, 5.0f, 1.0f);
    vector v2(7.0f, 1.0f, 9.0f, 1.0f);

    vector ans = v1 + v2;
}
---cut here---

Here is the output I get:

$ g++ -v -save-temps -Wall -g sse.cpp -o sse
Reading specs from /usr/lib/gcc-lib/i586-mandrake-linux/2.95.3/specs
gcc version 2.95.3 19991030 (prerelease)
 /usr/lib/gcc-lib/i586-mandrake-linux/2.95.3/cpp -lang-c++ -v -D__GNUC__=2 -D__GNUG__=2 -D__GNUC_MINOR__=95 -D__cplusplus -D__ELF__ -Dunix -D__i386__ -Dlinux -D__ELF__ -D__unix__ -D__i386__ -D__linux__ -D__unix -D__linux -Asystem(posix) -D__EXCEPTIONS -g -Wall -Acpu(i386) -Amachine(i386) -Di386 -D__i386 -D__i386__ -Di586 -Dpentium -D__i586 -D__i586__ -D__pentium -D__pentium__ sse.cpp sse.ii
GNU CPP version 2.95.3 19991030 (prerelease) (i386 Linux/ELF)
#include "..." search starts here:
#include <...> search starts here:
 /usr/lib/gcc-lib/i586-mandrake-linux/2.95.3/../../../../include/g++-3
 /usr/local/include
 /usr/lib/gcc-lib/i586-mandrake-linux/2.95.3/../../../../i586-mandrake-linux/include
 /usr/lib/gcc-lib/i586-mandrake-linux/2.95.3/include
 /usr/include
End of search list.
The following default directories have been omitted from the search path:
End of omitted list.
 /usr/lib/gcc-lib/i586-mandrake-linux/2.95.3/cc1plus sse.ii -quiet -dumpbase sse.cc -g -Wall -version -o sse.s
GNU C++ version 2.95.3 19991030 (prerelease) (i586-mandrake-linux) compiled by GNU C version 2.95.3 19991030 (prerelease).
sse.cpp: In function `int main(int, char **)':
sse.cpp:36: warning: unused variable `class vector ans'
 as -V -Qy -o sse.o sse.s
GNU assembler version 2.10.91 (i686-mandrake-linux) using BFD version 2.10.91.0.2
 /usr/lib/gcc-lib/i586-mandrake-linux/2.95.3/collect2 -m elf_i386 -dynamic-linker /lib/ld-linux.so.2 -o sse /usr/lib/crt1.o /usr/lib/crti.o /usr/lib/gcc-lib/i586-mandrake-linux/2.95.3/crtbegin.o -L/usr/lib/gcc-lib/i586-mandrake-linux/2.95.3 -L/usr/i586-mandrake-linux/lib sse.o -lstdc++ -lm -lgcc -lc -lgcc /usr/lib/gcc-lib/i586-mandrake-linux/2.95.3/crtend.o /usr/lib/crtn.o



I also tried the "equivalent" C code and compiling with gcc but I got the same problem.

This problem also exists for these other versions of gcc:

---
Reading specs from /usr/lib/gcc-lib/i686-pc-cygwin/2.95.3-5/specs
gcc version 2.95.3-5 (cygwin special)
---
Using builtin specs.
Configured with: ../configure --prefix=/usr --mandir=/usr/share/man --infodir=/usr/share/info --enable-shared --enable-threads=posix --disable-checking --enable
-long-long --enable-cstdio=stdio --enable-clocale=generic --enable-languages=c,c++,f77,objc,java --program-suffix=-3.0 --enable-objc-gc --host=i686-mandrake-linux
Thread model: posix
gcc version 3.0
---
(But my 3.0 installation may not be correct, so you may want to ignore that)
>How-To-Repeat:
- compile the sample code on a PIII
- run it.
>Fix:
By adding three ints before my vectors:

...
    int i,j,k;
    vector ret __attribute((aligned(16)));
...
    int i,j,k;
    vector v1(...);
    vector v2(...);

I get the proper alignment and the code runs successfully.
Of course, this is a less than good solution, since as soon as I use any sort of optimisation, the alignment is not good again.
>Release-Note:
>Audit-Trail:
>Unformatted: