[Bug target/33671] New: incorrect vector codegen on PS3 at O3 level optimization

["j dot m dot taylor at dur dot ac dot uk" <gcc-bugzilla at gcc dot gnu dot org>]

Summary: Incorrect code is generated by gcc 4.1.1 when compiling vectorized
code on a PS3 running yellowdog linux. 

Other platforms: the code compiles correctly on a ppc macintosh, and compiles
correctly on the PS3 if sony's ppu-gcc is used.

Discussion: It seems to be an ABI issue: the caller assumes vr0 is stable
across a function call, but the callee alters vr0 and doesn't restore it. If
the function is inlined then the problem is not seen. I have reduced the
problem to a small test case. Compiler options and C source are shown below.
The variable tkr should be unaffected by the function call, but its value is
being corrupted.


$ gcc -maltivec -O3 main.c -o my_program
$ ./my_program
1.000000 1.000000 1.000000 1.000000
0.000000 0.000000 0.000000 0.000000
$ ppu-gcc -maltivec -O3 main.c -o my_program
$ ./my_program
1.000000 1.000000 1.000000 1.000000
1.000000 1.000000 1.000000 1.000000

$ gcc -v
Using built-in specs.
Target: ppu
Configured with: ../src/configure --prefix=/usr --mandir=/usr/share/man
--infodir=/usr/share/info --enable-threads --disable-checking
--with-system-zlib --enable-__cxa_atexit --disable-libunwind-exceptions
--enable-languages=c,c++,fortran --disable-nls
--enable-version-specific-runtime-libs --enable-install-libbfd
--with-long-double-128 --program-prefix=ppu- --target=ppu --enable-targets=spu
Thread model: posix
gcc version 4.1.1

$ cat main.c

#include <altivec.h>
#include <stdio.h>

typedef vector float vFloat;

static void PrintFloat(float *temp)
{
printf("%f %f %f %f\n", temp[0], temp[1], temp[2], temp[3]);
}

static void Print(vFloat v)
{
PrintFloat((float *)&v);
}

void subfunc(vFloat *outABValues) __attribute__((noinline));

/* this function wipes out v0 */
void subfunc(vFloat *outABValues)
{
vFloat zero = (vFloat) { 0.0, 0.0, 0.0, 0.0 };
outABValues[0] = zero;
}

int main (int argc, char * argv[])
{
vFloat temp;
vFloat tkr = (vFloat) { 1.0, 1.0, 1.0, 1.0 };
Print(tkr);
/* the compiler seems to assume v0 is unchanged across this function call */
subfunc(&temp);
Print(tkr);
}


-- 
           Summary: incorrect vector codegen on PS3 at O3 level optimization
           Product: gcc
           Version: 4.1.1
            Status: UNCONFIRMED
          Severity: major
          Priority: P3
         Component: target
        AssignedTo: unassigned at gcc dot gnu dot org
        ReportedBy: j dot m dot taylor at dur dot ac dot uk
  GCC host triplet: ppc64-yellowdog-linux
GCC target triplet: ppc64-yellowdog-linux


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33671