Another AltiVec problem

[Daniel Egger <degger at fhm dot edu>]

Hija,

consider:
#include <altivec.h>

int main ()
{
  static float storage[4] __attribute__ ((aligned(16)));

  vector float one = (vector float) {1.0, 1.0, 1.0, 1.0};
  vector float two = (vector float) {2.0, 2.0, 2.0, 2.0};
  vector float three = (vector float) {3.0, 3.0, 3.0, 3.0};
  vector float result;

  result = vec_madd (three, two, one);
  vec_st (result, 0, storage);

  printf ("%f %f %f %f\n", storage[0], storage[1], storage[2],
storage[3]);
}

given that d = vec_madd (a, b, c) is specified to yield

d = {a0 * b0 + c0, a1 * b1 + c1, a2 * b2 + c2, a3 * b3 + c3}

according to the ALTIVEC PIM on page 102 the actual
result of {5.0, 5.0, 5.0, 5.0} is cleary wrong.

Looking at the assembly:

        .file   "maddfp.c"
        .lcomm  storage.0,16,16
        .section        .rodata
        .align 4
.LC0:
        .long   1065353216  <= 1.0
        .long   1065353216
        .long   1065353216
        .long   1065353216
        .align 4
.LC1:
        .long   1073741824  <= 2.0
        .long   1073741824
        .long   1073741824
        .long   1073741824
        .align 4
.LC2:
        .long   1077936128  <= 3.0
        .long   1077936128
        .long   1077936128
        .long   1077936128
        .align 2
.LC3:
        .string "%f %f %f %f\n"
        .section        ".text"
        .align 2
        .globl main
        .type   main,@function
main:
        lis 9,.LC0@ha
        lis 11,.LC1@ha
        la 9,.LC0@l(9)  <= baseaddress of the 1.0s
        stwu 1,-16(1)
        lvx 13,0,9
        la 11,.LC1@l(11)  <= baseaddress of the 2.0s
        lis 9,.LC2@ha
        mflr 0
        la 9,.LC2@l(9)  <= baseaddress of the 3.0s
        lvx 1,0,11
        lvx 0,0,9
        lis 10,storage.0@ha
        stw 0,20(1)
        la 11,storage.0@l(10)
        li 9,0
        vmaddfp 0,0,1,13  <= result <- madd ({3.0}, {2.0}, {1.0})
        stvx 0,9,11
        lis 3,.LC3@ha
        lfs 1,storage.0@l(10)
        la 3,.LC3@l(3)
        lfs 4,12(11)
        lfs 2,4(11)
        lfs 3,8(11)
        creqv 6,6,6
        bl printf
        mr 3,0
        lwz 0,20(1)
        addi 1,1,16
        mtlr 0
        blr
.Lfe1:
        .size   main,.Lfe1-main
        .ident  "GCC: (GNU) 3.2 20020313 (experimental)"

Looking at the ALTIVEC PEM page 6-72 we can see that the
mnemonic is defined to be:
vmaddfp  vD, vA, vC, vB
where vD will be the result of vA * vC + vB
(don't ask me what those guys drank when choosing the registernames 
and encoding) and thus the assembly also seems to be correct.

I don't want to bore you with more details but the binary encoding
of the mnemonic is also correct according to the very same page.

The questions ist: where exactly is the error? And resulting thereof:
Which part of the compiler chain needs to be fixed or should it 
be fixed in the sources? 

-- 
Servus,
       Daniel