Bug#836844: eigen3: autopkgtests fail on ppc64el - solution

[TFAUCK]

Hello,
I found the missing information in POWER ISA

    /To use a multiply-add to perform an IEEE or Java//
    //compliant multiply, the addend must be -0.0. This//
    //is necessary to insure that the sign of a zero result//
    //will be correct when the product is -0.0 (+0.0 + -0.0//
    //􀁴 +0.0, and -0.0 + -0.0􀁴 -0.0). When the sign of a//
    //resulting 0.0 is not important, then +0.0 can be//
    //used as an addend which may, in some cases,//
    //avoid the need for a second register to hold a -0.0//
    //in addition to the integer 0/floating-point +0.0 that//
    //may already be available./

so the example shows it:

$ cat t2.c
#include "stdio.h"
#include "altivec.h"
int main() {
    vector float X={ -1, 0, 0, 0 };
    vector float Y={ -1, -1, -1, -2 };
    vector float Z, V;
*    vector float p4f_ZERO={ 0, 0, 0, 0 };**
**    vector float p4f_mZERO={ -0.0, -0.0, -0.0, -0.0 };*
    Z = vec_madd(X, Y, p4f_ZERO);
    V = vec_madd(X, Y, p4f_mZERO);
    printf("%f %f %f %f \n",Z[0], Z[1], Z[2], Z[3]);
    printf("%f %f %f %f \n",V[0], V[1], V[2], V[3]);
    vector float U;
    U= X * Y ;
    printf("%f %f %f %f \n",U[0], U[1], U[2], U[3]);
    return(0);
}

$ gcc t2.c
debian at vm18:~$ ./a.out
1.000000 0.000000 0.000000 0.000000
*1.000000 -0.000000 -0.000000 -0.000000 *
1.000000 -0.000000 -0.000000 -0.000000

So the vector p4f_ZERO must be modified with a signed value of -0.0 in
order to have that function to work.

-- 

__________________________________________________________________________
thf - Thierry Fauck - tfauck at free.fr>
/pubkey: 4096R/FCC181CE/
/fingerprint: 5CCF 6B82 DE4E E72A A40B B63E A153 BF4F FCC1 81CE/
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.alioth.debian.org/pipermail/debian-science-maintainers/attachments/20161207/1ba93fd4/attachment.html>