On Tue, Feb 26, 2008 at 12:12 PM, Revital1 Eres <[EMAIL PROTECTED]> wrote:
>
> Hello,
>
> I am running the attached testcase (inspired from vect/vect-reduc-3.c
> testcase) with -O3 -fwrapv on powerpc64-linux with trunk 4.4.
>
> Here is a snippet from the testcase:
>
> ...
>
> unsigned short ub[N] = {0,3,6,9,12,15,18,21,24,27,30,33,36,39,42,45};
> unsigned short uc[N] = {0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15};
> unsigned short udiff;
>
>
> udiff = 0;
> for (i = 0; i < n; i++) {,
> udiff += (ub[i] - uc[i]);,
> }
>
This is done by fold (see the .original dump), both operands decay to int
according to C rules, so fold sees (unsigned short)((int)ub[i] -
(int)uc[i] + (int)udiff)
and thinks that (short unsigned int) ((short int) (ub[i] - uc[i]) +
(short int) udiff)
is a better way to compute it.
Richard.
