Disabling unroll-and-jam for unsigned loop is unnecessary. Maybe we
can set loop breaking
condition to "i < (n >= u) n - u : 0", which incur only a little cost
computing the upper bound
before the loop is entered.
On my Linux/x8664 box, this loop nest is unroll-and-jammed by factor
5. With -m32
flag, I got correct code other than the infinite upper bound. With
-m64, the sign of <n>
, <j> and the comparison are completely wrong, see bellow:
opencc ...-PHASE:clist=on l.c -LNO:simd=0 -LNO:pref=off ...
cat a.w2c.c:
...
for(i = 0U; (_INT32)(i) <= ((_INT32)(n) + -5); i = i + 5U)
....
Shuxin
On Tue, Sep 21, 2010 at 1:57 PM, Dror Maydan <may...@tensilica.com> wrote:
> We ran into the following bug.
>
> int a[1024];
> void r(unsigned int n)
> {
> unsigned int i,j;
> for (i=0; i<n; i++) {
> for (j=0; j<n; j++) {
> a[i]++;
> }
> }
> }
>
>
> At -O3, LNO outer unrolls the outer loop by setting the bound to
> "i<n-u+1", creating an infinite loop whenever "n" is less than the
> unroll factor. My memory is faded, but I thought it wasn't supposed to
> do that for unsigned loops. From a quick scan of the code, I didn't
> find anything that checks. Anyone have experience with this or a better
> recollection than I do?
>
> Dror
------------------------------------------------------------------------------
Nokia and AT&T present the 2010 Calling All Innovators-North America contest
Create new apps & games for the Nokia N8 for consumers in U.S. and Canada
$10 million total in prizes - $4M cash, 500 devices, nearly $6M in marketing
Develop with Nokia Qt SDK, Web Runtime, or Java and Publish to Ovi Store
http://p.sf.net/sfu/nokia-dev2dev
_______________________________________________
Open64-devel mailing list
Open64-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/open64-devel
