http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49513
--- Comment #2 from vincenzo Innocente <vincenzo.innocente at cern dot ch>
2011-06-23 12:16:59 UTC ---
thanks for the fast analysis.
the code in question was there also to test what the vectorized code would do
for x=0 and y=0 (it is extracted from a simplified version of atan2f that gcc
can vectorize).
Being a test, I can workaround (I've to test extreme conditions explicitely
anyhow).
In any case I think that any peeling will ruin alignmement.
hoping gcc can find ways to avoid PRE and vectorization to clash.
vincenzo
On 23 Jun, 2011, at 1:55 PM, rguenth at gcc dot gnu.org wrote:
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=49513
>
> Richard Guenther <rguenth at gcc dot gnu.org> changed:
>
> What |Removed |Added
> ----------------------------------------------------------------------------
> Status|UNCONFIRMED |NEW
> Keywords| |missed-optimization
> Last reconfirmed| |2011.06.23 11:55:09
> CC| |rguenth at gcc dot gnu.org
> Ever Confirmed|0 |1
> Summary|introducing a product |PRE inhibits if-conversion
> |inhibit vectorization |and vectorization
>
> --- Comment #1 from Richard Guenther <rguenth at gcc dot gnu.org> 2011-06-23
> 11:55:09 UTC ---
> That is because PRE decided to optimize the first iteration where it knows
> that z == 0 and thus z*s[i] and z*c[i] are 0 and thus a[i] will be 0. Which
> means we confuse if-conversion which in turn causes this missed vectorization.
> IL before if-conversion:
>
> <bb 2>:
> goto <bb 6>;
>
> <bb 3>:
> z_3 = (float) i_9;
> D.3280_4 = c[i_9];
> D.3281_5 = D.3280_4 * z_3;
> D.3282_6 = s[i_9];
> D.3283_7 = D.3282_6 * z_3;
> yy_13 = ABS_EXPR <D.3281_5>;
> yy_14 = ABS_EXPR <D.3283_7>;
> if (yy_13 < yy_14)
> goto <bb 5>;
> else
> goto <bb 4>;
>
> <bb 4>:
>
> <bb 5>:
> # yy_28 = PHI <yy_13(4), yy_14(3)>
> # yy_29 = PHI <yy_14(4), yy_13(3)>
>
> <bb 6>:
> # yy_16 = PHI <yy_28(5), 0.0(2)>
> # yy_15 = PHI <yy_29(5), 0.0(2)>
> # i_30 = PHI <i_9(5), 0(2)>
> # ivtmp.44_24 = PHI <ivtmp.44_10(5), 1024(2)>
> t_17 = yy_15 / yy_16;
> a[i_30] = t_17;
> i_9 = i_30 + 1;
> ivtmp.44_10 = ivtmp.44_24 - 1;
> if (ivtmp.44_10 != 0)
> goto <bb 3>;
> else
> goto <bb 7>;
>
> <bb 7>:
> return;
>
> The PHI nodes in bb 6 is what makes if-conversion fail (thus, the
> irregular loop entry which really should be peeled off).
>
> This situation commonly occurs when PRE can compute the first iterations
> result. Manually peeling off the iteration like the following is a
> workaround:
>
> void foo2() {
> a[0] = 0;
> for (int i=1; i!=1024; ++i) {
> float z = i;
> a[i] = bar(z*s[i],z*c[i]);
> }
> }
>
> Not sure if your original issue you derived this testcase from really
> matches the above problem though.
>
> --
> Configure bugmail: http://gcc.gnu.org/bugzilla/userprefs.cgi?tab=email
> ------- You are receiving this mail because: -------
> You reported the bug.
--
Il est bon de suivre sa pente, pourvu que ce soit en montant.
A.G.
http://www.flickr.com/photos/vin60/1320965757/
