[Bug middle-end/99633] New: s1113 benchmark of TSVC is unrolled by icc and not by gcc and runs faster on znver3

[Bug middle-end/99633] New: s1113 benchmark of TSVC is unrolled by icc and not by gcc and runs faster on znver3

[hubicka at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99633

            Bug ID: 99633
           Summary: s1113 benchmark of TSVC is unrolled by icc and not by
                    gcc and runs faster on znver3
           Product: gcc
           Version: 11.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: hubicka at gcc dot gnu.org
  Target Milestone: ---

typedef float real_t;

#define iterations 100000
#define LEN_1D 32000
#define LEN_2D 256
// array definitions

real_t a[LEN_1D],b[LEN_1D],c[LEN_1D],d[LEN_1D],e[LEN_1D];

int main(struct args_t * func_args)
{

//    linear dependence testing
//    one iteration dependency on a(LEN_1D/2) but still vectorizable

    //initialise_arrays(__func__);
    //gettimeofday(&func_args->t1, NULL);

    for (int nl = 0; nl < 2*iterations; nl++) {
        for (int i = 0; i < LEN_1D; i++) {
            a[i] = a[LEN_1D/2] + b[i];
        }
        //dummy(a, b, c, d, e, aa, bb, cc, 0.);
    }

    return a[10];
}

Is unrolled twice by icc and runs 1.5s instead of 2.6s when built with gcc.
-funroll-loops fixes the issue, but it suggests we may want to unroll by
default on zver3

[Bug middle-end/99633] s1113 benchmark of TSVC is unrolled by icc and not by gcc and runs faster on znver3

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99633

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
     Ever confirmed|0                           |1
           Keywords|                            |missed-optimization
             Target|                            |x86_64-*-*
             Status|UNCONFIRMED                 |NEW
   Last reconfirmed|                            |2021-03-18

--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
I guess a heuristic could be to use the available load/store bandwith (for
streaming loads/stores only) when load/store 'uops' (stmts/insns) dominate
the loop.  In the case of this loop we don't even need an epilogue so that's
a plus as well.

The inner loop could also be split at LEN_1D/2 to make the load of a[LEN_1D/2]
invariant in all but a single iteration (possibly not worth the trouble in this
case).

[Bug tree-optimization/99633] s1113 benchmark of TSVC is unrolled by icc and not by gcc and runs faster on znver3

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99633

Andrew Pinski <pinskia at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|middle-end                  |tree-optimization
           Severity|normal                      |enhancement