https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121290
Tamar Christina <tnfchris at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |tnfchris at gcc dot gnu.org
--- Comment #3 from Tamar Christina <tnfchris at gcc dot gnu.org> ---
(In reply to Soumya AR from comment #0)
>
> #define iterations 100000
> #define LEN_1D 32000
>
> float a[LEN_1D];
>
> int main()
> {
> for (int i = 0; i < LEN_1D; i++) {
> a[i] = (i * 7) % LEN_1D;
> }
> float x, chksum;
> int index;
> for (int nl = 0; nl < iterations; nl++) {
> x = a[0];
> index = 0;
> for (int i = 0; i < LEN_1D; ++i) {
> if (a[i] > x) {
> x = a[i];
> index = i;
> }
> }
> chksum = x + (float) index;
> }
>
> return index + x > 1;
> }
>
> Now:
>
> .L4:
> movi v23.4s, 0
> mov v24.16b, v26.16b
> mov x0, x3
> mov v22.16b, v23.16b
> .L3:
> ld1r {v1.4s}, [x0], 4
> fcmgt v20.4s, v1.4s, v24.4s
> bit v23.16b, v22.16b, v20.16b
> bsl v20.16b, v1.16b, v24.16b
> add v22.4s, v22.4s, v25.4s
> mov v24.16b, v20.16b
> cmp x1, x0
> bne .L3
> add w2, w2, 1
> cmp w2, w4
> bne .L4
> dup s23, v23.s[3]
> dup s20, v20.s[3]
> fmov s21, 1.0e+0
> scvtf s0, s23
> fadd s20, s0, s20
> fcmpe s20, s21
> cset w0, gt
> ret
>
At least this one seems costing related. The vectorized code
can't possible be profitable. it's performing scalar calculations
as vector by duplicating the single scalar.
Other operations like
> movi v23.4s, 0
> mov v22.16b, v23.16b
> bit v23.16b, v22.16b, v20.16b
is quite silly, as it's essentially a move of v20.16b.
But yeah this code should definitely be slower than scalar as it *is* scalar.
> Before:
> .L6:
> fmov s25, s1
> movi v26.2d, #0
> mov x0, 0
> .L5:
> ldr s0, [x1, x0, lsl 2]
> fcmpe s25, s0
> bmi .L7
> .L3:
> add x0, x0, 1
> cmp x0, x2
> bne .L5
> subs w3, w3, #1
> bne .L6
> scvtf s26, s26
> fmov s24, 1.0e+0
> fadd s26, s26, s25
> fcmpe s26, s24
> cset w0, gt
> ret
> .L7:
> fmov s26, w0
> fmov s25, s0
> b .L3
>
>
> Thanks,
> Soumya
