https://gcc.gnu.org/bugzilla/show_bug.cgi?id=107160
--- Comment #8 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Kewen Lin from comment #7)
> One reduced test case is:
>
> ============================================================
>
> #include <stdio.h>
> #include <math.h>
>
> #define N 128
> float fl[N];
>
> __attribute__ ((noipa, optimize (0))) void
> init ()
> {
> for (int i = 0; i < N; i++)
> fl[i] = i;
> }
>
> __attribute__ ((noipa)) float
> foo (int n1)
> {
> float sum0, sum1, sum2, sum3;
> sum0 = sum1 = sum2 = sum3 = 0.0f;
>
> int n = (n1 / 4) * 4;
> for (int i = 0; i < n; i += 4)
> {
> sum0 += fabs (fl[i]);
> sum1 += fabs (fl[i + 1]);
> sum2 += fabs (fl[i + 2]);
> sum3 += fabs (fl[i + 3]);
> }
>
> return sum0 + sum1 + sum2 + sum3;
> }
>
> __attribute__ ((optimize (0))) int
> main ()
> {
> init ();
> float res = foo (80);
> __builtin_printf ("res:%f\n", res);
> return 0;
> }
>
> ============================================================
> incorrect result "res:670.000000" vs expected result "res:3160.000000"
>
> It looks it exposes one bug in vectorization reduction support. The
> reduction epilogue handling looks wrong, it generates gimple code like:
>
> # vect_sum3_31.16_101 = PHI <vect_sum3_31.16_97(3)>
> # vect_sum3_31.16_102 = PHI <vect_sum3_31.16_98(3)>
> # vect_sum3_31.16_103 = PHI <vect_sum3_31.16_99(3)>
> # vect_sum3_31.16_104 = PHI <vect_sum3_31.16_100(3)>
> _105 = BIT_FIELD_REF <vect_sum3_31.16_101, 32, 0>;
> _106 = BIT_FIELD_REF <vect_sum3_31.16_101, 32, 32>;
> _107 = BIT_FIELD_REF <vect_sum3_31.16_101, 32, 64>;
> _108 = BIT_FIELD_REF <vect_sum3_31.16_101, 32, 96>;
> _109 = BIT_FIELD_REF <vect_sum3_31.16_102, 32, 0>;
> _110 = BIT_FIELD_REF <vect_sum3_31.16_102, 32, 32>;
> _111 = BIT_FIELD_REF <vect_sum3_31.16_102, 32, 64>;
> _112 = BIT_FIELD_REF <vect_sum3_31.16_102, 32, 96>;
> ...
>
> it doesn't consider the reduced results vect_sum3_31.16_10{1,2,3,4} from the
> loop can be reduced again in loop exit block as they are in the same slp
> group.
The above doesn't look wrong (but may miss the rest of the IL). On
x86_64 this looks like
<bb 4> [local count: 105119324]:
# sum0_41 = PHI <sum0_28(3)>
# sum1_39 = PHI <sum1_29(3)>
# sum2_37 = PHI <sum2_30(3)>
# sum3_35 = PHI <sum3_31(3)>
# vect_sum3_31.11_59 = PHI <vect_sum3_31.11_60(3)>
_58 = BIT_FIELD_REF <vect_sum3_31.11_59, 32, 0>;
_57 = BIT_FIELD_REF <vect_sum3_31.11_59, 32, 32>;
_56 = BIT_FIELD_REF <vect_sum3_31.11_59, 32, 64>;
_55 = BIT_FIELD_REF <vect_sum3_31.11_59, 32, 96>;
_74 = _58 + _57;
_76 = _56 + _74;
_78 = _55 + _76;
<bb 5> [local count: 118111600]:
# prephitmp_79 = PHI <_78(4), 0.0(2)>
return prephitmp_79;
when unrolling is applied, thus with a larger VF, you should ideally
see the vectors accumulated.
Btw, I've fixed a SLP reduction issue two days ago in r13-3226-gee467644c53ee2
though that looks unrelated?
When I force a larger VF on x86 by adding a int store in the loop I see
<bb 11> [local count: 94607391]:
# sum0_48 = PHI <sum0_29(3)>
# sum1_36 = PHI <sum1_30(3)>
# sum2_35 = PHI <sum2_31(3)>
# sum3_24 = PHI <sum3_32(3)>
# vect_sum3_32.16_110 = PHI <vect_sum3_32.16_106(3)>
# vect_sum3_32.16_111 = PHI <vect_sum3_32.16_107(3)>
# vect_sum3_32.16_112 = PHI <vect_sum3_32.16_108(3)>
# vect_sum3_32.16_113 = PHI <vect_sum3_32.16_109(3)>
_114 = BIT_FIELD_REF <vect_sum3_32.16_110, 32, 0>;
_115 = BIT_FIELD_REF <vect_sum3_32.16_110, 32, 32>;
_116 = BIT_FIELD_REF <vect_sum3_32.16_110, 32, 64>;
_117 = BIT_FIELD_REF <vect_sum3_32.16_110, 32, 96>;
_118 = BIT_FIELD_REF <vect_sum3_32.16_111, 32, 0>;
_119 = BIT_FIELD_REF <vect_sum3_32.16_111, 32, 32>;
_120 = BIT_FIELD_REF <vect_sum3_32.16_111, 32, 64>;
_121 = BIT_FIELD_REF <vect_sum3_32.16_111, 32, 96>;
_122 = BIT_FIELD_REF <vect_sum3_32.16_112, 32, 0>;
_123 = BIT_FIELD_REF <vect_sum3_32.16_112, 32, 32>;
_124 = BIT_FIELD_REF <vect_sum3_32.16_112, 32, 64>;
_125 = BIT_FIELD_REF <vect_sum3_32.16_112, 32, 96>;
_126 = BIT_FIELD_REF <vect_sum3_32.16_113, 32, 0>;
_127 = BIT_FIELD_REF <vect_sum3_32.16_113, 32, 32>;
_128 = BIT_FIELD_REF <vect_sum3_32.16_113, 32, 64>;
_129 = BIT_FIELD_REF <vect_sum3_32.16_113, 32, 96>;
_130 = _114 + _118;
_131 = _115 + _119;
_132 = _116 + _120;
_133 = _117 + _121;
_134 = _130 + _122;
_135 = _131 + _123;
_136 = _132 + _124;
_137 = _133 + _125;
_138 = _134 + _126;
see how the lanes from the different vectors are accumulated? (yeah,
we should simply add the vectors!)
