https://gcc.gnu.org/bugzilla/show_bug.cgi?id=119876
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Version|unknown |16.0
Target| |x86_64-*-*
--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
We're actually vectorizing this as
vect_pretmp_6.14_58 = MEM <vector(16) int> [(int *)vectp_c.12_56];
vect__28.15_59 = VIEW_CONVERT_EXPR<vector(16) unsigned
int>(vect_pretmp_6.14_58);
vect__31.18_62 = vect__28.15_59 + { 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
2, 2 };
vect_iftmp.19_63 = VIEW_CONVERT_EXPR<vector(16) int>(vect__31.18_62);
vect__29.16_60 = vect__28.15_59 + { 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1 };
vect_iftmp.17_61 = VIEW_CONVERT_EXPR<vector(16) int>(vect__29.16_60);
vect_iftmp.20_64 = VEC_COND_EXPR <mask__4.11_55, vect_iftmp.17_61,
vect_iftmp.19_63>;
MEM <vector(16) int> [(int *)vectp_a.21_65] = vect_iftmp.20_64;
where clang seems to vectorize
a[i] = c[i] + (b[i] > 0) ? 1 : 2;
Using a merge-masking add is of course expensive. Folding turns the above to
vect_pretmp_6.14_58 = MEM <vector(16) int> [(int *)&c + ivtmp.44_72 * 1];
vect__28.15_59 = VIEW_CONVERT_EXPR<vector(16) unsigned
int>(vect_pretmp_6.14_58);
vect__31.18_62 = vect__28.15_59 + { 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2,
2, 2 };
_52 = .COND_ADD (mask__4.11_55, vect__28.15_59, { 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1 }, vect__31.18_62);
vect_iftmp.20_64 = VIEW_CONVERT_EXPR<vector(16) int>(_52);
MEM <vector(16) int> [(int *)&a + ivtmp.44_72 * 1] = vect_iftmp.20_64;
I think if-conversion (not phiopt) should have linearized
pretmp_6 = c[i_14];
if (_1 > 0)
goto <bb 4>; [59.00%]
else
goto <bb 5>; [41.00%]
<bb 4> [local count: 627172604]:
iftmp.0_9 = pretmp_6 + 1;
goto <bb 6>; [100.00%]
<bb 5> [local count: 435831803]:
iftmp.0_8 = pretmp_6 + 2;
<bb 6> [local count: 1063004410]:
# iftmp.0_5 = PHI <iftmp.0_9(4), iftmp.0_8(5)>
as an add of a conditional 1 or 2 instead (possibly using folding and match
during building of the COND_EXPR). Note it requires PRE code hoisting to
hoist the load out of the conditional if/else. One could argue we miss
a phiopt after the PRE/SINK/DSE/DCE pass group before loop opts or that
ifcvt should also run (parts of) PHI-OPT first.
