https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92655
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
Last reconfirmed| |2019-11-25
CC| |jakub at gcc dot gnu.org,
| |rsandifo at gcc dot gnu.org
Ever confirmed|0 |1
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
The issue here is that we see
t.ii:10:76: note: ==> examining statement: _2 = x.val[j_22];
t.ii:10:76: note: precomputed vectype: vector(4) long unsigned int
t.ii:10:76: note: nunits = 4
t.ii:10:76: note: ==> examining statement: _10 = (unsigned int) _1;
t.ii:10:76: note: get vectype for scalar type: unsigned int
t.ii:10:76: note: vectype: vector(8) unsigned int
t.ii:10:76: note: nunits = 8
t.ii:10:76: note: ==> examining statement: _14 = _2 << _10;
t.ii:10:76: note: skip.
t.ii:10:76: note: ==> examining pattern def stmt: patt_38 = _1 & 4294967295;
t.ii:10:76: note: precomputed vectype: vector(4) long unsigned int
t.ii:10:76: note: nunits = 4
t.ii:10:76: note: ==> examining pattern statement: patt_37 = _2 << patt_38;
t.ii:10:76: note: precomputed vectype: vector(4) long unsigned int
...
t.ii:10:76: note: ==> examining statement: _16 = 61 - _10;
t.ii:10:76: note: get vectype for scalar type: unsigned int
t.ii:10:76: note: vectype: vector(8) unsigned int
t.ii:10:76: note: nunits = 8
t.ii:10:76: note: ==> examining statement: _17 = _2 >> _16;
t.ii:10:76: note: skip.
t.ii:10:76: note: ==> examining pattern def stmt: patt_36 = (long unsigned
int) _16;
t.ii:10:76: note: precomputed vectype: vector(4) long unsigned int
t.ii:10:76: note: get vectype for smallest scalar type: unsigned int
t.ii:10:76: note: nunits vectype: vector(8) unsigned int
t.ii:10:76: note: nunits = 8
t.ii:10:76: note: ==> examining pattern statement: patt_35 = _2 >> patt_36;
t.ii:10:76: note: precomputed vectype: vector(4) long unsigned int
t.ii:10:76: note: nunits = 4
because
<bb 3> [local count: 858993460]:
# j_22 = PHI <0(2), j_12(5)>
# ivtmp_40 = PHI <4(2), ivtmp_39(5)>
_1 = SPECIALMUL.val[j_22];
_2 = x.val[j_22];
_10 = (unsigned int) _1;
_14 = _2 << _10;
_15 = _14 & 2305843009213693951;
_16 = 61 - _10;
_17 = _2 >> _16;
_18 = _15 + _17;
<retval>.val[j_22] = _18;
j_12 = j_22 + 1;
ivtmp_39 = ivtmp_40 - 1;
if (ivtmp_39 == 0)
goto <bb 4>; [25.00%]
else
goto <bb 5>; [75.00%]
so _10 is used twice and the 2nd shift argument isn't promoted to unsigned long
by any pattern (instead it was demoted earlier). Even if fixing that
in source there's vect_recog_over_widening_pattern undoing that:
t.ii:10:76: note: vect_recog_over_widening_pattern: detected: xyz_10 = 61 -
_1;
t.ii:10:76: note: demoting long unsigned int to unsigned int
t.ii:10:76: note: created pattern stmt: patt_40 = 61 - patt_41;
:/
so it looks like those work against each other here.
