https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116784
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Last reconfirmed| |2024-09-20
Target Milestone|--- |15.0
Assignee|unassigned at gcc dot gnu.org |rguenth at gcc dot
gnu.org
Ever confirmed|0 |1
Status|UNCONFIRMED |ASSIGNED
Keywords| |testsuite-fail
--- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> ---
It looks like powerpc supports the 3-lane interleaving scheme we use for SLP.
It might fit vect_strided3 (but powerpc isn't amongst that).
In particular we use the following V8HImode permutes:
_39 = VEC_PERM_EXPR <vect_a_13.25_22, vect_a_13.26_8, { 2, 2, 5, 5, 8, 8, 11,
11 }>;
_40 = VEC_PERM_EXPR <vect_a_13.26_8, vect_a_13.27_6, { 6, 6, 9, 9, 12, 12,
15, 15 }>;
_41 = VEC_PERM_EXPR <_39, _40, { 0, 2, 4, 6, 8, 10, 12, 14 }>;
_35 = VEC_PERM_EXPR <vect_a_13.25_22, vect_a_13.26_8, { 1, 1, 4, 4, 7, 7, 10,
10 }>;
_36 = VEC_PERM_EXPR <vect_a_13.26_8, vect_a_13.27_6, { 5, 5, 8, 8, 11, 11,
14, 14 }>;
_37 = VEC_PERM_EXPR <_35, _36, { 0, 2, 4, 6, 8, 10, 12, 14 }>;
_32 = VEC_PERM_EXPR <vect_a_13.25_22, vect_a_13.26_8, { 0, 0, 3, 3, 6, 6, 9,
9 }>;
_33 = VEC_PERM_EXPR <vect_a_13.26_8, vect_a_13.27_6, { 4, 4, 7, 7, 10, 10,
13, 13 }>;
_34 = VEC_PERM_EXPR <_32, _33, { 0, 2, 4, 6, 8, 10, 12, 14 }>;
_45 = VEC_PERM_EXPR <vect__3.30_43, vect__4.31_44, { 0, 8, 0, 1, 9, 1, 2, 10
}>;
_46 = VEC_PERM_EXPR <vect__3.30_43, vect__4.31_44, { 2, 3, 11, 3, 4, 12, 4, 5
}>;
_47 = VEC_PERM_EXPR <vect__4.31_44, vect__3.30_43, { 5, 13, 14, 6, 14, 15, 7,
15 }>;
_49 = VEC_PERM_EXPR <_45, vect__5.32_48, { 0, 1, 8, 3, 4, 9, 6, 7 }>;
_50 = VEC_PERM_EXPR <vect__5.32_48, _46, { 2, 9, 10, 3, 12, 13, 4, 15 }>;
_51 = VEC_PERM_EXPR <_47, vect__5.32_48, { 0, 13, 2, 3, 14, 5, 6, 15 }>;
Note the testcase has the "obvious" problem that the vectorized test uses
vect_perm_short while the SLP check uses vect_perm3_short. In general
we should now be able to always SLP when we vectorize, so the checks should
match up (or be removed alltogether).
I'll see if matching up causes problems elsewhere.
