https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109971
Kewen Lin <linkw at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Ever confirmed|0 |1
Status|UNCONFIRMED |ASSIGNED
Last reconfirmed| |2023-05-26
--- Comment #5 from Kewen Lin <linkw at gcc dot gnu.org> ---
(In reply to JuzheZhong from comment #4)
> (In reply to Kewen Lin from comment #3)
> > I'll take a look first.
>
> Thanks a lot. I am sorry for causing such issue to you.
Never mind! Some failures can't be even caught by normal testings, or not
responsible by the culprit patch itself but just exposed by it instead.
As your comment #c2, it seems that you want to disable this on Power (and s390)
for now? (It's disabled for s390 apparently since it has
LOOP_VINFO_PARTIAL_LOAD_STORE_BIAS 1 always).
After some checkings, I found that:
1) for failures on p9-vec-length-full-{1,2,6}.c, the root cause is that the
main loop becomes neat and rtl pass bbro is able to duplicate it, the expected
counts on vector with length instructions change accordingly, I think they are
test issues.
With decrement IV, the optimized IR actually becomes better, it also aligns
with our discussion here:
https://gcc.gnu.org/pipermail/gcc-patches/2023-April/615629.html (Thanks for
the improvement!)
For example on full-1.c int8_t type:
<bb 3> [local count: 75161909]:
# vectp_a_int8_t.4_18 = PHI <vectp_a_int8_t.4_17(5), &a_int8_t(2)>
# vectp_b_int8_t.8_8 = PHI <vectp_b_int8_t.8_7(5), &b_int8_t(2)>
# vectp_c_int8_t.14_26 = PHI <vectp_c_int8_t.14_27(5), &c_int8_t(2)>
# ivtmp_29 = PHI <ivtmp_30(5), 0(2)>
# loop_len_16 = PHI <_34(5), 16(2)>
vect__1.6_13 = .LEN_LOAD (vectp_a_int8_t.4_18, 8B, loop_len_16, 0);
vect__2.7_12 = VIEW_CONVERT_EXPR<vector(16) unsigned char>(vect__1.6_13);
vect__3.10_22 = .LEN_LOAD (vectp_b_int8_t.8_8, 8B, loop_len_16, 0);
vect__4.11_23 = VIEW_CONVERT_EXPR<vector(16) unsigned char>(vect__3.10_22);
vect__5.12_24 = vect__2.7_12 + vect__4.11_23;
vect__6.13_25 = VIEW_CONVERT_EXPR<vector(16) signed char>(vect__5.12_24);
.LEN_STORE (vectp_c_int8_t.14_26, 8B, loop_len_16, vect__6.13_25, 0);
vectp_a_int8_t.4_17 = vectp_a_int8_t.4_18 + 16;
vectp_b_int8_t.8_7 = vectp_b_int8_t.8_8 + 16;
vectp_c_int8_t.14_27 = vectp_c_int8_t.14_26 + 16;
ivtmp_30 = ivtmp_29 + 16;
_32 = MIN_EXPR <ivtmp_30, 127>;
_33 = 127 - _32;
_34 = MIN_EXPR <_33, 16>;
if (ivtmp_30 <= 126)
goto <bb 5>; [85.71%]
else
goto <bb 4>; [14.29%]
vs.
<bb 3> [local count: 75161909]:
# vectp_a_int8_t.4_18 = PHI <vectp_a_int8_t.4_17(5), &a_int8_t(2)>
# vectp_b_int8_t.8_8 = PHI <vectp_b_int8_t.8_7(5), &b_int8_t(2)>
# vectp_c_int8_t.14_26 = PHI <vectp_c_int8_t.14_27(5), &c_int8_t(2)>
# ivtmp_29 = PHI <ivtmp_30(5), 127(2)>
loop_len_16 = MIN_EXPR <ivtmp_29, 16>;
vect__1.6_13 = .LEN_LOAD (vectp_a_int8_t.4_18, 8B, loop_len_16, 0);
vect__2.7_12 = VIEW_CONVERT_EXPR<vector(16) unsigned char>(vect__1.6_13);
vect__3.10_22 = .LEN_LOAD (vectp_b_int8_t.8_8, 8B, loop_len_16, 0);
vect__4.11_23 = VIEW_CONVERT_EXPR<vector(16) unsigned char>(vect__3.10_22);
vect__5.12_24 = vect__2.7_12 + vect__4.11_23;
vect__6.13_25 = VIEW_CONVERT_EXPR<vector(16) signed char>(vect__5.12_24);
.LEN_STORE (vectp_c_int8_t.14_26, 8B, loop_len_16, vect__6.13_25, 0);
vectp_a_int8_t.4_17 = vectp_a_int8_t.4_18 + 16;
vectp_b_int8_t.8_7 = vectp_b_int8_t.8_8 + 16;
vectp_c_int8_t.14_27 = vectp_c_int8_t.14_26 + 16;
ivtmp_30 = ivtmp_29 - loop_len_16;
if (ivtmp_30 != 0)
goto <bb 5>; [85.71%]
else
goto <bb 4>; [14.29%]
2) for failure on p9-vec-length-full-7.c ({u,}int8_t), the IR difference causes
cunroll not to unroll the loop further, so IR has some differences during
optimized dumpings:
<bb 4> [local count: 18146240]:
MEM <vector(16) signed char> [(signed char *)&x_int8_t + 16B] = { 15, 16, 17,
18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30 };
MEM <vector(16) signed char> [(signed char *)&x_int8_t + 32B] = { 31, 32, 33,
34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46 };
.LEN_STORE (&MEM <int8_t[64]> [(void *)&x_int8_t + 48B], 128B, 11, { 47, 48,
49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62 }, 0); [tail call]
return;
vs.
<bb 5> [local count: 72584963]:
# vect_vec_iv_.6_50 = PHI <_51(5), { 15, 16, 17, 18, 19, 20, 21, 22, 23, 24,
25, 26, 27, 28, 29, 30 }(4)>
# ivtmp_57 = PHI <ivtmp_58(5), 43(4)>
# ivtmp.12_11 = PHI <ivtmp.12_22(5), ivtmp.12_26(4)>
loop_len_55 = MIN_EXPR <ivtmp_57, 16>;
_51 = vect_vec_iv_.6_50 + { 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16, 16,
16, 16, 16, 16 };
_5 = (void *) ivtmp.12_11;
_14 = &MEM <vector(16) signed char> [(signed char *)_5];
.LEN_STORE (_14, 128B, loop_len_55, vect_vec_iv_.6_50, 0);
ivtmp_58 = ivtmp_57 - loop_len_55;
ivtmp.12_22 = ivtmp.12_11 + 16;
if (ivtmp_58 != 0)
goto <bb 5>; [75.00%]
else
goto <bb 6>; [25.00%]
It exposes something inefficient at -O2, it seems we can teach cunroll further
about this kind of new sequence.
If you meant to disable decrement IV on Power (but now actually enable it
unexpectedly), then probably we can just keep it (not disabling), for Power we
mainly adopt --param=vect-partial-vector-usage=1, it shouldn't be affected, for
--param=vect-partial-vector-usage=2, it does generate better code sequence for
most cases and we can improve the remaining worse one gradually.
