[Bug tree-optimization/83518] [8 Regression] Missing optimization: useless instructions should be dropped
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83518 Richard Biener changed: What|Removed |Added Priority|P3 |P2 Target Milestone|8.0 |9.0 --- Comment #6 from Richard Biener --- Re-targeting to GCC9.
[Bug tree-optimization/83518] [8 Regression] Missing optimization: useless instructions should be dropped
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83518
--- Comment #5 from rguenther at suse dot de ---
On Mon, 8 Jan 2018, jakub at gcc dot gnu.org wrote:
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83518
>
> --- Comment #4 from Jakub Jelinek ---
> Before store-merging we have:
> MEM[(int *)] = { 5, 4, 3, 2 };
> t_2 = arr[0];
> _65 = arr[1];
> _69 = arr[2];
> _73 = arr[3];
> arr[0] = _69;
> arr[1] = _73;
> arr[2] = 1;
> arr[3] = t_2;
> vect__2.5_38 = MEM[(int *)];
> and all store-merging can do with this is what it does:
> MEM[(int *)] = { 5, 4, 3, 2 };
> t_2 = arr[0];
> _65 = arr[1];
> _69 = arr[2];
> _73 = arr[3];
> _46 = MEM[(int *) + 8B];
> MEM[(int *)] = _46;
> arr[2] = 1;
> arr[3] = t_2;
> vect__2.5_38 = MEM[(int *)];
> where the _69 and _73 sets can be DCEd later. store-merging has no framework
> like FRE to do analysis what memory location contains at which point.
> So we'd need another late FRE pass to handle this?
Without enhancing FRE that wouldn't help. I think the best thing is
to try sanitizing the IL produced by unrolling so SLP vectorization
can do its job here.
Nothing for GCC 8 though.
[Bug tree-optimization/83518] [8 Regression] Missing optimization: useless instructions should be dropped
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83518
--- Comment #4 from Jakub Jelinek ---
Before store-merging we have:
MEM[(int *)] = { 5, 4, 3, 2 };
t_2 = arr[0];
_65 = arr[1];
_69 = arr[2];
_73 = arr[3];
arr[0] = _69;
arr[1] = _73;
arr[2] = 1;
arr[3] = t_2;
vect__2.5_38 = MEM[(int *)];
and all store-merging can do with this is what it does:
MEM[(int *)] = { 5, 4, 3, 2 };
t_2 = arr[0];
_65 = arr[1];
_69 = arr[2];
_73 = arr[3];
_46 = MEM[(int *) + 8B];
MEM[(int *)] = _46;
arr[2] = 1;
arr[3] = t_2;
vect__2.5_38 = MEM[(int *)];
where the _69 and _73 sets can be DCEd later. store-merging has no framework
like FRE to do analysis what memory location contains at which point.
So we'd need another late FRE pass to handle this?
[Bug tree-optimization/83518] [8 Regression] Missing optimization: useless instructions should be dropped
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83518
--- Comment #3 from Richard Biener ---
Yes, it's not really sth new but a known issue with late value-numbering. Note
that FRE wouldn't know how to simplify this either, we'd need store-merging
to effectively vectorize the earlier sets. BB vectorization doesn't do this
because after unrolling we see
vect_cst__46 = { 5, 4, 3, 2 };
MEM[(int *)] = vect_cst__46;
arr[4] = 1;
t_2 = arr[0];
arr[0] = 5;
arr[0] = t_2;
t_32 = arr[0];
_65 = arr[1];
arr[0] = _65;
arr[1] = t_32;
t_68 = arr[0];
_69 = arr[2];
arr[0] = _69;
arr[2] = t_68;
t_72 = arr[0];
_73 = arr[3];
arr[0] = _73;
arr[3] = t_72;
t_76 = arr[0];
_77 = arr[4];
arr[0] = _77;
arr[4] = t_76;
i_80 = 1;
ivtmp_81 = 4;
pretmp_82 = arr[0];
t_87 = arr[i_80];
arr[i_80] = pretmp_82;
...
and BB vectorization is confused by the dead stores (and DSE would be
by the missed constant propagations).
[Bug tree-optimization/83518] [8 Regression] Missing optimization: useless instructions should be dropped
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83518 Jakub Jelinek changed: What|Removed |Added CC||jakub at gcc dot gnu.org --- Comment #2 from Jakub Jelinek --- We do vectorize this and thus not unroll it since r253975. With -O3 -fno-vect-cost-model we've generated what GCC 8 emits or something similar for many years, including 4.4+ (r140264 already shows this behavior).
[Bug tree-optimization/83518] [8 Regression] Missing optimization: useless instructions should be dropped
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=83518
Richard Biener changed:
What|Removed |Added
Status|UNCONFIRMED |NEW
Known to work||7.2.1
Version|tree-ssa|8.0
Keywords||missed-optimization
Last reconfirmed||2018-01-05
CC||rguenth at gcc dot gnu.org
Ever confirmed|0 |1
Summary|Missing optimization: |[8 Regression] Missing
|useless instructions should |optimization: useless
|be dropped |instructions should be
||dropped
Target Milestone|--- |8.0
--- Comment #1 from Richard Biener ---
Works fine with GCC 7. I suppose unroller limits hit and/or we're just lucky
that GCC 7 doesn't vectorize the reduction loop ...
Trunk has
t.C:16:20: note: loop vectorized
t.C:21:10: note: basic block vectorized
resulting in
[local count: 178992762]:
MEM[(int *)] = { 5, 4, 3, 2 };
t_2 = arr[0];
_65 = arr[1];
_46 = MEM[(int *) + 8B];
MEM[(int *)] = _46;
arr[2] = 1;
arr[3] = t_2;
vect__2.5_38 = MEM[(int *)];
vect_sum_21.8_30 = VEC_PERM_EXPR ;
vect_sum_21.8_15 = vect_sum_21.8_30 + vect__2.5_38;
vect_sum_21.8_59 = VEC_PERM_EXPR ;
vect_sum_21.8_60 = vect_sum_21.8_15 + vect_sum_21.8_59;
stmp_sum_21.7_61 = BIT_FIELD_REF ;
sum_27 = stmp_sum_21.7_61 + _65;
_23 = (unsigned int) sum_27;
arr ={v} {CLOBBER};
return _23;
while GCC 7 simply unrolls the loop. DOM is not able to simplify the
vector load from MEM[(int *)] but the scalar loads from the unrolled
variant.
