[Bug tree-optimization/101105] [11/12 Regression] wrong code at -O3 on x86_64-linux-gnu
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101105 --- Comment #6 from CVS Commits --- The master branch has been updated by Richard Biener : https://gcc.gnu.org/g:50374fdacbd121bc4a61b073e559208ff61bee06 commit r12-1765-g50374fdacbd121bc4a61b073e559208ff61bee06 Author: Richard Biener Date: Wed Jun 23 12:43:03 2021 +0200 tree-optimization/101105 - fix runtime alias test optimization We were ignoring DR_STEP for VF == 1 which is OK only in case the scalar order is preserved or both DR steps are the same. 2021-06-23 Richard Biener PR tree-optimization/101105 * tree-vect-data-refs.c (vect_prune_runtime_alias_test_list): Only ignore steps when they are equal or scalar order is preserved. * gcc.dg/torture/pr101105.c: New testcase.
[Bug tree-optimization/101105] [11/12 Regression] wrong code at -O3 on x86_64-linux-gnu
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101105 --- Comment #5 from Richard Biener --- Created attachment 51055 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=51055=edit patch So like this.
[Bug tree-optimization/101105] [11/12 Regression] wrong code at -O3 on x86_64-linux-gnu
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101105 rsandifo at gcc dot gnu.org changed: What|Removed |Added Status|NEW |ASSIGNED Assignee|unassigned at gcc dot gnu.org |rsandifo at gcc dot gnu.org --- Comment #4 from rsandifo at gcc dot gnu.org --- Mine.
[Bug tree-optimization/101105] [11/12 Regression] wrong code at -O3 on x86_64-linux-gnu
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101105
--- Comment #3 from Richard Biener ---
So what happens is that vect_compile_time_alias fails to perform the offset
adjustment for the negative step DR
#(Data Ref:
# bb: 3
# stmt: b[g_40][0] = 0;
# ref: b[g_40][0];
# base_object: b;
# Access function 0: 0
# Access function 1: {3, +, -1}_1
which is done via
/* For negative step, we need to adjust address range by TYPE_SIZE_UNIT
bytes, e.g., int a[3] -> a[1] range is [a+4, a+16) instead of
[a, a+12) */
if (tree_int_cst_compare (DR_STEP (a->dr), size_zero_node) < 0)
{
const_length_a = (-wi::to_poly_wide (segment_length_a)).force_uhwi ();
offset_a -= const_length_a;
}
since we zero segment_length_a because of ignore_step_p. But that adjustment
cannot be ignored. I suppose we need to track a separate "offset segment
length"
for this purpose?
[Bug tree-optimization/101105] [11/12 Regression] wrong code at -O3 on x86_64-linux-gnu
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101105
Richard Biener changed:
What|Removed |Added
CC|rguenther at suse dot de |rguenth at gcc dot
gnu.org,
||rsandifo at gcc dot gnu.org
Status|UNCONFIRMED |NEW
Ever confirmed|0 |1
Last reconfirmed||2021-06-17
--- Comment #2 from Richard Biener ---
Confirmed.
short a;
int b[5][4] = {2, 2};
int d;
short e(int f) { return f == 0 || a && f == 1 ? 0 : a; }
int main() {
int g, h;
g = 3;
for (; g >= 0; g--) {
h = 3;
for (; h >= 0; h--)
b[g][h] = b[0][1] && e(1);
}
d = b[0][1];
if (d != 0)
__builtin_abort ();
return 0;
}
after if-conversion we have
[local count: 214748368]:
# g_40 = PHI
# ivtmp_2 = PHI
b[g_40][2] = 0;
b[g_40][3] = 0;
b[g_40][1] = 0;
_42 = b[0][1];
b[g_40][0] = 0;
g_11 = g_40 + -1;
ivtmp_27 = ivtmp_2 - 1;
if (ivtmp_27 != 0)
goto ; [75.00%]
else
goto ; [25.00%]
[local count: 161061274]:
goto ; [100.00%]
[local count: 53687093]:
# _18 = PHI <_42(3)>
d = _18;
if (_18 != 0)
and vect produces
[local count: 214748368]:
# g_40 = PHI
# ivtmp_2 = PHI
# ivtmp_48 = PHI
# ivtmp_32 = PHI
_42 = b[0][1];
MEM [(int *)ivtmp_48] = { 0, 0, 0, 0 };
ivtmp_31 = ivtmp_48 + 18446744073709551600(OVF);
g_11 = g_40 + -1;
ivtmp_27 = ivtmp_2 - 1;
ivtmp_26 = ivtmp_48 + 18446744073709551600(OVF);
ivtmp_12 = ivtmp_32 + 1;
if (ivtmp_12 < 4)
goto ; [75.00%]
else
goto ; [25.00%]
[local count: 161061274]:
goto ; [100.00%]
[local count: 53687093]:
# _18 = PHI <_42(3)>
We have
(Data Dep:
#(Data Ref:
# bb: 3
# stmt: b[g_40][1] = 0;
# ref: b[g_40][1];
# base_object: b;
# Access function 0: 1
# Access function 1: {3, +, -1}_1
#)
#(Data Ref:
# bb: 3
# stmt: _42 = b[0][1];
# ref: b[0][1];
# base_object: b;
# Access function 0: 1
# Access function 1: 0
#)
access_fn_A: 1
access_fn_B: 1
(subscript
iterations_that_access_an_element_twice_in_A: [0]
last_conflict: scev_not_known
iterations_that_access_an_element_twice_in_B: [0]
last_conflict: scev_not_known
(Subscript distance: 0 ))
access_fn_A: {3, +, -1}_1
access_fn_B: 0
(subscript
iterations_that_access_an_element_twice_in_A: [-3]
last_conflict: 1
iterations_that_access_an_element_twice_in_B: [0]
last_conflict: 1
(Subscript distance: -3 ))
loop nest: (1 )
)
and mark the DR for a runtime alias test. But then
t3.c:8:12: note: === vect_prune_runtime_alias_test_list ===
t3.c:8:12: note: can tell at compile time that b[g_40][0] and b[0][1] do not
alias
t3.c:8:12: note: improved number of alias checks from 1 to 0
as vect_compile_time_alias gets segment_lengths of zero (huh), since
/* Step values are irrelevant for aliasing if the number of vector
iterations is equal to the number of scalar iterations (which can
happen for fully-SLP loops). */
bool ignore_step_p = known_eq (LOOP_VINFO_VECT_FACTOR (loop_vinfo), 1U);
...
if (ignore_step_p)
{
segment_length_a = size_zero_node;
segment_length_b = size_zero_node;
}
but I'm not sure if a segment length of zero will do good in any case where
the alias check is needed? Richard - that's your code, can you have a look?
[Bug tree-optimization/101105] [11/12 Regression] wrong code at -O3 on x86_64-linux-gnu
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=101105 Andrew Pinski changed: What|Removed |Added Keywords||wrong-code Version|unknown |11.0 Target||x86_64-linux-gnu Summary|wrong code at -O3 on|[11/12 Regression] wrong |x86_64-linux-gnu|code at -O3 on ||x86_64-linux-gnu Target Milestone|--- |11.2
