https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98674
Bug ID: 98674
Summary: [10/11] Regression vectorizer failed for compilation
time alias
Product: gcc
Version: 11.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: crazylht at gmail dot com
Target Milestone: ---
Host: x86_64-pc-linux-gnu
Cat test.c, refer to https://godbolt.org/z/fx9cMq
---
void swap(short *p, int cnt) {
cnt = 1000;
while (cnt-- > 0) {
*p = ((*p << 8) & 0xFF00) |
((*p >> 8) & 0x00FF);
++p;
}
}
----
test.c:3:16: missed: couldn't vectorize loop
test.c:4:8: missed: not vectorized: compilation time alias: load_dst_23 =
MEM[(short int *)p_21];
*p_21 = _8;
but they have the same address, and of course they alias.
---
<bb 3> [local count: 1063004409]:
# p_21 = PHI <p_16(5), p_12(D)(2)>
# cnt_24 = PHI <cnt_14(5), 999(2)>
# ivtmp_26 = PHI <ivtmp_2(5), 1000(2)>
load_dst_23 = MEM[(short int *)p_21];
bswapdst_11 = load_dst_23 r>> 8;
_8 = (short int) bswapdst_11;
*p_21 = _8;
p_16 = p_21 + 2;
cnt_14 = cnt_24 + -1;
ivtmp_2 = ivtmp_26 - 1;
if (ivtmp_2 != 0)
----
So does vectorizer assume there won't be 2 DRs with same reference?
Successfully vectorized with below hack
---
modified gcc/tree-vect-data-refs.c
@@ -3302,6 +3302,10 @@ vect_compile_time_alias (dr_vec_info *a, dr_vec_info *b,
const_length_a += access_size_a;
const_length_b += access_size_b;
+ /* It's ok for the same reference. */
+ if (known_eq (const_length_a, const_length_b))
+ return 0;
+
if (ranges_known_overlap_p (offset_a, const_length_a,
offset_b, const_length_b))
return 1;
---
