https://gcc.gnu.org/bugzilla/show_bug.cgi?id=121225
Bug ID: 121225
Summary: Missed autovectorization of bswap8 in a loop
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: dusan.stojko...@rt-rk.com
Target Milestone: ---
Hi, so a bswap8 in a loop like this:
```
void vbswap8(unsigned int* in, int len)
{
for (int i = 0; i < len; i++)
in[i] = (in[i] & 0xffff0000) | ((in[i] & 0xff00) >> 8) | ((in[i] &
0xff) << 8);
}
```
When compiled with: -O3 -march=mavx2.
The loop is not being vectorized.
Bswap finds a suitable expression to transform like so:
...
16 bit bswap implementation found at: _13 = _7 | _10;
...
load_dst_8 = MEM <short unsigned int> [(unsigned int *)_3];
bswapdst_12 = load_dst_8 r>> 8;
_13 = (unsigned int) bswapdst_12;
...
And when it is time for the vectorization pass to execute, it fails with the
following message:
...
note: can tell at compile time that MEM <short unsigned int> [(unsigned int
*)_3] and *_3 alias
missed: not vectorized: compilation time alias: load_dst_8 = MEM <short
unsigned int> [(unsigned int *)_3];
...
It seems to be related to the bswap pass, since compiling with
-fno-expensive-optimizations produces the vectorized version:
https://godbolt.org/z/ba5rfTbPP
