https://gcc.gnu.org/bugzilla/show_bug.cgi?id=118311
Bug ID: 118311
Summary: Poorly optimized trivial integer serialization due to
vectorizer on aarch64
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: pinskia at gcc dot gnu.org
CC: marco.rubini08 at gmail dot com, unassigned at gcc dot
gnu.org
Target Milestone: ---
Target: aarch64
Codegen is suboptimal for the following code.
Note that the output improves drastically after uncommenting the commented
line.
```
void f(unsigned char* dst,
unsigned long long low,
unsigned long long high)
{
dst[0] = ((low >> 0) & 0xff);
dst[1] = ((low >> 8) & 0xff);
dst[2] = ((low >> 16) & 0xff);
dst[3] = ((low >> 24) & 0xff);
dst[4] = ((low >> 32) & 0xff);
dst[5] = ((low >> 40) & 0xff);
dst[6] = ((low >> 48) & 0xff);
dst[7] = ((low >> 56) & 0xff);
// asm volatile ("" ::: "memory");
dst[8] = ((high >> 0) & 0xff);
dst[9] = ((high >> 8) & 0xff);
dst[10] = ((high >> 16) & 0xff);
dst[11] = ((high >> 24) & 0xff);
dst[12] = ((high >> 32) & 0xff);
dst[13] = ((high >> 40) & 0xff);
dst[14] = ((high >> 48) & 0xff);
dst[15] = ((high >> 56) & 0xff);
}
```
For aarch64, the vectorizer cost model says vectoring is better for some
reason.
