https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103144
Bug ID: 103144
Summary: vectorizer failed to recognize shift>>=1 in loop as
shift>>i
Product: gcc
Version: 12.0
Status: UNCONFIRMED
Keywords: missed-optimization
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
#include<stdint.h>
void
foo1 (uint64_t* __restrict pdst, uint64_t* psrc, uint64_t shift)
{
for (int64_t i = 0; i != 64; i++)
{
uint64_t shift_i = shift >> i;
pdst[i] = psrc[i] + shift_i;
}
}
void
foo (uint64_t* __restrict pdst, uint64_t* psrc, uint64_t shift)
{
for (int64_t i = 0; i != 64; i++)
{
pdst[i] = psrc[i] + shift;
shift >>= 1;
}
}
gcc can vectorize foo1 but not foo since there's cross-iteration dep for shift
>>= 1, but shift >>= 1 is just shift >> i in loop.
Where should it be handled, can vect_recog_pattern handle this?
