https://gcc.gnu.org/bugzilla/show_bug.cgi?id=124032
Bug ID: 124032
Summary: Apply demorgan's rule to improve aliasing check
Product: gcc
Version: unknown
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: law at gcc dot gnu.org
Target Milestone: ---
Consider this on rv64gcbv_zicond at -O3 (from 525.x264_r):
#include <stdint.h>
void pixel_avg(uint8_t *dst,
uint8_t *src1,
uint8_t *src2)
{
for( int y = 0; y < 8; y++ )
dst[y] = (src1[y] + src2[y] + 1) >> 1;
}
The guard for the vector loop looks like:
sltiu a4,a4,7 # 10 [c=4 l=4] *sleu_didi
sltiu a5,a5,7 # 15 [c=4 l=4] *sleu_didi
seqz a4,a4 # 11 [c=4 l=4] *seq_zero_didi
seqz a5,a5 # 16 [c=4 l=4] *seq_zero_didi
and a5,a4,a5 # 21 [c=4 l=4] *anddi3/0
csrwi vxrm,0 # 118 [c=4 l=4] vxrmsi
beq a5,zero,.L2 # 22 [c=16 l=4] *branchdi
Note how we effectively have to invert the sltiu result (via seqz), feed those
into an AND then branch on equality.
It would be better to drop the seqz instructions and use IOR instead.
