https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112935
Bug ID: 112935
Summary: [14 Regression] Performance regression in Coremarks
crcu8 function
Product: gcc
Version: 14.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: xry111 at gcc dot gnu.org
Target Milestone: ---
typedef __UINT8_TYPE__ ee_u8;
typedef __UINT16_TYPE__ ee_u16;
ee_u16 crcu8(ee_u8 data, ee_u16 crc) {
ee_u8 i = 0, x16 = 0, carry = 0;
for (i = 0; i < 8; i++) {
x16 = (ee_u8)((data & 1) ^ ((ee_u8)crc & 1));
data >>= 1;
if (x16 == 1) {
crc ^= 0x4002;
carry = 1;
} else
carry = 0;
crc >>= 1;
if (carry)
crc |= 0x8000;
else
crc &= 0x7fff;
}
return crc;
}
With GCC 13.2.0 -O2, on LoongArch we get:
.L2:
xor $r12,$r4,$r14
andi $r12,$r12,1
sub.w $r12,$r0,$r12
srli.w $r4,$r4,1
and $r12,$r12,$r15
addi.w $r13,$r13,-1
xor $r12,$r12,$r4
bstrpick.w $r13,$r13,7,0
srli.d $r14,$r14,1
bstrpick.w $r4,$r12,15,0
bnez $r13,.L2
With GCC 14.0.0 -O2:
.L2:
xor $r12,$r4,$r14
andi $r12,$r12,1
mul.w $r12,$r12,$r15
srli.w $r4,$r4,1
addi.w $r13,$r13,-1
bstrpick.w $r13,$r13,7,0
srli.d $r14,$r14,1
xor $r12,$r12,$r4
bstrpick.w $r4,$r12,15,0
bnez $r13,.L2
mul.w is slower than sub.w + and.
I'm now setting components to tree-optimization because the difference already
exists in 254t.optimized vs 263t.optimized. But maybe the tree optimizer is
doing things correctly and we should just add a target-specific optimization.
