https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116840
Bug ID: 116840
Summary: Optimise __builtin_parity for aarch64
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: enhancement
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
Target Milestone: ---
Target: aarch64
int
foob (unsigned char x)
{
return __builtin_parityg (x);
}
int
fooh (unsigned short x)
{
return __builtin_parityg (x);
}
int
foo (unsigned int x)
{
return __builtin_parity (x);
}
int
fooll (unsigned long x)
{
return __builtin_parityll (x);
}
For -O2 aarch64 we emit:
foob:
and x0, x0, 255
fmov d31, x0
cnt v31.8b, v31.8b
addv b31, v31.8b
fmov w0, s31
and w0, w0, 1
ret
fooh:
and x0, x0, 65535
fmov d31, x0
cnt v31.8b, v31.8b
addv b31, v31.8b
fmov w0, s31
and w0, w0, 1
ret
foo:
fmov s31, w0
cnt v31.8b, v31.8b
addv b31, v31.8b
fmov x0, d31
and w0, w0, 1
ret
fooll:
fmov d31, x0
cnt v31.8b, v31.8b
addv b31, v31.8b
fmov x0, d31
and w0, w0, 1
ret
For all functions but fooll this looks suboptimal. For the 32-bit foo LLVM
emits:
foo:
eor w8, w0, w0, lsr #16
eor w8, w8, w8, lsr #8
eor w8, w8, w8, lsr #4
eor w8, w8, w8, lsr #2
eor w8, w8, w8, lsr #1
and w0, w8, #0x1
ret
This is one instruction longer than the GCC sequence but should be preferable
as they are all 1-cycle instructions on many modern AArch64 cores and does not
involve any cross-bank moves or vector reductions. I think we should use this
sequence for (!TARGET_CSSC || !TARGET_SIMD) expansions of the paritym2 optab,
maybe gated on optimize_function_for_speed_p (cfun) as well.