https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91201
--- Comment #6 from Joel Yliluoma <bisqwit at iki dot fi> ---
Maybe a horizontal checksum is a bit obscure term. A 8-bit checksum is what is
being accomplished, nonetheless. Yes, there are simpler ways to do it…
But I tried a number of different approaches in order to try and get maximum
performance SIMD code out of GCC, and I came upon this curious case that I
posted this bugreport about.
To another compiler, I reported a related bug concerning a code that looks like
this:
unsigned char calculate_checksum(const void* ptr)
{
unsigned char bytes[16], result = 0;
memcpy(bytes, ptr, 16);
// The reason the memcpy is there in place is because to
// my knowledge, it is the only _safe_ way permitted by
// the standard to do conversions between representations.
// Union, pointer casting, etc. are not safe.
for(unsigned n=0; n<16; ++n) result += bytes[n];
return result;
}
After my report, their compiler now generates:
vmovdqu xmm0, xmmword ptr [rdi]
vpshufd xmm1, xmm0, 78 # xmm1 = xmm0[2,3,0,1]
vpaddb xmm0, xmm0, xmm1
vpxor xmm1, xmm1, xmm1
vpsadbw xmm0, xmm0, xmm1
vpextrb eax, xmm0, 0
ret
This is what GCC generates for the same code.
vmovdqu xmm0, XMMWORD PTR [rdi]
vpsrldq xmm1, xmm0, 8
vpaddb xmm0, xmm0, xmm1
vpsrldq xmm1, xmm0, 4
vpaddb xmm0, xmm0, xmm1
vpsrldq xmm1, xmm0, 2
vpaddb xmm0, xmm0, xmm1
vpsrldq xmm1, xmm0, 1
vpaddb xmm0, xmm0, xmm1
vpextrb eax, xmm0, 0
ret
So the bottom line is, (v)psadbw reductions should be added as M. Glisse
correctly indicated.
