https://gcc.gnu.org/bugzilla/show_bug.cgi?id=90836
Bug ID: 90836
Summary: Missing popcount pattern matching
Product: gcc
Version: 9.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
Target Milestone: ---
The following source:
int
foo (unsigned long long a)
{
unsigned long long b = a;
b -= ((b>>1) & 0x5555555555555555ULL);
b = ((b>>2) & 0x3333333333333333ULL) + (b & 0x3333333333333333ULL);
b = ((b>>4) + b) & 0x0F0F0F0F0F0F0F0FULL;
b *= 0x0101010101010101ULL;
return (int)(b >> 56);
}
implements a popcount operation. Some compilers can emit for x86:
xorq %rax, %rax
popcnt %rdi, %rax
ret
but GCC doesn't detect this and emits the sequence of masks and shifts.
This code is important for some workloads, 531.deepsjeng_r from SPEC, for
example.
Can this be matched in a match.pd pattern?
