https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120647
Bug ID: 120647
Summary: [X86] Sub optimal code generated for counting the
number matches between two array elements
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: rtl-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: vekumar at gcc dot gnu.org
Target Milestone: ---
For the below test case
unsigned vector_comparison(const char * rhs, const char * lhs)
{
unsigned n = 0;
for (unsigned i = 0; i < 48; ++i)
if (lhs[i] == rhs[i])
++n;
return n;
}
GCC generated suboptimal code.
ref: https://godbolt.org/z/s1sxxq77s
When "pocnt" is available for ex. -O3 -march=znver5 we should be generating
simpler code.
vmovdqu (%rdi), %ymm0 ; Load first 32 bytes
vpcmpeqb (%rsi), %ymm0, %k0 ; Direct compare 32 bytes
vmovdqu 32(%rdi), %xmm0 ; Load next 16 bytes
vpcmpeqb 32(%rsi), %xmm0, %k1 ; Compare next 16 bytes
kmovd %k0, %eax ; Convert mask to register
popcntl %eax, %ecx ; Count matches (32 bytes)
kmovw %k1, %eax ; Convert second mask
popcntl %eax, %eax ; Count matches (16 bytes)
addl %ecx, %eax ; Sum results (total 48 bytes)
