Bug ID: 36212
           Summary: AVX-512: VPBROADCASTD from general purpose register is
           Product: libraries
           Version: 6.0
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Backend: X86

Hello, I notice that LLVM likes to emit instructions like:
mov r9d, ...
vpbroadcastd xmm0{k1}, r9d

On the other hand, broadcast from the xmm register should have better latency:
vpbroadcastd xmm0{k1}, xmm1

I'm not an expert, but from what I know it's better to avoid any gpr<>xmm
transition because it's penalized.

You are receiving this mail because:
You are on the CC list for the bug.
llvm-bugs mailing list

Reply via email to