https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114944
Alexander Monakov <amonakov at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |amonakov at gcc dot gnu.org
--- Comment #3 from Alexander Monakov <amonakov at gcc dot gnu.org> ---
Throughput-wise, the code in comment 2 has a significant bottleneck on port 5
on Haswell and Skylake (31 uops out of 70 go to port 5). Straightforward code
that does 16x movzx-movzx-movb for each byte should fare better, even
considering the load-store penalty for retrieving the vector from memory:
pand xmm1, XMMWORD PTR .LC0[rip]
movaps XMMWORD PTR [rsp-56], xmm0
movaps XMMWORD PTR [rsp-40], xmm1
movzx eax, BYTE PTR [rsp-40]
movzx eax, BYTE PTR [rsp-56+rax]
mov BYTE PTR [rsp-24], al
movzx eax, BYTE PTR [rsp-39]
movzx eax, BYTE PTR [rsp-56+rax]
mov BYTE PTR [rsp-23], al
movzx eax, BYTE PTR [rsp-38]
movzx eax, BYTE PTR [rsp-56+rax]
mov BYTE PTR [rsp-22], al
movzx eax, BYTE PTR [rsp-37]
movzx eax, BYTE PTR [rsp-56+rax]
mov BYTE PTR [rsp-21], al
movzx eax, BYTE PTR [rsp-36]
movzx eax, BYTE PTR [rsp-56+rax]
mov BYTE PTR [rsp-20], al
movzx eax, BYTE PTR [rsp-35]
movzx eax, BYTE PTR [rsp-56+rax]
mov BYTE PTR [rsp-19], al
movzx eax, BYTE PTR [rsp-34]
movzx eax, BYTE PTR [rsp-56+rax]
mov BYTE PTR [rsp-18], al
movzx eax, BYTE PTR [rsp-33]
movzx eax, BYTE PTR [rsp-56+rax]
mov BYTE PTR [rsp-17], al
movzx eax, BYTE PTR [rsp-32]
movzx eax, BYTE PTR [rsp-56+rax]
mov BYTE PTR [rsp-16], al
movzx eax, BYTE PTR [rsp-31]
movzx eax, BYTE PTR [rsp-56+rax]
mov BYTE PTR [rsp-15], al
movzx eax, BYTE PTR [rsp-30]
movzx eax, BYTE PTR [rsp-56+rax]
mov BYTE PTR [rsp-14], al
movzx eax, BYTE PTR [rsp-29]
movzx eax, BYTE PTR [rsp-56+rax]
mov BYTE PTR [rsp-13], al
movzx eax, BYTE PTR [rsp-28]
movzx eax, BYTE PTR [rsp-56+rax]
mov BYTE PTR [rsp-12], al
movzx eax, BYTE PTR [rsp-27]
movzx eax, BYTE PTR [rsp-56+rax]
mov BYTE PTR [rsp-11], al
movzx eax, BYTE PTR [rsp-26]
movzx eax, BYTE PTR [rsp-56+rax]
mov BYTE PTR [rsp-10], al
movzx eax, BYTE PTR [rsp-25]
movzx eax, BYTE PTR [rsp-56+rax]
mov BYTE PTR [rsp-9], al
movdqa xmm0, XMMWORD PTR [rsp-24]
If you want to avoid the load-store forwarding stall, perhaps you can assemble
two halves of the shuffled vector on GPRs (e.g. do 'movzx ecx, byte[...]; shl
eax, 8; mov al, byte [...+rcx]), then merge two 64-bit GPRs into one 128-bit
vector.
